{"id": "2506.14990", "title": "MEAL: A Benchmark for Continual Multi-Agent Reinforcement Learning", "abstract": "Benchmarks play a central role in reinforcement learning (RL) research, yet their computational constraints often shape what is studied. Despite the motivation of lifelong learning, most continual RL papers consider only 3-10 sequential tasks, as CPU-bound environments make longer sequences impractical. Meanwhile, continual learning in cooperative multi-agent settings remains largely unexplored. To address these gaps, we introduce MEAL (Multi-agent Environments for Adaptive Learning), the first benchmark for continual multi-agent RL. By leveraging JAX and GPU acceleration, MEAL enables training on sequences of 100 tasks in a few hours on a single GPU. We find that long task sequences reveal failure modes that do not appear at smaller scales.", "date": "2025-06-17", "categories": []}
{"id": "2507.23534", "title": "Continual Learning with Support Boundary Experience Blending", "abstract": "Continual learning (CL) seeks to mitigate catastrophic forgetting when models are trained with sequential tasks. A common approach, experience replay (ER), stores past exemplars but only sparsely approximates the data distribution, yielding fragile and oversimplified decision boundaries. We address this limitation by introducing Support Boundary Data (SBD), generated via differential-privacy-inspired noise into latent features to create boundary-adjacent representations that implicitly regularize decision boundaries. Building on this idea, we propose Experience Blending (EB), a framework that jointly trains on exemplars and SBD through a dual-model aggregation strategy. EB has two components: (1) latent-space noise injection to generate support boundary data, and (2) end-to-end training that jointly leverages exemplars and SBD. Unlike standard experience replay, SBD enriches the feature space near decision boundaries, leading to more stable and robust continual learning. Extensive experiments on CIFAR-10, CIFAR-100, Tiny ImageNet, and ImageNet1K demonstrate consistent accuracy improvements of 10%, 6%, 14%, 2%, respectively.", "date": "2025-07-31", "categories": []}
{"id": "2603.07236", "title": "HY-WU (Part I): An Extensible Functional Neural Memory Framework and An Instantiation in Text-Guided Image Editing", "abstract": "Foundation models are transitioning from offline predictors to deployed systems expected to operate over long time horizons. In real deployments, objectives are not fixed: domains drift, user preferences evolve, and new tasks appear after the model has shipped. This elevates continual learning and instant personalization from optional features to core architectural requirements. Yet most adaptation pipelines still follow a static weight paradigm: after training (or after any adaptation step), inference executes a single parameter vector regardless of user intent, domain, or instance-specific constraints. This treats the trained or adapted model as a single point in parameter space. In heterogeneous and continually evolving regimes, distinct objectives can induce separated feasible regions over parameters, forcing any single shared update into compromise, interference, or overspecialization. As a result, continual learning and personalization are often implemented as repeated overwriting of shared weights, risking degradation of previously learned behaviors. We propose HY-WU (Weight Unleashing), a memory-first adaptation framework that shifts adaptation pressure away from overwriting a single shared parameter point. HY-WU implements functional (operator-level) memory as a neural module: a generator that synthesizes weight updates on-the-fly from the instance condition, yielding instance-specific operators without test-time optimization.", "date": "2026-03-07", "categories": []}
{"id": "2606.20431", "title": "Sparsity, Superposition, and Forgetting: A Mechanistic Study of Representation Retention in Continual Learning", "abstract": "Continual learning (CL) systems often forget previously acquired knowledge, yet the mechanisms driving forgetting remain hard to isolate in practice because real datasets entangle many factors. We present a controlled, toy-world framework that makes these mechanisms observable and testable. Using a synthetic generator-separator pipeline, we define ground-truth latent features, build tasks with tunable sparsity and overlap, and introduce measurable quantities for representation strength and superposition (directional overlap among features). We then study retention dynamics-the temporal change of representation strength by fitting sparse dynamical relations (via SINDy) between retention, superposition, and exposure history. A complementary task-level analysis based on effective rank characterizes how representational capacity is allocated across tasks. Our controlled experiments yield three takeaways. (1) Superposition tends to increase over time with transient dips at task boundaries, suggesting boundary-specific interference rather than steady drift. (2) Higher feature sparsity induces more superposition yet does not inevitably cause forgetting; when representations remain strong, forgetting can be reduced despite overlap. (3) Task-level effective rank grows with sparsity, indicating broader capacity usage under sparse regimes. Together, these results nuance the common intuition that more superposition leads to more forgetting by showing that overlap interacts with representation strength and capacity allocation. Our toy analysis provides falsifiable hypotheses and diagnostic tools for CL.", "date": "2026-06-18", "categories": []}
{"id": "2602.09234", "title": "Do Neural Networks Lose Plasticity in a Gradually Changing World?", "abstract": "Continual learning has become a trending topic in machine learning. Recent studies have discovered an interesting phenomenon called loss of plasticity, referring to neural networks gradually losing the ability to learn new tasks. However, existing plasticity research largely relies on benchmarks with abrupt task transitions, without examining whether the abruptness itself contributes to the observed plasticity loss. In this paper, we investigate the role of transition abruptness by simulating gradually changing environments through input/output interpolation and task sampling. We perform theoretical and empirical analysis, showing that the severity of plasticity loss is closely tied to the abruptness of task transitions, and can be substantially reduced when the environment changes gradually.", "date": "2026-02-09", "categories": []}
{"id": "2602.03846", "title": "PLATE: Plasticity-Tunable Efficient Adapters for Geometry-Aware Continual Learning", "abstract": "We develop a continual learning method for pretrained models that \\emph{requires no access to old-task data}, addressing a practical barrier in foundation model adaptation where pretraining distributions are often unavailable. Our key observation is that pretrained networks exhibit substantial \\emph{geometric redundancy}, and that this redundancy can be exploited in two complementary ways. First, redundant neurons provide a proxy for dominant pretraining-era feature directions, enabling the construction of approximately protected update subspaces directly from pretrained weights. Second, redundancy offers a natural bias for \\emph{where} to place plasticity: by restricting updates to a subset of redundant neurons and constraining the remaining degrees of freedom, we obtain update families with reduced functional drift on the old-data distribution and improved worst-case retention guarantees. These insights lead to \\textsc{PLATE} (\\textbf{Pla}sticity-\\textbf{T}unable \\textbf{E}fficient Adapters), a continual learning method requiring no past-task data that provides explicit control over the plasticity-retention trade-off. PLATE parameterizes each layer with a structured low-rank update $\\Delta W = B A Q^\\top$, where $B$ and $Q$ are computed once from pretrained weights and kept frozen, and only $A$ is trained on the new task. The code is available at https://github.com/SalesforceAIResearch/PLATE.", "date": "2026-02-03", "categories": []}
{"id": "2603.01761", "title": "Position: Modular Memory is the Key to Continual Learning Agents", "abstract": "Foundation models have transformed machine learning through large-scale pretraining and increased test-time compute. Despite surpassing human performance in several domains, these models remain fundamentally limited in continuous operation, experience accumulation, and personalization, capabilities that are central to adaptive intelligence. While continual learning research has long targeted these goals, its historical focus on in-weight learning (IWL), i.e., updating a single model's parameters to absorb new knowledge, has rendered catastrophic forgetting a persistent challenge. Our position is that combining the strengths of In-Weight Learning (IWL) and the newly emerged capabilities of In-Context Learning (ICL) through the design of modular memory is the missing piece for continual adaptation at scale. We outline a conceptual framework for modular memory-centric architectures that leverage ICL for rapid adaptation and knowledge accumulation, and IWL for stable updates to model capabilities, charting a practical roadmap toward continually learning agents.", "date": "2026-03-02", "categories": []}
{"id": "2606.13258", "title": "MOSAIC: Modality-Specific Adaptation for Incremental Continual Learning in Parkinson's Disease Gait Assessment", "abstract": "Gait-based Parkinson's disease assessment increasingly relies on heterogeneous sensors, but clinical systems rarely collect all modalities simultaneously. New sensors may arrive through device upgrades, protocol changes, or multi-center deployment, while historical patient data are often unavailable because of privacy and storage constraints. This modality-incremental setting faces three challenges: unreliable cross-modal distillation, modality-specific statistical shifts, and reduced plasticity after preservation. We propose MOSAIC, a compact continual learning framework. First, we identify the Toxic Teacher phenomenon and introduce Modality-Specific Warm-Up to stabilize newly learned modality representations before distillation. Second, we propose a statistics-decoupled MSBN architecture that isolates sensor statistics while maintaining a shared semantic backbone. Third, we design a curriculum-guided repulsive objective for Plasticity Recovery, preserving legacy knowledge while recovering modality-specific capacity. Experiments on three multimodal Parkinson's gait datasets show that MOSAIC improves final performance and mitigates forgetting. Project code is available at: https://github.com/minlinzeng/MOSAIC_Modality-Specific-Adaptation-for-Incremental-Continual-Learning-in-PD-Gait-Assessment.git", "date": "2026-06-11", "categories": []}
{"id": "2606.17803", "title": "Continual Self-Improvement with Lightweight Experiential Latent Memories", "abstract": "Large language models achieve strong reasoning performance by scaling inference-time compute, yet remain fundamentally stateless, discarding the rich, self-produced reasoning traces generated during this process. We investigate whether models can instead learn online from this experience, converting transient computation (reasoning traces) into persistent reusable knowledge, and without external supervision or access to future data. We show that In-Context Learning (ICL) over raw reasoning traces fails to generalize, reflecting a fundamental limitation of token-level reuse: individual traces lack the abstraction needed for transfer, even after refinement (e.g. self-reflection). In contrast, drawing inspiration from recent works on unsupervised reinforcement learning, we find that lightweight per-instance training with self-generated test-time signals (majority voting) as rewards yields substantial gains, often surpassing full-dataset offline training, motivating a shift from raw traces to learned latent representations. Building on this insight, we propose an online method that distills inference-time compute spent on encountered problems into compact modular latent memories capturing the underlying reasoning structure. These memories are stored and retrieved for future inputs, enabling continual improvement while avoiding catastrophic forgetting through modular design. Importantly, our method is highly efficient, parametrized as extremely lightweight soft prompt memories (~0.001% of model parameters) and trained with only a few gradient steps, yet achieving performance competitive with full parametric updates and offline training. Across challenging mathematical reasoning benchmarks, our approach significantly outperforms zero-shot and raw data ICL baselines, while transferring effectively across datasets.", "date": "2026-06-16", "categories": []}
{"id": "2606.17889", "title": "Dimensionality Controls When Modularity Helps in Continual Learning", "abstract": "Compositional learning systems must balance plasticity, the ability to acquire new knowledge, with stability, the preservation of previously learned components, especially when tasks share structure and risk interference. We study how modular architecture, task similarity, and representational dimensionality jointly shape compositional continual learning in a sequential A-B-A paradigm, comparing a task-partitioned recurrent network to a single-network baseline while inducing high- and low-dimensional regimes via weight-scale manipulations. In a high-dimensional \"lazy\" regime, both architectures achieve similar performance and internal geometry, suggesting that explicit modular structure has little impact when representations are weakly constrained. In a lower-dimensional \"rich\" regime, modularity becomes decisive: the modular network develops graded task-specific subspaces that overlap for similar tasks, partially align for moderately dissimilar tasks, and separate for dissimilar tasks, yielding a more compositional and interpretable organization than the single network. These findings identify the representational regime induced by initialization scale, which co-varies with representational dimensionality, as a key factor governing when compositional, modular structure is functionally beneficial in continual learning, and support viewing safety and robustness as problems of adaptive allocation of representational subspaces rather than fixed separation versus sharing.", "date": "2026-06-16", "categories": []}
{"id": "2606.18003", "title": "C2FL: Clustered Continual Federated Learning under Spatial and Temporal Drift", "abstract": "Collective Adaptive Systems (CAS) increasingly rely on machine learning to let each node learn from locally sensed data, aligning its behavior with the surrounding environment. Scaling this intelligence, however, raises fundamental challenges: sensed data is often privacy-sensitive, preventing centralized collection; nodes are mobile, traversing regions where nearby nodes perceive similar phenomena while distant ones observe radically different conditions, creating natural spatial clusters; and these distributions evolve over time due to mobility, introducing temporal drift that makes local models progressively stale. These dynamics arise across domains - vehicular sensing, drone-based monitoring, smartphone crowdsensing - yet the interplay of privacy, spatial heterogeneity, and temporal drift severely undermines conventional learning strategies. Therefore, we propose C2FL, a fully distributed Federated Learning (FL) approach where nodes self-organize into learning groups through spatial clustering, reflecting the geographic structure of the environment. To counteract temporal drift, each node combines experience replay with a dwell-time-aware adaptive averaging step, progressively incorporating the regional consensus as it remains longer within the same area, while preserving previously acquired knowledge under evolving distributions. We evaluate our approach on synthetic experiments that systematically reproduce spatial and temporal shifts, showing that standard federated strategies degrade significantly under these conditions and that our method restores robust collective adaptation.", "date": "2026-06-16", "categories": []}
{"id": "2606.18024", "title": "Catastrophic Forgetting is Low-Rank: A Function-Space Theory for Continual Adaptation", "abstract": "Catastrophic forgetting in continual adaptation is usually studied through parameter drift, replay, or distillation, but these views do not identify which output-space directions are vulnerable. We give a function-space account in the NTK regime: new-task training induces old-task prediction drift through the cross-task kernel, yielding a closed-form predictor for the forgetting vector before any new-task gradient step. In frozen-backbone linear-head PEFT-CL, where the model is linear in the trainable parameters, the predictor is exact up to numerical precision; for nonlinear adapters/full fine-tuning, it is a local NTK approximation. The same expression reveals that forgetting concentrates in a small number of old-task NTK eigenmodes and under frozen linear heads gives a Kronecker scaling rule for the vulnerable rank. These results clarify the relation to prior NTK-overlap theory, explain why parameter-space regularizers can miss output-space interference, and motivate a targeted spectral regularizer.", "date": "2026-06-16", "categories": []}
{"id": "2512.18295", "title": "AL-GNN: Privacy-Preserving and Replay-Free Continual Graph Learning via Analytic Learning", "abstract": "Continual graph learning (CGL) aims to enable graph neural networks to incrementally learn from a stream of graph structured data without forgetting previously acquired knowledge. Existing methods particularly those based on experience replay typically store and revisit past graph data to mitigate catastrophic forgetting. However, these approaches pose significant limitations, including privacy concerns, inefficiency. In this work, we propose AL GNN, a novel framework for continual graph learning that eliminates the need for backpropagation and replay buffers. Instead, AL GNN leverages principles from analytic learning theory to formulate learning as a recursive least squares optimization process. It maintains and updates model knowledge analytically through closed form classifier updates and a regularized feature autocorrelation matrix. This design enables efficient one pass training for each task, and inherently preserves data privacy by avoiding historical sample storage. Extensive experiments on multiple dynamic graph classification benchmarks demonstrate that AL GNN achieves competitive or superior performance compared to existing methods. For instance, it improves average performance by 10% on CoraFull and reduces forgetting by over 30% on Reddit, while also reducing training time by nearly 50% due to its backpropagation free design.", "date": "2025-12-20", "categories": []}
{"id": "2601.18699", "title": "Mechanistic Analysis of Catastrophic Forgetting in Large Language Models During Continual Fine-tuning", "abstract": "Sequential fine-tuning of Large Language Models (LLMs) adaptation to target tasks often triggers catastrophic forgetting, where the acquisition of novel target skills degrades ancestral capabilities. This paper presents a systematic comparative study of catastrophic forgetting across twenty premier models representing the state-of-the-art in mid-2026. We categorize our investigation into two primary research lines: (i) a behavioral and semantic output drift analysis of ten leading closed-source models (including Claude Fable 5, GPT-5.5 High, and Gemini 3.5 Flash), and (ii) a deep mechanistic interpretation of ten prominent open-weight architectures (such as DeepSeek-V4-Pro, Llama 4 Maverick, and Qwen 3.6-27B). Through weight-space trajectory tracking, Centered Kernel Alignment (CKA), and routing gate drift calculations in Mixture-of-Experts (MoE) layers, we localize the neural circuits highly susceptible to parameter overwriting. Our findings indicate that early-layer attention heads exhibit systemic entropic dispersion, while mid-to-deep feed-forward networks (or sparse expert blocks) suffer localized representation collapse. Informed by these insights, we introduce Low-Rank Circuit Projection (LRCP), a subspace-regularized training intervention. Empirical evaluations show that LRCP successfully mitigates up to 94.2% of ancestral capabilities in open-weight configurations and matches the adaptation velocity of standard PEFT baselines.", "date": "2026-01-26", "categories": []}
{"id": "2603.24350", "title": "Evidence of an Emergent \"Self\" in Continual Robot Learning", "abstract": "A key challenge to understanding self-awareness has been a principled way of quantifying whether an intelligent system has a concept of a \"self\", and if so how to differentiate the \"self\" from other cognitive structures. We propose that the \"self\" can be isolated by seeking the invariant portion of cognitive process that changes relatively little compared to more rapidly acquired cognitive skills - because our self is the most persistent aspect of our experiences. We used this principle to analyze the cognitive structure of robots under two conditions: One robot learns a constant task, while a second undergoes continual learning under variable tasks. We find that robots subjected to continual learning develop an invariant subnetwork that is significantly more stable (p < 0.001) compared to the control, and that this subnetwork is also functionally important: preserving it aids adaptation while damaging it impairs performance. We validate this pattern across three different robots spanning locomotion and manipulation.", "date": "2026-03-25", "categories": []}
{"id": "2604.13085", "title": "Adaptive Memory Crystallization for Autonomous AI Agent Learning in Dynamic Environments", "abstract": "Autonomous AI agents operating in dynamic environments face a persistent challenge: acquiring new capabilities without erasing prior knowledge. We present Adaptive Memory Crystallization (AMC), a memory architecture for progressive experience consolidation in continual reinforcement learning.\n  AMC is conceptually inspired by the qualitative structure of synaptic tagging and capture (STC) theory, the idea that memories transition through discrete stability phases, but makes no claim to model the underlying molecular or synaptic mechanisms.\n  AMC models memory as a continuous crystallization process in which experiences migrate from plastic to stable states according to a multi-objective utility signal. The framework introduces a three-phase memory hierarchy (Liquid--Glass--Crystal) governed by an It\\^o stochastic differential equation (SDE) whose population-level behavior is captured by an explicit Fokker--Planck equation admitting a closed-form Beta stationary distribution.\n  We provide proofs of: (i) well-posedness and global convergence of the crystallization SDE to a unique Beta stationary distribution; (ii) exponential convergence of individual crystallization states to their fixed points, with explicit rates and variance bounds; and (iii) end-to-end Q-learning error bounds and matching memory-capacity lower bounds that link SDE parameters directly to agent performance.\n  Empirical evaluation on Meta-World MT50, Atari 20-game sequential learning, and MuJoCo continual locomotion consistently shows improvements in forward transfer (+34--43\\% over the strongest baseline), reductions in catastrophic forgetting (67--80\\%), and a 62\\% decrease in memory footprint.", "date": "2026-04-02", "categories": []}
{"id": "2606.14731", "title": "BBR-Net: Boundary-Balanced Replay for Continual Medical Image Segmentation", "abstract": "Continual learning for medical image segmentation remains challenging under domain shift because replay-based methods often preserve appearance information without explicitly modeling anatomical structure. This study investigates whether structural consistency governs knowledge retention in continual cardiac ultrasound segmentation. We propose the Boundary-Balanced Replay Network (BBR-Net), which selects replay samples using boundary-aware priority and class balance to preserve anatomically informative regions. The method is evaluated on CAMUS and CardiacNet under forward (CAMUS to CardiacNet) and reverse (CardiacNet to CAMUS) task orders. In the forward setting, BBR-Net retains source-task performance close to an offline joint-training reference, while markedly reducing catastrophic forgetting and preserving competitive target-task adaptation. Ablation results show that boundary-aware prioritization contributes to retention and improves the balance between source-task preservation and target-task adaptation when combined with class-aware sampling. In contrast, the reverse setting reveals that structure-aware replay fails when initial representations are learned from noisy and structurally inconsistent data. To isolate this effect, we conduct a controlled structural perturbation analysis by progressively corrupting source-task boundaries while keeping the dataset, architecture, and training protocol fixed. Forgetting increases consistently as structural reliability decreases, suggesting that replay effectiveness is strongly influenced by the quality of stored structural information, rather than by memory capacity alone. These findings indicate that preserving anatomical structure under domain shift is a central factor in continual medical image segmentation, and that replay mechanisms should account for structural reliability to support robust knowledge retention.", "date": "2026-06-02", "categories": []}
{"id": "2606.14883", "title": "Understanding Cross-Modal Contributions in Continual Vision-Language Models: A Theoretical Perspective", "abstract": "Continual vision-language models are commonly addressed through sequential fine-tuning; however, although this paradigm enables adaptation to new environments (tasks), it inherently emphasizes the contribution of previously learned environments (tasks) at the expense of the stability required to preserve previously acquired knowledge. While existing approaches have adequately studied continual learning and catastrophic forgetting in vision-language models (VLMs), the theoretical understanding of modality-specific contributions across a sequence of environments remains largely unexplored. In this paper, we present a new theoretical perspective to understand the cross-modal (vision-language) contributions to consecutive environments. We empirically evaluate our theoretical findings on large VLMs and demonstrate their effectiveness in capturing environment-level cross-modal contributions. Our analysis provides deeper insights into continual VLMs, highlighting their contribution robustness to varying task orders and inter-task similarities, and their improved generalization performance.", "date": "2026-06-12", "categories": []}
{"id": "2606.14987", "title": "Continual Backdoor Training in IoT/CPS", "abstract": "Internet of Things (IoT) and Cyber-physical systems (CPS) increasingly rely on continual learning (CL) to adapt to evolving environments, device heterogeneity, and concept drift, thereby improving overall utility. While continual adaptation is essential for long-lived IoT deployments where data patterns evolve, it also introduces new security vulnerabilities. In particular, backdoor attacks can exploit incremental updates, replay buffers, and representation reuse to implant persistent malicious behaviors that remain dormant during normal operation but activate upon specific triggers. In this paper, we present a backdoor attack in continual learning used in IoT/CPS systems. To this end, we formalize an IoT/CPS-specific threat model, analyze why continual learning amplifies backdoor persistence in IoT pipelines, and evaluate our technique under varying conditions. Our analysis highlights critical open challenges in securing lifelong learning in IoT/CPS and industrial IoT (IIoT) environments, as well as the need for heightened security controls.", "date": "2026-06-12", "categories": []}
{"id": "2606.15253", "title": "Focus, Align, and Sustain: Counteracting Gradient Dilution in Incremental Object Detection", "abstract": "Adapting Detection Transformers to Incremental Object Detection (IOD) poses a systemic challenge, as set-based optimization is inherently destabilized by sequential learning. In this work, we identify Gradient Dilution as the root cause of performance degradation, wherein optimization signals required to preserve old knowledge are progressively weakened. This phenomenon manifests as a cascading erosion of preservation gradients in magnitude, direction, and support coverage, driven by three tightly coupled factors: Signal Dispersion, where foreground gradients are overwhelmed by background noise; Assignment Drift, where stochastic query-target matching induces inconsistent gradient trajectories; and Support Attrition, where gradients from retained samples insufficiently cover the old-class feature space, weakening decision boundaries under interference from new classes. To counteract this, we propose FAS, a unified framework that Focuses, Aligns, and Sustains gradient flow throughout incremental learning. Specifically, we introduce prior-injected queries to focus discriminative signals by filtering background interference at the source. We further propose deterministic anchor distillation to align query-target assignments and enforce semantic consistency across stages under unstable matching. Finally, we devise manifold-support replay to sustain distributional support of old classes, counteracting representational erosion induced by continual updates. Extensive experiments show that FAS restores robust optimization dynamics and outperforms state-of-the-art methods, achieving over 5.0 AP improvement in the challenging 40+10x4 incremental setting.", "date": "2026-06-13", "categories": []}
{"id": "2606.15306", "title": "LatentGym: A Testbed For Cross-Task Experiential Learning With Controllable Latent Structure", "abstract": "We envision continually learning agentic systems that become more useful over time: as they encounter sequences of related tasks, they should infer the hidden structure shared across those tasks and use it to improve future decisions. This cross-task experiential learning capability is pivotal in domains such as personalization and interactive assistance, but existing training/evaluation frameworks do not provide shared, controllable latent structures and cannot measure whether or why agents improve. We introduce LatentGym: a controllable suite in which each environment is organized around a ground-truth latent variable governing the structure across tasks. Our construction yields metrics that separate exploration (whether the agent's actions gather information about the latent) from exploitation (whether the agent uses what it has gathered). We demonstrate our suite on empirical studies addressing three questions: how and why frontier models fail to adapt across related tasks; whether post-training on related task sequences improves general cross-task adaptation, and where those gains come from; and how design choices such as inter-task feedback shape training dynamics and generalization. Together, these results establish a controlled foundation for studying how LLM agents learn from experience across tasks, and for designing agents that adapt more reliably in sequential, personalized, and interactive settings.", "date": "2026-06-13", "categories": []}
{"id": "2606.15685", "title": "Learning New Tasks via Reusable Skills: Skill-Compositional Experts for Embodied Continual Learning", "abstract": "Embodied Continual Learning (ECL) aims to enable robots to continually acquire new manipulation tasks while retaining previously learned behaviors under closed-loop control. Compared with conventional continual learning, ECL suffers from more severe catastrophic forgetting. Feature drift accumulated under closed-loop control progressively propagates through sequential decision-making, leading to degradation of previously learned behaviors. A key challenge in ECL lies in structured skill reuse across continually evolving tasks, since existing methods primarily focus on skill learning without explicitly organizing them for coherent task execution. To address this issue, we propose SCE, a Skill-Compositional Experts framework for ECL. SCE builds a skill base via Compositional Skill Grounding (CSG), which decomposes task demonstrations into reusable skills. Based on this, Dual Execution-and-Transition Experts (DETE) enable new task learning through skill composition, where one branch ensures skill execution and the other supports transitions between skills for coherent behavior. Experiments on LIBERO benchmarks and real-world manipulation tasks demonstrate that SCE consistently improves retention and overall task performance. Further feature drift analyses and ablation studies verify the effectiveness of our method. Project website: https://eqcy.github.io/sce/.", "date": "2026-06-14", "categories": []}
{"id": "2606.15695", "title": "When Generator Replay Degrades: Projected Rehearsal Orchestration for Heterogeneous Federated Class-Incremental Learning", "abstract": "Federated class-incremental learning (FCIL) becomes substantially harder when clients observe different label subsets, progress through tasks at different stages, and provide uneven supervision for the same semantic concepts. Existing FCIL methods often preserve old knowledge through input-space synthesis, but they can be fragile under heterogeneous task streams and difficult to transfer across modalities. To alleviate such issues, we propose PRO, a framework that replaces synthetic input replay with projected rehearsal orchestration. To remove external pretraining, we evaluate all methods under the same warmup. After this, PRO maintains compact class-level projected memories on the server and allows clients perform balanced pseudo multi-task training over current examples and old projected memories. To handle stronger representation drift, we further introduce PRO-MAX, which augments PRO with neighborhood-weighted memory alignment while preserving the same server-light principle that the server only aggregates model updates and memory statistics. Across image, text, and graph benchmarks, PRO and PRO-MAX improve retention and final utility under heterogeneous streams while remaining competitive in homogeneous FCIL. Even when baselines are given expanded replay budgets, they degrade under supervision imbalance and stage misalignment, indicating that replay quantity alone does not resolve replay-quality failures. Additional weak-task diagnostics further show that larger replay mismatch is associated with larger downstream degradation, while our method keeps projected memories better aligned with the evolving representation.", "date": "2026-06-14", "categories": []}
{"id": "2606.15734", "title": "Retrievable Gradients: Continual Post-Training Without Cumulative Weight Drift", "abstract": "Continual post-training enables models to absorb emerging knowledge after deployment, but repeatedly updating shared parameters can accumulate weight drift, potentially causing catastrophic forgetting and degrading general capabilities. Retrieval-augmented generation avoids such parameter drift, yet often lacks the depth of parametric knowledge integration. In this paper, we propose ReGrad (Retrievable Gradients), a new paradigm that treats gradients as retrievable units of knowledge. ReGrad pre-computes document-specific gradients offline, stores them in an indexed Gradient Bank, and retrieves only query-relevant gradients at inference time for temporary weight adaptation. However, raw language-modeling gradients are optimized for token-level document reconstruction rather than for query-driven knowledge use. We therefore introduce a bi-level meta-learning objective that reshapes document-derived gradients into generalizable adaptation signals for downstream tasks. Experiments across general and domain-specific settings show that \\textsc{ReGrad} outperforms CPT and RAG baselines, enabling scalable and reversible parametric knowledge injection without accumulating weight drift.", "date": "2026-06-14", "categories": []}
{"id": "2606.16256", "title": "KeepLoRA++: Continual Learning with Layer-Scaled Residual Gradient Adaptation", "abstract": "Continual learning for pre-trained vision-language models requires balancing three competing objectives: retaining pre-trained knowledge, preserving knowledge from a sequence of learned tasks, and maintaining the plasticity to acquire new knowledge. This paper presents KeepLoRA++, balancing these objectives through a unified dual-dimensional knowledge retention mechanism. We analyze knowledge distribution of Transformer architecture from both inter-layer and intra-layer perspectives. The inter-layer perspective examines how retention is distributed across layers, while the intra-layer perspective focuses on the parameter space within each layer. Our analysis reveals a structural property: general transferable knowledge is mainly encoded in the shallow layers and the principal subspace of the parameters, while task-specific adaptations are localized in the deep layers and the residual subspace. Motivated by this insight, KeepLoRA++ introduces a layer-scaled residual gradient adaptation method. New tasks are learned by restricting LoRA parameter updates to the residual subspace, combined with a shallow-to-deep layer scaling, to prevent interference with previously acquired capabilities. Specifically, the gradient of a new task is projected onto a subspace orthogonal to both the principal subspace of the pre-trained model and the dominant directions of previous task features, while simultaneously assigning smaller update magnitudes to shallow layers and larger ones to deeper layers. Our theoretical analysis and empirical evaluations confirm that KeepLoRA++ successfully balances these three competing objectives, consistently outperforming representative baselines across image classification, visual question answering, and video understanding tasks.", "date": "2026-06-15", "categories": []}
{"id": "2306.01690", "title": "Context selectivity with dynamic availability enables lifelong continual learning", "abstract": "\"You never forget how to ride a bike\", -- but how is that possible? The brain is able to learn complex skills, stop the practice for years, learn other skills in between, and still retrieve the original knowledge when necessary. The mechanisms of this capability, referred to as lifelong learning (or continual learning, CL), are unknown. We suggest a bio-plausible meta-plasticity rule building on classical work in CL which we summarize in two principles: (i) neurons are context selective, and (ii) a local availability variable partially freezes the plasticity if the neuron was relevant for previous tasks. In a new neuro-centric formalization of these principles, we suggest that neuron selectivity and neuron-wide consolidation is a simple and viable meta-plasticity hypothesis to enable CL in the brain. In simulation, this simple model balances forgetting and consolidation leading to better transfer learning than contemporary CL algorithms on image recognition and natural language processing CL benchmarks.", "date": "2023-06-02", "categories": []}
{"id": "2603.11395", "title": "ARROW: Augmented Replay for RObust World models", "abstract": "Continual reinforcement learning challenges agents to acquire new skills while retaining previously learned ones with the goal of improving performance in both past and future tasks. Most existing approaches rely on model-free methods with replay buffers to mitigate catastrophic forgetting; however, these solutions often face significant scalability challenges due to large memory demands. Drawing inspiration from neuroscience, where the brain replays experiences to a predictive World Model rather than directly to the policy, we present ARROW (Augmented Replay for RObust World models), a model-based continual RL algorithm that extends DreamerV3 with a memory-efficient, distribution-matching replay buffer. Unlike standard fixed-size FIFO buffers, ARROW maintains two complementary buffers: a short-term buffer for recent experiences and a long-term buffer that preserves task diversity through intelligent sampling. We evaluate ARROW on two challenging continual RL settings: Tasks without shared structure (Atari), and tasks with shared structure, where knowledge transfer is possible (Procgen CoinRun variants). Compared to model-free and model-based baselines with replay buffers of the same-size, ARROW demonstrates substantially less forgetting on tasks without shared structure, while maintaining comparable forward transfer. Our findings highlight the potential of model-based RL and bio-inspired approaches for continual reinforcement learning, warranting further research.", "date": "2026-03-12", "categories": []}
{"id": "2606.12488", "title": "A Stationary (and Therefore Compatible) Representation is All You Need", "abstract": "Learning compatible representations aims to learn feature representations that can be used interchangeably over time whenever a model undergoes updates. In this paper, we demonstrate that stationary representations learned by d-Simplex fixed classifiers imply compatibility as in its formal definition. This result establishes a foundation for future works and can be directly exploited in practical learning scenarios. We address the challenge of learning compatibility using $d$-Simplex fixed classifiers when the model is sequentially fine-tuned. Learning according to a d-Simplex fixed classifier with the cross-entropy loss aligns feature distributions at the first-order statistics. Consequently, it may not fully capture higher-order dependencies in the representation between model updates. To address this issue, we demonstrate that training the model using a $d$-Simplex fixed classifier through a convex combination of the cross-entropy loss and a contrastive loss not only captures higher-order dependencies, but is also equivalent to learning with the cross-entropy under the compatibility constraints. We confirm our findings with extensive experiments also considering a new scenario where a pre-trained model is sequentially fine-tuned and occasionally replaced with an improved model. We show that stationary representations enable uninterrupted retrieval services (without reprocessing gallery images) while improving performance during model updates and replacements, achieving state-of-the-art. Code at https://github.com/miccunifi/iamcl2r.", "date": "2026-06-10", "categories": []}
{"id": "2606.12633", "title": "ECA: Efficient Continual Alignment for Open-Ended Image-to-Text Generation", "abstract": "Incremental Learning (IL) for Open-ended Image-to-Text Generation (OpenITG) enables models to continuously generate accurate, contextually relevant text for new images while preserving previously acquired knowledge. Unlike prior studies, this paper addresses a more practical scenario in which the predominant category of visual data shifts over time as environments evolve. In this context, we introduce a new notion of continual alignment, which incrementally adapts the alignment module within pre-trained VLMs to preserve high-quality cross-modal representations. Based on this idea, we propose Efficient Continual Alignment (ECA), a novel exemplar-free IL approach for OpenITG. The key challenge is enabling the model to acquire new, task-specific features while minimizing interference with the established alignment without accessing raw data from previous tasks. To address this, ECA employs three core mechanisms: a Mixture of Query (MoQ) module that adapts task-specific query tokens, a Fisher Dynamic Expansion (FeDEx) that dynamically expands model structure based on a Fisher Information Matrix (FIM)-based metric, and an embedding dictionary with Dictionary Replay (DR) to retain past knowledge. To evaluate ECA's performance, we construct four new IL OpenITG benchmarks that better reflect real-world scenarios. Experimental results demonstrate that ECA significantly mitigates catastrophic forgetting and improves IL performance compared to baseline methods. Code and benchmarks are available at https://github.com/Snowball0823/ECA.", "date": "2026-06-10", "categories": []}
{"id": "2606.12655", "title": "Amnesia: A Stealthy Replay Attack on Continual Learning Dreams", "abstract": "Continual learning (CL) models often use experience replay to reduce catastrophic forgetting, but their robustness to replay sampling interference remains underexplored. Existing CL attacks alter inputs or training pipelines (poisoning/backdoors) and rarely include explicit auditable constraints, limiting realism. Here, auditability means a monitor can verify compliance from sampler-visible telemetry - e.g., logged replay index/label statistics - by checking that the realized replay class histogram stays close to a nominal baseline and that replay rate is unchanged per batch and/or over a rolling window. We study a limited-privilege insider who controls only replay index selection, not pixels, labels, or model parameters, while staying within auditable limits such as queue priorities. We introduce Amnesia, a replay composition attack that maximizes degradation under two budgets: a visibility budget delta bounding the TV/KL divergence from a nominal class histogram p0, and a mass budget f fixing the replay rate. Amnesia has two steps: (i) compute lightweight class utilities, such as EMA loss or confidence, to tilt p0 toward harmful classes; and (ii) project the tilt back into the delta-ball using efficient KL (exponential tilt) or TV (balanced mass redistribution) optimizers. A windowed scheduler enforces rolling audits. Across challenging CL benchmarks and strong replay baselines, Amnesia consistently lowers final accuracy (ACC) and worsens backward transfer (-BWT). The KL variant delivers high impact while remaining largely undetected under multiple audit schemes, including per-batch and rolling-window checks. The TV variant is more damaging but easier to detect, especially under tight per-class constraints. These results expose index-only replay control as a practical, auditable threat surface in CL systems and establish a principled impact-visibility trade-off.", "date": "2026-06-10", "categories": []}
{"id": "2606.13637", "title": "The Stable Recovery Manifold: Geometric Principles Governing Recoverability in Continual Learning", "abstract": "Catastrophic forgetting is often viewed as the destruction of previously learned knowledge during sequential learning. Building on the Accessibility Collapse framework, we investigate the geometric structure of recoverability in continual learning. Using Split CIFAR-100 and a sequentially trained ResNet-18, we analyze recoverability, representational drift, and recovery complexity across ten tasks. We introduce Recovery Subspace Dimensionality (k_t), a measure of the minimum number of singular directions required to preserve 90 percent of full probe performance. Contrary to our Recoverability Diffusion hypothesis, recovery dimensionality remains stable throughout training (mean k_t = 8.0) despite substantial representational drift. Principal-angle drift strongly predicts recoverability (r = -0.862), and a simple geometric model explains 82.2 percent of recoverability variance. These findings support the Stable Recovery Manifold hypothesis, suggesting that forgotten knowledge remains compactly decodable despite representational reorganization. The results indicate that catastrophic forgetting is primarily an accessibility and manifold-alignment problem rather than information destruction.", "date": "2026-06-11", "categories": []}
{"id": "2606.08102", "title": "Continual Quadruped Robots Coordination via Semantic Skill Discovery", "abstract": "Multi-quadruped coordination has attracted increasing attention due to its enhanced payload capacity, broader contact coverage, and improved adaptability to challenging tasks. Existing methods for multi-quadruped manipulation typically focus on predefined or closed task families, often relying on multi-agent reinforcement learning (MARL) to train task-specific coordination policies. However, such methods struggle in open-ended continual learning settings, where tasks arrive sequentially and robots are expected to acquire new coordination skills while reusing previously learned ones without catastrophic forgetting. To address this challenge, we propose Conquer, a semantic skill-library framework that formulates continual multi-quadruped coordination as a retrieve-adapt-update process. First, to accommodate varying team sizes across tasks, we design a team-structured Self-Allies-Goal (SAG) backbone that supports variable-cardinality robot teams by explicitly modeling each robot's own state, teammate context, and task goal. For each incoming task, Conquer constructs a task-level semantic descriptor from pre-execution information and retrieves a relevant skill from the library for adaptation. After successful execution, Conquer updates the skill library by extracting trajectory-level semantic descriptors and organizing them according to semantic distance, thereby enabling continual skill accumulation and cross-task knowledge transfer. Simulation experiments show that Conquer achieves a final average success rate of 95.6%, demonstrating strong forward transfer and negligible catastrophic forgetting. Real-world rollouts on Unitree Go2 teams further validate the deployment feasibility of Conquer for practical multi-quadruped coordination. Simulation and real-robot demonstration videos are available at: https://conquer-project.pages.dev/.", "date": "2026-06-06", "categories": []}
{"id": "2606.11272", "title": "Federated continual learning: A comprehensive survey on lifelong and privacy-preserving learning over distributed and non-stationary data", "abstract": "Federated Learning (FL) enables collaborative and privacy-preserving model training across distributed clients, but most existing FL systems implicitly assume data stationarity. In real-world settings-such as healthcare, industrial IoT (IIOT), cybersecurity, and smart cities-data streams are inherently non-stationary, leading classical FL methods to suffer from performance degradation, instability, and catastrophic forgetting.\n  Continual Learning (CL) addresses learning under evolving data distributions but has been largely studied in centralized settings, overlooking key constraints of federated systems, including privacy, limited communication, and client heterogeneity. Federated Continual Learning (FCL) emerges at the intersection of FL and CL, aiming to support lifelong, adaptive, and privacy-aware learning over distributed and non-stationary data.\n  This survey provides a comprehensive and systematic overview of FCL. We first present a formal definition of the FCL problem and clarify its distinctive characteristics. We then analyze the limitations of classical FL under non-stationary conditions, highlighting how CL principles support long-term adaptation. To organize the rapidly growing literature, we propose a multi-dimensional taxonomy of FCL approaches. Furthermore, we review representative application domains and data modalities, summarize commonly used evaluation metrics, and discuss experimental perspectives for assessing long-term performance and forgetting. Finally, we highlight key open challenges, including handling extreme heterogeneity under temporal drift, designing scalable and privacy-preserving memory mechanisms, and establishing standardized benchmarks. This survey aims to serve as a reference and a roadmap for advancing FCL toward robust and deployable real-world systems.", "date": "2026-06-09", "categories": []}
{"id": "2606.11480", "title": "Accurate and Resource-Efficient Federated Continual Learning", "abstract": "Federated continual learning (FCL) must learn from distributed task streams under limited resources, such as communication, computation, memory, and label availability. Existing FCL methods often rely on repeated local optimization, replay, and full supervision. Analytic alternatives avoid iterative training and replay, but using high-dimensional random features to improve accuracy requires a second-order feature statistic, the Gram matrix, which has a quadratic communication cost in the random feature size $M$. We propose FedRAN, a resource-aware analytic FCL framework that replaces gradient-based updates with compact random feature statistics. Each client transmits a truncated-SVD summary of its Gram matrix, reducing the dominant second-order upload from quadratic to linear in $M$ for fixed rank. The server performs a two-level QR-SVD subspace merge, spatially across clients and temporally across tasks, and solves a ridge classifier in closed form. FedRAN further supports label scarcity through prototype-based pseudo-labeling. Across CIFAR-100, ImageNet-R, and VTAB datasets, FedRAN improves average accuracy by up to 4.8 percentage points over the strongest baseline, uses 30.6-121.8$\\times$ less per-client communication than optimization-based FCL, and is 190.3$\\times$ faster on average than gradient-based baselines; with only 20% labels, pseudo-labeling improves average accuracy by up to 6.61 points. These results show that FedRAN enables accurate and resource-efficient FCL under communication, computation, and label constraints. The source code is available at https://github.com/JebacyrilArockiaraj/Fed-RAN-SSL.", "date": "2026-06-09", "categories": []}
{"id": "2606.11844", "title": "TaskFusion: Continual Anomaly Detection for Heterogeneous Tabular Data", "abstract": "Continual anomaly detection in tabular data is challenging and remains largely underexplored, particularly in settings with heterogeneous feature schemas, distribution shifts, and severe class imbalance. In many real-world applications, data arrive sequentially from diverse domains, rendering conventional continual learning methods ineffective due to their reliance on a fixed input space. We propose a continual learning (CL) method, which can overcome these challenges and continually learn from different tasks. Our method consists of three main parts: our AGF model, Taskfusion augmentation, and outlier exposure. The AGF-model maps task-specific features into a shared space, then aligns distributions to reduce representation drift, and learns anomaly decision boundaries in the aligned space. To improve stability, we introduce Taskfusion augmentation, combining boundary-aware interpolation within tasks to refine the model anomaly boundaries and cross-task mixing to transfer anomaly structure across datasets. To handle class imbalance and memory constraints, we employ tabular dataset distillation to store compact synthetic replay samples, which are jointly used with augmented data in an outlier exposure objective for robust anomaly detection. We evaluate the approach on 21 heterogeneous datasets across multiple domains. Results show that our approach substantially improves continual anomaly detection performance over sequential fine-tuning and other CL baselines while reducing catastrophic forgetting and maintaining stable detection across heterogeneous datasets.", "date": "2026-06-10", "categories": []}
{"id": "2507.14725", "title": "GRID: Scaling Task-Agnostic Inference in Continual Prompt Tuning", "abstract": "Prompt-based continual learning (CL) offers a parameter-efficient way to adapt large language models (LLMs) across task sequences. However, existing methods often rely on task-aware inference and maintain an expanding set of task-specific prompts, leading to (1) severe performance degradation on earlier tasks when task identifiers are unavailable for prompt selection at inference time, and (2) limited scalability as task sequence grows. We propose GRID, a unified framework designed to address these challenges. GRID incorporates an output-space-aware decoding mechanism that enhances backward transfer by leveraging representative inputs and automatic label semantic normalization, alongside a gradient-guided prompt selection strategy that compresses less informative prompts into a single aggregated representation for scalable, memory-efficient continual learning. Extensive experiments on long-sequence and negative-transfer benchmarks show that GRID improves backward transfer, achieves competitive forward transfer, and substantially reduces prompt memory across encoder-decoder and decoder-only architectures, including T5, Qwen, and LLaMA. Source code is available at https://github.com/AnushkaTi/GRID.", "date": "2025-07-19", "categories": []}
{"id": "2604.15414", "title": "Beyond Single-Model Optimization: Preserving Plasticity in Continual Reinforcement Learning", "abstract": "Continual reinforcement learning must balance retention with adaptation, yet many methods still rely on \\emph{single-model preservation}, committing to one evolving policy as the main reusable solution across tasks. Even when a previously successful policy is retained, it may no longer provide a reliable starting point for rapid adaptation after interference, reflecting a form of \\emph{loss of plasticity} that single-policy preservation cannot address. Inspired by quality-diversity methods, we introduce \\textsc{TeLAPA} (Transfer-Enabled Latent-Aligned Policy Archives), a continual RL framework that organizes behaviorally diverse policy neighborhoods into per-task archives and maintains a shared latent space so that archived policies remain comparable and reusable under non-stationary drift. This perspective shifts continual RL from retaining isolated solutions to maintaining \\emph{skill-aligned neighborhoods} with competent and behaviorally related policies that support future relearning. In our MiniGrid CL setting, \\textsc{TeLAPA} learns more tasks successfully, recovers competence faster on revisited tasks after interference, and retains higher performance across a sequence of tasks. Our analyses show that source-optimal policies are often not transfer-optimal, even within a local competent neighborhood, and that effective reuse depends on retaining and selecting among multiple nearby alternatives rather than collapsing them to one representative. Together, these results reframe continual RL around reusable and competent policy neighborhoods, providing a route beyond single-model preservation toward more plastic lifelong agents.", "date": "2026-04-16", "categories": []}
{"id": "2606.06698", "title": "RECAP: Regression Evaluation for Continual Adaptation of Prompts", "abstract": "Production agentic systems routinely face evolving constraints and must comply from the very next interaction. Scenarios like a tool-call notification changing a compliance threshold or a policy update adding disclosure requirements fit this criteria, having close to no room for errors in production. This proactive adaptation setting is common in deployment, but absent from current benchmarks, which assume either static constraint sets or reactive protocols with evaluation feedback. We introduce RECAP, a benchmark that measures continual-learning phenomena (forgetting, regression, forward transfer) at the constraint level under a strictly proactive adapt-then-test protocol: prompt optimization methods receive only the constraint specification and must generalize before seeing any test data. Evaluating six methods across four LLMs and three schedules with evolving constraints, we find that these methods show no significant improvement in performance, even after incurring a higher latency. These methods, designed for offline or reactive settings, are inadequate for the proactive paradigm. Our work emphasizes the growing need for designing proactive prompt adaptation methods, where the models must remain robust to evolving needs in deployment.", "date": "2026-06-04", "categories": []}
{"id": "2606.09960", "title": "HydraCIL: Decoupled Class-Incremental Learning through Prototype-Guided Multi-Head Classifiers", "abstract": "We present HydraCIL, a decoupled continual learning model based on prototype-guided multi-head classifiers, targeting sustainable deployment in embedded and resource-constrained environments. While most Class-Incremental Learning (CIL) methods rely on powerful hardware and long retraining cycles, real-world systems, such as robots or edge AI devices, must adapt quickly with limited resources. HydraCIL addresses this gap by freezing the backbone and decoupling feature extraction from learning. For each task, features are extracted once and a lightweight, task-specific classifier head is created, avoiding costly backbone retraining. At inference, HydraCIL selects the appropriate head via similarity with prototypes. Experiments on CIFAR-100, ImageNet-100, CoRe50, and Flowers102 datasets show that HydraCIL matches or outperforms state-of-the-art CIL methods while significantly reducing training time and carbon footprint, making it a practical solution for continual learning in real-world and embedded settings, where energy efficiency and rapid adaptation are critical.", "date": "2026-06-08", "categories": []}
{"id": "2606.10522", "title": "GUI-AC: Enhancing Continual Learning in GUI Agents", "abstract": "Graphical User Interfaces (GUIs) serve as the dominant medium for human-computer interaction, yet building GUI agents that generalize across the vast diversity of real-world interface environments, with the same flexibility and robustness that humans naturally exhibit, remains unsolved. Notably, GUI data are inherently non-stationary: the continual emergence of previously unseen interface instances (e.g., novel domains and resolutions) induces persistent distribution shifts, significantly impeding the continual learning of existing GUI agents. Reinforcement fine-tuning (RFT) has attracted considerable attention as a promising approach. Nevertheless, RFT exhibits pronounced instability in its grounding capability, manifested as sharp reward discontinuities and high-variance oscillations. The imbalanced distribution of rollout outcomes introduces substantial noise into advantage estimation, leading to policy overconfidence. The fixed clipping bound suppresses the increase in policy probabilities needed to adapt to new distributions, leading to a collapse in exploration capacity. To address these challenges, we propose GUI-AC, a method that enhances the continual learning capability of GUI agents. GUI-AC introduces grounding certainty to support two core mechanisms: (i) Adaptive Advantage, which down-weights noisy advantage estimates to prevent policy overconfidence; and (ii) Dynamic Clipping, which relaxes the clipping bound to encourage exploration range. Extensive experiments show that these mechanisms jointly improve performance, enabling our method to surpass state-of-the-art baselines. Code is available anonymously at https://anonymous.4open.science/r/GUI-AC.", "date": "2026-06-09", "categories": []}
{"id": "2606.10887", "title": "Listen, Look, and Learn: Learning Without Forgetting through SAM-Audio", "abstract": "Class-Incremental Learning (CIL) aims to continuously learn new classes without forgetting previously acquired knowledge. While recent CIL advances have spurred significant interest across various modalities, the audio-visual setting remains underexplored. Furthermore, although foundational multimodal models like SAM-Audio encapsulate rich static priors, our empirical analysis reveals that these representations struggle in incremental settings. This work bridges this gap by integrating SAM-Audio's audio-visual priors into the CIL setting. Specifically, we leverage its dense audio and visual representations and employ a novel guided attention strategy where the audio features contextually guide the visual representations. To further mitigate catastrophic forgetting, we introduce dual-level distillation objectives at both the feature and logit levels. Extensive evaluations on audio-visual CIL benchmarks demonstrate that our approach consistently outperforms state-of-the-art methods.", "date": "2026-06-09", "categories": []}
{"id": "2511.07938", "title": "Decision-Focused Continual Learning for Seaport Power-Logistics Scheduling: Generalization across Varying Tasks", "abstract": "Power-logistics scheduling in modern seaports typically follows a predict-then-optimize pipeline. To enhance the decision quality of predictions, decision-focused learning has been proposed, which aligns the training of forecasting models with downstream decision outcomes. However, this end-to-end design inherently restricts the value of forecasting models to a specific task structure and therefore generalizes poorly to evolving tasks induced by varying vessel arrivals. We address this gap with a decision-focused continual learning framework that adapts online to a stream of scheduling tasks. Specifically, we introduce Fisher-information-based regularization to enhance cross-task generalization by preserving parameters critical to prior tasks. A differentiable convex surrogate is also developed to stabilize gradient backpropagation. The proposed approach enables learning a decision-aligned forecasting model across a varying task stream with sustainable long-term computational and memory requirements. Experiments calibrated to Jurong Port show improved decision performance and cross-task generalization over existing methods, together with reduced computational cost and a bounded memory footprint.", "date": "2025-11-11", "categories": []}
{"id": "2606.01379", "title": "Turning Back Without Forgetting: Selective Backward Refinement for Parameter-Efficient Continual Learning", "abstract": "While prompt-based parameter-efficient continual learning mitigates catastrophic forgetting by isolating task-specific prompts, this isolation also limits later tasks from improving earlier ones, leaving backward knowledge transfer underexplored. We address this limitation by proposing Selective bAckward refinement for positive Backward knowledge transfER (SABER), a replay-free framework that enables controlled backward transfer in prompt-based continual learning. SABER determines when backward refinement is beneficial using complementary task-correlation criteria based on prompt-gradient geometry and loss-distribution similarity, and how to perform refinement safely by restricting updates to non-interfering directions in the prompt parameter space. Extensive experiments across multiple continual learning benchmarks and diverse pretrained backbones, including T5-Large, LLaMA, and Qwen, demonstrate that SABER consistently achieves positive backward transfer while maintaining strong overall average performance. Code is available at https://github.com/OptMN-Lab/SABER-ICML-2026/.", "date": "2026-05-31", "categories": []}
{"id": "2606.04029", "title": "Position: Deployed Reinforcement Learning should be Continual", "abstract": "Reinforcement Learning (RL) has received increasing attention and adoption in real-world use cases. Most of these systems follow a train-then-fix paradigm, where trained agents do not learn while interacting with the world until performance degrades and retraining becomes necessary. In this position paper, we argue that deploying an agent that is incapable of optimality, but receives an evaluative reward signal, is inherently a continual RL problem. We identify four sources of non-stationarity after deployment that necessitate never-ending learning, and highlight why the best deployed agents never stop adapting. We analyze successful examples of continual RL in the real world, and present the community with the advantages and measures to move away from the current train-then-fix paradigm.", "date": "2026-06-01", "categories": []}
{"id": "2606.08013", "title": "Evaluating the Impact of Task Granularity on Catastrophic Forgetting in Continual Learning", "abstract": "Catastrophic forgetting, the abrupt loss of previously acquired knowledge upon learning new information, remains the central challenge in Continual Learning. This project investigates whether the order in which a model learns information affects how well it retains knowledge. Specifically, we ask: does learning general categories first (like \"animals\" vs \"vehicles\") before learning specific classes (like \"dog\" vs \"cat\") reduce forgetting compared to learning all classes at once?\n  We test three approaches on CIFAR-100: (1) Coarse-to-Fine: train on 2 super-classes, then expand to 10 specific sub-classes, (2) Fine-to-Coarse: train on 10 sub-classes, then group into 2 super-classes, and (3) Flat: train on all 10 classes from the start. We use Elastic Weight Consolidation (EWC) to prevent forgetting during transitions. Our hypothesis is that learning general patterns first creates a stable foundation that helps the model retain knowledge when learning more detailed distinctions. We evaluate using standard metrics (accuracy, precision, recall, F1) plus continual learning metrics like backward transfer and forgetting rates. This work could inform how we design learning sequences for real-world systems that need to learn incrementally.", "date": "2026-06-06", "categories": []}
{"id": "2606.08447", "title": "Not Just After One: Sleep-Inspired Replay Prevents Catastrophic Forgetting After Sequential Tasks", "abstract": "One of the critical limitations of artificial neural networks is their lack of ability to continually learn: training on new tasks often leads to interference and forgetting of the previous ones. While several algorithms have been proposed to protect old memories from interference, they are typically applied during or immediately after each new episode of training. In contrast, humans and animals can learn continuously, acquiring multiple new memories during active learning before consolidating all of them into long-term storage. Here we show that multiple new tasks can be trained sequentially before an unsupervised sleep-like replay phase is applied to partially restore performance across all previously learned tasks. Our study further suggests that task-specific information remains resilient to new training but decays gradually as network is trained on new tasks. These findings point to novel principles for developing a broad range of continual learning AI solutions.", "date": "2026-06-07", "categories": []}
{"id": "2606.08452", "title": "Theoretical Foundations of Continual Learning via Drift-Plus-Penalty", "abstract": "In many real-world settings, data streams are nonstationary and arrive sequentially, requiring learning systems to adapt continuously without retraining from scratch. Continual learning (CL) addresses this challenge by incorporating new tasks while mitigating catastrophic forgetting, where learning new information degrades performance on previously acquired knowledge. We introduce a control-theoretic perspective on CL that explicitly regulates the evolution of forgetting, framing adaptation as a controlled process subject to long-term stability constraints. We focus on replay-based CL, where a finite memory buffer stores representative samples from prior tasks. We propose COntinual Learning with Drift-Plus-Penalty (COLD), a continual learning framework based on the Drift-Plus-Penalty (DPP) principle from stochastic optimization. To facilitate analysis, we also consider an oracle variant, COLD-ORACLE, as a reference benchmark. At each task, both methods minimize the current task loss while maintaining a virtual queue that tracks deviations from long-term stability on previously learned tasks, capturing the stability-plasticity trade-off as a regulated dynamical process. We establish stability and convergence guarantees that characterize this trade-off through a tunable control parameter. Experiments on standard benchmarks demonstrate that COLD consistently outperforms a broad range of state-of-the-art CL methods while providing competitive and controllable forgetting behavior through explicit regulation of stability and plasticity.", "date": "2026-06-07", "categories": []}
{"id": "2606.09430", "title": "LargeMonitor: Monitoring Online Task-Free Continual Learning via Large Pretrained Models", "abstract": "Online task-free continual learning (TFCL) requires intelligent agents to sequentially accumulate knowledge from an unbounded, non-stationary data stream under strict single-pass constraints and without any explicit task identifiers. Existing online TFCL paradigms primarily rely on parameter-efficient prompt tuning or dynamic structure expansion driven by training-coupled optimization dynamics, such as empirical loss fluctuations or evolving latent distances. As a result, these training-coupled solvers remain agnostic to the structural origins of distribution drift, mechanically enforcing a fixed strategy across fundamentally distinct streaming variations. To address this gap, we propose LargeMonitor, a framework that leverages large pretrained foundation models to autonomously orchestrate task-free continuous adaptation. Specifically, LargeMonitor introduces a decoupled detection module utilizing the frozen, stable representation space of large vision models (LVMs) to achieve robust, zero-shot drift detection without training-dependent interference or brittle threshold tuning. Upon a confirmed drift, the framework activates a context-aware diagnostic module driven by large multimodal models (LMMs) to interpret the precise semantic etiologies of the stream variation (e.g., novel class emergence vs. environmental domain shift). This dual-stage capability empowers the continuous learner to dynamically deploy adaptive and shift-specific optimization strategies. Extensive experiments across multiple TFCL settings and benchmarks demonstrate that LargeMonitor achieves precise, robust detection and diagnosis of complex data streams while consistently improving the performance of existing online TFCL algorithms.", "date": "2026-06-08", "categories": []}
{"id": "2606.09762", "title": "Preserving Plasticity in Continual Learning via Dynamical Isometry", "abstract": "Continual training of deep neural networks under non-stationarity often leads to a progressive loss of plasticity, eventually limiting further learning. We relate plasticity to the empirical Neural Tangent Kernel, and identify dynamical isometry (the condition that layer-wise Jacobian singular values remain close to one) as a key mechanism for preserving plasticity in continual learning. We revisit a class of networks that are almost-everywhere isometric while remaining universal Lipschitz function approximators, demonstrating that near-dynamical isometry is compatible with expressive nonlinear representations. For general architectures, we propose an efficient isometry-promoting regularization scheme and identify a novel mechanism by which it can reactivate dormant ReLU units. Building on this, we introduce AdamO, an Adam-style adaptive optimizer that decouples isometry regularization from gradient updates, analogous to AdamW. We further reinterpret prior plasticity-preserving approaches through the lens of dynamical isometry, showing that they target only a partial measure of isometry. Across supervised and reinforcement-learning continual-learning benchmarks designed to induce plasticity loss, our methods consistently match or outperform existing approaches.", "date": "2026-06-08", "categories": []}
{"id": "2505.12239", "title": "Towards Efficient and Exact Forgetting Services in Pre-Trained-Model-based Continual Learning", "abstract": "In Continual Learning (CL), using a Pre-Trained Model (PTM) as the feature extractor has become a popular practice. Accompanied by analytic classifiers, the PTM-based methods have achieved state-of-the-art performance in CL, in pursuit of the non-forgetting goal. Meanwhile, actively forgetting specific knowledge acquired during the CL phase is also essential in most service construction paradigms, for example, Mobile Crowd Sensing (MCS), where mobile edge nodes continuously collect sensory data and demand not only non-forgetting adaptation but also specific knowledge forgetting for privacy preservation. Thus, a unique problem, called Continual Unlearning (CU), arises when the forgetting requests show sequentially in CL. However, existing unlearning methods focus on single-shot joint forgetting and prove highly inadequate when applied to CU, including (1) violating the historical data privacy in CL and (2) vulnerably being overwhelmed or degraded with adversarially frequent requests. To handle the challenges of CU, we propose a gradient-free approach, called Analytic Continual Unlearning (ACU), for efficient and exact forgetting with historical data privacy preservation in PTM-based CL. In response to each unlearning request, our ACU recursively derives the analytical (i.e., closed-form) solutions via least squares in an interpretable manner. By meticulous design, our ACU is compatible with both sample-level and class-level unlearning requests. The theoretical and experimental evaluations validate our ACU's superiority in unlearning effectiveness, model fidelity, and system efficiency.", "date": "2025-05-18", "categories": []}
{"id": "2606.06978", "title": "CL-CLIP: CLIP-Based Continual Learning Framework with Cost-Volume Category Decoupling for Object Detection", "abstract": "Continual Object Detection (COD) requires a detector to acquire new categories over time while preserving previously learned ones. This goal is closely related to open-vocabulary detection, since both settings require reasoning over categories that are not fully covered by the annotations available at the current training stage. Recent CLIP-based open-vocabulary detectors have shown strong zero-shot generalization, and frameworks such as F-ViT demonstrate that vision-language pretraining can provide powerful zero-shot detection ability for unseen categories. However, real-world deployments cannot remain purely zero-shot: once these detectors are continually updated on newly introduced categories, they suffer severe catastrophic forgetting and quickly lose their previously calibrated detection ability. We therefore propose CL-CLIP, a CLIP-based COD framework that equips open-vocabulary detectors with better continual learning ability through cost-volume-guided category decoupling. Specifically, following CAT-Seg, we compute a CLIP image-text similarity cost volume, defined as dense category-wise response maps between visual tokens and class text embeddings. This zero-shot spatial prior decomposes shared region features into class-specific pathways, which are then processed by a Multi-Expert RoI head. Extensive experiments on PASCAL VOC and MS-COCO show that CL-CLIP substantially improves the F-ViT baseline under continual fine-tuning and achieves competitive performance with existing continual object detectors, especially in adapting to newly introduced categories while preserving competitive base-class performance.", "date": "2026-06-05", "categories": []}
{"id": "2606.07474", "title": "Unsupervised Continual Clustering via Forward-Backward Knowledge Distillation", "abstract": "Unsupervised Continual Learning (UCL) aims to enable neural networks to learn sequential tasks without labels or access to past data. A major challenge in this setting is Catastrophic Forgetting, where models forget previously learned tasks upon learning new ones. This challenge is amplified in UCL due to the absence of labels to guide learning and memory retention. Existing mitigation strategies, such as knowledge distillation and replay buffers, often raise memory and privacy concerns. Moreover, current UCL methods largely overlook clustering-specific objectives. To fill this gap, we introduce Unsupervised Continual Clustering (UCC) and propose Forward-Backward Knowledge Distillation for Continual Clustering (FBCC). FBCC employs a continual teacher network with a clustering projector and lightweight task-specific students. Through a dual-phase forward-backward distillation process, the teacher learns new clusters while preserving previously discovered cluster structure without storing past data. FBCC represents a pioneering approach to UCC, demonstrating improved clustering performance across sequential tasks. Experiments on four benchmark datasets demonstrate that FBCC consistently outperforms existing continual learning baselines in clustering accuracy while significantly reducing catastrophic forgetting.", "date": "2026-06-05", "categories": []}
{"id": "2606.07500", "title": "Sparse Subspace-to-Expert Sharing for Task-Agnostic Continual Learning", "abstract": "Continual learning in Large Language Models (LLMs) is hindered by the plasticity-stability dilemma, where acquiring new capabilities often leads to catastrophic forgetting of previous knowledge. Existing methods typically treat parameters uniformly, failing to distinguish between specific task knowledge and shared capabilities. We introduce Mixture of Sparse Experts for Task Agnostic Continual Learning (SETA), a framework that resolves the plasticity-stability conflict through adaptive sparse subspace decomposition into task-specific expert modules. Unlike standard updates, where tasks compete for the same parameters, SETA separates knowledge into unique experts, designed to isolate task-specific patterns, and shared experts, responsible for capturing common features. This structure is maintained through adaptive elastic anchoring and a routing-aware regularization that jointly protect shared knowledge at both the weight and routing levels and enable a unified gating network to automatically retrieve the correct expert combination during inference. Extensive experiments across diverse domain-specific benchmarks demonstrate that SETA achieves competitive or superior overall performance relative to state-of-the-art continual learning baselines, with particularly strong retention of early-task knowledge and improved backward transfer on LLaMA-2 7B and Qwen3-4B.", "date": "2026-06-05", "categories": []}
{"id": "2511.20158", "title": "Harmonious Parameter Adaptation in Continual Visual Instruction Tuning for Safety-Aligned MLLMs", "abstract": "While continual visual instruction tuning (CVIT) has shown promise in adapting multimodal large language models (MLLMs), existing studies predominantly focus on models without safety alignment. This critical oversight ignores the fact that real-world MLLMs inherently require such mechanisms to mitigate potential risks. In this work, we shift our focus to CVIT for safety-aligned MLLMs and observe that during continual adaptation, the model not only suffers from task forgetting but also exhibits degradation in its safety. Achieving a harmonious balance between safety and task performance remains a crucial challenge. To address this, we propose Harmonious Parameter Adaptation (HPA), a post-training framework composed of focusing-based parameter partition, harmoniously balanced parameter selection, and orthogonal parameter adjustment. Specifically, HPA partitions parameters into two types based on their focus on safety or task performance, and selects the focused ones to preserve from a balanced perspective. In addition, HPA imposes orthogonality constraints on parameter updates to further alleviate catastrophic forgetting. Extensive experiments on the CVIT benchmark and safety evaluation datasets demonstrate that HPA better maintains high safety and mitigates forgetting than existing baselines. Code is available at https://github.com/Minato-Zackie/HPA.", "date": "2025-11-25", "categories": []}
{"id": "2604.20572", "title": "Ask Only When Needed: Proactive Retrieval from Memory and Skills for Experience-Driven Lifelong Agents", "abstract": "Online lifelong learning agents must decide not only how to act but also when to consult prior experience to continually improve on long-horizon tasks. Existing methods typically retrieve memories passively, such as at task initialization or after each step, and therefore miss knowledge gaps that arise during interaction. We propose ProactAgent, an experience-driven lifelong learning framework for proactive retrieval over a structured Experience Base. ProactAgent continually improves through ExpOnEvo, which jointly updates policies and refines memory, organizing past interactions into factual, episodic, and skill repositories. It further introduces ProactRL, which treats retrieval as an explicit policy action and learns when and what to retrieve. By comparing paired continuations from identical interaction prefixes with and without retrieval, ProactRL provides step-level process rewards that encourage retrieval only when it improves task outcomes or efficiency. Experiments on SciWorld, AlfWorld, and StuLife show that ProactAgent consistently outperforms all baselines, achieving up to 32% relative improvement in success rate and over 33% reduction in interaction rounds. Our code will be publicly available at GitHub.", "date": "2026-04-22", "categories": []}
{"id": "2605.13075", "title": "Scaling few-shot spoken word classification with generative meta-continual learning", "abstract": "Few-shot spoken word classification has largely been developed for applications where a small number of classes is considered, and so the potential of larger-scale few-shot spoken word classification remains untapped. This paper investigates the potential of a spoken word classifier to sequentially learn to distinguish between 1000 classes when it is given only five shots per class. We demonstrate that this scaling capability exists by training a model using the Generative Meta-Continual Learning (GeMCL) algorithm and comparing it to repeatedly trained or finetuned baselines. We find that GeMCL produces exceptionally stable performance, and although it does not always outperform a repeatedly fully-finetuned HuBERT model nor a frozen HuBERT model with a repeatedly trained classifier head, it produces comparable performance to the latter while adapting 2000 times faster, having been trained less than half of the data for two orders of magnitude less time.", "date": "2026-05-13", "categories": []}
{"id": "2606.05559", "title": "CLaaS: Continual learning as a service for sample efficient online learning", "abstract": "Deployed large language model agents must adapt to distribution shift in dynamic environments. Ideally, adaptation can be performed from accumulated agent experiences and retain prior capabilities while transferring to future tasks. However, agent actions and environmental transitions can only be sampled once per scenario, as real-world environments cannot be trivially reset. To this end, we investigate an experiential and online continual learning setting in which agents learn from a stream of scenarios. We propose continual learning as-a-service (CLaaS), a system which enables agents to improve during deployment, abstracted behind a chat API. To increase sample efficiency, CLaaS stores rollouts in an experience replay buffer for gradient reuse during asynchronous training. We evaluate CLaaS on an adversarial task, demonstrating that parametric updates lead to superior forward transfer and less forgetting than in-context learning, with replay being a critical choice for sample efficiency.", "date": "2026-06-04", "categories": []}
{"id": "2606.05661", "title": "Continual Learning Bench: Evaluating Frontier AI Systems in Real-World Stateful Environments", "abstract": "Continual learning, the ability of AI systems to improve through sequential experience, has attracted substantial interest, but no high-quality benchmark exists to evaluate it. We introduce Continual Learning Bench (CL-Bench), the first difficult, expert-validated benchmark designed to measure whether LLM-based systems genuinely improve with experience. CL-Bench spans six diverse domains (software engineering, signal processing, disease outbreak forecasting, database querying, strategic game-playing, and demand forecasting), each validated by domain experts and designed so that tasks share a learnable latent structure (codebase layout, disease outbreak dynamics, opponent strategies) that a stateful system can discover online but a stateless one cannot. We evaluate frontier models across several agent architectures, from naive in-context learning (ICL) to dedicated memory systems, introducing a gain metric to isolate learning from prior capabilities. We find that these systems leave headroom for improved continual learning: agents frequently overfit to immediate observations or fail to reuse knowledge across instances, and dedicated memory systems do not fix this -- in fact, naive ICL outperforms systems dedicated to memory management. CL-Bench is the first benchmark to evaluate continual learning across diverse real-world domains with expert-validated tasks and isolate online learning from underlying model capability, showing a need for better continual learning systems.", "date": "2026-06-04", "categories": []}
{"id": "2606.05675", "title": "Two-Way Is Better Than One: Bidirectional Alignment with Cycle Consistency for Exemplar-Free Class-Incremental Learning", "abstract": "Continual learning (CL) seeks models that acquire new skills without erasing prior knowledge. In exemplar-free class-incremental learning (EFCIL), this challenge is amplified because past data cannot be stored, making representation drift for old classes particularly harmful. Prototype-based EFCIL is attractive for its efficiency, yet prototypes drift as the embedding space evolves; therefore, projection-based drift compensation has become a popular remedy. We show, however, that existing one-directional projections introduce systematic bias: they either retroactively distort the current feature geometry or align past classes only locally, leaving cycle inconsistencies that accumulate across tasks. We introduce BiCyc, a bidirectional projector alignment approach with a cycle-consistency objective. BiCyc jointly optimizes two maps, old-to-new and new-to-old, with stop-gradient gating so that transport and representation co-evolve. Analytically, we show that the cycle loss contracts the singular spectrum toward unity in whitened space, and that improved transport of class means and covariances yields smaller perturbations of classification log-odds, preserving old-class decisions and mitigating catastrophic forgetting. Empirically, across standard EFCIL benchmarks, BiCyc substantially reduces forgetting and improves accuracy in from-scratch settings, while remaining competitive in the pretrained fine-grained regime.", "date": "2026-06-04", "categories": []}
{"id": "2606.05695", "title": "Revisiting Prototype Rehearsal for Exemplar-Free Continual Learning: Manifold-Aware Boundary Sampling with Adaptive Class-Balanced Loss", "abstract": "Exemplar-free class-incremental learning (EFCIL) aims to acquire new classes over time without storing raw data. Historically, prototype rehearsal, which samples around stored class prototypes and mixes them with current-task data, has been a popular strategy to reduce catastrophic forgetting. However, recent drift-compensation methods that explicitly realign prototypes in the evolving feature space consistently outperform prototype-based rehearsal, raising the question of whether rehearsal itself is fundamentally limited. We argue that the performance gap stems not from the idea of prototype rehearsal per se, but from how it is typically instantiated: existing approaches treat prototypes as isolated class summaries that ignore information from nearby enemy classes, and fail to correct the emerging class imbalance between a handful of synthetic old-class samples and hundreds of real instances from newly introduced classes. Building on this hypothesis, we revisit prototype rehearsal and propose a manifold-aware variant that restores its competitiveness in EFCIL. First, we introduce Constrained Expansive Over-Sampling, which interpolates each old-class prototype toward its nearest enemy features from new classes, generating boundary-aware rehearsal samples that better follow the underlying data manifold while preserving inter-class separation. Second, we design an Adaptive Class-Balanced loss that performs time-based class weighting, amplifying gradients from older prototypes when they are most informative and gradually annealing their influence as richer supervision from later tasks accumulates. Together, these components turn prototype rehearsal into a drift-resilient, imbalance-aware mechanism that closes, and often reverses, the gap to recent drift-compensation methods, achieving state-of-the-art performance across multiple EFCIL benchmarks.", "date": "2026-06-04", "categories": []}
{"id": "2606.06032", "title": "Catastrophic Forgetting as Accessibility Collapse: A Three-Level Framework for Knowledge Persistence in Continual Learning", "abstract": "Catastrophic forgetting is commonly interpreted as the irreversible erasure of previously acquired knowledge during sequential learning. In this work, we investigate an alternative perspective: that forgetting may arise not from complete destruction of task representations but from a loss of accessibility to preserved information. We introduce a three-level framework separating knowledge storage, representation, and accessibility, and evaluate each component through a series of continual-learning experiments on sequential CIFAR-100 classification using ResNet-18. Our analysis combines checkpoint persistence, linear probing, representation geometry, classifier-reset recovery, and layer-wise recoverability experiments. We observe complete behavioral forgetting of earlier tasks, with task accuracy collapsing from 54.8% to 0%, while linear probe performance retains approximately 76% of the original representational information. Furthermore, retraining only the final classifier restores 75.7% of the original task performance without modifying the backbone network. Layer-wise analysis reveals that early and intermediate layers preserve highly recoverable task information despite severe degradation at later stages. Projection-energy and principal-angle analyses indicate that retained knowledge persists as distributed high-dimensional representations rather than through preservation of a small dominant subspace. These findings suggest that catastrophic forgetting is better characterized as an accessibility failure than complete representational erasure, and that substantial task-relevant information remains embedded within neural representations even after functional forgetting has occurred.", "date": "2026-06-04", "categories": []}
{"id": "2606.06494", "title": "TailLoR: Protecting Principal Components in Parameter-Efficient Continual Learning", "abstract": "Parameter-efficient finetuning methods based on spectral decomposition have enabled progress in Continual Learning. In this paper we introduce TailLoR, which utilizes the singular bases U and V of the pre-trained weights as a fixed reference frame to learn a low-rank update applied to the singular value matrix. A soft spectral penalty discourages updates aligned with dominant singular directions, reducing interference while routing fine-grained adaptation into the highly flexible, long-tail spectral coordinates.", "date": "2026-06-04", "categories": []}
{"id": "2510.17281", "title": "MemoryBench: A Benchmark for Memory and Continual Learning in LLM Systems", "abstract": "Scaling up data, parameters, and test-time computation has been the mainstream methods to improve LLM systems (LLMsys), but their upper bounds are almost reached due to the gradual depletion of high-quality data and marginal gains obtained from larger computational resource consumption. Inspired by the abilities of human and traditional AI systems in learning from practice, constructing memory and continual learning frameworks for LLMsys has become an important and popular research direction in recent literature. Yet, existing benchmarks for LLM memory often focus on evaluating the system on homogeneous reading comprehension tasks with long-form inputs rather than testing their abilities to learn from accumulated user feedback in service time. Therefore, we propose a user feedback simulation framework and a comprehensive benchmark covering multiple domains, languages, and types of tasks to evaluate the continual learning abilities of LLMsys. Experiments show that the effectiveness and efficiency of state-of-the-art baselines are far from satisfying, and we hope this benchmark could pave the way for future studies on LLM memory and optimization algorithms.\n  Website: https://memorybench.thuir.cn\n  Code: https://github.com/THUIR/MemoryBench\n  Data: https://huggingface.co/datasets/THUIR/MemoryBench\n  Data-Full: https://huggingface.co/datasets/THUIR/MemoryBench-Full", "date": "2025-10-20", "categories": []}
{"id": "2606.02576", "title": "ProtoAda: Prototype-Guided Adaptive Adapter Expansion and Geometric Consolidation for Multimodal Continual Instruction Tuning", "abstract": "Multimodal Large Language Models (MLLMs) achieve strong performance through instruction tuning, but real-world deployment requires them to continually acquire new vision-language capabilities, making Multimodal Continual Instruction Tuning (MCIT) essential. To reduce inter-task interference and promote collaboration, recent methods often employ sparse architectures like Mixture of LoRA Experts with image-text similarity routing. However, tasks with distinct response structures could share highly similar visual-linguistic semantics and thus be wrongly routed to the same expert; image-text similarity alone is insufficient for reliable task assignment. For example, an expert in a grounding task requiring coordinate prediction may be biased toward producing short textual answers after learning semantically similar VQA tasks. This format-blind task assignment integrates heterogeneous response types into shared parameters, inducing gradient interference and ineffective expert collaboration. To address this problem, we propose ProtoAda, a prototype-guided adaptive tuning framework. ProtoAda introduces format-aware task prototypes to align task assignment and routing with both task semantics and output structure, and further consolidates format-compatible updates in a geometry-aware manner to effectively reuse and progressively refine existing parameters. Extensive experiments on multiple benchmarks demonstrate that ProtoAda achieves superior performance, especially on tasks whose answer structures are easily corrupted by sequential tuning.", "date": "2026-06-01", "categories": []}
{"id": "2606.03598", "title": "PHASER: Phase-Aware and Semantic Experience Replay for Vision-Language-Action Models", "abstract": "Vision-Language-Action (VLA) models have achieved remarkable success in language-conditioned robotic manipulation. However, deploying these models in open-ended environments requires continuously acquiring novel skills, a process that inevitably triggers severe catastrophic forgetting of previously learned behaviors. While experience replay (ER) serves as a standard mitigating strategy, naive uniform sampling fundamentally misaligns with the temporal characteristics of manipulation trajectories. It systematically under-samples brief but causally critical sub-skills, leading to phase starvation, and completely overlooks the varying degrees of forgetting across historical tasks. To overcome these limitations, we introduce PHASER, an architecture-agnostic continual learning framework. PHASER employs a phase-centric capacity allocation to guarantee equal memory support for all sub-skills, coupled with a multi-modal interference routing strategy that dynamically prioritizes historical phases at high risk of forgetting. Furthermore, to enable fully autonomous lifelong adaptation, we integrate Auto-PC, a lightweight pipeline combining unsupervised action-signal change-point detection with VLM-based semantic verification to extract temporal boundaries without intensive manual supervision. Evaluated across three VLA backbones on LIBERO continual learning suites, PHASER yields substantial empirical improvements, increasing Average Success Rate (ASR) by up to 31% over matched-budget ER and achieving an 87.8% final ASR on the LIBERO-Goal CL setting.", "date": "2026-06-02", "categories": []}
{"id": "2606.04060", "title": "Weakly Supervised Incremental Segmentation via Semantic Anchors and Spatial Arbitration", "abstract": "Weakly Incremental Learning for Semantic Segmentation (WILSS) suffers from the continuous introduction of noisy supervision, which progressively corrupts class-level representations, leading to severe feature drift and semantic corruption, thereby causing newly learned classes to overwrite old ones. To address these issues, we propose a drift-resilient WILSS approach, named SASA, designed to stabilize semantic learning via Semantic Anchors and Spatial Arbitration. Specifically, at the representation level, we introduce semantic anchors of learnable tokens as rigid class-level references to preserve long-term semantic identity. Complementary to this, an elastic residual adaptation facilitates controlled, instance-specific refinement, ensuring a stable yet flexible learning trajectory. At the supervision level, we develop a Spatial Label Arbitration mechanism that performs geometry-aware decisions to directly filter unreliable signals and enforce a strict \"one object, one class\" constraint. By synergistically stabilizing representations and improving supervision reliability, SASA effectively mitigates feature drift under weak supervision. Extensive experiments on standard benchmarks demonstrate that our approach consistently outperforms existing state-of-the-art methods, particularly in challenging multi-step incremental settings. The code is available at https://github.com/ZhonggaiWang/SASA.", "date": "2026-06-02", "categories": []}
{"id": "2606.04528", "title": "Optical-Guided Neural Collapse for SAR Few-Shot Class Incremental Learning", "abstract": "Few-shot class-incremental learning (FSCIL) in synthetic aperture radar imagery presents unique challenges due to severe data scarcity and SAR-specific variability. In particular, strong azimuth sensitivity in SAR induces large intra-class variation and inter-class confusion, and FSCIL sequential updates further lead to catastrophic forgetting of previously learned classes. Inspired by neural collapse, we propose an optical-guided SAR FSCIL framework, which derives orthogonal feature subspaces from a data-rich optical ATR dataset and uses them as geometric priors to guide SAR feature learning. SAR features are projected onto these orthogonal subspaces via principal angle constraints, effectively transferring discriminative structure from the optical to the SAR domain. Specifically, our projection loss and the classifier loss optimized with a frozen simplex-ETF geometry jointly induce neural collapse by concentrating features around class means while maintaining large inter-class angles. We evaluate the approach on a benchmark comprising an optical ATR dataset and a SAR ATR dataset with 24 target classes, organized into a base training session and seven incremental sessions. Compared with recent FSCIL methods including NCFSCIL and so on, our method achieves the highest final accuracy and a favorable trade-off between final performance and performance degradation. Moreover, neural collapse metrics show improved intra-class compactness and inter-class separability, indicating that the learned features more closely approximate the ideal simplex-ETF geometry.", "date": "2026-06-03", "categories": []}
{"id": "2606.04703", "title": "Rethinking Continual Experience Internalization for Self-Evolving LLM Agents", "abstract": "Experience internalization converts contextual experience from past interactions into reusable parametric capability, offering a promising path toward continual learning in large language models (LLMs). While prior work has predominantly focused on single-iteration transfer, we discover that under multi-iteration experience learning, existing methods suffer from a progressive capability collapse rather than compounding improvement. We systematically examine this failure through three vital dimensions of experience internalization: (1) Experience Granularity: We find that principle-level experience is more durable than instance-level experience, as it effectively abstracts transferable strategies away from trajectory-specific details. (2) Experience Injection Pattern: Our analysis reveals that step-wise injection significantly outperforms global injection by aligning experience with intermediate decision states, a property that is critical for long-horizon tool use. (3) Internalization Regime: We demonstrate that off-policy context-distillation on high-quality teacher trajectories provides a substantially more stable training signal than on-policy context-distillation, which is inherently limited by local corrections on student-induced flawed states. Together, these insights yield a simple yet robust recipe for stable and sustainable experience internalization, providing concrete guidance for engineering self-evolving and continually learning LLMs.", "date": "2026-06-03", "categories": []}
{"id": "2606.04797", "title": "Crafting Your Evolving Dreams: Concept-Incremental Versatile Customization", "abstract": "Custom diffusion models (CDMs) have garnered significant interest owing to their remarkable capacity for generating personalized concepts. However, the majority of CDMs unrealistically presume that the user's collection of personalized concepts is static and incapable of incremental growth over time. Furthermore, they exhibit significant catastrophic forgetting and concept neglect of previously learned concepts when incrementally learning a sequence of new ones. To resolve the above challenges, we develop a novel Continually Customizable Diffusion Model (CCDM), enabling users to perform concept-incremental versatile customization. Specifically, we design an attribute-decoupled LoRA (AD-LoRA) module and a relevance-guided AD-LoRA aggregation strategy to mitigate catastrophic forgetting. They can preserve concept-specific attributes of each task and leverage beneficial inter-task correlations to enhance the continual learning of new customization tasks. Additionally, to address the challenge of concept neglect, we propose a controllable regional context synthesis strategy that performs multi-concept composition in alignment with user-provided conditions. This strategy enhances the overall consistency in multi-concept synthesis by guaranteeing semantic independence between user-defined regions and their smooth boundary transitions. Experiments show our CCDM exhibits significant improvements over baseline methods.", "date": "2026-06-03", "categories": []}
{"id": "2606.04815", "title": "Learning While Acting: A Skill-Enhanced Test-Time Co-Evolution Framework for Online Lifelong Learning Agents", "abstract": "Lifelong learning is essential for Large Language Model (LLM) agents operating in dynamic, interactive environments. However, existing lifelong learning agents for long-horizon tasks typically depend on discrete skill or past experiences retrieval with static parameters during inference, which prevents them from continuously internalizing test-time feedback like human learners. To bridge this gap, we propose Skill-enhanced Test-Time Co-Evolution (\\texttt{LifeSkill}), a two-stage reinforcement learning framework for Online Lifelong Learning Agents. Specifically, we design Verifier-Guided Skill Learning that addresses the lack of direct supervision for skill extraction by rewarding candidate skills according to the average verifier success of multiple skill-conditioned policy rollouts, encouraging the model to generate skills that are useful for solving tasks rather than merely plausible in text. Furthermore, we introduce Online Skill Internalization, which continuously improves the policy model during test-time interaction by transforming skill-conditioned trajectories into reward signals. This enables the agent to directly internalize reasoning capabilities into its parameters, avoiding the context bloat of experience retrieval. Experiments on LifelongAgentBench show that LifeSkill improves average performance by 7 absolute points by comparing with existing lifelong agent baselines.", "date": "2026-06-03", "categories": []}
{"id": "2606.05115", "title": "Continual Visual and Verbal Learning Through a Child's Egocentric Input", "abstract": "Children learn the meanings of words from a continuous, temporally structured stream of egocentric experience. Recent work shows that neural networks can also learn word-referent mappings from a child's egocentric video recordings, but they cycle through the shuffled data for hundreds of epochs, contrasting with how children actually encounter their environment. We introduce BabyCL, a continual multimodal learning framework that processes the SAYCam dataset in a single chronological pass, combining streaming visual representation learning with an image-text contrastive objective. BabyCL combines a multi-stage temporal segmentation of the stream with a dual replay buffer that independently manages visual and multimodal histories, and it is jointly trained with three contrastive losses on a shared backbone. Under a matched optimization budget, BabyCL outperforms streaming learning baselines on the SAYCam Labeled-S 4AFC benchmark, substantially narrowing the gap to an upper bound of offline training. Ablations show that the gains are robust to the length of the online temporal segmentation window and the eviction rule of the replay buffer. Together, these results show that meaningful word-referent mappings can emerge under training conditions much closer to a child's actual experience.", "date": "2026-06-03", "categories": []}
{"id": "2512.03627", "title": "MemVerse: Multimodal Memory for Lifelong Learning Agents", "abstract": "Despite rapid progress in large-scale language and vision models, AI agents still suffer from a fundamental limitation: they cannot remember. Without reliable memory, agents catastrophically forget past experiences, struggle with long-horizon reasoning, and fail to operate coherently in multimodal or interactive environments. We introduce MemVerse, a model-agnostic, plug-and-play memory framework that bridges fast parametric recall with hierarchical retrieval-based memory, enabling scalable and adaptive multimodal intelligence. MemVerse maintains short-term memory for recent context while transforming raw multimodal experiences into structured long-term memories organized as hierarchical knowledge graphs. This design supports continual consolidation, adaptive forgetting, and bounded memory growth. To handle real-time demands, MemVerse introduces a periodic distillation mechanism that compresses essential knowledge from long-term memory into the parametric model, allowing fast, differentiable recall while preserving interpretability. Extensive experiments demonstrate that MemVerse significantly improves multimodal reasoning and continual learning efficiency, empowering agents to remember, adapt, and reason coherently across extended interactions.", "date": "2025-12-03", "categories": []}
{"id": "2606.01767", "title": "EvoBrain: Continual Learning of EEG Foundation Models Across Heterogeneous BCI Tasks", "abstract": "Electroencephalography (EEG) is the cornerstone of non-invasive brain-computer interfaces (BCIs), yet conventional decoding relies on fragmented, task-specific architectures that severely limit cross-task scalability. While EEG foundation models pre-trained on massive corpora promise universal brain decoding, current post-training depends on task-isolated fine-tuning. This static paradigm restricts knowledge transfer across heterogeneous tasks, hinders model scalability, and incurs computational and storage overheads that scale linearly with task count. To overcome these bottlenecks, we formulate downstream adaptation as a cross-task continual learning problem and propose EvoBrain, a dynamic, task-aware continual learning framework for unified EEG decoding. EvoBrain addresses the plasticity-stability trade-off via two complementary components: (1) Neuro-Spectral Task Normalization (NSN) aligns incoming tasks with historical statistics while recalibrating spectral responses to handle distributional and neuro-spectral shifts; and (2) Response-Affinity Distillation (RAD), combined with time-dependent replay, preserves old-task response geometry and promotes selective knowledge transfer between spectrally compatible tasks, effectively mitigating forgetting. Extensive evaluations across six distinct BCI tasks demonstrate that EvoBrain consistently surpasses state-of-the-art methods across diverse foundation backbones, optimally balancing plasticity and stability. To our knowledge, this work pioneers cross-task continual learning in the EEG domain, advancing the realization of a unified, one-for-all brain decoding system.", "date": "2026-06-01", "categories": []}
{"id": "2606.02461", "title": "AgentCL: Toward Rigorous Evaluation of Continual Learning in Language Agents", "abstract": "Language agents spend substantial inference time solving individual tasks, yet the experience acquired in one episode is often underutilized in future episodes. Continual learning expects an agent to accumulate reusable experience across a stream of tasks, improve over time, and avoid interference from irrelevant experiences. Unfortunately, existing benchmarks struggle to evaluate continual learning in language agents rigorously. Most efforts focus on retrieval and reasoning over long-context conversations or documents, while recent lifelong-adaptation benchmarks often rely on naive task streams with limited analysis of cross-task relationships, making it difficult to understand what an agent learns and reuses over time. This paper presents an evaluation framework AgentCL for continual learning in agents, centered on controlled task streams and metrics for transfer gains. AgentCL constructs compositional streams where earlier sub-solutions, evidence, or workflows are intentionally reusable in later tasks, and contrasts them with naive streams where such reusability is not guaranteed. We use the benchmark to evaluate non-parametric memory designs for continual learning. To diagnose how memory design choices affect continual learning, we develop MemProbe, a probing method that stores interactions, insights, and skills, while filtering unreliable experiences during consolidation. Empirical analysis across coding, deep research, and language understanding/reasoning tasks shows that naive streams offer limited ability to distinguish memory designs, whereas controlled streams more clearly distinguish their plasticity. Meanwhile, naive and held-out settings often yield limited gains and can expose memory-induced degradation. These results highlight the need for stronger memory designs that balance plasticity and stable reuse.", "date": "2026-06-01", "categories": []}
{"id": "2606.02860", "title": "Forgetting is Not Erasure: Recovering Latent Knowledge via Transport Keys", "abstract": "Catastrophic forgetting is often framed as a representational problem: after sequential training, a model appears to lose the features that supported performance on earlier tasks. We challenge the stronger form of this view. Across controlled continual-learning settings, we find that a significant portion of apparent forgetting can be attributed to interface drift between internal stages rather than permanent erasure of task-relevant computation. We study this phenomenon through a stitched evaluation protocol that combines early computation from a post-update network with late computation from its predecessor, optionally mediated by a compact, task-specific transport key. We describe transport keys at a systems level as compact interface-alignment operators estimated from a small set of paired anchor activations and evaluated through model stitching. On split CIFAR-100 with a ResNet-style network, transport keys recover most of the original Task A performance after sequential training on Task B. On a compact vision transformer, we observe a similar recovery pattern. These results suggest that continual learning may require better mechanisms for indexing and re-accessing latent computations, not only methods that prevent weight change.", "date": "2026-06-01", "categories": []}
{"id": "2606.03321", "title": "Validation-Gated Multi-Agent Governance for Online Adaptation of Thermal-Hydraulic Surrogate Models under Operating-Regime Shift", "abstract": "Artificial-intelligence surrogates can support second-by-second thermal-hydraulic forecasting, but models selected and frozen offline may become condition-locked once deployed outside their pretraining envelope. This study develops a guarded continual-adaptation framework for experimental thermal-hydraulic loop data in which role-separated agents - Monitor, Diagnosis, Adaptation, Safety-Auditor, and Orchestrator - diagnose error signatures, prioritize candidate model families, and review promotions, while deterministic champion-challenger gates and background shadow learning retain final authority over model replacement. Seven surrogate families were screened by blocked three-fold cross-validation, and a temporal Fourier neural operator was selected as the initial champion for 60-s-history-to-10-s-trajectory forecasting on two held-out transients, with three seeds per adaptive mode. Static deployment gave a channel-averaged MAE of 7.06 and a 56.8% warning-exceedance ratio; rule-based adaptation reduced MAE to 6.54, whereas shadow refresh alone remained close to Static. The MA-Full mode, in which the role-separated multi-agent council reviews every evaluated stream step, achieved the lowest mean error, 5.72, and 35.8% exceedance, corresponding to a 19.0% improvement over Static. Paired bootstrap intervals against Static excluded zero, although intervals among adaptive modes overlapped and the six paired units limit broad statistical claims. Validated promotions from the neural operator to Transformer and graph neural network indicate that logged, gate-controlled adaptation can support auditable surrogate evolution while deterministic gates retain deployment authority.", "date": "2026-06-02", "categories": []}
{"id": "2606.03843", "title": "Re-Evaluating Continual Learning with Few-Shot Adaptation", "abstract": "Continual learning methods aim to maximize the stability and plasticity of machine learning models that are trained on a sequence of tasks. The standard measure of stability (i.e., forgetting) is the 0-shot performance of a model on previously learned tasks, and plasticity, the performance on the most recently learned task. However, 0-shot evaluation does not fully measure a model or method's ability to retain learned information or adapt quickly to new information, as it requires perfect recall across multiple tasks. In this paper, we propose few-shot evaluation as a more comprehensive assessment of the stability and plasticity of a continual learning system. We conduct a fine-grained assessment on task sequences for continual image classification and find that this paradigm produces novel insights into the performance of popular continual learning strategies. Through few-shot evaluation with a novel metric -- per-shot plasticity -- we show that adding `foresight' to continual learning methods via the meta-learning of a short sequence of future tasks induces learning-to-learn behavior over the task sequence.", "date": "2026-06-02", "categories": []}
{"id": "2606.03939", "title": "FlashbackCL: Mitigating Temporal Forgetting in Federated Learning", "abstract": "Federated Learning (FL) of foundation and edge models increasingly targets deployments where client data distributions drift over time, yet existing forgetting-mitigation methods assume each client's distribution is stationary. Flashback, the strongest recent FL method against cross-client (spatial) forgetting, uses monotonically accumulating per-class label counts as a knowledge proxy; this proxy becomes miscalibrated under temporal distribution shift and anchors the global model to an outdated class balance. We formalise temporal forgetting in FL with a per-phase metric isolated from protocol-level fluctuations and propose Flashback Continual Learning (FlashbackCL), a drop-in extension of Flashback with (i) temporally-decayed label counts; (ii) a device-aware replay buffer with Class-Balanced Reservoir Sampling (CBRS); and (iii) server-side active coreset curation on the public distillation set. The results show that FlashbackCL achieves 6.9% to 10.0% relative improvement relative to Flashback, on CIFAR-10 with 50 clients and three controlled temporal shift modes, while simultaneously reducing temporal forgetting by up to 68%. A 5-variant ablation identifies CBRS replay as the critical component. FlashbackCL also improves Flashback by 3.5 points on stationary CIFAR-100, suggesting that class-balanced replay regularises spatial heterogeneity as well as temporal shift.", "date": "2026-06-02", "categories": []}
{"id": "2606.03979", "title": "Language Models Need Sleep: Learning to Self-Modify and Consolidate Memories", "abstract": "The past few decades have witnessed significant advances in the design of machine learning algorithms, from early studies on task-specific shallow models to more general deep Large Language Models (LLMs). Despite showing promising results in tasks that require instant prediction or in-context learning, existing models lack the ability to continually learn and effectively transfer their temporal in-context knowledge to their long-term parameters. Inspired by human learning process, we introduce a ''Sleep'' paradigm that allows the models to continually learn, distill their short-term fragile memories into stable long-term knowledge with replay, and recursively improve themselves with ''Dreaming'' process. In more detail, sleep consists of two stages: (1) Memory Consolidation: an upward distillation process, called Knowledge Seeding, where the memories of a smaller-self are distilled into a larger network to provide more capacity while preserving the knowledge. As a proof of concept, we present a new Generalized Distillation process for {Knowledge Seeding} (i.e., the combination of on-policy distillation with Reinforcement Learning (RL)-based imitation learning); (2) Dreaming: a self-improvement phase, where the model uses RL to generate a curriculum of synthetic data to rehearse new knowledge and refine existing capabilities without human supervision. Our experiments on long-horizon, continual learning, knowledge incorporation, and few-shot generalization tasks support the importance of the sleep stage.", "date": "2026-06-02", "categories": []}
{"id": "2603.11653", "title": "Simple Recipe Works: Vision-Language-Action Models are Natural Continual Learners with Reinforcement Learning", "abstract": "Continual Reinforcement Learning (CRL) for Vision-Language-Action (VLA) models is a promising direction toward self-improving embodied agents that can adapt in openended, evolving environments. However, conventional wisdom from continual learning suggests that naive Sequential Fine-Tuning (Seq. FT) leads to catastrophic forgetting, necessitating complex CRL strategies. In this work, we take a step back and conduct a systematic study of CRL for large pretrained VLAs across diverse lifelong RL benchmarks. We find that, contrary to established belief, simple Seq. FT with low-rank adaptation (LoRA) is remarkably strong: it achieves high plasticity, exhibits little to no forgetting, and retains strong zero-shot generalization, frequently outperforming more sophisticated CRL methods. Through detailed analysis, we show that this robustness arises from a synergy between the large pretrained model, parameter-efficient adaptation, and on-policy RL. Together, these components reshape the stability-plasticity trade-off, making continual adaptation both stable and scalable. Our results position Sequential Fine-Tuning as a powerful method for continual RL with VLAs and provide new insights into lifelong learning in the large model era. Code is available at github.com/UT-Austin-RobIn/continual-vla-rl.", "date": "2026-03-12", "categories": []}
{"id": "2605.29183", "title": "TIMEGATE: Sustainable Time-Boxed Promotion Gates for Continual ML Adaptation Under Resource Constraints", "abstract": "As machine learning(ML) systems evolve to continual adaptation, each re-training cycle uses compute, annotation, and energy. We introduce TIMEGATE, a policy layer managing adaptation by budgeting time, labeling, training, and evaluation. TIMEGATE emits a metric-availability signal M for partial vs. full-evaluation decisions. We validate: (i) labeling outperforms training by 2.3x on Adult tabular; (ii) it transfers to LLaMA-3.1-8B + QLoRA on SST-2 (accuracy 0.80 to 0.96; M =1 in 35/36 runs); (iii) M is informative, 28-cell sensitivity shows M drops to 0.81 at tight thresholds; (iv) 100-cycle simulation achieves 66% evaluation-compute savings with no silent mis-promotions; (v) 10%-slice evaluation on LLaMA uses 89% less wall-clock and energy on a single H200 (ratios agree to 0.2%).", "date": "2026-05-27", "categories": []}
{"id": "2606.00143", "title": "Regime-Adaptive Continual Learning for Portfolio Management", "abstract": "Financial markets are inherently non-stationary, exhibiting frequent regime shifts and structural changes that render traditional Portfolio Management (PM) approaches ineffective. Existing remedies, such as rolling-window retraining and naive online fine-tuning, are hindered by high computational costs and insufficient knowledge utilization, respectively, resulting in low returns and limited adaptability. Continual learning (CL) offers a promising paradigm by enabling trading agents to accumulate and transfer knowledge across sequential tasks. In this paper, we propose \\textbf{Re}gime-aware \\textbf{C}ontinual \\textbf{A}daptive \\textbf{P}ortfolio management (\\textbf{ReCAP}), a novel framework that integrates CL into PM to address the challenges of dynamic financial environments. ReCAP employs an adaptive regime detection module to segment historical market data into variable-length regimes, enabling regime-specific learning of policy vectors and the construction of a policy library. During continual trading, a regime-gate module adaptively combines policy vectors from the library based on the current market state, facilitating rapid adaptation to newly detected regimes. Only the regime-gate and the current regime's policy vector are continually updated to preserve useful knowledge effectively. Extensive experiments on five real-world datasets demonstrate that ReCAP consistently outperforms popular baselines, achieving superior returns in long-term investment horizons and rapid adaptation to regime shifts.", "date": "2026-05-29", "categories": []}
{"id": "2606.00284", "title": "Parameter Alignment Mitigates Catastrophic Forgetting in Multilingual Expert Language Models", "abstract": "While continual pretraining~(CPT) is a practical way to extend large language models to new languages, na\\\"ive finetuning on targeted data erodes existing capabilities through catastrophic forgetting. Organizing training around language families reduces cross-language interference but cannot alone prevent forgetting of the general knowledge needed for downstream tasks. We link this forgetting to parameter drift in multilingual CPT and present a suite of five layer-aware parameter alignment strategies: hard layer freezing, soft regularization, post-hoc weight reversion, and model merging. We systematically compare our alignment strategies against two unregularized CPT baselines on benchmarks spanning 32 training languages from five language families, plus held-out languages, across four evaluation axes: perplexity, reading comprehension, physical reasoning, and translation. Parameter alignment substantially reduces forgetting at minimal cost to language acquisition: layer freezing and regularization best preserve comprehension, whereas post-hoc reversion yields the strongest translation gains. Together, these results map the acquisition--forgetting frontier for family-expert CPT and offer practical deployment guidelines pairing each strategy to the tasks it best serves.", "date": "2026-05-29", "categories": []}
{"id": "2606.00382", "title": "CRMA: A Spectrally-Bounded Backbone for Modular Continual Fine-Tuning of LLMs", "abstract": "Sequential fine-tuning of large language models forces a choice: let the shared substrate keep learning and accept catastrophic forgetting, or freeze it after task one and foreclose cross-task refinement. Per-task adapter methods (LoRAHub, AdapterFusion, PackNet, Progressive Networks) take the second path. We introduce CRMA (Constrained Residual Mixing Adapter), a residual adapter whose internal mixing matrix M is doubly-stochastic at every forward pass via Sinkhorn normalization, so by Birkhoff's theorem ||M||_2 <= 1 holds by construction -- a structural bound, not a penalty. CRMA's spectrally bounded backbone provides a continuously trained shared substrate that earlier modular methods could not, while preserving their forgetting guarantees. On Mistral-7B across 5 sequential domains and 3 seeds, modular per-task LoRA on a CRMA backbone reduces loss-relative drift from +42.96% +/- 5.5 (naive sequential fine-tuning) to -0.17% +/- 0.17, with disjoint per-seed ranges, and improves prior-task holdout loss by 1.99% +/- 0.54 over a matched frozen-substrate baseline. Three independent experimental setups (Mistral-7B 4-domain controlled ablation, TinyLlama 3-domain contamination-controlled replication, Mistral-7B cross-domain probes at 7B) all show positive backward transfer -- without replay buffers, without growing per-task memory, and without distillation. An inference-time ablation on Gemma-2-9B confirms CRMA mediates access to sequentially trained knowledge: 98/100 vs. 38/100 on the same weights and same questions with only CRMA injection toggled. 867 logged training steps verify ||M||_2 = 1.0 within float32 precision (max deviation 1.2 x 10^-7). The forgetting-prevention effect holds across 1.1B-9.2B parameters and four architecture families.", "date": "2026-05-29", "categories": []}
{"id": "2606.00400", "title": "Dynamic Proxy-Mixing: Transferring Replay Controllers from Small to Large Models for Continual Instruction Tuning", "abstract": "Continual instruction tuning updates a language model through a sequence of new domains, yet each update can progressively erode previously learned capabilities and alignment behavior. Replay is the standard mitigation, but fixed replay ratios are inherently limited because the optimal mixture varies with the current domain, the training stage, and the evolving vulnerability of prior behaviors. We propose PROX-YMIX, a framework that learns a dynamic replay controller on a small proxy model and transfers the frozen controller to a larger target. The controller never observes future tasks and constructs its state from normalized validation losses and their temporal dynamics, producing a masked mixture over the current task and accessible replay buffers. Our core empirical hypothesis is forgetting mirroring: task vulnerability rankings remain largely consistent across model scales even when absolute loss magnitudes differ. We validate this assumption empirically before transferring controllers across scales. On LLaMA-3-8B across five continual instruction tuning sequences, PROXYMIX improves average accuracy by 3.4 points, reduces final forgetting by 3.5 points, and raises safety score by 5.8 points over the strongest non-oracle baseline, at roughly 50x lower policy learning cost than Oracle Target RL. The framework is leakage free and architecture independent at the interface level, and we also identify settings where the proxy assumption breaks down, highlighting limitations for robust deployment.", "date": "2026-05-29", "categories": []}
{"id": "2606.00426", "title": "Canonicalized Stable-List Replay for Private Federated Continual Learning over Language-Model Embeddings", "abstract": "Federated continual learning (FCL) lets distributed clients adapt language-model heads to evolving NLP tasks without sharing raw text. Under user-level differential privacy (DP), replay-based continual learning faces a structural obstacle: clients can release only small noisy lists of candidate replay summaries, and those lists are unordered across clients. We introduce Canonicalized Stable-List Replay (CSLR), where clients privately produce candidate replay distributions over a shared sentence-embedding space and the server aligns them using signatures induced by public anchor sentences. The anchors provide identifiability for aggregation rather than additional replay data. We prove that, under an observable anchor-signature margin, $O(\\log(N/\\eta)/p)$ anchors distinguish $N$ candidate list elements with probability at least $1-\\eta$, and we give a scoped anchorless non-identifiability result for unordered-label oracle models. Across five seeds on continual classification, NER, and dialogue benchmarks, CSLR improves the final average task metric by 3.9--5.6 points over the strongest non-CSLR DP baseline at $\\eps=4$ under the reported replay-release budget, while also outperforming Hungarian and optimal-transport matchers. The formal privacy guarantee covers replay release; end-to-end private training additionally requires composition with a private optimizer for task-head updates.", "date": "2026-05-29", "categories": []}
{"id": "2606.00880", "title": "Task diversity produces systematic transfer but inhibits continual reinforcement learning", "abstract": "Continual reinforcement learning aims to produce agents that learn not only to improve at their current tasks but also to adapt as task distributions change. Training an agent on many diverse tasks can induce zero-shot generalization, but previous work generally evaluates this generalization after training -- with frozen weights. Whether task diversity also improves an agent's ability to continue learning across distribution shifts remains unclear. We introduce Banyan, a GPU-accelerated continual RL domain in which task diversity factors into three independently controllable axes: the map layouts an agent must navigate, the objects it must interact with, and the hierarchical structures of sub-goal dependencies. Across individual distribution shifts, increasing diversity along each axis causes agents to begin training on the new tasks near the performance attained on the previous one, even when the shift changes the structure of the optimal policy. However, as the number of shifts increases, this local transfer does not by itself yield sustained continual learning: longer-horizon tasks plateau, and earlier task distributions are forgotten after later training. Banyan is a benchmark for studying when controlled task diversity produces transferable learning, when that transfer persists, and where it falls short of proper continual learning.", "date": "2026-05-30", "categories": []}
{"id": "2606.01863", "title": "Continual Learning as a Multiphase Moving-Boundary Problem", "abstract": "Continual learning struggles to balance retaining past knowledge with absorbing new tasks. Stefan-CL elegantly resolves this stability-plasticity dilemma through the physics of melting. It frames consolidated knowledge as a protected \"solid\" and unused capacity as an adaptable \"liquid.\" As the network learns, this boundary expands, governed by a \"latent heat\" tuning dial. By mathematically freezing the learned interior, Stefan-CL cuts forgetting to near zero, matching memory-heavy baselines without storing raw data, forging a beautiful, physics-grounded path for AI.", "date": "2026-06-01", "categories": []}
{"id": "2606.01873", "title": "G2LoRA: Gradient Orthogonal Low-Rank Adaptation Framework for Graph Continual Learning on Text-Attributed Graphs", "abstract": "LLM-as-Aligner has emerged as a prevalent pre-training paradigm for Text-Attributed Graphs(TAGS), aligning graph and text modalities into a shared embedding space via CLIP-style contrastive learning. While effective on individual downstream tasks, we observe severe catastrophic forgetting when such models are sequentially fine-tuned on streaming tasks. Although parameter-efficient fine-tuning alleviates forgetting to some extent, it remains insufficient to resolve task interference and ineffective knowledge transfer. In this work, we study graph continual learning for LLM-as-Aligner models on TAGs, with the goal of mitigating interference while promoting positive transfer across tasks. This setting introduces two fundamental challenges: (1) heterogeneous downstream tasks induce shifting optimization objectives, hindering unified fine-tuning; and (2) graph and text encoders exhibit different sensitivities to adaptation, making uncoordinated updates prone to misalignment. To address these challenges, we propose G2LoRA, a continual learning framework for TAGs. G2LoRA unifies node-, link-, and graph-level tasks under a single graph--text alignment objective, and enables consistent optimization across domain/class/task incremental modes. To reduce task interference while encouraging positive transfer, G2LoRA performs category-aware gradient projection in structured subspaces, resolving conflicting updates and enabling conditional backward transfer to balance forward and backward knowledge flow. To further prevent cross-modal drift, G2LoRA introduces gradient magnitude modulation to coordinate update rates between graph and text encoders. Extensive experiments on benchmark datasets demonstrate that G2LoRA consistently outperforms strong baselines across different backbone architectures, achieving superior continual performance and transferability.", "date": "2026-06-01", "categories": []}
{"id": "2606.02322", "title": "Repurposing Adversarial Perturbations for Continual Learning: From Defense to Active Alignment", "abstract": "In dynamic environments, large language models need to keep adapting to new tasks, but continual learning often suffers from forgetting, limited transfer, and vulnerability to adversarial perturbations. To address this, we present AdvCL, which repurposes adversarial perturbations as a geometric control signal for stable continual adaptation. AdvCL combines three plug-in modules: Intra-Smooth promotes local smoothness via small adversarial perturbations; Proto-Clip uses similarity clipping to prevent excessive alignment to current task prototype; and Inter-Align applies directional alignment toward previous task prototype to reduce representational gaps. Experiments show consistent gains in both standard performance and robustness, with lower forgetting and stronger transfer. We further analyze key mechanisms by quantifying the sensitivity of Intra-Smooth to perturbation settings and the effect of Inter-Align on task similarity and geometric distance. In summary, the modules provide complementary gains when combined, and each can also be integrated individually into diverse CL paradigms, including replay, regularization, and dynamic architectures, thereby offering a geometric control mechanism for continual learning.", "date": "2026-06-01", "categories": []}
{"id": "2606.02502", "title": "CRAM: Centroid-Routing and Adaptive MoE for Multimodal Continual Instruction Tuning", "abstract": "Multimodal Large Language Models (MLLMs) unify heterogeneous vision-language tasks under a shared generative framework via instruction tuning, yet real-world deployment demands continuous capability expansion, making Multimodal Continual Instruction Tuning (MCIT) essential. Existing methods either update all tasks with a shared parameter set or allocate dedicated modules for each new task. Shared updates force heterogeneous tasks to compete, causing forgetting of learned capabilities. Conversely, isolated expansion prevents interference but severely limits parameter efficiency over long task streams. To address this dilemma, we propose CRAM. Specifically, by isolating task-specific patterns into independent modules, CRAM mitigates catastrophic forgetting across tasks. To further boost parameter efficiency, we utilize adaptive-rank instantiation to identify the capability gap between existing expert capability and new task demands, and dynamically allocate only the necessary parameters. To ensure stable reuse among tasks, centroid-guided routing recognizes and activates existing experts' capabilities, while an orthogonality penalty confines new updates to task-specific directions, preventing re-learning general capability. Extensive experiments across diverse benchmarks consistently demonstrate its superiority over existing methods.", "date": "2026-06-01", "categories": []}
{"id": "2506.08255", "title": "SHIELD: Secure Hypernetworks for Incremental Expansion Learning Defense", "abstract": "Continual learning under adversarial conditions remains an open problem, as existing methods often compromise either robustness, scalability, or both. We propose a novel framework that integrates Interval Bound Propagation (IBP) with a hypernetwork-based architecture to enable certifiably robust continual learning across sequential tasks. Our method, SHIELD, generates task-specific model parameters via a shared hypernetwork conditioned solely on compact task embeddings, eliminating the need for replay buffers or full model copies and enabling efficient over time. To further enhance robustness, we introduce Interval MixUp, a novel training strategy that blends virtual examples represented as $\\ell_{\\infty}$ balls centered around MixUp points. Leveraging interval arithmetic, this technique guarantees certified robustness while mitigating the wrapping effect, resulting in smoother decision boundaries. We evaluate SHIELD under strong white-box adversarial attacks, including PGD and AutoAttack, across multiple benchmarks. It consistently outperforms existing robust continual learning methods, achieving state-of-the-art average accuracy while maintaining both scalability and certification. These results represent a significant step toward practical and theoretically grounded continual learning in adversarial settings.", "date": "2025-06-09", "categories": []}
{"id": "2506.21035", "title": "Little by Little: Continual Learning via Incremental Mixture of Rank-1 Associative Memory Experts", "abstract": "Continual learning (CL) with large pre-trained models aims to incrementally acquire knowledge without catastrophic forgetting. Existing LoRA-based Mixture-of-Experts (MoE) methods expand capacity by adding isolated new experts while freezing old ones, but still suffer from redundancy, interference, routing ambiguity, and consequent forgetting. We investigate the issues stemming from coarse-grained expert granularity. Coarse-grained experts (e.g., high-rank LoRA) encode low-specialty information, leading to expert duplication/interference and routing degradation/confusion as experts accumulate. In this work, we propose MoRAM (Mixture of Rank-1 Associative Memory). Grounded in the view that weight matrices act as linear associative memories, MoRAM achieves CL as incremental expansion of reusable atomic rank-1 experts as memory. Each rank-1 adapter acts as a fine-grained MoE expert or an associative memory unit. By viewing rank-1 experts as key-value memory pairs, we eliminate explicit MoE-LoRA routers with self-activation, where each memory atom evaluates its relevance via its intrinsic key. The inference process thus becomes a content-addressable retrieval and recall over the incrementally accumulated memory of learning snapshots. Extensive experiments on CLIP and LLMs show that MoRAM significantly outperforms state-of-the-art methods, achieving a better plasticity-stability trade-off, stronger generalization, and reduced forgetting. Project Page: https://artificer-ai-lab.github.io/MoRAM/.", "date": "2025-06-26", "categories": []}
{"id": "2509.22335", "title": "Spectral Collapse Drives Loss of Plasticity in Deep Continual Learning", "abstract": "We investigate why deep neural networks suffer from loss of plasticity in continual learning, and thus fail to learn new tasks without reinitializing parameters. We show that this failure is preceded by Hessian spectral collapse at new-task initialization, where meaningful curvature directions vanish and gradient descent becomes ineffective. Analyzing a linearized ReLU network, we derive explicit $\\epsilon$-rank conditions for successful training and prove that the loss-weighted Gram matrix is spectrally equivalent to the Generalized Gauss-Newton approximation, thereby relating NTK dynamics to Hessian curvature. Targeting spectral collapse directly, we then discuss the Kronecker factored approximation of the Hessian, which motivates two regularization enhancements: maintaining high effective feature rank and applying L2 penalties. Experiments on continual supervised and reinforcement learning tasks confirm that combining these two regularizers effectively preserves plasticity.", "date": "2025-09-26", "categories": []}
{"id": "2605.19145", "title": "PMF-CL: Pareto-Minimal-Forgetting Continual Learner for Conflicting Tasks", "abstract": "In the literature, many continual learning (CL) algorithms have been proposed to address the issue of catastrophic forgetting in ML models (i.e., learning new tasks leads to the loss of performance on previously learned tasks). Although all CL approaches use some form of memory to retain information about past tasks, a grounded understanding of what information needs to be stored to minimize catastrophic forgetting remains elusive. Recently, it has been recognized that under the strong assumption of the existence of a common global minimizer over all tasks, catastrophic forgetting can be completely avoided. However, in practice, tasks rarely have a common global minimizer, and a certain amount of forgetting is inevitable. In this paper, we propose a foundational framework for principled and systematic CL of conflicting tasks using a multi-task learning (MTL) perspective. The approach is based on finding Pareto-optimal solutions, i.e., the solutions which, by definition, minimally forget the previous tasks in the Pareto sense. We derive Pareto-minimal-forgetting CL algorithms for linear and basis-function regression, and general loss functions which have a quadratic upper bound, e.g., logistic regression. For quadratic problems, PMF-CL uses memory-efficient iterative updates with a static memory footage of $\\mathcal{O}(d^2)$ for models with $d$ parameters.", "date": "2026-05-18", "categories": []}
{"id": "2605.31025", "title": "TRACE: Discovering Task-Specific Parameter via Adaptation-Aware Probing for Continual Fine-Tuning", "abstract": "In real-world deployment, LLMs are often adapted continually across tasks to keep LLMs up-to-date in production, where new fine-tuning should preserve previously learned skills. However, indiscriminately mixing tasks can dilute task specialization, while sequential fine-tuning (full-parameter or low rank adaptation) often causes catastrophic forgetting due to destructive overwriting. Replay-based continual tuning and maintaining separate task-specific adapters can mitigate forgetting, but introduce additional compute, storage, and management overhead. Recognizing the redundancy of LLM parameters for any single task, we reframe continual task adaptation as task-specific parameter discovery via adaptation-aware probing: a short warm-start probe exposes a task's adaptation trace, enabling us to identify and isolate the small subset of parameters essential for each task to mitigate catastrophic forgetting. Building on this view, we introduce TRACE, a novel approach for discovering Task-specific paRameters via Adaptation-aware probing for Continual finE-tuning. We perform a short warm-start fine-tune to derive task-specific core parameters by comparing the warm-started and pre-trained models. Core parameters are identified via two strategies: importance scoring (L$_2$ norm and Fisher Information) and specificity analysis (cosine similarity of parameter updates). In continual fine-tuning settings, only the active task's core parameters are updated while others remain frozen, preserving prior knowledge. We conduct extensive experiments across multiple standard benchmarks to demonstrate the superior performance of our proposed method. Additionally, we validate the generalization of our method through a cross-model and scale transferability study, demonstrating a \"small-to-large\" paradigm that guides the fine-tuning of large-scale models under resource constraints.", "date": "2026-05-29", "categories": []}
{"id": "2605.31108", "title": "Remembering by Reconstructing: Domain Incremental Learning With Test-Time Training on Video Streams", "abstract": "In this work we introduce a novel approach to domain incremental learning, adapting models over time to evolving, non-stationary data. In contrast to other works, we do not attempt to avoid catastrophic forgetting, but rather allow it and exploit it. Our model combines a main task head with a self-supervised masked autoencoder (MAE) head. We then learn domain-specific LoRA adapters during incremental training. Each adapter specializes to its domain, naturally inducing forgetting on other domains in both heads. At inference, we perform online test-time training on the self-supervised MAE head to identify which LoRAs best matches the current input, so the model can `remember' the domain again. Our scheme is especially well-suited to real-world streaming data, such as video, where consecutive samples are highly correlated and domain shifts are gradual. We demonstrate our method on domain-incremental action recognition and semantic segmentation tasks.", "date": "2026-05-29", "categories": []}
{"id": "2605.31229", "title": "Beyond Classification: Dynamic Adapter Routing for Continual Multimodal Retrieval", "abstract": "While retrieval is a core function of vision-language models, continually updating these models for retrieval tasks remains critically underexplored. Existing work often approaches continual retrieval through the lens of class-incremental learning (CIL), evaluating both standard CIL methods and retrieval-oriented adaptations in settings that may not fully capture the retrieval-specific dynamics. To address this, we introduce a new, principled evaluation framework for continual multimodal retrieval (CMR) spanning diverse visual domains, and systematically evaluate common approaches within this setting. Our empirical analysis shows that standard CIL methods fail to yield meaningful gains in our more challenging scenario. Therefore, we propose Dynamic Adapter Routing (DAR), a novel approach based on adapters selected through prototype-based routing and combined via model merging.DAR achieves superior performance over the previous baselines and demonstrates strong generalization under out-of-distribution evaluation. Our results highlights the unique challenges of CMR and encourages further research in this direction.", "date": "2026-05-29", "categories": []}
{"id": "2506.05985", "title": "Dynamic Mixture of Progressive Parameter-Efficient Expert Library for Lifelong Robot Learning", "abstract": "A generalist agent must continuously learn and adapt throughout its lifetime, achieving efficient forward transfer while minimizing catastrophic forgetting. Previous work within the dominant pretrain-then-finetune paradigm has explored parameter-efficient fine-tuning for single-task adaptation, effectively steering a frozen pretrained model with a small number of parameters. However, in the context of lifelong learning, these methods rely on the impractical assumption of a test-time task identifier and restrict knowledge sharing among isolated adapters. To address these limitations, we propose Dynamic Mixture of Progressive Parameter-Efficient Expert Library (DMPEL) for lifelong robot learning. DMPEL progressively builds a low-rank expert library and employs a lightweight router to dynamically combine experts into an end-to-end policy, enabling flexible and efficient lifelong forward transfer. Furthermore, by leveraging the modular structure of the fine-tuned parameters, we introduce expert coefficient replay, which guides the router to accurately retrieve frozen experts for previously encountered tasks. This technique mitigates forgetting while being significantly more storage- and computation-efficient than experience replay over the entire policy. Extensive experiments on the lifelong robot learning benchmark LIBERO demonstrate that our framework outperforms state-of-the-art lifelong learning methods in success rates during continual adaptation, while utilizing minimal trainable parameters and storage.", "date": "2025-06-06", "categories": []}
{"id": "2508.08677", "title": "Multi-level Collaborative Distillation Meets Global Workspace Model: A Unified Framework for OCIL", "abstract": "Online Class-Incremental Learning (OCIL) enables models to learn continuously from non-i.i.d. data streams. Since samples of the data streams can be seen only once, it is more suitable for real-world scenarios compared to offline learning. However, this constraint intensifies the challenge for OCIL in maintaining an appropriate balance between stability and plasticity. Moreover, under stricter memory buffer constraints in real world, current replay-based methods are less effective. While ensemble methods improve plasticity, they often struggle with stability. Inspired by the Global Workspace Theory (GWT), we propose a novel approach that enhances ensemble learning through a Global Workspace Model (GWM)-a shared, implicit memory that guides the learning of multiple student models. The GWM is formed by fusing the parameters of all students within each training batch, capturing the historical learning trajectory and serving as a dynamic anchor for knowledge consolidation. Like the broadcasting mechanism of GWT, the GWM is redistributed periodically to students, stabilizing learning and promoting cross-task consistency. In addition, we introduce a multi-level collaborative distillation mechanism. It enforces peer-to-peer consistency among students and preserves historical knowledge by aligning each student with the GWM. As a result, student models remain adaptable to new tasks while maintaining previously learned knowledge, striking a better balance between stability and plasticity. Extensive experiments on three standard OCIL benchmarks show that our method delivers significant performance improvement for several OCIL models across various memory budgets. The code is available at https://github.com/susususushi/GWM.", "date": "2025-08-12", "categories": []}
{"id": "2601.22274", "title": "Server-Proximal Aggregation for Federated Domain-Incremental Learning under Partial Participation: Task-Uniform Convergence and Backward Transfer", "abstract": "Real-world federated systems seldom operate on static data: input distributions drift while privacy rules forbid raw-data sharing. We study this setting as Federated Domain-Incremental Learning (FDIL), where (i) clients are heterogeneous, (ii) tasks arrive sequentially with shifting domains, yet (iii) the label space remains fixed. Two theoretical pillars remain missing for FDIL under realistic deployment: a guarantee of backward knowledge transfer (BKT) and a convergence rate that holds across the sequence of all tasks with partial participation. We introduce SPECIAL (Server-Proximal Efficient Continual Aggregation for Learning), a simple, memory-free FDIL algorithm that adds a single server-side ``anchor'' to vanilla FedAvg: in each round, the server nudges the uniformly sampled participated clients update toward the previous global model with a lightweight proximal term. This anchor curbs cumulative drift without replay buffers, synthetic data, or task-specific heads, keeping communication and model size unchanged. Our theory shows that SPECIAL (i) preserves earlier tasks: a BKT bound caps any increase in prior-task loss by a drift-controlled term that shrinks with more rounds, local epochs, and participating clients; and (ii) learns efficiently across all tasks: the first communication-efficient non-convex convergence rate for FDIL with partial participation, O((E/NT)^(1/2)), with E local epochs, T communication rounds, and N participated clients per round, matching single-task FedAvg while explicitly separating optimization variance from inter-task drift. Experimental results further demonstrate the effectiveness of SPECIAL.", "date": "2026-01-29", "categories": []}
{"id": "2605.27975", "title": "Continual Learning in Modern Hopfield Networks with an Application to Diffusion Models", "abstract": "Generative models, including diffusion models, are increasingly used as foundation models and adapted through sequential fine-tuning, making continual learning an essential problem setting. However, continual learning in such generative models remains poorly understood: after a task change, what aspects of the learned distribution are most easily lost, and what replay samples should be prioritized? We address these questions through the modern Hopfield energy. Recent links between modern Hopfield networks (MHNs) and diffusion models allow analyses in MHNs to be transferred to diffusion models. We introduce intrinsic forgetting as an increase in Hopfield energy after the task change. In tractable settings in an MHN, we prove that high-energy, outlier-like samples undergo a larger energy increase than cluster-like samples, implying that samples located in sharp, isolated basins are more forgettable. We further analyze memory replay and show that replay is particularly effective for high-energy samples, enabling an energy-based selection of replay samples. We validate these predictions in experiments on MHNs and two diffusion models under continual-learning settings: Stable Diffusion and a pixel-space DDPM. In these diffusion models, Hopfield energy tracks reconstruction-based forgetting, and replay experiments reveal energy-dependent mitigation of forgetting that is consistent with the MHN analysis.", "date": "2026-05-27", "categories": []}
{"id": "2605.29495", "title": "On-Policy Replay for Continual Supervised Fine-Tuning", "abstract": "Continual supervised fine-tuning (SFT) is the de facto recipe for adapting large language models (LLMs) to a stream of downstream tasks, but it suffers from catastrophic forgetting of earlier capabilities. Recent work shows that on-policy signals -- training on the model's own outputs -- reduce forgetting more reliably than off-policy supervision. Existing on-policy methods route this signal through a new training objective (e.g., self-distillation losses with a teacher copy), inheriting an extra forward pass, schedule sensitivity, and stylistic drift from the teacher.We instead route the on-policy signal through the training data source. Our method, On-Policy Replay (OPR), rolls out the most recent checkpoint on a small budget of historical prompts, filters the generations by a task reward, and replays the surviving (prompt, model response) pairs as ordinary SFT examples. There is no teacher, no auxiliary loss, and no on-the-fly distillation. Across three 7--8B instruction-tuned backbones (Qwen2.5-7B-Instruct, Qwen3-8B, Llama3.1-8B-Instruct) on the TRACE continual-learning benchmark, OPR consistently reduces forgetting; on the sharpest stress test (Qwen2.5-7B-Instruct, Sequential SFT BWT -13.93), OPR lifts BWT to -0.65 at a 10% replay budget and to -2.29 at a 1% budget -- a 46% reduction in |BWT| over a tuned Vanilla Replay baseline, with 42--46% reductions observed across all three backbones. We give a KL-shrinkage interpretation that places OPR and prior on-policy distillation methods on a single axis, and we present a counterintuitive finding that explains why Vanilla Replay is already a strong baseline: low-score replay is uniformly worse than Vanilla Replay, demonstrating that the active ingredient in OPR is the on-policy distribution, not the response quality alone.Our code is available at https://github.com/Yancey2024/OnPolicyReplay.", "date": "2026-05-28", "categories": []}
{"id": "2605.29592", "title": "Non-Forgetting Knowledge Allocation with Bi-level Competition for Class-Incremental Learning", "abstract": "Class-Incremental Learning (CIL) with pre-trained models (PTMs) aims to sequentially adapt PTMs to new categories without forgetting old knowledge. Built upon PTMs, existing adapter-based methods mainly train models via distinct task-specific adapters, and present a uniform knowledge allocation for each adapter during inference. However, this allocation mechanism ignores the nature of task discrepancy and leads to suboptimal utilization of adapters. Also, under CIL constraint, an allocator is prone to forgetting when tasks evolve. To address these issues, we propose a Non-Forgetting Allocation with Bi-Level Competition (NoFA-BC). NoFA-BC constructs a non-forgetting allocator (NFA) by transforming the allocator training into a recursive least-squares problem and achieves an allocator equivalent to that trained with all data. Based on the NFA, a Bi-Level Competition (BLC) including an intra-task level Winner-Takes-All (WTA) mechanism and inter-task Last-Ones-Fall (LOF) elimination is proposed to provide better allocation of adapter knowledge. WTA extracts the most significant logit within a task to represent the adapter's contribution and LOF suppresses the irrelevant adapters. With BLC, participation ratio of each adapter can be tailored for each input. Moreover, a Stability Enhancement (SE) process is incorporated to further improve the performance of old tasks.", "date": "2026-05-28", "categories": []}
{"id": "2605.30198", "title": "Active Continual Learning with Metaplastic Binary Bayesian Neural Networks", "abstract": "Always-on edge systems must keep learning as conditions change under tight compute budgets and must detect unreliable predictions. Bayesian binary neural networks are attractive in this setting, but mean-field Bernoulli posteriors can saturate on long non-stationary streams, wiping out epistemic uncertainty and freezing plasticity. We propose BiMU, derived from a bounded-memory variational objective that balances stability, plasticity, and forgetting. BiMU combines a data term with controlled relaxation toward the prior and an uncertainty-dependent step size that prevents saturation and sustains informative uncertainty. This non-degenerate posterior enables fully online, buffer-free active querying via Monte Carlo disagreement, reducing label queries and backpropagation updates under imbalance. BiMU sustains learning and strong OOD detection on 1000-tasks Permuted-MNIST, and on OpenLORIS-Object achieves up to 32$\\times$ label/update savings at matched accuracy under class imbalance and feature compression.", "date": "2026-05-28", "categories": []}
{"id": "2412.01004", "title": "Take Only What You Need: Rank Minimization as an Implicit Forgetting Regularizer in Continual Learning", "abstract": "The central tension in continual learning (CL) is the trade-off between plasticity (acquiring new knowledge) and stability (retaining prior knowledge). We study how a pre-trained backbone can be continually updated to absorb new knowledge while preserving existing capabilities, via capacity control: regulating the effective rank of each parameter update, a per-step quantity directly controllable inside a LoRA update. A controlled probe of LoRA rank and placement across modules and tasks reveals a consistent trade-off, with a moderate-rank sweet spot that varies by placement and task, leaving no universally optimal fixed rank; a formal bound shows forgetting grows with rank. Building on these findings, we propose Continual Dynamic Rank-Selective LoRA (CoDyRA), which jointly trains each LoRA update with rank minimization via sparsity-promoting regularization on per-component importance weights. The supervised objective drives plasticity; rank minimization regularizes forgetting. We show that rank minimization serves as an implicit forgetting regularizer in the CL regime, protecting general capability and prior-task knowledge simultaneously by controlling forgetting against the current model state. Across MTIL, X-TAIL, and TRACE (CLIP, LLaMA, Gemma), CoDyRA outperforms prior CL methods on new knowledge learning and forgetting, achieving a strong plasticity-stability balance. Code is available at https://github.com/jeff024/codyra.", "date": "2024-12-01", "categories": []}
{"id": "2602.01990", "title": "SAME: Stabilized Mixture-of-Experts for Multimodal Continual Instruction Tuning", "abstract": "Multimodal Large Language Models (MLLMs) achieve strong performance through instruction tuning, but real-world deployment requires them to continually expand their capabilities, making Multimodal Continual Instruction Tuning (MCIT) essential. Recent methods leverage sparse expert routing to promote task specialization, but we find that the expert routing process suffers from drift as the data distribution evolves. For example, a grounding query that previously activated localization experts may instead be routed to irrelevant experts after learning OCR tasks. Meanwhile, the grounding-related experts can be overwritten by new tasks and lose their original functionality. Such failure reflects two problems: router drift, where expert selection becomes inconsistent over time, and expert drift, where shared experts are overwritten across tasks. Therefore, we propose StAbilized Mixture-of-Experts (SAME) for MCIT. To address router drift, SAME stabilizes expert selection by decomposing routing dynamics into orthogonal subspaces and updating only task-relevant directions. To mitigate expert drift, we regulate expert updates via curvature-aware scaling using historical input covariance in a rehearsal-free manner. SAME also introduces adaptive expert activation to freeze selected experts during training, reducing redundant computation and cross-task interference. We also introduce a new benchmark to evaluate MCIT with long task sequence, and extensive experiments demonstrate SAME's SOTA performance. Code is available at https://github.com/LAMDA-CL/Prism.", "date": "2026-02-02", "categories": []}
{"id": "2602.02417", "title": "Trust Region Continual Learning as an Implicit Meta-Learner", "abstract": "Continual learning aims to acquire tasks sequentially without catastrophic forgetting, yet standard strategies face a core tradeoff: regularization-based methods (e.g., EWC) can overconstrain updates when task optima are weakly overlapping, while replay-based methods can retain performance but drift due to imperfect replay. We study a hybrid perspective: \\emph{trust region continual learning} that combines generative replay with a Fisher-metric trust region constraint. We show that, under local approximations, the resulting update admits a MAML-style interpretation with a single implicit inner step: replay supplies an old-task gradient signal (query-like), while the Fisher-weighted penalty provides an efficient offline curvature shaping (support-like). This yields an emergent meta-learning property in continual learning: the model becomes an initialization that rapidly \\emph{re-converges} to prior task optima after each task transition, without explicitly optimizing a bilevel objective. Empirically, on task-incremental diffusion image generation and continual diffusion-policy control, trust region continual learning achieves the best final performance and retention, and consistently recovers early-task performance faster than EWC, replay, and continual meta-learning baselines.", "date": "2026-02-02", "categories": []}
{"id": "2605.26357", "title": "Balancing Plasticity and Stability with Fast and Slow Successor Features", "abstract": "A hallmark of intelligence is the ability to adapt in non-stationary environments, yet deep Reinforcement Learning (RL) agents often struggle in such settings. Prior studies introduce non-stationarity through abrupt shifts in features or dynamics, whereas real-world environments often evolve gradually through continual drift. This distinction has important implications for the \"stability-plasticity dilemma\" in RL, as abrupt task changes may demand more plasticity than naturalistic settings. To address this, we modify existing 3D Miniworld and MuJoCo environments to incorporate naturalistic, continual non-stationarity, and use them to examine how stability and adaptation affect performance under continuous environmental change. We find that methods favoring stability, such as synaptic consolidation, outperform approaches focused on plasticity, such as parameters resetting. Motivated by this result, and prior evidence that Successor Features (SFs) reduce interference, we investigate whether SFs are better consolidation targets than Q-values. Across both environments, applying neuro-inspired synaptic consolidation to SFs yields superior performance on continually changing settings. Moreover, consolidation is most effective when SFs are stabilized across multiple timescales, which capture complementary aspects of gradual environmental change. Together, these results suggest that stability is more critical in continual learning when changes are gradual, and that multi-timescale consolidation of predictive representations is an effective approach.", "date": "2026-05-25", "categories": []}
{"id": "2605.27469", "title": "Architecture-driven Shift: towards a lightweight selector for capturing the trends of logit shift", "abstract": "Continual Learning (CL) is a practical paradigm to utilize power of deep pre-trained neural networks, but which pre-trained model has a better ability to balance ``Plasticity-Stability\", deserving to be chosen? The logit shift serves as a natural proxy because it represents the logit shift in CL scenarios. However, obtaining the logit shift requires huge computational cost, which hinders large-scale model selection. Existing theoretical analyses fail to offer an efficient alternative because of the assumption of uniform hidden layer widths, which ignores the structural heterogeneity (variable width and depth) of real-world architectures. This raises a critical question: what theoretically relationship can be identified between heterogeneous architecture and logit shift on prior tasks (that the model has been trained on)? To answer the question, we decouple logit shift into architecture dependency and data dependency to establish our framework, which reveals that the combination of two dependency, defined as Architecture-driven Shift (ADS), that can capture the logit shift tendency well computable with few data samples. Specifically, for a well-optimized model on prior tasks, higher ADS is associated with a larger logit shift after training on the current task, which derived based on three mechanistic components: (1) spectral norm scaling of weight matrix gradients with layer width, (2) the optimization path length of the new task, and (3) the asymptotic task conflict in wide networks. Extensive empirical results across more than 175 diverse architectures demonstrate a strong monotonic correlation (the weakest Spearman's $r_s=0.731$) between ADS and logit shift. Practically, we demonstrate that ADS can serve as a lightweight proxy of the expected calibration error, which is a widely used metric for reliable CL model selection, on three datasets across six scenarios.", "date": "2026-05-26", "categories": []}
{"id": "2605.27482", "title": "Energy-Structured Low-Rank Adaptation for Continual Learning", "abstract": "While orthogonal subspace methods try to mitigate task interference in Continual Learning (CL), they often suffer from energy diffusion across the basis, hindering knowledge compaction and exhausting capacity for future tasks. We observe that output feature drift induced by parameter updates is inherently low-rank, and theoretically prove that preserving parameters along the principal directions of this drift minimizes the output reconstruction error. Motivated by this, we propose \\textbf{E}nergy-Concentrated and \\textbf{E}nergy-Ordered \\textbf{Lo}w-\\textbf{R}ank \\textbf{A}daptation (E$^2$-LoRA). By explicitly ordering and concentrating knowledge into leading ranks, E$^2$-LoRA frees capacity for subsequent tasks. Furthermore, we design a dynamic rank allocation strategy to balance stability and plasticity by jointly optimizing energy retention and model plasticity. Extensive experiments across multiple benchmarks demonstrate that E$^2$-LoRA achieves state-of-the-art performance.", "date": "2026-05-26", "categories": []}
{"id": "2605.28387", "title": "CLANE: Continual Learning of Actions on Neuromorphic Hardware from Event Cameras", "abstract": "Recognizing and continuously learning novel human actions without forgetting prior classes is a requirement for emerging AR/VR and robotics applications. For these applications, both on-device processing and learning are essential for privacy and low-latency adaptation. Event cameras address the efficiency of visual sensing with sparse, asynchronous output that is naturally compatible with neuromorphic processing. Yet no prior system has deployed a continual on-device learning pipeline for event-based action recognition using neuromorphic hardware. We present CLANE, Continual Learning of Actions on Neuromorphic Hardware from Event Cameras, deployed end-to-end on Intel Loihi 2. CLANE combines a spiking 2D CNN for spatiotemporal feature extraction with CLP-SNN as its on-chip learning head, extended to action clips via a Temporal Aggregation Layer and a fixed-point Normalization Layer, both novel Loihi 2 modules. On THU E-ACT-50, a 50-class dataset captured under real-world conditions, CLANE achieves 70.4% accuracy in a continual learning task while delivering more than 100x energy reduction and 16x lower latency over a sequential CNN+GRU+CLP edge GPU baseline, validated through iso-algorithm cross-platform benchmarking across three evaluation levels.", "date": "2026-05-27", "categories": []}
{"id": "2605.28495", "title": "Janus-LoRA: A Balanced Low-Rank Adaptation for Continual Learning", "abstract": "Low-Rank Adaptation (LoRA) has emerged as a promising paradigm for Continual Learning. It independently updates its low-rank factors ($A$ and $B$), creating a composite update to the full weight matrix through their interaction. To prevent catastrophic forgetting, this update should remain orthogonal to the task-specific subspace that contains previously learned knowledge. However, we identify that this composite update systematically violates this orthogonality, reintroducing interference and undermining stability. Furthermore, naively enforcing this orthogonality compromises plasticity, disrupting the delicate stability-plasticity trade-off. To resolve these issues, we propose \\textbf{Janus-LoRA}, a framework that restores this balance through two novel components. Specifically, we first introduce Gradient Rectification, a closed-form solution that mathematically decouples LoRA's factor updates, enforcing orthogonality against the historical knowledge subspace identified by an efficient Online Estimation. Next, to enhance plasticity, we introduce a Decoupled Margin Loss that promotes feature-level separation by pushing new feature representations away from old ones, thus creating distinct, low-interference regions for new learning. Comprehensive experiments on challenging benchmarks demonstrate that by harmonizing parameter-level orthogonality with feature-level separation, Janus-LoRA achieves a superior balance and establishes new state-of-the-art performance.", "date": "2026-05-27", "categories": []}
{"id": "2605.28577", "title": "Continual Model Routing in Evolving Model Hubs", "abstract": "AI model hubs provide access to a rapidly growing collection of powerful pre-trained models, enabling off-the-shelf mixture-of-experts systems with different routing strategies. However, this rapid growth poses two fundamental challenges: scaling model selection across thousands of experts and continually updating routing mechanisms as new models and tasks are introduced. In this paper, we formalise this setting as Continual Model Routing (CMR) and propose CMRBench, a new large-scale benchmark simulating realistic hub expansion and including over 2,000 candidate models. Finally, we introduce CARvE, a contrastive embedding approach for efficient continual model routing via checkpoint-based anchoring and structured replay. Extensive empirical results and ablations show that CARvE significantly outperforms zero-shot retrieval, fine-tuning, and adapter-merging baselines in model, family, and domain-level accuracy.", "date": "2026-05-27", "categories": []}
{"id": "2605.28705", "title": "Understanding Generalization and Forgetting in In-Context Continual Learning", "abstract": "In-context learning (ICL) derives its power from enabling Large Language Models to adapt to new tasks via prompt-based reasoning alone, entirely bypassing the need for parameter updates. Existing theories primarily study ICL in single-task settings, while real-world prompts often contain sequences of heterogeneous tasks, leaving a gap in understanding whether Large Language Models implicitly perform continual learning during inference. To bridge this gap, we propose the first theoretical framework for in-context continual learning, modeling how a pretrained Transformer processes multiple sequential tasks within a single prompt through shared attention mechanisms. Focusing on linear and masked linear self-attention, we derive error expressions for model predictions under sequential task prompts and analyze their generalization and forgetting behavior. Our results reveal that standard attention mechanisms inevitably induce intertask interference by uniformly or causally aggregating historical contexts, leading to systematic bias. We further provide a bias-variance-interference decomposition of prediction error, characterizing when historical in-context information yields positive transfer or provable negative transfer. This analysis exposes fundamental limits of attention-based continual inference and offers theoretical explanations for order sensitivity and performance degradation in long prompts.", "date": "2026-05-27", "categories": []}
{"id": "2605.28809", "title": "AREA: Attribute Extraction and Aggregation for CLIP-Based Class-Incremental Learning", "abstract": "Class-Incremental Learning (CIL) is important in building real-world learning systems. In CLIP-based CIL, the model performs classification by comparing similarity between visual and textual embeddings obtained from template prompts, e.g., ``a photo of a [CLASS]''. This seemingly monolithic matching process can be decomposed into two conceptually distinct stages: attribute extraction and attribute aggregation. For example, a model may recognize cat using attributes such as fur texture and whiskers. When learning a new class like car, the model must extract additional attributes like wheels and adjust how they are aggregated in the shared representation space. However, since only data from the current task is available, incremental updates can bias both attribute extraction and aggregation toward new classes, leading to catastrophic forgetting. Therefore, we propose AREA for attribute extraction and aggregation in CLIP-based CIL. To stabilize extraction, we anchor class-level visual and textual attributes on the hyperspherical embedding space via principal geodesic analysis. To stabilize aggregation, we learn lightweight task-specific experts with scoring and residual refinement, regularized by a variational information bottleneck objective. During inference, we perform routing over task attribute manifolds via optimal transport for more concise prediction. Experiments show that AREA consistently outperforms SOTA methods. Code is available at https://github.com/LAMDA-CL/ICML2026-AREA.", "date": "2026-05-27", "categories": []}
{"id": "2605.27116", "title": "COVD: Continual Open-Vocabulary Object Detection with Novel Concept Injection", "abstract": "Open-vocabulary object detection (OVD) has made significant progress, enabling detectors to generalize from seen to unseen categories. However, real-world category spaces continually evolve, and existing OVD models still struggle with newly emerging concepts, while repeated full retraining is prohibitively expensive. To this end, we introduce a new task setting, termed Continual OVD with Novel Concept Injection (COVD), where models sequentially learn incoming novel concept groups while preserving prior concepts and original open-vocabulary knowledge, along with a new benchmark, Novel-114. Our key observation is that pretrained visual encoders often already perceive and represent many novel concepts, and the main bottleneck lies in the lack of stable semantic alignment between visual representations and textual concepts. Based on this, we propose NoIn-Det, an efficient continual injection framework without additional parameters. NoIn-Det freezes the visual encoder, preserves the text representation space using only texts of common concepts and previously injected concepts, and injects novel concepts by updating only a small subset of text-branch parameters beneficial to novel concept learning. Extensive experiments show that NoIn-Det effectively learns novel concepts, preserves old knowledge, and consistently outperforms existing continual learning methods for VLMs without introducing additional parameters.Novel-114 and the code will be released.", "date": "2026-05-26", "categories": []}
{"id": "2605.27128", "title": "PILOT: A Data-Free Continual Learning Approach for Real-Time Semantic Segmentation via Boundary Guidance", "abstract": "Real-time semantic segmentation models offer an excellent balance between accuracy and inference speed. However, deploying these models in dynamic real world environments often requires the ability to learn novel classes incrementally without retraining on the entire dataset. This capability is known as continual learning. In this regard, the standard fine-tuning methods in deep learning often fail due to catastrophic forgetting, where the model learns new information but forgets previously trained and learned classes. Contributing to this crucial domain, the current paper proposes a novel continual learning framework tailored for PIDNet, which is a widely cited state-of-the-art real-time semantic segmentation model. Our method, PILOT(Parallel Incremental Learning Over Time), introduces a real-time and lightweight strategy by implementing a parallel Derivative-branch (D-branch) designed to capture the high frequency boundary information of novel classes while freezing the trained parameters of the original segmentation network. This novel setup allows the model to adapt to new semantic categories while preserving the knowledge of previously learned classes. By using only data associated with the new class, our model significantly reduces training overhead. Experimental results demonstrate that our approach successfully segments new classes while maintaining high mean Intersection over Union (mIoU) on the original base classes, thereby comfortably outperforming all major continual learning approaches in this domain. Overall, PILOT is shown to effectively mitigate catastrophic forgetting with minimal impact on inference latency, thus maintaining real-time performance.", "date": "2026-05-26", "categories": []}
{"id": "2601.15544", "title": "RDumb++: Drift-Aware Continual Test-Time Adaptation", "abstract": "Continual Test-Time Adaptation (CTTA) seeks to update a pretrained model during deployment using only the incoming, unlabeled data stream. Although prior approaches such as Tent, EATA etc. provide meaningful improvements under short evolving shifts, they struggle when the test distribution changes rapidly or over extremely long horizons. This challenge is exemplified by the CCC benchmark, where models operate over streams of 7.5M samples with continually changing corruption types and severities. We propose RDumb++, a principled extension of RDumb that introduces two drift-detection mechanisms i.e entropy-based drift scoring and KL-divergence drift scoring, together with adaptive reset strategies. These mechanisms allow the model to detect when accumulated adaptation becomes harmful and to recover before prediction collapse occurs. Across CCC-medium with three speeds and three seeds (nine runs, each containing one million samples), RDumb++ consistently surpasses RDumb, yielding approx 3% absolute accuracy gains while maintaining stable adaptation throughout the entire stream. Ablation experiments on drift thresholds and reset strengths further show that drift-aware resetting is essential for preventing collapse and achieving reliable long-horizon CTTA.", "date": "2026-01-22", "categories": []}
{"id": "2602.15811", "title": "CARL-CXR: Continual Adapter-Based Routing for Task-Unknown Chest Radiograph Classification", "abstract": "Clinical deployment of chest radiograph classifiers requires models that can be updated as new datasets become available without retraining on previously observed data or degrading validated performance. We study a task-incremental continual learning setting for chest radiograph classification under task-unknown inference, where heterogeneous chest X-ray datasets arrive sequentially and task identity is unavailable at deployment time. We propose CARL-CXR, a continual adapter-based routing framework that maintains a fixed high-capacity backbone while incrementally introducing lightweight task-specific adapters and classifier heads. A latent task selector operates on adapter-conditioned features to dynamically route each input to the most relevant task pathway, leveraging compact task prototypes and feature-level experience replay to preserve task identity across sequential updates without storing raw images. Experiments on MIMIC-CXR and CheXpert two large-scale datasets with distinct patient populations, imaging devices, and annotation pipelines demonstrate that CARL-CXR achieves minimal catastrophic forgetting (0.012 AUROC drop), representing a 6X and 11X reduction over established continual learning baselines LwF and EWC respectively, while maintaining competitive diagnostic performance (AUROC 0.74). Under task unknown deployment, CARL-CXR outperforms joint training by 12.5 points in routing accuracy (75.0% vs. 62.5%): unlike LwF and EWC, which require explicit task identifiers at inference and provide no routing mechanism.", "date": "2026-02-17", "categories": []}
{"id": "2603.00191", "title": "Task-Driven Subspace Decomposition for Knowledge Sharing and Isolation in LoRA-based Continual Learning", "abstract": "Continual Learning (CL) requires models to sequentially adapt to new tasks without forgetting old knowledge. Recently, Low-Rank Adaptation (LoRA), a representative Parameter-Efficient Fine-Tuning (PEFT) method, has gained increasing attention in CL. Several LoRA-based CL methods reduce interference across tasks by separating their update spaces, typically building the new space from the estimated null space of past tasks. However, they (i) overlook task-shared directions, which suppresses knowledge transfer, and (ii) fail to capture truly effective task-specific directions since these ``null bases\" of old tasks can remain nearly inactive for new task under correlated tasks. To address this, we study LoRA learning capability from a projection energy perspective, and propose Low-rank Decomposition and Adaptation (LoDA). It performs a task-driven decomposition to build general and truly task-specific LoRA subspaces by solving two energy-based objectives, decoupling directions for knowledge sharing and isolation. LoDA fixes LoRA down-projections on two subspaces and learns robust up-projections via a Gradient-Aligned Optimization (GAO) approach. After each task, before integrating the LoRA updates into the backbone, LoDA derives a closed-form recalibration for the general update, approximating a feature-level joint optimum along this task-shared direction. Experiments indicate that LoDA outperforms existing CL methods. Our code is available at https://github.com/HHHLF/LoDA_ICML2026.", "date": "2026-02-27", "categories": []}
{"id": "2605.19409", "title": "Unlocking the Potential of Continual Model Merging: An ODE Perspective", "abstract": "Continual Model Merging (CMM) enables rapid customization of foundation models by sequentially incorporating task-adapted models without repeated retraining. However, existing merging rules usually update the deployed model through fixed algebraic or projection-based operations, providing limited control over how much previously accumulated knowledge should be retained relative to the incoming task model. This limitation leads to unstable retention and performance degradation in long task streams, and becomes more pronounced when tasks have heterogeneous utilities. We propose ODE-driven Merging (ODE-M), a controllable framework that formulates each continual merge as a trajectory in parameter space rather than a one-step endpoint update. Motivated by mode connectivity, ODE-M constructs a barrier-aware trajectory using a rectified time-dependent velocity field, where lightweight first-order feedback from a small calibration set suppresses loss-increasing motion while preserving progress toward the incoming model. The next merged model is then obtained by selecting an operating point along this trajectory through a utility-aware time schedule, providing an explicit mechanism for balancing retained historical knowledge and incoming task expertise. Extensive experiments on standard CMM benchmarks show that ODE-M consistently improves over strong continual merging baselines across CLIP ViT backbones, stream lengths, and heterogeneous task-utility settings.", "date": "2026-05-19", "categories": []}
{"id": "2605.25170", "title": "Grow-Prune-Freeze Networks: Adaptive & Continual Learning Technique for Olfactory Navigation", "abstract": "Training data for olfaction is scattered through disparate, non-standardized datasets that limit the ability to build representative world models. Olfactory navigation is a highly dynamic and non-stationary task that benefits from real-time continual learning. We introduce an adaptive framework called Grow-Prune-Freeze (GPF) networks that enable an agent to continually learn through growing, pruning, and freezing early layers of its policy in response to world complexity. Grounding GPFs in non-linear random matrix theory, we show that the work of Pennington & Worth (2017) can be extended from single hidden layers to n-layer continual-learning models, and that eigenvalue composition of network weights is preserved as successive layers are added. We show that GPFs based on Expected SARSA achieve a 94% success rate on turbulent plume navigation - a partially observable, non-stationary task representative of the \"big world\" challenges that motivate adaptive learning in robotics - and provide supporting methodology for applying GPFs in other world models. Further experiments amount evidence that GPFs may generalize well to other machine learning tasks such as reinforcement learning in Atari, image classification, and autoregressive language models. We open source all code and data to encourage improvements on and more research in olfactory robotics.", "date": "2026-05-24", "categories": []}
{"id": "2605.25525", "title": "SAE-FD: Sparse Autoencoder Feature Distillation for Continual Learning of Large Language Models", "abstract": "Continual learning enables large language models to adapt to evolving tasks without retraining from scratch, yet catastrophic forgetting remains a central obstacle. Among continual learning methods, regularization-based approaches are widely used to constrain model updates and reduce forgetting, operating in weight space, gradient space, or output space. However, these dense representation spaces suffer from feature superposition, where multiple concepts are encoded in overlapping dimensions, making it difficult to selectively protect previously learned knowledge without impeding new-task learning. To address this issue, we propose \\method (Sparse Autoencoder Feature Distillation), which anchors model representations in the sparse feature space of a pre-trained Sparse Autoencoder, where dense activations are decomposed into a sparse overcomplete basis that reduces representational entanglement, enabling more targeted regularization with less interference to new-task learning. Experiments on two continual learning benchmarks across three model architectures show that \\method consistently outperforms existing regularization-based methods, achieving up to 52.70% average accuracy with only -0.46 backward transfer.", "date": "2026-05-25", "categories": []}
{"id": "2605.25708", "title": "CMAP: Cross-Modal Adaptive Prompting for Multi-Domain Task-Incremental Learning", "abstract": "Multi-domain task-incremental learning requires a model to sequentially acquire knowledge across visually diverse domains without forgetting prior tasks, and without access to task identity at inference. Parameter-efficient methods built on frozen vision-language models have made strong progress, yet all existing approaches rely exclusively on visual features for task routing, confidence estimation, and encoder adaptation, leaving CLIP's cross-modal text embedding space entirely unexploited. We address this gap through three contributions. Text-space task routing replaces visual Gaussian matching with cosine similarity to frozen CLIP text prototypes, giving order-independent routing robust to data scarcity at zero parameter cost. Multi-prototype visual-textual confidence replaces single-Gaussian class modeling with K-means visual prototypes and cross-modal alignment scores under task-calibrated thresholds. Symmetric cross-modal gating extends per-layer Gumbel gates to the text encoder conditioned on batch image features, preserving cross-modal alignment on out-of-distribution inputs. On the MTIL benchmark spanning 11 datasets and 1201 classes, our method achieves 74.2% Transfer, 80.5% Average, and 88.7% Last under Order-I, surpassing the prior state of the art by 5.0, 3.7, and 3.0 percentage points with only 2.5M trainable parameters and no external data.", "date": "2026-05-25", "categories": []}
{"id": "2605.26097", "title": "Forgetting in Language Models: Capacity, Optimization, and Self-Generated Replay", "abstract": "Models trained on a new task typically degrade on prior tasks, a phenomenon known as forgetting. Traditionally, mitigating forgetting has required replaying stored exemplars from prior tasks, which is often impractical. By contrast, language models can sample from their own training distribution, and we show that these self-generated samples serve as effective replay data, nearly eliminating forgetting. We find that forgetting nonetheless persists when the model has little remaining capacity: models pretrained close to saturation cannot absorb new information without overwriting prior knowledge. When capacity is not the limiting factor, low learning rates reduce forgetting but require substantially more training steps. Replay breaks this tradeoff, enabling fast, high-learning-rate finetuning without forgetting.", "date": "2026-05-25", "categories": []}
{"id": "2605.26110", "title": "Prism: A Plug-in Reproducible Infrastructure for Scalable Multimodal Continual Instruction Tuning", "abstract": "Multimodal Large Language Models (MLLMs) achieve versatility by reformulating diverse tasks into a unified instruction-following framework via instruction tuning. However, real-world deployment requires continuous adaptation to emerging tasks, motivating Multimodal Continual Instruction Tuning (MCIT). Despite its growing importance, current MCIT research is hindered by severe engineering bottlenecks. Existing methods are typically implemented by directly modifying the base MLLM codebase, which imposes substantial implementation overhead and yields method-specific architectures that severely limit code reuse and fair comparison. To address this, we introduce Prism, a plug-in reproducible codebase specifically designed for scalable MCIT research. It separates algorithmic development from the backbone implementation via a lightweight plugin registration mechanism, enabling new strategies to be integrated as independent plugins without modifying the underlying MLLM codebase, thereby eliminating structural fragmentation and accelerating method development. Prism natively supports widely used large-scale training pipeline, thereby enabling reproducible and scalable MCIT experimentation. Code is available at https://github.com/LAMDA-CL/Prism.", "date": "2026-05-25", "categories": []}
{"id": "2509.12958", "title": "Forget What's Sensitive, Remember What Matters: Token-Level Differential Privacy in Memory Sculpting for Continual Learning", "abstract": "Continual Learning (CL) models, while adept at sequential knowledge acquisition, face significant and often overlooked privacy challenges due to accumulating diverse information. Traditional privacy methods, like a uniform Differential Privacy (DP) budget, indiscriminately protect all data, leading to substantial model utility degradation and hindering CL deployment in privacy-sensitive areas. To overcome this, we propose a privacy-enhanced continual learning (PeCL) framework that forgets what's sensitive and remembers what matters. Our approach first introduces a token-level dynamic Differential Privacy strategy that adaptively allocates privacy budgets based on the semantic sensitivity of individual tokens. This ensures robust protection for private entities while minimizing noise injection for non-sensitive, general knowledge. Second, we integrate a privacy-guided memory sculpting module. This module leverages the sensitivity analysis from our dynamic DP mechanism to intelligently forget sensitive information from the model's memory and parameters, while explicitly preserving the task-invariant historical knowledge crucial for mitigating catastrophic forgetting. Extensive experiments show that PeCL achieves a superior balance between privacy preserving and model utility, outperforming baseline models by maintaining high accuracy on previous tasks while ensuring robust privacy.", "date": "2025-09-16", "categories": []}
{"id": "2411.02813", "title": "Sparse Orthogonal Parameters Tuning for Continual Learning", "abstract": "Continual learning methods based on pre-trained models (PTM) have recently gained attention which adapt to successive downstream tasks without catastrophic forgetting. These methods typically refrain from updating the pre-trained parameters and instead employ additional adapters, prompts, and classifiers. In this paper, we from a novel perspective investigate the benefit of sparse orthogonal parameters for continual learning. We found that merging sparse orthogonality of models learned from multiple streaming tasks has great potential in addressing catastrophic forgetting. Leveraging this insight, we propose a novel yet effective method called SoTU (Sparse Orthogonal Parameters TUning). We hypothesize that the effectiveness of SoTU lies in the transformation of knowledge learned from multiple domains into the fusion of orthogonal delta parameters. Experimental evaluations on diverse CL benchmarks demonstrate the effectiveness of the proposed approach. Notably, SoTU achieves optimal feature representation for streaming data without necessitating complex classifier designs, making it a Plug-and-Play solution.", "date": "2024-11-05", "categories": []}
{"id": "2605.21951", "title": "Dynamic Mixture of Latent Memories for Self-Evolving Agents", "abstract": "Achieving self-evolution in intelligent agents requires the continual accumulation of new knowledge across changing task sequences without forgetting previously acquired abilities. Existing approaches either internalize knowledge by updating model parameters, which induces catastrophic forgetting, or rely on external memory, which fails to genuinely enhance the model's intrinsic capabilities. We propose MoLEM, a generative mixture of latent memory framework based on a dynamic mixture-of-experts (MoE). We treat multiple experts as independent carriers to generate memory. A router selects and weights experts through key-query matching, and the aggregated latent memory is injected into the reasoning process. The base model for reasoning remains entirely frozen, with all experiential knowledge internalized into the additional modules, avoiding catastrophic forgetting. For continual learning, each training stage is paired with a lightweight autoencoder that selects the appropriate routing group at inference, and inputs that match no stage fall back to the pretrained model. Experiments train the framework on continual-learning sequences spanning math, science, and code domains. After training, we evaluate the framework on the corresponding test sets to measure task learning and competence preservation across continual adaptation stages. After the full continual-learning sequence, our method improves the average accuracy by 10.40% over the Vanilla pretrained baseline, while none of the competing methods consistently exceed this baseline across different training orders.", "date": "2026-05-21", "categories": []}
{"id": "2605.22035", "title": "HyLoVQA: Dynamic Hypernetwork-Generated Low-Rank Adaptation for Continual Visual Question Answering", "abstract": "Continual Visual Question Answering (VQA) requires learning from non-stationary streams of visual inputs and questions while preserving past knowledge. Most prior methods adapt by updating a largely shared parameter set. This often leads to cross-level task interference, hindering accurate adaptation to the current task and object. To address this limitation, we propose HyLoVQA. It maintains a drift-resilient memory bank of anchors. The bank stores the content of visual objects and textual tasks, and they are updated using current input features. Conditioned on retrieved anchors, a hypernetwork generates lightweight Low-Rank Adaptation (LoRA) adapters. This ensures parameter efficiency, allowing the model to adapt to each task and object dynamically. Additionally, we formulate an alignment loss that aligns semantic discrepancies in the feature space with functional changes in the parameter space, thereby constraining LoRA adapters to remain focused on the current task and object. Extensive experiments on VQA v2 and NExT-QA under both standard and compositional settings demonstrate the superiority of HyLoVQA over prior state-of-the-art methods.", "date": "2026-05-21", "categories": []}
{"id": "2605.22454", "title": "Don't Forget the Critic: Value-Based Data Rehearsal for Multi-Cyclic Continual Reinforcement Learning", "abstract": "Data rehearsal has emerged as a leading approach for mitigating catastrophic forgetting in Continual Reinforcement Learning (CRL). However, existing work remains confined to policy gradient frameworks, regularizing only actors due to the performance degradation incurred by critic regularization. This actor-centric approach overlooks the potential of data rehearsal for value function approximation. Moreover, existing evaluations in CRL rarely consider multi-cyclic environments where task sequences repeat, a critical real-world scenario that exacerbates forgetting and plasticity. We investigate data rehearsal for Deep Q-Networks using Q-value regularization in multi-cyclic settings and propose Qreg+NWLU which introduces two simple modifications: (1) continuous data rehearsal that dynamically collects and updates stored Q-values throughout training, and (2) \"No-Wait\" regularization that applies immediately rather than after the first task. Together, these modifications yield improvements in learning efficiency, forgetting mitigation, and knowledge transfer over Qreg and conventional CRL methods within value function approximation settings.", "date": "2026-05-21", "categories": []}
{"id": "2605.22743", "title": "SeqLoRA: Bilevel Orthogonal Adaptation for Continual Multi-Concept Generation", "abstract": "Parameter-efficient fine-tuning enables fast personalization of text-to-image diffusion models, but composing multiple custom concepts remains challenging due to representation interference. Existing modular methods either rely on expensive post-hoc fusion or freeze adaptation subspaces, which limit expressiveness and concept fidelity. To address this trade-off, we propose Sequential regularized LoRA (SeqLoRA), a constrained continual learning framework that jointly optimizes both LoRA factors via bilevel optimization. Theoretically, we establish strong convergence guarantees for our algorithm and model the residual layer activations as a matrix sub-Gaussian process to derive high-probability bounds on catastrophic forgetting. We further prove that learning the LoRA basis from data minimizes residual interference energy more effectively than frozen-basis methods. Experiments on multi-concept image generation demonstrate that SeqLoRA improves identity preservation and scalability across up to 101 concepts, while avoiding costly fusion and reducing attribute interference in composed generations.", "date": "2026-05-21", "categories": []}
{"id": "2605.14364", "title": "MoRe: Modular Representations for Principled Continual Representation Learning on Sequential Data", "abstract": "Continual learning requires models to adapt to new data while preserving previously acquired knowledge. At its core, this challenge can be viewed as principled one-step adaptation: incorporating new information with minimal interference to existing representations. Most existing approaches address this challenge by modifying model parameters or architectures in a supervised, task-specific manner. However, the underlying issue is representational: tasks require distinct yet structured representations that can be selectively updated without disrupting representations, while structure should reflect intrinsic organization in the data rather than task boundaries. In sequential data, time-delayed dependencies provide a natural signal for uncovering this organization, revealing how fundamental representations give rise to more specific ones. Inspired by the modular organization of the human brain, we propose MoRe, a framework that identifies modularity in the representation itself rather than allocating it at the architectural level. MoRe decomposes knowledge into a hierarchy of fundamental and specific modules with identifiability guarantees, enabling principled module reuse, alignment, and expansion during adaptation while preserving old modules by construction. Experiments on synthetic benchmarks and real-world LLM activations demonstrate interpretable hierarchical structure, improved plasticity-stability trade-offs, suggesting MoRe as a principled foundation for continual adaptation", "date": "2026-05-14", "categories": []}
{"id": "2605.20189", "title": "SOLAR: A Self-Optimizing Open-Ended Autonomous Agent for Lifelong Learning and Continual Adaptation", "abstract": "Despite the remarkable success of large language models (LLMs), they still face bottlenecks while deploying in dynamic, real-world settings with primary challenges being concept drift and the high cost of gradient-based adaptation. Traditional fine-tuning (FT) struggles to adapt to non-stationary data streams without resulting in catastrophic for getting or requiring extensive manual data curation. To address these limitations within the streaming and continual learning paradigm, we propose the Self-Optimizing Lifelong Autonomous Reasoner (SOLAR) which is an open-ended autonomous agent that leverages parameter-level meta-learning to self-improve, treating model weights as an environment for exploration. It initiates the process by consolidating a strong prior over common-sense knowledge making it effective for transfer-learning. By utilizing a multi-level reinforcement learning approach, SOLAR autonomously discovers adaptation strategies, enabling efficient test-time adaptation to unseen domains. Crucially, SOLAR maintains an evolving knowledge base of valid modification strategies, implicitly acting as an episodic memory buffer to balance plasticity (adaptation to new tasks) and stability (retention of meta-knowledge). Experiments demonstrate that SOLAR outperforms strong baselines on common-sense, mathematical, medical, coding, social and logical reasoning tasks, marking a significant step toward autonomous agents capable of lifelong adaptation in evolving environments.", "date": "2026-03-23", "categories": []}
{"id": "2605.20247", "title": "CP-MoE: Consistency-Preserving Mixture-of-Experts for Continual Learning", "abstract": "Catastrophic forgetting remains a major obstacle to continual learning in large language models (LLMs) and vision--language models (VLMs). Although Mixture-of-Experts (MoE) architectures offer an efficient path to scaling, existing LoRA-based MoE continual learning methods still face a fundamental trade-off: they either isolate experts too aggressively, limiting knowledge transfer across tasks, or allow task-specific updates to overwrite important existing parameters, leading to severe forgetting. To address this, we propose CP-MoE, a continual learning framework built around a transient expert that captures early task-specific updates and guides their integration into stable experts. CP-MoE introduces a consistency-preserving routing bias, which uses the transient expert to estimate representation similarity with stable experts and steer routing towards more compatible expert selection, and a transient expert-guided regularisation mechanism, which selectively protects important historical parameters during merging. Together, these components reduce parameter interference and forgetting while preserving cross-task knowledge transfer. We validate CP-MoE on both unimodal and multimodal continual learning benchmarks with LLM-based and VLM-based MoE models. On SuperNI benchmark, spanning diverse sequential language tasks, CP-MoE achieves state-of-the-art performance and stronger zero-shot transfer to unseen tasks. On VQA v2 dataset, it scales effectively to multimodal visual reasoning, consistently reduces forgetting, and outperforms strong MoE baselines.", "date": "2026-05-18", "categories": []}
{"id": "2605.20296", "title": "Spectral Unforgetting: Post-Hoc Recovery of Damaged Capabilities Without Retraining", "abstract": "Fine-tuning a language model for a target task routinely degrades capabilities the training data never explicitly threatened. We study this phenomenon, known as catastrophic forgetting, and propose a post-hoc repair solution that uses only the pretrained checkpoint $W_{\\mathrm{base}}$ and its fine-tuned descendant $W_{\\mathrm{ft}}$. The goal is not merely to revert the model toward the base checkpoint, but to recover capabilities damaged by fine-tuning while preserving both the target-task gains and any beneficial held-out improvements. We introduce DG-Hard, a checkpoint-only spectral repair method for the fine-tuning update $\\Delta = W_{\\mathrm{ft}} - W_{\\mathrm{base}}$. DG-Hard treats $\\Delta$ as a low-rank task-aligned signal embedded in an IID-like noise residual that gradient descent has no incentive to remove, and applies the Donoho-Gavish hard singular-value threshold to each weight-delta matrix, keeping the structured high-energy part of the update and removing the spectral bulk. This reduces repair to a closed-form SVD filtering step requiring no data-dependent tuning. A central difficulty is evaluation: average accuracy hides per-benchmark failures, while naive recovery scores reward models that simply revert toward the base. We therefore introduce a partition-conditional metric that separately tracks healing, preservation, non-damage, and target-task retention. Across $14$ (model, task) settings and nine cross-domain held-out benchmarks, DG-Hard achieves the strongest balanced repair among post-hoc baselines. DG-Hard also restores safety alignment degraded by benign fine-tuning on three independent safety axes, despite using no alignment data. These results suggest that part of fine-tuning-induced capability loss is not an unavoidable consequence of specialization, but a removable spectral residue in the weight update itself.", "date": "2026-05-19", "categories": []}
{"id": "2605.20538", "title": "Continual Segmentation under Joint Nonstationarity", "abstract": "Evolving data streams induce joint nonstationarity in continual semantic segmentation, where semantic classes, input distributions, and supervision availability change simultaneously over time. This setting reflects practical structured prediction systems, yet remains largely unexplored in prior continual learning work, which typically studies these factors in isolation. We formalize continual segmentation under coupled class, domain, and label shifts and investigate learning in heterogeneous dense prediction environments with limited annotations and abundant unlabeled data. To address instability and overfitting arising from few-shot supervision under distribution drift, we introduce gradient-adaptive stabilization, a parameter-wise regularization mechanism implemented via gradient-scaled stochastic perturbations that promotes a principled stability-plasticity tradeoff. We further leverage unlabeled data through semi-supervised learning and introduce prototype anchored supervision that validates pseudo-labels via joint confidence and prototype consistency. Together, these mechanisms enable learning under joint nonstationarity in continual segmentation. Extensive empirical evaluation across class-incremental, domain-incremental, and few-shot regimes demonstrates consistent improvements over prior methods in heterogeneous structured prediction settings. Our results expose fundamental failure modes of existing continual segmentation approaches and provide insight into learning robust dense predictors in dynamically evolving environments.", "date": "2026-05-19", "categories": []}
{"id": "2605.20803", "title": "Tunable MAGMAX: Preference-Aware Model Merging for Continual Learning", "abstract": "Continual learning (CL) aims to train models sequentially on multiple tasks while mitigating catastrophic forgetting of previously learned knowledge. Recent advances in large pre-trained models (LPMs) and model merging techniques, such as MAGMAX, have demonstrated effective CL performance by combining task-specific parameters. However, existing methods primarily focus on average performance across all tasks and do not adequately address how to construct models accommodating different deployment environments or varying user preferences. This paper proposes a model merging framework, termed Tunable MAGMAX, which enables preference-aware control of task-specific performance in CL. Our method introduces a preference vector that controls the number of elements selected from each task vector during model merging, allowing us to adjust the merged model performance according to their deployment needs. We further propose a method for automatically constructing appropriate preference vectors by leveraging small amounts of target environment data and datasets from model training tasks, thereby eliminating the need for manual specification. The experimental result on CL benchmark tasks demonstrates that Tunable MAGMAX effectively controls task-wise performance and successfully adapts merged models to various target environments. The proposed Tunable MAGMAX achieves superior or comparable performance to baseline methods, making it a practical solution for deploying CL models to various environments where the preferences of each task performance differ.", "date": "2026-05-20", "categories": []}
{"id": "2602.04555", "title": "Finding Structure in Continual Learning", "abstract": "Learning from a stream of tasks usually pits plasticity against stability: acquiring new knowledge often causes catastrophic forgetting of past information. Most methods address this by summing competing loss terms, creating gradient conflicts that are managed with complex and often inefficient strategies such as external memory replay or parameter regularization. We propose a reformulation of the continual learning objective using Douglas-Rachford Splitting (DRS). This reframes the learning process not as a direct trade-off, but as a negotiation between two decoupled objectives: one promoting plasticity for new tasks and the other enforcing stability of old knowledge. By iteratively finding a consensus through their proximal operators, DRS provides a more principled and stable learning dynamic. Our approach achieves an efficient balance between stability and plasticity without the need for auxiliary modules or complex add-ons, providing a simpler yet more powerful paradigm for continual learning systems.", "date": "2026-02-04", "categories": []}
{"id": "2605.04970", "title": "Skill Neologisms: Towards Skill-based Continual Learning", "abstract": "Modern LLMs show mastery over an ever-growing range of skills, as well as the ability to compose them flexibly. However, extending model capabilities to new skills in a scalable manner is an open problem: fine-tuning and parameter-efficient variants risk catastrophic forgetting, while context-based approaches have limited expressiveness and are constrained by the model's effective context. We explore skill neologisms--soft tokens integrated in the model's vocabulary and optimized to improve capabilities over a specific skill--as a way to selectively acquire new skills without weight updates. We first observe that pre-trained LLMs already exhibit tokens associated with procedural knowledge. We then show on a controlled synthetic task that skill neologisms can be learned to improve model capabilities on specific skills while being composable with out-of-distribution skills, and that independently trained skill neologisms can be composed zero-shot. Finally, we validate zero-shot composition of independently learned skill neologisms on the more realistic natural language setting of the Skill-Mix benchmark. These results suggest that skill neologisms may provide a scalable path towards skill-based continual learning.", "date": "2026-05-06", "categories": []}
{"id": "2605.16353", "title": "StrLoRA: Towards Streaming Continual Visual Instruction Tuning for MLLMs", "abstract": "Continual Visual Instruction Tuning (CVIT) enables Multimodal Large Language Models to incrementally acquire new abilities. However, existing CVIT methods operate under a restrictive task-incremental setting, where each training phase corresponds to a single, predefined task. This does not reflect real-world conditions, where data arrives as a continuous stream of interleaved and dynamically evolving tasks. To bridge this gap, we introduce Streaming CVIT (StrCVIT), a more general and realistic setting where models learn from a stream of data chunks containing a dynamic mixture of tasks. In StrCVIT, a model must simultaneously acquire new abilities, reinforce recurring abilities, and mitigate forgetting. Existing CVIT methods fail here as they cannot reliably distinguish or adapt to the heterogeneous task samples within each chunk. We therefore propose StrLoRA, a regularized two-stage expert routing framework. StrLoRA first performs task-aware expert selection using the textual instruction to activate a sparse subset of relevant experts, reducing cross-task interference. It then applies token-wise expert weighting within this subset, where contribution weights are computed via cross-modal attention between local visual tokens and the global instruction representation. To maintain stability across the non-stationary stream, a routing-stability regularization aligns current routing distributions with a historical exponential moving average reference. Extensive experiments on a newly developed StrCVIT benchmark show that StrLoRA substantially outperforms existing methods, effectively enhancing model's abilities from continuously evolving data streams. The code is available at https://github.com/chanceche/StrCVIT.", "date": "2026-05-08", "categories": []}
{"id": "2605.18842", "title": "Safe Continual Reinforcement Learning under Nonstationarity via Adaptive Safety Constraints", "abstract": "Safe reinforcement learning in nonstationary environments requires safety mechanisms that adapt as environmental conditions change. Standard safe reinforcement learning methods often assume fixed constraints or stable environmental conditions, which can become inadequate under distribution shift. We propose LILAC+, a framework for safe continual reinforcement learning under nonstationarity that combines three adaptive safety mechanisms: context-based safety constraints, adaptation-speed constraints, and budget-to-state safety enforcement. Context-based constraints adjust safety requirements using inferred and predicted environmental context. Adaptation-speed constraints tighten safety requirements when the rate of environmental change exceeds the agent's ability to adapt safely. Budget-to-state enforcement converts cumulative safety requirements into local state-level control constraints that can be enforced at decision time. Together, these mechanisms provide a unified approach for proactive and reactive safety adaptation in continual reinforcement learning. We evaluate the framework in simulated driving environments under stationary, seen nonstationary, and unseen nonstationary conditions. The results show that adaptive safety constraints substantially reduce safety violations under distribution shift while maintaining competitive task performance compared with unconstrained and fixed-constraint baselines. These findings suggest that safe continual reinforcement learning requires adaptive constraint mechanisms that respond not only to current state information but also to predicted environmental context, adaptation demand, and remaining safety budget.", "date": "2026-05-13", "categories": []}
{"id": "2605.18903", "title": "Reasoning Portability: Guiding Continual Learning for MLLMs in the RLVR Era", "abstract": "Vision-Language Models in Continual Learning (VLM-CL) aim to continuously adapt to new multimodal tasks while retaining prior knowledge. The emerging paradigm that couples Multimodal Large Language Models (MLLMs) with Reinforcement Learning with Verifiable Rewards (RLVR) calls for a new pattern to guide continual adaptation. Advances in reasoning capability now make it feasible to impose constraints at the reasoning level. We formalize portability, a sample-level measure of how reusable the previous policy's behavior is on a new task, and empirically show that reasoning-level signals remain reliable on out-of-distribution samples while answer-level signals do not. We instantiate this as Reasoning Portability (RP) and propose Reasoning-based Dynamic Balance Continual Learning (RDB-CL), which modulates the per-sample Kullback-Leibler regularization in RLVR according to RP: a tight anchor preserves reusable reasoning on high-RP samples, while a relaxed anchor on low-RP samples permits exploration of new reasoning pathways. Experiments show that RDB-CL consistently outperforms baselines, improving Last accuracy by +12.0% over the vanilla RLVR baseline.", "date": "2026-05-17", "categories": []}
{"id": "2605.19080", "title": "MANGO: Meta-Adaptive Network Gradient Optimization for Online Continual Learning", "abstract": "In Online Continual Learning (OCL), a neural network sequentially learns from a non-stationary data stream in a single-pass with access only to a limited memory replay buffer. This contrasts sharply with off-line continual learning where training is multiple epoch dependent on large datasets. The main challenge faced by OCL is to overcome catastrophic forgetting of past tasks (stability) while learning new ones efficiently (plasticity). Existing methods counter forgetting via replay-based rehearsal, output level distillation, fixed regularization, or meta-learning on the current data. However, these methods have limitations: rehearsal introduces a stored sample bias; distillation operates on output-distributions without modulating parameter updates; fixed-regularization penalizes parameters irrespective of sensitivity; stream-only meta-learning lacks a feedback controlled parameter update. We propose Meta-Adaptive Network Gradient Optimization (MANGO), an OCL framework that balances stability-plasticity via gradient-gating and meta-learned regularization. Gradient-gating scales parameter updates based on sensitivity, preventing destructive updates. Meta-learned regularization adapts stability coefficients, evaluating the effect of parameter update on replay. In MANGO, replay acts as both a training signal and a forgetting evaluator. We evaluated our method on three standard OCL benchmark datasets. MANGO outperforms strong baselines, achieving state-of-the-art results with consistent performance across replay sizes. In domain incremental learning on CLEAR-10 and class incremental learning on CIFAR-100 and Tiny-ImageNet, it achieves highest accuracy among all baselines and achieves positive Backward Transfer, overcoming forgetting on CLEAR-10.", "date": "2026-05-18", "categories": []}
{"id": "2605.19301", "title": "iGSP:Implicit Gradient Subspace Projection for Efficient Continual Learning of Vision-Language Models", "abstract": "Vision-Language Models require efficient adaptation to continually emerging downstream tasks. While Parameter-Efficient Fine-Tuning mitigates catastrophic forgetting, assigning isolated modules per task leads to parameter explosion. Conversely, recent similarity-driven sharing mechanisms falsely equate superficial visual similarity with underlying alignment consistency. This fundamental mismatch triggers severe negative transfer between visually similar but logically distinct tasks and fails to exploit alignment reuse across visually diverse ones. We argue thatalignment sharing is fundamentally a geometric problem of overlapping optimization trajectories within shared low-rank subspaces. Grounded in this insight, we propose iGSP, a novel framework that achieves efficient adaptation via implicit gradient subspace projection. Leveraging the early convergence of MoE routers to establish the subspace basis, iGSP bifurcates the adaptation process into two phases. First, the Subspace Identification phase introduces candidate experts via basis pre-expansion, applies a novel subspace-constrained regularization to implicitly project new task gradients onto the historical subspace, and precisely prunes redundant dimensions by treating routing probabilities as gradient flow indicators, ultimately to maximize knowledge reuse. Second, the Orthogonal Subspace Fine-Tuning phase fixes this structural basis and removes the regularization to rapidly fit the task-specific residual loss. Extensive experiments on the MTIL benchmark demonstrate that iGSP achieves state-of-the-art accuracy while significantly improving training efficiency, reducing the average trainable parameters by 42.7\\% compared to current SOTA methods, and decreasing the final total parameters by 86.9\\% relative to counterparts. The source code is available at https://github.com/GeoX-Lab/iGSP.", "date": "2026-05-19", "categories": []}
{"id": "2508.04227", "title": "Continual Learning for VLMs: A Survey and Taxonomy Beyond Forgetting", "abstract": "Vision-language models (VLMs) and the recent surge of Multimodal Large Language Models (MLLMs) have revolutionized artificial intelligence with unprecedented cross-modal alignment and zero-shot generalization. However, enabling them to learn continually from non-stationary data remains a major challenge, as their cross-modal alignment and generalization capabilities are particularly vulnerable to catastrophic forgetting. Unlike traditional unimodal continual learning (CL), VLMs face unique challenges such as cross-modal feature drift, parameter interference due to shared architectures, and zero-shot capability erosion. Furthermore, generative MLLMs exhibit a unique ``alignment tax,'' where catastrophic forgetting manifests not merely as factual amnesia, but as a systemic collapse of deep Chain-of-Thought (CoT) reasoning. This survey presents the first comprehensive, diagnostic review bridging continual learning for both predictive VLMs and generative MLLMs. We systematically deconstruct the aforementioned failure modes and propose a challenge-driven taxonomy comprising four core paradigms: (1) Multi-Modal Replay Strategies addressing explicit and implicit memory drift; (2) Cross-Modal Regularization enforcing topological and geometric alignment; (3) Parameter-Efficient Adaptation} utilizing dynamic routing and subspace projections; and the emerging (4) Model Fusion and Decoupling paradigms. We critically analyze the evolution of evaluation protocols, highlighting the essential shift toward dual-track benchmarks (Domain vs. Ability CL) and micro-diagnostic CoT evaluations. Finally, we chart a roadmap for future research, emphasizing compositional zero-shot learning, embodied AI with sensor fusion, and autonomous agentic ecosystems. All resources are available at: https://github.com/YuyangSunshine/Awesome-Continual-learning-of-Vision-Language-Models.", "date": "2025-08-06", "categories": []}
{"id": "2510.00304", "title": "Barriers for Learning in an Evolving World: Mathematical Understanding of Loss of Plasticity", "abstract": "Deep learning models excel in stationary data but struggle in non-stationary environments due to a phenomenon known as loss of plasticity (LoP), the degradation of their ability to learn in the future. This work presents a first-principles investigation of LoP in gradient-based learning. Grounded in dynamical systems theory, we formally define LoP by identifying stable manifolds in the parameter space that trap gradient trajectories. Our analysis reveals two primary mechanisms that create these traps: frozen units from activation saturation and cloned-unit manifolds from representational redundancy. Our framework uncovers a fundamental tension: properties that promote generalization in static settings, such as low-rank representations and simplicity biases, directly contribute to LoP in continual learning scenarios. We validate our theoretical analysis with numerical simulations and explore architectural choices or targeted perturbations as potential mitigation strategies.", "date": "2025-09-30", "categories": []}
{"id": "2605.11617", "title": "MIST: Reliable Streaming Decision Trees for Online Class-Incremental Learning via McDiarmid Bound", "abstract": "Streaming decision trees are natural candidates for open-world continual learning, as they perform local updates, enjoy bounded memory, and static decision boundaries. Despite these, they still fail in online class-incremental learning due to two coupled miscalibrations: (i) their split criterion grows unreliable as the class count K expands, and (ii) the absence of knowledge transfer at split time. Both failures share a common root: the range of Information Gain intrinsically scales with log2 K. Consequently, any Hoeffding-style confidence radius derived from it must inevitably grow with the class count, making a K-independent split criterion structurally impossible, taking away the potential benefits of applying streaming decision trees to continual learning. To fix this issue, we present MIST (McDiarmid Incremental Streaming Tree), which resolves both failures through three integrated components: (i) a tight, K-independent McDiarmid confidence radius for Gini splitting that acts as a structural regulariser; (ii) a Bayesian inheritance protocol that projects parent statistics to child nodes via truncated-Gaussian moments, with variance reduction guarantees strongest precisely when splitting is most conservative; and (iii) per-leaf KLL quantile sketches that support both continuous threshold evaluation and geometry-adaptive leaf prediction from a single data structure. On standard and stress-test tabular streams, MIST is competitive with global parametric methods on near-Gaussian benchmarks and uniquely robust on non-Gaussian geometry where SOTA benchmarks collapse.", "date": "2026-05-12", "categories": []}
{"id": "2605.11710", "title": "Unlocking Compositional Generalization in Continual Few-Shot Learning", "abstract": "Object-centric representations promise a key property for few-shot learning: Rather than treating a scene as a single unit, a model can decompose it into individual object-level parts that can be matched and compared across different concepts. In practice, this potential is rarely realized. Continual learners either collapse scenes into global embeddings, or train with part-level matching objectives that tie representations too closely to seen patterns, leaving them unable to generalize to truly novel concepts. In this paper, we identify this fundamental structural conflict and pioneer a new paradigm that strictly decouples representation learning from compositional inference. Leveraging the inherent patch-level semantic geometry of self-supervised Vision Transformers (ViTs), our framework employs a dual-phase strategy. During training, slot representations are optimized entirely toward holistic class identity, preserving highly generalizable, object-level geometries. At inference, preserved slots are dynamically composed to match novel scenes. We demonstrate that this paradigm offers dual structural benefits: The frozen backbone naturally prevents representation drift, while our lightweight, holistic optimization preserves the features' capacity for novel-concept transfer. Extensive experiments validate this approach, achieving state-of-the-art unseen-concept generalization and minimal forgetting across standard continual learning benchmarks.", "date": "2026-05-12", "categories": []}
{"id": "2605.12451", "title": "FuTCR: Future-Targeted Contrast and Repulsion for Continual Panoptic Segmentation", "abstract": "Continual Panoptic Segmentation (CPS) requires methods that can quickly adapt to new categories over time. The nature of this dense prediction task means that training images may contain a mix of labeled and unlabeled objects. As nothing is known about these unlabeled objects a priori, existing methods often simply group any unlabeled pixel into a single \"background\" class during training. In effect, during training, they repeatedly tell the model that all the different background categories are the same (even when they aren't). This makes learning to identify different background categories as they are added challenging since these new categories may require using information the model was previously told was unimportant and ignored. Thus, we propose a Future-Targeted Contrastive and Repulsive (FuTCR) framework that addresses this limitation by restructuring representations before new classes are introduced. FuTCR first discovers confident future-like regions by grouping model-predicted masks whose pixels are consistently classified as background but exhibit non-background logits. Next, FuTCR applies pixel-to-region contrast to build coherent prototypes from these unlabeled regions, while simultaneously repelling background features away from known-class prototypes to explicitly reserve representational space for future categories. Experiments across six CPS settings and a range of dataset sizes show FuTCR improves relative new-class panoptic quality over the state-of-the-art by up to 28%, while preserving or improving base-class performance with gains up to 4%.", "date": "2026-05-12", "categories": []}
{"id": "2605.16374", "title": "Lost or Hidden? A Concept-Level Forgetting in Supervised Continual Learning", "abstract": "Continual learning studies how models can adapt to new tasks while retaining previously acquired knowledge. Although a broad spectrum of methods has been proposed to mitigate catastrophic forgetting, the field remains predominantly performance-driven, with limited insight into what forgetting actually corresponds to within the vision model's representation space. Prior work has primarily analyzed forgetting through task-level performance or coarse measures of representational drift, without disentangling output-level accessibility from changes in finer-grained internal structure. To this end, we propose a diagnostic framework that leverages Sparse Autoencoders (SAEs) to define a task-anchored latent feature space, enabling analysis of how task-specific information evolves at a finer granularity, where individual SAE latents are treated as concept proxies for recurring and relatively disentangled visual patterns in the model's internal computations. Within this framework, we decompose forgetting into apparent concept deletion, recoverability, and decodability. We show that a large portion of seemingly lost concept-level information can often be recovered under linearity assumption, with concept decodability degrading as more tasks are introduced. Overall, our findings suggest that a significant part of concept-level forgetting can be attributed to changes in the representational accessibility rather than complete information erasure.", "date": "2026-05-10", "categories": []}
{"id": "2605.17361", "title": "\\textsc{MasFACT}: Continual Multi-Agent Topology Learning via Geometry-Aware Posterior Transfer", "abstract": "Multi-agent systems (MAS) powered by large language models (LLMs) have emerged as a powerful paradigm for complex problem solving, where performance critically depends on the underlying inter-agent communication topology. However, existing topology generation methods mainly optimize for isolated tasks, while real-world deployments involve streams of evolving tasks, requiring previously effective collaboration patterns to be retained and reused rather than rediscovered or overwritten. We identify a previously underexplored failure mode, \\emph{topology forgetting}, in which adapting to new tasks shifts the topology generator away from communication structures required by earlier tasks. This issue stems from cross-task misalignment in both agent-level functional semantics and relational communication structures. To address this challenge, we propose \\textbf{\\textsc{MasFACT}}, a geometry-aware posterior transfer framework that preserves and reuses historical collaboration knowledge as transferable topology priors. We transfer these priors across task-specific agent spaces through Fused Gromov-Wasserstein optimal transport and perform PAC-Bayes-guided conservative posterior adaptation to balance task-specific plasticity with structural stability. Experiments across class-, domain-, and task-level continual settings demonstrate that \\textsc{MasFACT} consistently improves average accuracy while reducing topology forgetting compared to strong topology generation and replay-based baselines, and can be seamlessly integrated with different MAS topology generators.", "date": "2026-05-17", "categories": []}
{"id": "2605.17571", "title": "Stable Routing for Mixture-of-Experts in Class-Incremental Learning", "abstract": "Class-incremental learning (CIL) requires models to learn new classes sequentially while preserving prior knowledge. Recently, approaches that combine pre-trained models with mixture-of-experts (MoE) have received increasing attention in CIL: they typically expand experts during learning and employ a router to assign weights across experts. However, existing MoE methods often overlook routing drift induced by expert expansion. Once new experts are introduced, the router may reassign samples from earlier classes to newly added experts, thereby perturbing previously established expert compositions and causing interference even when old experts remain frozen. We argue that expandable MoE in CIL requires two complementary properties: stable old-class routing for knowledge preservation and sufficient capacity utilization for new-class adaptation. To this end, we propose Stable Routing for MoE (StaR-MoE), a routing-level framework for expandable MoE in CIL. By incorporating sensitivity-aware routing alignment, StaR-MoE aligns current old-class routing behavior with historical routing distributions through sensitivity-guided constraints. Complementarily, StaR-MoE introduces asymmetric capacity regularization to encourage effective utilization of the expanded expert pool without compromising class-specific routing specialization. Extensive experiments across four standard CIL benchmarks demonstrate that StaR-MoE consistently improves both average and last accuracy over state-of-the-art methods, highlighting the importance of stable routing.", "date": "2026-05-17", "categories": []}
{"id": "2605.17729", "title": "Domain Incremental Learning for Pandemic-Resilient Chest X-Ray Analysis", "abstract": "Deep learning models achieved high accuracy in pneumonia detection from chest X-rays. However, their generalization across clinical domains remains limited due to variations in imaging devices, acquisition protocols, and institutional conditions. This study introduces a replay-based domain-incremental continual learning designed to enable continual adaptation to cross-domain variations without catastrophic forgetting. The proposed method incorporates a class-aware balanced replay to maintain balanced class representation within a constrained memory and a class-aware loss to dynamically reweight class imbalance during training. Experiments conducted on a domain-shifted PneumoniaMNIST dataset consisting of five simulated domains demonstrate that the proposed method achieves an average accuracy of 88.66%, outperforming Experience Replay, Fine-Tuning, and Joint Training baselines. These findings highlight the efficacy of the proposed approach in achieving robust and consistent pneumonia detection across clinical environment variations.", "date": "2026-05-18", "categories": []}
{"id": "2601.09512", "title": "CLARE: Continual Learning for Vision-Language-Action Models via Autonomous Adapter Routing and Expansion", "abstract": "To teach robots complex manipulation tasks, a common approach is to fine-tune a pre-trained vision-language-action model (VLA) on task-specific data. However, since this recipe updates existing representations, it is unsuitable for long-term operation in the real world, where robots must continually adapt to new tasks and environments while retaining the knowledge they have already acquired. Existing continual learning methods for robotics commonly require storing previous data (exemplars), struggle with long task sequences, or rely on task identifiers for deployment. To address these limitations, we propose CLARE, a general, parameter-efficient framework for exemplar-free continual learning with VLAs. CLARE introduces lightweight modular adapters into selected VLA modules and autonomously expands the model only where necessary when learning a new task, guided by layer-wise feature similarity. During deployment, an autoencoder-based routing mechanism dynamically activates the most relevant adapters without requiring task labels. Through extensive experiments on the LIBERO benchmark and five real-world tasks, we show that CLARE achieves high performance on new tasks without catastrophic forgetting of earlier tasks, significantly outperforming even exemplar-based methods. Code, data, and videos are available at our website: https://tum-lsy.github.io/clare.", "date": "2026-01-14", "categories": []}
{"id": "2605.08949", "title": "Muon-OGD: Muon-based Spectral Orthogonal Gradient Projection for LLM Continual Learning", "abstract": "A central challenge in continual learning for large language models (LLMs) is catastrophic forgetting, where adapting to new tasks can substantially degrade performance on previously learned ones. Existing projection-based methods mitigate such interference by restricting parameter updates to subspaces that are orthogonal to directions associated with past tasks. However, these methods are typically formulated under Euclidean parameter geometry, with update magnitudes and projections governed by the Frobenius norm. The recent empirical success of the Muon optimizer, which applies orthogonalized matrix updates and admits a spectral-norm interpretation, suggests that Frobenius geometry may not be the most effective choice for matrix-valued LLM parameters. Motivated by this observation, we propose Muon-OGD, a spectral-norm-aware continual learning framework that integrates Muon-style operator-norm geometry with orthogonal projection constraints. Our method formulates each update as a spectral-norm-constrained optimization problem with linear non-interference constraints, and solves it efficiently through dual iterations and Newton--Schulz matrix-sign approximations. By applying orthogonalized momentum updates that avoid protected directions associated with prior tasks, Muon-OGD aims to improve the stability--plasticity trade-off in sequential LLM adaptation. We evaluate the proposed method on standard continual learning benchmarks, TRACE, and domain-specific Coding--Math--Medical curricula using both encoder--decoder and decoder-only architectures. Empirically, Muon-OGD consistently improves over sequential fine-tuning and competitive orthogonal-gradient baselines, while remaining computationally scalable. These results suggest that spectral-norm-aware update geometry provides a practical and effective alternative to Frobenius-norm projection for continual learning in LLMs.", "date": "2026-05-09", "categories": []}
{"id": "2605.15053", "title": "TFGN: Task-Free, Replay-Free Continual Pre-Training Without Catastrophic Forgetting at LLM Scale", "abstract": "Continually pre-training a large language model on heterogeneous text domains, without replay or task labels, has remained an unsolved architectural problem at LLM scale. Existing methods rely on replay buffers, task identifiers, regularization penalties that scale poorly, or sentence-classification-scale evaluation. We introduce TFGN, an architectural overlay for transformer language models that produces input-conditioned, parameter-efficient updates while leaving the rest of the transformer unchanged. On six heterogeneous text domains (Prose, Python, Math, Biomedical, Chinese, JavaScript) at 1B tokens per phase across three model scales (~398M, ~739M, ~9B) and two regimes (From-Scratch and Retrofit), TFGN achieves backward transfer of -0.007 at LLaMA 3.1 8B Retrofit, HellaSwag retention 0.506/0.504/0.510, and >=99.59% L2-orthogonal gradient separation between domain pairs - with no replay, no task IDs, no Fisher penalty. The same matrices show positive cross-domain forward transfer: held-out JavaScript PPL drops 26.8% at LLaMA-8B Retrofit and 62.0% at GPT-2 Medium From-Scratch purely from Python training. Two extensions on the same substrate close further open problems. A closed-loop meta-control layer (Extension A) reduces forgetting by an additional 81% at ~398M, mapping onto the System A and System M roles of Dupoux et al. (arXiv:2603.15381). An operator-level plan vector (Extension B) reshapes forward-pass behavior at 99.96% cosine fidelity over 30 source->target pairs. The architectural insight is a Read/Write decomposition: the forward pass is fully dense, while cross-domain parameter updates are structured so prior-domain subspaces are not written to. To our knowledge, TFGN is the first architecture that simultaneously closes catastrophic forgetting at LLM scale, realizes a closed-loop autonomous-learning meta-controller, and carries an operator-level latent planner.", "date": "2026-05-14", "categories": []}
{"id": "2605.15775", "title": "Continual Learning of Domain-Invariant Representations", "abstract": "Continual learning (CL) aims to train models sequentially over multiple domains without forgetting previously learned knowledge. However, existing CL methods optimize for in-domain performance and are therefore prone to learning spurious, domain-specific cues (``shortcut learning''), which limits generalization to unseen domains after deployment. In this paper, we address this limitation through continual learning of domain-invariant representation. We introduce a broad class of CL methods that sequentially learn representations capturing invariant structures across domains. Our methods are motivated by the observation that such invariant structures often preserve the underlying causal mechanisms, which can reduce the risk of overfitting to domain-specific cues and thus offer better out-of-domain generalization. Our proposed CL methods combine replay-based training with a tailored sequential invariance alignment to learn -- and preserve -- invariant structures over time. We evaluate our methods under a deployment-oriented protocol that measures performance on unseen target domains. Across six benchmark and real-world datasets spanning vision, medicine, manufacturing, and ecology, our methods consistently outperform existing CL baselines in terms of generalization to unseen target domains. As an ablation, we further show that na\\\"ive extensions of sequential training with existing domain-invariant representation learning (DIRL) methods provide only limited benefits. To the best of our knowledge, this is the first work to develop domain-invariant representation methods for CL.", "date": "2026-05-15", "categories": []}
{"id": "2605.15877", "title": "Shapley Neuron Values for Continual Learning: Which Neurons Matter Most?", "abstract": "Continual learning enables neural networks to learn tasks sequentially without forgetting previously acquired knowledge. However, neural networks suffer from catastrophic forgetting, where learning new tasks degrades performance on earlier ones. We address this problem with Shapley Neuron Valuation (SNV), a principled framework that quantifies Neuron importance in continual learning, grounded in cooperative game theory. SNV selectively freezes important Neurons while keeping others plastic, enabling buffer-free continual learning without expanding architecture. Experiments on ImageNet-1k show that SNV consistently outperforms existing buffer-free methods. In particular, SNV improves accuracy by +2.88% in the class incremental learning and +6.46% in the task incremental learning scenarios compared to the second baseline.", "date": "2026-05-15", "categories": []}
{"id": "2508.06202", "title": "LoRA in LoRA: Towards Parameter-Efficient Architecture Expansion for Continual Visual Instruction Tuning", "abstract": "Continual Visual Instruction Tuning (CVIT) enables Multimodal Large Language Models (MLLMs) to incrementally learn new tasks over time. However, this process is challenged by catastrophic forgetting, where performance on previously learned tasks deteriorates as the model adapts to new ones. A common approach to mitigate forgetting is architecture expansion, which introduces task-specific modules to prevent interference. Yet, existing methods often expand entire layers for each task, leading to significant parameter overhead and poor scalability. To overcome these issues, we introduce LoRA in LoRA (LiLoRA), a highly efficient architecture expansion method tailored for CVIT in MLLMs. LiLoRA shares the LoRA matrix A across tasks to reduce redundancy, applies an additional low-rank decomposition to matrix B to minimize task-specific parameters, and incorporates a cosine-regularized stability loss to preserve consistency in shared representations over time. Extensive experiments on a diverse CVIT benchmark show that LiLoRA consistently achieves superior performance in sequential task learning while significantly improving parameter efficiency compared to existing approaches. The code is available at https://github.com/chanceche/LiLoRA.", "date": "2025-08-08", "categories": []}
{"id": "2511.14823", "title": "Dynamic Nested Hierarchies: Pioneering Self-Evolution in Machine Learning Architectures for Lifelong Intelligence", "abstract": "Contemporary machine learning models, including large language models, exhibit remarkable capabilities in static tasks yet falter in non-stationary environments due to rigid architectures that hinder continual adaptation and lifelong learning. Building upon the nested learning paradigm, which decomposes models into multi-level optimization problems with fixed update frequencies, this work proposes dynamic nested hierarchies as the next evolutionary step in advancing artificial intelligence and machine learning. Dynamic nested hierarchies empower models to autonomously adjust the number of optimization levels, their nesting structures, and update frequencies during training or inference, inspired by neuroplasticity to enable self-evolution without predefined constraints. This innovation addresses the anterograde amnesia in existing models, facilitating true lifelong learning by dynamically compressing context flows and adapting to distribution shifts. Through rigorous mathematical formulations, theoretical proofs of convergence, expressivity bounds, and sublinear regret in varying regimes, alongside empirical demonstrations of superior performance in language modeling, continual learning, and long-context reasoning, dynamic nested hierarchies establish a foundational advancement toward adaptive, general-purpose intelligence.", "date": "2025-11-18", "categories": []}
{"id": "2605.09664", "title": "FreeMOCA: Memory-Free Continual Learning for Malicious Code Analysis", "abstract": "As over 200 million new malware samples are identified each year, antivirus systems must continuously adapt to the evolving threat landscape. However, retraining solely on new samples leads to catastrophic forgetting and exploitable blind spots, while retraining on the entire dataset incurs substantial computational cost. We propose FreeMOCA, a memory- and compute-efficient continual learning framework for malicious code analysis that preserves prior knowledge via adaptive layer-wise interpolation between consecutive task updates, leveraging the fact that warm-started task optima are connected by low-loss paths in parameter space.\n  We evaluate FreeMOCA in both class-incremental (Class-IL) and domain-incremental (Domain-IL) settings on large-scale Windows (EMBER) and Android (AZ) malware benchmarks. FreeMOCA achieves substantial gains in Class-IL, outperforming 11 baselines on both EMBER and AZ benchmarks. It also significantly reduces forgetting, achieving the best retention across baselines, and improving accuracy by up to 42% and 37% on EMBER and AZ, respectively. These results demonstrate that warm-started interpolation in parameter space provides a scalable and effective alternative to replay for continual malware detection.\n  Code is available at: https://github.com/IQSeC-Lab/FreeMOCA.", "date": "2026-05-10", "categories": []}
{"id": "2605.12484", "title": "Learning, Fast and Slow: Towards LLMs That Adapt Continually", "abstract": "Large language models (LLMs) are trained for downstream tasks by updating their parameters (e.g., via RL). However, updating parameters forces them to absorb task-specific information, which can result in catastrophic forgetting and loss of plasticity. In contrast, in-context learning with fixed LLM parameters can cheaply and rapidly adapt to task-specific requirements (e.g., prompt optimization), but cannot by itself typically match the performance gains available through updating LLM parameters. There is no good reason for restricting learning to being in-context or in-weights. Moreover, humans also likely learn at different time scales (e.g., System 1 vs 2). To this end, we introduce a fast-slow learning framework for LLMs, with model parameters as \"slow\" weights and optimized context as \"fast\" weights. These fast \"weights\" can learn from textual feedback to absorb the task-specific information, while allowing slow weights to stay closer to the base model and persist general reasoning behaviors. Fast-Slow Training (FST) is up to 3x more sample-efficient than only slow learning (RL) across reasoning tasks, while consistently reaching a higher performance asymptote. Moreover, FST-trained models remain closer to the base LLM (up to 70% less KL divergence), resulting in less catastrophic forgetting than RL-training. This reduced drift also preserves plasticity: after training on one task, FST trained models adapt more effectively to a subsequent task than parameter-only trained models. In continual learning scenarios, where task domains change on the fly, FST continues to acquire each new task while parameter-only RL stalls.", "date": "2026-05-12", "categories": []}
{"id": "2605.12998", "title": "DRIFT: A Benchmark for Task-Free Continual Graph Learning with Continuous Distribution Shifts", "abstract": "Continual graph learning (CGL) aims to learn from dynamically evolving graphs while mitigating catastrophic forgetting. Existing CGL approaches typically adopt a task-based formulation, where the data stream is partitioned into a sequence of discrete tasks with pre-defined boundaries. However, such assumptions rarely hold in real-world environments, where data distributions evolve continuously and task identity is often unavailable. To better reflect realistic non-stationary environments, we revisit continual graph learning from a task-free perspective. We propose a unified formulation that models the data stream as a time-varying mixture of latent task distributions, enabling continuous modeling of distribution drift. Based on this formulation, we construct \\emph{DRIFT}, a benchmark that spans a spectrum of transition dynamics ranging from hard task switches to smooth distributional drift through a Gaussian parameterization. We evaluate representative continual learning methods under this task-free setting and observe substantial performance degradation compared to traditional task-based protocols. Our findings indicate that many existing approaches implicitly rely on task boundary information and struggle under realistic task-free graph streams. This work highlights the importance of studying continual graph learning under realistic non-stationary conditions and provides a benchmark for future research in this direction. Our code is available at https://github.com/UConn-DSIS/DRIFT.", "date": "2026-05-13", "categories": []}
{"id": "2605.14785", "title": "Understanding Imbalanced Forgetting in Rehearsal-Based Class-Incremental Learning", "abstract": "Neural networks suffer from catastrophic forgetting in class-incremental learning (CIL) settings. Rehearsal$\\unicode{x2013}$replaying a subset of past samples$\\unicode{x2013}$is a well-established mitigation strategy. However, recent results suggest that, despite balanced rehearsal allocation, some classes are forgotten substantially more than others. Despite its relevance, this imbalanced forgetting phenomenon remains underexplored. This work shows that imbalanced forgetting arises systematically and severely in rehearsal-based CIL and investigates it extensively. Specifically, we construct, from a principled analysis, three last-layer coefficients that capture different gradient-level sources of interference affecting each past class during an incremental step. We then demonstrate that, together, they reliably predict how past classes will rank in terms of forgetting at the end of that step. While predictive performance alone does not establish causality, these results support the interpretation of the coefficients as a plausible mechanistic account linking last-layer gradient-level interactions during training to class-level forgetting outcomes. Notably, one coefficient$\\unicode{x2013}$capturing self-induced interference$\\unicode{x2013}$emerges as the strongest predictor, with controlled experiments providing evidence consistent with this coefficient being influenced by the new-class interference coefficient. Overall, our findings provide valuable insights and suggest promising directions for mitigating imbalanced forgetting by reducing class-wise disparities in the identified sources of interference.", "date": "2026-05-14", "categories": []}
{"id": "2605.14938", "title": "Octopus: History-Free Gradient Orthogonalization for Continual Learning in Multimodal Large Language Models", "abstract": "Continual learning in multimodal large language models (MLLMs) aims to sequentially acquire knowledge while mitigating catastrophic forgetting, yet existing methods face inherent limitations: architecture-based approaches incur additional computational overhead and often generalize poorly to new tasks, rehearsal-based methods rely on storing historical data, raising privacy and storage concerns, and conventional regularization-based strategies alone are insufficient to fully prevent parameter interference. We propose Octopus, a two-stage continual learning framework based on History-Free Gradient Orthogonalization (HiFGO), which enforces gradient-level orthogonality without historical task data. Our proposed two-stage finetuning strategy decouples task adaptation from regularization, achieving a principled balance between plasticity and stability. Experiments on UCIT show that Octopus establishes state-of-the-art performance, surpassing prior SOTA by 2.14% and 6.82% in terms of Avg and Last.", "date": "2026-05-14", "categories": []}
{"id": "2605.14948", "title": "ACE-LoRA: Adaptive Orthogonal Decoupling for Continual Image Editing", "abstract": "State-of-the-art diffusion models often rely on parameter-efficient fine-tuning to perform specialized image editing tasks. However, real-world applications require continual adaptation to new tasks while preserving previously learned knowledge. Despite the practical necessity, continual learning for image editing remains largely underexplored. We propose ACE-LoRA, a dynamic regularization framework for continual image editing that effectively mitigates catastrophic forgetting. ACE-LoRA leverages Adaptive Orthogonal Decoupling to identify and orthogonalize task interference, and introduces a Rank-Invariant Historical Information Compression strategy to address scalability issues in continual updates. To facilitate continual learning in image editing and provide a standardized evaluation protocol, we introduce CIE-Bench, the first comprehensive benchmark in this domain. CIE-Bench encompasses diverse and practically relevant image editing scenarios with a balanced level of difficulty to effectively expose limitations of existing models while remaining compatible with parameter-efficient fine-tuning. Extensive experiments demonstrate that our method consistently outperforms existing baselines in terms of instruction fidelity, visual realism, and robustness to forgetting, establishing a strong foundation for continual learning in image editing.", "date": "2026-05-14", "categories": []}
{"id": "2505.18604", "title": "Exemplar-Free Continual Learning for State Space Models", "abstract": "State-Space Models (SSMs) excel at capturing long-range dependencies with structured recurrence, making them well-suited for sequence modeling. However, their evolving internal states pose challenges in adapting them under Continual Learning (CL). This is particularly difficult in exemplar-free settings, where the absence of prior data leaves updates to the dynamic SSM states unconstrained, resulting in catastrophic forgetting. To address this, we propose Inf-SSM, a novel and simple geometry-aware regularization method that utilizes the geometry of the infinite-dimensional Grassmannian to constrain state evolution during CL. Unlike classical continual learning methods that constrain weight updates, Inf-SSM regularizes the infinite-horizon evolution of SSMs encoded in their extended observability subspace. We show that enforcing this regularization requires solving a matrix equation known as the Sylvester equation, which typically incurs $\\mathcal{O}(n^3)$ complexity. We develop a $\\mathcal{O}(n^2)$ solution by exploiting the structure and properties of SSMs. This leads to an efficient regularization mechanism that can be seamlessly integrated into existing CL methods. Comprehensive experiments on challenging benchmarks, including ImageNet-R and Caltech-256, demonstrate a significant reduction in forgetting while improving accuracy across sequential tasks.", "date": "2025-05-24", "categories": []}
{"id": "2601.21577", "title": "Collaborative Parameter Learning: Mitigating Forgetting via Parameter-Level Gradient Analysis", "abstract": "Catastrophic forgetting during knowledge injection impairs the ability of large language models to acquire new knowledge without overwriting previously mastered knowledge. Recent studies analyze forgetting from a gradient similarity perspective and mitigate forgetting through vector projection. However, these methods primarily characterize gradient similarity at the aggregate direction level, leaving the parameter wise contributions to forgetting underexplored. In this paper, we decompose gradient similarity into parameter wise contributions and identify two types of parameters during forgetting: Conflicting Parameters, whose updates contribute to forgetting and typically account for 50 percent to 75 percent of parameters, and Collaborative Parameters, whose updates mitigate forgetting and account for 25 percent to 50 percent. Based on this analysis, we propose Collaborative Parameter Learning, CPL, a parameter wise training rule that freezes Conflicting Parameters and updates only Collaborative Parameters. Experiments comparing CPL with seven baseline methods show that CPL learns 20.2% to 48.2% more questions with negligible forgetting, while reducing peak VRAM by approximately 3 GB per billion model parameters and computation time by 16.5 percent. Extensive evaluations on parameter consumption, out of set generalization, cross prompt generalization, multimodal tasks, open ended question answering, and multilingual settings demonstrate that CPL effectively mitigates forgetting across diverse scenarios.", "date": "2026-01-29", "categories": []}
{"id": "2605.12705", "title": "Early Data Exposure Improves Robustness to Subsequent Fine-Tuning", "abstract": "How can we train models whose post-trained capabilities survive subsequent fine-tuning? Rather than focusing on downstream interventions to mitigate forgetting of upstream capabilities, we study how upstream training choices - that is, the manner in which a capability is acquired - shape how robustly that capability is retained. We investigate this question in a controlled three-stage language-model pipeline: pretraining, post-training to acquire a target capability, and downstream fine-tuning on a new objective. Across 135M and 1B models, two post-training domains, and two downstream fine-tuning tasks, we find that immediate post-training performance does not reliably predict retention after subsequent fine-tuning: training recipes that look equivalent immediately after post-training can retain the target capability very differently after subsequent fine-tuning. In particular, early exposure - mixing post-training data into pretraining - consistently improves the frontier between retained upstream performance and downstream performance. In compute-matched experiments, where the target data must be allocated between pretraining and post-training, we find that the optimum lies at neither extreme. Together with our other empirical and theoretical findings, this supports the view that post-training drives immediate specialization while early exposure improves robustness to later forgetting. Replay and dropout, typically used to mitigate forgetting as it occurs during fine-tuning, provide complementary gains to early exposure when applied during post-training. Our findings suggest that robustness to subsequent fine-tuning should be treated as a first-class objective of upstream training, addressed preventatively through choices like early exposure rather than reactively during fine-tuning itself.", "date": "2026-05-12", "categories": []}
{"id": "2605.12752", "title": "Low-Rank Adapters Initialization via Gradient Surgery for Continual Learning", "abstract": "LoRA is widely adopted for continual fine-tuning of Large Language Models due to its parameter efficiency, modularity across tasks, and compatibility with replay strategies. However, LoRA-based continual learning remains vulnerable to catastrophic forgetting, whose severity depends on how successive task gradients interact: when consecutive task gradients conflict, standard adapter initializations channel updates into subspaces that overwrite previously learned directions. We propose SLICE, a gradient-surgery-based initialization for LoRA adapters in continual learning. SLICE accumulates gradients from both the current task and a replay buffer of prior tasks, reconciles them through a projection operator, and decomposes the result via truncated SVD to initialize the adapter weights. We evaluate SLICE on the TRACE benchmark and sequences of Super-NI tasks, including a set of adversarial Super-NI sequences that we construct by mining task pairs with maximally opposing gradients. Compared to vanilla LoRA, LoRA-GA, and LoRAM, SLICE consistently achieves a better stability-plasticity trade-off, improving Average Performance, Final Performance and Forgetting metrics while preserving General Performance and In Context Performance across both standard and adversarial continual learning sequences.", "date": "2026-05-12", "categories": []}
{"id": "2605.13162", "title": "Continual Fine-Tuning of Large Language Models via Program Memory", "abstract": "Parameter-Efficient Fine-Tuning (PEFT), particularly Low-Rank Adaptation (LoRA), has become a standard approach for adapting Large Language Models (LLMs) under limited compute. However, in continual settings where models are updated sequentially with small datasets, conventional LoRA updates struggle to balance rapid adaptation and knowledge retention. Existing methods typically treat the low-rank space as a homogeneous update region, lacking mechanisms to regulate how short-term updates are consolidated over time. We propose a continual LoRA framework with \\textbf{Pro}gram memory, inspired by \\textbf{C}omplementary \\textbf{L}earning Systems in neuroscience. Our approach, dubbed \\textbf{ProCL}, organizes LoRA adapters into structured program memory slots that are dynamically retrieved through input-conditioned attention. This enables rapid and localized adaptation, encouraging similar inputs to reuse shared adapter regions while reserving unused capacity for future data. The slots are then combined with the underlying adapter, which maintains a distributed representation that gradually accumulates knowledge across tasks to balance plasticity and stability. Our method operates entirely within the LoRA parameterization and incurs no additional inference cost. Experiments on diverse benchmarks demonstrate improved retention and reduced catastrophic forgetting over other continual LoRA strategies.", "date": "2026-05-13", "categories": []}
{"id": "2605.13835", "title": "Unlocking Patch-Level Features for CLIP-Based Class-Incremental Learning", "abstract": "Class-Incremental Learning (CIL) enables models to continuously integrate new knowledge while mitigating catastrophic forgetting. Driven by the remarkable generalization of CLIP, leveraging pre-trained vision-language models has become a dominant paradigm in CIL. However, current work primarily focuses on aligning global image embeddings (i.e., [CLS] token) with their corresponding text prompts (i.e., [EOS] token). Despite their good performance, we find that they discard the rich patch-level semantic information inherent in CLIP's encoders. For instance, when recognizing a rabbit, local patches may encode its distinctive cues, such as long ears and a fluffy tail, which can provide complementary evidence for recognition. Based on the above observation, we propose SPA (Semantic-guided Patch-level Alignment) for CLIP-based CIL, which aims to awaken long-neglected local representations within CLIP. Specifically, for each class, we first construct representative and diverse visual samples and feed them to GPT-5 as visual guidance to generate class-wise semantic descriptions. These descriptions are used to guide the selection of discriminative patch-level visual features. Building upon these selected patches, we further employ optimal transport to align selected patch tokens with semantic tokens from class-wise descriptions, yielding a structured cross-modal alignment that improves recognition. Furthermore, we introduce task-specific projectors for effective adaptation to downstream incremental tasks, and sample pseudo-features from stored class-wise Gaussian statistics to calibrate old-class representations, thereby mitigating catastrophic forgetting. Extensive experiments demonstrate that SPA achieves state-of-the-art performance.", "date": "2026-05-13", "categories": []}
{"id": "2412.13050", "title": "Modality-Inconsistent Continual Learning of Multimodal Large Language Models", "abstract": "In this paper, we introduce Modality-Inconsistent Continual Learning (MICL), a new continual learning scenario for Multimodal Large Language Models (MLLMs) that involves tasks with inconsistent modalities (image, audio, or video) and varying task types (captioning or question-answering). Unlike existing vision-only or modality-incremental settings, MICL combines modality and task type shifts, both of which drive catastrophic forgetting. To address these challenges, we propose MoInCL, which employs a Pseudo Targets Generation Module to mitigate forgetting caused by task type shifts in previously seen modalities. It also incorporates Instruction-based Knowledge Distillation to preserve the model's ability to handle previously learned modalities when new ones are introduced. We benchmark MICL using a total of six tasks and conduct experiments to validate the effectiveness of our MoInCL. The experimental results highlight the superiority of MoInCL, showing significant improvements over representative and state-of-the-art continual learning baselines.", "date": "2024-12-17", "categories": []}
{"id": "2602.07892", "title": "Safety Alignment as Continual Learning: Mitigating the Alignment Tax via Orthogonal Gradient Projection", "abstract": "Safety post-training can improve the harmfulness and policy compliance of Large Language Models (LLMs), but it may also reduce general utility, a phenomenon often described as the \\emph{alignment tax}. We study this trade-off through the lens of continual learning: sequential alignment stages expose the model to shifted data distributions and objectives, and their gradients may interfere with directions that support previously acquired general capabilities. This view does not claim that all alignment degradation has a single cause; rather, it provides a useful first-order mechanism for mitigating one important source of capability regression. We propose \\textbf{O}rthogonal \\textbf{G}radient \\textbf{P}rojection for \\textbf{S}afety \\textbf{A}lignment (\\textbf{OGPSA}), a lightweight update rule that estimates a low-rank reference subspace from gradients on a small set of general-capability data and removes from each safety gradient the component lying in this subspace. The resulting update is the steepest local safety-descent direction subject to first-order preservation constraints on the reference objectives. OGPSA is compatible with standard post-training pipelines and avoids large-scale replay, although it introduces periodic reference-gradient computation. Across Supervised Fine-Tuning (SFT), Direct Preference Optimization (DPO), and sequential SFT$\\rightarrow$DPO settings, OGPSA improves the observed safety--utility trade-off over standard baselines. Under the sequential SFT$\\rightarrow$DPO pipeline, the average performance gain increases from 33.98\\% to 42.74\\% on Qwen2.5-7B-Instruct and from 19.74\\% to 32.98\\% on Llama3.1-8B-Instruct. We have open sourced our code at https://github.com/SunGL001/OGPSA.", "date": "2026-02-08", "categories": []}
{"id": "2605.07076", "title": "Self-Consolidating Language Models: Continual Knowledge Incorporation from Context", "abstract": "Large language models (LLMs) increasingly receive information as streams of passages, conversations, and long-context workflows. While longer context windows expose more evidence, they do not ensure that useful information is preserved and reused. We study continual context consolidation: writing current context into model weights while limiting interference with previously consolidated information. We propose \\textbf{S}elf-\\textbf{Co}nsolidating \\textbf{L}anguage Models (SCoL), a post-training framework in which, given current context, an LLM learns to generate textual update instructions specifying which of its own Transformer layers should be updated. Because committed updates change the model that later generates future selections, we train SCoL with meta-reinforcement learning over an evolving model state. We instantiate SCoL with supervised QA rewards on SQuAD knowledge incorporation and intrinsic likelihood-based rewards for LongBench v2 long-context consolidation. Across both settings, SCoL improves acquisition and retention over prompting, summarization, batch test-time training, and sequential finetuning baselines. Analysis of learned selection patterns shows that SCoL encourages the LLM to generate sparse update locations that align with layers of high Fisher information, suggesting that the model learns to route plasticity toward loss-sensitive regions while limiting interference. Moreover, SCoL transfers from shorter meta-training streams to longer LongBench v2 streams at evaluation, suggesting that our framework supports scalable streaming consolidation.", "date": "2026-05-08", "categories": []}
{"id": "2605.11742", "title": "Online Continual Learning with Dynamic Label Hierarchies", "abstract": "Online Continual Learning (OCL) aims to learn from endless non\\text{-}stationary data streams, yet most existing methods assume a flat label space and overlook the hierarchical organization of real\\text{-}world concepts that evolves both horizontally (sibling classes) and vertically (coarse or fine categories). To better reflect this context, we introduce a new problem setting, DHOCL (Online Continual Learning from Dynamic Hierarchies), where taxonomies evolve across granularities and each sample provides supervision at a single hierarchical level. In this setting, we find two fundamental issues: (i) partial supervision under mixed granularities provides only point-wise signals over an evolving path-wise hierarchy, which constrains plasticity and undermines cross-level semantic consistency, and (ii) the dynamically evolving hierarchies induce granularity-dependent interference, destabilizing popular replay and regularization mechanisms and thereby exacerbating catastrophic forgetting. To tackle these issues, we propose HALO (Hierarchical Adaptive Learning with Organized Prototypes), which adaptively combines complementary classification heads, regularized by organized learnable hierarchical prototypes, enabling rapid adaptation, hierarchical consistency, and structured knowledge consolidation as the taxonomy evolves. Extensive experiments on multiple benchmarks demonstrate that HALO consistently outperforms existing methods across hierarchical accuracy, mistake severity, and continual performance.", "date": "2026-05-12", "categories": []}
{"id": "2605.11804", "title": "Stop Marginalizing My Dreams: Model Inversion via Laplace Kernel for Continual Learning", "abstract": "Data-free continual learning (DFCIL) relies on model inversion to synthesize pseudo-samples and mitigate catastrophic forgetting. However, existing inversion methods are fundamentally limited by a simplifying assumption: they model feature distributions using diagonal covariance, effectively ignoring correlations that define the geometry of learned representations. As a result, synthesized samples often lack fidelity, limiting knowledge retention. In this work, we show that modeling feature dependencies is a key ingredient for effective DFCIL. We introduce REMIX, a structured covariance modeling framework that enables scalable full-covariance modeling without the prohibitive cost of dense matrix inversion and log-determinant computation. By leveraging a Laplace kernel parameterization, REMIX captures structured feature dependencies using memory that scales linearly with the feature dimensionality, while requiring only an additional logarithmic factor in computation. Modeling these correlations produces more coherent synthetic samples and consistently improves performance across standard DFCIL benchmarks. Our results demonstrate that moving beyond diagonal assumptions is essential for effective and scalable data-free continual learning. Our code is available at https://github. com/pkrukowski1/REMIX-Model-Inversion-via-Laplace-Kernel.", "date": "2026-05-12", "categories": []}
{"id": "2605.11904", "title": "Beyond Point-wise Neural Collapse: A Topology-Aware Hierarchical Classifier for Class-Incremental Learning", "abstract": "The Nearest Class Mean (NCM) classifier is widely favored in Class-Incremental Learning (CIL) for its superior resistance to catastrophic forgetting compared to Fully Connected layers. While Neural Collapse (NC) theory supports NCM's optimality by assuming features collapse into single points, non-linear feature drift and insufficient training in CIL often prevent this ideal state. Consequently, classes manifest as complex manifolds rather than collapsed points, rendering the single-point NCM suboptimal. To address this, we propose Hierarchical-Cluster SOINN (HC-SOINN), a novel classifier that captures the topological structure of these manifolds via a ``local-to-global'' representation. Furthermore, we introduce Structure-Topology Alignment via Residuals (STAR) method, which employs a fine-grained pointwise trajectory tracking mechanism to actively deform the learned topology, allowing it to adapt precisely to complex non-linear feature drift. Theoretical analysis and Procrustes distance experiments validate our framework's resilience to manifold deformations. We integrated HC-SOINN into seven state-of-the-art methods by replacing their original classifiers, achieving consistent improvements that highlight the effectiveness and robustness of our approach. Code is available at https://github.com/yhyet/HC_SOINN.", "date": "2026-05-12", "categories": []}
{"id": "2605.12306", "title": "KAN-CL: Per-Knot Importance Regularization for Continual Learning with Kolmogorov-Arnold Networks", "abstract": "Catastrophic forgetting remains the central obstacle in continual learning (CL): parameters shared across tasks interfere with one another, and existing regularization methods such as EWC and SI apply uniform penalties without awareness of which input region a parameter serves. We propose KAN-CL, a continual learning framework that exploits the compact-support spline parameterization of Kolmogorov-Arnold Networks (KANs) to perform importance-weighted anchoring at per-knot granularity. Deployed as a classification head on a convolutional backbone with standard EWC regularization on the backbone (bbEWC) KAN-CL achieves forgetting reductions of 88% and 93% over a head-only KAN baseline on Split-CIFAR-10/5T and Split-CIFAR-100/10T respectively, while matching or exceeding the accuracy of all baselines on both benchmarks. We further provide a Neural Tangent Kernel (NTK) analysis showing that KAN's spline locality induces a structural rank deficit in the cross-task NTK, yielding a forgetting bound that holds even in the feature-learning regime. These results establish that combining an architecture with natural parameter locality (KAN head) with a complementary backbone regularizer (bbEWC) yields a compositional and principled approach to catastrophic forgetting.", "date": "2026-05-12", "categories": []}
{"id": "2605.12419", "title": "ORBIT: Preserving Foundational Language Capabilities in GenRetrieval via Origin-Regulated Merging", "abstract": "Despite the rapid advancements in large language model (LLM) development, fine-tuning them for specific tasks often results in the catastrophic forgetting of their general, language-based reasoning abilities. This work investigates and addresses this challenge in the context of the Generative Retrieval (GenRetrieval) task. During GenRetrieval fine-tuning, we find this forgetting occurs rapidly and correlates with the distance between the fine-tuned and original model parameters. Given these observations, we propose ORBIT, a novel approach that actively tracks the distance between fine-tuned and initial model weights, and uses a weight averaging strategy to constrain model drift during GenRetrieval fine-tuning when this inter-model distance exceeds a maximum threshold. Our results show that ORBIT retains substantial text and retrieval performance by outperforming both common continual learning baselines and related regularization methods that also employ weight averaging.", "date": "2026-05-12", "categories": []}
{"id": "2602.07940", "title": "MePo: Meta Post-Refinement for Rehearsal-Free General Continual Learning", "abstract": "To cope with uncertain changes of the external world, intelligent systems must continually learn from complex, evolving environments and respond in real time. This ability, collectively known as general continual learning (GCL), encapsulates practical challenges such as online datastreams and blurry task boundaries. Although leveraging pretrained models (PTMs) has greatly advanced conventional continual learning (CL), these methods remain limited in reconciling the diverse and temporally mixed information along a single pass, resulting in sub-optimal GCL performance. Inspired by meta-plasticity and reconstructive memory in neuroscience, we introduce here an innovative approach named Meta Post-Refinement (MePo) for PTMs-based GCL. This approach constructs pseudo task sequences from pretraining data and develops a bi-level meta-learning paradigm to refine the pretrained backbone, which serves as a prolonged pretraining phase but greatly facilitates rapid adaptation of representation learning to downstream GCL tasks. MePo further initializes a meta covariance matrix as the reference geometry of pretrained representation space, enabling GCL to exploit second-order statistics for robust output alignment. MePo serves as a plug-in strategy that achieves significant performance gains across a variety of GCL benchmarks and pretrained checkpoints in a rehearsal-free manner (e.g., 15.10\\%, 13.36\\%, and 12.56\\% on CIFAR-100, ImageNet-R, and CUB-200 under Sup-21/1K). Our source code is available at \\href{https://github.com/SunGL001/MePo}{MePo}", "date": "2026-02-08", "categories": []}
{"id": "2602.10356", "title": "Autonomous Continual Learning for Environment Adaptation of Computer-Use Agents", "abstract": "Real-world digital environments are highly diverse and dynamic. These characteristics cause agents to frequently encounter unseen environments and distribution shifts, making continual learning in such environments essential for computer-use agents (CUAs). However, a key challenge lies in obtaining high-quality and environment-grounded training data without relying on costly human annotation. In this work, we introduce ACuRL, an Autonomous Curriculum Reinforcement Learning framework that continually adapts agents to specific environments with zero human data. The agent first explores an environment to acquire initial experiences. During subsequent iterative training, a curriculum task generator leverages these experiences together with feedback from the previous iteration to synthesize new tasks tailored for the agent's current capabilities. To provide reliable reward signals, we introduce CUAJudge, a robust automatic evaluator for CUAs that achieves 93% agreement with human judgments. Empirically, our method effectively enables both intra-environment and cross-environment continual learning, yielding 3-29% absolute performance gains on the target environments without catastrophic forgetting on others. We also show that it can mitigate performance degradation under environment changes (e.g., version updates, platform migration, and resolution shifts). Further analyses show highly sparse updates (e.g., only 20% parameters), which helps explain the effective and robust adaptation.", "date": "2026-02-10", "categories": []}
{"id": "2605.08311", "title": "Revitalizing the Beginning: Avoiding Storage Dependency for Model Merging in Continual Learning", "abstract": "Model merging provides a compelling paradigm for integrating specialized expertise into a unified multi-task model, a goal that aligns naturally with the sequential knowledge acquisition in continual learning (CL). However, the requirement for preserving diverse forms of previous knowledge conflicts with the storage limitations inherent to CL. In this paper, we systematically analyze existing model merging methods under the constraints of CL. We find that current methods prioritize global alignment, which often leads to the accumulation and amplification of task-specific errors within the continuous data stream; and the vanishing gradients at the onset of subsequent tasks frequently cause optimization to stagnate. These leave the merged model in a suboptimal state at the beginning of the next training phase. To address these challenges, we propose Trajectory Regularized Merging (TRM), a framework that reformulates the merging phase as an optimization process within an augmented trajectory subspace. Our framework integrates three synergistic objectives including task alignment, prediction consistency, and gradient responsiveness to concurrently preserve merged model's historical stability and re-activate optimization dynamics. Extensive experimental results demonstrate that our method achieves state-of-the-art performance across multiple benchmarks.", "date": "2026-05-08", "categories": []}
{"id": "2605.08839", "title": "Cross-Sample Relational Fusion: Unifying Domain Generalization and Class-Incremental Learning", "abstract": "Class-Incremental Learning (CIL) requires a learning system to learn new classes while retaining previously learned knowledge. However, in real-world scenarios such as autonomous driving, a system trained on urban roads in sunny weather may later need to operate in rural or highway environments with different traffic patterns and weather conditions. This requires the model not only to overcome catastrophic forgetting, but also to effectively handle domain shifts. In this paper, we propose CrOss-sample Relational Fusion (CORF), a unified framework to address domain shift and catastrophic forgetting simultaneously. To enhance generalizability, we perform selective refinement of training samples by leveraging spatial contribution maps to highlight semantically informative regions. Furthermore, we incorporate predictive confidence to adaptively weigh samples, thereby facilitating the learning of domain-agnostic representations. To alleviate forgetting, we propose a cascaded distillation framework that captures cross-sample relational dependencies across multiple feature hierarchies, enabling multi-grained knowledge transfer from previous tasks. CORF can be seamlessly integrated into existing CIL algorithms to enhance their generalizability, achieving competitive performance across various benchmark datasets. Code is available at https://github.com/LAMDA-CL/TMM26-CORF .", "date": "2026-05-09", "categories": []}
{"id": "2605.09044", "title": "Predicting Plasticity in Deep Continual Learning: A Theoretical Perspective", "abstract": "Deep continual learning requires models to adapt to new tasks without retraining from scratch. However, neural networks can lose their ability to adapt to new tasks after training on previous ones, a phenomenon known as loss of plasticity. There have been several explanations and diagnostics proposed for plasticity loss. Motivated by the philosophy that \"all models are wrong, but some are useful\", we ask: can existing diagnostics predict a neural network's plasticity? In this work, we take a practical view to interpret plasticity as trainability, i.e., a neural network's future optimization gain on a target task. We first take a theoretical approach, showing, by constructing a few counterexamples, that some widely adopted diagnostics of plasticity, including representation rank and neural tangent kernel rank, can fail to predict the loss of trainability in both regression and classification settings. We instead propose a novel metric, called optimization readiness, which combines gradient strength and gradient reliability. We prove that optimization readiness lower bounds one-step optimization gain under standard smoothness assumptions, providing a theoretical guarantee for its predictive power. Empirically, we show that across commonly used deep continual learning settings, such as Slowly-Changing Regression and Permuted MNIST, optimization readiness more reliably ranks checkpoints by trainability than prior diagnostics, even with substantially fewer samples.", "date": "2026-05-09", "categories": []}
{"id": "2605.09355", "title": "FLAME: Adaptive Mixture-of-Experts for Continual Multimodal Multi-Task Learning", "abstract": "Real-world model deployment across multiple domains requires multimodal models to operate under two complementary regimes: (1) multi-task pretraining, tasks are co-available at design time where related tasks could borrow representational strength from one another, (2) continual adaptation, in which new tasks emerge after deployment with previously unseen modality combinations. However, neither regime alone suffices: the pretraining task set is never exhaustive, while bypassing joint training forfeits the transfer gains and efficiency among co-trainable tasks. Sparse Mixture-of-Experts (MoE) is a natural fit for this dual requirement: sparse activation enables modular capacity expansion as new tasks arrive, while routing decouples modality-level computation from task-level composition. In this work, we propose a scalable MoE framework for multitask pretraining and continual learning across flexible modality combinations. The framework is designed to support training on multimodal tasks with diverse modality configurations by leveraging modality-specific routers that process tokens from each modality across tasks. Furthermore, it enables continual learning over sequential multimodal tasks within a fixed-capacity MoE by compressing accumulated expert knowledge into low-rank memory subspaces, while expanding only the lightweight routers. We validate the effectiveness of our method on multiple healthcare multimodal benchmarks. It demonstrates competitive multitask pretraining performance while alleviating catastrophic forgetting and improving parameter efficiency.", "date": "2026-05-10", "categories": []}
{"id": "2605.09608", "title": "Geometry Conflict: Explaining and Controlling Forgetting in LLM Continual Post-Training", "abstract": "Continual post-training aims to extend large language models (LLMs) with new knowledge, skills, and behaviors, yet it remains unclear when sequential updates enable capability transfer and when they cause catastrophic forgetting. Existing methods mitigate forgetting through sequential fine-tuning, replay, regularization, or model merging, but offer limited criteria for determining when incorporating new updates is beneficial or harmful. In this work, we study LLM continual post-training through three questions: What drives forgetting? When do sequentially acquired capabilities transfer or interfere? How can compatibility be used to control update integration? We address these questions through task geometry: we represent each post-training task by its parameter update and study the covariance geometry induced by the update. Our central finding is that: forgetting can be considered as a state-relative update-integration failure, it arises when the covariance geometries induced by tasks misalign with the geometry of the evolving model state. Sequential updates transfer when they remain compatible with the model state shaped by previous updates, and interfere when state-relative geometry conflict becomes high. Motivated by this finding, we propose Geometry-Conflict Wasserstein Merging (GCWM), a data-free update-integration method that constructs a shared Wasserstein metric via Gaussian Wasserstein barycenters and uses geometry conflict to gate geometry-aware correction. Across Qwen3 0.6B--14B on domain-continual and capability-continual settings, GCWM consistently outperforms data-free baselines, improving retention and final performance without replay data. These results identify geometry conflict as both an explanatory signal for forgetting and a practical control signal for LLM continual post-training.", "date": "2026-05-10", "categories": []}
{"id": "2605.09640", "title": "Overcoming Catastrophic Forgetting in Visual Continual Learning with Reinforcement Fine-Tuning", "abstract": "Recent studies suggest that Reinforcement Fine-Tuning (RFT) is inherently more resilient to catastrophic forgetting than Supervised Fine-Tuning (SFT). However, whether RFT (e.g., GRPO) can effectively overcome forgetting in challenging visual continual learning settings, such as class-incremental learning (CIL) and domain-incremental learning (DIL), remains an open problem. Through a pilot study, we confirm that while RFT consistently outperforms SFT, it still suffers from non-negligible forgetting. We empirically trace this bottleneck to Trajectory-level Drift Agnosticism: among candidate rollouts achieving identical task rewards, the KL divergence from the preceding-task policy varies substantially, which strongly correlates with catastrophic forgetting across sequential tasks. Motivated by this insight, we propose Retention-aware Policy Optimization (RaPO), a simple yet effective RFT method that explicitly mitigates forgetting through trajectory-level reward shaping. Specifically, RaPO comprises two core components: (1) Retention Reward that converts trajectory-level distribution drift into a continuous reward signal, preferentially reinforcing knowledge-preserving rollouts within each group; (2) Cross-Task Advantage Normalization (CTAN), which maintains a persistent exponential moving average of reward statistics across task boundaries to stabilize the optimization progress during continual learning. Leveraging the free-form textual generalization of MLLMs, we comprehensively evaluate RaPO across five visual continual learning settings. Extensive experiments demonstrate that RaPO achieves leading performance, substantially reducing catastrophic forgetting while preserving strong plasticity. To the best of our knowledge, this work represents the first systematic exploration of RFT in visual continual learning, offering insights that we hope will inspire future research.", "date": "2026-05-10", "categories": []}
{"id": "2605.09734", "title": "Trajectory Supervision for Continual Tool-Use Learning in LLMs", "abstract": "Most language-model training data shows final artifacts, not the process that produced them. We study a tractable version of this question in tool use: when a model learns a stream of new API domains, does keeping tool-use trajectories help compared with stripping the intermediate API trace? We fine-tune Llama 3.1 8B Instruct with QLoRA on API-Bank using four sequential domain blocks. Condition A strips previous API request/response lines from the prompt and trains the model to predict the next API call. Condition B keeps the trajectory context. In a single-seed pilot, full held-out generation evaluation shows that Condition B reaches 56.9\\% final exact full-call accuracy compared with 39.2\\% for Condition A. B also improves final API-name accuracy by 7.7 points. However, B uses 25.1\\% more training tokens, the run uses one seed, and the task is next-call prediction rather than full dialogue success.", "date": "2026-05-10", "categories": []}
{"id": "2605.09862", "title": "UFO: A Unified Flow-Oriented Framework for Robust Continual Graph Learning", "abstract": "Graph learning research has increasingly shifted toward continual graph learning (CGL), which better reflects real-world scenarios where graphs evolve over time. However, existing CGL methods largely assume clean supervision and overlook a critical challenge: the newly arriving portions of the graph are often noisy, due to annotation errors or adversarial corruption. This mismatch limits their applicability in practice. In this work, we study robust continual graph learning, where models must simultaneously handle catastrophic forgetting and noisy supervision in evolving graph data. We show that label noise introduces a new failure mode, catastrophic remembering, where models persistently reinforce corrupted knowledge across tasks. To address these challenges, we propose a Unified Flow-Oriented framework (UFO). First, UFO models conditional feature distributions via flow-based generative modeling and produces replay representations, mitigating forgetting without storing historical data. Second, UFO estimates instance-level reliability scores to distinguish clean from noisy nodes, reducing the impact of corrupted supervision and alleviating catastrophic remembering. Extensive experiments on four benchmark graph datasets under varying noise ratios demonstrate that UFO consistently outperforms existing methods in both accuracy and forgetting metrics. Code is available at: https://anonymous.4open.science/r/UFO.", "date": "2026-05-11", "categories": []}
{"id": "2605.10510", "title": "CMKL: Modality-Aware Continual Learning for Evolving Biomedical Knowledge Graphs", "abstract": "Biomedical knowledge graphs are increasingly large, dynamic, and multimodal, driven by rapid advances in biotechnology such as high-throughput sequencing. Machine learning models can infer previously unobserved biomedical relationships and characterize biomedical entities in these graphs, but existing knowledge graph embedding methods and their continual learning extensions either assume static graph structure or fail to exploit multimodal information under evolving data distributions. They also apply uniform regularization across all model parameters, ignoring that different modalities may exhibit distinct forgetting dynamics as the graph evolves. We propose the Continual Multimodal Knowledge Graph Learner (CMKL), a CL framework for biomedical KGs that natively encodes structure, text, and molecules, fuses them through a Mixture-of-Experts (MoE) router, and protects previously learned knowledge with standard EWC regularization and a K-means-diverse multimodal replay buffer. We evaluate CMKL on a 129K-entity biomedical continual benchmark with 10 tasks. On continual biomedical entity classification, CMKL reaches AP 0.591 versus 0.370 for the strongest structural baseline, a 60% gain that is driven by access to multimodal features and preserved across the sequence with near-zero forgetting (AF 0.008). On continual relationship prediction, CMKL reaches AP $0.062$, matching Naive Sequential and EWC (0.058) within seed noise and outperforming Joint Training (0.047, p=0.045) and LKGE (0.039). A frozen-text ablation reaches AP 0.136, more than double any jointly trained model, yet that signal is unreachable by margin-ranking gradients: the greedy-modality asymmetry lives at the representation level, not the fusion level, and MoE routing manages it by suppressing the unreachable modality without forcing it through a learned bottleneck. Code: github.com/yradwan147/cmkl-neurips2026", "date": "2026-05-11", "categories": []}
{"id": "2605.10529", "title": "PrimeKG-CL: A Continual Graph Learning Benchmark on Evolving Biomedical Knowledge Graphs", "abstract": "Biomedical knowledge graphs underwrite drug repurposing and clinical decision support, yet the upstream ontologies they depend on update on independent cycles that add millions of edges and deprecate hundreds of thousands more between releases. Yet existing continual graph learning has been studied almost exclusively on synthetic random splits of static, generic KGs, a regime that cannot reproduce the asynchronous, structured evolution real biomedical KGs undergo. To this end, we introduce PrimeKG-CL, a CGL benchmark built from nine authoritative biomedical databases (129K+ nodes, 8.1M+ edges, 10 node types, 30 relation types) with two genuine temporal snapshots (June 2021, July 2023; 5.83M edges added, 889K removed, 7.21M persistent), 10 entity-type-grouped tasks, multimodal node features, and a per-task persistent/added/removed test stratification. On three tasks (biomedical relationship prediction, entity classification, KGQA), we evaluate six CL strategies across four KGE decoders, plus LKGE, an LLM-RAG agent, and CMKL. We find that decoder choice and continual learning strategy interact strongly: no single strategy performs best across all decoders, and mismatched combinations can significantly degrade performance. Moreover, only DistMult exhibits a clear separation between persistent and deprecated knowledge, indicating that standard metrics conflate retention of still-valid facts with failure to forget outdated ones; this effect is absent under RotatE. In addition, multimodal features improve entity-level tasks by up to 60%, and a recent CKGE framework (IncDE) failed to scale to our 5.67M-triple base task across five attempts up to 350GB RAM. Data, pipeline, baselines, and the stratified split are released openly. Dataset:huggingface.co/datasets/yradwan147/PrimeKGCL|Code:github.com/yradwan147/primekg-cl-neurips2026", "date": "2026-05-11", "categories": []}
{"id": "2605.10640", "title": "Towards Understanding Continual Factual Knowledge Acquisition of Language Models: From Theory to Algorithm", "abstract": "Continual Pre-Training (CPT) is essential for enabling Language Models (LMs) to integrate new knowledge without erasing old. While classical CPT techniques like data replay have become the standard paradigm, the mechanisms underlying how LMs acquire and retain facts over time, termed as continual Factual Knowledge Acquisition (cFKA), remain unclear. In this work, we present a theoretical framework that characterizes the training dynamics of cFKA using a single-layer Transformer, offering a unified explanation for the behavior of representative CPT methods. Our analysis reveals that regularization-based methods merely adjust the convergence rate of parameters without altering the inherent forgetting tendency, whereas data replay methods succeed in shifting convergence dynamics and stabilizing pretrained knowledge. Building on these insights, we propose a novel generative data replay approach, called \\textbf{S}electing \\textbf{T}okens via attenti\\textbf{O}n \\textbf{C}ontribution~(STOC), which identifies influential factual snippets to guide replay data generation. Extensive experiments on both synthetic and real-world datasets validate our findings and demonstrate that STOC effectively enhances cFKA by mitigating catastrophic forgetting.", "date": "2026-05-11", "categories": []}
{"id": "2605.10658", "title": "Why Zeroth-Order Adaptation May Forget Less: A Randomized Shaping Theory", "abstract": "Continual learning requires new-task adaptation without damaging previously acquired capabilities. Recent forward-pass and zeroth-order (ZO) results show that low-query adaptation may retain better than first-order (FO) descent, but the usual view of ZO as noisy FO estimation does not explain why. We give a local randomized gradient-shaping analysis: finite differences expose a raw shape that is mean-aligned with FO, while the norm-matched comparator fixes the expected squared adaptation norm. Under this controlled comparison, forgetting depends on how the adaptation shape exposes retention curvature. For norm-matched ZO, the expected shaped retention curvature obeys an exact identity that preserves the isotropic retention floor while contracting only the anisotropic component. Projecting this identity onto the incoming gradient yields the observable FO--ZO quadratic forgetting gap: ZO improves mean forgetting precisely when the FO direction has above-average retention curvature, by a query-dependent fraction of that curvature excess. A practical finite-query accounting separates the mean mechanism from one-batch sampling and smoothing perturbations. As an algorithmic transfer, RISE applies the calibrated ZO shape to exact FO gradients inside parameter blocks. Its target is a stability--plasticity tradeoff: randomized shaping may reduce the retention exposure paid by FO, exact gradients remove finite-smoothing bias from finite-difference ZO, and blockwise sampling supplies many local shaping directions after one gradient computation. The blockwise analysis separates mean-step damage from centered random exposure, showing how block-diagonal curvature, cross-block coupling, and local shaping diagnostics specify where this exact-gradient transfer is most likely to be visible.", "date": "2026-05-11", "categories": []}
{"id": "2605.10765", "title": "Dynamic Cross-Modal Prompt Generation for Multimodal Continual Instruction Tuning", "abstract": "Multimodal Large Language Models (MLLMs) achieve strong performance through instruction tuning, yet real-world deployment often requires continual capability expansion across sequential tasks. In such scenarios, Multimodal Continual Instruction Tuning (MCIT) aims to acquire new capabilities while limiting catastrophic forgetting. Existing methods mainly follow a module-composition paradigm: they maintain task-level prompts or LoRA experts and dynamically route or aggregate a subset of them at inference. However, samples within the same task can still differ substantially in visual scenes, question intents, and reasoning demands. This motivates instance-level adaptation to individual query-image pairs rather than only selecting or combining task-level modules. To this end, we propose DRAPE (Dynamic Cross-Modal Prompt Generation), a prompt-learning framework that synthesizes continuous instance-specific soft prompts for MCIT. Instead of selecting prompts from a fixed pool, DRAPE derives prompt queries from the textual instruction and cross-attends to visual patch features, producing query-image conditioned prompts that are prepended to the frozen LLM. To mitigate forgetting during sequential updates, DRAPE applies null-space gradient projection to the shared projector and uses CLIP-based prototype routing for task-label-free generator selection at inference. Extensive experiments on MCIT benchmarks show that DRAPE achieves state-of-the-art performance among representative prompt-based and LoRA-based continual-learning baselines.", "date": "2026-05-11", "categories": []}
{"id": "2503.04638", "title": "No Forgetting Learning: Buffer-free Continual Learning Classification", "abstract": "Most Continual Learning (CL) methods maintain performance on earlier tasks by storing exemplars in a replay buffer, introducing memory overhead that scales with the number of tasks and raising privacy concerns in regulated domains. We propose No Forgetting Learning (NFL), a buffer-free framework for class- and task-incremental learning that instead exploits the inherent redundancy of overparameterized networks.\n  NFL decomposes the network into a shared backbone and task-specific heads, then applies a stepwise freezing protocol: new capabilities are first isolated, shared representations are adapted under knowledge distillation, and all components are jointly refined with dual soft-target anchoring. NFL+ augments this pipeline with an under-complete auto-encoder that preserves informative features from previous tasks and corrects the prediction bias caused by class imbalance. NFL+LoRA further extends the framework to pre-trained Vision Transformers by confining updates to a low-rank subspace with Fisher-weighted regularization, maintaining constant backbone memory cost regardless of the number of tasks.\n  On CIFAR-100, Tiny-ImageNet, and ImageNet-1000 across up to 50 incremental tasks, NFL+ outperforms all buffer-free baselines and matches memory-based methods while requiring only 2.53\\% of their model size. We also propose a Plasticity--Stability score for more balanced trade-off evaluation.", "date": "2025-03-06", "categories": []}
{"id": "2511.12090", "title": "Teaching Prompts to Coordinate: Hierarchical Layer-Grouped Prompt Tuning for Continual Learning", "abstract": "Prompt-based continual learning methods fine-tune only a small set of additional learnable parameters while keeping the pre-trained model's parameters frozen. It enables efficient adaptation to new tasks while mitigating the risk of catastrophic forgetting. These methods typically attach one independent task-specific prompt to each layer of pre-trained models to locally modulate its features, ensuring that the layer's representation aligns with the requirements of the new task. However, although introducing learnable prompts independently at each layer provides high flexibility for adapting to new tasks, this overly flexible tuning could make certain layers susceptible to unnecessary updates. As all prompts till the current task are added together as a final prompt for all seen tasks, the model may easily overwrite feature representations essential to previous tasks, which increases the risk of catastrophic forgetting. To address this issue, we propose a novel hierarchical layer-grouped prompt tuning method for continual learning. It improves model stability in two ways: (i) Layers in the same group share roughly the same prompts, which are adjusted by position encoding. This helps preserve the intrinsic feature relationships and propagation pathways of the pre-trained model within each group. (ii) It utilizes a single task-specific root prompt to learn to generate sub-prompts for each layer group. In this way, all sub-prompts are conditioned on the same root prompt, enhancing their synergy and reducing independence. Extensive experiments across four benchmarks demonstrate that our method achieves favorable performance compared with several state-of-the-art methods.", "date": "2025-11-15", "categories": []}
{"id": "2511.18085", "title": "Continually Evolving Skill Knowledge in Vision Language Action Model", "abstract": "Vision-language-action (VLA) models show promising knowledge accumulation ability from pretraining, yet continual learning in VLA remains challenging, especially for efficient adaptation. Existing continual imitation learning (CIL) methods often rely on additional parameters or external modules, limiting scalability for large VLA models. We propose Stellar VLA, a knowledge-driven CIL framework without increasing network parameters. Two progressively extended variants are designed: T-Stellar for flat task-centric modeling and TS-Stellar for hierarchical task-skill structure. Stellar VLA enables self-evolving knowledge learning by jointly optimizing task representations and a learned knowledge space. We propose a knowledge-guided expert routing mechanism conditioned on knowledge relation and Top-K semantic embeddings, enabling task specialization without increasing model size. Experiments on the LIBERO benchmark show that Stellar VLAs achieve strong performance among both VLA and CIL baselines, using only 1 % data replay. Real-world evaluation on a dual-arm platform with distinct embodiment and scene configurations validates effective knowledge transfer. TS-Stellar excels in hierarchical manipulation, and visualizations reveal robust knowledge retention and task discovery. Project Website: https://stellarvla.github.io/", "date": "2025-11-22", "categories": []}
{"id": "2602.03473", "title": "Scaling Continual Learning to 300+ Tasks with Bi-Level Routing Mixture-of-Experts", "abstract": "Continual learning, especially class-incremental learning (CIL), on the basis of a pre-trained model (PTM) has garnered substantial research interest in recent years. However, how to effectively learn both discriminative and comprehensive feature representations while maintaining stability and plasticity over very long task sequences remains an open problem. We propose CaRE, a scalable {C}ontinual Le{a}rner with efficient Bi-Level {R}outing Mixture-of-{E}xperts (BR-MoE). The core idea of BR-MoE is a bi-level routing mechanism: a router selection stage that dynamically activates relevant task-specific routers, followed by an expert routing phase that dynamically activates and aggregates experts, aiming to inject discriminative and comprehensive representations into every intermediate network layer. On the other hand, we introduce a challenging dataset, OmniBenchmark-1K, for CIL performance evaluation on very long task sequences with hundreds of tasks. Extensive experiments show that CaRE demonstrates leading performance across a variety of datasets and task settings, including commonly used CIL datasets with classical CIL settings (e.g., 5-20 tasks). To the best of our knowledge, CaRE is the first continual learner that scales to very long task sequences (ranging from 100 to over 300 non-overlapping tasks), while outperforming all baselines by a large margin on such task sequences. We hope that this work will inspire further research into continual learning over extremely long task sequences. Code and dataset are publicly released at https://github.com/LMMMEng/CaRE.", "date": "2026-02-03", "categories": []}
{"id": "2605.05732", "title": "CRAFT: Forgetting-Aware Intervention-Based Adaptation for Continual Learning", "abstract": "Large language models (LLMs) can acquire new capabilities through fine-tuning, but continual adaptation often leads to catastrophic forgetting. We propose CRAFT, a continual learning framework that avoids updating model weights by instead learning low-rank interventions on hidden representations. CRAFT proceeds in three stages: it first routes each task to a group of similar tasks based on output-distribution divergence; it then fine-tunes the model using a Kullback-Leibler (KL) divergence against the group's prior state, which directly controls forgetting and determines convergence; finally, it merges interventions for the updated task into the shared representation using the same KL signal. This design unifies routing, regularization, and merging through a single KL-based objective. CRAFT improves overall performance and reduces forgetting compared to strong LoRA-based approaches across multiple benchmarks and model scales, while remaining robust to task ordering. These results suggest that controlling adaptation in representation space, guided by output-space divergence, provides a scalable and principled approach to continual learning in LLMs.", "date": "2026-05-07", "categories": []}
{"id": "2605.06702", "title": "CASCADE: Case-Based Continual Adaptation for Large Language Models During Deployment", "abstract": "Large language models (LLMs) have become a central foundation of modern artificial intelligence, yet their lifecycle remains constrained by a rigid separation between training and deployment, after which learning effectively ceases. This limitation contrasts with natural intelligence, which continually adapts through interaction with its environment. In this paper, we formalise deployment-time learning (DTL) as the third stage in the LLM lifecycle that enables LLM agents to improve from experience during deployment without modifying model parameters. We present CASCADE (CASe-based Continual Adaptation during DEployment), a general and principled framework that equips LLM agents with an explicit, evolving episodic memory. CASCADE formulates experience reuse as a contextual bandit problem, enabling principled exploration-exploitation trade-offs and establishing no-regret guarantees over long-term interactions. This design allows agents to accumulate, select, and refine task-relevant cases, transforming past experience into actionable knowledge. Across 16 diverse tasks spanning medical diagnosis, legal analysis, code generation, web search, tool use, and embodied interaction, CASCADE improves macro-averaged success rate by 20.9% over zero-shot prompting while consistently outperforming gradient-based and memory-based baselines. By reframing deployment as an adaptive learning process, this work establishes a foundation for continually improving AI systems.", "date": "2026-05-05", "categories": []}
{"id": "2605.06834", "title": "Attribution-Based Neuron Utility for Plasticity Restoration in Deep Networks", "abstract": "Continual learning research attempts to conserve two fundamental capabilities: new knowledge acquisition and the preservation of previously acquired knowledge. While knowledge in this case can be measured through performance over an implicit or explicit task space, model plasticity generally concerns adaptability as data distributions evolve. Though much of the literature has focused on catastrophic forgetting, deep networks can also suffer from loss of plasticity, becoming progressively harder to update under continued training. Recent research has identified multiple mechanisms underlying this phenomenon, including neuron saturation, parameter norm growth, and loss of useful curvature directions. Adaptive reset-based interventions, which selectively reinitialize low-utility network parameters, have emerged as practical solutions to restore trainability. Existing utility measures used to guide resets, such as activation magnitude, contribution utility, or gradient-based activity, rely on proxy signals that can become misaligned with the intervention they are meant to guide. In this paper, we introduce gradient times difference from reference (GXD), a theoretically motivated utility measure based on reference-based gradient attribution that estimates the first-order functional cost of replacing a unit. Our results show that utility measures aligned with the functional cost of the reset can make interventions more reliable in settings where existing reset criteria degrade. GXD reframes adaptive resetting as an intervention cost estimation problem, providing a practical path toward more robust continual learning systems.", "date": "2026-05-07", "categories": []}
{"id": "2605.07420", "title": "SR$^2$-LoRA: Self-Rectifying Inter-layer Relations in Low-Rank Adaptation for Class-Incremental Learning", "abstract": "Pre-trained models with parameter-efficient fine-tuning (PEFT) have demonstrated promising potential for class-incremental learning (CIL), yet catastrophic forgetting still persists when adapting models to new tasks. In this paper, we present a novel perspective on catastrophic forgetting through the analysis of inter-layer relation drift, i.e., the progressive disruption of relationships among layer-wise representations during the learning of new tasks. We theoretically show that the increase of such drift reduces the classification margins of previously learned tasks, thereby degrading overall model performance. To address this issue, we propose \\underline{S}elf-\\underline{R}ectifying inter-layer \\underline{R}elation Low-Rank Adaptation~(SR$^2$-LoRA), a simple yet effective method that mitigates catastrophic forgetting by constraining inter-layer relation drift. Specifically, SR$^2$-LoRA constructs the relation matrices induced by the previous and current models on current-task samples, and aligns the corresponding singular values. We further theoretically show that this alignment exhibits greater robustness to estimation perturbations than direct entry-wise alignment. Extensive experiments on standard CIL benchmarks demonstrate that SR$^2$-LoRA effectively mitigates catastrophic forgetting, with its advantages becoming more pronounced as the number of tasks increases. Code is available in the \\href{https://github.com/FqWan24/SR-2-LoRA}{repository}.", "date": "2026-05-08", "categories": []}
{"id": "2605.07494", "title": "DIMoE-Adapters: Dynamic Expert Evolution for Continual Learning in Vision-Language Models", "abstract": "Continual learning enables vision-language models to accumulate knowledge and adapt to evolving tasks without retraining from scratch. However, in multi-domain task-incremental learning, large domain shifts intensify the stability-plasticity dilemma. Most existing methods rely on fixed architectures with statically allocated parameters, which limits adaptation to new domains and aggravates catastrophic forgetting. To address these challenges, we propose DIMoE-Adapters, a Dynamic Incremental Mixture-of-Experts Adapters framework that introduces a dynamic expert evolution paradigm to balance stability and plasticity. This paradigm is implemented through two collaborative components: Self-Calibrated Expert Evolution (SCEE) and Prototype-Guided Expert Selection (PGES). SCEE constructs and evolves a sparse expert pool through expert optimization dynamics, improving plasticity while reducing redundant capacity. PGES controls expert utilization based on the pool shaped by SCEE, improving stability across both previously encountered and unseen tasks. Extensive experiments show that DIMoE-Adapters outperforms previous state-of-the-art methods across various settings.", "date": "2026-05-08", "categories": []}
{"id": "2605.07512", "title": "Hierarchical Dual-Subspace Decoupling for Continual Learning in Vision-Language Models", "abstract": "Class-incremental learning aims to continuously acquire new knowledge while preserving previously learned information, thereby mitigating catastrophic forgetting. Existing methods primarily restrict parameter updates but often overlook their structural properties in high-dimensional spaces. From a subspace perspective, updates induced by different tasks tend to lie in multiple overlapping low-rank subspaces, leading to cross-task subspace interference and severe forgetting. To address this issue, we propose HDSD, a Hierarchical Dual-Subspace Decoupling framework for continual learning in vision-language models. Specifically, we introduce a lightweight Feature Modulation Module (FMM) that explicitly decomposes the parameter space into general and task-specific subspaces. Building on this design, we develop two complementary components. First, a General Fusion Module (GFM) evaluates relative parameter changes across tasks and uses an adaptive threshold to capture stable and transferable knowledge. Second, a Hierarchical Learning Module (HLM) performs structured parameter decomposition via Singular Value Decomposition (SVD) and uses a scaling mechanism to constrain updates within distinct subspace scales. Together, these designs reduce subspace interference and parameter drift. Extensive experiments on conventional benchmarks show that HDSD achieves state-of-the-art results.", "date": "2026-05-08", "categories": []}
{"id": "2602.13670", "title": "Advancing Analytic Class-Incremental Learning through Vision-Language Calibration", "abstract": "Class-incremental learning (CIL) with pre-trained models (PTMs) faces a critical trade-off between efficient adaptation and long-term stability. While analytic learning enables rapid, recursive closed-form updates, its efficacy is often compromised by accumulated errors and feature incompatibility. In this paper, we first conduct a systematic study to dissect the failure modes of PTM-based analytic CIL, identifying representation rigidity as the primary bottleneck. Motivated by this insight, we propose VILA, a novel dual-branch framework that advances analytic CIL via a two-level vision-language calibration strategy. Specifically, we coherently fuse plastic, task-adapted features with a frozen, universal visual anchor at the feature level through geometric calibration, and leverage cross-modal semantic priors at the decision level to rectify prediction bias. This confluence maintains analytic-learning's extreme efficiency while overcoming its inherent brittleness. Extensive experiments across eight benchmarks demonstrate that VILA consistently yields superior performance, particularly in fine-grained and long-sequence scenarios. Our framework harmonizes high-fidelity prediction with the simplicity of analytic learning. Our code is available at https://github.com/byzhaoAI/VILA.", "date": "2026-02-14", "categories": []}
{"id": "2605.05285", "title": "Attribution-Guided Continual Learning for Large Language Models", "abstract": "Large language models (LLMs) often suffer from catastrophic forgetting in continual learning: after learning new tasks sequentially, they perform worse on earlier tasks. Existing methods mitigate catastrophic forgetting by data replay, parameter freezing, or regularization. However, these methods lack semantic awareness of internal knowledge distribution in LLMs. As a result, they cannot distinguish parameters that should be preserved or updated. We propose an attribution-guided continual fine-tuning framework for LLMs. Our method estimates task-specific, element-wise parameter importance in each Transformer layer and uses these scores to modulate gradients. Parameters important to previous tasks receive smaller updates, while less relevant ones remain plastic for learning new tasks. Experiments on continual learning benchmarks show that our method consistently outperforms baselines, achieving better retention of old tasks while maintaining competitive performance on new tasks.", "date": "2026-05-06", "categories": []}
{"id": "2605.05495", "title": "Shortcut Solutions Learned by Transformers Impair Continual Compositional Reasoning", "abstract": "Identifying and exploiting common features across domains is at the heart of the human ability to make analogies, and is believed to be crucial for the ability to continually learn. To do this successfully, general and flexible computational strategies must be developed. While the extent to which Transformer neural network models can perform compositional reasoning has been the subject of intensive recent investigation, little work has been done to systematically understand how well these models can leverage their representations to learn new, related experiences. To address this gap, we expand the previously developed Learning Equality and Group Operations (LEGO) framework to a continual learning (CL) setting (\"continual LEGO\"). Using this continual LEGO experimental paradigm, we study the capability of feedforward and recurrent Transformer models to perform CL. We find that BERT, a canonical feedforward Transformer model, learns shortcut solutions that limits its ability to generalize and prevents strong forward transfer to new experiences. In contrast, we find evidence supporting the hypothesis that ALBERT, a recurrent version of BERT, learns a For loop-esque solution, which leads to better CL performance. When applying BERT and ALBERT models to a CL setting that requires composition across experiences, we find that both model families fail. Our investigation suggests that ALBERT models can have their performance drop rescued by use of training strategies that combine data across experiences, but this is not true for BERT models, where a detrimental shortcut solution becomes entrenched with initial training. Our results demonstrate that the recurrent ALBERT model may have an inductive bias better suited for CL and motivate future investigation of the interplay between Transformer architecture and computational solutions that emerge in modern models and tasks.", "date": "2026-05-06", "categories": []}
{"id": "2605.05738", "title": "CoMemNet: Contrastive Sampling with Memory Replay Network for Continual Traffic Prediction", "abstract": "In recent years, the integration of non-topological space modeling with temporal learning methods has emerged as an effective approach for capturing spatio-temporal information in non-Euclidean graphs. However, most existing methods rely on static underlying graph structures, which are inadequate for capturing the continuously expanding and evolving patterns in streaming traffic networks. To address this challenge, we propose a simple yet efficient dual-branch continual learning framework for traffic prediction, named CoMemNet. The fast-converging Online branch undertakes the primary prediction tasks, while the momentum-updated Target branch extracts historical information using Wasserstein Distance features to create a Dynamic Contrastive Sampler (DC Sampler). This sampler selects a node set with significant dynamic network feature changes for training, effectively mitigating the issue of catastrophic forgetting. Additionally, the backbone incorporates a lightweight Node-Adaptive Temporal Memory Buffer (TMRB-N) to consolidate old knowledge through memory replay and address the risk of memory explosion. Finally, we provide two newly curated open-source datasets. Experimental results demonstrate that CoMemNet achieves state-of-the-art (SOTA) performance across all three large-scale real-world datasets. The code is available at: https://github.com/meiwu5/CoMemNet.", "date": "2026-05-07", "categories": []}
{"id": "2605.06447", "title": "Scene-Adaptive Continual Learning for CSI-based Human Activity Recognition with Mixture of Experts", "abstract": "Channel state information (CSI)-based human activity recognition (HAR) is vulnerable to performance degradation under domain shifts across varying physical environments. Continual learning (CL) offers a principled way to learn new domains sequentially while preserving past knowledge, but existing CL solutions for CSI-based HAR scale poorly with accumulating domains, rely on a large replay buffer, or incur linearly growing inference cost. In this letter, we propose Scene-Adaptive Mixture of Experts with Clustered Specialists (SAMoE-C), which formulates cross-domain CSI-based HAR as a mixture-of-experts system that enables scene-specific adaptation, via an attention-based semantic router that activates only selected experts for each input. Moreover, we develop a novel training protocol, which requires only a tiny replay buffer for stabilizing domain discrimination of the router. Experimental results on a four-scene CSI dataset demonstrate that SAMoE-C approaches the state-of-the-art accuracy, while maintaining a significantly lower inference cost. By jointly combining modular experts, selective activation with router and a lightweight training protocol, SAMoE-C enables scalable cross-domain CSI-based HAR deployment with low training overhead and high computational efficiency in real-world settings.", "date": "2026-05-07", "categories": []}
{"id": "2605.06477", "title": "GeoStack: A Framework for Quasi-Abelian Knowledge Composition in VLMs", "abstract": "We address the challenge of knowledge composition in Vision-Language Models (VLMs), where accumulating expertise across multiple domains or tasks typically leads to catastrophic forgetting. We introduce GeoStack (Geometric Stacking), a modular framework that allows independently trained domain experts to be composed into a unified model. By imposing geometric and structural constraints on the adapter manifold, GeoStack ensures the foundational knowledge of the base model is preserved. Furthermore, we mathematically demonstrate a weight-folding property that achieves constant-time inference complexity ($O(1)$), regardless of the number of integrated experts. Experimental results across multi-domain adaptation and class-incremental learning show that GeoStack provides an efficient mechanism for long-term knowledge composition while significantly mitigating catastrophic forgetting. Code is available at https://github.com/QuantitativeImagingLaboratory/GeoStack.", "date": "2026-05-07", "categories": []}
{"id": "2511.01553", "title": "Online Continual Learning on Intel Loihi 2 via a Co-designed Spiking Neural Network", "abstract": "AI systems on edge devices require online continual learning -- adapting to non-stationary streams and unfamiliar classes without catastrophic forgetting -- under strict power constraints. We present CLP-SNN, a spiking neural network with a self-normalizing local learning rule and a spike-driven neural state machine for autonomous on-chip learning, implemented on Intel's Loihi 2 neuromorphic processor. On OpenLORIS few-shot experiments, CLP-SNN matches replay-based accuracy rehearsal-free. On Loihi 2, CLP-SNN achieves 113x lower latency (0.33 ms vs. 37.3 ms) and 6,600x lower energy (0.05 mJ vs. 333 mJ) than the strongest edge-GPU baseline. This gain decomposes into algorithmic efficiency (~14.5x latency, ~22.6x energy on the same GPU) and neuromorphic hardware co-design (~7.8x latency, ~295x energy) exploiting event-driven learning and sparse graded-spike communication. We show that co-designed brain-inspired algorithms and neuromorphic hardware can break traditional accuracy-efficiency trade-offs in edge AI.", "date": "2025-11-03", "categories": []}
{"id": "2605.04103", "title": "HERCULES: Hardware-Efficient, Robust, Continual Learning Neural Architecture Search", "abstract": "Neural Architecture Search (NAS) has emerged as a powerful framework for automatically discovering neural architectures that balance accuracy and efficiency. However, as AI transitions from static benchmarks to real-world deployment, the traditional focus on hardware-aware efficiency is no longer sufficient. We observe that modern NAS methods, especially those that target edge AI, are evolving to address a triple objective: Efficiency, Robustness, and Continual Learning. While efficiency ensures feasibility in resource-constrained environments, robustness guarantees reliability under environmental variabilities, and continual learning enables adaptation to sequential tasks without catastrophic forgetting. We propose a taxonomy of NAS approaches through this triple lens, distinguishing between methods targeting resource optimization, environmental resilience, and architectural plasticity. This unified perspective reveals that these axes, though often studied in isolation, are mutually reinforcing. Building on this taxonomy, we map the current landscape of these NAS methods into a new framework called Hardware-Efficient, Robust, and ContinUal LEarning Search (HERCULES). We define the desiderata, the twelve labours of HERCULES, addressing the non-trivial challenge of balancing an adequate search-space exploration with the immense computational costs of a multi-objective NAS, accounting for these crucial objectives of current AI systems. By identifying critical gaps in existing research, this survey outlines a roadmap toward integrated algorithmic, architectural, and hardware-software co-design for truly deployable, lifelong-learning AI systems.", "date": "2026-05-03", "categories": []}
{"id": "2605.04308", "title": "Memory as a Markov Matrix: Sample Efficient Knowledge Expansion via Token-to-Dictionary Mapping", "abstract": "Continual incorporation of new knowledge is essential for the long-term evolution of large language models (LLMs). Existing approaches typically rely on parameter-update algorithms to mitigate catastrophic forgetting, yet they suffer from fundamental limitations: 1) forgetting is unavoidable as the amount of newly injected knowledge grows; and 2) model updates are often irreversible. As modern LLMs become increasingly expressive, it is natural to question whether large-scale weight updates are necessary for acquiring a small amount of new knowledge. In this work, we propose a principled framework that models autoregressive language generation as a Markov process over tokens, where model memory is represented by a Markov transition matrix. Under this formulation, incorporating new knowledge/tokens corresponds to extending the state space, and preserving existing transitions guarantees retention of previously learned knowledge. We then prove a sample complexity bound for incorporating new tokens via a token-to-dictionary mapping strategy. In particular, for learning the transition behavior of each new token, the required number of samples scales linearly with the number of existing tokens it is mapped to. To realize this mapping, we propose an embedding-tuning algorithm that requires minimal parameter updates and induces zero forgetting. Experimental results further demonstrate the effectiveness of our method and validate our theoretical findings.", "date": "2026-05-05", "categories": []}
{"id": "2605.03085", "title": "Adaptive Data Compression and Reconstruction for Memory-Bounded EEG Continual Learning", "abstract": "Electroencephalography (EEG) signals provide millisecond-level temporal resolution but their analysis is limited by remarkable noise and inter-subject variability, making robust personalization difficult under limited annotations. Unsupervised Individual Continual Learning (UICL) has been proposed to address this practical challenge, where a model pretrained on a labeled cohort must adapt online to unlabeled subject streams under strict memory constraints. However, existing UICL methods typically store full past samples, which undermine the continual learning goal of avoiding retraining. Observing that EEG signals exhibit well-structured morphologies to be exploited via morphology-aware selection, compression, and reconstruction, here we propose Adaptive Data Compression and Reconstruction (ADaCoRe) for UICL. This is a memory-efficient pipeline composed of saliency-driven keyframe protection, rational polyphase compression, adjoint reconstruction with verbatim overwrite on protected indices, and prototype-confidence selection for adaptive exemplar maintenance. Across three representative benchmarks, ADaCoRe consistently outperforms recent strong baselines under tight buffer regimes (eg., the performance gains are at least +2.7 and +15.3 ACC on ISRUC and FACED datasets, respectively). Ablation studies quantify compression-fidelity trade-offs and highlight the contribution of each design, while visualizations confirm the preservation of key EEG morphology during compression and reconstruction.", "date": "2026-05-04", "categories": []}
{"id": "2605.03153", "title": "OCRR: A Benchmark for Online Correction Recovery under Distribution Shift", "abstract": "Static benchmarks measure a model frozen at training time. Real systems face distribution shift: new categories, paraphrased queries, drift: and must recover online via user corrections. No existing benchmark measures recovery speed under correction streams. We introduce OCRR (Online Correction Recovery Rate): a benchmark that streams a corpus through a classification system, applies oracle or stochastic corrections to wrong predictions, and reports two curves: novel-class accuracy and original-distribution accuracy versus correction count. We evaluate the substrate alongside nine baseline algorithms from five families plus seven bounded-storage variants of the substrate for the Pareto sweep, including standard online-learning baselines (river), continual-learning methods (EWC, A-GEM, LwF), retrieval/parametric hybrids (kNN-LM), parameter-efficient fine-tuning of a 1.5 B-parameter encoder (LoRA on DeBERTa-v3-large), and a hash-chained append-only substrate (Substrate). On Banking77 and CLINC150, under oracle and sparse correction policies, the substrate is the only system that simultaneously recovers novel-class accuracy (88.7 +/- 2.9 %) and retains original-distribution accuracy (95.4 +/- 0.8 %) beating the next-best published continual-learning baseline by 32.6 percentage points at equal memory budget, and beating LoRA-on-DeBERTa-v3-large by 84.6 percentage points on retention. We further find that classification accuracy remains stable at 99 % even as approximate-nearest-neighbour recall@5 degrades from 0.69 to 0.23 across 10 k to 10 M corpus scales, suggesting the substrate's margin-band majority vote is robust to retrieval imperfection in a way that pure top-k recall metrics do not predict. Code and data are available at https://github.com/adriangrassi/ocrr-benchmark.", "date": "2026-05-04", "categories": []}
{"id": "2605.03364", "title": "Dynamic Distillation and Gradient Consistency for Robust Long-Tailed Incremental Learning", "abstract": "The task of Long-tailed Class Incremental Learning (LT-CIL) addresses the sequential learning of new classes from datasets with imbalanced class distributions. This scenario intensifies the fundamental problem of catastrophic forgetting, inherent to continual learning, with the dual challenges of under-learning minority classes and overfitting majority classes. To tackle these combined issues, this paper proposes two main techniques. First, we introduce gradient consistency regularization, which leverages the moving average of gradients to suppress abrupt fluctuations and stabilize the training process. Second, we dynamically adjust the weight of the distillation loss by measuring the degree of class imbalance with normalized entropy. This adaptive weighting establishes an optimal balance between retaining old knowledge and acquiring new information. Experiments on the CIFAR-100-LT, ImageNetSubset-LT, and Food101-LT benchmarks show that our method achieves consistent accuracy improvements of up to 5.0\\%. Furthermore, we demonstrate dramatic gains in the challenging 'In-ordered' setting, where tasks progress from majority to minority classes, highlighting our method's robustness in mitigating forgetting under unfavorable learning dynamics. This enhanced performance is achieved without a significant increase in computational overhead, demonstrating the practicality of our framework.", "date": "2026-05-05", "categories": []}
{"id": "2605.03555", "title": "MILE: Mixture of Incremental LoRA Experts for Continual Semantic Segmentation across Domains and Modalities", "abstract": "Continual semantic segmentation requires models to adapt to new domains or modalities without sacrificing performance on previously learned tasks. Expert-based learning, in which task-specific modules specialize in different domains, has proven effective in mitigating forgetting. These methods include dynamic expansion, which suffers from scalability issues, or parameter isolation, which constrains the ability to learn new tasks. We introduce Mixture of Incremental LoRA Experts (MILE), a modular and parameter-efficient framework for continual segmentation across both domains and modalities. MILE leverages Low-Rank Adaptation (LoRA) to instantiate lightweight experts for each new task while keeping the pretrained base network frozen. Each expert is trained exclusively on its task data, thus avoids overwriting previously learned information. A prototype-guided gating mechanism dynamically selects the most appropriate expert at inference. MILE achieves the benefits of expert-based learning while overcoming its scalability limitations. It requires only a marginal parameter increase per task and tens of LoRA adapters are needed before matching the size of a single full model, making it highly efficient in both training and storage. Across domain- and modality-incremental benchmarks, MILE achieves strong performance while ensuring better stability, plasticity, and scalability.", "date": "2026-05-05", "categories": []}
{"id": "2605.03832", "title": "A Domain Incremental Continual Learning Benchmark for ICU Time Series Model Transportability", "abstract": "In recent years, machine learning has made significant progress in clinical outcome prediction, demonstrating increasingly accurate results. However, the substantial resources required for hospitals to train these models, such as data collection, labeling, and computational power, limit the feasibility for smaller hospitals to develop their own models. An alternative approach involves transferring a machine learning model trained by a large hospital to smaller hospitals, allowing them to fine-tune the model on their specific patient data.\n  However, these models are often trained and validated on data from a single hospital, raising concerns about their generalizability to new data. Our research shows that there are notable differences in measurement distributions and frequencies across various regions in the United States. To address this, we propose a benchmark that tests a machine learning model's ability to transfer from a source domain to different regions across the country. This benchmark assesses a model's capacity to learn meaningful information about each new domain while retaining key features from the original domain.\n  Using this benchmark, we frame the transfer of a machine learning model from one region to another as a domain incremental learning problem. While the task of patient outcome prediction remains the same, the input data distribution varies, necessitating a model that can effectively manage these shifts. We evaluate two popular domain incremental learning methods: data replay, which stores examples from previous data sources for fine-tuning on the current source, and Elastic Weight Consolidation (EWC), a model parameter regularization method that maintains features important for both data sources.", "date": "2026-05-05", "categories": []}
{"id": "2605.03866", "title": "Memory-Efficient Continual Learning with CLIP Models", "abstract": "Contrastive Language-Image Pretraining (CLIP) models excel at understanding image-text relationships but struggle with adapting to new data without forgetting prior knowledge. To address this, models are typically fine-tuned using both new task data and a memory buffer of past tasks. However, CLIP's contrastive loss suffers when the memory buffer is small, leading to performance degradation on previous tasks. We propose a memory-efficient, distributionally robust method that dynamically reweights losses per class during training. Our approach, tested on class incremental settings (CIFAR-100, ImageNet1K) and a domain incremental setting (DomainNet) adapts CLIP models quickly while minimizing catastrophic forgetting, even with minimal memory usage.", "date": "2026-05-05", "categories": []}
{"id": "2510.02945", "title": "Ergodic Risk Measures: Towards a Risk-Aware Foundation for Continual Reinforcement Learning", "abstract": "Continual reinforcement learning (continual RL) seeks to formalize the notions of lifelong learning and endless adaptation in RL. In particular, the aim of continual RL is to develop RL agents that can maintain a careful balance between retaining useful information and adapting to new situations. To date, continual RL has been explored almost exclusively through the lens of risk-neutral decision-making, in which the agent aims to optimize the expected long-run performance. In this work, we present the first formal theoretical treatment of continual RL through the lens of risk-aware decision-making, in which the behaviour of the agent is directed towards optimizing a measure of long-run performance beyond the mean. In particular, we show that the classical theory of risk measures, widely used as a theoretical foundation in non-continual risk-aware RL, is, in its current form, incompatible with continual learning. Then, building on this insight, we extend risk measure theory into the continual setting by introducing a new class of ergodic risk measures, and showing that it is compatible with continual learning. Finally, we provide a case study of continual risk-aware learning, along with empirical results, which show the intuitive appeal of ergodic risk measures in continual settings.", "date": "2025-10-03", "categories": []}
{"id": "2601.17616", "title": "Split-on-Share: Mixture of Sparse Experts for Task-Agnostic Continual Learning", "abstract": "Continual learning in Large Language Models (LLMs) is hindered by the plasticity-stability dilemma, where acquiring new capabilities often leads to catastrophic forgetting of previous knowledge. Existing methods typically treat parameters uniformly, failing to distinguish between specific task knowledge and shared capabilities. We introduce Mixture of Sparse Experts for Task-Agnostic Continual Learning, referred to as SETA, a framework that resolves the plasticity-stability conflict by decomposing the model into modular subspaces. Unlike standard updates, where tasks compete for the same parameters, SETA separates knowledge into unique experts, designed to isolate task-specific patterns, and shared experts, responsible for capturing common features. This structure is maintained through elastic weight anchoring, which protects critical shared knowledge and enables a unified gating network to automatically retrieve the correct expert combination for each task during inference. Extensive experiments across diverse domain-specific and general benchmarks demonstrate that SETA consistently outperforms state-of-the-art parameter-efficient fine-tuning-based continual learning methods.", "date": "2026-01-24", "categories": []}
{"id": "2605.01020", "title": "Continual Learning of Feedback-based Molecular Communication", "abstract": "This paper proposes and evaluates a new performance estimation method that leverages continual learning (CL) algorithms to carry out sequential simulation experiments for a feedback-based molecular communication protocol. As the protocol is sequentially examined in various experimental settings, the proposed CL-based performance estimators incrementally learn a series of unexperienced estimation tasks without compromising those that have been learned in the past. They are designed to work on a standard neural network architecture by customizing regularization and replay strategies in the loss function. Experimental results demonstrate that the proposed estimators can effectively learn on a continuous stream of simulation results and enhance the baseline neural network by improving estimation accuracy at a variety of computational costs. This paper's contribution is to establish the implications of CL in the field of molecular communication.", "date": "2026-05-01", "categories": []}
{"id": "2605.01131", "title": "Forager: a lightweight testbed for continual learning with partial observability in RL", "abstract": "In continual reinforcement learning (CRL), good performance requires never-ending learning, acting, and exploration in a big, partially observable world. Most CRL experiments have focused on loss of plasticity -- the inability to keep learning -- in one-off experiments where some unobservable non-stationarity is added to classic fully observable MDPs. Further, these experiments rarely consider the role of partial observability and the importance of CRL agents that use memory or recurrence. One potential reason for this focus on mitigating loss of plasticity without considering partial observability is that many partially-observable CRL environments are prohibitively expensive. In this paper, we introduce Forager, a light-weight partially-observable CRL environment with a constant memory footprint. We provide a set of experiments and sample tasks demonstrating that Forager is challenging for current CRL agents and yet also allows for in-depth study of those agents. We demonstrate that agents exhibit loss of plasticity, proposed mitigations can help, but that most useful is to leverage state construction. We conclude with a variant of Forager that generates an unending stream of new tasks to learn that clearly highlights the limitations of current CRL agents.", "date": "2026-05-01", "categories": []}
{"id": "2605.01383", "title": "Sequential Learning and Catastrophic Forgetting in Differentiable Resistor Networks", "abstract": "Differentiable physical networks provide a simple setting in which learning can be studied through the interaction between trainable parameters and physical equilibrium constraints. We investigate sequential learning in differentiable resistor networks governed by Kirchhoff's laws. Although individual input--output mappings can be learned by gradient-based adjustment of edge conductances, sequential training on conflicting tasks produces catastrophic forgetting. We show that forgetting is controlled by task conflict and by the degree of adaptation to the new task. Uniform anchoring and normalised gradient-weighted anchoring reduce forgetting only by increasing the final loss on the new task, giving a clear forgetting--adaptation trade-off. We also show that forgetting is associated with localised conductance changes on high-current edges, giving a physical interpretation as reconfiguration of dominant transport pathways. Broader random-task ensembles show that the strongest forgetting occurs when the second task reverses the output ordering imposed by the first task. Finally, comparisons across Erd\\H{o}s--R\\'enyi, small-world, scale-free, and random-geometric graph ensembles show that topology changes the forgetting--adaptation balance. These results position differentiable resistor networks as compact, physically interpretable testbeds for studying continual learning in tunable matter.", "date": "2026-05-02", "categories": []}
{"id": "2605.02105", "title": "Sharpness-Aware Pretraining Mitigates Catastrophic Forgetting", "abstract": "Pretraining optimizers are tuned to produce the strongest possible base model, on the assumption that a stronger starting point yields a stronger model after subsequent changes like post-training and quantization. This overlooks the geometry of the base model which controls how much of the base model's capabilities survive subsequent parameter updates. We study three pretraining optimization approaches that bias optimization toward flatter minima: Sharpness-Aware Minimization (SAM), large learning rates, and shortened learning rate annealing periods. Across model sizes ranging from 20M to 150M parameters, we find that these interventions consistently improve downstream performance after post-training on five common datasets with up to 80% less forgetting. These principles hold at scale: a short SAM mid-training phase applied to an existing OLMo-2-1B checkpoint reduces forgetting by 31% after MetaMath post-training and by 40% after 4-bit quantization.", "date": "2026-05-04", "categories": []}
{"id": "2605.02509", "title": "MPCS: Neuroplastic Continual Learning via Multi-Component Plasticity and Topology-Aware EWC", "abstract": "Continual learning systems face a fundamental tension between plasticity -- acquiring new knowledge  --  and stability  --  retaining prior knowledge. We introduce MPCS (Multi-Plasticity Continual System), a neuroplastic architecture that integrates eleven complementary mechanisms: task-driven neurogenesis, Fourier-encoded inputs, EWC regularization, meta-replay, mixed consolidation, hybrid gating, synapse pruning/regeneration, Hebbian updates, task similarity routing, adaptive growth control, and continuous neuron importance tracking. We evaluate MPCS on MEP-BENCH, a multi-track benchmark spanning 31 tasks across regression, classification, logic, and mixed domains, using a three-dimensional Pareto criterion over task performance (Perf), representation diversity (RD), and gradient conflict rate (GCR). Across 15 ablation configurations (3 seeds x 4 tracks x 2000 epochs), MPCS achieves a Normalized Efficiency Score of 94.2, placing it on the Pareto frontier among 9 of 14 gate-passing systems. Key findings: (i) Fourier encoding is the single most critical component (removal drops Perf by 30.7 pp and fails the MEP gate on 14% of tasks); (ii) global EWC degrades performance (NES = -4.2); topology-local EWC reduces this penalty (NES 90.5->91.8) but does not eliminate it; removing EWC entirely yields MPCS_EFFICIENT, the highest-Perf system -- establishing a monotone relationship in the high task-similarity regime (s_bar ~= 0.95): global EWC < topology EWC < no EWC; (iii) the Pareto status assessment is predictive: removing the two Pareto-dominated components (EWC + Hebbian) jointly yields MPCS_EFFICIENT, which improves Perf by 0.6 pp at 4.7x lower compute cost (127 vs. 602 min), validating the Pareto frontier as an actionable model-compression guide.", "date": "2026-05-04", "categories": []}
{"id": "2407.19001", "title": "Effective Prompt Pool Learning for Continual Category Discovery", "abstract": "This paper studies effective prompt pool learning for Continual Category Discovery (CCD), a challenging open-world setting where a model must discover novel categories from a continuous stream of unlabelled data containing both known and novel classes, while mitigating catastrophic forgetting of previously learned concepts. We introduce a series of novel prompt-pool-based frameworks for CCD, each exploring a different design of prompt pools. First, we propose PromptCCD, which focuses on global class prototypes via a Gaussian Mixture Prompt (GMP) module. GMP fits a generative Gaussian mixture model over feature embeddings, where each mixture component serves as both a class prototype and a dynamic prompt that conditions the backbone's representations. This design enables label-free prompt selection and on-the-fly estimation of the number of emerging categories. Through a systematic spectrum study, we then show that category count, rather than sample size, is the primary bottleneck for discovery performance, motivating the need for finer-grained representations. Building on this finding, we propose PromptCCD++, which focuses on object-part prototypes via Part-level Prompting (PLP) modules. PLP decomposes prompt pool into multiple, specialized part-level prompt pools. During discovery phase, these pools dynamically assign part-specific prompts to local object regions without the need for manual part annotations, enabling the model to learn object-part representations that boost category discovery. Extensive evaluations on both generic and fine-grained benchmarks, supported by comprehensive ablation studies, demonstrate the effectiveness of our framework for CCD.", "date": "2024-07-26", "categories": []}
{"id": "2506.02671", "title": "Test-Time Distillation for Continual Model Adaptation", "abstract": "Deep neural networks often suffer performance degradation upon deployment due to distribution shifts. Continual Test-Time Adaptation (CTTA) aims to address this issue in an unsupervised manner. However, existing methods that rely on self-supervision are prone to an inherent self-referential feedback loop that amplifies initial prediction errors, leading to model drift. We revisit this limitation and propose Test-Time Distillation (TTD), which reframes adaptation as a distillation process guided by a frozen Vision-Language Model (VLM) as an external signal. While promising, we find that direct distillation is fraught with two pitfalls: (1) the Generalist Trap, where the VLM's broad but non-specialized knowledge leads to suboptimal performance on specific tasks and shifts; and (2) the Entropy Bias, where naive model fusion techniques based on entropy fail due to the disparate calibration of heterogeneous models. These pitfalls highlight the need to build a robust supervisory signal and leverage it to guide the target model toward stable adaptation. Hence, we present CoDiRe, a Continual Distillation and Rectification framework for TTD. CoDiRe first constructs a robust blended teacher by dynamically fusing the predictions of the VLM and the target model. Critically, it circumvents the Entropy Bias by leveraging Maximum Softmax Probability (MSP) as a more reliable confidence metric for weighting each model's expertise. Then it applies an Optimal Transport-based rectification to further align predictions with the blended teacher, enabling continuous and stable adaptation. Extensive experiments show that CoDiRe outperforms state-of-the-art baselines, exceeding CoTTA by 10.55% with only 48% of its time cost on ImageNet-C. Project page is publicly available at https://github.com/walawalagoose/TTD.", "date": "2025-06-03", "categories": []}
{"id": "2509.22562", "title": "Activation Function Design Sustains Plasticity in Continual Learning", "abstract": "In independent, identically distributed (i.i.d.) training regimes, activation functions have been benchmarked extensively, and their differences often shrink once model size and optimization are tuned. In continual learning, however, the picture is different: beyond catastrophic forgetting, models can progressively lose the ability to adapt (referred to as loss of plasticity) and the role of the non-linearity in this failure mode remains underexplored. We show that activation choice is a primary, architecture-agnostic lever for mitigating plasticity loss. Building on a property-level analysis of negative-branch shape and saturation behavior, we introduce two drop-in nonlinearities (Smooth-Leaky and Randomized Smooth-Leaky) and evaluate them in two complementary settings: (i) supervised class-incremental benchmarks and (ii) reinforcement learning with non-stationary MuJoCo environments designed to induce controlled distribution and dynamics shifts. We also provide a simple stress protocol and diagnostics that link the shape of the activation to the adaptation under change. The takeaway is straightforward: thoughtful activation design offers a lightweight, domain-general way to sustain plasticity in continual learning without extra capacity or task-specific tuning.", "date": "2025-09-26", "categories": []}
{"id": "2604.24637", "title": "Cortex-Inspired Continual Learning: Unsupervised Instantiation and Recovery of Functional Task Networks", "abstract": "Block-sequential continual learning demands that a single model both protect prior solutions from catastrophic forgetting and efficiently infer at inference time which prior solution matches the current input without task labels. We present Functional Task Networks (FTN), a parameter-isolation method inspired by structural and dynamical motifs found in the mammalian neocortex. Similar to mixture-of-experts, this method uses a high dimensional, self-organizing binary mask over a large population of small but deep networks, inspired by dendritic models of pyramidal neurons. The mask is produced by a three-stage procedure: (1) gradient descent on a continuous mask identifies task-relevant neurons, (2) a smoothing kernel biases the result toward spatial contiguity, (3) and k-winner-take-all binarizes the resulting group at a fixed capacity budget. Like mixture-of-experts, each neuron is an independent deep network, so disjoint masks give exactly disjoint gradient updates, providing structural guarantees against catastrophic forgetting. This three-stage procedure recovers the sub-network of a previously-trained task in a single gradient step, providing unsupervised task segmentation at inference time. We test it on three continual-learning benchmarks: (1) a synthetic multi-task classification/regression generator, (2) MNIST with shuffled class labels (pure concept shift), and (3) Permuted MNIST (domain shift). On all three, FTN with fine grained smoothing (FTN-Slow) results in nearly zero forgetting. FTN with a large kernel and only 2 iterations of smoothing (FTN-Fast) trades off some retention for increased speed. We show that the spatial organization mechanism reduces the effective mask search from the combinatorial top-k subset problem in O(C(H,K)) to the complexity of a near-linear scan in O(H) over compact cortical neighborhoods, which is parallelized by the gradient-based update.", "date": "2026-04-27", "categories": []}
{"id": "2604.27031", "title": "NORACL: Neurogenesis for Oracle-free Resource-Adaptive Continual Learning", "abstract": "In a continual learning setting, we require a model to be plastic enough to learn a new task and stable enough to not disturb previously learned capabilities. We argue that this dilemma has an architectural root. A finite network has limited representational and plastic resources, yet the required capacity depends on properties of the future task stream that are unknown: how many tasks will be encountered, and how much they overlap in feature space. Regularization-based methods preserve past knowledge within fixed-capacity architectures and therefore implicitly rely on an oracle architecture sized for this unknown future. When tasks are only weakly related, fixed architectures progressively run out of plastic resources; when tasks are few or strongly overlapping, models are often over-provisioned. Inspired by neurogenesis in biology, we propose NORACL to address the stability-plasticity dilemma by tackling the oracle architecture problem through neuronal growth. Starting from a compact network, NORACL grows only when needed by monitoring two complementary signals for representational and plasticity saturation. We evaluate NORACL against oracle-sized static baselines across varying task counts and geometries. Across all settings, NORACL achieves final average accuracies that are better than or on par with oracle-provisioned static baselines while using fewer parameters. Additionally, NORACL yields architectures with interpretable growth, i.e. dissimilar tasks predominantly expand feature-extraction layers, whereas tasks which rely on common features shift growth toward later feature-combination layers. Our analysis further explains why fixed-capacity networks lose plasticity as tasks accumulate, whereas NORACL creates fresh capacity for new tasks through growth. Together, these results show that adaptive neurogenesis pushes the stability-plasticity Pareto frontier of continual learning.", "date": "2026-04-29", "categories": []}
{"id": "2604.27063", "title": "Learning to Forget: Continual Learning with Adaptive Weight Decay", "abstract": "Continual learning agents with finite capacity must balance acquiring new knowledge with retaining the old. This requires controlled forgetting of knowledge that is no longer needed, freeing up capacity to learn. Weight decay, viewed as a mechanism for forgetting, can serve this role by gradually discarding information stored in the weights. However, a fixed scalar weight decay drives this forgetting uniformly over time and uniformly across all parameters, even when some encode stable knowledge while others track rapidly changing targets. We introduce Forgetting through Adaptive Decay (FADE), which adapts per-parameter weight decay rates online via approximate meta-gradient descent. We derive FADE for the online linear setting and apply it to the final layer of neural networks. Our empirical analysis shows that FADE automatically discovers distinct decay rates for different parameters, complements step-size adaptation, and consistently improves over fixed weight decay across online tracking and streaming classification problems.", "date": "2026-04-29", "categories": []}
{"id": "2604.27656", "title": "When Does Structure Matter in Continual Learning? Dimensionality Controls When Modularity Shapes Representational Geometry", "abstract": "To preserve previously learned representations, continual learning systems must strike a balance between plasticity, the ability to acquire new knowledge, and stability. This stability-plasticity dilemma affects how representations can be reused across tasks: shared structure enables transfer when tasks are similar but may also induce interference when new learning disrupts existing representations. However, it remains unclear when and why structural separation influences this trade-off. In this study, we examine how network architecture, task similarity, and representational dimensionality jointly shape learning in a sequential task paradigm inspired by transfer-interference studies. We compare a task-partitioned modular recurrent network with a single-module baseline by systematically varying task similarity (low, medium, high) and the scale of weight initialization, which induces different learning regimes that we empirically characterize through the effective dimensionality of the learned representations. We find that architecture has minimal impact in high-dimensional regimes where representations are sufficiently unconstrained to accommodate multiple tasks without strong interference. In contrast, in lower-dimensional (rich) regimes, architectural separation is decisive: modular networks exhibit graded alignment of task-specific subspaces with overlap for similar tasks, partial orthogonalization for moderately dissimilar tasks, and stronger separation for dissimilar tasks. This graded geometry is absent in the single network baseline. Our findings suggest that representational dimensionality acts as a key organizing variable governing when structural separation becomes functionally relevant, and highlight adaptive geometry as a central principle for designing continual learning systems.", "date": "2026-04-30", "categories": []}
{"id": "2603.09145", "title": "Causally Sufficient and Necessary Feature Expansion for Class-Incremental Learning", "abstract": "Current expansion-based methods for Class Incremental Learning (CIL) effectively mitigate catastrophic forgetting by freezing old features. However, such task-specific features learned from the new task may collide with the old features. From a causal perspective, spurious feature correlations are the main cause of this collision, manifesting in two scopes: (i) guided by empirical risk minimization (ERM), intra-task spurious correlations cause task-specific features to rely on shortcut features. These non-robust features are vulnerable to interference, inevitably drifting into the feature space of other tasks; (ii) inter-task spurious correlations induce semantic confusion between visually similar classes across tasks. To address this, we propose a Probability of Necessity and Sufficiency (PNS)-based regularization method to guide feature expansion in CIL. Specifically, we first extend the definition of PNS to expansion-based CIL, termed CPNS, which quantifies both the causal completeness of intra-task representations and the separability of inter-task representations. We then introduce a dual-scope counterfactual generator based on twin networks to ensure the measurement of CPNS, which simultaneously generates: (i) intra-task counterfactual features to minimize intra-task PNS risk and ensure causal completeness of task-specific features, and (ii) inter-task interfering features to minimize inter-task PNS risk, ensuring the separability of inter-task representations. Theoretical analyses confirm its reliability. The regularization is a plug-and-play method for expansion-based CIL to mitigate feature collision. Extensive experiments demonstrate the effectiveness of the proposed method.", "date": "2026-03-10", "categories": []}
{"id": "2411.16073", "title": "Soft-TransFormers for Continual Learning", "abstract": "Inspired by the \\emph{Well-initialized Lottery Ticket Hypothesis (WLTH)}, we introduce Soft-Transformer (Soft-TF), a parameter-efficient framework for continual learning that leverages soft, real-valued subnetworks over a frozen pre-trained Transformer. Instead of relying on manually designed prompts or adapters, Soft-TF learns task-specific multiplicative masks applied to the key, query, value, and output projections in self-attention. These masks enable smooth and stable task adaptation while preserving shared representations. Combined with a lightweight dual-prompt mechanism, Soft-TF maintains strong knowledge retention and mitigates Catastrophic Forgetting (CF). Across multiple continual learning benchmarks, Soft-TF achieves state-of-the-art performance, consistently outperforming prompt-based, adapter-based, and LoRA-style baselines while requiring minimal additional parameters.", "date": "2024-11-25", "categories": []}
{"id": "2604.16171", "title": "JumpLoRA: Sparse Adapters for Continual Learning in Large Language Models", "abstract": "Adapter-based methods have become a cost-effective approach to continual learning (CL) for Large Language Models (LLMs), by sequentially learning a low-rank update matrix for each task. To mitigate catastrophic forgetting, state-of-the-art approaches impose constraints on new adapters with respect to the previous ones, by targeting either subspace or coordinate-wise interference. In this paper, we propose JumpLoRA, a novel framework to adaptively induce sparsity in the Low-Rank Adaptation (LoRA) blocks through the use of JumpReLU gating. The method achieves dynamic parameter isolation, which helps prevent task interference. We demonstrate that our method is highly modular and compatible with LoRA-based CL approaches. Specifically, it significantly boosts the performance of IncLoRA and outperforms the leading state-of-the-art CL method, ELLA.", "date": "2026-04-17", "categories": []}
{"id": "2604.25376", "title": "CoRE: Concept-Reasoning Expansion for Continual Brain Lesion Segmentation", "abstract": "Accurate brain lesion segmentation in MRI is vital for effective clinical diagnosis and treatment planning. Due to high annotation costs and strict data privacy regulations, universal models require employing Continual Learning (CL) to adapt to evolving clinical tasks without losing previously acquired knowledge. However, existing CL paradigms often suffer from capacity limits or redundant parameter growth, and even advanced dynamic methods rely mostly on image-perception strategies that struggle to handle the substantial pathological and multimodal heterogeneity inherent in brain imaging. To address this issue, we propose Concept-Reasoning Expansion (CoRE) framework, which establishes a joint decision-making mechanism by integrating visual features with structured concepts. Through the alignment of image tokens with a hierarchical concept library, CoRE simulates clinical reasoning to guide both interpretable expert routing and demand-based model growth. This collaborative process ensures model evolution is grounded in clinical priors, preventing redundant parameter expansion while maximizing knowledge reuse. Extensive evaluations across 12 sequential brain lesion MRI tasks demonstrate that CoRE achieves state-of-the-art performance and provides a high knowledge starting point for efficient future adaptation. Its superior few-shot transferability and clinical interpretability further validate its effectiveness in managing non-stationary clinical data streams. Our code will be released soon.", "date": "2026-04-28", "categories": []}
{"id": "2604.25898", "title": "TSN-Affinity: Similarity-Driven Parameter Reuse for Continual Offline Reinforcement Learning", "abstract": "Continual offline reinforcement learning (CORL) aims to learn a sequence of tasks from datasets collected over time while preserving performance on previously learned tasks. This setting corresponds to domains where new tasks arise over time, but adapting the model in live environment interactions is expensive, risky, or impossible. However, CORL inherits the dual difficulty of offline reinforcement learning and adapting while preventing catastrophic forgetting. Replay-based continual learning approaches remain a strong baseline but incur memory overhead and suffer from a distribution mismatch between replayed samples and newly learned policies. At the same time, architectural continual learning methods have shown strong potential in supervised learning but remain underexplored in CORL. In this work, we propose TSN-Affinity, a novel CORL method based on TinySubNetworks and Decision Transformer. The method enables task-specific parameterization and controlled knowledge sharing through a RL-aware reuse strategy that routes tasks according to action compatibility and latent similarity. We evaluate the approach on benchmarks based on Atari games and simulations of manipulation tasks with the Franka Emika Panda robotic arm, covering both discrete and continuous control. Results show strong retention from sparse SubNetworks, with routing further improving multi-task performance. Our findings suggest that similarity-guided architectural reuse is a strong and viable alternative to replay-based strategies in a CORL setting. Our code is available at: https://github.com/anonymized-for-submission123/tsn-affinity.", "date": "2026-04-28", "categories": []}
{"id": "2601.20597", "title": "StructAlign: Structured Cross-Modal Alignment for Continual Text-to-Video Retrieval", "abstract": "Continual Text-to-Video Retrieval (CTVR) is a challenging multimodal continual learning setting, where models must incrementally learn new semantic categories while maintaining accurate text-video alignment for previously learned ones, thus making it particularly prone to catastrophic forgetting. A key challenge in CTVR is feature drift, which manifests in two forms: intra-modal feature drift caused by continual learning within each modality, and non-cooperative feature drift across modalities that leads to modality misalignment. To mitigate these issues, we propose StructAlign, a structured cross-modal alignment method for CTVR. First, StructAlign introduces a simplex Equiangular Tight Frame (ETF) geometry as a unified geometric prior to mitigate modality misalignment. Building upon this geometric prior, we design a cross-modal ETF alignment loss that aligns text and video features with category-level ETF prototypes, encouraging the learned representations to form an approximate simplex ETF geometry. In addition, to suppress intra-modal feature drift, we design a Cross-modal Relation Preserving loss, which leverages complementary modalities to preserve cross-modal similarity relations, providing stable relational supervision for feature updates. By jointly addressing non-cooperative feature drift across modalities and intra-modal feature drift, StructAlign effectively alleviates catastrophic forgetting in CTVR. Extensive experiments on benchmark datasets demonstrate that our method consistently outperforms state-of-the-art continual retrieval approaches.", "date": "2026-01-28", "categories": []}
{"id": "2604.21927", "title": "Fine-Tuning Regimes Define Distinct Continual Learning Problems", "abstract": "Continual learning (CL) studies how models acquire tasks sequentially while retaining previously learned knowledge. Despite substantial progress in benchmarking CL methods, comparative evaluations typically keep the fine-tuning regime fixed. In this paper, we argue that the fine-tuning regime, defined by the trainable parameter subspace, is itself a key evaluation variable. We formalize adaptation regimes as projected optimization over fixed trainable subspaces, showing that changing the trainable depth alters the effective update signal through which both current task fitting and knowledge preservation operate. This analysis motivates the hypothesis that method comparisons need not be invariant across regimes. We test this hypothesis in task incremental CL, five trainable depth regimes, and four standard methods: online EWC, LwF, SI, and GEM. Across five benchmark datasets, namely MNIST, Fashion MNIST, KMNIST, QMNIST, and CIFAR-100, and across 11 task orders per dataset, we find that the relative ranking of methods is not consistently preserved across regimes. We further show that deeper adaptation regimes are associated with larger update magnitudes, higher forgetting, and a stronger relationship between the two. These results show that comparative conclusions in CL can depend strongly on the chosen fine-tuning regime, motivating regime-aware evaluation protocols that treat trainable depth as an explicit experimental factor.", "date": "2026-04-23", "categories": []}
{"id": "2604.23388", "title": "A Parametric Memory Head for Continual Generative Retrieval", "abstract": "Generative information retrieval (GenIR) consolidates retrieval into a single neural model that decodes document identifiers (docids) directly from queries. While this model-as-index paradigm offers architectural simplicity, it is poorly suited to dynamic document collections. Unlike modular systems, where indexes are easily updated, GenIR's knowledge is parametrically encoded in its weights; consequently, standard adaptation methods such as full and parameter-efficient fine-tuning can induce catastrophic forgetting. We show that sequential adaptation improves retrieval on newly added documents but substantially degrades performance on earlier slices, exposing a pronounced stability-plasticity trade-off. To address this, we propose post-adaptation memory tuning (PAMT), a memory-only stabilization stage that augments an adapted model with a modular parametric memory head (PMH). PAMT freezes the backbone and attaches a product-key memory with fixed addressing. During prefix-trie constrained decoding, decoder hidden states sparsely query PMH to produce residual corrections in hidden space; these corrections are mapped to score adjustments via the frozen output embedding matrix, computed only over trie-valid tokens. This guides docid generation while keeping routing and backbone parameters fixed. To limit cross-slice interference, PAMT updates only a fixed budget of memory values selected using decoding-time access statistics, prioritizing entries frequently activated by the current slice and rarely used in prior sessions. Experiments on MS MARCO and Natural Questions under sequential, disjoint corpus increments show that PAMT substantially improves retention on earlier slices with minimal impact on retrieval performance for newly added documents, while modifying only a sparse subset of memory values per session.", "date": "2026-04-25", "categories": []}
{"id": "2604.23987", "title": "Continual Calibration: Coverage Can Collapse Before Accuracy in Lifelong LLM Fine-Tuning", "abstract": "Continual learning for large language models is typically evaluated through accuracy retention under sequential fine-tuning. We argue that this perspective is incomplete, because uncertainty reliability can degrade earlier and more sharply than top-1 performance. We study this empirically by measuring conformal coverage and calibration error on sequentially fine-tuned models across three model families and eight task sequences drawn primarily from classification and multiple-choice benchmarks. Across the classification-style settings we study, coverage loss exceeds accuracy loss by a factor of roughly \\(3.4\\times \\pm 0.5\\times\\) on average across seeds; in the most pronounced case, coverage drops from \\(0.92\\) to \\(0.61\\), while accuracy remains within three points of baseline. Standard continual-learning methods that preserve accuracy do not automatically preserve coverage, and naive calibration baselines recover only part of the gap. We propose calibration replay, a lightweight post-hoc procedure that maintains a task-specific held-out buffer and refits a task-specific conformal threshold under the current model after each update. It adds no training-time gradient cost, uses less than one percent of the memory of ordinary experience replay, and typically restores coverage to within two points of nominal at buffer size \\(m = 200\\). We accompany the empirical study with a drift decomposition, a finite-sample recovery theorem showing exact conformal validity under exchangeability, and a mixture-validity proposition explaining why pooled thresholds do not suffice. Our guarantees are stated for classification-style tasks with task-specific buffers; extensions to open-ended generation are exploratory.", "date": "2026-04-27", "categories": []}
{"id": "2411.04680", "title": "Privacy Leakage via Output Label Space and Differentially Private Continual Learning", "abstract": "Differential privacy (DP) is a formal privacy framework that enables training machine learning (ML) models while protecting individuals' data. As pointed out by prior work, ML models are part of larger systems, which can lead to so-called privacy side-channels even if the model training itself is DP. We identify the output label space of a classification model as such a privacy side-channel and show a concrete privacy attack that exploits it. The side-channel becomes highly relevant in continual learning (CL), where the output label space changes over time. To reason about privacy guarantees in CL, we introduce a formalisation of DP for CL, which also clarifies how our approach differs from existing approaches. We propose and evaluate two methods for eliminating this side-channel: applying an optimal DP mechanism to release the labels in the sensitive data, and using a large public label space. We explore the trade-offs of these methods through adapting pre-trained models. We demonstrate empirically that our models consistently achieve higher accuracy under DP than previous work over both Split-CIFAR-100 and Split-ImageNet-R, with a stronger privacy model.", "date": "2024-11-07", "categories": []}
{"id": "2604.17578", "title": "Recovery Guarantees for Continual Learning of Dependent Tasks: Memory, Data-Dependent Regularization, and Data-Dependent Weights", "abstract": "Continual learning (CL) is concerned with learning multiple tasks sequentially without forgetting previously learned tasks. Despite substantial empirical advances over recent years, the theoretical development of CL remains in its infancy. At the heart of developing CL theory lies the challenge that the data distribution varies across tasks, and we argue that properly addressing this challenge requires understanding this variation--dependency among tasks. To explicitly model task dependency, we consider nonlinear regression tasks and propose the assumption that these tasks are dependent in such a way that the data of the current task is a nonlinear transformation of previous data. With this model and under natural assumptions, we prove statistical recovery guarantees (more specifically, bounds on estimation errors) for several CL paradigms in practical use, including experience replay with data-independent regularization and data-independent weights that balance the losses of tasks, replay with data-dependent weights, and continual learning with data-dependent regularization (e.g., knowledge distillation). To the best of our knowledge, our bounds are informative in cases where prior work gives vacuous bounds.", "date": "2026-04-19", "categories": []}
{"id": "2604.22407", "title": "Hidden Failure Modes of Gradient Modification under Adam in Continual Learning, and Adaptive Decoupled Moment Routing as a Repair", "abstract": "Many continual-learning methods modify gradients upstream (e.g., projection, penalty rescaling, replay mixing) while treating Adam as a neutral backend. We show this composition has a hidden failure mode. In a high-overlap, non-adaptive 8-domain continual LM, all shared-routing projection baselines collapse close to vanilla forgetting (12.5--12.8 vs. 13.2). A 0.5% replay buffer is the strongest shared alternative but still reaches 11.6, while fixed-strength decoupling falls below vanilla at 14.1. Only adaptive decoupled routing remains stable at 9.4, improving over vanilla by 3.8 units. On a 16-domain stream, its gain over the strongest shared-routing projection baseline grows to 4.5--4.8 units. The failure is largely invisible on clean benchmarks.\n  We explain this effect through Adam's second-moment pathway: in the tested regime, projection induces a 1/(1-alpha) inflation of the old-direction effective learning rate, matching measurements within 8% across eight alpha values. The same conflict appears with penalty methods, replay mixing, and at 7B scale under LoRA. Our fix routes the modified gradient only to the first moment while preserving magnitude-faithful second-moment statistics, with overlap-aware adaptive strength. This simple change is the only tested configuration that consistently avoids collapse across methods, optimizers, and scale.", "date": "2026-04-24", "categories": []}
{"id": "2604.22464", "title": "Towards Adaptive Continual Model Merging via Manifold-Aware Expert Evolution", "abstract": "Continual Model Merging (CMM) sequentially integrates task-specific models into a unified architecture without intensive retraining. However, existing CMM methods are hindered by a fundamental saturation-redundancy dilemma: backbone-centric approaches face parameter saturation and representation interference within fixed capacities, whereas Mixture-of-Experts (MoE) variants resort to indiscriminate expansion, incurring expert redundancy and a routing bottleneck reliant on additional data-driven optimization. To resolve these challenges, we propose MADE-IT (Manifold-Aware Dynamic Expert Evolution and Implicit rouTing), an adaptive CMM method that orchestrates expert management and activation by grounding intrinsic expert representations in manifold geometry. We introduce a projection-based subspace affinity metric coupled with a distribution-aware adaptive threshold mechanism to guide autonomous expert evolution, harmonizing diversity with architectural parsimony. Furthermore, to bypass parameterized gating networks, we design a data-free and training-free implicit routing mechanism that activates experts via feature-subspace alignment. Extensive experiments demonstrate that MADE-IT consistently outperforms strong baselines in accuracy and robustness across long-horizon and shuffled task sequences, while significantly pruning redundant experts, particularly within generic modules and early layers.", "date": "2026-04-24", "categories": []}
{"id": "2406.11354", "title": "Preserving Knowledge in Large Language Model with Model-Agnostic Self-Decompression", "abstract": "Humans can retain old knowledge while learning new information, but Large Language Models (LLMs) often suffer from catastrophic forgetting when post-pretrained or supervised fine-tuned (SFT) on domain-specific data. Moreover, for Multimodal Large Language Models (MLLMs) which are composed of the LLM base and visual projector (e.g. LLaVA), a significant decline in performance on language benchmarks was observed compared to their single-modality counterparts. To address these challenges, we introduce a novel model-agnostic self-decompression method, Tree Generation (TG), that decompresses knowledge within LLMs into the training corpus. This paper focuses on TG-SFT, which can synthetically generate SFT data for the instruction tuning steps. By incorporating the dumped corpus during SFT for MLLMs, we significantly reduce the forgetting problem.", "date": "2024-06-17", "categories": []}
{"id": "2410.16006", "title": "Exploring Continual Fine-Tuning for Enhancing Language Ability in Large Language Model", "abstract": "A common challenge towards the adaptability of Large Language Models (LLMs) is their ability to learn new languages over time without hampering the model's performance on languages in which the model is already proficient (usually English). Continual fine-tuning (CFT) is the process of sequentially fine-tuning an LLM to enable the model to adapt to downstream tasks with varying data distributions and time shifts. This paper focuses on the language adaptability of LLMs through CFT. We study a two-phase CFT process in which an English-only end-to-end fine-tuned LLM from Phase 1 (predominantly Task Ability) is sequentially fine-tuned on a multilingual dataset -- comprising task data in new languages -- in Phase 2 (predominantly Language Ability). We observe that the ``similarity'' of Phase 2 tasks with Phase 1 determines the LLM's adaptability. For similar phase-wise datasets, the LLM after Phase 2 does not show deterioration in task ability. In contrast, when the phase-wise datasets are not similar, the LLM's task ability deteriorates. We test our hypothesis on the open-source \\mis\\ and \\llm\\ models with multiple phase-wise dataset pairs. To address the deterioration, we analyze tailored variants of two CFT methods: layer freezing and generative replay. Our findings demonstrate their effectiveness in enhancing the language ability of LLMs while preserving task performance, in comparison to relevant baselines.", "date": "2024-10-21", "categories": []}
{"id": "2511.11439", "title": "Retrofit: Continual Learning with Controlled Forgetting for Binary Security Detection and Analysis", "abstract": "Binary security has increasingly relied on deep learning to reason about malware behavior and program semantics. However, the performance often degrades as threat landscapes evolve and code representations shift. While continual learning (CL) offers a natural solution through sequential updates, most existing approaches rely on data replay or unconstrained updates, limiting their applicability and effectiveness in data-sensitive security environments. We propose RETROFIT, which regulates knowledge retention and adaptation with controlled forgetting at each update, without requiring historical data. Our key idea is to consolidate previously trained and newly fine-tuned models, serving as teachers of legacy and emergent knowledge, through retrospective-free parameter merging. Forgetting control is achieved by 1) constraining parameter changes to low-rank and sparse subspaces for approximate orthogonality, and 2) employing a confidence-guided arbitration mechanism to dynamically aggregate knowledge from both teachers.\n  Our evaluation on two representative applications demonstrates that RETROFIT consistently mitigates forgetting while maintaining adaptability. In malware detection under temporal drift, it substantially improves the retention score, from 20.2% to 38.6% over CL baselines, and exceeds the oracle upper bound on new data. In binary summarization across decompilation levels, where analyzing stripped binaries is especially challenging, RETROFIT achieves over 2x the BLEU score of transfer learning used in prior work and surpasses all baselines in cross-representation generalization.", "date": "2025-11-14", "categories": []}
{"id": "2604.21280", "title": "ImageHD: Energy-Efficient On-Device Continual Learning of Visual Representations via Hyperdimensional Computing", "abstract": "On-device continual learning (CL) is critical for edge AI systems operating on non-stationary data streams, but most existing methods rely on backpropagation or exemplar-heavy classifiers, incurring substantial compute, memory, and latency overheads. Hyperdimensional computing (HDC) offers a lightweight alternative through fast, non-iterative online updates. Combined with a compact convolutional neural network (CNN) feature extractor, HDC enables efficient on-device adaptation with strong visual representations. However, prior HDC-based CL systems often depend on multi-tier memory hierarchies and complex cluster management, limiting deployability on resource-constrained hardware.\n  We present ImageHD, an FPGA accelerator for on-device continual learning of visual data based on HDC. ImageHD targets streaming CL under strict latency and on-chip memory constraints, avoiding costly iterative optimization. At the algorithmic level, we introduce a hardware-aware CL method that bounds class exemplars through a unified exemplar memory and a hardware-efficient cluster merging strategy, while incorporating a quantized CNN front-end to reduce deployment overhead without sacrificing accuracy. At the system level, ImageHD is implemented as a streaming dataflow architecture on the AMD Zynq ZCU104 FPGA, integrating HDC encoding, similarity search, and bounded cluster management using word-packed binary hypervectors for massively parallel bitwise computation within tight on-chip resource budgets. On CORe50, ImageHD achieves up to 40.4x (4.84x) speedup and 383x (105.1x) energy efficiency over optimized CPU (GPU) baselines, demonstrating the practicality of HDC-enabled continual learning for real-time edge AI.", "date": "2026-04-23", "categories": []}
{"id": "2604.21930", "title": "Temporal Taskification in Streaming Continual Learning: A Source of Evaluation Instability", "abstract": "Streaming Continual Learning (CL) typically converts a continuous stream into a sequence of discrete tasks through temporal partitioning. We argue that this temporal taskification step is not a neutral preprocessing choice, but a structural component of evaluation: different valid splits of the same stream can induce different CL regimes and therefore different benchmark conclusions. To study this effect, we introduce a taskification-level framework based on plasticity and stability profiles, a profile distance between taskifications, and Boundary-Profile Sensitivity (BPS), which diagnoses how strongly small boundary perturbations alter the induced regime before any CL model is trained. We evaluate continual finetuning, Experience Replay, Elastic Weight Consolidation, and Learning without Forgetting on network traffic forecasting with CESNET-Timeseries24, keeping the stream, model, and training budget fixed while varying only the temporal taskification. Across 9-, 30-, and 44-day splits, we observe substantial changes in forecasting error, forgetting, and backward transfer, showing that taskification alone can materially affect CL evaluation. We further find that shorter taskifications induce noisier distribution-level patterns, larger structural distances, and higher BPS, indicating greater sensitivity to boundary perturbations. These results show that benchmark conclusions in streaming CL depend not only on the learner and the data stream, but also on how that stream is taskified, motivating temporal taskification as a first-class evaluation variable.", "date": "2026-04-23", "categories": []}
{"id": "2510.15751", "title": "SAMix: Calibrated and Accurate Continual Learning via Sphere-Adaptive Mixup and Neural Collapse", "abstract": "While most continual learning methods focus on mitigating forgetting and improving accuracy, they often overlook the critical aspect of network calibration, despite its importance. Neural collapse, a phenomenon where last-layer features collapse to their class means, has demonstrated advantages in continual learning by reducing feature-classifier misalignment. Few works aim to improve the calibration of continual models for more reliable predictions. Our work goes a step further by proposing a novel method that not only enhances calibration but also improves performance by reducing overconfidence, mitigating forgetting, and increasing accuracy. We introduce Sphere-Adaptive Mixup (SAMix), an adaptive mixup strategy tailored for neural collapse-based methods. SAMix adapts the mixing process to the geometric properties of feature spaces under neural collapse, ensuring more robust regularization and alignment. Experiments show that SAMix significantly boosts performance, surpassing SOTA methods in continual learning while also improving model calibration. SAMix enhances both across-task accuracy and the broader reliability of predictions, making it a promising advancement for robust continual learning systems.", "date": "2025-10-17", "categories": []}
{"id": "2604.20087", "title": "SkillLearnBench: Benchmarking Continual Learning Methods for Agent Skill Generation on Real-World Tasks", "abstract": "Skills have become the de facto way to enable LLM agents to perform complex real-world tasks with customized instructions, workflows, and tools, but how to learn them automatically and effectively remains unclear. We introduce SkillLearnBench, the first benchmark for evaluating continual skill learning methods, comprising 20 verified, skill-dependent tasks across 15 sub-domains derived from a real-world skill taxonomy , evaluated at three levels: skill quality, execution trajectory, and task outcome. Using this benchmark, we evaluate recent continual learning techniques, those leveraging one-shot, self/teacher feedback, and skill creator to generate skills from agent experiences. We find that all continual learning methods improve over the no-skill baseline, yet consistent gains remain elusive: no method leads across all tasks and LLMs, and scaling to stronger LLMs does not reliably help. Continual learning improves tasks with clear, reusable workflows but struggles on open-ended tasks, and using stronger LLM backbones does not consistently produce better skills. Our analysis also revealed that multiple iterations in continual learning facilitate genuine improvement via external feedback, whereas self-feedback alone induces recursive drift. Our data and code are open-source at https://github.com/cxcscmu/SkillLearnBench to enable further studies of automatic skill generation and continual learning techniques.", "date": "2026-04-22", "categories": []}
{"id": "2604.20720", "title": "COMPASS: COntinual Multilingual PEFT with Adaptive Semantic Sampling", "abstract": "Large language models (LLMs) often exhibit performance disparities across languages, with naive multilingual fine-tuning frequently degrading performance due to negative cross-lingual interference. To address this, we introduce COMPASS (COntinual Multilingual PEFT with Adaptive Semantic Sampling), a novel data-centric framework for adapting LLMs to target languages. COMPASS leverages parameter-efficient fine-tuning (PEFT) by training lightweight, language-specific adapters on a judiciously selected subset of auxiliary multilingual data. The core of our method is a distribution-aware sampling strategy that uses multilingual embeddings and clustering to identify semantic gaps between existing training data and a target usage distribution. By prioritizing auxiliary data from under-represented semantic clusters, COMPASS maximizes positive cross-lingual transfer while minimizing interference. We extend this into a continual learning framework, COMPASS-ECDA, which monitors for data distribution shifts in production and dynamically updates adapters to prevent model staleness, balancing adaptation to new data with the preservation of existing knowledge. Across three different model architectures (Phi-4-Mini, Llama-3.1-8B, and Qwen2.5-7B) and multiple challenging multilingual benchmarks (Global-MMLU, MMLU-ProX), including unseen long-context tasks (OneRuler), we demonstrate that COMPASS consistently outperforms baseline methods guided by linguistic similarity, providing an effective, efficient, and sustainable solution for developing and maintaining high-performing multilingual models in dynamic environments.", "date": "2026-04-22", "categories": []}
{"id": "2604.20745", "title": "Lifecycle-Aware Federated Continual Learning in Mobile Autonomous Systems", "abstract": "Federated continual learning (FCL) allows distributed autonomous fleets to adapt collaboratively to evolving terrain types across extended mission lifecycles. However, current approaches face several key challenges: 1) they use uniform protection strategies that do not account for the varying sensitivities to forgetting on different network layers; 2) they focus primarily on preventing forgetting during training, without addressing the long-term effects of cumulative drift; and 3) they often depend on idealized simulations that fail to capture the real-world heterogeneity present in distributed fleets. In this paper, we propose a lifecycle-aware dual-timescale FCL framework that incorporates training-time (pre-forgetting) prevention and (post-forgetting) recovery. Under this framework, we design a layer-selective rehearsal strategy that mitigates immediate forgetting during local training, and a rapid knowledge recovery strategy that restores degraded models after long-term cumulative drift. We present a theoretical analysis that characterizes heterogeneous forgetting dynamics and establishes the inevitability of long-term degradation. Our experimental results show that this framework achieves up to 8.3\\% mIoU improvement over the strongest federated baseline and up to 31.7\\% over conventional fine-tuning. We also deploy the FCL framework on a real-world rover testbed to assess system-level robustness under realistic constraints; the testing results further confirm the effectiveness of our FCL design.", "date": "2026-04-22", "categories": []}
{"id": "2510.08564", "title": "How to Teach Large Multimodal Models New Skills", "abstract": "How can we teach large multimodal models (LMMs) new skills without erasing prior abilities? We study sequential fine-tuning on five target skills while monitoring general ability on eight held-out benchmarks across three model families. Surprisingly, we find that performance lost on held-out tasks after fine-tuning on one skill can partly recover when the model is subsequently tuned on a different skill. We trace this behavior to a measurable shift in the output token distribution, manifested through a simple counting-bias probe that shows the shift co-varies with forgetting. Guided by this insight, we identify two simple, robust tuning recipes that learn strongly while limiting drift: (i) updating only the self-attention projection layers (SA Proj., $\\Delta$ learning +24.9 / $\\Delta$ held-out forgetting -0.6), and (ii) updating only the MLP Gate&Up while freezing the Down projection (+30.5 / -2.1). Both substantially outperform full-LLM tuning (+31.8 / -23.3) in the learning-forgetting trade-off. We also compare against common forgetting mitigation methods: Learning without Forgetting (LwF), LoRA, Mixture-of-Experts, and weight-space interpolation (WiSE-FT), and find that our selective tuning recipes match or exceed their learning-stability balance while remaining simpler, requiring no replay, auxiliary parameters, or per-stage tuning. These results hold across LLaVA-OneVision, LLaVA-NeXT, and Qwen2.5-VL, confirming that the key to teaching LMMs new skills without forgetting lies in controlling output distribution shift by choosing which components to tune. Code will be made available.", "date": "2025-10-09", "categories": []}
{"id": "2604.17736", "title": "IncreFA: Breaking the Static Wall of Generative Model Attribution", "abstract": "As AI generative models evolve at unprecedented speed, image attribution has become a moving target. New diffusion, adversarial and autoregressive generators appear almost monthly, making existing watermark, classifier and inversion methods obsolete upon release. The core problem lies not in model recognition, but in the inability to adapt attribution itself. We introduce IncreFA, a framework that redefines attribution as a structured incremental learning problem, allowing the system to learn continuously as new generative models emerge. IncreFA departs from conventional incremental learning by exploiting the hierarchical relationships among generative architectures and coupling them with continual adaptation. It integrates two mutually reinforcing mechanisms: (1) Hierarchical Constraints, which encode architectural hierarchies through learnable orthogonal priors to disentangle family-level invariants from model-specific idiosyncrasies; and (2) a Latent Memory Bank, which replays compact latent exemplars and mixes them to generate pseudo-unseen samples, stabilising representation drift and enhancing open-set awareness. On the newly constructed Incremental Attribution Benchmark (IABench) covering 28 generative models released between 2022 and 2025, IncreFA achieves state-of-the-art attribution accuracy and 98.93% unseen detection under a temporally ordered open-set protocol. Code will be available at https://github.com/Ant0ny44/IncreFA.", "date": "2026-04-20", "categories": []}
{"id": "2604.18611", "title": "Neuromorphic Continual Learning for Sequential Deployment of Nuclear Plant Monitoring Systems", "abstract": "Anomaly detection in nuclear industrial control systems (ICS) requires continuous, energy-efficient monitoring across multiple subsystems that are often deployed at different stages of plant commissioning. When a conventional neural network is sequentially trained to monitor new subsystems, it catastrophically forgets previously learned anomaly patterns, a safety-critical failure mode. We present the first spiking neural network (SNN)-based anomaly detection system with continual learning for nuclear ICS, addressing both challenges simultaneously. Our approach introduces spike-encoded asynchronous sensor fusion, a delta-based encoding that converts heterogeneous sensor streams into sparse spike trains at rates dictated by each sensor's natural dynamics, achieving 92.7% input sparsity. We evaluate five continual learning strategies, including sequential fine-tuning, Elastic Weight Consolidation (EWC), Synaptic Intelligence (SI), experience replay, and a hybrid EWC+Replay approach, on the HAI 21.03 nuclear ICS security dataset across three sequentially deployed subsystems (boiler, turbine, water treatment). The hybrid EWC+Replay method achieves an average F1 score of 0.979 with near-zero average forgetting (AF = 0.000 single seed; 0.035 +/- 0.039 across three seeds), while requiring 12.6x fewer operations (an estimated 2.5x in energy based on published hardware specifications) than an equivalent artificial neural network. The system detects all tested attacks with a mean latency of 0.6 seconds. These results demonstrate that neuromorphic computing offers a viable path toward always-on, energy-efficient, and adaptable safety monitoring for next-generation nuclear facilities.", "date": "2026-04-13", "categories": []}
{"id": "2604.18857", "title": "Task Switching Without Forgetting via Proximal Decoupling", "abstract": "In continual learning, the primary challenge is to learn new information without forgetting old knowledge. A common solution addresses this trade-off through regularization, penalizing changes to parameters critical for previous tasks. In most cases, this regularization term is directly added to the training loss and optimized with standard gradient descent, which blends learning and retention signals into a single update and does not explicitly separate essential parameters from redundant ones. As task sequences grow, this coupling can over-constrain the model, limiting forward transfer and leading to inefficient use of capacity. We propose a different approach that separates task learning from stability enforcement via operator splitting. The learning step focuses on minimizing the current task loss, while a proximal stability step applies a sparse regularizer to prune unnecessary parameters and preserve task-relevant ones. This turns the stability-plasticity into a negotiated update between two complementary operators, rather than a conflicting gradient. We provide theoretical justification for the splitting method on the continual-learning objective, and demonstrate that our proposed solver achieves state-of-the-art results on standard benchmarks, improving both stability and adaptability without the need for replay buffers, Bayesian sampling, or meta-learning components.", "date": "2026-04-20", "categories": []}
{"id": "2604.19401", "title": "Revisiting Catastrophic Forgetting in Continual Knowledge Graph Embedding", "abstract": "Knowledge Graph Embeddings (KGEs) support a wide range of downstream tasks over Knowledge Graphs (KGs). In practice, KGs evolve as new entities and facts are added, motivating Continual Knowledge Graph Embedding (CKGE) methods that update embeddings over time. Current CKGE approaches address catastrophic forgetting (i.e., the performance degradation on previously learned tasks) primarily by limiting changes to existing embeddings.\n  However, we show that this view is incomplete. When new entities are introduced, their embeddings can interfere with previously learned ones, causing the model to predict them in place of previously correct answers. This phenomenon, which we call entity interference, has been largely overlooked and is not accounted for in current CKGE evaluation protocols. As a result, the assessment of catastrophic forgetting becomes misleading, and CKGE methods performance is systematically overestimated.\n  To address this issue, we introduce a corrected CKGE evaluation protocol that accounts for entity interference. Through experiments on multiple benchmarks, we show that ignoring this effect can lead to performance overestimation of up to 25%, particularly in scenarios with significant entity growth. We further analyze how different CKGE methods and KGE models are affected by the different sources of forgetting, and introduce a catastrophic forgetting metric tailored to CKGE.", "date": "2026-04-21", "categories": []}
{"id": "2604.19737", "title": "Safe Continual Reinforcement Learning in Non-stationary Environments", "abstract": "Reinforcement learning (RL) offers a compelling data-driven paradigm for synthesizing controllers for complex systems when accurate physical models are unavailable; however, most existing control-oriented RL methods assume stationarity and, therefore, struggle in real-world non-stationary deployments where system dynamics and operating conditions can change unexpectedly. Moreover, RL controllers acting in physical environments must satisfy safety constraints throughout their learning and execution phases, rendering transient violations during adaptation unacceptable. Although continual RL and safe RL have each addressed non-stationarity and safety, respectively, their intersection remains comparatively unexplored, motivating the study of safe continual RL algorithms that can adapt over the system's lifetime while preserving safety. In this work, we systematically investigate safe continual reinforcement learning by introducing three benchmark environments that capture safety-critical continual adaptation and by evaluating representative approaches from safe RL, continual RL, and their combinations. Our empirical results reveal a fundamental tension between maintaining safety constraints and preventing catastrophic forgetting under non-stationary dynamics, with existing methods generally failing to achieve both objectives simultaneously. To address this shortcoming, we examine regularization-based strategies that partially mitigate this trade-off and characterize their benefits and limitations. Finally, we outline key open challenges and research directions toward developing safe, resilient learning-based controllers capable of sustained autonomous operation in changing environments.", "date": "2026-04-21", "categories": []}
{"id": "2510.05573", "title": "On the Theory of Continual Learning with Gradient Descent for Neural Networks", "abstract": "Continual learning, the ability of a model to adapt to an ongoing sequence of tasks without forgetting earlier ones, is a central goal of artificial intelligence. To better understand its underlying mechanisms, we study the limitations of continual learning in a tractable yet representative setting. Specifically, we analyze one-hidden-layer quadratic neural networks trained by gradient descent on a sequence of XOR-cluster datasets with Gaussian noise, where different tasks correspond to clusters with orthogonal means. Our analysis is based on a tight characterization of gradient descent dynamics for the training loss, which yields explicit bounds on the rate of train-time forgetting as functions of the number of iterations, sample size, number of tasks, and hidden-layer width. We then leverage an algorithmic stability framework to bound the generalization gap, leading to corresponding guarantees on test-time forgetting. Together, our results provide the first closed-form guarantees for forgetting in continual learning with neural networks and show how key problem parameters jointly govern forgetting dynamics. Numerical experiments corroborate our theoretical results.", "date": "2025-10-07", "categories": []}
{"id": "2601.03938", "title": "FOREVER: Forgetting Curve-Inspired Memory Replay for Language Model Continual Learning", "abstract": "Continual learning (CL) for large language models (LLMs) aims to enable sequential knowledge acquisition without catastrophic forgetting. Memory replay methods are widely used for their practicality and effectiveness, but most rely on fixed, step-based heuristics that often misalign with the model's actual learning progress, since identical training steps can result in varying degrees of parameter change. Motivated by recent findings that LLM forgetting mirrors the Ebbinghaus human forgetting curve, we propose FOREVER (FORgEtting curVe-inspired mEmory Replay), a novel CL framework that aligns replay schedules with a model-centric notion of time. FOREVER defines model time using the magnitude of optimizer updates, allowing forgetting curve-inspired replay intervals to align with the model's internal evolution rather than raw training steps. Building on this approach, FOREVER incorporates a forgetting curve-based replay scheduler to determine when to replay and an intensity-aware regularization mechanism to adaptively control how to replay. Extensive experiments on three CL benchmarks and models ranging from 0.6B to 13B parameters demonstrate that FOREVER consistently mitigates catastrophic forgetting.", "date": "2026-01-07", "categories": []}
{"id": "2601.22012", "title": "Putting a Face to Forgetting: Continual Learning meets Mechanistic Interpretability", "abstract": "Catastrophic forgetting in continual learning is often measured at the performance or last-layer representation level, overlooking the underlying mechanisms. We introduce a mechanistic framework that offers a geometric interpretation of catastrophic forgetting as the result of transformations to the encoding of individual features. These transformations can lead to forgetting by reducing the allocated capacity of features or by disrupting their readout by downstream computations. Analysis of a tractable toy model formalizes this view, allowing us to identify best- and worst-case scenarios. Through experiments on this model, we empirically test our formal analysis and highlight the detrimental effect of depth. Finally, we demonstrate how our framework can be used in the analysis of practical models through the use of Crosscoders. We do so through a case study example of a Vision Transformer trained on sequential CIFAR-10. Our work provides a new, feature-centric vocabulary for continual learning.", "date": "2026-01-29", "categories": []}
{"id": "2604.16496", "title": "Gradient-Free Continual Learning in Spiking Neural Networks via Inter-Spike Interval Regularization", "abstract": "Continual learning, the ability to acquire new tasks sequentially without forgetting prior knowledge, is essential for deploying neural networks in dynamic real-world environments, from nuclear digital twin monitoring to grid-edge fault detection. Existing synaptic importance methods, such as Elastic Weight Consolidation (EWC) and Synaptic Intelligence (SI), rely on gradient computation, making them incompatible with neuromorphic hardware that lacks backpropagation support. We propose ISI-CV, the first gradient-free synaptic importance metric for SNN continual learning, derived from the Coefficient of Variation (CV) of Inter-Spike Intervals (ISIs). Neurons that fire regularly (low CV) encode stable, task-relevant features and are protected from overwriting; neurons with irregular firing are permitted to adapt freely. ISI-CV requires only spike time counters and integer arithmetic, all of which are native to every neuromorphic chip. We evaluate on four benchmarks of increasing difficulty: Split-MNIST, Permuted-MNIST, Split-FashionMNIST, and Split-N-MNIST using real Dynamic Vision Sensor (DVS) event data. Across three seeds, ISI-CV achieves zero forgetting (AF = 0.000 +/- 0.000) on Split-MNIST and Split-FashionMNIST, near-zero forgetting on Permuted-MNIST (AF = 0.001 +/- 0.000), and the highest accuracy with the lowest forgetting on real neuromorphic DVS data (AA = 0.820 +/- 0.012, AF = 0.221 +/- 0.014). On N-MNIST, gradient-based methods produce unreliable importance estimates and perform worse than no regularization; ISI-CV avoids this failure by design.", "date": "2026-04-14", "categories": []}
{"id": "2604.17089", "title": "Tree of Concepts: Interpretable Continual Learners in Non-Stationary Clinical Domains", "abstract": "Continual learning aims to update models under distribution shift without forgetting, yet many high-stakes deployments, such as healthcare, also require interpretability. In practice, models that adapt well (e.g., deep networks) are often opaque, while models that are interpretable (e.g., decision trees) are brittle under shift, making it difficult to achieve both properties simultaneously. In response, we propose Tree of Concepts, an interpretable continual learning framework that uses a shallow decision tree to define a fixed, rule-based concept interface and trains a concept bottleneck model to predict these concepts from raw features. Continual updates act on the concept extractor and label head while keeping concept semantics stable over time, yielding explanations that do not drift across sequential updates. On multiple tabular healthcare benchmarks under continual learning protocols, our method achieves a stronger stability-plasticity trade-off than existing baselines, including replay-enhanced variants. Our results suggest that structured concept interfaces can support continual adaptation while preserving a consistent audit interface in non-stationary, high-stakes domains.", "date": "2026-04-18", "categories": []}
{"id": "2604.17822", "title": "GR4CIL: Gap-compensated Routing for CLIP-based Class Incremental Learning", "abstract": "Class-Incremental Learning (CIL) aims to continuously acquire new categories while preserving previously learned knowledge. Recently, Contrastive Language-Image Pre-trained (CLIP) models have shown strong potential for CIL due to their powerful generalization ability. However, existing methods still face two key challenges: shared-parameter adaptation tends to cause old-knowledge drift, and task-specific knowledge organization often leads to poorly calibrated cross-task responses, making reliable routing difficult. To address these issues, we propose GR4CIL, a framework combining task discrimination and knowledge routing for CLIP-based CIL. GR4CIL preserves task-specific visual knowledge while maintaining an incrementally stable shared textual semantic space, thereby reducing interference across tasks. Moreover, we introduce an orthogonal compensation mechanism to mitigate modality-gap-induced bias, enhance within-task discrimination, and enlarge the score margin between the ground-truth task and competing tasks. As a result, GR4CIL enables more reliable task-aware routing over learned knowledge while retaining the zero-shot generalization capability. Experiments on multiple benchmarks show that GR4CIL consistently outperforms strong baselines.", "date": "2026-04-20", "categories": []}
{"id": "2604.18075", "title": "Enhancing Continual Learning of Vision-Language Models via Dynamic Prefix Weighting", "abstract": "We investigate recently introduced domain-class incremental learning scenarios for vision-language models (VLMs). Recent works address this challenge using parameter-efficient methods, such as prefix-tuning or adapters, which facilitate model adaptation to downstream tasks by incorporating task-specific information into input tokens through additive vectors. However, previous approaches often normalize the weights of these vectors, disregarding the fact that different input tokens require different degrees of adjustment. To overcome this issue, we propose Dynamic Prefix Weighting (DPW), a framework that dynamically assigns weights to prefixes, complemented by adapters. DPW consists of 1) a gating module that adjusts the weights of each prefix based on the importance of the corresponding input token, and 2) a weighting mechanism that derives adapter output weights as a residual of prefix-tuning weights, ensuring that adapters are utilized only when necessary. Experimental results demonstrate that our method achieves state-of-the-art performance in domain-class incremental learning scenarios for VLMs. The code is available at: https://github.com/YonseiML/dpw.", "date": "2026-04-20", "categories": []}
{"id": "2604.18270", "title": "Incremental learning for audio classification with Hebbian Deep Neural Networks", "abstract": "The ability of humans for lifelong learning is an inspiration for deep learning methods and in particular for continual learning. In this work, we apply Hebbian learning, a biologically inspired learning process, to sound classification. We propose a kernel plasticity approach that selectively modulates network kernels during incremental learning, acting on selected kernels to learn new information and on others to retain previous knowledge. Using the ESC-50 dataset, the proposed method achieves 76.3% overall accuracy over five incremental steps, outperforming a baseline without kernel plasticity (68.7%) and demonstrating significantly greater stability across tasks.", "date": "2026-04-20", "categories": []}
{"id": "2602.22479", "title": "Efficient Continual Learning in Language Models via Thalamically Routed Cortical Columns", "abstract": "Large language models deployed in the wild must adapt to evolving data, user behavior, and task mixtures without erasing previously acquired capabilities. In practice, this remains difficult: sequential updates induce catastrophic forgetting, while many stabilization methods rely on external procedures that are costly, brittle, or difficult to scale. We present TRC$^{2}$ (Thalamically Routed Cortical Columns), a decoder-only architecture that makes continual learning a property of the backbone itself. TRC$^{2}$ combines stacked cortical columns with a thalamic modulatory pathway for selective inter-column communication and a hippocampal pathway for event selective retrieval, delayed surprise-based writing, and replay-driven consolidation. This design localizes fast plasticity while preserving a slower stable computation pathway. We further introduce a causal memory-update scheme and an online replay controller that adjusts consolidation strength from measured forgetting. Across a task-sequential language-modeling stream over C4, WikiText-103, and GSM8K, TRC$^{2}$ consistently improves task-boundary modeling quality and substantially reduces cumulative forgetting relative to Transformer, Mamba, MoE, DeepSeek and continual learning baselines trained under the same pipeline. Ablations show that the thalamic and hippocampal components are central to the retention gains, while the full model remains competitive in throughput and training cost.", "date": "2026-02-25", "categories": []}
{"id": "2604.15678", "title": "HyCal: A Training-Free Prototype Calibration Method for Cross-Discipline Few-Shot Class-Incremental Learning", "abstract": "Pretrained Vision-Language Models (VLMs) like CLIP show promise in continual learning, but existing Few-Shot Class-Incremental Learning (FSCIL) methods assume homogeneous domains and balanced data distributions, limiting real-world applicability where data arises from heterogeneous disciplines with imbalanced sample availability and varying visual complexity. We identify Domain Gravity, a representational asymmetry where data imbalance across heterogeneous domains causes overrepresented or low-entropy domains to disproportionately influence the embedding space, leading to prototype drift and degraded performance on underrepresented or high-entropy domains. To address this, we introduce Cross-Discipline Variable Few-Shot Class-Incremental Learning (XD-VSCIL), a benchmark capturing real-world heterogeneity and imbalance where Domain Gravity naturally intensifies. We propose Hybrid Prototype Calibration (HyCal), a training-free method combining cosine similarity and Mahalanobis distance to capture complementary geometric properties-directional alignment and covariance-aware magnitude-yielding stable prototypes under imbalanced heterogeneous conditions. Operating on frozen CLIP embeddings, HyCal achieves consistent retention-adaptation improvements while maintaining efficiency. Experiments show HyCal effectively mitigates Domain Gravity and outperforms existing methods in imbalanced cross-domain incremental learning.", "date": "2026-04-17", "categories": []}
{"id": "2604.15814", "title": "Continual Hand-Eye Calibration for Open-world Robotic Manipulation", "abstract": "Hand-eye calibration through visual localization is a critical capability for robotic manipulation in open-world environments. However, most deep learning-based calibration models suffer from catastrophic forgetting when adapting into unseen data amongst open-world scene changes, while simple rehearsal-based continual learning strategy cannot well mitigate this issue. To overcome this challenge, we propose a continual hand-eye calibration framework, enabling robots to adapt to sequentially encountered open-world manipulation scenes through spatially replay strategy and structure-preserving distillation. Specifically, a Spatial-Aware Replay Strategy (SARS) constructs a geometrically uniform replay buffer that ensures comprehensive coverage of each scene pose space, replacing redundant adjacent frames with maximally informative viewpoints. Meanwhile, a Structure-Preserving Dual Distillation (SPDD) is proposed to decompose localization knowledge into coarse scene layout and fine pose precision, and distills them separately to alleviate both types of forgetting during continual adaptation. As a new manipulation scene arrives, SARS provides geometrically representative replay samples from all prior scenes, and SPDD applies structured distillation on these samples to retain previously learned knowledge. After training on the new scene, SARS incorporates selected samples from the new scene into the replay buffer for future rehearsal, allowing the model to continuously accumulate multi-scene calibration capability. Experiments on multiple public datasets show significant anti scene forgetting performance, maintaining accuracy on past scenes while preserving adaptation to new scenes, confirming the effectiveness of the framework.", "date": "2026-04-17", "categories": []}
{"id": "2209.14774", "title": "RECALL: Rehearsal-free Continual Learning for Object Classification", "abstract": "Convolutional neural networks show remarkable results in classification but struggle with learning new things on the fly. We present a novel rehearsal-free approach, where a deep neural network is continually learning new unseen object categories without saving any data of prior sequences. Our approach is called RECALL, as the network recalls categories by calculating logits for old categories before training new ones. These are then used during training to avoid changing the old categories. For each new sequence, a new head is added to accommodate the new categories. To mitigate forgetting, we present a regularization strategy where we replace the classification with a regression. Moreover, for the known categories, we propose a Mahalanobis loss that includes the variances to account for the changing densities between known and unknown categories. Finally, we present a novel dataset for continual learning, especially suited for object recognition on a mobile robot (HOWS-CL-25), including 150,795 synthetic images of 25 household object categories. Our approach RECALL outperforms the current state of the art on CORe50 and iCIFAR-100 and reaches the best performance on HOWS-CL-25.", "date": "2022-09-29", "categories": []}
{"id": "2211.16780", "title": "An Optimal Transport-driven Approach for Cultivating Latent Space in Online Incremental Learning", "abstract": "In online incremental learning, data continuously arrives with substantial distributional shifts, creating a significant challenge because previous samples have limited replay value when learning a new task. Prior research has typically relied on either a single adaptive centroid or multiple fixed centroids to represent each class in the latent space. However, such methods struggle when class data streams are inherently multimodal and require continual centroid updates. To overcome this, we introduce an online Mixture Model learning framework grounded in Optimal Transport theory (MMOT), where centroids evolve incrementally with new data. This approach offers two main advantages: (i) it provides a more precise characterization of complex data streams, and (ii) it enables improved class similarity estimation for unseen samples during inference through MMOT-derived centroids. Furthermore, to strengthen representation learning and mitigate catastrophic forgetting, we design a Dynamic Preservation strategy that regulates the latent space and maintains class separability over time. Experimental evaluations on benchmark datasets confirm the superior effectiveness of our proposed method.", "date": "2022-11-30", "categories": []}
{"id": "2604.11508", "title": "Not All Forgetting Is Equal: Architecture-Dependent Retention Dynamics in Fine-Tuned Image Classifiers", "abstract": "Fine-tuning pretrained image classifiers is standard practice, yet which individual samples are forgotten during this process, and whether forgetting patterns are stable or architecture dependent, remains unclear. Understanding these dynamics has direct implications for curriculum design, data pruning, and ensemble construction. We track per-sample correctness at every epoch during fine-tuning of ResNet-18 and DeiT-Small on a retinal OCT dataset (7 classes, 56:1 imbalance) and CUB-200-2011 (200 bird species), fitting Ebbinghaus-style exponential decay curves to each sample's retention trace. Five findings emerge. First, the two architectures forget fundamentally different samples: Jaccard overlap of the top 10 percent most-forgotten is 0.34 on OCTDL and 0.15 on CUB-200. Second, ViT forgetting is more structured (mean $R^2 = 0.74$) than CNN forgetting ($R^2 = 0.52$). Third, per-sample forgetting is stochastic across random seeds (Spearman $\\rho \\approx 0.01$), challenging the assumption that sample difficulty is an intrinsic property. Fourth, class-level forgetting is consistent and semantically interpretable: visually similar species are forgotten most, distinctive ones least. Fifth, a sample's loss after head warmup predicts its long-term decay constant ($\\rho = 0.30$ to $0.50$, $p < 10^{-45}$). These findings suggest that architectural diversity in ensembles provides complementary retention coverage, and that curriculum or pruning methods based on per-sample difficulty may not generalize across runs. A spaced repetition sampler built on these decay constants does not outperform random sampling, indicating that static scheduling cannot exploit unstable per-sample signals.", "date": "2026-04-13", "categories": []}
{"id": "2604.14336", "title": "Mistake gating leads to energy and memory efficient continual learning", "abstract": "Synaptic plasticity is metabolically expensive, yet animals continuously update their internal models without exhausting energy reserves. However, when artificial neural networks are trained, the network parameters are typically updated on every sample that is presented, even if the sample was classified correctly. Inspired by the human negativity bias and error-related negativity, we propose 'memorized mistake-gated learning' -- a biologically plausible plasticity rule where synaptic updates are strictly gated by current and past classification errors. This reduces the number of updates the network needs to make by $50\\%\\sim80\\%$. Mistake gating is particularly well suited in two cases: 1) For incremental learning where new knowledge is acquired on a background of pre-existing knowledge, 2) For online learning scenarios when data needs to be stored for later replay, as mistake-gating reduces storage buffer requirements. The algorithm can be implemented in a few lines of code, adds no hyper-parameters, and comes at negligible computational overhead. Learning on mistakes is an energy efficient and biologically relevant modification to commonly used learning rules that is well suited for continual learning.", "date": "2026-04-15", "categories": []}
{"id": "2604.14375", "title": "Modular Continual Learning via Zero-Leakage Reconstruction Routing and Autonomous Task Discovery", "abstract": "Catastrophic forgetting remains a primary hurdle in sequential task learning for artificial neural networks. We propose a silicon-native modular architecture that achieves structural parameter isolation using Task-Specific Experts and a distributed, outlier-based Gatekeeper. Moving beyond traditional sequential consolidation, our framework utilizes a Simultaneous Pipeline where Teacher learning, Student distillation, and Router manifold acquisition occur in parallel while raw data is present in a localized training session. This approach ensures computational efficiency and complies with privacy mandates like GDPR by deleting raw data as soon as a task is learned. We demonstrate that a Tight-Bottleneck Autoencoder (TB-AE) can effectively distinguish semantically crowded manifolds in high-dimensional latent spaces, overcoming the posterior collapse inherent to standard variational methods. By establishing strict topological boundaries, our TB-AE resolves latent space crowding in 4096-D LLM embeddings to provide a robust, unsupervised novelty signal. Furthermore, we validate an Autonomous Retrieval mechanism that confidently identifies returning manifolds, enabling stable lifelong learning without redundant module instantiation. Empirical results demonstrate that our ``Live Distillation'' approach acts as a natural regularizer, achieving strong retention across computer vision and natural language processing domains without suffering a student fidelity gap.", "date": "2026-04-15", "categories": []}
{"id": "2604.14519", "title": "CI-CBM: Class-Incremental Concept Bottleneck Model for Interpretable Continual Learning", "abstract": "Catastrophic forgetting remains a fundamental challenge in continual learning, in which models often forget previous knowledge when fine-tuned on a new task. This issue is especially pronounced in class incremental learning (CIL), which is the most challenging setting in continual learning. Existing methods to address catastrophic forgetting often sacrifice either model interpretability or accuracy. To address this challenge, we introduce ClassIncremental Concept Bottleneck Model (CI-CBM), which leverage effective techniques, including concept regularization and pseudo-concept generation to maintain interpretable decision processes throughout incremental learning phases. Through extensive evaluation on seven datasets, CI-CBM achieves comparable performance to black-box models and outperforms previous interpretable approaches in CIL, with an average 36% accuracy gain. CICBM provides interpretable decisions on individual inputs and understandable global decision rules, as shown in our experiments, thereby demonstrating that human understandable concepts can be maintained during incremental learning without compromising model performance. Our approach is effective in both pretrained and non-pretrained scenarios; in the latter, the backbone is trained from scratch during the first learning phase. Code is publicly available at github.com/importAmir/CI-CBM.", "date": "2026-04-16", "categories": []}
{"id": "2604.14779", "title": "AIM: Asymmetric Information Masking for Visual Question Answering Continual Learning", "abstract": "In continual visual question answering (VQA), existing Continual Learning (CL) methods are mostly built for symmetric, unimodal architectures. However, modern Vision-Language Models (VLMs) violate this assumption, as their trainable components are inherently asymmetric. This structural mismatch renders VLMs highly prone to catastrophic forgetting when learning from continuous data streams. Specifically, the asymmetry causes standard global regularization to favor the massive language decoder during optimization, leaving the smaller but critical visual projection layers highly vulnerable to interference. Consequently, this localized degradation leads to a severe loss of compositional reasoning capabilities. To address this, we propose Asymmetric Information Masking (AIM), which balances stability and plasticity by applying targeted masks based on modality-specific sensitivity. Experiments on VQA v2 and GQA under continual VQA settings show that AIM achieves state-of-the-art performance in both Average Performance (AP) and Average Forgetting (AF), while better preserving generalization to novel skill-concept compositions.", "date": "2026-04-16", "categories": []}
{"id": "2604.11064", "title": "A Faster Path to Continual Learning", "abstract": "Continual Learning (CL) aims to train neural networks on a dynamic stream of tasks without forgetting previously learned knowledge. Among optimization-based approaches, C-Flat has emerged as a promising solution due to its plug-and-play nature and its ability to encourage uniformly low-loss regions for both new and old tasks. However, C-Flat requires three additional gradient computations per iteration, imposing substantial overhead on the optimization process. In this work, we propose C-Flat Turbo, a faster yet stronger optimizer that significantly reduces the training cost. We show that the gradients associated with first-order flatness contain direction-invariant components relative to the proxy-model gradients, enabling us to skip redundant gradient computations in the perturbed ascent steps. Moreover, we observe that these flatness-promoting gradients progressively stabilize across tasks, which motivates a linear scheduling strategy with an adaptive trigger to allocate larger turbo steps for later tasks. Experiments show that C-Flat Turbo is 1.0$\\times$ to 1.25$\\times$ faster than C-Flat across a wide range of CL methods, while achieving comparable or even improved accuracy.", "date": "2026-04-13", "categories": []}
{"id": "2604.13460", "title": "From Order to Distribution: A Spectral Characterization of Forgetting in Continual Learning", "abstract": "A central challenge in continual learning is forgetting, the loss of performance on previously learned tasks induced by sequential adaptation to new ones. While forgetting has been extensively studied empirically, rigorous theoretical characterizations remain limited. A notable step in this direction is \\citet{evron2022catastrophic}, which analyzes forgetting under random orderings of a fixed task collection in overparameterized linear regression. We shift the perspective from order to distribution. Rather than asking how a fixed task collection behaves under random orderings, we study an exact-fit linear regime in which tasks are sampled i.i.d.\\ from a task distribution~$\\Pi$, and ask how the generating distribution itself governs forgetting. In this setting, we derive an exact operator identity for the forgetting quantity, revealing a recursive spectral structure. Building on this identity, we establish an unconditional upper bound, identify the leading asymptotic term, and, in generic nondegenerate cases, characterize the convergence rate up to constants. We further relate this rate to geometric properties of the task distribution, clarifying what drives slow or fast forgetting in this model.", "date": "2026-04-15", "categories": []}
{"id": "2604.13730", "title": "ReConText3D: Replay-based Continual Text-to-3D Generation", "abstract": "Continual learning enables models to acquire new knowledge over time while retaining previously learned capabilities. However, its application to text-to-3D generation remains unexplored. We present ReConText3D, the first framework for continual text-to-3D generation. We first demonstrate that existing text-to-3D models suffer from catastrophic forgetting under incremental training. ReConText3D enables generative models to incrementally learn new 3D categories from textual descriptions while preserving the ability to synthesize previously seen assets. Our method constructs a compact and diverse replay memory through text-embedding k-Center selection, allowing representative rehearsal of prior knowledge without modifying the underlying architecture. To systematically evaluate continual text-to-3D learning, we introduce Toys4K-CL, a benchmark derived from the Toys4K dataset that provides balanced and semantically diverse class-incremental splits. Extensive experiments on the Toys4K-CL benchmark show that ReConText3D consistently outperforms all baselines across different generative backbones, maintaining high-quality generation for both old and new classes. To the best of our knowledge, this work establishes the first continual learning framework and benchmark for text-to-3D generation, opening a new direction for incremental 3D generative modeling. Project page is available at: https://mauk95.github.io/ReConText3D/.", "date": "2026-04-15", "categories": []}
{"id": "2604.14016", "title": "MAny: Merge Anything for Multimodal Continual Instruction Tuning", "abstract": "Multimodal Continual Instruction Tuning (MCIT) is essential for sequential task adaptation of Multimodal Large Language Models (MLLMs) but is severely restricted by catastrophic forgetting. While existing literature focuses on the reasoning language backbone, in this work, we expose a critical yet neglected dual-forgetting phenomenon across both perception drift in Cross-modal Projection Space and reasoning collapse in Low-rank Parameter Space. To resolve this, we present \\textbf{MAny} (\\textbf{M}erge \\textbf{Any}thing), a framework that merges task-specific knowledge through \\textbf{C}ross-modal \\textbf{P}rojection \\textbf{M}erging (\\textbf{CPM}) and \\textbf{L}ow-rank \\textbf{P}arameter \\textbf{M}erging (\\textbf{LPM}). Specifically, CPM recovers perceptual alignment by adaptively merging cross-modal visual representations via visual-prototype guidance, ensuring accurate feature recovery during inference. Simultaneously, LPM eliminates mutual interference among task-specific low-rank modules by recursively merging low-rank weight matrices. By leveraging recursive least squares, LPM provides a closed-form solution that mathematically guarantees an optimal fusion trajectory for reasoning stability. Notably, MAny operates as a training-free paradigm that achieves knowledge merging via efficient CPU-based algebraic operations, eliminating additional gradient-based optimization beyond initial tuning. Our extensive evaluations confirm the superior performance and robustness of MAny across multiple MLLMs and benchmarks. Specifically, on the UCIT benchmark, MAny achieves significant leads of up to 8.57\\% and 2.85\\% in final average accuracy over state-of-the-art methods across two different MLLMs, respectively.", "date": "2026-04-15", "categories": []}
{"id": "2505.15467", "title": "Joint Flashback Adaptation for Forgetting-Resistant Instruction Tuning", "abstract": "Large language models have achieved remarkable success in various tasks. However, it is challenging for them to learn new tasks incrementally due to catastrophic forgetting. Existing approaches rely on experience replay, optimization constraints, or task differentiation, which encounter strict limitations in real-world scenarios. To address these issues, we propose Joint Flashback Adaptation. We first introduce flashbacks -- a limited number of prompts from old tasks -- when adapting to new tasks and constrain the deviations of the model outputs compared to the original one. We then interpolate latent tasks between flashbacks and new tasks to enable jointly learning relevant latent tasks, new tasks, and flashbacks, alleviating data sparsity in flashbacks and facilitating knowledge sharing for smooth adaptation. Our method requires only a limited number of flashbacks without access to the replay data and is task-agnostic. We conduct extensive experiments on state-of-the-art large language models across 1000+ instruction-following tasks, arithmetic reasoning tasks, and general reasoning tasks. The results demonstrate the superior performance of our method in improving generalization on new tasks and reducing forgetting in old tasks.", "date": "2025-05-21", "categories": []}
{"id": "2604.12526", "title": "Orthogonal Subspace Projection for Continual Machine Unlearning via SVD-Based LoRA", "abstract": "Continual machine unlearning aims to remove the influence of data that should no longer be retained, while preserving the usefulness of the model on everything else. This setting becomes especially difficult when deletion requests arrive sequentially, because the model must repeatedly adapt without erasing previously retained knowledge. Low-Rank Adaptation (LoRA) offers an efficient way to implement such updates, but naively combining many sequential LoRA modules leads to parameter collision, causing \\textit{strong interference} between tasks. We propose a static alternative based on Singular Value Decomposition (SVD)-guided orthogonal subspace projection. Our method constrains each new LoRA update during training so that it lies in the orthogonal complement of the subspaces used by earlier unlearning tasks. This preserves task isolation without requiring dynamic routing at deployment. Experiments on CIFAR-100 with ResNet-20 and on MNIST show stable behavior across long sequences of unlearning tasks. After thirty sequential unlearning tasks, state-of-the-art static fusion reduces retained accuracy from 60.39\\% to 12.70\\%, whereas the proposed in-training constrained optimization maintains baseline performance ($\\sim$58.1\\%) while preserving strong unlearning efficacy.", "date": "2026-04-14", "categories": []}
{"id": "2604.12686", "title": "BID-LoRA: A Parameter-Efficient Framework for Continual Learning and Unlearning", "abstract": "Recent advances in deep learning underscore the need for systems that can not only acquire new knowledge through Continual Learning (CL) but also remove outdated, sensitive, or private information through Machine Unlearning (MU). However, while CL methods are well-developed, MU techniques remain in early stages, creating a critical gap for unified frameworks that depend on both capabilities. We find that naively combining existing CL and MU approaches results in knowledge leakage a gradual degradation of foundational knowledge across repeated adaptation cycles. To address this, we formalize Continual Learning Unlearning (CLU) as a unified paradigm with three key goals: (i) precise deletion of unwanted knowledge, (ii) efficient integration of new knowledge while preserving prior information, and (iii) minimizing knowledge leakage across cycles. We propose Bi-Directional Low-Rank Adaptation (BID-LoRA), a novel framework featuring three dedicated adapter pathways-retain, new, and unlearn applied to attention layers, combined with escape unlearning that pushes forget-class embeddings to positions maximally distant from retained knowledge, updating only 5% of parameters. Experiments on CIFAR-100 show that BID-LoRA outperforms CLU baselines across multiple adaptation cycles. We further evaluate on CASIA-Face100, a curated face recognition subset, demonstrating practical applicability to real-world identity management systems where new users must be enrolled and withdrawn users removed.", "date": "2026-04-14", "categories": []}
{"id": "2604.12874", "title": "LIFE -- an energy efficient advanced continual learning agentic AI framework for frontier systems", "abstract": "The rapid advancement of AI has changed the character of HPC usage such as dimensioning, provisioning, and execution. Not only has energy demand been amplified, but existing rudimentary continual learning capabilities limit ability of AI to effectively manage HPCs. This paper reviews emerging directions beyond monolithic transformers, emphasizing agentic AI and brain inspired architectures as complementary paths toward sustainable, adaptive systems. We propose LIFE, a reasoning and Learning framework that is Incremental, Flexible, and Energy efficient that is implemented as an agent centric system rather than a single monolithic model. LIFE uniquely combines four components to realize self evolving network management and operations in HPCs. The components are an orchestrator, Agentic Context Engineering, a novel memory system, and information lattice learning. LIFE can also generalize to enable a variety of orthogonal use cases. We ground LIFE in a specific closed loop HPC operations example for detecting and mitigating latency spikes experienced by critical micro services running on a Kubernetes like cluster.", "date": "2026-04-14", "categories": []}
{"id": "2604.12941", "title": "Direct Discrepancy Replay: Distribution-Discrepancy Condensation and Manifold-Consistent Replay for Continual Face Forgery Detection", "abstract": "Continual face forgery detection (CFFD) requires detectors to learn emerging forgery paradigms without forgetting previously seen manipulations. Existing CFFD methods commonly rely on replaying a small amount of past data to mitigate forgetting. Such replay is typically implemented either by storing a few historical samples or by synthesizing pseudo-forgeries from detector-dependent perturbations. Under strict memory budgets, the former cannot adequately cover diverse forgery cues and may expose facial identities, while the latter remains strongly tied to past decision boundaries. We argue that the core role of replay in CFFD is to reinstate the distributions of previous forgery tasks during subsequent training. To this end, we directly condense the discrepancy between real and fake distributions and leverage real faces from the current stage to perform distribution-level replay. Specifically, we introduce Distribution-Discrepancy Condensation (DDC), which models the real-to-fake discrepancy via a surrogate factorization in characteristic-function space and condenses it into a tiny bank of distribution discrepancy maps. We further propose Manifold-Consistent Replay (MCR), which synthesizes replay samples through variance-preserving composition of these maps with current-stage real faces, yielding samples that reflect previous-task forgery cues while remaining compatible with current real-face statistics. Operating under an extremely small memory budget and without directly storing raw historical face images, our framework consistently outperforms prior CFFD baselines and significantly mitigates catastrophic forgetting. Replay-level privacy analysis further suggests reduced identity leakage risk relative to selection-based replay.", "date": "2026-04-14", "categories": []}
{"id": "2502.07432", "title": "CapyMOA: Efficient Machine Learning for Data Streams and Online Continual Learning in Python", "abstract": "CapyMOA is an open-source Python library for efficient machine learning on data streams and online continual learning. It provides a structured framework for real-time learning, supporting adaptive models that evolve over time. CapyMOA's architecture allows integration with frameworks such as MOA, scikit-learn and PyTorch, enabling the combination of high-performance online algorithms with modern deep learning techniques. By emphasizing efficiency, scalability, and usability, CapyMOA allows researchers and practitioners to tackle dynamic learning challenges across various domains. Website: https://capymoa.org. GitHub: https://github.com/adaptive-machine-learning/CapyMOA.", "date": "2025-02-11", "categories": []}
{"id": "2507.15640", "title": "Data Mixing Agent: Learning to Re-weight Domains for Continual Pre-training", "abstract": "Continual pre-training on small-scale task-specific data is an effective method for improving large language models in new target fields, yet it risks catastrophic forgetting of their original capabilities. A common solution is to re-weight training data mixtures from source and target fields on a domain space to achieve balanced performance. Previous domain reweighting strategies rely on manual designation with certain heuristics based on human intuition or empirical results. In this work, we prove that more general heuristics can be parameterized by proposing Data Mixing Agent, the first model-based, end-to-end framework that learns to re-weight domains. The agent learns generalizable heuristics through reinforcement learning on large quantities of data mixing trajectories with corresponding feedback from an evaluation environment. Experiments in continual pre-training on math reasoning show that Data Mixing Agent outperforms strong baselines in achieving balanced performance across source and target field benchmarks. Furthermore, it generalizes well across unseen source fields, target models, and domain spaces without retraining. Direct application to the code generation field also indicates its adaptability across target domains. Further analysis showcases the agents' well-aligned heuristics with human intuitions and their efficiency in achieving superior model performance with less source-field data.", "date": "2025-07-21", "categories": []}
{"id": "2601.03641", "title": "Agent-Dice: Disentangling Knowledge Updates via Geometric Consensus for Agent Continual Learning", "abstract": "Large Language Model (LLM)-based agents significantly extend the utility of LLMs by interacting with dynamic environments. However, enabling agents to continually learn new tasks without catastrophic forgetting remains a critical challenge, known as the stability-plasticity dilemma. In this work, we argue that this dilemma fundamentally arises from the failure to explicitly distinguish between common knowledge shared across tasks and conflicting knowledge introduced by task-specific interference. To address this, we propose Agent-Dice, a parameter fusion framework based on directional consensus evaluation. Concretely, Agent-Dice disentangles knowledge updates through a two-stage process: geometric consensus filtering to prune conflicting gradients, and curvature-based importance weighting to amplify shared semantics. We provide a rigorous theoretical analysis that establishes the validity of the proposed fusion scheme and offers insight into the origins of the stability-plasticity dilemma. Extensive experiments on GUI agents and tool-use agent domains demonstrate that Agent-Dice exhibits outstanding continual learning performance with minimal computational overhead and parameter updates. The codes are available at https://github.com/Wuzheng02/Agent-Dice.", "date": "2026-01-07", "categories": []}
{"id": "2601.13844", "title": "Optimal L2 Regularization in High-dimensional Continual Linear Regression", "abstract": "We study generalization in an overparameterized continual linear regression setting, where a model is trained with L2 (isotropic) regularization across a sequence of tasks. We derive a closed-form expression for the expected generalization loss in the high-dimensional regime that holds for arbitrary linear teachers. We demonstrate that isotropic regularization mitigates label noise under both single-teacher and multiple i.i.d. teacher settings, whereas prior work accommodating multiple teachers either did not employ regularization or used memory-demanding methods. Furthermore, we prove that the optimal fixed regularization strength scales nearly linearly with the number of tasks $T$, specifically as $T/\\ln T$. To our knowledge, this is the first such result in theoretical continual learning. Finally, we validate our theoretical findings through experiments on linear regression and neural networks, illustrating how this scaling law affects generalization and offering a practical recipe for the design of continual learning systems.", "date": "2026-01-20", "categories": []}
{"id": "2604.09576", "title": "AHC: Meta-Learned Adaptive Compression for Continual Object Detection on Memory-Constrained Microcontrollers", "abstract": "Deploying continual object detection on microcontrollers (MCUs) with under 100KB memory requires efficient feature compression that can adapt to evolving task distributions. Existing approaches rely on fixed compression strategies (e.g., FiLM conditioning) that cannot adapt to heterogeneous task characteristics, leading to suboptimal memory utilization and catastrophic forgetting. We introduce Adaptive Hierarchical Compression (AHC), a meta-learning framework featuring three key innovations: (1) true MAML-based compression that adapts via gradient descent to each new task in just 5 inner-loop steps, (2) hierarchical multi-scale compression with scale-aware ratios (8:1 for P3, 6.4:1 for P4, 4:1 for P5) matching FPN redundancy patterns, and (3) a dual-memory architecture combining short-term and long-term banks with importance-based consolidation under a hard 100KB budget. We provide formal theoretical guarantees bounding catastrophic forgetting as O({\\epsilon}{sq.root(T)} + 1/{sq.root(M)}) where {\\epsilon} is compression error, T is task count, and M is memory size. Experiments on CORe50, TiROD, and PASCAL VOC benchmarks with three standard baselines (Fine-tuning,EWC, iCaRL) demonstrate that AHC enables practical continual detection within a 100KB replay budget, achieving competitive accuracy through mean-pooled compressed feature replay combined with EWC regularization and feature distillation.", "date": "2026-02-24", "categories": []}
{"id": "2604.10586", "title": "Preventing Latent Rehearsal Decay in Online Continual SSL with SOLAR", "abstract": "This paper explores Online Continual Self-Supervised Learning (OCSSL), a scenario in which models learn from continuous streams of unlabeled, non-stationary data, where methods typically employ replay and fast convergence is a central desideratum. We find that OCSSL requires particular attention to the stability-plasticity trade-off: stable methods (e.g. replay with Reservoir sampling) are able to converge faster compared to plastic ones (e.g. FIFO buffer), but incur in performance drops under certain conditions. We explain this collapse phenomenon with the Latent Rehearsal Decay hypothesis, which attributes it to latent space degradation under excessive stability of replay. We introduce two metrics (Overlap and Deviation) that diagnose latent degradation and correlate with accuracy declines. Building on these insights, we propose SOLAR, which leverages efficient online proxies of Deviation to guide buffer management and incorporates an explicit Overlap loss, allowing SOLAR to adaptively managing plasticity. Experiments demonstrate that SOLAR achieves state-of-the-art performance on OCSSL vision benchmarks, with both high convergence speed and final performance.", "date": "2026-04-12", "categories": []}
{"id": "2604.11091", "title": "LDEPrompt: Layer-importance guided Dual Expandable Prompt Pool for Pre-trained Model-based Class-Incremental Learning", "abstract": "Prompt-based class-incremental learning methods typically construct a prompt pool consisting of multiple trainable key-prompts and perform instance-level matching to select the most suitable prompt embeddings, which has shown promising results. However, existing approaches face several limitations, including fixed prompt pools, manual selection of prompt embeddings, and strong reliance on the pretrained backbone for prompt selection. To address these issues, we propose a \\textbf{L}ayer-importance guided \\textbf{D}ual \\textbf{E}xpandable \\textbf{P}rompt Pool (\\textbf{LDEPrompt}), which enables adaptive layer selection as well as dynamic freezing and expansion of the prompt pool. Extensive experiments on widely used class-incremental learning benchmarks demonstrate that LDEPrompt achieves state-of-the-art performance, validating its effectiveness and scalability.", "date": "2026-04-13", "categories": []}
{"id": "2604.11112", "title": "Quantum-Gated Task-interaction Knowledge Distillation for Pre-trained Model-based Class-Incremental Learning", "abstract": "Class-incremental learning (CIL) aims to continuously accumulate knowledge from a stream of tasks and construct a unified classifier over all seen classes. Although pretrained models (PTMs) have shown promising performance in CIL, they still struggle with the entanglement of multi-task subspaces, leading to catastrophic forgetting when task routing parameters are poorly calibrated or task-level representations are rigidly fixed. To address this issue, we propose a novel Quantum-Gated Task-interaction Knowledge Distillation (QKD) framework that leverages quantum gating to guide inter-task knowledge transfer. Specifically, we introduce a quantum-gated task modulation gating mechanism to model the relational dependencies among task embedding, dynamically capturing the sample-to-task relevance for both joint training and inference across streaming tasks. Guided by the quantum gating outputs, we perform task-interaction knowledge distillation guided by these task-embedding-level correlation weights from old to new adapters, enabling the model to bridge the representation gaps between independent task subspaces. Extensive experiments demonstrate that QKD effectively mitigates forgetting and achieves state-of-the-art performance.", "date": "2026-04-13", "categories": []}
{"id": "2505.12318", "title": "Task-agnostic Low-rank Residual Adaptation for Efficient Federated Continual Fine-Tuning", "abstract": "Federated Parameter-Efficient Fine-Tuning (Fed-PEFT) enables lightweight adaptation of large pre-trained models in federated learning settings by updating only a small subset of parameters. However, Fed-PEFT methods typically assume a fixed label space and static downstream tasks, which is restrictive in realistic application scenarios where clients continuously encounter new classes over time. This leads to an emerging problem, known as \\emph{Federated Continual Fine-Tuning} (FCFT). In FCFT, clients collaboratively fine-tune a pre-trained model over a sequence of tasks, where each client observes disjoint sets of new classes over time, and task identity is unavailable at inference time. FCFT is challenging because it simultaneously suffers from severe forgetting under non-IID client data distributions, parameter growth and task-specific inference caused by task-wise modules, and aggregation inconsistency across heterogeneous clients. To address these challenges, we propose Federated Task-agnostic Low-rank Residual Adaptation (Fed-TaLoRA), a novel approach for efficient FCFT built on task-agnostic adaptation, post-aggregation model calibration, and strategic low-rank adaptation placement. Fed-TaLoRA continuously fine-tunes a single shared module across sequential tasks to avoid task-wise parameter growth, and further introduces a theoretically grounded residual weight update mechanism to calibrate the aggregated global model and improve aggregation fidelity. We provide a theoretical analysis of the convergence and aggregation behavior of Fed-TaLoRA. Extensive experiments on four benchmark datasets demonstrate that Fed-TaLoRA consistently outperforms strong baselines while reducing communication and computation costs significantly.", "date": "2025-05-18", "categories": []}
{"id": "2603.13804", "title": "Memory-efficient Continual Learning with Prototypical Exemplar Condensation", "abstract": "Rehearsal-based continual learning (CL) mitigates catastrophic forgetting by maintaining a subset of samples from previous tasks for replay. Existing studies primarily focus on optimizing memory storage through coreset selection strategies. While these methods are effective, they typically require storing a substantial number of samples per class (SPC), often exceeding 20, to maintain satisfactory performance. In this work, we propose to further compress the memory footprint by synthesizing and storing prototypical exemplars, which can form representative prototypes when passed through the feature extractor. Owing to their representative nature, these exemplars enable the model to retain previous knowledge using only a small number of samples while preserving privacy. Moreover, we introduce a perturbation-based augmentation mechanism that generates synthetic variants of previous data during training, thereby enhancing CL performance. Extensive evaluations on widely used benchmark datasets and settings demonstrate that the proposed algorithm achieves superior performance compared to existing baselines, particularly in scenarios involving large-scale datasets and a high number of tasks.", "date": "2026-03-14", "categories": []}
{"id": "2604.08617", "title": "From Selection to Scheduling: Federated Geometry-Aware Correction Makes Exemplar Replay Work Better under Continual Dynamic Heterogeneity", "abstract": "Exemplar replay has become an effective strategy for mitigating catastrophic forgetting in federated continual learning (FCL) by retaining representative samples from past tasks. Existing studies focus on designing sample-importance estimation mechanisms to identify information-rich samples. However, they typically overlook strategies for effectively utilizing the selected exemplars, which limits their performance under continual dynamic heterogeneity across clients and tasks. To address this issue, this paper proposes a Federated gEometry-Aware correcTion method, termed FEAT, which alleviates imbalance-induced representation collapse that drags rare-class features toward frequent classes across clients. Specifically, it consists of two key modules: 1) the Geometric Structure Alignment module performs structural knowledge distillation by aligning the pairwise angular similarities between feature representations and their corresponding Equiangular Tight Frame prototypes, which are fixed and shared across clients to serve as a class-discriminative reference structure. This encourages geometric consistency across tasks and helps mitigate representation drift; 2) the Energy-based Geometric Correction module removes task-irrelevant directional components from feature embeddings, which reduces prediction bias toward majority classes. This improves sensitivity to minority classes and enhances the model's robustness under class-imbalanced distributions.", "date": "2026-04-09", "categories": []}
{"id": "2604.09038", "title": "Towards Lifelong Aerial Autonomy: Geometric Memory Management for Continual Visual Place Recognition in Dynamic Environments", "abstract": "Robust geo-localization in changing environmental conditions is critical for long-term aerial autonomy. While visual place recognition (VPR) models perform well when airborne views match the training domain, adapting them to shifting distributions during sequential missions triggers catastrophic forgetting. Existing continual learning (CL) methods often fail here because geographic features exhibit severe intra-class variations. In this work, we formulate aerial VPR as a mission-based domain-incremental learning (DIL) problem and propose a novel heterogeneous memory framework. To respect strict onboard storage constraints, our \"Learn-and-Dispose\" pipeline decouples geographic knowledge into static satellite anchors (preserving global geometric priors) and a dynamic experience replay buffer (retaining domain-specific features). We introduce a spatially-constrained allocation strategy that optimizes buffer selection based on sample difficulty or feature space diversity. To facilitate systematic assessment, we provide three evaluation criteria and a comprehensive benchmark derived from 21 diverse mission sequences. Extensive experiments demonstrate that our architecture significantly boosts spatial generalization; our diversity-driven buffer selection outperforms the random baseline by 7.8% in knowledge retention. Unlike class-mean preservation methods that fail in unstructured environments, maximizing structural diversity achieves a superior plasticity-stability balance and ensures order-agnostic robustness across randomized sequences. These results prove that maintaining structural feature coverage is more critical than sample difficulty for resolving catastrophic forgetting in lifelong aerial autonomy.", "date": "2026-04-10", "categories": []}
{"id": "2604.07399", "title": "Critical Patch-Aware Sparse Prompting with Decoupled Training for Continual Learning on the Edge", "abstract": "Continual learning (CL) on edge devices requires not only high accuracy but also training-time efficiency to support on-device adaptation under strict memory and computational constraints. While prompt-based continual learning (PCL) is parameter-efficient and achieves competitive accuracy, prior work has focused mainly on accuracy or inference-time performance, often overlooking the memory and computational costs of on-device training. In this paper, we propose CPS-Prompt, a critical patch-aware sparse prompting framework that explicitly targets training-time memory usage and computational cost by integrating critical patch sampling (CPS) for task-aware token reduction and decoupled prompt and classifier training (DPCT) to reduce backpropagation overhead. Experiments on three public benchmarks and real edge hardware show that CPS-Prompt improves peak memory, training time, and energy efficiency by about 1.6x over the balanced CODA-Prompt baseline, while maintaining accuracy within 2% of the state-of-the-art C-Prompt on average and remaining competitive with CODA-Prompt in accuracy. The code is available at https://github.com/laymond1/cps-prompt.", "date": "2026-04-08", "categories": []}
{"id": "2604.08159", "title": "Face-D(^2)CL: Multi-Domain Synergistic Representation with Dual Continual Learning for Facial DeepFake Detection", "abstract": "The rapid advancement of facial forgery techniques poses severe threats to public trust and information security, making facial DeepFake detection a critical research priority. Continual learning provides an effective approach to adapt facial DeepFake detection models to evolving forgery patterns. However, existing methods face two key bottlenecks in real-world continual learning scenarios: insufficient feature representation and catastrophic forgetting. To address these issues, we propose Face-D(^2)CL, a framework for facial DeepFake detection. It leverages multi-domain synergistic representation to fuse spatial and frequency-domain features for the comprehensive capture of diverse forgery traces, and employs a dual continual learning mechanism that combines Elastic Weight Consolidation (EWC), which distinguishes parameter importance for real versus fake samples, and Orthogonal Gradient Constraint (OGC), which ensures updates to task-specific adapters do not interfere with previously learned knowledge. This synergy enables the model to achieve a dynamic balance between robust anti-forgetting capabilities and agile adaptability to emerging facial forgery paradigms, all without relying on historical data replay. Extensive experiments demonstrate that our method surpasses current SOTA approaches in both stability and plasticity, achieving 60.7% relative reduction in average detection error rate, respectively. On unseen forgery domains, it further improves the average detection AUC by 7.9% compared to the current SOTA method.", "date": "2026-04-09", "categories": []}
{"id": "2604.08336", "title": "Leveraging Complementary Embeddings for Replay Selection in Continual Learning with Small Buffers", "abstract": "Catastrophic forgetting remains a key challenge in Continual Learning (CL). In replay-based CL with severe memory constraints, performance critically depends on the sample selection strategy for the replay buffer. Most existing approaches construct memory buffers using embeddings learned under supervised objectives. However, class-agnostic, self-supervised representations often encode rich, class-relevant semantics that are overlooked. We propose a new method, Multiple Embedding Replay Selection, MERS, which replaces the buffer selection module with a graph-based approach that integrates both supervised and self-supervised embeddings. Empirical results show consistent improvements over SOTA selection strategies across a range of continual learning algorithms, with particularly strong gains in low-memory regimes. On CIFAR-100 and TinyImageNet, MERS outperforms single-embedding baselines without adding model parameters or increasing replay volume, making it a practical, drop-in enhancement for replay-based continual learning.", "date": "2026-04-09", "categories": []}
{"id": "2505.20049", "title": "Data-Free Class-Incremental Gesture Recognition with Prototype-Guided Pseudo Feature Replay", "abstract": "Gesture recognition is an important research area in the field of computer vision. Most gesture recognition efforts focus on close-set scenarios, thereby limiting the capacity to effectively handle unseen or novel gestures. We aim to address class-incremental gesture recognition, which entails the ability to accommodate new and previously unseen gestures over time. Specifically, we introduce a Prototype-Guided Pseudo Feature Replay (PGPFR) framework for data-free class-incremental gesture recognition. This framework comprises four components: Pseudo Feature Generation with Batch Prototypes (PFGBP), Variational Prototype Replay (VPR) for old classes, Truncated Cross-Entropy (TCE) for new classes, and Continual Classifier Re-Training (CCRT). To tackle the issue of catastrophic forgetting, the PFGBP dynamically generates a diversity of pseudo features in an online manner, leveraging class prototypes of old classes along with batch class prototypes of new classes. Furthermore, the VPR enforces consistency between the classifier's weights and the prototypes of old classes, leveraging class prototypes and covariance matrices to enhance robustness and generalization capabilities. The TCE mitigates the impact of domain differences of the classifier caused by pseudo features. Finally, the CCRT training strategy is designed to prevent overfitting to new classes and ensure the stability of features extracted from old classes. Extensive experiments conducted on two widely used gesture recognition datasets, namely SHREC 2017 3D and EgoGesture 3D, demonstrate that our approach outperforms existing state-of-the-art methods by 11.8\\% and 12.8\\% in terms of mean global accuracy, respectively. The code is available on https://github.com/sunao-101/PGPFR-3/.", "date": "2025-05-26", "categories": []}
{"id": "2604.07108", "title": "Information as Structural Alignment: A Dynamical Theory of Continual Learning", "abstract": "Catastrophic forgetting is not an engineering failure. It is a mathematical consequence of storing knowledge as global parameter superposition. Existing methods, such as regularization, replay, and frozen subnetworks, add external mechanisms to a shared-parameter substrate. None derives retention from the learning dynamics themselves.\n  This paper introduces the Informational Buildup Framework (IBF), an alternative substrate for continual learning, based on the premise that information is the achievement of structural alignment rather than stored content. In IBF, two equations govern the dynamics: a Law of Motion that drives configuration toward higher coherence, and Modification Dynamics that persistently deform the coherence landscape in response to localized discrepancies. Memory, agency, and self-correction arise from these dynamics rather than being added as separate modules.\n  We first demonstrate the full lifecycle in a transparent two-dimensional toy model, then validate across three domains: a controlled non-stationary world, chess evaluated independently by Stockfish, and Split-CIFAR-100 with a frozen ViT encoder. Across all three, IBF achieves replay-superior retention without storing raw data. We observe near-zero forgetting on CIFAR-100 (BT = -0.004), positive backward transfer in chess (+38.5 cp), and 43% less forgetting than replay in the controlled domain. In chess, the framework achieves a mean behavioral advantage of +88.9 +/- 2.8 cp under independent evaluation, exceeding MLP and replay baselines.", "date": "2026-04-08", "categories": []}
{"id": "2506.21872", "title": "A Survey of Continual Reinforcement Learning", "abstract": "Reinforcement Learning (RL) is an important machine learning paradigm for solving sequential decision-making problems. Recent years have witnessed remarkable progress in this field due to the rapid development of deep neural networks. However, the success of RL currently relies on extensive training data and computational resources. In addition, RL's limited ability to generalize across tasks restricts its applicability in dynamic and real-world environments. With the arisen of Continual Learning (CL), Continual Reinforcement Learning (CRL) has emerged as a promising research direction to address these limitations by enabling agents to learn continuously, adapt to new tasks, and retain previously acquired knowledge. In this survey, we provide a comprehensive examination of CRL, focusing on its core concepts, challenges, and methodologies. Firstly, we conduct a detailed review of existing works, organizing and analyzing their metrics, tasks, benchmarks, and scenario settings. Secondly, we propose a new taxonomy of CRL methods, categorizing them into four types from the perspective of knowledge storage and/or transfer. Finally, our analysis highlights the unique challenges of CRL and provides practical insights into future directions.", "date": "2025-06-27", "categories": []}
{"id": "2511.01831", "title": "Routing-Based Continual Learning for Multimodal Large Language Models", "abstract": "Multimodal Large Language Models (MLLMs) struggle with continual learning, often suffering from catastrophic forgetting when adapting to sequential tasks. We introduce a routing-based architecture that integrates new capabilities while robustly preserving foundational knowledge. While Multi-Task Learning (MTL) offers a theoretical performance upper bound, it incurs a linearly scaling computational overhead as the number of tasks increases. In contrast, our method maintains fixed data and compute requirements regardless of the task sequence length. Across models ranging from 2B to 8B parameters, we demonstrate that our routing approach performs on par with MTL while retaining the training efficiency of sequential fine-tuning. Beyond merely mitigating forgetting, we observe that token-level routing facilitates cross-modal transfer, leveraging knowledge from one modality to bolster performance in another. Ablation studies confirm the approach's scalability: routing remains robust even with large expert pools and effectively capitalizes on task relatedness. Finally, we show that our method scales favorably, with larger models exhibiting minimal degradation compared to fully specialized fine-tuning.", "date": "2025-11-03", "categories": []}
{"id": "2604.05248", "title": "Improving Sparse Memory Finetuning", "abstract": "Large Language Models (LLMs) are typically static after training, yet real-world applications require continual adaptation to new knowledge without degrading existing capabilities. Standard approaches to updating models, like full finetuning or parameter-efficient methods (e.g., LoRA), face a fundamental trade-off: catastrophic forgetting. They modify shared dense representations, causing interference across tasks. Sparse Memory Finetuning (SMF) offers a promising alternative by localizing updates to a small subset of parameters in explicit memory layers. In this work, we present an open-source pipeline to retrofit existing pretrained models (Qwen-2.5-0.5B) with sparse memory modules, enabling effective continual learning on consumer hardware. We extend prior work by introducing a theoretically grounded slot-selection mechanism based on Kullback-Leibler (KL) divergence, which prioritizes memory updates for informationally \"surprising\" tokens relative to a background distribution. Our experiments demonstrate that our retrofitted models can acquire new factual knowledge with minimal forgetting of held-out capabilities, validating the sparse update hypothesis in a practical setting.", "date": "2026-04-06", "categories": []}
{"id": "2603.28858", "title": "OptiMer: Optimal Distribution Vector Merging Is Better than Data Mixing for Continual Pre-Training", "abstract": "Continual pre-training is widely used to adapt LLMs to target languages and domains, yet the mixture ratio of training data remains a sensitive hyperparameter that is expensive to tune: they must be fixed before training begins, and a suboptimal choice can waste weeks of compute. In this work, we propose OptiMer, which decouples ratio selection from training: we train one CPT model per dataset, extract each model's distribution vector, which represents the parameter shift induced by that dataset, and search for optimal composition weights post-hoc via Bayesian optimization. Experiments on Gemma 3 27B across languages (Japanese, Chinese) and domains (Math, Code) show that OptiMer consistently outperforms data mixture and model averaging baselines with 15-35 times lower search cost. Key findings reveal that 1) the optimized weights can be interpreted as data mixture ratios, and retraining with these ratios improves data mixture CPT, and 2) the same vector pool can be re-optimized for a given objective without any retraining, producing target-tailored models on demand. Our work establishes that data mixture ratio selection, traditionally a pre-training decision, can be reformulated as a post-hoc optimization over distribution vectors, offering a more flexible paradigm for continual pre-training.", "date": "2026-03-30", "categories": []}
{"id": "2604.04154", "title": "Non-Equilibrium Stochastic Dynamics as a Unified Framework for Insight and Repetitive Learning: A Kramers Escape Approach to Continual Learning", "abstract": "Continual learning in artificial neural networks is fundamentally limited by the stability--plasticity dilemma: systems that retain prior knowledge tend to resist acquiring new knowledge, and vice versa. Existing approaches, most notably elastic weight consolidation~(EWC), address this empirically without a physical account of why plasticity eventually collapses as tasks accumulate. Separately, the distinction between sudden insight and gradual skill acquisition through repetitive practice has lacked a unified theoretical description. Here, we show that both problems admit a common resolution within non-equilibrium statistical physics. We model the state of a learning system as a particle evolving under Langevin dynamics on a double-well energy landscape, with the noise amplitude governed by a time-dependent effective temperature $T(t)$. The probability density obeys a Fokker--Planck equation, and transitions between metastable states are governed by the Kramers escape rate $k = (\\omega_0\\omega_b/2\\pi)\\,e^{-\\Delta E/T}$. We make two contributions. First, we identify the EWC penalty term as an energy barrier whose height grows linearly with the number of accumulated tasks, yielding an exponential collapse of the transition rate predicted analytically and confirmed numerically. Second, we show that insight and repetitive learning correspond to two qualitatively distinct temperature protocols within the same Fokker--Planck equation: insight events produce transient spikes in $T(t)$ that drive rapid barrier crossing, whereas repetitive practice operates at a modestly elevated but fixed temperature, achieving transitions through sustained stochastic diffusion. These results establish a physically grounded framework for understanding plasticity and its failure in continual learning systems, and suggest principled design criteria for adaptive noise schedules in artificial intelligence.", "date": "2026-04-05", "categories": []}
{"id": "2604.04420", "title": "Is Prompt Selection Necessary for Task-Free Online Continual Learning?", "abstract": "Task-free online continual learning has recently emerged as a realistic paradigm for addressing continual learning in dynamic, real-world environments, where data arrive in a non-stationary stream without clear task boundaries and can only be observed once. To consider such challenging scenarios, many recent approaches have employed prompt selection, an adaptive strategy that selects prompts from a pool based on input signals. However, we observe that such selection strategies often fail to select appropriate prompts, yielding suboptimal results despite additional training of key parameters. Motivated by this observation, we propose a simple yet effective SinglePrompt that eliminates the need for prompt selection and focuses on classifier optimization. Specifically, we simply (i) inject a single prompt into each self-attention block, (ii) employ a cosine similarity-based logit design to alleviate the forgetting effect inherent in the classifier weights, and (iii) mask logits for unexposed classes in the current minibatch. With this simple task-free design, our framework achieves state-of-the-art performance across various online continual learning benchmarks. Source code is available at https://github.com/efficient-learning-lab/SinglePrompt.", "date": "2026-04-06", "categories": []}
{"id": "2512.03537", "title": "Pushing the Limits of Distillation-Based Continual Learning via Classifier-Proximal Lightweight Plugins", "abstract": "Continual learning requires models to learn continuously while preserving prior knowledge under evolving data streams. Distillation-based methods are appealing for retaining past knowledge in a shared single-model framework with low storage overhead. However, they remain constrained by the stability-plasticity dilemma: knowledge acquisition and preservation are still optimized through coupled objectives, and existing enhancement methods do not alter this underlying bottleneck. To address this issue, we propose a plugin extension paradigm termed Distillation-aware Lightweight Components (DLC) for distillation-based CL. DLC deploys lightweight residual plugins into the base feature extractor's classifier-proximal layer, enabling semantic-level residual correction for better classification accuracy while minimizing disruption to the overall feature extraction process. During inference, plugin-enhanced representations are aggregated to produce classification predictions. To mitigate interference from non-target plugins, we further introduce a lightweight weighting unit that learns to assign importance scores to different plugin-enhanced representations. DLC could deliver a significant 8% accuracy gain on large-scale benchmarks while introducing only a 4% increase in backbone parameters, highlighting its exceptional efficiency. Moreover, DLC is compatible with other plug-and-play CL enhancements and delivers additional gains when combined with them.", "date": "2025-12-03", "categories": []}
{"id": "2604.02356", "title": "MLFCIL: A Multi-Level Forgetting Mitigation Framework for Federated Class-Incremental Learning in LEO Satellites", "abstract": "Low-Earth-orbit (LEO) satellite constellations are increasingly performing on-board computing. However, the continuous emergence of new classes under strict memory and communication constraints poses major challenges for collaborative training. Federated class-incremental learning (FCIL) enables distributed incremental learning without sharing raw data, but faces three LEO-specific challenges: non-independent and identically distributed data heterogeneity caused by orbital dynamics, amplified catastrophic forgetting during aggregation, and the need to balance stability and plasticity under limited resources. To tackle these challenges, we propose MLFCIL, a multi-level forgetting mitigation framework that decomposes catastrophic forgetting into three sources and addresses them at different levels: class-reweighted loss to reduce local bias, knowledge distillation with feature replay and prototype-guided drift compensation to preserve cross-task knowledge, and class-aware aggregation to mitigate forgetting during federation. In addition, we design a dual-granularity coordination strategy that combines round-level adaptive loss balancing with step-level gradient projection to further enhance the stability-plasticity trade-off. Experiments on the NWPU-RESISC45 dataset show that MLFCIL significantly outperforms baselines in both accuracy and forgetting mitigation, while introducing minimal resource overhead.", "date": "2026-03-14", "categories": []}
{"id": "2604.02633", "title": "Analytic Drift Resister for Non-Exemplar Continual Graph Learning", "abstract": "Non-Exemplar Continual Graph Learning (NECGL) seeks to eliminate the privacy risks intrinsic to rehearsal-based paradigms by retaining solely class-level prototype representations rather than raw graph examples for mitigating catastrophic forgetting. However, this design choice inevitably precipitates feature drift. As a nascent alternative, Analytic Continual Learning (ACL) capitalizes on the intrinsic generalization properties of frozen pre-trained models to bolster continual learning performance. Nonetheless, a key drawback resides in the pronounced attenuation of model plasticity. To surmount these challenges, we propose Analytic Drift Resister (ADR), a novel and theoretically grounded NECGL framework. ADR exploits iterative backpropagation to break free from the frozen pre-trained constraint, adapting to evolving task graph distributions and fortifying model plasticity. Since parameter updates trigger feature drift, we further propose Hierarchical Analytic Merging (HAM), performing layer-wise merging of linear transformations in Graph Neural Networks (GNNs) via ridge regression, thereby ensuring absolute resistance to feature drift. On this basis, Analytic Classifier Reconstruction (ACR) enables theoretically zero-forgetting class-incremental learning. Empirical evaluation on four node classification benchmarks demonstrates that ADR maintains strong competitiveness against existing state-of-the-art methods.", "date": "2026-04-03", "categories": []}
{"id": "2604.02650", "title": "Revealing the Learning Dynamics of Long-Context Continual Pre-training", "abstract": "Existing studies on Long-Context Continual Pre-training (LCCP) mainly focus on small-scale models and limited data regimes (tens of billions of tokens). We argue that directly migrating these small-scale settings to industrial-grade models risks insufficient adaptation and premature training termination. Furthermore, current evaluation methods rely heavily on downstream benchmarks (e.g., Needle-in-a-Haystack), which often fail to reflect the intrinsic convergence state and can lead to \"deceptive saturation\". In this paper, we present the first systematic investigation of LCCP learning dynamics using the industrial-grade Hunyuan-A13B (80B total parameters), tracking its evolution across a 200B-token training trajectory. Specifically, we propose a hierarchical framework to analyze LCCP dynamics across behavioral (supervised fine-tuning probing), probabilistic (perplexity), and mechanistic (attention patterns) levels. Our findings reveal: (1) Necessity of Massive Data Scaling: Training regimes of dozens of billions of tokens are insufficient for industrial-grade LLMs' LCCP (e.g., Hunyuan-A13B reaches saturation after training over 150B tokens). (2) Deceptive Saturation vs. Intrinsic Saturation: Traditional NIAH scores report \"fake saturation\" early, while our PPL-based analysis reveals continuous intrinsic improvements and correlates more strongly with downstream performance. (3) Mechanistic Monitoring for Training Stability: Retrieval heads act as efficient, low-resource training monitors, as their evolving attention scores reliably track LCCP progress and exhibit high correlation with SFT results. This work provides a comprehensive monitoring framework, evaluation system, and mechanistic interpretation for the LCCP of industrial-grade LLM.", "date": "2026-04-03", "categories": []}
{"id": "2604.02765", "title": "Towards Realistic Class-Incremental Learning with Free-Flow Increments", "abstract": "Class-incremental learning (CIL) is typically evaluated under predefined schedules with equal-sized tasks, leaving more realistic and complex cases unexplored. However, a practical CIL system should learns immediately when any number of new classes arrive, without forcing fixed-size tasks. We formalize this setting as Free-Flow Class-Incremental Learning (FFCIL), where data arrives as a more realistic stream with a highly variable number of unseen classes each step. It will make many existing CIL methods brittle and lead to clear performance degradation. We propose a model-agnostic framework for robust CIL learning under free-flow arrivals. It comprises a class-wise mean (CWM) objective that replaces sample frequency weighted loss with uniformly aggregated class-conditional supervision, thereby stabilizing the learning signal across free-flow class increments, as well as method-wise adjustments that improve robustness for representative CIL paradigms. Specifically, we constrain distillation to replayed data, normalize the scale of contrastive and knowledge transfer losses, and introduce Dynamic Intervention Weight Alignment (DIWA) to prevent over-adjustment caused by unstable statistics from small class increments. Experiments confirm a clear performance degradation across various CIL baselines under FFCIL, while our strategies yield consistent gains.", "date": "2026-04-03", "categories": []}
{"id": "2604.00653", "title": "Chameleons do not Forget: Prompt-Based Online Continual Learning for Next Activity Prediction", "abstract": "Predictive process monitoring (PPM) focuses on predicting future process trajectories, including next activity predictions. This is crucial in dynamic environments where processes change or face uncertainty. However, current frameworks often assume a static environment, overlooking dynamic characteristics and concept drifts. This results in catastrophic forgetting, where training while focusing merely on new data distribution negatively impacts the performance on previously learned data distributions. Continual learning addresses, among others, the challenges related to mitigating catastrophic forgetting. This paper proposes a novel approach called Continual Next Activity Prediction with Prompts (CNAPwP), which adapts the DualPrompt algorithm for next activity prediction to improve accuracy and adaptability while mitigating catastrophic forgetting. We introduce new datasets with recurring concept drifts, alongside a task-specific forgetting metric that measures the prediction accuracy gap between initial occurrence and subsequent task occurrences. Extensive testing on three synthetic and two real-world datasets representing several setups of recurrent drifts shows that CNAPwP achieves SOTA or competitive results compared to five baselines, demonstrating its potential applicability in real-world scenarios. An open-source implementation of our method, together with the datasets and results, is available at: https://github.com/SvStraten/CNAPwP.", "date": "2026-04-01", "categories": []}
{"id": "2303.08250", "title": "CHEEM: Continual Learning by Reuse, New, Adapt and Skip -- A Hierarchical Exploration-Exploitation Approach", "abstract": "To effectively manage the complexities of real-world dynamic environments, continual learning must incrementally acquire, update, and accumulate knowledge from a stream of tasks of different nature without suffering from catastrophic forgetting of prior knowledge. While this capability is innate to human cognition, it remains a significant challenge for modern deep learning systems. At the heart of this challenge lies the stability-plasticity dilemma: the need to balance leveraging prior knowledge, integrating novel information, and allocating model capacity adaptively based on task complexity and synergy. In this paper, we propose a novel exemplar-free class-incremental continual learning (ExfCCL) framework that addresses these issues through a Hierarchical Exploration-Exploitation (HEE) approach. The core of our method is a HEE-guided efficient neural architecture search (HEE-NAS) that enables a learning-to-adapt backbone via four primitive operations - reuse, new, adapt, and skip - thereby serving as an internal memory that dynamically updates selected components across streaming tasks. To address the task ID inference problem in ExfCCL, we exploit an external memory of task centroids proposed in the prior art. We term our method CHEEM (Continual Hierarchical-Exploration-Exploitation Memory). CHEEM is evaluated on the challenging MTIL and VDD benchmarks using both Tiny and Base Vision Transformers and a proposed holistic Figure-of-Merit (FoM) metric. It significantly outperforms state-of-the-art prompting-based continual learning methods, closely approaching full fine-tuning upper bounds. Furthermore, it learns adaptive model structures tailored to individual tasks in a semantically meaningful way. Our code is available at https://github.com/savadikarc/cheem .", "date": "2023-03-14", "categories": []}
{"id": "2602.08040", "title": "FIRE: Frobenius-Isometry Reinitialization for Balancing the Stability-Plasticity Tradeoff", "abstract": "Deep neural networks trained on nonstationary data must balance stability (i.e., retaining prior knowledge) and plasticity (i.e., adapting to new tasks). Standard reinitialization methods, which reinitialize weights toward their original values, are widely used but difficult to tune: conservative reinitializations fail to restore plasticity, while aggressive ones erase useful knowledge. We propose FIRE, a principled reinitialization method that explicitly balances the stability-plasticity tradeoff. FIRE quantifies stability through Squared Frobenius Error (SFE), measuring proximity to past weights, and plasticity through Deviation from Isometry (DfI), reflecting weight isotropy. The reinitialization point is obtained by solving a constrained optimization problem, minimizing SFE subject to DfI being zero, which is efficiently approximated by Newton-Schulz iteration. FIRE is evaluated on continual visual learning (CIFAR-10 with ResNet-18), language modeling (OpenWebText with GPT-0.1B), and reinforcement learning (HumanoidBench with SAC and Atari games with DQN). Across all domains, FIRE consistently outperforms both naive training without intervention and standard reinitialization methods, demonstrating effective balancing of the stability-plasticity tradeoff.", "date": "2026-02-08", "categories": []}
{"id": "2603.08426", "title": "Grow, Assess, Compress: Adaptive Backbone Scaling for Memory-Efficient Class Incremental Learning", "abstract": "Class Incremental Learning (CIL) poses a fundamental challenge: maintaining a balance between the plasticity required to learn new tasks and the stability needed to prevent catastrophic forgetting. While expansion-based methods effectively mitigate forgetting by adding task-specific parameters, they suffer from uncontrolled architectural growth and memory overhead. In this paper, we propose a novel dynamic scaling framework that adaptively manages model capacity through a cyclic \"GRow, Assess, ComprEss\" (GRACE) strategy. Crucially, we supplement backbone expansion with a novel saturation assessment phase that evaluates the utilization of the model's capacity. This assessment allows the framework to make informed decisions to either expand the architecture or compress the backbones into a streamlined representation, preventing parameter explosion. Experimental results demonstrate that our approach achieves state-of-the-art performance across multiple CIL benchmarks, while reducing memory footprint by up to a 73% compared to purely expansionist models.", "date": "2026-03-09", "categories": []}
{"id": "2604.00067", "title": "Temporal Memory for Resource-Constrained Agents: Continual Learning via Stochastic Compress-Add-Smooth", "abstract": "An agent that operates sequentially must incorporate new experience without forgetting old experience, under a fixed memory budget. We propose a framework in which memory is not a parameter vector but a stochastic process: a Bridge Diffusion on a replay interval $[0,1]$, whose terminal marginal encodes the present and whose intermediate marginals encode the past. New experience is incorporated via a three-step \\emph{Compress--Add--Smooth} (CAS) recursion. We test the framework on the class of models with marginal probability densities modeled via Gaussian mixtures of fixed number of components~$K$ in $d$ dimensions; temporal complexity is controlled by a fixed number~$L$ of piecewise-linear protocol segments whose nodes store Gaussian-mixture states. The entire recursion costs $O(LKd^2)$ flops per day -- no backpropagation, no stored data, no neural networks -- making it viable for controller-light hardware.\n  Forgetting in this framework arises not from parameter interference but from lossy temporal compression: the re-approximation of a finer protocol by a coarser one under a fixed segment budget. We find that the retention half-life scales linearly as $a_{1/2}\\approx c\\,L$ with a constant $c>1$ that depends on the dynamics but not on the mixture complexity~$K$, the dimension~$d$, or the geometry of the target family. The constant~$c$ admits an information-theoretic interpretation analogous to the Shannon channel capacity. The stochastic process underlying the bridge provides temporally coherent ``movie'' replay -- compressed narratives of the agent's history, demonstrated visually on an MNIST latent-space illustration. The framework provides a fully analytical ``Ising model'' of continual learning in which the mechanism, rate, and form of forgetting can be studied with mathematical precision.", "date": "2026-03-31", "categories": []}
{"id": "2604.00677", "title": "CL-VISTA: Benchmarking Continual Learning in Video Large Language Models", "abstract": "Video Large Language Models (Video-LLMs) require continual learning to adapt to non-stationary real-world data. However, existing benchmarks fall short of evaluating modern foundation models: many still rely on models without large-scale pre-training, and prevailing benchmarks typically partition a single dataset into sub-tasks, resulting in high task redundancy and negligible forgetting on pre-trained Video-LLMs. To address these limitations, we propose CL-VISTA, a benchmark tailored for continual video understanding of Video-LLMs. By curating 8 diverse tasks spanning perception, understanding, and reasoning, CL-VISTA induces substantial distribution shifts that effectively expose catastrophic forgetting. To systematically assess CL methods, we establish a comprehensive evaluation framework comprising 6 distinct protocols across 3 critical dimensions: performance, computational efficiency, and memory footprint. Notably, the performance dimension incorporates a general video understanding assessment to assess whether CL methods genuinely enhance foundational intelligence or merely induce task-specific overfitting. Extensive benchmarking of 10 mainstream CL methods reveals a fundamental trade-off: no single approach achieves universal superiority across all dimensions. Methods that successfully mitigate catastrophic forgetting tend to compromise generalization or incur prohibitive computational and memory overheads. We hope CL-VISTA provides critical insights for advancing continual learning in multimodal foundation models.", "date": "2026-04-01", "categories": []}
{"id": "2604.00820", "title": "Continual Vision-Language Learning for Remote Sensing: Benchmarking and Analysis", "abstract": "Current remote sensing vision-language models (RS VLMs) demonstrate impressive performance in image interpretation but rely on static training data, limiting their ability to accommodate continuously emerging sensing modalities and downstream tasks. This exposes a fundamental challenge: enabling RS VLMs to continually adapt without catastrophic forgetting. Despite its practical importance, the continual learning capability of RS VLMs remains underexplored, and no dedicated benchmark currently exists. In this work, we present CLeaRS, a comprehensive benchmark for continual vision-language learning in remote sensing. CLeaRS comprises 10 curated subsets with over 207k image-text pairs, spanning diverse interpretation tasks, sensing modalities, and application scenarios. We further define three evaluation protocols: long-horizon, modality-incremental, and task-incremental settings, to systematically assess continual adaptation. Extensive benchmarking of diverse vision-language models reveals catastrophic forgetting across all settings. Moreover, representative continual learning methods, when adapted to RS VLMs, exhibit limited effectiveness in handling task, instruction, and modality transitions. Our findings underscore the need for developing continual learning methods tailored to RS VLMs.", "date": "2026-04-01", "categories": []}
{"id": "2604.01116", "title": "ProTPS: Prototype-Guided Text Prompt Selection for Continual Learning", "abstract": "For continual learning, text-prompt-based methods leverage text encoders and learnable prompts to encode semantic features for sequentially arrived classes over time. A common challenge encountered by existing works is how to learn unique text prompts, which implicitly carry semantic information of new classes, so that the semantic features of newly arrived classes do not overlap with those of trained classes, thereby mitigating the catastrophic forgetting problem. To address this challenge, we propose a novel approach Prototype-guided Text Prompt Selection (ProTPS)'' to intentionally increase the training flexibility thus encouraging the learning of unique text prompts. Specifically, our ProTPS learns class-specific vision prototypes and text prompts. Vision prototypes guide the selection and learning of text prompts for each class. We first evaluate our ProTPS in both class incremental (CI) setting and cross-datasets continual (CDC) learning setting. Because our ProTPS achieves performance close to the upper bounds, we further collect a real-world dataset with 112 marine species collected over a span of six years, named Marine112, to bring new challenges to the community. Marine112 is authentically suited for the class and domain incremental (CDI) learning setting and is under natural long-tail distribution. The results under three settings show that our ProTPS performs favorably against the recent state-of-the-art methods. The implementation code and Marine112 dataset will be released upon the acceptance of our paper.", "date": "2026-04-01", "categories": []}
{"id": "2604.01152", "title": "Brainstacks: Cross-Domain Cognitive Capabilities via Frozen MoE-LoRA Stacks for Continual LLM Learning", "abstract": "We present Brainstacks, a modular architecture for continual multi-domain fine-tuning of large language models that packages domain expertise as frozen adapter stacks composing additively on a shared frozen base at inference. Five interlocking components: (1) MoE-LoRA with Shazeer-style noisy top-2 routing across all seven transformer projections under QLoRA 4-bit quantization with rsLoRA scaling; (2) an inner loop performing residual boosting by freezing trained stacks and adding new ones; (3) an outer loop training sequential domain-specific stacks with curriculum-ordered dependencies; (4) null-space projection via randomized SVD constraining new stacks to subspaces orthogonal to prior directions, achieving zero forgetting in isolation; (5) an outcome-based sigmoid meta-router trained on empirically discovered domain-combination targets that selectively weights stacks, enabling cross-domain composition. Two boundary experiments: (6) PSN pretraining on a randomly initialized model; (7) per-domain RL (DPO/GRPO) validating compatibility with post-SFT alignment. Validated on TinyLlama-1.1B (4 domains, 9 stacks) and Gemma 3 12B IT (5 domains, 10 stacks), MoE-LoRA achieves 2.5x faster convergence than parameter-matched single LoRA, residual boosting breaks through the single-stack ceiling, and the routed system recovers generation quality destroyed by ungated stack accumulation. The central finding: the outcome-based router discovers that domain stacks encode transferable cognitive primitives (instruction-following clarity, numerical reasoning, procedural logic, chain-of-thought structure) rather than domain-specific knowledge, with medical prompts routing to chat+math stacks in 97% of cases despite zero medical data in those stacks.", "date": "2026-04-01", "categories": []}
{"id": "2508.12692", "title": "Multi-Level Knowledge Distillation and Dynamic Self-Supervised Learning for Continual Learning", "abstract": "Class-incremental with repetition (CIR), where previously trained classes repeatedly introduced in future tasks, is a more realistic scenario than the traditional class incremental setup, which assumes that each task contains unseen classes. CIR assumes that we can easily access abundant unlabeled data from external sources, such as the Internet. Therefore, we propose two components that efficiently use the unlabeled data to ensure the high stability and the plasticity of models trained in CIR setup. First, we introduce multi-level knowledge distillation (MLKD) that distills knowledge from multiple previous models across multiple perspectives, including features and logits, so the model can maintain much various previous knowledge. Moreover, we implement dynamic self-supervised loss (SSL) to utilize the unlabeled data that accelerates the learning of new classes, while dynamic weighting of SSL keeps the focus of training to the primary task. Both of our proposed components significantly improve the performance in CIR setup, achieving 2nd place in the CVPR 5th CLVISION Challenge.", "date": "2025-08-18", "categories": []}
{"id": "2603.29133", "title": "Dual-Imbalance Continual Learning for Real-World Food Recognition", "abstract": "Visual food recognition in real-world dietary logging scenarios naturally exhibits severe data imbalance, where a small number of food categories appear frequently while many others occur rarely, resulting in long-tailed class distributions. In practice, food recognition systems often operate in a continual learning setting, where new categories are introduced sequentially over time. However, existing studies typically assume that each incremental step introduces a similar number of new food classes, which rarely happens in real world where the number of newly observed categories can vary significantly across steps, leading to highly uneven learning dynamics. As a result, continual food recognition exhibits a dual imbalance: imbalanced samples within each food class and imbalanced numbers of new food classes to learn at each incremental learning step. In this work, we introduce DIME, a Dual-Imbalance-aware Adapter Merging framework for continual food recognition. DIME learns lightweight adapters for each task using parameter-efficient fine-tuning and progressively integrates them through a class-count guided spectral merging strategy. A rank-wise threshold modulation mechanism further stabilizes the merging process by preserving dominant knowledge while allowing adaptive updates. The resulting model maintains a single merged adapter for inference, enabling efficient deployment without accumulating task-specific modules. Experiments on realistic long-tailed food benchmarks under our step-imbalanced setup show that the proposed method consistently improves by more than 3% over the strongest existing continual learning baselines. Code is available at https://github.com/xiaoyanzhang1/DIME.", "date": "2026-03-31", "categories": []}
{"id": "2301.12230", "title": "Continual Graph Learning: A Survey", "abstract": "Continual Graph Learning (CGL) enables models to incrementally learn from streaming graph-structured data without forgetting previously acquired knowledge. Experience replay is a common solution that reuses a subset of past samples during training. However, it may lead to information loss and privacy risks. Generative replay addresses these concerns by synthesizing informative subgraphs for rehearsal. Existing generative replay approaches often rely on graph condensation via distribution matching, which faces two key challenges: (1) the use of random feature encodings may fail to capture the characteristic kernel of the discrepancy metric, weakening distribution alignment; and (2) matching over a fixed small subgraph cannot guarantee low risk on previous tasks, as indicated by domain adaptation theory. To overcome these limitations, we propose an Adversarial Condensation based Generative Replay (ACGR) framwork. It reformulates graph condensation as a min-max optimization problem to achieve better distribution matching. Moreover, instead of learning a single subgraph, we learn its distribution, allowing for the generation of multiple samples and improved empirical risk minimization. Experiments on three benchmark datasets demonstrate that ACGR outperforms existing methods in both accuracy and stability.", "date": "2023-01-28", "categories": []}
{"id": "2410.05352", "title": "Recent Advances of Multimodal Continual Learning: A Comprehensive Survey", "abstract": "Continual learning (CL) aims to empower machine learning models to learn continually from new data, while building upon previously acquired knowledge without forgetting. As models have evolved from small to large pre-trained architectures, and from supporting unimodal to multimodal data, multimodal continual learning (MMCL) methods have recently emerged. The primary complexity of MMCL is that it extends beyond a simple stacking of unimodal CL methods. Such straightforward approaches often suffer from multimodal catastrophic forgetting, yielding unsatisfactory performance. In addition, MMCL introduces new challenges that unimodal CL methods fail to adequately address, including modality imbalance, complex modality interaction, high computational costs, and degradation of pre-trained zero-shot capability of multimodal backbones. In this work, we present the first comprehensive survey on MMCL. We provide essential background knowledge and MMCL settings, as well as a structured taxonomy of MMCL methods. We categorize MMCL methods into four categories, i.e., regularization-based, architecture-based, replay-based, and prompt-based methods, explaining their methodologies and highlighting their key innovations. Additionally, to prompt further research in this field, we summarize open MMCL datasets and benchmarks, provide an in-depth discussion, and discuss several promising future directions. We have also created a GitHub repository for indexing relevant MMCL papers and open resources available at https://github.com/LucyDYu/Awesome-Multimodal-Continual-Learning.", "date": "2024-10-07", "categories": []}
{"id": "2602.22601", "title": "$\\phi$-DPO: Fairness Direct Preference Optimization Approach to Continual Learning in Large Multimodal Models", "abstract": "Fairness in Continual Learning for Large Multimodal Models (LMMs) is an emerging yet underexplored challenge, particularly in the presence of imbalanced data distributions that can lead to biased model updates and suboptimal performance across tasks. While recent continual learning studies have made progress in addressing catastrophic forgetting, the problem of fairness caused the imbalanced data remains largely underexplored. This paper presents a novel Fairness Direct Preference Optimization (FaiDPO or $\\phi$-DPO) framework for continual learning in LMMs. In particular, we first propose a new continual learning paradigm based on Direct Preference Optimization (DPO) to mitigate catastrophic forgetting by aligning learning with pairwise preference signals. Then, we identify the limitations of conventional DPO in imbalanced data and present a new $\\phi$-DPO loss that explicitly addresses distributional biases. We provide a comprehensive theoretical analysis demonstrating that our approach addresses both forgetting and data imbalance. Additionally, to enable $\\phi$-DPO-based continual learning, we construct pairwise preference annotations for existing benchmarks in the context of continual learning. Extensive experiments and ablation studies show the proposed $\\phi$-DPO achieves State-of-the-Art performance across multiple benchmarks, outperforming prior continual learning methods of LMMs.", "date": "2026-02-26", "categories": []}
{"id": "2603.14354", "title": "Deconfounded Lifelong Learning for Autonomous Driving via Dynamic Knowledge Spaces", "abstract": "End-to-End autonomous driving (E2E-AD) systems face challenges in lifelong learning, including catastrophic forgetting, difficulty in knowledge transfer across diverse scenarios, and spurious correlations between unobservable confounders and true driving intents. To address these issues, we propose DeLL, a Deconfounded Lifelong Learning framework that integrates a Dirichlet process mixture model (DPMM) with the front-door adjustment mechanism from causal inference. The DPMM is employed to construct two dynamic knowledge spaces: a trajectory knowledge space for clustering explicit driving behaviors and an implicit feature knowledge space for discovering latent driving abilities. Leveraging the non-parametric Bayesian nature of DPMM, our framework enables adaptive expansion and incremental updating of knowledge without predefining the number of clusters, thereby mitigating catastrophic forgetting. Meanwhile, the front-door adjustment mechanism utilizes the DPMM-derived knowledge as valid mediators to deconfound spurious correlations, such as those induced by sensor noise or environmental changes, and enhances the causal expressiveness of the learned representations. Additionally, we introduce an evolutionary trajectory decoder that enables non-autoregressive planning. To evaluate the lifelong learning performance of E2E-AD, we propose new evaluation protocols and metrics based on Bench2Drive. Extensive evaluations in the closed-loop CARLA simulator demonstrate that our framework significantly improves adaptability to new driving scenarios and overall driving performance, while effectively retaining previous acquired knowledge.", "date": "2026-03-15", "categories": []}
{"id": "2603.26671", "title": "Mitigating Forgetting in Continual Learning with Selective Gradient Projection", "abstract": "As neural networks are increasingly deployed in dynamic environments, they face the challenge of catastrophic forgetting, the tendency to overwrite previously learned knowledge when adapting to new tasks, resulting in severe performance degradation on earlier tasks. We propose Selective Forgetting-Aware Optimization (SFAO), a dynamic method that regulates gradient directions via cosine similarity and per-layer gating, enabling controlled forgetting while balancing plasticity and stability. SFAO selectively projects, accepts, or discards updates using a tunable mechanism with efficient Monte Carlo approximation. Experiments on standard continual learning benchmarks show that SFAO achieves competitive accuracy with markedly lower memory cost, a 90$\\%$ reduction, and improved forgetting on MNIST datasets, making it suitable for resource-constrained scenarios.", "date": "2026-02-08", "categories": []}
{"id": "2603.27481", "title": "On Token's Dilemma: Dynamic MoE with Drift-Aware Token Assignment for Continual Learning of Large Vision Language Models", "abstract": "Multimodal Continual Instruction Tuning aims to continually enhance Large Vision Language Models (LVLMs) by learning from new data without forgetting previously acquired knowledge. Mixture of Experts (MoE) architectures naturally facilitate this by incrementally adding new experts and expanding routers while keeping the existing ones frozen. However, despite expert isolation, MoE-based continual learners still suffer from forgetting due to routing-drift: old-task tokens become mistakenly attracted to newly added experts, degrading performance on prior tasks. We analyze the failure mode at the token level and reveal the token's dilemma: ambiguous and old tokens in new-task data offer minimal learning benefit yet induce forgetting when routed to new experts, due to their ambiguous routing assignment during training. Motivated by this, we propose LLaVA-DyMoE, a dynamic MoE framework that incrementally expands the MoE with drift-aware token assignment. We characterize token types via their routing score distributions and apply targeted regularization. Specifically, a token-level assignment guidance steers ambiguous and old tokens away from new experts to preserve established routing patterns and alleviate routing-drift, while complementary routing score regularizations enforce expert-group separation and promote new-expert specialization. Extensive experiments demonstrate that our LLaVA-DyMoE effectively mitigates routing-drift-induced forgetting, achieving over a 7% gain in mean final accuracy and a 12% reduction in forgetting compared to baselines. The project page is https://zhaoc5.github.io/DyMoE.", "date": "2026-03-29", "categories": []}
{"id": "2603.27678", "title": "Prototype-Aligned Federated Soft-Prompts for Continual Web Personalization", "abstract": "Continual web personalization is essential for engagement, yet real-world non-stationarity and privacy constraints make it hard to adapt quickly without forgetting long-term preferences. We target this gap by seeking a privacy-conscious, parameter-efficient interface that controls stability-plasticity at the user/session level while tying user memory to a shared semantic prior. We propose ProtoFed-SP, a prompt-based framework that injects dual-timescale soft prompts into a frozen backbone: a fast, sparse short-term prompt tracks session intent, while a slow long-term prompt is anchored to a small server-side prototype library that is continually refreshed via differentially private federated aggregation. Queries are routed to Top-M prototypes to compose a personalized prompt. Across eight benchmarks, ProtoFed-SP improves NDCG@10 by +2.9% and HR@10 by +2.0% over the strongest baselines, with notable gains on Amazon-Books (+5.0% NDCG vs. INFER), H&M (+2.5% vs. Dual-LoRA), and Taobao (+2.2% vs. FedRAP). It also lowers forgetting (AF) and Steps-to-95% and preserves accuracy under practical DP budgets. Our contribution is a unifying, privacy-aware prompting interface with prototype anchoring that delivers robust continual personalization and offers a transparent, controllable mechanism to balance stability and plasticity in deployment.", "date": "2026-03-29", "categories": []}
{"id": "2603.27707", "title": "Low-Rank Adaptation Reduces Catastrophic Forgetting in Sequential Transformer Encoder Fine-Tuning: Controlled Empirical Evidence and Frozen-Backbone Representation Probes", "abstract": "Sequential fine-tuning of pretrained language encoders often overwrites previously acquired capabilities, but the forgetting behavior of parameter-efficient updates remains under-characterized. We present a controlled empirical study of Low-Rank Adaptation (LoRA) in sequential transformer encoder fine-tuning with companion representation probes that test a frozen-backbone explanation of its robustness. In five full-validation BERT-base reruns on an RTE->MRPC->CoLA->SST-2 sequence, full fine-tuning yields 19.9%+/-4.8% average forgetting, whereas standard LoRA (r=8, query/value modules) yields 0.6%+/-1.4% (paired t-test, p=0.002, Cohen's d_s=3.12). Task-level analyses confirm this reduction is not merely an aggregate effect. Secondary experiments on RoBERTa-base show the same pattern, and the strongest EWC baseline remains at 15.5%+/-1.4% forgetting. A six-task extension reveals that low average forgetting can hide strong task-level heterogeneity. Fine-grained freezing ablations show a marked forgetting drop once frozen parameters exceed roughly 95%, with classifier-only and shallow-adapter baselines approaching LoRA. Companion task-similarity probes in GPT-2 and RoBERTa show the same directional story: frozen-backbone regimes preserve higher inter-task similarity than full fine-tuning, gradual unfreezing weakens stability, and full fine-tuning exhibits its clearest divergence at the final transformer layer. These results support a restrained mechanistic interpretation: LoRA helps largely because backbone freezing preserves a more stable shared feature scaffold. We position standard LoRA as both a strong empirical baseline for sequential encoder adaptation and a useful probe of how selective plasticity shapes interference in transformer continual learning.", "date": "2026-03-29", "categories": []}
{"id": "2603.28455", "title": "FeDMRA: Federated Incremental Learning with Dynamic Memory Replay Allocation", "abstract": "In federated healthcare systems, Federated Class-Incremental Learning (FCIL) has emerged as a key paradigm, enabling continuous adaptive model learning among distributed clients while safeguarding data privacy. However, in practical applications, data across agent nodes within the distributed framework often exhibits non-independent and identically distributed (non-IID) characteristics, rendering traditional continual learning methods inapplicable. To address these challenges, this paper covers more comprehensive incremental task scenarios and proposes a dynamic memory allocation strategy for exemplar storage based on the data replay mechanism. This strategy fully taps into the inherent potential of data heterogeneity, while taking into account the performance fairness of all participating clients, thereby establishing a balanced and adaptive solution to mitigate catastrophic forgetting. Unlike the fixed allocation of client exemplar memory, the proposed scheme emphasizes the rational allocation of limited storage resources among clients to improve model performance. Furthermore, extensive experiments are conducted on three medical image datasets, and the results demonstrate significant performance improvements compared to existing baseline models.", "date": "2026-03-30", "categories": []}
{"id": "2603.28611", "title": "LACE: Loss-Adaptive Capacity Expansion for Continual Learning", "abstract": "Fixed representational capacity is a fundamental constraint in continual learning: practitioners must guess an appropriate model width before training, without knowing how many distinct concepts the data contains. We propose LACE (Loss-Adaptive Capacity Expansion), a simple online mechanism that expands a model's representational capacity during training by monitoring its own loss signal. When sustained loss deviation exceeds a threshold - indicating that the current capacity is insufficient for newly encountered data - LACE adds new dimensions to the projection layer and trains them jointly with existing parameters. Across synthetic and real-data experiments, LACE triggers expansions exclusively at domain boundaries (100% boundary precision, zero false positives), matches the accuracy of a large fixed-capacity model while starting from a fraction of its dimensions, and produces adapter dimensions that are collectively critical to performance (3% accuracy drop when all adapters removed). We further demonstrate unsupervised domain separation in GPT-2 activations via layer-wise clustering, showing a U-shaped separability curve across layers that motivates adaptive capacity allocation in deep networks. LACE requires no labels, no replay buffers, and no external controllers, making it suitable for on-device continual learning under resource constraints.", "date": "2026-03-30", "categories": []}
{"id": "2603.26190", "title": "Dual-Stage Invariant Continual Learning under Extreme Visual Sparsity", "abstract": "Continual learning seeks to maintain stable adaptation under non-stationary environments, yet this problem becomes particularly challenging in object detection, where most existing methods implicitly assume relatively balanced visual conditions. In extreme-sparsity regimes, such as those observed in space-based resident space object (RSO) detection scenarios, foreground signals are overwhelmingly dominated by background observations. Under such conditions, we analytically demonstrate that background-driven gradients destabilize the feature backbone during sequential domain shifts, causing progressive representation drift. This exposes a structural limitation of continual learning approaches relying solely on output-level distillation, as they fail to preserve intermediate representation stability. To address this, we propose a dual-stage invariant continual learning framework via joint distillation, enforcing structural and semantic consistency on both backbone representations and detection predictions, respectively, thereby suppressing error propagation at its source while maintaining adaptability. Furthermore, to regulate gradient statistics under severe imbalance, we introduce a sparsity-aware data conditioning strategy combining patch-based sampling and distribution-aware augmentation. Experiments on a high-resolution space-based RSO detection dataset show consistent improvement over established continual object detection methods, achieving an absolute gain of +4.0 mAP under sequential domain shifts.", "date": "2026-03-27", "categories": []}
{"id": "2603.26192", "title": "HAD: Heterogeneity-Aware Distillation for Lifelong Heterogeneous Learning", "abstract": "Lifelong learning aims to preserve knowledge acquired from previous tasks while incorporating knowledge from a sequence of new tasks. However, most prior work explores only streams of homogeneous tasks (\\textit{e.g.}, only classification tasks) and neglects the scenario of learning across heterogeneous tasks that possess different structures of outputs. In this work, we formalize this broader setting as lifelong heterogeneous learning (LHL). Departing from conventional lifelong learning, the task sequence of LHL spans different task types, and the learner needs to retain heterogeneous knowledge for different output space structures. To instantiate the LHL, we focus on LHL in the context of dense prediction (LHL4DP), a realistic and challenging scenario. To this end, we propose the Heterogeneity-Aware Distillation (HAD) method, an exemplar-free approach that preserves previously gained heterogeneous knowledge by self-distillation in each training phase. The proposed HAD comprises two complementary components, including a distribution-balanced heterogeneity-aware distillation loss to alleviate the global imbalance of prediction distribution and a salience-guided heterogeneity-aware distillation loss that concentrates learning on informative edge pixels extracted with the Sobel operator. Extensive experiments demonstrate that the proposed HAD method significantly outperforms existing methods in this new scenario.", "date": "2026-03-27", "categories": []}
{"id": "2603.18596", "title": "Elastic Weight Consolidation Done Right for Continual Learning", "abstract": "Weight regularization methods in continual learning (CL) alleviate catastrophic forgetting by assessing and penalizing changes to important model weights. Elastic Weight Consolidation (EWC) is a foundational and widely used approach within this framework that estimates weight importance based on gradients. However, it has consistently shown suboptimal performance. In this paper, we conduct a systematic analysis of importance estimation in EWC from a gradient-based perspective. For the first time, we find that EWC's reliance on the Fisher Information Matrix (FIM) results in gradient vanishing and inaccurate importance estimation in certain scenarios. Our analysis also reveals that Memory Aware Synapses (MAS), a variant of EWC, imposes unnecessary constraints on parameters irrelevant to prior tasks, termed the redundant protection. Consequently, both EWC and its variants exhibit fundamental misalignments in estimating weight importance, leading to inferior performance. To tackle these issues, we propose the Logits Reversal (LR) operation, a simple yet effective modification that rectifies EWC's importance estimation. Specifically, reversing the logit values during the calculation of FIM can effectively prevent both gradient vanishing and redundant protection. Extensive experiments across various CL tasks and datasets show that the proposed method significantly outperforms existing EWC and its variants. Therefore, we refer to it as EWC Done Right (EWC-DR). Code is available at https://github.com/scarlet0703/EWC-DR.", "date": "2026-03-19", "categories": []}
{"id": "2603.25115", "title": "When Sensing Varies with Contexts: Context-as-Transform for Tactile Few-Shot Class-Incremental Learning", "abstract": "Few-Shot Class-Incremental Learning (FSCIL) can be particularly susceptible to acquisition contexts with only a few labeled samples. A typical scenario is tactile sensing, where the acquisition context ({\\it e.g.}, diverse devices, contact state, and interaction settings) degrades performance due to a lack of standardization. In this paper, we propose Context-as-Transform FSCIL (CaT-FSCIL) to tackle the above problem. We decompose the acquisition context into a structured low-dimensional component and a high-dimensional residual component. The former can be easily affected by tactile interaction features, which are modeled as an approximately invertible Context-as-Transform family and handled via inverse-transform canonicalization optimized with a pseudo-context consistency loss. The latter mainly arises from platform and device differences, which can be mitigated with an Uncertainty-Conditioned Prototype Calibration (UCPC) that calibrates biased prototypes and decision boundaries based on context uncertainty. Comprehensive experiments on the standard benchmarks HapTex and LMT108 have demonstrated the superiority of the proposed CaT-FSCIL.", "date": "2026-03-26", "categories": []}
{"id": "2603.23729", "title": "Bi-CRCL: Bidirectional Conservative-Radical Complementary Learning with Pre-trained Foundation Models for Class-incremental Medical Image Analysis", "abstract": "Class-incremental learning (CIL) in medical image-guided diagnosis requires retaining prior diagnostic knowledge while adapting to newly emerging disease categories, which is critical for scalable clinical deployment. This problem is particularly challenging due to heterogeneous data and privacy constraints that prevent memory replay. Although pretrained foundation models (PFMs) have advanced general-domain CIL, their potential in medical imaging remains underexplored, where domain-specific adaptation is essential yet difficult due to anatomical complexity and inter-institutional heterogeneity. To address this gap, we conduct a systematic benchmark of recent PFM-based CIL methods and propose Bidirectional Conservative-Radical Complementary Learning (Bi-CRCL), a dual-learner framework inspired by complementary learning systems. Bi-CRCL integrates a conservative learner that preserves prior knowledge through stability-oriented updates and a radical learner that rapidly adapts to new categories via plasticity-oriented learning. A bidirectional interaction mechanism enables forward transfer and backward consolidation, allowing continual integration of new knowledge while mitigating catastrophic forgetting. During inference, outputs from both learners are adaptively fused for robust predictions. Experiments on five medical imaging datasets demonstrate consistent improvements over state-of-the-art methods under diverse settings, including cross-dataset shifts and varying task configurations.", "date": "2026-03-24", "categories": []}
{"id": "2411.00623", "title": "Replay-Free Continual Low-Rank Adaptation with Dynamic Memory", "abstract": "We revisit continual learning~(CL), which enables pre-trained vision transformers (ViTs) to sequentially fine-tune on new downstream tasks over time. However, as the scale of these models increases, catastrophic forgetting remains a more serious challenge. Recent studies highlight a crossover between CL techniques and parameter-efficient fine-tuning (PEFT), which focuses on fine-tuning only a small set of trainable parameters to adapt to downstream tasks, such as low-rank adaptation (LoRA). While LoRA achieves faster convergence and requires fewer trainable parameters, it has seldom been explored in the context of continual learning. To address this gap, we propose a novel PEFT-CL method called Dual Low-Rank Adaptation (DualLoRA), which introduces both an orthogonal LoRA adapter and a residual LoRA adapter parallel to pre-trained weights in each layer. These components are orchestrated by a dynamic memory mechanism to strike a balance between stability and plasticity. Additionally, we propose a scheme to predict task identity with confidence and calibrate the model's outputs accordingly. On ViT-based models, we demonstrate that DualLoRA offers significant advantages in accuracy, inference speed, and computation efficiency in training over existing CL methods across multiple benchmarks.", "date": "2024-11-01", "categories": []}
{"id": "2602.01976", "title": "FlyPrompt: Brain-Inspired Random-Expanded Routing with Temporal-Ensemble Experts for General Continual Learning", "abstract": "General continual learning (GCL) challenges intelligent systems to learn from single-pass, non-stationary data streams without clear task boundaries. While recent advances in continual parameter-efficient tuning (PET) of pretrained models show promise, they typically rely on multiple training epochs and explicit task cues, limiting their effectiveness in GCL scenarios. Moreover, existing methods often lack targeted design and fail to address two fundamental challenges in continual PET: how to allocate expert parameters to evolving data distributions, and how to improve their representational capacity under limited supervision. Inspired by the fruit fly's hierarchical memory system characterized by sparse expansion and modular ensembles, we propose FlyPrompt, a brain-inspired framework that decomposes GCL into two subproblems: expert routing and expert competence improvement. FlyPrompt introduces a randomly expanded analytic router for instance-level expert activation and a temporal ensemble of output heads to dynamically adapt decision boundaries over time. Extensive theoretical and empirical evaluations demonstrate FlyPrompt's superior performance, achieving up to 11.23%, 12.43%, and 7.62% gains over state-of-the-art baselines on CIFAR-100, ImageNet-R, and CUB-200, respectively. Our source code is available at https://github.com/AnAppleCore/FlyGCL.", "date": "2026-02-02", "categories": []}
{"id": "2603.23436", "title": "Similarity-Aware Mixture-of-Experts for Data-Efficient Continual Learning", "abstract": "Machine learning models often need to adapt to new data after deployment due to structured or unstructured real-world dynamics. The Continual Learning (CL) framework enables continuous model adaptation, but most existing approaches either assume each task contains sufficiently many data samples or that the learning tasks are non-overlapping. In this paper, we address the more general setting where each task may have a limited dataset, and tasks may overlap in an arbitrary manner without a priori knowledge. This general setting is substantially more challenging for two reasons. On the one hand, data scarcity necessitates effective contextualization of general knowledge and efficient knowledge transfer across tasks. On the other hand, unstructured task overlapping can easily result in negative knowledge transfer. To address the above challenges, we propose an adaptive mixture-of-experts (MoE) framework over pre-trained models that progressively establishes similarity awareness among tasks. Our design contains two innovative algorithmic components: incremental global pooling and instance-wise prompt masking. The former mitigates prompt association noise through gradual prompt introduction over time. The latter decomposes incoming task samples into those aligning with current prompts (in-distribution) and those requiring new prompts (out-of-distribution). Together, our design strategically leverages potential task overlaps while actively preventing negative mutual interference in the presence of per-task data scarcity. Experiments across varying data volumes and inter-task similarity show that our method enhances sample efficiency and is broadly applicable.", "date": "2026-03-24", "categories": []}
{"id": "2302.13001", "title": "Better Generative Replay for Continual Federated Learning", "abstract": "Federated learning is a technique that enables a centralized server to learn from distributed clients via communications without accessing the client local data. However, existing federated learning works mainly focus on a single task scenario with static data. In this paper, we introduce the problem of continual federated learning, where clients incrementally learn new tasks and history data cannot be stored due to certain reasons, such as limited storage and data retention policy. Generative replay based methods are effective for continual learning without storing history data, but adapting them for this setting is challenging. By analyzing the behaviors of clients during training, we find that the unstable training process caused by distributed training on non-IID data leads to a notable performance degradation. To address this problem, we propose our FedCIL model with two simple but effective solutions: model consolidation and consistency enforcement. Our experimental results on multiple benchmark datasets demonstrate that our method significantly outperforms baselines.", "date": "2023-02-25", "categories": []}
{"id": "2501.00725", "title": "Automatic Construction of Pattern Classifiers Capable of Continuous Incremental Learning and Unlearning Tasks Based on Compact-Sized Probabilistic Neural Network", "abstract": "This paper proposes a novel approach to pattern classification using a probabilistic neural network model. The strategy is based on a compact-sized probabilistic neural network capable of continuous incremental learning and unlearning tasks. The network is constructed/reconstructed using a simple, one-pass network-growing algorithm with no hyperparameter tuning. Then, given the training dataset, its structure and parameters are automatically determined and can be dynamically varied in continual incremental and decremental learning situations. The algorithm proposed in this work involves no iterative or arduous matrix-based parameter approximations but a simple data-driven updating scheme. Simulation results using nine publicly available databases demonstrate the effectiveness of this approach, showing that compact-sized probabilistic neural networks constructed have a much smaller number of hidden units compared to the original probabilistic neural network model and yet can achieve a similar classification performance to that of multilayer perceptron neural networks in standard classification tasks, while also exhibiting sufficient capability in continuous class incremental learning and unlearning tasks.", "date": "2025-01-01", "categories": []}
{"id": "2502.00618", "title": "DesCLIP: Robust Continual Learning via General Attribute Descriptions for VLM-Based Visual Recognition", "abstract": "Continual learning of vision-language models (VLMs) focuses on leveraging cross-modal pretrained knowledge to incrementally adapt to expanding downstream tasks and datasets, while tackling the challenge of knowledge forgetting. Existing research often focuses on connecting visual features with specific class text in downstream tasks, overlooking the latent relationships between general and specialized knowledge. Our findings reveal that forcing models to optimize inappropriate visual-text matches exacerbates forgetting of VLM's recognition ability. To tackle this issue, we propose DesCLIP, which leverages general attribute (GA) descriptions to guide the understanding of specific class objects, enabling VLMs to establish robust vision-GA-class trilateral associations rather than relying solely on vision-class connections. Specifically, we introduce a language assistant to generate concrete GA description candidates via proper request prompts. Then, an anchor-based embedding filter is designed to obtain highly relevant GA description embeddings, which are leveraged as the paired text embeddings for visual-textual instance matching, thereby tuning the visual encoder. Correspondingly, the class text embeddings are gradually calibrated to align with these shared GA description embeddings. Extensive experiments demonstrate the advancements and efficacy of our proposed method, with comprehensive empirical evaluations highlighting its superior performance in VLM-based recognition compared to existing continual learning methods.", "date": "2025-02-02", "categories": []}
{"id": "2506.05736", "title": "Generalized Incremental Learning under Concept Drift across Evolving Data Streams", "abstract": "Real-world data streams exhibit inherent non-stationarity characterized by concept drift, posing significant challenges for adaptive learning systems. While existing methods address isolated distribution shifts, they overlook the critical co-evolution of label spaces and distributions under limited supervision and persistent uncertainty. To address this, we formalize Generalized Incremental Learning under Concept Drift (GILCD), characterizing the joint evolution of distributions and label spaces in open-environment streaming contexts, and propose a novel framework called Calibrated Source-Free Adaptation (CSFA). First, CSFA introduces a training-free prototype calibration mechanism that dynamically fuses emerging prototypes with base representations, enabling stable new-class identification without optimization overhead. Second, we design a novel source-free adaptation algorithm, i.e., Reliable Surrogate Gap Sharpness-aware (RSGS) minimization. It integrates sharpness-aware perturbation loss optimization with surrogate gap minimization, while employing entropy-based uncertainty filtering to discard unreliable samples. This mechanism ensures robust distribution alignment and mitigates generalization degradation caused by uncertainties. Thus, CSFA establishes a unified framework for stable adaptation to evolving semantics and distributions in open-world streaming scenarios. Extensive experiments validate the superior performance and effectiveness of CSFA compared to SOTA approaches.", "date": "2025-06-06", "categories": []}
{"id": "2511.05421", "title": "Sharing the Learned Knowledge-base to Estimate Convolutional Filter Parameters for Continual Image Restoration", "abstract": "Continual learning is an emerging topic in the field of deep learning, where a model is expected to learn continuously for new upcoming tasks without forgetting previous experiences. This field has witnessed numerous advancements, but few works have been attempted in the direction of image restoration. Handling large image sizes and the divergent nature of various degradation poses a unique challenge in the restoration domain. However, existing works require heavily engineered architectural modifications for new task adaptation, resulting in significant computational overhead. Regularization-based methods are unsuitable for restoration, as different restoration challenges require different kinds of feature processing. In this direction, we propose a simple modification of the convolution layer to adapt the knowledge from previous restoration tasks without touching the main backbone architecture. Therefore, it can be seamlessly applied to any deep architecture without any structural modifications. Unlike other approaches, we demonstrate that our model can increase the number of trainable parameters without significantly increasing computational overhead or inference time. Experimental validation demonstrates that new restoration tasks can be introduced without compromising the performance of existing tasks. We also show that performance on new restoration tasks improves by adapting the knowledge from the knowledge base created by previous restoration tasks. The code is available at https://github.com/aupendu/continual-restore.", "date": "2025-11-07", "categories": []}
{"id": "2603.12055", "title": "Continual Learning with Vision-Language Models via Semantic-Geometry Preservation", "abstract": "Continual learning of pretrained vision-language models (VLMs) is prone to catastrophic forgetting, yet current approaches adapt to new tasks without explicitly preserving the cross-modal semantic geometry inherited from pretraining and previous stages, allowing new-task supervision to induce geometric distortion. We observe that the most pronounced drift tends to concentrate in vulnerable neighborhoods near the old-new semantic interface, where shared visual patterns are easily re-explained by new textual semantics. To address this under an exemplar-free constraint, we propose Semantic Geometry Preservation for Continual Learning (SeGP-CL). SeGP-CL first probes the drift-prone region by constructing a compact set of adversarial anchors with dual-targeted projected gradient descent (DPGD), which drives selected new-task seeds toward old-class semantics while remaining faithful in raw visual space. During training, we preserve cross-modal structure by anchor-guided cross-modal geometry distillation (ACGD), and stabilize the textual reference frame across tasks via a lightweight text semantic-geometry regularization (TSGR). After training, we estimate anchor-induced raw-space drift to transfer old visual prototypes and perform dual-path inference by fusing cross-modal and visual cues. Extensive experiments on five continual learning benchmarks demonstrate that SeGP-CL consistently improves stability and forward transfer, achieving state-of-the-art performance while better preserving semantic geometry of VLMs.", "date": "2026-03-12", "categories": []}
{"id": "2603.20292", "title": "HSI Image Enhancement Classification Based on Knowledge Distillation: A Study on Forgetting", "abstract": "In incremental classification tasks for hyperspectral images, catastrophic forgetting is an unavoidable challenge. While memory recall methods can mitigate this issue, they heavily rely on samples from old categories. This paper proposes a teacher-based knowledge retention method for incremental image classification. It alleviates model forgetting of old category samples by utilizing incremental category samples, without depending on old category samples. Additionally, this paper introduces a mask-based partial category knowledge distillation algorithm. By decoupling knowledge distillation, this approach filters out potentially misleading information that could misguide the student model, thereby enhancing overall accuracy. Comparative and ablation experiments demonstrate the proposed method's robust performance.", "date": "2026-03-19", "categories": []}
{"id": "2603.20648", "title": "A Multihead Continual Learning Framework for Fine-Grained Fashion Image Retrieval with Contrastive Learning and Exponential Moving Average Distillation", "abstract": "Most fine-grained fashion image retrieval (FIR) methods assume a static setting, requiring full retraining when new attributes appear, which is costly and impractical for dynamic scenarios. Although pretrained models support zero-shot inference, their accuracy drops without supervision, and no prior work explores class-incremental learning (CIL) for fine-grained FIR. We propose a multihead continual learning framework for fine-grained fashion image retrieval with contrastive learning and exponential moving average (EMA) distillation (MCL-FIR). MCL-FIR adopts a multi-head design to accommodate evolving classes across increments, reformulates triplet inputs into doublets with InfoNCE for simpler and more effective training, and employs EMA distillation for efficient knowledge transfer. Experiments across four datasets demonstrate that, beyond its scalability, MCL-FIR achieves a strong balance between efficiency and accuracy. It significantly outperforms CIL baselines under similar training cost, and compared with static methods, it delivers comparable performance while using only about 30% of the training cost. The source code is publicly available in https://github.com/Dr-LingXiao/MCL-FIR.", "date": "2026-03-21", "categories": []}
{"id": "2603.20898", "title": "Natural Gradient Descent for Online Continual Learning", "abstract": "Online Continual Learning (OCL) for image classification represents a challenging subset of Continual Learning, focusing on classifying images from a stream without assuming data independence and identical distribution (i.i.d). The primary challenge in this context is to prevent catastrophic forgetting, where the model's performance on previous tasks deteriorates as it learns new ones. Although various strategies have been proposed to address this issue, achieving rapid convergence remains a significant challenge in the online setting. In this work, we introduce a novel approach to training OCL models that utilizes the Natural Gradient Descent optimizer, incorporating an approximation of the Fisher Information Matrix (FIM) through Kronecker Factored Approximate Curvature (KFAC). This method demonstrates substantial improvements in performance across all OCL methods, particularly when combined with existing OCL tricks, on datasets such as Split CIFAR-100, CORE50, and Split miniImageNet.", "date": "2026-03-21", "categories": []}
{"id": "2603.21170", "title": "Pruned Adaptation Modules: A Simple yet Strong Baseline for Continual Foundation Models", "abstract": "The continual learning literature has rapidly shifted from traditional class incremental learning (CIL) techniques to foundation model (FM)-based CIL methods without a clear understanding of how these newer approaches compare to strong, lightweight convolutional baselines. This abrupt transition has created a substantial methodological gap, making it difficult to assess whether recent FM-based CIL progress reflects genuine advances or merely the absence of rigorous baselines. To address this gap, we introduce Pruned Adaptation Modules (PAM), a simple yet effective method that freezes the vast majority of the pre-trained ResNet while enabling scalable continual adaptation through sparse task-specific layers. PAM yields up to a ~5x reduction in trainable parameters and a ~6x reduction in total parameters, significantly reducing the cost of continual updates. Across diverse benchmarks, PAM consistently mitigates catastrophic forgetting and outperforms state-of-the-art FM-based CIL approaches. Our findings position PAM as a strong and transparent baseline that helps bridge the gap between traditional and FM-based CIL, guiding future research for a more accurate assessment of true progress in continual adaptation. The code can be found at: https://github.com/ElifCerenGokYildirim/PAM.", "date": "2026-03-22", "categories": []}
{"id": "2603.21484", "title": "Which Concepts to Forget and How to Refuse? Decomposing Concepts for Continual Unlearning in Large Vision-Language Models", "abstract": "Continual unlearning poses the challenge of enabling large vision-language models to selectively refuse specific image-instruction pairs in response to sequential deletion requests, while preserving general utility. However, sequential unlearning updates distort shared representations, creating spurious associations between vision-language pairs and refusal behaviors that hinder precise identification of refusal targets, resulting in inappropriate refusals. To address this challenge, we propose a novel continual unlearning framework that grounds refusal behavior in fine-grained descriptions of visual and textual concepts decomposed from deletion targets. We first identify which visual-linguistic concept combinations characterize each forget category through a concept modulator, then determine how to generate appropriate refusal responses via a mixture of refusal experts, termed refusers, each specialized for concept-aligned refusal generation. To generate concept-specific refusal responses across sequential tasks, we introduce a multimodal, concept-driven routing scheme that reuses refusers for tasks sharing similar concepts and adapts underutilized ones for novel concepts. Extensive experiments on vision-language benchmarks demonstrate that the proposed framework outperforms existing methods by generating concept-grounded refusal responses and preserving the general utility across unlearning sequences.", "date": "2026-03-23", "categories": []}
{"id": "2603.21708", "title": "Compensating Visual Insufficiency with Stratified Language Guidance for Long-Tail Class Incremental Learning", "abstract": "Long-tail class incremental learning (LT CIL) remains highly challenging because the scarcity of samples in tail classes not only hampers their learning but also exacerbates catastrophic forgetting under continuously evolving and imbalanced data distributions. To tackle these issues, we exploit the informativeness and scalability of language knowledge. Specifically, we analyze the LT CIL data distribution to guide large language models (LLMs) in generating a stratified language tree that hierarchically organizes semantic information from coarse to fine grained granularity. Building upon this structure, we introduce stratified adaptive language guidance, which leverages learnable weights to merge multi-scale semantic representations, thereby enabling dynamic supervisory adjustment for tail classes and alleviating the impact of data imbalance. Furthermore, we introduce stratified alignment language guidance, which exploits the structural stability of the language tree to constrain optimization and reinforce semantic visual alignment, thereby alleviating catastrophic forgetting. Extensive experiments on multiple benchmarks demonstrate that our method achieves state of the art performance.", "date": "2026-03-23", "categories": []}
{"id": "2511.15164", "title": "Multimodal Continual Instruction Tuning with Dynamic Gradient Guidance", "abstract": "Multimodal continual instruction tuning enables multimodal large language models to sequentially adapt to new tasks while building upon previously acquired knowledge. However, this continual learning paradigm faces the significant challenge of catastrophic forgetting, where learning new tasks leads to performance degradation on previous ones. In this paper, we introduce a novel insight into catastrophic forgetting by conceptualizing it as a problem of missing gradients from old tasks during new task learning. Our approach approximates these missing gradients by leveraging the geometric properties of the parameter space, specifically using the directional vector between current parameters and previously optimal parameters as gradient guidance. This approximated gradient can be further integrated with real gradients from a limited replay buffer and regulated by a Bernoulli sampling strategy that dynamically balances model stability and plasticity. Extensive experiments on multimodal continual instruction tuning datasets demonstrate that our method achieves state-of-the-art performance without model expansion, effectively mitigating catastrophic forgetting while maintaining a compact architecture.", "date": "2025-11-19", "categories": []}
{"id": "2603.19624", "title": "Continual Learning for Food Category Classification Dataset: Enhancing Model Adaptability and Performance", "abstract": "Conventional machine learning pipelines often struggle to recognize categories absent from the original trainingset. This gap typically reduces accuracy, as fixed datasets rarely capture the full diversity of a domain. To address this, we propose a continual learning framework for text-guided food classification. Unlike approaches that require retraining from scratch, our method enables incremental updates, allowing new categories to be integrated without degrading prior knowledge. For example, a model trained on Western cuisines could later learn to classify dishes such as dosa or kimchi. Although further refinements are needed, this design shows promise for adaptive food recognition, with applications in dietary monitoring and personalized nutrition planning.", "date": "2026-03-20", "categories": []}
{"id": "2603.20036", "title": "Continual Learning as Shared-Manifold Continuation Under Compatible Shift", "abstract": "Continual learning methods usually preserve old behavior by regularizing parameters, matching old outputs, or replaying previous examples. These strategies can reduce forgetting, but they do not directly specify how the latent representation should evolve. We study a narrower geometric alternative for the regime where old and new data should remain on the same latent support: continual learning as continuation of a shared manifold. We instantiate this view within Support-Preserving Manifold Assimilation (SPMA) and evaluate a geometry-preserving variant, SPMA-OG, that combines sparse replay, output distillation, relational geometry preservation, local smoothing, and chart-assignment regularization on old anchors. On representative compatible-shift CIFAR10 and Tiny-ImageNet runs, SPMA-OG improves over sparse replay baselines in old-task retention and representation-preservation metrics while remaining competitive on new-task accuracy. On a controlled synthetic atlas-manifold benchmark, it achieves near-perfect anchor-geometry preservation while also improving new-task accuracy over replay. These results provide evidence that geometry-aware anchor regularization is a useful inductive bias when continual learning should preserve a shared latent support rather than create a new one.", "date": "2026-03-20", "categories": []}
{"id": "2409.16215", "title": "TiROD: Tiny Robotics Dataset and Benchmark for Continual Object Detection", "abstract": "Detecting objects with visual sensors is crucial for numerous mobile robotics applications, from autonomous navigation to inspection. However, robots often need to operate under significant domains shifts from those they were trained in, requiring them to adjust to these changes. Tiny mobile robots, subject to size, power, and computational constraints, face even greater challenges when running and adapting detection models on low-resolution and noisy images. Such adaptability, though, is crucial for real-world deployment, where robots must operate effectively in dynamic and unpredictable settings. In this work, we introduce a new vision benchmark to evaluate lightweight continual learning strategies tailored to the unique characteristics of tiny robotic platforms. Our contributions include: (i) Tiny Robotics Object Detection~(TiROD), a challenging video dataset collected using the onboard camera of a small mobile robot, designed to test object detectors across various domains and classes; (ii) a comprehensive benchmark of several continual learning strategies on different scenarios using NanoDet, a lightweight, real-time object detector for resource-constrained devices.. Our results highlight some key challenges in developing robust and efficient continual learning strategies for object detectors in tiny robotics.es; (ii) a benchmark of different continual learning strategies on this dataset using NanoDet, a lightweight object detector. Our results highlight key challenges in developing robust and efficient continual learning strategies for object detectors in tiny robotics.", "date": "2024-09-24", "categories": []}
{"id": "2512.07400", "title": "Heads collapse, features stay: Why Replay needs big buffers", "abstract": "A persistent paradox in continual learning (CL) is that neural networks often retain linearly separable representations of past tasks even when their output predictions fail. We formalize this distinction as the gap between deep (feature-space) and shallow (classifier-level) forgetting. We reveal a critical asymmetry in Experience Replay: while minimal buffers successfully anchor feature geometry and prevent deep forgetting, mitigating shallow forgetting typically requires substantially larger buffer capacities. To explain this, we extend the Neural Collapse framework to the sequential setting. We characterize deep forgetting as a geometric drift toward out-of-distribution subspaces and prove that any non-zero replay fraction asymptotically guarantees the retention of linear separability. Conversely, we identify that the ``strong collapse'' induced by small buffers leads to rank-deficient covariances and inflated class means, effectively blinding the classifier to true population boundaries. By unifying CL with out-of-distribution detection, our work challenges the prevailing reliance on large buffers, suggesting that explicitly correcting these statistical artifacts could unlock robust performance with minimal replay.", "date": "2025-12-08", "categories": []}
{"id": "2603.11201", "title": "Representation Finetuning for Continual Learning", "abstract": "The world is inherently dynamic, and continual learning aims to enable models to adapt to ever-evolving data streams. While pre-trained models have shown powerful performance in continual learning, they still require finetuning to adapt effectively to downstream tasks. However, prevailing Parameter-Efficient Fine-Tuning (PEFT) methods operate through empirical, black-box optimization at the weight level. These approaches lack explicit control over representation drift, leading to sensitivity to domain shifts and catastrophic forgetting in continual learning scenarios. In this work, we introduce Continual Representation Learning (CoRe), a novel framework that for the first time shifts the finetuning paradigm from weight space to representation space. Unlike conventional methods, CoRe performs task-specific interventions within a low-rank linear subspace of hidden representations, adopting a learning process with explicit objectives, which ensures stability for past tasks while maintaining plasticity for new ones. By constraining updates to a low-rank subspace, CoRe achieves exceptional parameter efficiency. Extensive experiments across multiple continual learning benchmarks demonstrate that CoRe not only preserves parameter efficiency but also significantly outperforms existing state-of-the-art methods. Our work introduces representation finetuning as a new, more effective and interpretable paradigm for continual learning.", "date": "2026-03-11", "categories": []}
{"id": "2603.11211", "title": "A Simple Efficiency Incremental Learning Framework via Vision-Language Model with Nonlinear Multi-Adapters", "abstract": "Incremental Learning (IL) aims to learn new tasks while preserving previously acquired knowledge. Integrating the zero-shot learning capabilities of pre-trained vision-language models into IL methods has marked a significant advancement. However, these methods face three primary challenges: (1) the need for improved training efficiency; (2) reliance on a memory bank to store previous data; and (3) the necessity of a strong backbone to augment the model's capabilities. In this paper, we propose SimE, a Simple and Efficient framework that employs a vision-language model with adapters designed specifically for the IL task. We report a remarkable phenomenon: there is a nonlinear correlation between the number of adaptive adapter connections and the model's IL capabilities. While increasing adapter connections between transformer blocks improves model performance, adding more adaptive connections within transformer blocks during smaller incremental steps does not enhance, and may even degrade the model's IL ability. Extensive experimental results show that SimE surpasses traditional methods by 9.6% on TinyImageNet and outperforms other CLIP-based methods by 5.3% on CIFAR-100. Furthermore, we conduct a systematic study to enhance the utilization of the zero-shot capabilities of CLIP. We suggest replacing SimE's encoder with a CLIP model trained on larger datasets (e.g., LAION2B) and stronger architectures (e.g., ViT-L/14).", "date": "2026-03-11", "categories": []}
{"id": "2603.18641", "title": "A Comparative Empirical Study of Catastrophic Forgetting Mitigation in Sequential Task Adaptation for Continual Natural Language Processing Systems", "abstract": "Neural language models deployed in real-world applications must continually adapt to new tasks and domains without forgetting previously acquired knowledge. This work presents a comparative empirical study of catastrophic forgetting mitigation in continual intent classification. Using the CLINC150 dataset, we construct a 10-task label-disjoint scenario and evaluate three backbone architectures: a feed-forward Artificial Neural Network (ANN), a Gated Recurrent Unit (GRU), and a Transformer encoder, under a range of continual learning (CL) strategies. We consider one representative method from each major CL family: replay-based Maximally Interfered Retrieval (MIR), regularization-based Learning without Forgetting (LwF), and parameter-isolation via Hard Attention to Task (HAT), both individually and in all pairwise and triple combinations. Performance is assessed with average accuracy, macro F1, and backward transfer, capturing the stability-plasticity trade-off across the task sequence. Our results show that naive sequential fine-tuning suffers from severe forgetting for all architectures and that no single CL method fully prevents it. Replay emerges as a key ingredient: MIR is the most reliable individual strategy, and combinations that include replay (MIR+HAT, MIR+LwF, MIR+LwF+HAT) consistently achieve high final performance with near-zero or mildly positive backward transfer. The optimal configuration is architecture-dependent. MIR+HAT yields the best result for ANN and Transformer, MIR+LwF+HAT, on the other hand, works the best for GRU, and in several cases CL methods even surpass joint training, indicating a regularization effect. These findings highlight the importance of jointly selecting backbone architecture and CL mechanism when designing continual intent-classification systems.", "date": "2026-03-19", "categories": []}
{"id": "2603.18743", "title": "Memento-Skills: Let Agents Design Agents", "abstract": "We introduce \\emph{Memento-Skills}, a generalist, continually-learnable LLM agent system that functions as an \\emph{agent-designing agent}: it autonomously constructs, adapts, and improves task-specific agents through experience. The system is built on a memory-based reinforcement learning framework with \\emph{stateful prompts}, where reusable skills (stored as structured markdown files) serve as persistent, evolving memory. These skills encode both behaviour and context, enabling the agent to carry forward knowledge across interactions.\n  Starting from simple elementary skills (like Web search and terminal operations), the agent continually improves via the \\emph{Read--Write Reflective Learning} mechanism introduced in \\emph{Memento~2}~\\cite{wang2025memento2}. In the \\emph{read} phase, a behaviour-trainable skill router selects the most relevant skill conditioned on the current stateful prompt; in the \\emph{write} phase, the agent updates and expands its skill library based on new experience. This closed-loop design enables \\emph{continual learning without updating LLM parameters}, as all adaptation is realised through the evolution of externalised skills and prompts.\n  Unlike prior approaches that rely on human-designed agents, Memento-Skills enables a generalist agent to \\emph{design agents end-to-end} for new tasks. Through iterative skill generation and refinement, the system progressively improves its own capabilities. Experiments on the \\emph{General AI Assistants} benchmark and \\emph{Humanity's Last Exam} demonstrate sustained gains, achieving 26.2\\% and 116.2\\% relative improvements in overall accuracy, respectively. Code is available at https://github.com/Memento-Teams/Memento-Skills.", "date": "2026-03-19", "categories": []}
{"id": "2603.18872", "title": "DriftGuard: Mitigating Asynchronous Data Drift in Federated Learning", "abstract": "In real-world Federated Learning (FL) deployments, data distributions on devices that participate in training evolve over time. This leads to asynchronous data drift, where different devices shift at different times and toward different distributions. Mitigating such drift is challenging: frequent retraining incurs high computational cost on resource-constrained devices, while infrequent retraining degrades performance on drifting devices. We propose DriftGuard, a federated continual learning framework that efficiently adapts to asynchronous data drift. DriftGuard adopts a Mixture-of-Experts (MoE) inspired architecture that separates shared parameters, which capture globally transferable knowledge, from local parameters that adapt to group-specific distributions. This design enables two complementary retraining strategies: (i) global retraining, which updates the shared parameters when system-wide drift is identified, and (ii) group retraining, which selectively updates local parameters for clusters of devices identified via MoE gating patterns, without sharing raw data. Experiments across multiple datasets and models show that DriftGuard matches or exceeds state-of-the-art accuracy while reducing total retraining cost by up to 83%. As a result, it achieves the highest accuracy per unit retraining cost, improving over the strongest baseline by up to 2.3x. DriftGuard is available for download from https://github.com/blessonvar/DriftGuard.", "date": "2026-03-19", "categories": []}
{"id": "2603.19145", "title": "Enhancing Pretrained Model-based Continual Representation Learning via Guided Random Projection", "abstract": "Recent paradigms in Random Projection Layer (RPL)-based continual representation learning have demonstrated superior performance when building upon a pre-trained model (PTM). These methods insert a randomly initialized RPL after a PTM to enhance feature representation in the initial stage. Subsequently, a linear classification head is used for analytic updates in the continual learning stage. However, under severe domain gaps between pre-trained representations and target domains, a randomly initialized RPL exhibits limited expressivity under large domain shifts. While largely scaling up the RPL dimension can improve expressivity, it also induces an ill-conditioned feature matrix, thereby destabilizing the recursive analytic updates of the linear head. To this end, we propose the Stochastic Continual Learner with MemoryGuard Supervisory Mechanism (SCL-MGSM). Unlike random initialization, MGSM constructs the projection layer via a principled, data-guided mechanism that progressively selects target-aligned random bases to adapt the PTM representation to downstream tasks. This facilitates the construction of a compact yet expressive RPL while improving the numerical stability of analytic updates. Extensive experiments on multiple exemplar-free Class Incremental Learning (CIL) benchmarks demonstrate that SCL-MGSM achieves superior performance compared to state-of-the-art methods.", "date": "2026-03-19", "categories": []}
{"id": "2506.10680", "title": "SatSOM: Saturation Self-Organizing Maps for Continual Learning", "abstract": "Continual learning poses a fundamental challenge for neural systems, which often suffer from catastrophic forgetting when exposed to sequential tasks. Self-Organizing Maps (SOMs), despite their interpretability and efficiency, are not immune to this issue. In this paper, we introduce Saturation Self-Organizing Maps (SatSOM)-an extension of SOMs designed to improve knowledge retention in continual learning scenarios. SatSOM incorporates a novel saturation mechanism that gradually reduces the learning rate and neighborhood radius of neurons as they accumulate information. This effectively freezes well-trained neurons and redirects learning to underutilized areas of the map.", "date": "2025-06-12", "categories": []}
{"id": "2508.05059", "title": "Learning from Oblivion: Predicting Knowledge Overflowed Weights via Retrodiction of Forgetting", "abstract": "Pre-trained weights have become a cornerstone of modern deep learning, enabling efficient knowledge transfer and improving downstream task performance, especially in data-scarce scenarios. However, a fundamental question remains: how can we obtain better pre-trained weights that encapsulate more knowledge beyond the given dataset? In this work, we introduce KNowledge-Overflowed Weights (KNOW) prediction, a novel strategy that leverages structured forgetting and its inversion to synthesize knowledge-enriched weights. Our key insight is that sequential fine-tuning on progressively downsized datasets induces a structured forgetting process, which can be modeled and reversed to recover knowledge as if trained on a larger dataset. We construct a dataset of weight transitions governed by this controlled forgetting and employ meta-learning to model weight prediction effectively. Specifically, our KNowledge-Overflowed Weights Nowcaster (KNOWN) acts as a hyper-model that learns the general evolution of weights and predicts enhanced weights with improved generalization. Extensive experiments across diverse datasets and architectures demonstrate that KNOW prediction consistently outperforms Naive fine-tuning and simple weight prediction, leading to superior downstream performance. Our work provides a new perspective on reinterpreting forgetting dynamics to push the limits of knowledge transfer. The code and pre-trained model are available at https://github.com/jjh6297/KNOW", "date": "2025-08-07", "categories": []}
{"id": "2510.25093", "title": "Continual Low-Rank Adapters for LLM-based Generative Recommender Systems", "abstract": "While large language models (LLMs) achieve strong performance in recommendation, they face challenges in continual learning as users, items, and user preferences evolve over time. Existing LoRA-based continual methods primarily focus on preserving performance on previous tasks, but this overlooks the unique nature of recommendation: the goal is not to predict past preferences, and outdated preferences can even harm performance when current interests shift significantly. To address this, we propose PESO (Proximally rEgularized Single evolving lOra, a continual adaptation method for LoRA in recommendation. PESO introduces a proximal regularizer that anchors the current adapter to its most recent frozen state, enabling the model to flexibly balance adaptation and preservation, and to better capture recent user behaviors. Theoretically, we show that this proximal design provides data-aware, direction-wise guidance in the LoRA subspace. Empirically, PESO consistently outperforms existing LoRA-based continual learning methods.", "date": "2025-10-29", "categories": []}
{"id": "2603.17187", "title": "MetaClaw: Just Talk -- An Agent That Meta-Learns and Evolves in the Wild", "abstract": "Large language model (LLM) agents are increasingly used for complex tasks, yet deployed agents often remain static, failing to adapt as user needs evolve. This creates a tension between the need for continuous service and the necessity of updating capabilities to match shifting task distributions. On platforms like OpenClaw, which handle diverse workloads across 20+ channels, existing methods either store raw trajectories without distilling knowledge, maintain static skill libraries, or require disruptive downtime for retraining. We present MetaClaw, a continual meta-learning framework that jointly evolves a base LLM policy and a library of reusable behavioral skills. MetaClaw employs two complementary mechanisms. Skill-driven fast adaptation analyzes failure trajectories via an LLM evolver to synthesize new skills, enabling immediate improvement with zero downtime. Opportunistic policy optimization performs gradient-based updates via cloud LoRA fine-tuning and Reinforcement Learning with a Process Reward Model (RL-PRM). This is triggered during user-inactive windows by the Opportunistic Meta-Learning Scheduler (OMLS), which monitors system inactivity and calendar data. These mechanisms are mutually reinforcing: a refined policy generates better trajectories for skill synthesis, while richer skills provide higher-quality data for policy optimization. To prevent data contamination, a versioning mechanism separates support and query data. Built on a proxy-based architecture, MetaClaw scales to production-size LLMs without local GPUs. Experiments on MetaClaw-Bench and AutoResearchClaw show that skill-driven adaptation improves accuracy by up to 32% relative. The full pipeline advances Kimi-K2.5 accuracy from 21.4% to 40.6% and increases composite robustness by 18.3%. Code is available at https://github.com/aiming-lab/MetaClaw.", "date": "2026-03-17", "categories": []}
{"id": "2603.17548", "title": "CLeAN: Continual Learning Adaptive Normalization in Dynamic Environments", "abstract": "Artificial intelligence systems predominantly rely on static data distributions, making them ineffective in dynamic real-world environments, such as cybersecurity, autonomous transportation, or finance, where data shifts frequently. Continual learning offers a potential solution by enabling models to learn from sequential data while retaining prior knowledge. However, a critical and underexplored issue in this domain is data normalization. Conventional normalization methods, such as min-max scaling, presuppose access to the entire dataset, which is incongruent with the sequential nature of continual learning. In this paper we introduce Continual Learning Adaptive Normalization (CLeAN), a novel adaptive normalization technique designed for continual learning in tabular data. CLeAN involves the estimation of global feature scales using learnable parameters that are updated via an Exponential Moving Average (EMA) module, enabling the model to adapt to evolving data distributions. Through comprehensive evaluations on two datasets and various continual learning strategies, including Resevoir Experience Replay, A-GEM, and EwC we demonstrate that CLeAN not only improves model performance on new data but also mitigates catastrophic forgetting. The findings underscore the importance of adaptive normalization in enhancing the stability and effectiveness of tabular data, offering a novel perspective on the use of normalization to preserve knowledge in dynamic learning environments.", "date": "2026-03-18", "categories": []}
{"id": "2509.21617", "title": "LANCE: Low Rank Activation Compression for Efficient On-Device Continual Learning", "abstract": "On-device learning is essential for personalization, privacy, and long-term adaptation in resource-constrained environments. Achieving this requires efficient learning, both fine-tuning existing models and continually acquiring new tasks without catastrophic forgetting. Yet both settings are constrained by high memory cost of storing activations during backpropagation. Existing activation compression methods reduce this cost but rely on repeated low-rank decompositions, introducing computational overhead. Also, such methods have not been explored for continual learning. We propose LANCE (Low-rank Activation Compression), a framework that performs one-shot higher-order Singular Value Decomposition (SVD) to obtain a reusable low-rank subspace for activation projection. This eliminates repeated decompositions, reducing both memory and computation. Moreover, fixed low-rank subspaces further enable on-device continual learning by allocating tasks to orthogonal subspaces without storing large task-specific matrices. Experiments show that LANCE reduces activation storage up to 250$\\times$ while maintaining accuracy comparable to full backpropagation on CIFAR-10/100, Oxford-IIIT Pets, Flowers102, and CUB-200 datasets. On continual learning benchmarks (Split CIFAR-100, Split MiniImageNet, 5-Datasets), it performs competitively with orthogonal gradient projection methods at a fraction of the memory cost. These results position LANCE as a practical and scalable solution for efficient fine-tuning and continual learning on edge devices.", "date": "2025-09-25", "categories": []}
{"id": "2603.16738", "title": "MedCL-Bench: Benchmarking stability-efficiency trade-offs and scaling in biomedical continual learning", "abstract": "Medical language models must be updated as evidence and terminology evolve, yet sequential updating can trigger catastrophic forgetting. Although biomedical NLP has many static benchmarks, no unified, task-diverse benchmark exists for evaluating continual learning under standardized protocols, robustness to task order and compute-aware reporting. We introduce MedCL-Bench, which streams ten biomedical NLP datasets spanning five task families and evaluates eleven continual learning strategies across eight task orders, reporting retention, transfer, and GPU-hour cost. Across backbones and task orders, direct sequential fine-tuning on incoming tasks induces catastrophic forgetting, causing update-induced performance regressions on prior tasks. Continual learning methods occupy distinct retention-compute frontiers: parameter-isolation provides the best retention per GPU-hour, replay offers strong protection at higher cost, and regularization yields limited benefit. Forgetting is task-dependent, with multi-label topic classification most vulnerable and constrained-output tasks more robust. MedCL-Bench provides a reproducible framework for auditing model updates before deployment.", "date": "2026-03-17", "categories": []}
{"id": "2509.21413", "title": "Null-Space Filtering for Data-Free Continual Model Merging: Preserving Stability, Promoting Plasticity", "abstract": "Data-free continual model merging (DFCMM) aims to fuse independently fine-tuned models into a single backbone that evolves with incoming tasks without accessing task data. This paper revisits two fundamental desiderata for DFCMM: stability, avoiding interference with earlier tasks, and plasticity, adapting faithfully to each new task. This poses a challenge that existing approaches fail to address: how to bridge data-level desiderata with parameter-space optimization to ensure stability and plasticity in the absence of task data. To this end, we propose NUFILT (NUll-space FILTering), a data-free framework that directly links these desiderata into parameter-space optimization. Our key observation is that task vectors approximately align with representation subspaces, providing structural surrogates for enforcing stability and plasticity. Accordingly, we design a null-space projector that preserves prior responses by filtering overlapping components of new task vectors, ensuring stability. We further introduce a lightweight LoRA adapter that injects complementary task-specific signals to enable plasticity. The adapter is trained with a projection-based surrogate loss that preserves consistency with prior knowledge while introducing novel directions. This joint filtering-adaptation process enables the backbone to absorb new knowledge while retaining existing behaviors, with updates fused back in a layer-wise linear fashion without extra parameters or inference cost. Theoretically, we establish approximate subspace alignment guarantees that justify null-space filtering. Empirically, NUFILT achieves state-of-the-art performance with minimal forgetting on both vision and NLP benchmarks, improving average accuracy by 4-7% over OPCM and WUDI-Merging, while narrowing the gap to fine-tuning and reducing computation overhead. The code is available at: https://github.com/zihuanqiu/NUFILT", "date": "2025-09-25", "categories": []}
{"id": "2603.06662", "title": "HyperTokens: Controlling Token Dynamics for Continual Video-Language Understanding", "abstract": "Continual VideoQA with multimodal LLMs is hindered by interference between tasks and the prohibitive cost of storing task-specific prompts. We introduce HyperTokens, a transformer-based token generator that produces fine-tuning tokens on demand, giving explicit control over prompt updates while keeping memory fixed. To suppress forgetting, we propose meta-inspired regularisers that look ahead to avoid task-specific sharp directions and anchor the evolving generator to prior tasks. We further connect our objective to sharpness-aware optimisation, providing insight into why it encourages flatter cross-task minima and improves retention. Beyond regularisation, HyperTokens exploits lightweight auxiliary multimodal supervision through shared generation weights; guided by a causal perspective, we design feasible objectives and surrogate mutual-information losses to regularise anti-causal cross-modal directions. Across two standard continual VideoQA benchmarks, HyperTokens achieves higher average accuracy with substantially lower forgetting. Finally, we introduce a challenging cross-modal ImageQA->VideoQA protocol and show that HyperTokens enables robust continual transfer in this setting.", "date": "2026-03-02", "categories": []}
{"id": "2603.13235", "title": "Continual Fine-Tuning with Provably Accurate and Parameter-Free Task Retrieval", "abstract": "Continual fine-tuning aims to adapt a pre-trained backbone to new tasks sequentially while preserving performance on earlier tasks whose data are no longer available. Existing approaches fall into two categories which include input- and parameter-adaptation. Input-adaptation methods rely on retrieving the most relevant prompts at test time, but require continuously learning a retrieval function that is prone to forgetting. Parameter-adaptation methods instead use a fixed input embedding function to enable retrieval-free prediction and avoid forgetting, but sacrifice representation adaptability. To combine their best strengths, we propose a new parameter-adaptation method that enables adaptive use of input embeddings during test time with parameter-free retrieval. We derive task-retrieval error bounds for a clustering-based, parameter-free paradigm, providing theoretical guarantees that link low retrieval error to structural properties of task-specific representation clusters, revealing a fresh insight into how well-organized clustering structure will enable reliable retrieval. Motivated by this insight, our method is designed with two key components: (i) an adaptive module composition strategy that learns informative task-specific updates to preserve and complement prior knowledge, and (ii) a clustering-based retrieval mechanism that captures distinct representation signatures for each task, enabling adaptive representation use at test time. Extensive experiments show that these components work synergistically to improve retrieval and predictive performance under large shifts in task semantics.", "date": "2026-01-28", "categories": []}
{"id": "2603.13276", "title": "FastODT: A tree-based framework for efficient continual learning", "abstract": "Machine learning models deployed in real-world settings must operate under evolving data distributions and constrained computational resources. This challenge is particularly acute in non-stationary domains such as energy time series, weather monitoring, and environmental sensing. To remain effective, models must support adaptability, continuous learning, and long-term knowledge retention. This paper introduces a oblivious tree-based model with Hoeffding bound controlling its growth. It seamlessly integrates rapid learning and inference with efficient memory management and robust knowledge preservation, thus allowing for online learning. Extensive experiments across energy and environmental sensing time-series benchmarks demonstrate that the proposed framework achieves performance competitive with, and in several cases surpassing, existing online and batch learning methods, while maintaining superior computational efficiency. Collectively, these results demonstrate that the proposed approach fulfills the core objectives of adaptability, continual updating, and efficient retraining without full model retraining. The framework provides a scalable and resource-aware foundation for deployment in real-world non-stationary environments where resources are constrained and sustained adaptation is essential.", "date": "2026-02-27", "categories": []}
{"id": "2603.13335", "title": "Information-Theoretic Constraints for Continual Vision-Language-Action Alignment", "abstract": "When deployed in open-ended robotic environments, Vision--Language--Action (VLA) models need to continually acquire new skills, yet suffer from severe catastrophic forgetting. We observe that this degradation is related to the deterioration of cross-modal information structure, where dependencies among visual observations, language instructions, and actions progressively diffuse during continual adaptation. But existing continual learning methods fail to preserve such cross-modal information dependencies. Thus, we propose Info-VLA, an information-preserving continual learning framework that maintains cross-modal information structure through two complementary constraints. Replay Anchor Contrastive Learning constructs stable alignment anchors from a frozen teacher model, preserving cross-modal alignment in the representation space. Cross-Modal Mutual Information Maximization further preserves dependency structure between visual and language representations through mutual information constraints. By jointly preserving historical alignment and cross-modal dependency information, Info-VLA balances stability and plasticity during continual learning. Furthermore, experiments on the LIBERO show that Info-VLA significantly outperforms existing methods in both task retention and adaptation.", "date": "2026-03-06", "categories": []}
{"id": "2603.13874", "title": "Zero-Forgetting CISS via Dual-Phase Cognitive Cascades", "abstract": "Continual semantic segmentation (CSS) is a cornerstone task in computer vision that enables a large number of downstream applications, but faces the catastrophic forgetting challenge. In conventional class-incremental semantic segmentation (CISS) frameworks using Softmax-based classification heads, catastrophic forgetting originates from Catastrophic forgetting and task affiliation probability. We formulate these problems and provide a theoretical analysis to more deeply understand the limitations in existing CISS methods, particularly Strict Parameter Isolation (SPI). To address these challenges, we follow a dual-phase intuition from human annotators, and introduce Cognitive Cascade Segmentation (CogCaS), a novel dual-phase cascade formulation for CSS tasks in the CISS setting. By decoupling the task into class-existence detection and class-specific segmentation, CogCaS enables more effective continual learning, preserving previously learned knowledge while incorporating new classes. Using two benchmark datasets PASCAL VOC 2012 and ADE20K, we have shown significant improvements in a variety of challenging scenarios, particularly those with long sequence of incremental tasks, when compared to exsiting state-of-the-art methods. Our code will be made publicly available upon paper acceptance.", "date": "2026-03-14", "categories": []}
{"id": "2603.14484", "title": "Unlearning-based sliding window for continual learning under concept drift", "abstract": "Traditional machine learning assumes a stationary data distribution, yet many real-world applications operate on nonstationary streams in which the underlying concept evolves over time. This problem can also be viewed as task-free continual learning under concept drift, where a model must adapt sequentially without explicit task identities or task boundaries. In such settings, effective learning requires both rapid adaptation to new data and forgetting of outdated information. A common solution is based on a sliding window, but this approach is often computationally demanding because the model must be repeatedly retrained from scratch on the most recent data. We propose a different perspective based on machine unlearning. Instead of rebuilding the model each time the active window changes, we remove the influence of outdated samples using unlearning and then update the model with newly observed data. This enables efficient, targeted forgetting while preserving adaptation to evolving distributions. To the best of our knowledge, this is the first work to connect machine unlearning with concept drift mitigation for task-free continual learning. Empirical results on image stream classification across multiple drift scenarios demonstrate that the proposed approach offers a competitive and computationally efficient alternative to standard sliding-window retraining. Our implementation can be found at \\hrehttps://anonymous.4open.science/r/MUNDataStream-60F3}{https://anonymous.4open.science/r/MUNDataStream-60F3}.", "date": "2026-03-15", "categories": []}
{"id": "2603.14843", "title": "ContiGuard: A Framework for Continual Toxicity Detection Against Evolving Evasive Perturbations", "abstract": "Toxicity detection mitigates the dissemination of toxic content (e.g., hateful comments, posts, and messages within online social actions) to safeguard a healthy online social environment. However, malicious users persistently develop evasive perturbations to disguise toxic content and evade detectors. Traditional detectors or methods are static over time and are inadequate in addressing these evolving evasion tactics. Thus, continual learning emerges as a logical approach to dynamically update detection ability against evolving perturbations. Nevertheless, disparities across perturbations hinder the detector's continual learning on perturbed text. More importantly, perturbation-induced noises distort semantics to degrade comprehension and also impair critical feature learning to render detection sensitive to perturbations. These amplify the challenge of continual learning against evolving perturbations. In this work, we present ContiGuard, the first framework tailored for continual learning of the detector on time-evolving perturbed text (termed continual toxicity detection) to enable the detector to continually update capability and maintain sustained resilience against evolving perturbations. Specifically, to boost the comprehension, we present an LLM-powered semantic enriching strategy, where we dynamically incorporate possible meaning and toxicity-related clues excavated by LLM into the perturbed text to improve the comprehension. To mitigate non-critical features and amplify critical ones, we propose a discriminability-driven feature learning strategy, where we strengthen discriminative features while suppressing the less-discriminative ones to shape a robust classification boundary for detection...", "date": "2026-03-16", "categories": []}
{"id": "2603.15184", "title": "CATFormer: When Continual Learning Meets Spiking Transformers With Dynamic Thresholds", "abstract": "Although deep neural networks perform extremely well in controlled environments, they fail in real-world scenarios where data isn't available all at once, and the model must adapt to a new data distribution that may or may not follow the initial distribution. Previously acquired knowledge is lost during subsequent updates based on new data. a phenomenon commonly known as catastrophic forgetting. In contrast, the brain can learn without such catastrophic forgetting, irrespective of the number of tasks it encounters. Existing spiking neural networks (SNNs) for class-incremental learning (CIL) suffer a sharp performance drop as tasks accumulate. We here introduce CATFormer (Context Adaptive Threshold Transformer), a scalable framework that overcomes this limitation. We observe that the key to preventing forgetting in SNNs lies not only in synaptic plasticity but also in modulating neuronal excitability. At the core of CATFormer is the Dynamic Threshold Leaky Integrate-and-Fire (DTLIF) neuron model, which leverages context-adaptive thresholds as the primary mechanism for knowledge retention. This is paired with a Gated Dynamic Head Selection (G-DHS) mechanism for task-agnostic inference. Extensive evaluation on both static (CIFAR-10/100/Tiny-ImageNet) and neuromorphic (CIFAR10-DVS/SHD) datasets reveals that CATFormer outperforms existing rehearsal-free CIL algorithms across various task splits, establishing it as an ideal architecture for energy-efficient, true-class incremental learning.", "date": "2026-03-16", "categories": []}
{"id": "2508.01587", "title": "Distilling the Past: Information-Dense and Style-Aware Replay for Lifelong Person Re-Identification", "abstract": "Lifelong person re-identification (LReID) aims to continuously adapt to new domains while mitigating catastrophic forgetting. While replay-based methods effectively alleviate forgetting, they are constrained by strict memory budgets, leading to limited sample diversity. Conversely, exemplar-free approaches bypass memory constraints entirely but struggle to preserve the fine-grained identity semantics crucial for Re-ID tasks. To resolve this fundamental dilemma, we propose an Information-Dense and Style-Aware Replay framework. Instead of storing a sparse set of raw historical images, we fuse the knowledge of sequential data into the pixel space of a compact replay buffer via multi-stage gradient matching and identity supervision. This condensation process not only maximizes the semantic representativeness of limited memory but also naturally conceals original visual details, inherently preserving data privacy. Furthermore, to combat forgetting induced by cross-domain shifts, we introduce a dual-alignment style replay strategy that adapts both current and fused replay samples, harmonizing feature representations across disparate domains. Extensive experiments on multiple LReID benchmarks demonstrate that our method significantly outperforms existing approaches, achieving improvements of +5.0% and +6.0% in Seen-Avg mAP over current state-of-the-art and traditional replay-based methods, respectively, thereby establishing an efficient and robust new baseline for lifelong learning.", "date": "2025-08-03", "categories": []}
{"id": "2508.10954", "title": "UniPrompt-CL: Sustainable Continual Learning in Medical AI with Unified Prompt Pools", "abstract": "Modern AI models are typically trained on static datasets, limiting their ability to continuously adapt to rapidly evolving real-world environments. While continual learning (CL) addresses this limitation, most CL methods are designed for natural images and often underperform or fail to transfer to medical data due to domain bias, institutional constraints, and subtle inter-stage boundaries. We propose UniPrompt-CL, a medical-oriented prompt-based continual learning method that improves prompt pool design via a minimally expanding unified prompt pool and a new regularization term, achieving a better stability-plasticity trade-off with lower computational cost. Across two domain-incremental learning settings, UniPrompt-CL effectively reduces inference cost while improving AvgACC by 1-3 percentage points. In addition to strong performance, extensive experiments clearly validate the motivation and effectiveness of the proposed improvements.", "date": "2025-08-14", "categories": []}
{"id": "2508.12932", "title": "SEDEG:Sequential Enhancement of Decoder and Encoder's Generality for Class Incremental Learning with Small Memory", "abstract": "In incremental learning, enhancing the generality of knowledge is crucial for adapting to dynamic data inputs. It can develop generalized representations or more balanced decision boundaries, preventing the degradation of long-term knowledge over time and thus mitigating catastrophic forgetting. Some emerging incremental learning methods adopt an encoder-decoder architecture and have achieved promising results. In the encoder-decoder achitecture, improving the generalization capabilities of both the encoder and decoder is critical, as it helps preserve previously learned knowledge while ensuring adaptability and robustness to new, diverse data inputs. However, many existing continual methods focus solely on enhancing one of the two components, which limits their effectiveness in mitigating catastrophic forgetting. And these methods perform even worse in small-memory scenarios, where only a limited number of historical samples can be stored. To mitigate this limitation, we introduces SEDEG, a two-stage training framework for vision transformers (ViT), focusing on sequentially improving the generality of both Decoder and Encoder. Initially, SEDEG trains an ensembled encoder through feature boosting to learn generalized representations, which subsequently enhance the decoder's generality and balance the classifier. The next stage involves using knowledge distillation (KD) strategies to compress the ensembled encoder and develop a new, more generalized encoder. This involves using a balanced KD approach and feature KD for effective knowledge transfer. Extensive experiments on three benchmark datasets show SEDEG's superior performance, and ablation studies confirm the efficacy of its components. The code is available at https://github.com/ShaolingPu/CIL.", "date": "2025-08-18", "categories": []}
{"id": "2510.27316", "title": "Parameterized Prompt for Incremental Object Detection", "abstract": "Recent studies have demonstrated that incorporating trainable prompts into pretrained models enables effective incremental learning. However, the application of prompts in incremental object detection (IOD) remains underexplored. Our study reveals that existing prompt-pool-based approaches assume disjoint class sets across incremental tasks, which are unsuitable for IOD as they overlook the inherent co-occurrence phenomenon in detection. In co-occurring scenarios, unlabeled objects from previous tasks may appear in current task images, leading to confusion in prompts pool. In this paper, we hold that prompt structures should exhibit adaptive consolidation properties across tasks, with constrained updates to prevent confusion and catastrophic forgetting. Motivated by this, we introduce Parameterized Prompts for Incremental Object Detection (P$^2$IOD). Leveraging neural networks global evolution properties, P$^2$IOD employs networks as the parameterized prompts to adaptively consolidate knowledge across tasks. To constrain prompts structure updates, P$^2$IOD further engages a parameterized prompts fusion strategy. Extensive experiments on PASCAL VOC2007 and MS COCO datasets demonstrate that P$^2$IOD's effectiveness in IOD and achieves the state-of-the-art performance among existing baselines. Code is available at https://github.com/EMLS-ICTCAS/P2IOD.", "date": "2025-10-31", "categories": []}
{"id": "2511.18507", "title": "Multimodal Continual Learning with MLLMs from Multi-scenario Perspectives", "abstract": "Multimodal large language models (MLLMs) deployed on devices must adapt to continuously changing visual scenarios such as variations in background and perspective, to effectively perform complex visual tasks. To investigate catastrophic forgetting under real-world scenario shifts, we construct a multimodal visual understanding dataset (MSVQA), covering four distinct scenarios and perspectives: high-altitude, underwater, low-altitude, and indoor environments. Furthermore, we propose UNIFIER (mUltimodal coNtInual learning with MLLMs From multi-scenarIo pERspectives), a continual learning (CL) framework designed to address visual discrepancies while learning different scenarios. Compared to existing CL methods, UNIFIER enables knowledge accumulation within the same scenario and mutual enhancement across different scenarios via Vision Representation Expansion (VRE) and Vision Consistency Constraint (VCC). Experimental results show that UNIFIER improves the last-step VQA scores by 2.70%~10.62% and the last-step F1 scores by 3.40%~7.69% compared to the state-of-the-art method, QUAD, in 20-step cross-scenario continual learning tasks. MSVQA dataset is available at https://huggingface.co/datasets/Kaij00/MSVQA.", "date": "2025-11-23", "categories": []}
{"id": "2601.04864", "title": "Key-Value Pair-Free Continual Learner via Task-Specific Prompt-Prototype", "abstract": "Continual learning aims to enable models to acquire new knowledge while retaining previously learned information. Prompt-based methods have shown remarkable performance in this domain; however, they typically rely on key-value pairing, which can introduce inter-task interference and hinder scalability. To overcome these limitations, we propose a novel approach employing task-specific Prompt-Prototype (ProP), thereby eliminating the need for key-value pairs. In our method, task-specific prompts facilitate more effective feature learning for the current task, while corresponding prototypes capture the representative features of the input. During inference, predictions are generated by binding each task-specific prompt with its associated prototype. Additionally, we introduce regularization constraints during prompt initialization to penalize excessively large values, thereby enhancing stability. Experiments on several widely used datasets demonstrate the effectiveness of the proposed method. In contrast to mainstream prompt-based approaches, our framework removes the dependency on key-value pairs, offering a fresh perspective for future continual learning research.", "date": "2026-01-08", "categories": []}
{"id": "2603.12658", "title": "Continual Learning in Large Language Models: Methods, Challenges, and Opportunities", "abstract": "Continual learning (CL) has emerged as a pivotal paradigm to enable large language models (LLMs) to dynamically adapt to evolving knowledge and sequential tasks while mitigating catastrophic forgetting-a critical limitation of the static pre-training paradigm inherent to modern LLMs. This survey presents a comprehensive overview of CL methodologies tailored for LLMs, structured around three core training stages: continual pre-training, continual fine-tuning, and continual alignment.Beyond the canonical taxonomy of rehearsal-, regularization-, and architecture-based methods, we further subdivide each category by its distinct forgetting mitigation mechanisms and conduct a rigorous comparative analysis of the adaptability and critical improvements of traditional CL methods for LLMs. In doing so, we explicitly highlight core distinctions between LLM CL and traditional machine learning, particularly with respect to scale, parameter efficiency, and emergent capabilities. Our analysis covers essential evaluation metrics, including forgetting rates and knowledge transfer efficiency, along with emerging benchmarks for assessing CL performance. This survey reveals that while current methods demonstrate promising results in specific domains, fundamental challenges persist in achieving seamless knowledge integration across diverse tasks and temporal scales. This systematic review contributes to the growing body of knowledge on LLM adaptation, providing researchers and practitioners with a structured framework for understanding current achievements and future opportunities in lifelong learning for language models.", "date": "2026-03-13", "categories": []}
{"id": "2603.12816", "title": "Residual SODAP: Residual Self-Organizing Domain-Adaptive Prompting with Structural Knowledge Preservation for Continual Learning", "abstract": "Continual learning (CL) suffers from catastrophic forgetting, which is exacerbated in domain-incremental learning (DIL) where task identifiers are unavailable and storing past data is infeasible. While prompt-based CL (PCL) adapts representations with a frozen backbone, we observe that prompt-only improvements are often insufficient due to suboptimal prompt selection and classifier-level instability under domain shifts. We propose Residual SODAP, which jointly performs prompt-based representation adaptation and classifier-level knowledge preservation. Our framework combines $\\alpha$-entmax sparse prompt selection with residual aggregation, data-free distillation with pseudo-feature replay, prompt-usage--based drift detection, and uncertainty-aware multi-loss balancing. Across three DIL benchmarks without task IDs or extra data storage, Residual SODAP achieves state-of-the-art AvgACC/AvgF of 0.850/0.047 (DR), 0.760/0.031 (Skin Cancer), and 0.995/0.003 (CORe50).", "date": "2026-03-13", "categories": []}
{"id": "2510.21019", "title": "More Than Memory Savings: Zeroth-Order Optimization Mitigates Forgetting in Continual Learning", "abstract": "Zeroth-order (ZO) optimization has gained attention as a memory-efficient alternative to first-order (FO) methods, particularly in settings where gradient computation is expensive or even impractical. Beyond its memory efficiency, in this work, we investigate ZO optimization for continual learning (CL) as a novel approach to address the plasticity-stability-efficiency trilemma. Through theoretical analysis and empirical evidence, we show that ZO optimization naturally leads to flatter loss landscapes, which in turn reduce forgetting in CL. However, this stability comes at a cost of plasticity: due to its imprecise gradient estimates and slower convergence, ZO optimization tends to be less effective than FO in acquiring new task-specific knowledge, particularly under constrained training budgets. To better understand this trade-off, we conduct a holistic evaluation of ZO optimization applied to various existing CL methods. Our findings reveal that ZO optimization enhances stability but often undermines plasticity, particularly when used with learnable classifiers. Motivated by this insight, we propose ZO-FC, a simple but effective approach that applies ZO optimization to a single adapter-based PEFT module with FO optimized classifier. This design leverages the stability benefits of ZO while preserving the adaptability of FO updates with negligible memory overhead. Experiments demonstrate that ZO-FC achieves an effective balance between stability and plasticity, offering a practical and memory-efficient solution for on-device CL.", "date": "2025-10-23", "categories": []}
{"id": "2603.10929", "title": "Lifelong Imitation Learning with Multimodal Latent Replay and Incremental Adjustment", "abstract": "We introduce a lifelong imitation learning framework that enables continual policy refinement across sequential tasks under realistic memory and data constraints. Our approach departs from conventional experience replay by operating entirely in a multimodal latent space, where compact representations of visual, linguistic, and robot's state information are stored and reused to support future learning. To further stabilize adaptation, we introduce an incremental feature adjustment mechanism that regularizes the evolution of task embeddings through an angular margin constraint, preserving inter-task distinctiveness. Our method establishes a new state of the art in the LIBERO benchmarks, achieving 10-17 point gains in AUC and up to 65% less forgetting compared to previous leading methods. Ablation studies confirm the effectiveness of each component, showing consistent gains over alternative strategies. The code is available at: https://github.com/yfqi/lifelong_mlr_ifa.", "date": "2026-03-11", "categories": []}
{"id": "2603.12163", "title": "A Quantitative Characterization of Forgetting in Post-Training", "abstract": "Continual post-training of generative models is widely used, yet a principled understanding of when and why forgetting occurs remains limited. We develop theoretical results under a two-mode mixture abstraction (representing old and new tasks), proposed by Chen et al. (2025) (arXiv:2510.18874), and formalize forgetting in two forms: (i) mass forgetting, where the old mixture weight collapses to zero, and (ii) old-component drift, where an already-correct old component shifts during training. For equal-covariance Gaussian modes, we prove that forward-KL objectives trained on data from the new distribution drive the old weight to zero, while reverse-KL objectives converge to the true target (thereby avoiding mass forgetting) and perturb the old mean only through overlap-gated misassignment probabilities controlled by the Bhattacharyya coefficient, yielding drift that decays exponentially with mode separation and a locally well-conditioned geometry with exponential convergence. We further quantify how replay interacts with these objectives. For forward-KL, replay must modify the training distribution to change the population optimum; for reverse-KL, replay leaves the population objective unchanged but prevents finite-batch old-mode starvation through bounded importance weighting. Finally, we analyze three recently proposed near-on-policy post-training methods, SDFT (arxiv:2601.19897), TTT-Discover (arxiv:2601.16175), and OAPL (arxiv:2602.19362), via the same lens and derive explicit conditions under which each retains old mass and exhibits overlap-controlled drift. Overall, our results show that forgetting can by precisely quantified based on the interaction between divergence direction, geometric behavioral overlap, sampling regime, and the visibility of past behavior during training.", "date": "2026-03-12", "categories": []}
{"id": "2411.00430", "title": "Class Incremental Learning with Task-Specific Batch Normalization and Out-of-Distribution Detection", "abstract": "This study focuses on incremental learning for image classification, exploring how to reduce catastrophic forgetting of all learned knowledge when access to old data is restricted. The challenge lies in balancing plasticity (learning new knowledge) and stability (retaining old knowledge). Based on whether the task identifier (task-ID) is available during testing, incremental learning is divided into task incremental learning (TIL) and class incremental learning (CIL). The TIL paradigm often uses multiple classifier heads, selecting the corresponding head based on the task-ID. Since the CIL paradigm cannot access task-ID, methods originally developed for TIL require explicit task-ID prediction to bridge this gap and enable their adaptation to the CIL paradigm. {In this study, a novel continual learning framework extends the TIL method for CIL by introducing out-of-distribution detection for task-ID prediction. Our framework utilizes task-specific Batch Normalization (BN) and task-specific classification heads to effectively adjust feature map distributions for each task, enhancing plasticity. With far fewer parameters than convolutional kernels, task-specific BN helps minimize parameter growth, preserving stability. Based on multiple task-specific classification heads, we introduce an ``unknow'' class for each head. During training, data from other tasks are mapped to this unknown class. During inference, the task-ID is predicted by selecting the classification head with the lowest probability assigned to the unknown class. Our method achieves state-of-the-art performance on two medical image datasets and two natural image datasets. The source code is available at https://github.com/z1968357787/mbn_ood_git_main.", "date": "2024-11-01", "categories": []}
{"id": "2503.10705", "title": "Enhanced Continual Learning of Vision-Language Models with Model Fusion", "abstract": "Vision-Language Models (VLMs) represent a significant breakthrough in artificial intelligence by integrating visual and textual modalities to achieve impressive zero-shot capabilities. However, VLMs are susceptible to catastrophic forgetting when sequentially fine-tuned on multiple downstream tasks. Existing continual learning methods for VLMs face various limitations, often relying on additional reference datasets, compromising zero-shot performance, or being restricted to parameter-efficient fine-tuning scenarios. In this paper, we propose a novel Continual Decoupling-Unifying (ConDU) approach that pioneers the use of model fusion for continual learning in VLMs. Specifically, ConDU maintains a unified model along with task triggers and prototype sets, employing an iterative process of decoupling task experts for previous tasks and unifying them with the task expert for the newly learned task. Additionally, we introduce an inference strategy for zero-shot scenarios by aggregating predictions from multiple decoupled task experts. Extensive experiments on the MTIL benchmark show that ConDU achieves up to a 2\\% improvement in average performance across all seen tasks compared to state-of-the-art baselines, while also enhancing zero-shot capabilities relative to the original VLM. Our code is available at https://github.com/zhangzicong518/ConDU.", "date": "2025-03-12", "categories": []}
{"id": "2509.24483", "title": "One-Prompt Strikes Back: Sparse Mixture of Experts for Prompt-based Continual Learning", "abstract": "Prompt-based methods have recently gained prominence in Continual Learning (CL) due to their strong performance and memory efficiency. A prevalent strategy in this paradigm assigns a dedicated subset of prompts to each task, which, while effective, incurs substantial computational overhead and causes memory requirements to scale linearly with the number of tasks. Conversely, approaches employing a single shared prompt across tasks offer greater efficiency but often suffer from degraded performance due to knowledge interference. To reconcile this trade-off, we propose SMoPE, a novel framework that integrates the benefits of both task-specific and shared prompt strategies. Inspired by recent findings on the relationship between Prefix Tuning and Mixture of Experts (MoE), SMoPE organizes a shared prompt into multiple \"prompt experts\" within a sparse MoE architecture. For each input, only a select subset of relevant experts is activated, effectively mitigating interference. To facilitate expert selection, we introduce a prompt-attention score aggregation mechanism that computes a unified proxy score for each expert, enabling dynamic and sparse activation. Additionally, we propose an adaptive noise mechanism to encourage balanced expert utilization while preserving knowledge from prior tasks. To further enhance expert specialization, we design a prototype-based loss function that leverages prefix keys as implicit memory representations. Extensive experiments across multiple CL benchmarks demonstrate that SMoPE consistently outperforms task-specific prompt methods and achieves performance competitive with state-of-the-art approaches, all while significantly reducing parameter counts and computational costs.", "date": "2025-09-29", "categories": []}
{"id": "2603.09888", "title": "LCA: Local Classifier Alignment for Continual Learning", "abstract": "A fundamental requirement for intelligent systems is the ability to learn continuously under changing environments. However, models trained in this regime often suffer from catastrophic forgetting. Leveraging pre-trained models has recently emerged as a promising solution, since their generalized feature extractors enable faster and more robust adaptation. While some earlier works mitigate forgetting by fine-tuning only on the first task, this approach quickly deteriorates as the number of tasks grows and the data distributions diverge. More recent research instead seeks to consolidate task knowledge into a unified backbone, or adapting the backbone as new tasks arrive. However, such approaches may create a (potential) \\textit{mismatch} between task-specific classifiers and the adapted backbone. To address this issue, we propose a novel \\textit{Local Classifier Alignment} (LCA) loss to better align the classifier with backbone. Theoretically, we show that this LCA loss can enable the classifier to not only generalize well for all observed tasks, but also improve robustness. Furthermore, we develop a complete solution for continual learning, following the model merging approach and using LCA. Extensive experiments on several standard benchmarks demonstrate that our method often achieves leading performance, sometimes surpasses the state-of-the-art methods with a large margin.", "date": "2026-03-10", "categories": []}
{"id": "2603.10046", "title": "Gated Adaptation for Continual Learning in Human Activity Recognition", "abstract": "Wearable sensors in Internet of Things (IoT) ecosystems increasingly support applications such as remote health monitoring, elderly care, and smart home automation, all of which rely on robust human activity recognition (HAR). Continual learning systems must balance plasticity (learning new tasks) with stability (retaining prior knowledge), yet AI models often exhibit catastrophic forgetting, where learning new tasks degrades performance on earlier ones. This challenge is especially acute in domain-incremental HAR, where on-device models must adapt to new subjects with distinct movement patterns while maintaining accuracy on prior subjects without transmitting sensitive data to the cloud. We propose a parameter-efficient continual learning framework based on channel-wise gated modulation of frozen pretrained representations. Our key insight is that adaptation should operate through feature selection rather than feature generation: by restricting learned transformations to diagonal scaling of existing features, we preserve the geometry of pretrained representations while enabling subject-specific modulation. We provide a theoretical analysis showing that gating implements a bounded diagonal operator that limits representational drift compared to unconstrained linear transformations. Empirically, freezing the backbone substantially reduces forgetting, and lightweight gates restore lost adaptation capacity, achieving stability and plasticity simultaneously. On PAMAP2 with 8 sequential subjects, our approach reduces forgetting from 39.7% to 16.2% and improves final accuracy from 56.7% to 77.7%, while training less than 2% of parameters. Our method matches or exceeds standard continual learning baselines without replay buffers or task-specific regularization, confirming that structured diagonal operators are effective and efficient under distribution shift.", "date": "2026-03-08", "categories": []}
{"id": "2603.10237", "title": "One Adapter for All: Towards Unified Representation in Step-Imbalanced Class-Incremental Learning", "abstract": "Class-incremental learning (CIL) aims to acquire new classes over time while retaining prior knowledge, yet most setups and methods assume balanced task streams. In practice, the number of classes per task often varies significantly. We refer to this as step imbalance, where large tasks that contain more classes dominate learning and small tasks inject unstable updates. Existing CIL methods assume balanced tasks and therefore treat all tasks uniformly, producing imbalanced updates that degrade overall learning performance. To address this challenge, we propose One-A, a unified and imbalance-aware framework that incrementally merges task updates into a single adapter, maintaining constant inference cost. One-A performs asymmetric subspace alignment to preserve dominant subspaces learned from large tasks while constraining low-information updates within them. An information-adaptive weighting balances the contribution between base and new adapters, and a directional gating mechanism selectively fuses updates along each singular direction, maintaining stability in head directions and plasticity in tail ones. Across multiple benchmarks and step-imbalanced streams, One-A achieves competitive accuracy with significantly low inference overhead, showing that a single, asymmetrically fused adapter can remain both adaptive to dynamic task sizes and efficient at deployment.", "date": "2026-03-10", "categories": []}
{"id": "2511.00097", "title": "GraphKeeper: Graph Domain-Incremental Learning via Knowledge Disentanglement and Preservation", "abstract": "Graph incremental learning (GIL), which continuously updates graph models by sequential knowledge acquisition, has garnered significant interest recently. However, existing GIL approaches focus on task-incremental and class-incremental scenarios within a single domain. Graph domain-incremental learning (Domain-IL), aiming at updating models across multiple graph domains, has become critical with the development of graph foundation models (GFMs), but remains unexplored in the literature. In this paper, we propose Graph Domain-Incremental Learning via Knowledge Dientanglement and Preservation (GraphKeeper), to address catastrophic forgetting in Domain-IL scenario from the perspectives of embedding shifts and decision boundary deviations. Specifically, to prevent embedding shifts and confusion across incremental graph domains, we first propose the domain-specific parameter-efficient fine-tuning together with intra- and inter-domain disentanglement objectives. Consequently, to maintain a stable decision boundary, we introduce deviation-free knowledge preservation to continuously fit incremental domains. Additionally, for graphs with unobservable domains, we perform domain-aware distribution discrimination to obtain precise embeddings. Extensive experiments demonstrate the proposed GraphKeeper achieves state-of-the-art results with 6.5%~16.6% improvement over the runner-up with negligible forgetting. Moreover, we show GraphKeeper can be seamlessly integrated with various representative GFMs, highlighting its broad applicative potential.", "date": "2025-10-30", "categories": []}
{"id": "2602.12236", "title": "Energy-Aware Spike Budgeting for Continual Learning in Spiking Neural Networks for Neuromorphic Vision", "abstract": "Neuromorphic vision systems based on spiking neural networks (SNNs) offer ultra-low-power perception for event-based and frame-based cameras, yet catastrophic forgetting remains a critical barrier to deployment in continually evolving environments. Existing continual learning methods, developed primarily for artificial neural networks, seldom jointly optimize accuracy and energy efficiency, with particularly limited exploration on event-based datasets. We propose an energy-aware spike budgeting framework for continual SNN learning that integrates experience replay, learnable leaky integrate-and-fire neuron parameters, and an adaptive spike scheduler to enforce dataset-specific energy constraints during training. Our approach exhibits modality-dependent behavior: on frame-based datasets (MNIST, CIFAR-10), spike budgeting acts as a sparsity-inducing regularizer, improving accuracy while reducing spike rates by up to 47\\%; on event-based datasets (DVS-Gesture, N-MNIST, CIFAR-10-DVS), controlled budget relaxation enables accuracy gains up to 17.45 percentage points with minimal computational overhead. Across five benchmarks spanning both modalities, our method demonstrates consistent performance improvements while minimizing dynamic power consumption, advancing the practical viability of continual learning in neuromorphic vision systems.", "date": "2026-02-12", "categories": []}
{"id": "2603.08763", "title": "SPREAD: Subspace Representation Distillation for Lifelong Imitation Learning", "abstract": "A key challenge in lifelong imitation learning (LIL) is enabling agents to acquire new skills from expert demonstrations while retaining prior knowledge. This requires preserving the low-dimensional manifolds and geometric structures that underlie task representations across sequential learning. Existing distillation methods, which rely on L2-norm feature matching in raw feature space, are sensitive to noise and high-dimensional variability, often failing to preserve intrinsic task manifolds. To address this, we introduce SPREAD, a geometry-preserving framework that employs singular value decomposition (SVD) to align policy representations across tasks within low-rank subspaces. This alignment maintains the underlying geometry of multimodal features, facilitating stable transfer, robustness, and generalization. Additionally, we propose a confidence-guided distillation strategy that applies a Kullback-Leibler divergence loss restricted to the top-M most confident action samples, emphasizing reliable modes and improving optimization stability. Experiments on the LIBERO, lifelong imitation learning benchmark, show that SPREAD substantially improves knowledge transfer, mitigates catastrophic forgetting, and achieves state-of-the-art performance.", "date": "2026-03-09", "categories": []}
{"id": "2603.09576", "title": "Routing without Forgetting", "abstract": "Continual learning in transformers is commonly addressed through parameter-efficient adaptation: prompts, adapters, or LoRA modules are specialized per task while the backbone remains frozen. Although effective in controlled multi-epoch settings, these approaches rely on gradual gradient-based specialization and struggle in Online Continual Learning (OCL), where data arrive as a non-stationary stream and each sample may be observed only once. We recast continual learning in transformers as a routing problem: under strict online constraints, the model must dynamically select the appropriate representational subspace for each input without explicit task identifiers or repeated optimization. We thus introduce Routing without Forgetting (RwF), a transformer architecture augmented with energy-based associative retrieval layers inspired by Modern Hopfield Networks. Instead of storing or merging task-specific prompts, RwF generates dynamic prompts through single-step associative retrieval over the transformer token embeddings at each layer. Retrieval corresponds to the closed-form minimization of a strictly convex free-energy functional, enabling input-conditioned routing within each forward pass, independently of iterative gradient refinement. Across challenging class-incremental benchmarks, RwF improves over existing prompt-based methods. On Split-ImageNet-R and Split-ImageNet-S, RwF outperforms prior prompt-based approaches by a large margin, even in few-shot learning regimes. These results indicate that embedding energy-based associative routing directly within the transformer backbone provides a principled and effective foundation for OCL.", "date": "2026-03-10", "categories": []}
{"id": "2603.09684", "title": "On Catastrophic Forgetting in Low-Rank Decomposition-Based Parameter-Efficient Fine-Tuning", "abstract": "Parameter-efficient fine-tuning (PEFT) based on low-rank decomposition, such as LoRA, has become a standard for adapting large pretrained models. However, its behavior in sequential learning -- specifically regarding catastrophic forgetting -- remains insufficiently understood. In this work, we present an empirical study showing that forgetting is strongly influenced by the geometry and parameterization of the update subspace. While methods that restrict updates to small, shared matrix subspaces often suffer from task interference, tensor-based decompositions (e.g., LoRETTA) mitigate forgetting by capturing richer structural information within ultra-compact budgets, and structurally aligned parameterizations (e.g., WeGeFT) preserve pretrained representations. Our findings highlight update subspace design as a key factor in continual learning and offer practical guidance for selecting efficient adaptation strategies in sequential settings.", "date": "2026-03-10", "categories": []}
{"id": "2603.09892", "title": "MSSR: Memory-Aware Adaptive Replay for Continual LLM Fine-Tuning", "abstract": "Continual fine-tuning of large language models (LLMs) is becoming increasingly crucial as these models are deployed in dynamic environments where tasks and data distributions evolve over time. While strong adaptability enables rapid acquisition of new knowledge, it also exposes LLMs to catastrophic forgetting, where previously learned skills degrade during sequential training. Existing replay-based strategies, such as fixed interleaved replay, accuracy-supervised, and loss-driven scheduling, remain limited: some depend on heuristic rules and provide only partial mitigation of forgetting, while others improve performance but incur substantial computational overhead. Motivated by retention dynamics under sequential fine-tuning, we propose Memory-Inspired Sampler and Scheduler Replay (MSSR), an experience replay framework that estimates sample-level memory strength and schedules rehearsal at adaptive intervals to mitigate catastrophic forgetting while maintaining fast adaptation. Extensive experiments across three backbone models and 11 sequential tasks show that MSSR consistently outperforms state-of-the-art replay baselines, with particularly strong gains on reasoning-intensive and multiple-choice benchmarks.", "date": "2026-03-10", "categories": []}
{"id": "2412.08528", "title": "Efficient Continual Learning for Small Language Models with a Discrete Key-Value Bottleneck", "abstract": "Continual learning remains a challenge across various natural language processing (NLP) tasks, as models updated with new training data often risk catastrophic forgetting of previously acquired knowledge. We introduce a discrete key-value bottleneck (DKVB) for encoder-only language models, enabling efficient continual learning through localized updates. Inspired by a discrete key-value bottleneck in vision, we consider new and NLP-specific challenges. We compare different bottleneck architectures for NLP and introduce a new, task-independent initialization technique for the discrete keys. We evaluate our DKVB for NLP in four continual learning scenarios and show that it alleviates catastrophic forgetting. Our experiments demonstrate that the proposed approach achieves competitive performance compared to popular continual learning methods while incurring lower computational costs. Furthermore, we show that DKVB remains effective even in challenging single-head continual learning scenarios where no task ID is provided.", "date": "2024-12-11", "categories": []}
{"id": "2505.18347", "title": "The Cell Must Go On: Agar.io for Continual Reinforcement Learning", "abstract": "Continual reinforcement learning (RL) concerns agents that are expected to learn continually, rather than converge to a policy that is then fixed for evaluation. This setting is well-suited to environments that the agent perceives as changing over time, rendering any static policy ineffective. In continual RL, researchers often simulate such changes either by modifying episodic environments to incorporate task shifts during interaction or by designing simulators that explicitly model continual dynamics. However, transforming episodic problems into continual ones primarily captures scenarios involving abrupt changes in the data stream and still relies on episodic structure. Meanwhile, the few simulators explicitly designed for empirical continual RL research are often limited in scope or complexity. In this paper, we introduce AgarCL, a research platform for continual RL that enables agents to progress toward increasingly sophisticated behaviour. AgarCL is based on the game Agar.io, a non-episodic, high-dimensional problem with stochastic, ever-evolving dynamics, continuous actions, and partial observability. We provide benchmark results for DQN, PPO, and SAC on the primary continual RL challenge, as well as across a suite of smaller tasks within AgarCL. These smaller tasks isolate aspects of the full environment and allow us to characterize the distinct challenges posed by different components of the game. We further evaluate three continual learning methods-Shrink and Perturb, ReDo, and Continual Backpropagation-and observe little improvement over standard RL algorithms, suggesting that the challenges posed by AgarCL extend beyond the stability-plasticity dilemma.", "date": "2025-05-23", "categories": []}
{"id": "2505.24254", "title": "Rethinking Continual Learning with Progressive Neural Collapse", "abstract": "Continual Learning (CL) seeks to build an agent that can continuously learn a sequence of tasks, where a key challenge, namely Catastrophic Forgetting, persists due to the potential knowledge interference among different tasks. On the other hand, deep neural networks (DNNs) are shown to converge to a terminal state termed Neural Collapse during training, where all class prototypes geometrically form a static simplex equiangular tight frame (ETF). These maximally and equally separated class prototypes make the ETF an ideal target for model learning in CL to mitigate knowledge interference. Thus inspired, several studies have emerged very recently to leverage a fixed global ETF in CL, which however suffers from key drawbacks, such as impracticability and limited performance.To address these challenges and fully unlock the potential of ETF in CL, we propose Progressive Neural Collapse (ProNC), a novel framework that completely removes the need of a fixed global ETF in CL. Specifically, ProNC progressively expands the ETF target in a principled way by adding new class prototypes as vertices for new tasks, ensuring maximal separability across all encountered classes with minimal shifts from the previous ETF. We next develop a new CL framework by plugging ProNC into commonly used CL algorithm designs, where distillation is further leveraged to balance between target shifting for old classes and target aligning for new classes. Extensive experiments show that our approach significantly outperforms related baselines while maintaining superior flexibility, simplicity, and efficiency. Our code is available at https://github.com/Continue-Edge-AI-Lab/ProNC", "date": "2025-05-30", "categories": []}
{"id": "2509.23077", "title": "CLAD-Net: Continual Activity Recognition in Multi-Sensor Wearable Systems", "abstract": "The rise of deep learning has greatly advanced human behavior monitoring using wearable sensors, particularly human activity recognition (HAR). While deep models have been widely studied, most assume stationary data distributions - an assumption often violated in real-world scenarios. For example, sensor data from one subject may differ significantly from another, leading to distribution shifts. In continual learning, this shift is framed as a sequence of tasks, each corresponding to a new subject. Such settings suffer from catastrophic forgetting, where prior knowledge deteriorates as new tasks are learned. This challenge is compounded by the scarcity and inconsistency of labeled data in human studies. To address these issues, we propose CLAD-Net (Continual Learning with Attention and Distillation), a framework enabling wearable-sensor models to be updated continuously without sacrificing performance on past tasks. CLAD-Net integrates a self-supervised transformer, acting as long-term memory, with a supervised Convolutional Neural Network (CNN) trained via knowledge distillation for activity classification. The transformer captures global activity patterns through cross-attention across body-mounted sensors, learning generalizable representations without labels. Meanwhile, the CNN leverages knowledge distillation to retain prior knowledge during subject-wise fine-tuning. On PAMAP2, CLAD-Net achieves 91.36 percent final accuracy with only 8.78 percent forgetting, surpassing memory-based and regularization-based baselines such as Experience Replay and Elastic Weight Consolidation. In semi-supervised settings with only 10-20 percent labeled data, CLAD-Net still delivers strong performance, demonstrating robustness to label scarcity. Ablation studies further validate each module's contribution.", "date": "2025-09-27", "categories": []}
{"id": "2510.17234", "title": "Taming Modality Entanglement in Continual Audio-Visual Segmentation", "abstract": "Recently, significant progress has been made in multi-modal continual learning, aiming to learn new tasks sequentially in multi-modal settings while preserving performance on previously learned ones. However, existing methods mainly focus on coarse-grained tasks, with limitations in addressing modality entanglement in fine-grained continual learning settings. To bridge this gap, we introduce a novel Continual Audio-Visual Segmentation (CAVS) task, aiming to continuously segment new classes guided by audio. Through comprehensive analysis, two critical challenges are identified: 1) multi-modal semantic drift, where a sounding objects is labeled as background in sequential tasks; 2) co-occurrence confusion, where frequent co-occurring classes tend to be confused. In this work, a Collision-based Multi-modal Rehearsal (CMR) framework is designed to address these challenges. Specifically, for multi-modal semantic drift, a Multi-modal Sample Selection (MSS) strategy is proposed to select samples with high modal consistency for rehearsal. Meanwhile, for co-occurence confusion, a Collision-based Sample Rehearsal (CSR) mechanism is designed, allowing for the increase of rehearsal sample frequency of those confusable classes during training process. Moreover, we construct three audio-visual incremental scenarios to verify effectiveness of our method. Comprehensive experiments demonstrate that our method significantly outperforms single-modal continual learning methods.", "date": "2025-10-20", "categories": []}
{"id": "2603.02951", "title": "CGL: Advancing Continual GUI Learning via Reinforcement Fine-Tuning", "abstract": "Graphical User Interface (GUI) Agents, benefiting from recent advances in multimodal large language models (MLLM), have achieved significant development. However, due to the frequent updates of GUI applications, adapting to new tasks without forgetting old tasks in GUI continual learning remains an open problem. In this work, we reveal that while Supervised Fine-Tuning (SFT) facilitates fast adaptation, it often triggers knowledge overwriting, whereas Reinforcement Learning (RL) demonstrates an inherent resilience that shields prior interaction logic from erasure. Based on this insight, we propose a \\textbf{C}ontinual \\textbf{G}UI \\textbf{L}earning (CGL) framework that dynamically balances adaptation efficiency and skill retention by enhancing the synergy between SFT and RL. Specifically, we introduce an SFT proportion adjustment mechanism guided by policy entropy to dynamically control the weight allocation between the SFT and RL training phases. To resolve explicit gradient interference, we further develop a specialized gradient surgery strategy. By projecting exploratory SFT gradients onto GRPO-based anchor gradients, our method explicitly clips the components of SFT gradients that conflict with GRPO. On top of that, we establish an AndroidControl-CL benchmark, which divides GUI applications into distinct task groups to effectively simulate and evaluate the performance of continual GUI learning. Experimental results demonstrate the effectiveness of our proposed CGL framework across continual learning scenarios. The benchmark, code, and model will be made publicly available.", "date": "2026-03-03", "categories": []}
{"id": "2603.07392", "title": "Can Large Language Models Keep Up? Benchmarking Online Adaptation to Continual Knowledge Streams", "abstract": "LLMs operating in dynamic real-world contexts often encounter knowledge that evolves continuously or emerges incrementally. To remain accurate and effective, models must adapt to newly arriving information on the fly. We introduce Online Adaptation to Continual Knowledge Streams(OAKS) to evaluate this capability, establishing a benchmark for online adaptation over streaming, continually updating knowledge. Specifically, the benchmark is structured as a sequence of fine-grained context chunks where facts change dynamically across time intervals. OAKS comprises two datasets: OAKS-BABI and OAKS-Novel, where individual facts evolve multiple times across context chunks. These datasets include dense annotations to measure whether models track changes accurately. Evaluating 14 models with varied inference approaches, we observe significant limitations in current methodologies. Both state-of-the-art models and agentic memory systems fail to adapt robustly on OAKS, demonstrating delays in state-tracking and susceptibility to distraction within streaming environments.", "date": "2026-03-08", "categories": []}
{"id": "2603.07415", "title": "Context Channel Capacity: An Information-Theoretic Framework for Understanding Catastrophic Forgetting", "abstract": "Catastrophic forgetting remains a central challenge in continual learning (CL), yet lacks a unified information-theoretic explanation for why some architectures forget catastrophically while others do not. We introduce \\emph{Context Channel Capacity} ($C_\\mathrm{ctx}$), the mutual information between a CL architecture's context signal and its generated parameters, and prove that zero forgetting requires $C_\\mathrm{ctx} \\geq H(T)$, where $H(T)$ is the task identity entropy. We establish an \\emph{Impossibility Triangle} -- zero forgetting, online learning, and finite parameters cannot be simultaneously satisfied by sequential state-based learners -- and show that conditional regeneration architectures (HyperNetworks) bypass this triangle by redefining parameters as function values rather than states.\n  We validate this framework across 8 CL methods on Split-MNIST (1,130+ experiments over 86 days, 4 seeds each), showing that $C_\\mathrm{ctx}$ perfectly predicts forgetting behavior: methods with $C_\\mathrm{ctx} = 0$ (NaiveSGD, EWC, SI, LwF, CFlow) exhibit catastrophic forgetting (6--97\\%), while methods with $C_\\mathrm{ctx} \\approx 1$ (HyperNetwork) achieve zero forgetting (98.8\\% ACC). We further propose \\emph{Wrong-Context Probing} (P5), a practical diagnostic protocol for measuring $C_\\mathrm{ctx}$, and extend the framework to CIFAR-10 via a novel \\emph{Gradient Context Encoder} that closes the oracle gap from 23.3pp to 0.7pp. A systematic taxonomy of 15+ closed research directions -- including the Hebbian null result (frozen random features outperform learned features), CFlow's $\\theta_0$-memorizer phenomenon, and the $S_N$ symmetry barrier to column specialization -- provides the community with precisely diagnosed negative results. Our central design principle: \\emph{architecture over algorithm} -- the context pathway must be structurally unbypassable.", "date": "2026-03-08", "categories": []}
{"id": "2603.07497", "title": "AMR-CCR: Anchored Modular Retrieval for Continual Chinese Character Recognition", "abstract": "Ancient Chinese character recognition is a core capability for cultural heritage digitization, yet real-world workflows are inherently non-stationary: newly excavated materials are continuously onboarded, bringing new classes in different scripts, and expanding the class space over time. We formalize this process as Continual Chinese Character Recognition (Continual CCR), a script-staged, class-incremental setting that couples two challenges: (i) scalable learning under continual class growth with subtle inter-class differences and scarce incremental data, and (ii) pronounced intra-class diversity caused by writing-style variations across writers and carrier conditions. To overcome the limitations of conventional closed-set classification, we propose AMR-CCR, an anchored modular retrieval framework that performs recognition via embedding-based dictionary matching in a shared multimodal space, allowing new classes to be added by simply extending the dictionary. AMR-CCR further introduces a lightweight script-conditioned injection module (SIA+SAR) to calibrate newly onboarded scripts while preserving cross-stage embedding compatibility, and an image-derived multi-prototype dictionary that clusters within-class embeddings to better cover diverse style modes. To support systematic evaluation, we build EvoCON, a six-stage benchmark for continual script onboarding, covering six scripts (OBC, BI, SS, SAC, WSC, CS), augmented with meaning/shape descriptions and an explicit zero-shot split for unseen characters without image exemplars.", "date": "2026-03-08", "categories": []}
{"id": "2603.07507", "title": "Online Continual Learning for Anomaly Detection in IoT under Data Distribution Shifts", "abstract": "In this work, we present OCLADS, a novel communication framework with continual learning (CL) for Internet of Things (IoT) anomaly detection (AD) when operating in non-stationary environments. As the statistical properties of the observed data change with time, the on-device inference model becomes obsolete, which necessitates strategic model updating. OCLADS keeps track of data distribution shifts to timely update the on-device IoT AD model. To do so, OCLADS introduces two mechanisms during the interaction between the resource-constrained IoT device and an edge server (ES): i) an intelligent sample selection mechanism at the device for data transmission, and ii) a distribution-shift detection mechanism at the ES for model updating. Experimental results with TinyML demonstrate that our proposed framework achieves high inference accuracy while realizing a significantly smaller number of model updates compared to the baseline schemes.", "date": "2026-03-08", "categories": []}
{"id": "2603.07784", "title": "ProgAgent:A Continual RL Agent with Progress-Aware Rewards", "abstract": "We present ProgAgent, a continual reinforcement learning (CRL) agent that unifies progress-aware reward learning with a high-throughput, JAX-native system architecture. Lifelong robotic learning grapples with catastrophic forgetting and the high cost of reward specification. ProgAgent tackles these by deriving dense, shaped rewards from unlabeled expert videos through a perceptual model that estimates task progress across initial, current, and goal observations. We theoretically interpret this as a learned state-potential function, delivering robust guidance in line with expert behaviors. To maintain stability amid online exploration - where novel, out-of-distribution states arise - we incorporate an adversarial push-back refinement that regularizes the reward model, curbing overconfident predictions on non-expert trajectories and countering distribution shift. By embedding this reward mechanism into a JIT-compiled loop, ProgAgent supports massively parallel rollouts and fully differentiable updates, rendering a sophisticated unified objective feasible: it merges PPO with coreset replay and synaptic intelligence for an enhanced stability-plasticity balance. Evaluations on ContinualBench and Meta-World benchmarks highlight ProgAgent's advantages: it markedly reduces forgetting, boosts learning speed, and outperforms key baselines in visual reward learning (e.g., Rank2Reward, TCN) and continual learning (e.g., Coreset, SI) - surpassing even an idealized perfect memory agent. Real-robot trials further validate its ability to acquire complex manipulation skills from noisy, few-shot human demonstrations.", "date": "2026-03-08", "categories": []}
{"id": "2603.07787", "title": "Vision Transformers that Never Stop Learning", "abstract": "Loss of plasticity refers to the progressive inability of a model to adapt to new tasks and poses a fundamental challenge for continual learning. While this phenomenon has been extensively studied in homogeneous neural architectures, such as multilayer perceptrons, its mechanisms in structurally heterogeneous, attention-based models such as Vision Transformers (ViTs) remain underexplored. In this work, we present a systematic investigation of loss of plasticity in ViTs, including a fine-grained diagnosis using local metrics that capture parameter diversity and utilization. Our analysis reveals that stacked attention modules exhibit increasing instability that exacerbates plasticity loss, while feed-forward network modules suffer even more pronounced degradation. Furthermore, we evaluate several approaches for mitigating plasticity loss. The results indicate that methods based on parameter re-initialization fail to recover plasticity in ViTs, whereas approaches that explicitly regulate the update process are more effective. Motivated by this insight, we propose ARROW, a geometry-aware optimizer that preserves plasticity by adaptively reshaping gradient directions using an online curvature estimate for the attention module. Extensive experiments show that ARROW effectively improves plasticity and maintains better performance on newly encountered tasks.", "date": "2026-03-08", "categories": []}
{"id": "2603.08600", "title": "Don't Look Back in Anger: MAGIC Net for Streaming Continual Learning with Temporal Dependence", "abstract": "Concept drift, temporal dependence, and catastrophic forgetting represent major challenges when learning from data streams. While Streaming Machine Learning and Continual Learning (CL) address these issues separately, recent efforts in Streaming Continual Learning (SCL) aim to unify them. In this work, we introduce MAGIC Net, a novel SCL approach that integrates CL-inspired architectural strategies with recurrent neural networks to tame temporal dependence. MAGIC Net continuously learns, looks back at past knowledge by applying learnable masks over frozen weights, and expands its architecture when necessary. It performs all operations online, ensuring inference availability at all times. Experiments on synthetic and real-world streams show that it improves adaptation to new concepts, limits memory usage, and mitigates forgetting.", "date": "2026-03-09", "categories": []}
{"id": "2603.08647", "title": "Grow, Don't Overwrite: Fine-tuning Without Forgetting", "abstract": "Adapting pre-trained models to specialized tasks often leads to catastrophic forgetting, where new knowledge overwrites foundational capabilities. Existing methods either compromise performance on the new task or struggle to balance training stability with efficient reuse of pre-trained knowledge. We introduce a novel function-preserving expansion method that resolves this dilemma. Our technique expands model capacity by replicating pre-trained parameters within transformer submodules and applying a scaling correction that guarantees the expanded model is mathematically identical to the original at initialization, enabling stable training while exploiting existing knowledge. Empirically, our method eliminates the trade-off between plasticity and stability, matching the performance of full fine-tuning on downstream tasks without any degradation of the model's original capabilities. Furthermore, we demonstrate the modularity of our approach, showing that by selectively expanding a small subset of layers we can achieve the same performance as full fine-tuning at a fraction of the computational cost.", "date": "2026-03-09", "categories": []}
{"id": "2509.19674", "title": "C^2Prompt: Class-aware Client Knowledge Interaction for Federated Continual Learning", "abstract": "Federated continual learning (FCL) tackles scenarios of learning from continuously emerging task data across distributed clients, where the key challenge lies in addressing both temporal forgetting over time and spatial forgetting simultaneously. Recently, prompt-based FCL methods have shown advanced performance through task-wise prompt communication.In this study, we underscore that the existing prompt-based FCL methods are prone to class-wise knowledge coherence between prompts across clients. The class-wise knowledge coherence includes two aspects: (1) intra-class distribution gap across clients, which degrades the learned semantics across prompts, (2) inter-prompt class-wise relevance, which highlights cross-class knowledge confusion. During prompt communication, insufficient class-wise coherence exacerbates knowledge conflicts among new prompts and induces interference with old prompts, intensifying both spatial and temporal forgetting. To address these issues, we propose a novel Class-aware Client Knowledge Interaction (C${}^2$Prompt) method that explicitly enhances class-wise knowledge coherence during prompt communication. Specifically, a local class distribution compensation mechanism (LCDC) is introduced to reduce intra-class distribution disparities across clients, thereby reinforcing intra-class knowledge consistency. Additionally, a class-aware prompt aggregation scheme (CPA) is designed to alleviate inter-class knowledge confusion by selectively strengthening class-relevant knowledge aggregation. Extensive experiments on multiple FCL benchmarks demonstrate that C${}^2$Prompt achieves state-of-the-art performance. Our source code is available at https://github.com/zhoujiahuan1991/NeurIPS2025-C2Prompt", "date": "2025-09-24", "categories": []}
{"id": "2509.23335", "title": "DeCLIP: Decoupled Prompting for CLIP-based Multi-Label Class-Incremental Learning", "abstract": "Multi-label class-incremental learning (MLCIL) continuously expands the label space while recognizing multiple co-occurring classes, making it prone to catastrophic forgetting and high false-positive rates (FPR). Extending CLIP to MLCIL is non-trivial because co-occurring categories violate CLIP's single image-text alignment paradigm and task-level partial labeling induces high FPR. We propose DeCLIP, a replay-free and parameter-efficient framework that decouples CLIP representations via a one-to-one class-specific prompting scheme. By assigning each category its own prompt space, DeCLIP prevents semantic confusion across labels and decouples multi-label images into per-class views compatible with CLIP pre-training. The learned prompts are preserved as knowledge anchors, mitigating catastrophic forgetting without replay. We further introduce Adaptive Similarity Tempering (AST), a task-aware strategy that suppresses FPR without dataset-specific tuning. Experiments on MS-COCO and PASCAL VOC show that DeCLIP consistently outperforms prior methods with minimal trainable parameters.", "date": "2025-09-27", "categories": []}
{"id": "2603.06073", "title": "Lifelong Embodied Navigation Learning", "abstract": "Embodied navigation agents powered by large language models have shown strong performance on individual tasks but struggle to continually acquire new navigation skills, which suffer from catastrophic forgetting. We formalize this challenge as lifelong embodied navigation learning (LENL), where an agent is required to adapt to a sequence of navigation tasks spanning multiple scenes and diverse user instruction styles, while retaining previously learned knowledge. To tackle this problem, we propose Uni-Walker, a lifelong embodied navigation framework that decouples navigation knowledge into task-shared and task-specific components with Decoder Extension LoRA (DE-LoRA). To learn the shared knowledge, we design a knowledge inheritance strategy and an experts co-activation strategy to facilitate shared knowledge transfer and refinement across multiple navigation tasks. To learn the specific knowledge, we propose an expert subspace orthogonality constraint together and a navigation-specific chain-of-thought reasoning mechanism to capture specific knowledge and enhance instruction-style understanding. Extensive experiments demonstrate the superiority of Uni-Walker for building universal navigation agents with lifelong learning.", "date": "2026-03-06", "categories": []}
{"id": "2603.04580", "title": "Why Do Neural Networks Forget: A Study of Collapse in Continual Learning", "abstract": "Catastrophic forgetting is a major problem in continual learning, and lots of approaches arise to reduce it. However, most of them are evaluated through task accuracy, which ignores the internal model structure. Recent research suggests that structural collapse leads to loss of plasticity, as evidenced by changes in effective rank (eRank). This indicates a link to forgetting, since the networks lose the ability to expand their feature space to learn new tasks, which forces the network to overwrite existing representations. Therefore, in this study, we investigate the correlation between forgetting and collapse through the measurement of both weight and activation eRank. To be more specific, we evaluated four architectures, including MLP, ConvGRU, ResNet-18, and Bi-ConvGRU, in the split MNIST and Split CIFAR-100 benchmarks. Those models are trained through the SGD, Learning-without-Forgetting (LwF), and Experience Replay (ER) strategies separately. The results demonstrate that forgetting and collapse are strongly related, and different continual learning strategies help models preserve both capacity and performance in different efficiency.", "date": "2026-03-04", "categories": []}
{"id": "2603.05160", "title": "Lifelong Language-Conditioned Robotic Manipulation Learning", "abstract": "Traditional language-conditioned manipulation agent sequential adaptation to new manipulation skills leads to catastrophic forgetting of old skills, limiting dynamic scene practical deployment. In this paper, we propose SkillsCrafter, a novel robotic manipulation framework designed to continually learn multiple skills while reducing catastrophic forgetting of old skills. Specifically, we propose a Manipulation Skills Adaptation to retain the old skills knowledge while inheriting the shared knowledge between new and old skills to facilitate learning of new skills. Meanwhile, we perform the singular value decomposition on the diverse skill instructions to obtain common skill semantic subspace projection matrices, thereby recording the essential semantic space of skills. To achieve forget-less and generalization manipulation, we propose a Skills Specialization Aggregation to compute inter-skills similarity in skill semantic subspaces, achieving aggregation of the previously learned skill knowledge for any new or unknown skill. Extensive experiments demonstrate the effectiveness and superiority of our proposed SkillsCrafter.", "date": "2026-03-05", "categories": []}
{"id": "2509.22580", "title": "The Lie of the Average: How Class Incremental Learning Evaluation Deceives You?", "abstract": "Class Incremental Learning (CIL) requires models to continuously learn new classes without forgetting previously learned ones, while maintaining stable performance across all possible class sequences. In real-world settings, the order in which classes arrive is diverse and unpredictable, and model performance can vary substantially across different sequences. Yet mainstream evaluation protocols calculate mean and variance from only a small set of randomly sampled sequences. Our theoretical analysis and empirical results demonstrate that this sampling strategy fails to capture the full performance range, resulting in biased mean estimates and a severe underestimation of the true variance in the performance distribution. We therefore contend that a robust CIL evaluation protocol should accurately characterize and estimate the entire performance distribution. To this end, we introduce the concept of extreme sequences and provide theoretical justification for their crucial role in the reliable evaluation of CIL. Moreover, we observe a consistent positive correlation between inter-task similarity and model performance, a relation that can be leveraged to guide the search for extreme sequences. Building on these insights, we propose EDGE (Extreme case-based Distribution & Generalization Evaluation), an evaluation protocol that adaptively identifies and samples extreme class sequences using inter-task similarity, offering a closer approximation of the ground-truth performance distribution. Extensive experiments demonstrate that EDGE effectively captures performance extremes and yields more accurate estimates of distributional boundaries, providing actionable insights for model selection and robustness checking. Our code is available at https://github.com/AIGNLAI/EDGE.", "date": "2025-09-26", "categories": []}
{"id": "2603.04321", "title": "SPRINT: Semi-supervised Prototypical Representation for Few-Shot Class-Incremental Tabular Learning", "abstract": "Real-world systems must continuously adapt to novel concepts from limited data without forgetting previously acquired knowledge. While Few-Shot Class-Incremental Learning (FSCIL) is established in computer vision, its application to tabular domains remains largely unexplored. Unlike images, tabular streams (e.g., logs, sensors) offer abundant unlabeled data, a scarcity of expert annotations and negligible storage costs, features ignored by existing vision-based methods that rely on restrictive buffers. We introduce SPRINT, the first FSCIL framework tailored for tabular distributions. SPRINT introduces a mixed episodic training strategy that leverages confidence-based pseudo-labeling to enrich novel class representations and exploits low storage costs to retain base class history. Extensive evaluation across six diverse benchmarks spanning cybersecurity, healthcare, and ecological domains, demonstrates SPRINT's cross-domain robustness. It achieves a state-of-the-art average accuracy of 77.37% (5-shot), outperforming the strongest incremental baseline by 4.45%.", "date": "2026-03-04", "categories": []}
{"id": "2602.23201", "title": "Tell Me What To Learn: Generalizing Neural Memory to be Controllable in Natural Language", "abstract": "Modern machine learning models are deployed in diverse, non-stationary environments where they must continually adapt to new tasks and evolving knowledge. Continual fine-tuning and in-context learning are costly and brittle, whereas neural memory methods promise lightweight updates with minimal forgetting. However, existing neural memory models typically assume a single fixed objective and homogeneous information streams, leaving users with no control over what the model remembers or ignores over time. To address this challenge, we propose a generalized neural memory system that performs flexible updates based on learning instructions specified in natural language. Our approach enables adaptive agents to learn selectively from heterogeneous information sources, supporting settings, such as healthcare and customer service, where fixed-objective memory updates are insufficient.", "date": "2026-02-26", "categories": []}
{"id": "2603.00624", "title": "IDER: IDempotent Experience Replay for Reliable Continual Learning", "abstract": "Catastrophic forgetting, the tendency of neural networks to forget previously learned knowledge when learning new tasks, has been a major challenge in continual learning (CL). To tackle this challenge, CL methods have been proposed and shown to reduce forgetting. Furthermore, CL models deployed in mission-critical settings can benefit from uncertainty awareness by calibrating their predictions to reliably assess their confidences. However, existing uncertainty-aware continual learning methods suffer from high computational overhead and incompatibility with mainstream replay methods. To address this, we propose idempotent experience replay (IDER), a novel approach based on the idempotent property where repeated function applications yield the same output. Specifically, we first adapt the training loss to make model idempotent on current data streams. In addition, we introduce an idempotence distillation loss. We feed the output of the current model back into the old checkpoint and then minimize the distance between this reprocessed output and the original output of the current model. This yields a simple and effective new baseline for building reliable continual learners, which can be seamlessly integrated with other CL approaches. Extensive experiments on different CL benchmarks demonstrate that IDER consistently improves prediction reliability while simultaneously boosting accuracy and reducing forgetting. Our results suggest the potential of idempotence as a promising principle for deploying efficient and trustworthy continual learning systems in real-world applications.Our code is available at https://github.com/YutingLi0606/Idempotent-Continual-Learning.", "date": "2026-02-28", "categories": []}
{"id": "2603.02224", "title": "Subspace Geometry Governs Catastrophic Forgetting in Low-Rank Adaptation", "abstract": "Low-Rank Adaptation (LoRA) has emerged as a parameter-efficient approach for adapting large pre-trained models, yet its behavior under continual learning remains poorly understood. We present a geometric theory characterizing catastrophic forgetting in LoRA through the lens of gradient subspace interactions. Our central finding is that forgetting is governed by a simple geometric law: $\\mathcal{F} = \\alpha(1 - \\cos^2\\theta_{\\min}) + \\beta$, where $\\theta_{\\min}$ is the minimum principal angle between task gradient subspaces. This formulation reveals an approximate rank-invariance property, at high subspace angles, forgetting becomes largely independent of the adapter rank (coefficient of variation $\\approx 0.8\\%$ in controlled synthetic settings; CV $\\approx 10$-$19\\%$ on real benchmarks, suggesting this is regime-dependent rather than absolute). We validate our theory on synthetic tasks ($r=0.994$ correlation), Split-CIFAR100 with ViT-LoRA, and sequential GLUE with RoBERTa-LoRA. Our analysis reconciles seemingly contradictory findings in the literature: we show that rank affects forgetting only when task subspaces are similar (low angle), while orthogonal methods like O-LoRA provide minimal benefit when natural orthogonality is already high. These insights provide principled guidance for continual learning with parameter-efficient fine-tuning.", "date": "2026-02-10", "categories": []}
{"id": "2603.02280", "title": "Temporal Imbalance of Positive and Negative Supervision in Class-Incremental Learning", "abstract": "With the widespread adoption of deep learning in visual tasks, Class-Incremental Learning (CIL) has become an important paradigm for handling dynamically evolving data distributions. However, CIL faces the core challenge of catastrophic forgetting, often manifested as a prediction bias toward new classes. Existing methods mainly attribute this bias to intra-task class imbalance and focus on corrections at the classifier head. In this paper, we highlight an overlooked factor -- temporal imbalance -- as a key cause of this bias. Earlier classes receive stronger negative supervision toward the end of training, leading to asymmetric precision and recall. We establish a temporal supervision model, formally define temporal imbalance, and propose Temporal-Adjusted Loss (TAL), which uses a temporal decay kernel to construct a supervision strength vector and dynamically reweight the negative supervision in cross-entropy loss. Theoretical analysis shows that TAL degenerates to standard cross-entropy under balanced conditions and effectively mitigates prediction bias under imbalance. Extensive experiments demonstrate that TAL significantly reduces forgetting and improves performance on multiple CIL benchmarks, underscoring the importance of temporal modeling for stable long-term learning.", "date": "2026-03-02", "categories": []}
{"id": "2501.06762", "title": "Improving the adaptive and continuous learning capabilities of artificial neural networks: Lessons from multi-neuromodulatory dynamics", "abstract": "Continuous, adaptive learning, the ability to adapt to the environment and keep improving performance, is a hallmark of natural intelligence. Biological organisms excel in acquiring, transferring, and retaining knowledge while adapting to volatile environments, making them a source of inspiration for artificial neural networks (ANNs). This study explores how neuromodulation, a building block of learning in biological systems, can help address catastrophic forgetting and enhance the robustness of ANNs in continual learning. Driven by neuromodulators including dopamine (DA), acetylcholine (ACh), serotonin (5-HT) and noradrenaline (NA), neuromodulatory processes in the brain operate at multiple scales, facilitating dynamic responses to environmental changes through mechanisms ranging from local synaptic plasticity to global network-wide adaptability. Importantly, the relationship between neuromodulators and their interplay in modulating sensory and cognitive processes is more complex than previously expected, demonstrating a \"many-to-one\" neuromodulator-to-task mapping. To inspire neuromodulation-aware learning rules, we highlight (i) how multi-neuromodulatory interactions enrich single-neuromodulator-driven learning, (ii) the impact of neuromodulators across multiple spatio-temporal scales, and correspondingly, (iii) strategies for approximating and integrating neuromodulated learning processes in ANNs. To illustrate these principles, we present a conceptual study to showcase how neuromodulation-inspired mechanisms, such as DA-driven reward processing and NA-based cognitive flexibility, can enhance ANN performance in a Go/No-Go task. Though multi-scale neuromodulation, we aim to bridge the gap between biological and artificial learning, paving the way for ANNs with greater flexibility, robustness, and adaptability.", "date": "2025-01-12", "categories": []}
{"id": "2506.19558", "title": "Consistency-Driven Calibration and Matching for Few-Shot Class-Incremental Learning", "abstract": "Few-Shot Class Incremental Learning (FSCIL) is crucial for adapting to the complex open-world environments. Contemporary prospective learning-based space construction methods struggle to balance old and new knowledge, as prototype bias and rigid structures limit the expressive capacity of the embedding space. Different from these strategies, we rethink the optimization dilemma from the perspective of feature-structure dual consistency, and propose a Consistency-driven Calibration and Matching (ConCM) framework that systematically mitigates the knowledge conflict inherent in FSCIL. Specifically, inspired by hippocampal associative memory, we design a memory-aware prototype calibration that extracts generalized semantic attributes from base classes and reintegrates them into novel classes to enhance the conceptual center consistency of features. Further, to consolidate memory associations, we propose dynamic structure matching, which adaptively aligns the calibrated features to a session-specific optimal manifold space, ensuring cross-session structure consistency. This process requires no class-number priors and is theoretically guaranteed to achieve geometric optimality and maximum matching. On large-scale FSCIL benchmarks including mini-ImageNet, CIFAR100 and CUB200, ConCM achieves state-of-the-art performance, with harmonic accuracy gains of up to 3.41% in incremental sessions. Code is available at: https://github.com/wire-wqz/ConCM", "date": "2025-06-24", "categories": []}
{"id": "2508.17261", "title": "CLIFF: Continual Learning for Incremental Flake Features in 2D Material Identification", "abstract": "Identifying quantum flakes is crucial for scalable quantum hardware; however, automated layer classification from optical microscopy remains challenging due to substantial appearance shifts across different materials. This paper proposes a new Continual-Learning Framework for Flake Layer Classification (CLIFF). To the best of our knowledge, this work represents the first systematic study of continual learning in two-dimensional (2D) materials. The proposed framework enables the model to distinguish materials and their physical and optical properties by freezing the backbone and base head, which are trained on a reference material. For each new material, it learns a material-specific prompt, embedding, and a delta head. A prompt pool and a cosine-similarity gate modulate features and compute material-specific corrections. Additionally, memory replay with knowledge distillation is incorporated. CLIFF achieves competitive accuracy with significantly lower forgetting than naive fine-tuning and a prompt-based baseline.", "date": "2025-08-24", "categories": []}
{"id": "2510.16877", "title": "Fly-CL: A Fly-Inspired Framework for Enhancing Efficient Decorrelation and Reduced Training Time in Pre-trained Model-based Continual Representation Learning", "abstract": "Using a nearly-frozen pretrained model, the continual representation learning paradigm reframes parameter updates as a similarity-matching problem to mitigate catastrophic forgetting. However, directly leveraging pretrained features for downstream tasks often suffers from multicollinearity in the similarity-matching stage, and more advanced methods can be computationally prohibitive for real-time, low-latency applications. Inspired by the fly olfactory circuit, we propose Fly-CL, a bio-inspired framework compatible with a wide range of pretrained backbones. Fly-CL substantially reduces training time while achieving performance comparable to or exceeding that of current state-of-the-art methods. We theoretically show how Fly-CL progressively resolves multicollinearity, enabling more effective similarity matching with low time complexity. Extensive simulation experiments across diverse network architectures and data regimes validate Fly-CL's effectiveness in addressing this challenge through a biologically inspired design. Code is available at https://github.com/gfyddha/Fly-CL.", "date": "2025-10-19", "categories": []}
{"id": "2603.00138", "title": "Latent Replay Detection: Memory-Efficient Continual Object Detection on Microcontrollers via Task-Adaptive Compression", "abstract": "Deploying object detection on microcontrollers (MCUs) enables intelligent edge devices but current models cannot learn new object categories after deployment. Existing continual learning methods require storing raw images far exceeding MCU memory budgets of tens of kilobytes. We present Latent Replay Detection (LRD), the first framework for continual object detection under MCU memory constraints. Our key contributions are: 1. Task-Adaptive Compression: Unlike fixed PCA, we propose learnable compression with FiLM (Feature-wise Linear Modulation) conditioning, where task specific embeddings modulate the compression to preserve discriminative features for each task's distribution; 2. Spatial-Diverse Exemplar Selection: Traditional sampling ignores spatial information critical for detection - we select exemplars maximizing bounding box diversity via farthest-point sampling in IoU space, preventing localization bias in replay; 3. MCU-Deployable System: Our latent replay stores 150 bytes per sample versus >10KB for images, enabling a 64KB buffer to hold 400+ exemplars. Experiments on CORe50 (50 classes, 5 tasks) demonstrate that LRD achieves mAP@50 on the initial task and maintains strong performance across subsequent tasks - a significant improvement over naive fine-tuning while operating within strict MCU constraints. Our task-adaptive FiLM compression and spatial diverse exemplar selection work synergistically to preserve detection capabilities. Deployed on STM32H753ZI, ESP32-S3, and MAX78000 MCUs, LRD achieves 4.9-97.5ms latency per inference within a 64KB memory budget-enabling practical continual detection on edge devices for the first time.", "date": "2026-02-24", "categories": []}
{"id": "2603.00903", "title": "Principled Fast and Meta Knowledge Learners for Continual Reinforcement Learning", "abstract": "Inspired by the human learning and memory system, particularly the interplay between the hippocampus and cerebral cortex, this study proposes a dual-learner framework comprising a fast learner and a meta learner to address continual Reinforcement Learning~(RL) problems. These two learners are coupled to perform distinct yet complementary roles: the fast learner focuses on knowledge transfer, while the meta learner ensures knowledge integration. In contrast to traditional multi-task RL approaches that share knowledge through average return maximization, our meta learner incrementally integrates new experiences by explicitly minimizing catastrophic forgetting, thereby supporting efficient cumulative knowledge transfer for the fast learner. To facilitate rapid adaptation in new environments, we introduce an adaptive meta warm-up mechanism that selectively harnesses past knowledge. We conduct experiments in various pixel-based and continuous control benchmarks, revealing the superior performance of continual learning for our proposed dual-learner approach relative to baseline methods. The code is released in https://github.com/datake/FAME.", "date": "2026-03-01", "categories": []}
{"id": "2603.01524", "title": "Better Matching, Less Forgetting: A Quality-Guided Matcher for Transformer-based Incremental Object Detection", "abstract": "Incremental Object Detection (IOD) aims to continuously learn new object classes without forgetting previously learned ones. A persistent challenge is catastrophic forgetting, primarily attributed to background shift in conventional detectors. While pseudo-labeling mitigates this in dense detectors, we identify a novel, distinct source of forgetting specific to DETR-like architectures: background foregrounding. This arises from the exhaustiveness constraint of the Hungarian matcher, which forcibly assigns every ground truth target to one prediction, even when predictions primarily cover background regions (i.e., low IoU). This erroneous supervision compels the model to misclassify background features as specific foreground classes, disrupting learned representations and accelerating forgetting. To address this, we propose a Quality-guided Min-Cost Max-Flow (Q-MCMF) matcher. To avoid forced assignments, Q-MCMF builds a flow graph and prunes implausible matches based on geometric quality. It then optimizes for the final matching that minimizes cost and maximizes valid assignments. This strategy eliminates harmful supervision from background foregrounding while maximizing foreground learning signals. Extensive experiments on the COCO dataset under various incremental settings demonstrate that our method consistently outperforms existing state-of-the-art approaches.", "date": "2026-03-02", "categories": []}
{"id": "2603.01632", "title": "DeLo: Dual Decomposed Low-Rank Experts Collaboration for Continual Missing Modality Learning", "abstract": "Adapting Large Multimodal Models (LMMs) to real-world scenarios poses the dual challenges of learning from sequential data streams while handling frequent modality incompleteness, a task known as Continual Missing Modality Learning (CMML). However, existing works on CMML have predominantly relied on prompt tuning, a technique that struggles with this task due to cross-task interference between its learnable prompts in their shared embedding space. A naive application of Low-Rank Adaptation (LoRA) with modality-shared module will also suffer modality interference from competing gradients. To this end, we propose DeLo, the first framework to leverage a novel dual-decomposed low-rank expert architecture for CMML. Specifically, this architecture resolves modality interference through decomposed LoRA expert, dynamically composing LoRA update matrix with rank-one factors from disentangled modality-specific factor pools. Embedded within a task-partitioned framework that structurally prevents catastrophic forgetting, this expert system is supported by two key mechanisms: a Cross-Modal Guided Routing strategy to handle incomplete data and a Task-Key Memory for efficient, task-agnostic inference. Extensive experiments on established CMML benchmarks demonstrate that our method significantly outperforms state-of-the-art approaches. This highlights the value of a principled, architecturally-aware LoRA design for real-world multimodal challenges.", "date": "2026-03-02", "categories": []}
{"id": "2603.01677", "title": "A Practical Guide to Streaming Continual Learning", "abstract": "Continual Learning (CL) and Streaming Machine Learning (SML) study the ability of agents to learn from a stream of non-stationary data. Despite sharing some similarities, they address different and complementary challenges. While SML focuses on rapid adaptation after changes (concept drifts), CL aims to retain past knowledge when learning new tasks. After a brief introduction to CL and SML, we discuss Streaming Continual Learning (SCL), an emerging paradigm providing a unifying solution to real-world problems, which may require both SML and CL abilities. We claim that SCL can i) connect the CL and SML communities, motivating their work towards the same goal, and ii) foster the design of hybrid approaches that can quickly adapt to new information (as in SML) without forgetting previous knowledge (as in CL). We conclude the paper with a motivating example and a set of experiments, highlighting the need for SCL by showing how CL and SML alone struggle in achieving rapid adaptation and knowledge retention.", "date": "2026-03-02", "categories": []}
{"id": "2603.01695", "title": "Streaming Continual Learning for Unified Adaptive Intelligence in Dynamic Environments", "abstract": "Developing effective predictive models becomes challenging in dynamic environments that continuously produce data and constantly change. Continual Learning (CL) and Streaming Machine Learning (SML) are two research areas that tackle this arduous task. We put forward a unified setting that harnesses the benefits of both CL and SML: their ability to quickly adapt to non-stationary data streams without forgetting previous knowledge. We refer to this setting as Streaming Continual Learning (SCL). SCL does not replace either CL or SML. Instead, it extends the techniques and approaches considered by both fields. We start by briefly describing CL and SML and unifying the languages of the two frameworks. We then present the key features of SCL. We finally highlight the importance of bridging the two communities to advance the field of intelligent systems.", "date": "2026-03-02", "categories": []}
{"id": "2603.01935", "title": "Dream2Learn: Structured Generative Dreaming for Continual Learning", "abstract": "Continual learning requires balancing plasticity and stability while mitigating catastrophic forgetting. Inspired by human dreaming as a mechanism for internal simulation and knowledge restructuring, we introduce Dream2Learn (D2L), a framework in which a model autonomously generates structured synthetic experiences from its own internal representations and uses them for self-improvement. Rather than reconstructing past data as in generative replay, D2L enables a classifier to create novel, semantically distinct dreamed classes that are coherent with its learned knowledge yet do not correspond to previously observed data. These dreamed samples are produced by conditioning a frozen diffusion model through soft prompt optimization driven by the classifier itself. The generated data are not used to replace memory, but to expand and reorganize the representation space, effectively allowing the network to self-train on internally synthesized concepts. By integrating dreamed classes into continual training, D2L proactively structures latent features to support forward knowledge transfer and adaptation to future tasks. This prospective self-training mechanism mirrors the role of sleep in consolidating and reorganizing memory, turning internal simulations into a tool for improved generalization. Experiments on Mini-ImageNet, FG-ImageNet, and ImageNet-R demonstrate that D2L consistently outperforms strong rehearsal-based baselines and achieves positive forward transfer, confirming its ability to enhance adaptability through internally generated training signals.", "date": "2026-03-02", "categories": []}
{"id": "2602.23533", "title": "Few-Shot Continual Learning for 3D Brain MRI with Frozen Foundation Models", "abstract": "Foundation models pretrained on large-scale 3D medical imaging data face challenges when adapted to multiple downstream tasks under continual learning with limited labeled data. We address few-shot continual learning for 3D brain MRI by combining a frozen pretrained backbone with task-specific Low-Rank Adaptation (LoRA) modules. Tasks arrive sequentially -- tumor segmentation (BraTS) and brain age estimation (IXI) -- with no replay of previous task data. Each task receives a dedicated LoRA adapter; only the adapter and task-specific head are trained while the backbone remains frozen, thereby eliminating catastrophic forgetting by design (BWT=0). In continual learning, sequential full fine-tuning suffers severe forgetting (T1 Dice drops from 0.80 to 0.16 after T2), while sequential linear probing achieves strong T1 (Dice 0.79) but fails on T2 (MAE 1.45). Our LoRA approach achieves the best balanced performance across both tasks: T1 Dice 0.62$\\pm$0.07, T2 MAE 0.16$\\pm$0.05, with zero forgetting and $<$0.1\\% trainable parameters per task, though with noted systematic age underestimation in T2 (Wilcoxon $p<0.001$). Frozen foundation models with task-specific LoRA adapters thus offer a practical solution when both tasks must be maintained under few-shot continual learning.", "date": "2026-02-26", "categories": []}
{"id": "2602.23817", "title": "Footprint-Guided Exemplar-Free Continual Histopathology Report Generation", "abstract": "Rapid progress in vision-language modeling has enabled pathology report generation from gigapixel whole-slide images, but most approaches assume static training with simultaneous access to all data. In clinical deployment, however, new organs, institutions, and reporting conventions emerge over time, and sequential fine-tuning can cause catastrophic forgetting. We introduce an exemplar-free continual learning framework for WSI-to-report generation that avoids storing raw slides or patch exemplars. The core idea is a compact domain footprint built in a frozen patch-embedding space: a small codebook of representative morphology tokens together with slide-level co-occurrence summaries and lightweight patch-count priors. These footprints support generative replay by synthesizing pseudo-WSI representations that reflect domain-specific morphological mixtures, while a teacher snapshot provides pseudo-reports to supervise the updated model without retaining past data. To address shifting reporting conventions, we distill domain-specific linguistic characteristics into a compact style descriptor and use it to steer generation. At inference, the model identifies the most compatible descriptor directly from the slide signal, enabling domain-agnostic setup without requiring explicit domain identifiers. Evaluated across multiple public continual learning benchmarks, our approach outperforms exemplar-free and limited-buffer rehearsal baselines, highlighting footprint-based generative replay as a practical solution for deployment in evolving clinical settings.", "date": "2026-02-27", "categories": []}
{"id": "2407.17120", "title": "Parameter-Efficient Fine-Tuning for Continual Learning: A Neural Tangent Kernel Perspective", "abstract": "Parameter-efficient fine-tuning for continual learning (PEFT-CL) has shown promise in adapting pre-trained models to sequential tasks while mitigating catastrophic forgetting problem. However, understanding the mechanisms that dictate continual performance in this paradigm remains elusive. To unravel this mystery, we undertake a rigorous analysis of PEFT-CL dynamics to derive relevant metrics for continual scenarios using Neural Tangent Kernel (NTK) theory. With the aid of NTK as a mathematical analysis tool, we recast the challenge of test-time forgetting into the quantifiable generalization gaps during training, identifying three key factors that influence these gaps and the performance of PEFT-CL: training sample size, task-level feature orthogonality, and regularization. To address these challenges, we introduce NTK-CL, a novel framework that eliminates task-specific parameter storage while adaptively generating task-relevant features. Aligning with theoretical guidance, NTK-CL triples the feature representation of each sample, theoretically and empirically reducing the magnitude of both task-interplay and task-specific generalization gaps. Grounded in NTK analysis, our framework imposes an adaptive exponential moving average mechanism and constraints on task-level feature orthogonality, maintaining intra-task NTK forms while attenuating inter-task NTK forms. Ultimately, by fine-tuning optimizable parameters with appropriate regularization, NTK-CL achieves state-of-the-art performance on established PEFT-CL benchmarks. This work provides a theoretical foundation for understanding and improving PEFT-CL models, offering insights into the interplay between feature representation, task orthogonality, and generalization, contributing to the development of more efficient continual learning systems.", "date": "2024-07-24", "categories": []}
{"id": "2503.10503", "title": "Sample Compression for Self Certified Continual Learning", "abstract": "Continual learning algorithms aim to learn from a sequence of tasks. In order to avoid catastrophic forgetting, most existing approaches rely on heuristics and do not provide computable learning guarantees. In this paper, we introduce Continual Pick-to-Learn (CoP2L), a method grounded in sample compression theory that retains representative samples for each task in a principled and efficient way. This allows us to derive non-vacuous, numerically computable upper bounds on the generalization loss of the learned predictors after each task. We evaluate CoP2L on standard continual learning benchmarks under Class-Incremental and Task-Incremental settings, showing that it effectively mitigates catastrophic forgetting. It turns out that CoP2L is empirically competitive with baseline methods while certifying predictor reliability in continual learning with a non-vacuous bound.", "date": "2025-03-13", "categories": []}
{"id": "2602.22884", "title": "Unsupervised Continual Learning for Amortized Bayesian Inference", "abstract": "Amortized Bayesian Inference (ABI) enables efficient posterior estimation using generative neural networks trained on simulated data, but often suffers from performance degradation under model misspecification. While self-consistency (SC) training on unlabeled empirical data can enhance network robustness, current approaches are limited to static, single-task settings and fail to handle sequentially arriving data or distribution shifts. We propose a continual learning framework for ABI that decouples simulation-based pre-training from unsupervised sequential SC fine-tuning on real-world data. To address the challenge of catastrophic forgetting, we introduce two adaptation strategies: (1) SC with episodic replay, utilizing a memory buffer of past observations, and (2) SC with elastic weight consolidation, which regularizes updates to preserve task-critical parameters. Across three diverse case studies, our methods significantly mitigate forgetting and yield posterior estimates that outperform standard simulation-based training, achieving estimates closer to MCMC reference, providing a viable path for trustworthy ABI across a range of different tasks.", "date": "2026-02-26", "categories": []}
{"id": "2602.21919", "title": "Learning in the Null Space: Small Singular Values for Continual Learning", "abstract": "Alleviating catastrophic forgetting while enabling further learning is a primary challenge in continual learning (CL). Orthogonal-based training methods have gained attention for their efficiency and strong theoretical properties, and many existing approaches enforce orthogonality through gradient projection. In this paper, we revisit orthogonality and exploit the fact that small singular values correspond to directions that are nearly orthogonal to the input space of previous tasks. Building on this principle, we introduce NESS (Null-space Estimated from Small Singular values), a CL method that applies orthogonality directly in the weight space rather than through gradient manipulation. Specifically, NESS constructs an approximate null space using the smallest singular values of each layer's input representation and parameterizes task-specific updates via a compact low-rank adaptation (LoRA-style) formulation constrained to this subspace. The subspace basis is fixed to preserve the null-space constraint, and only a single trainable matrix is learned for each task. This design ensures that the resulting updates remain approximately in the null space of previous inputs while enabling adaptation to new tasks. Our theoretical analysis and experiments on three benchmark datasets demonstrate competitive performance, low forgetting, and stable accuracy across tasks, highlighting the role of small singular values in continual learning. The code is available at https://github.com/pacman-ctm/NESS.", "date": "2026-02-25", "categories": []}
{"id": "2504.20932", "title": "Improvements of Dark Experience Replay and Reservoir Sampling towards Better Balance between Consolidation and Plasticity", "abstract": "Continual learning is the one of the most essential abilities for autonomous agents, which can incrementally learn daily-life skills. For this ultimate goal, a simple but powerful method, dark experience replay (DER), has been proposed recently. DER mitigates catastrophic forgetting, in which the skills acquired in the past are unintentionally forgotten, by stochastically storing the streaming data in a reservoir sampling (RS) buffer and by relearning them or retaining the past outputs for them. However, since DER considers multiple objectives, it will not function properly without appropriate weighting of them. In addition, the ability to retain past outputs inhibits learning if the past outputs are incorrect due to distribution shift or other effects. This is due to a tradeoff between memory consolidation and plasticity. The tradeoff is hidden even in the RS buffer, which gradually stops storing new data for new skills in it as data is continuously passed to it. To alleviate the tradeoff and achieve better balance, this paper proposes improvement strategies to each of DER and RS. Specifically, DER is improved with automatic adaptation of weights, block of replaying erroneous data, and correction of past outputs. RS is also improved with generalization of acceptance probability, stratification of plural buffers, and intentional omission of unnecessary data. These improvements are verified through multiple benchmarks including regression, classification, and reinforcement learning problems. As a result, the proposed methods achieve steady improvements in learning performance by balancing the memory consolidation and plasticity.", "date": "2025-04-29", "categories": []}
{"id": "2602.20791", "title": "Understanding the Role of Rehearsal Scale in Continual Learning under Varying Model Capacities", "abstract": "Rehearsal is one of the key techniques for mitigating catastrophic forgetting and has been widely adopted in continual learning algorithms due to its simplicity and practicality. However, the theoretical understanding of how rehearsal scale influences learning dynamics remains limited. To address this gap, we formulate rehearsal-based continual learning as a multidimensional effectiveness-driven iterative optimization problem, providing a unified characterization across diverse performance metrics. Within this framework, we derive a closed-form analysis of adaptability, memorability, and generalization from the perspective of rehearsal scale. Our results uncover several intriguing and counterintuitive findings. First, rehearsal can impair model's adaptability, in sharp contrast to its traditionally recognized benefits. Second, increasing the rehearsal scale does not necessarily improve memory retention. When tasks are similar and noise levels are low, the memory error exhibits a diminishing lower bound. Finally, we validate these insights through numerical simulations and extended analyses on deep neural networks across multiple real-world datasets, revealing statistical patterns of rehearsal mechanisms in continual learning.", "date": "2026-02-24", "categories": []}
{"id": "2602.20796", "title": "Exploring the Impact of Parameter Update Magnitude on Forgetting and Generalization of Continual Learning", "abstract": "The magnitude of parameter updates are considered a key factor in continual learning. However, most existing studies focus on designing diverse update strategies, while a theoretical understanding of the underlying mechanisms remains limited. Therefore, we characterize model's forgetting from the perspective of parameter update magnitude and formalize it as knowledge degradation induced by task-specific drift in the parameter space, which has not been fully captured in previous studies due to their assumption of a unified parameter space. By deriving the optimal parameter update magnitude that minimizes forgetting, we unify two representative update paradigms, frozen training and initialized training, within an optimization framework for constrained parameter updates. Our theoretical results further reveals that sequence tasks with small parameter distances exhibit better generalization and less forgetting under frozen training rather than initialized training. These theoretical insights inspire a novel hybrid parameter update strategy that adaptively adjusts update magnitude based on gradient directions. Experiments on deep neural networks demonstrate that this hybrid approach outperforms standard training strategies, providing new theoretical perspectives and practical inspiration for designing efficient and scalable continual learning algorithms.", "date": "2026-02-24", "categories": []}
{"id": "2602.20911", "title": "From Isolation to Integration: Building an Adaptive Expert Forest for Pre-Trained Model-based Class-Incremental Learning", "abstract": "Class-Incremental Learning (CIL) requires models to learn new classes without forgetting old ones. A common method is to freeze a pre-trained model and train a new, lightweight adapter for each task. While this prevents forgetting, it treats the learned knowledge as a simple, unstructured collection and fails to use the relationships between tasks. To this end, we propose the Semantic-guided Adaptive Expert Forest (SAEF), a new method that organizes adapters into a structured hierarchy for better knowledge sharing. SAEF first groups tasks into conceptual clusters based on their semantic relationships. Then, within each cluster, it builds a balanced expert tree by creating new adapters from merging the adapters of similar tasks. At inference time, SAEF finds and activates a set of relevant experts from the forest for any given input. The final prediction is made by combining the outputs of these activated experts, weighted by how confident each expert is. Experiments on several benchmark datasets show that SAEF achieves SOTA performance.", "date": "2026-02-24", "categories": []}
{"id": "2503.21258", "title": "Learn by Reasoning: Analogical Weight Generation for Few-Shot Class-Incremental Learning", "abstract": "Few-shot class-incremental Learning (FSCIL) enables models to learn new classes from limited data while retaining performance on previously learned classes. Traditional FSCIL methods often require fine-tuning parameters with limited new class data and suffer from a separation between learning new classes and utilizing old knowledge. Inspired by the analogical learning mechanisms of the human brain, we propose a novel analogical generative method. Our approach includes the Brain-Inspired Analogical Generator (BiAG), which derives new class weights from existing classes without parameter fine-tuning during incremental stages. BiAG consists of three components: Weight Self-Attention Module (WSA), Weight & Prototype Analogical Attention Module (WPAA), and Semantic Conversion Module (SCM). SCM uses Neural Collapse theory for semantic conversion, WSA supplements new class weights, and WPAA computes analogies to generate new class weights. Experiments on miniImageNet, CUB-200, and CIFAR-100 datasets demonstrate that our method achieves higher final and average accuracy compared to SOTA methods.", "date": "2025-03-27", "categories": []}
{"id": "2510.21491", "title": "Benchmarking Catastrophic Forgetting Mitigation Methods in Federated Time Series Forecasting", "abstract": "Catastrophic forgetting (CF) poses a persistent challenge in continual learning (CL), especially within federated learning (FL) environments characterized by non-i.i.d. time series data. While existing research has largely focused on classification tasks in vision domains, the regression-based forecasting setting prevalent in IoT and edge applications remains underexplored. In this paper, we present the first benchmarking framework tailored to investigate CF in federated continual time series forecasting. Using the Beijing Multi-site Air Quality dataset across 12 decentralized clients, we systematically evaluate several CF mitigation strategies, including Replay, Elastic Weight Consolidation, Learning without Forgetting, and Synaptic Intelligence. Key contributions include: (i) introducing a new benchmark for CF in time series FL, (ii) conducting a comprehensive comparative analysis of state-of-the-art methods, and (iii) releasing a reproducible open-source framework. This work provides essential tools and insights for advancing continual learning in federated time-series forecasting systems.", "date": "2025-10-24", "categories": []}
{"id": "2511.17439", "title": "InTAct: Interval-based Task Activation Consolidation for Continual Learning", "abstract": "Continual learning is a fundamental challenge in artificial intelligence that requires networks to acquire new knowledge while preserving previously learned representations. Despite the success of various approaches, most existing paradigms do not provide rigorous mathematical guarantees against catastrophic forgetting. Current methods that offer such guarantees primarily focus on analyzing the parameter space using \\textit{interval arithmetic (IA)}, as seen in frameworks such as InterContiNet. However, restricting high-dimensional weight updates can be computationally expensive. In this work, we propose InTAct (Interval-based Task Activation Consolidation), a method that mitigates catastrophic forgetting by enforcing functional invariance at the neuron level. We identify specific activation intervals where previous tasks reside and constrain updates within these regions while allowing for flexible adaptation elsewhere. By ensuring that predictions remain stable within these nested activation intervals, we provide a tractable mathematical guarantee of functional invariance. We emphasize that regulating the activation space is significantly more efficient than parameter-based constraints, because the dimensionality of internal signals is much lower than that of the vast space of model weights. While our approach is architecture-agnostic and applicable to various continual learning settings, its integration with prompt-based methods enables it to achieve state-of-the-art performance on challenging benchmarks.", "date": "2025-11-21", "categories": []}
{"id": "2602.18528", "title": "Audio-Visual Continual Test-Time Adaptation without Forgetting", "abstract": "Audio-visual continual test-time adaptation involves continually adapting a source audio-visual model at test-time, to unlabeled non-stationary domains, where either or both modalities can be distributionally shifted, which hampers online cross-modal learning and eventually leads to poor accuracy. While previous works have tackled this problem, we find that SOTA methods suffer from catastrophic forgetting, where the model's performance drops well below the source model due to continual parameter updates at test-time. In this work, we first show that adapting only the modality fusion layer to a target domain not only improves performance on that domain but can also enhance performance on subsequent domains. Based on this strong cross-task transferability of the fusion layer's parameters, we propose a method, $\\texttt{AV-CTTA}$, that improves test-time performance of the models without access to any source data. Our approach works by using a selective parameter retrieval mechanism that dynamically retrieves the best fusion layer parameters from a buffer using only a small batch of test data. These parameters are then integrated into the model, adapted to the current test distribution, and saved back for future use. Extensive experiments on benchmark datasets involving unimodal and bimodal corruptions show our proposed $\\texttt{AV-CTTA}$ significantly outperforms existing methods while minimizing catastrophic forgetting.", "date": "2026-02-20", "categories": []}
{"id": "2602.18628", "title": "Non-Interfering Weight Fields: Treating Model Parameters as a Continuously Extensible Function", "abstract": "Large language models store all learned knowledge in a single, fixed weight vector. Teaching a model new capabilities requires modifying those same weights, inevitably degrading previously acquired knowledge. This fundamental limitation, known as catastrophic forgetting, has resisted principled solutions for decades. Existing approaches treat weights as immutable artifacts that must be protected through techniques like regularization heuristics, replay buffers, or isolated adapter modules. The problem is none of these provide a structural guarantee against forgetting. In this work, we propose Non-Interfering Weight Fields (NIWF), a framework that replaces the fixed weight paradigm with a learned function that generates weight configurations on demand from a continuous capability coordinate space. After training on a task, we commit the occupied coordinate region by snapshotting the fields outputs on anchor points to enforce a functional lock during all future training. We validate NIWF on sequential instructionfollowing and code generation tasks using Mistral-7B, demonstrating zero forgetting on committed tasks with competitive perplexity on new tasks. The framework introduces the notion of software-like versioning for neural network intelligence, where capabilities can be committed, extended, composed, and rolled back without retraining.", "date": "2026-02-20", "categories": []}
{"id": "2602.19355", "title": "Active perception and disentangled representations allow continual, episodic zero and few-shot learning", "abstract": "Generalization is often regarded as an essential property of machine learning systems. However, perhaps not every component of a system needs to generalize. Training models for generalization typically produces entangled representations at the boundaries of entities or classes, which can lead to destructive interference when rapid, high-magnitude updates are required for continual or few-shot learning. Techniques for fast learning with non-interfering representations exist, but they generally fail to generalize. Here, we describe a Complementary Learning System (CLS) in which the fast learner entirely foregoes generalization in exchange for continual zero-shot and few-shot learning. Unlike most CLS approaches, which use episodic memory primarily for replay and consolidation, our fast, disentangled learner operates as a parallel reasoning system. The fast learner can overcome observation variability and uncertainty by leveraging a conventional slow, statistical learner within an active perception system: A contextual bias provided by the fast learner induces the slow learner to encode novel stimuli in familiar, generalized terms, enabling zero-shot and few-shot learning. This architecture demonstrates that fast, context-driven reasoning can coexist with slow, structured generalization, providing a pathway for robust continual learning.", "date": "2026-02-22", "categories": []}
{"id": "2602.19655", "title": "Representation Stability in a Minimal Continual Learning Agent", "abstract": "Continual learning systems are increasingly deployed in environments where retraining or reset is infeasible, yet many approaches emphasize task performance rather than the evolution of internal representations over time. In this work, we study a minimal continual learning agent designed to isolate representational dynamics from architectural complexity and optimization objectives. The agent maintains a persistent state vector across executions and incrementally updates it as new textual data is introduced. We quantify representational change using cosine similarity between successive normalized state vectors and define a stability metric over time intervals. Longitudinal experiments across eight executions reveal a transition from an initial plastic regime to a stable representational regime under consistent input. A deliberately introduced semantic perturbation produces a bounded decrease in similarity, followed by recovery and restabilization under subsequent coherent input. These results demonstrate that meaningful stability plasticity tradeoffs can emerge in a minimal, stateful learning system without explicit regularization, replay, or architectural complexity. The work establishes a transparent empirical baseline for studying representational accumulation and adaptation in continual learning systems.", "date": "2026-02-23", "categories": []}
{"id": "2602.19872", "title": "GOAL: Geometrically Optimal Alignment for Continual Generalized Category Discovery", "abstract": "Continual Generalized Category Discovery (C-GCD) requires identifying novel classes from unlabeled data while retaining knowledge of known classes over time. Existing methods typically update classifier weights dynamically, resulting in forgetting and inconsistent feature alignment. We propose GOAL, a unified framework that introduces a fixed Equiangular Tight Frame (ETF) classifier to impose a consistent geometric structure throughout learning. GOAL conducts supervised alignment for labeled samples and confidence-guided alignment for novel samples, enabling stable integration of new classes without disrupting old ones. Experiments on four benchmarks show that GOAL outperforms the prior method Happy, reducing forgetting by 16.1% and boosting novel class discovery by 3.2%, establishing a strong solution for long-horizon continual discovery.", "date": "2026-02-23", "categories": []}
{"id": "2509.23592", "title": "Toward a Holistic Approach to Continual Model Merging", "abstract": "We present a holistic framework for Continual Model Merging (CMM) that intervenes at three critical stages: pre-merging, during merging, and post-merging-to address two fundamental challenges in continual learning. In particular, conventional approaches either maintain a growing list of per-domain task vectors, leading to scalability issues or rely solely on weight-space merging when old data is inaccessible, thereby losing crucial functional information. Our method overcomes these limitations by first fine-tuning the main model within its tangent space on domain-specific data; this linearization amplifies per-task weight disentanglement, effectively mitigating across-task interference. During merging, we leverage functional information from available optimizer states beyond mere parameter averages to avoid the need to revisit old data. Finally, a post-merging correction aligns the representation discrepancy between pre- and post-merged models, reducing bias and enhancing overall performance-all while operating under constant memory constraints without accessing historical data. Extensive experiments on standard class-incremental and domain-incremental benchmarks demonstrate that our approach not only achieves competitive performance but also provides a scalable and efficient solution to the catastrophic forgetting problem.", "date": "2025-09-28", "categories": []}
{"id": "2602.18055", "title": "Continual-NExT: A Unified Comprehension And Generation Continual Learning Framework", "abstract": "Dual-to-Dual MLLMs refer to Multimodal Large Language Models, which can enable unified multimodal comprehension and generation through text and image modalities. Although exhibiting strong instantaneous learning and generalization capabilities, Dual-to-Dual MLLMs still remain deficient in lifelong evolution, significantly affecting continual adaptation to dynamic real-world scenarios. One of the challenges is that learning new tasks inevitably destroys the learned knowledge. Beyond traditional catastrophic forgetting, Dual-to-Dual MLLMs face other challenges, including hallucination, instruction unfollowing, and failures in cross-modal knowledge transfer. However, no standardized continual learning framework for Dual-to-Dual MLLMs has been established yet, leaving these challenges unexplored. Thus, in this paper, we establish Continual-NExT, a continual learning framework for Dual-to-Dual MLLMs with deliberately-architected evaluation metrics. To improve the continual learning capability of Dual-to-Dual MLLMs, we propose an efficient MAGE (Mixture and Aggregation of General LoRA and Expert LoRA) method to further facilitate knowledge transfer across modalities and mitigate forgetting. Extensive experiments demonstrate that MAGE outperforms other continual learning methods and achieves state-of-the-art performance.", "date": "2026-02-20", "categories": []}
{"id": "2502.14762", "title": "Unlocking [CLS] Features for Continual Post-Training", "abstract": "Continual learning requires models to integrate new classes or domains over time while preserving previously acquired knowledge. Within this paradigm, foundation models often achieve strong performance, but they still remain subject to the stability-plasticity trade-off, where excessive plasticity leads to forgetting of prior knowledge, and excessive stability constrains the adaptation. This necessitates an effective post-training strategy that introduces minimal yet functional modifications. To address this challenge, we first introduce a new parameter-efficient fine-tuning module 'Learn and Calibrate', or LuCA, designed to acquire task-specific knowledge through an adapter-calibrator couple, enabling well-refined feature representations. Then, for each task, we deploy a sparse LuCA module on top of the last classification token [CLS] just before the classifier, which we refer to as 'Token-level Sparse Calibration and Adaptation', or TOSCA. By leaving the generalization capabilities of the foundation models intact and adapting exclusively via the last token, our approach achieves a harmonious balance between stability and plasticity while reducing both training and inference complexity. We demonstrate that TOSCA yields state-of-the-art performance while introducing ~8 times fewer parameters compared to prior methods.", "date": "2025-02-20", "categories": []}
{"id": "2602.17559", "title": "Revisiting Weight Regularization for Low-Rank Continual Learning", "abstract": "Continual Learning (CL) with large-scale pre-trained models (PTMs) has recently gained wide attention, shifting the focus from training from scratch to continually adapting PTMs. This has given rise to a promising paradigm: parameter-efficient continual learning (PECL), where task interference is typically mitigated by assigning a task-specific module during training, such as low-rank adapters. However, weight regularization techniques, such as Elastic Weight Consolidation (EWC)-a key strategy in CL-remain underexplored in this new paradigm. In this paper, we revisit weight regularization in low-rank CL as a new perspective for mitigating task interference in PECL. Unlike existing low-rank CL methods, we mitigate task interference by regularizing a shared low-rank update through EWC, thereby keeping the storage requirement and inference costs constant regardless of the number of tasks. Our proposed method EWC-LoRA leverages a low-rank representation to estimate parameter importance over the full-dimensional space. This design offers a practical, computational- and memory-efficient solution for CL with PTMs, and provides insights that may inform the broader application of regularization techniques within PECL. Extensive experiments on various benchmarks demonstrate the effectiveness of EWC-LoRA, achieving a stability-plasticity trade-off superior to existing low-rank CL approaches. These results indicate that, even under low-rank parameterizations, weight regularization remains an effective mechanism for mitigating task interference. Code is available at: https://github.com/yaoyz96/low-rank-cl.", "date": "2026-02-19", "categories": []}
{"id": "2602.17625", "title": "Catastrophic Forgetting Resilient One-Shot Incremental Federated Learning", "abstract": "Modern big-data systems generate massive, heterogeneous, and geographically dispersed streams that are large-scale and privacy-sensitive, making centralization challenging. While federated learning (FL) provides a privacy-enhancing training mechanism, it assumes a static data flow and learns a collaborative model over multiple rounds, making learning with \\textit{incremental} data challenging in limited-communication scenarios. This paper presents One-Shot Incremental Federated Learning (OSI-FL), the first FL framework that addresses the dual challenges of communication overhead and catastrophic forgetting. OSI-FL communicates category-specific embeddings, devised by a frozen vision-language model (VLM) from each client in a single communication round, which a pre-trained diffusion model at the server uses to synthesize new data similar to the client's data distribution. The synthesized samples are used on the server for training. However, two challenges still persist: i) tasks arriving incrementally need to retrain the global model, and ii) as future tasks arrive, retraining the model introduces catastrophic forgetting. To this end, we augment training with Selective Sample Retention (SSR), which identifies and retains the top-p most informative samples per category and task pair based on sample loss. SSR bounds forgetting by ensuring that representative retained samples are incorporated into training in further iterations. The experimental results indicate that OSI-FL outperforms baselines, including traditional and one-shot FL approaches, in both class-incremental and domain-incremental scenarios across three benchmark datasets.", "date": "2026-02-19", "categories": []}
{"id": "2502.07274", "title": "Forget Forgetting: Continual Learning in a World of Abundant Memory", "abstract": "Continual learning (CL) has traditionally focused on minimizing exemplar memory, a constraint often misaligned with modern systems where GPU time, not storage, is the primary bottleneck. This paper challenges this paradigm by investigating a more realistic regime: one where memory is abundant enough to mitigate forgetting, but full retraining from scratch remains prohibitively expensive. In this practical \"middle ground\", we find that the core challenge shifts from stability to plasticity, as models become biased toward prior tasks and struggle to learn new ones. Conversely, improved stability allows simple replay baselines to outperform the state-of-the-art methods at a fraction of the GPU cost. To address this newly surfaced trade-off, we propose Weight Space Consolidation, a lightweight method that combines (1) rank-based parameter resets to restore plasticity with (2) weight averaging to enhance stability. Validated on both class-incremental learning with image classifiers and continual instruction tuning with large language models, our approach outperforms strong baselines while matching the low computational cost of replay, offering a scalable alternative to expensive full-retraining. These findings challenge long-standing CL assumptions and establish a new, cost-efficient baseline for real-world CL systems where exemplar memory is no longer the limiting factor.", "date": "2025-02-11", "categories": []}
{"id": "2602.16101", "title": "Axle Sensor Fusion for Online Continual Wheel Fault Detection in Wayside Railway Monitoring", "abstract": "Reliable and cost-effective maintenance is essential for railway safety, particularly at the wheel-rail interface, which is prone to wear and failure. Predictive maintenance frameworks increasingly leverage sensor-generated time-series data, yet traditional methods require manual feature engineering, and deep learning models often degrade in online settings with evolving operational patterns. This work presents a semantic-aware, label-efficient continual learning framework for railway fault diagnostics. Accelerometer signals are encoded via a Variational AutoEncoder into latent representations capturing the normal operational structure in a fully unsupervised manner. Importantly, semantic metadata, including axle counts, wheel indexes, and strain-based deformations, is extracted via AI-driven peak detection on fiber Bragg grating sensors (resistant to electromagnetic interference) and fused with the VAE embeddings, enhancing anomaly detection under unknown operational conditions. A lightweight gradient boosting supervised classifier stabilizes anomaly scoring with minimal labels, while a replay-based continual learning strategy enables adaptation to evolving domains without catastrophic forgetting. Experiments show the model detects minor imperfections due to flats and polygonization, while adapting to evolving operational conditions, such as changes in train type, speed, load, and track profiles, captured using a single accelerometer and strain gauge in wayside monitoring.", "date": "2026-02-18", "categories": []}
{"id": "2602.15156", "title": "Panini: Continual Learning in Token Space via Structured Memory", "abstract": "Language models are increasingly used to reason over content they were not trained on, such as new documents, evolving knowledge, and user-specific data. A common approach is retrieval-augmented generation (RAG), which stores verbatim documents externally (as chunks) and retrieves only a relevant subset at inference time for an LLM to reason over. However, this results in inefficient usage of test-time compute (LLM repeatedly reasons over the same documents); moreover, chunk retrieval can inject irrelevant context that increases unsupported generation. We propose a human-like non-parametric continual learning framework, where the base model remains fixed, and learning occurs by integrating each new experience into an external semantic memory state that accumulates and consolidates itself continually. We present Panini, which realizes this by representing documents as Generative Semantic Workspaces (GSW) -- an entity- and event-aware network of question-answer (QA) pairs, sufficient for an LLM to reconstruct the experienced situations and mine latent knowledge via reasoning-grounded inference chains on the network. Given a query, Panini only traverses the continually-updated GSW (not the verbatim documents or chunks), and retrieves the most likely inference chains. Across six QA benchmarks, Panini achieves the highest average performance, 5%-7% higher than other competitive baselines, while using 2-30x fewer answer-context tokens, supports fully open-source pipelines, and reduces unsupported answers on curated unanswerable queries. The results show that efficient and accurate structuring of experiences at write time -- as achieved by the GSW framework -- yields both efficiency and reliability gains at read time. Code is available at https://github.com/roychowdhuryresearch/gsw-memory.", "date": "2026-02-16", "categories": []}
{"id": "2601.21345", "title": "Semantic-Guided Dynamic Sparsification for Pre-Trained Model-based Class-Incremental Learning", "abstract": "Class-Incremental Learning (CIL) requires a model to continually learn new classes without forgetting old ones. A common and efficient solution freezes a pre-trained model and employs lightweight adapters, whose parameters are often forced to be orthogonal to prevent inter-task interference. However, we argue that this parameter-constraining method is detrimental to plasticity. To this end, we propose Semantic-Guided Dynamic Sparsification (SGDS), a novel method that proactively guides the activation space by governing the orientation and rank of its subspaces through targeted sparsification. Specifically, SGDS promotes knowledge transfer by encouraging similar classes to share a compact activation subspace, while simultaneously preventing interference by assigning non-overlapping activation subspaces to dissimilar classes. By sculpting class-specific sparse subspaces in the activation space, SGDS effectively mitigates interference without imposing rigid constraints on the parameter space. Extensive experiments on various benchmark datasets demonstrate the state-of-the-art performance of SGDS.", "date": "2026-01-29", "categories": []}
{"id": "2602.13440", "title": "Learning on the Fly: Replay-Based Continual Object Perception for Indoor Drones", "abstract": "Autonomous agents such as indoor drones must learn new object classes in real-time while limiting catastrophic forgetting, motivating Class-Incremental Learning (CIL). However, most unmanned aerial vehicle (UAV) datasets focus on outdoor scenes and offer limited temporally coherent indoor videos. We introduce an indoor dataset of $14,400$ frames capturing inter-drone and ground vehicle footage, annotated via a semi-automatic workflow with a $98.6\\%$ first-pass labeling agreement before final manual verification. Using this dataset, we benchmark 3 replay-based CIL strategies: Experience Replay (ER), Maximally Interfered Retrieval (MIR), and Forgetting-Aware Replay (FAR), using YOLOv11-nano as a resource-efficient detector for deployment-constrained UAV platforms. Under tight memory budgets ($5-10\\%$ replay), FAR performs better than the rest, achieving an average accuracy (ACC, $mAP_{50-95}$ across increments) of $82.96\\%$ with $5\\%$ replay. Gradient-weighted class activation mapping (Grad-CAM) analysis shows attention shifts across classes in mixed scenes, which is associated with reduced localization quality for drones. The experiments further demonstrate that replay-based continual learning can be effectively applied to edge aerial systems. Overall, this work contributes an indoor UAV video dataset with preserved temporal coherence and an evaluation of replay-based CIL under limited replay budgets. Project page: https://spacetime-vision-robotics-laboratory.github.io/learning-on-the-fly-cl", "date": "2026-02-13", "categories": []}
{"id": "2602.14078", "title": "Policy Gradient with Adaptive Entropy Annealing for Continual Fine-Tuning", "abstract": "Despite their success, large pretrained vision models remain vulnerable to catastrophic forgetting when adapted to new tasks in class-incremental settings. Parameter-efficient fine-tuning (PEFT) alleviates this by restricting trainable parameters, yet most approaches still rely on cross-entropy (CE) loss, a surrogate for the 0-1 loss, to learn from new data. We revisit this choice and revive the true objective (0-1 loss) through a reinforcement learning perspective. By formulating classification as a one-step Markov Decision Process, we derive an Expected Policy Gradient (EPG) method that directly minimizes misclassification error with a low-variance gradient estimation. Our analysis shows that CE can be interpreted as EPG with an additional sample-weighting mechanism: CE encourages exploration by emphasizing low-confidence samples, while EPG prioritizes high-confidence ones. Building on this insight, we propose adaptive entropy annealing (aEPG), a training strategy that transitions from exploratory (CE-like) to exploitative (EPG-like) learning. aEPG-based methods outperform CE-based methods across diverse benchmarks and with various PEFT modules. More broadly, we evaluate various entropy regularization methods and demonstrate that lower entropy of the output prediction distribution enhances adaptation in pretrained vision models.", "date": "2026-02-15", "categories": []}
{"id": "2602.14767", "title": "SAILS: Segment Anything with Incrementally Learned Semantics for Task-Invariant and Training-Free Continual Learning", "abstract": "Continual learning remains constrained by the need for repeated retraining, high computational costs, and the persistent challenge of forgetting. These factors significantly limit the applicability of continuous learning in real-world settings, as iterative model updates require significant computational resources and inherently exacerbate forgetting. We present SAILS -- Segment Anything with Incrementally Learned Semantics, a training-free framework for Class-Incremental Semantic Segmentation (CISS) that sidesteps these challenges entirely. SAILS leverages foundational models to decouple CISS into two stages: Zero-shot region extraction using Segment Anything Model (SAM), followed by semantic association through prototypes in a fixed feature space. SAILS incorporates selective intra-class clustering, resulting in multiple prototypes per class to better model intra-class variability. Our results demonstrate that, despite requiring no incremental training, SAILS typically surpasses the performance of existing training-based approaches on standard CISS datasets, particularly in long and challenging task sequences where forgetting tends to be most severe. By avoiding parameter updates, SAILS completely eliminates forgetting and maintains consistent, task-invariant performance. Furthermore, SAILS exhibits positive backward transfer, where the introduction of new classes can enhance performance on previous classes.", "date": "2026-02-16", "categories": []}
{"id": "2507.02310", "title": "Holistic Continual Learning under Concept Drift with Adaptive Memory Realignment", "abstract": "Traditional continual learning methods prioritize knowledge retention and focus primarily on mitigating catastrophic forgetting, implicitly assuming that the data distribution of previously learned tasks remains static. This overlooks the dynamic nature of real-world data streams, where concept drift permanently alters previously seen data and demands both stability and rapid adaptation. We introduce a holistic framework for continual learning under concept drift that simulates realistic scenarios by evolving task distributions. As a baseline, we consider Full Relearning (FR), in which the model is retrained from scratch on newly labeled samples from the drifted distribution. While effective, this approach incurs substantial annotation and computational overhead. To address these limitations, we propose Adaptive Memory Realignment (AMR), a lightweight alternative that equips rehearsal-based learners with a drift-aware adaptation mechanism. AMR selectively removes outdated samples of drifted classes from the replay buffer and repopulates it with a small number of up-to-date instances, effectively realigning memory with the new distribution. This targeted resampling matches the performance of FR while reducing the need for labeled data and computation by orders of magnitude. To enable reproducible evaluation, we introduce four concept drift variants of standard vision benchmarks, where previously seen classes reappear with shifted representations. Comprehensive experiments on these datasets using several rehearsal-based baselines show that AMR consistently counters concept drift, maintaining high accuracy with minimal overhead. These results position AMR as a scalable solution that reconciles stability and plasticity in non-stationary continual learning environments. Full implementation of our framework and benchmark datasets is available at: github.com/AlifAshrafee/CL-Under-Concept-Drift.", "date": "2025-07-03", "categories": []}
{"id": "2508.08275", "title": "MLLM-CTBench: A Benchmark for Continual Instruction Tuning with Reasoning Process Diagnosis", "abstract": "Continual instruction tuning(CIT) during the post-training phase is crucial for adapting multimodal large language models (MLLMs) to evolving real-world demands. However, the progress is hampered by the lack of benchmarks with rigorous, protocol-consistent evaluation. To bridge this gap, we introduce MLLM-CTBench, a comprehensive benchmark for CIT of MLLMs, covering seven challenging tasks across six diverse domains. MLLM-CTBench makes three key contributions. First, we establish a multidimensional evaluation framework that jointly assesses final-answer accuracy and process-level reasoning quality, where Chain-of-Thought (CoT) traces serve as an observable signal to diagnose catastrophic forgetting beyond answer-only evaluation. Second, we conduct a large-scale evaluation of continual learning methods by systematically assessing eight representative algorithms from four major families under a unified protocol across task orders, providing actionable insights for algorithm design. Third, we expand the scope from Supervised Fine-Tuning (SFT) to Reinforcement Fine-Tuning (RFT) in CIT. By investigating GRPO, an on-policy RL algorithm that stabilizes updates through explicit KL-divergence control to a prior policy, we aim to analyze how this mechanism affects cross-task knowledge retention. Our experiments yield several findings:(1) Process-level reasoning quality is often more resilient to catastrophic forgetting than final-answer accuracy, and forgetting is primarily driven by degradation in domain knowledge. (2) Model capability is critical factor influencing continual learning outcomes, with stronger baseline models exhibiting greater resistance to catastrophic forgetting. (3) On-policy RFT (GRPO), with its inherent KL control, achieves more stable cross-task retention than SFT. While removing KL control can amplify forgetting despite potential gains on new ones.", "date": "2025-07-31", "categories": []}
{"id": "2602.12746", "title": "Lamer-SSL: Layer-aware Mixture of LoRA Experts for Continual Multilingual Expansion of Self-supervised Models without Forgetting", "abstract": "Despite their impressive performance, self-supervised speech models often struggle to generalize to new languages and tend to forget previously acquired knowledge during continual training. To address this, we propose Lamer-SSL, a parameter-efficient framework that integrates a Layer-Aware MixturE of LoRA Experts (Lamer) module with a replay strategy. The Lamer module enables flexible balancing between shared and language-specific representations, while layer-aware expert allocation assigns more experts to deeper layers where semantic information is richer. Meanwhile, the replay strategy retains prior knowledge using minimal data, mitigating forgetting during continual training. Experiments on automatic speech recognition (ASR) and language identification (LID) demonstrate that Lamer-SSL extends self-supervised models to new languages effectively while maintaining strong performance on previously learned languages with only 2.14% parameters being trainable.", "date": "2026-02-13", "categories": []}
{"id": "2602.11516", "title": "Human-Inspired Continuous Learning of Internal Reasoning Processes: Learning How to Think for Adaptive AI Systems", "abstract": "Learning internal reasoning processes is crucial for developing AI systems capable of sustained adaptation in dynamic real-world environments. However, most existing approaches primarily emphasize learning task-specific outputs or static knowledge representations, while overlooking the continuous refinement of internal reasoning structures, action scheduling policies, and learning mechanisms themselves. In this paper, we propose a human-inspired continuous learning framework that unifies reasoning, action, reflection, and verification within a sequential reasoning model enhanced by parallel learning. The framework explicitly treats internal thinking processes as primary learning objects. It systematically records internal reasoning trajectories and environmental interactions as structured learning material, enabling the system to optimize not only task-level content but also the organization, scheduling, and evolution of reasoning activities. This design realizes learning alongside processing, allowing cognitive structures to improve during execution. Furthermore, the framework supports controlled replacement of predefined logic with learned procedures and introduces a hierarchical learning-to-learn mechanism that jointly adapts task-level parameters and learning strategies. As a result, the system progressively evolves its internal cognitive architecture while preserving operational stability. Experimental results on a temperature sensor abnormality detection task show that incorporating internal-process learning reduces average runtime by 23.9%.", "date": "2026-02-12", "categories": []}
{"id": "2505.13944", "title": "WAVE++: Capturing Within-Task Variance for Continual Relation Extraction with Adaptive Prompting", "abstract": "Memory-based approaches have shown strong performance in Continual Relation Extraction (CRE). However, storing examples from previous tasks increases memory usage and raises privacy concerns. Recently, prompt-based methods have emerged as a promising alternative, as they do not rely on storing past samples. Despite this progress, current prompt-based techniques face several core challenges in CRE, particularly in accurately identifying task identities and mitigating catastrophic forgetting. Existing prompt selection strategies often suffer from inaccuracies, lack robust mechanisms to prevent forgetting in shared parameters, and struggle to handle both cross-task and within-task variations. In this paper, we propose WAVE++, a novel approach inspired by the connection between prefix-tuning and mixture of experts. Specifically, we introduce task-specific prompt pools that enhance flexibility and adaptability across diverse tasks while avoiding boundary-spanning risks; this design more effectively captures both within-task and cross-task variations. To further refine relation classification, we incorporate label descriptions that provide richer, more global context, enabling the model to better distinguish among different relations. We also propose a training-free mechanism to improve task prediction during inference. Moreover, we integrate a generative model to consolidate prior knowledge within the shared parameters, thereby removing the need for explicit data storage. Extensive experiments demonstrate that WAVE++ outperforms state-of-the-art prompt-based and rehearsal-based methods, offering a more robust solution for continual relation extraction. Our code is publicly available at https://github.com/PiDinosauR2804/WAVE-CRE-PLUS-PLUS.", "date": "2025-05-20", "categories": []}
{"id": "2506.03956", "title": "Adapt before Continual Learning", "abstract": "Continual Learning (CL) seeks to enable neural networks to incrementally acquire new knowledge (plasticity) while retaining existing knowledge (stability). Although pre-trained models (PTMs) have provided a strong foundation for CL, existing approaches face a fundamental challenge in balancing these two competing objectives. Current methods typically address stability by freezing the PTM backbone, which severely limits the model's plasticity, particularly when incoming data distribution diverges largely from the pre-training data. Alternatively, sequentially fine-tuning the entire PTM can adapt to new knowledge but often leads to catastrophic forgetting, highlighting the critical stability-plasticity trade-off in PTM-based CL. To address this limitation, we propose Adapting PTMs before the core CL} process (ACL), a novel framework that introduces a plug-and-play adaptation phase prior to learning each new task. During this phase, ACL refines the PTM backbone by aligning embeddings with their original class prototypes while distancing them from irrelevant classes. This mechanism theoretically and empirically demonstrates desirable balance between stability and plasticity, significantly improving CL performance across benchmarks and integrated methods. Code is available at https://github.com/byyx666/ACL_code.", "date": "2025-06-04", "categories": []}
{"id": "2601.00725", "title": "Multi-Level Feature Fusion for Continual Learning in Visual Quality Inspection", "abstract": "Deep neural networks show great potential for automating various visual quality inspection tasks in manufacturing. However, their applicability is limited in more volatile scenarios, such as remanufacturing, where the inspected products and defect patterns often change. In such settings, deployed models require frequent adaptation to novel conditions, effectively posing a continual learning problem. To enable quick adaptation, the necessary training processes must be computationally efficient while still avoiding effects like catastrophic forgetting. This work presents a multi-level feature fusion (MLFF) approach that aims to improve both aspects simultaneously by utilizing representations from different depths of a pretrained network. We show that our approach is able to match the performance of end-to-end training for different quality inspection problems while using significantly less trainable parameters. Furthermore, it reduces catastrophic forgetting and improves generalization robustness to new product types or defects.", "date": "2026-01-02", "categories": []}
{"id": "2602.07950", "title": "A Thermodynamic Theory of Learning Part II: Critical Period Closure and Continual Learning Failure", "abstract": "Learning performed over finite time is inherently irreversible. In Part~I of this series, we modeled learning as a transport process in the space of parameter distributions and derived the Epistemic Speed Limit (ESL), which lower-bounds entropy production under finite-time dynamics.\n  In this work (Part~II), we show that irreversibility imposes a geometric restriction on future adaptability through the compositional structure of learning dynamics. Successive learning phases compose multiplicatively as transport maps, and their Jacobians form a semigroup whose rank and singular values are submultiplicative. As a result, dynamically usable degrees of reconfiguration can only decrease under composition.\n  We formalize the remaining adaptability of a model in terms of compatible effective rank, defined as the log-volume of task-preserving directions that remain dynamically accessible. Although task performance may remain unchanged, finite-time learning can progressively reduce this reconfiguration capacity.\n  We prove a capacity-threshold criterion for continual learning: let m_B denote the stable rank of the Hessian of a new task B restricted to the task-preserving manifold of a previously learned task A. If m_B exceeds the residual compatible effective rank, then task B is trajectory-level incompatible with task A; any sufficient adaptation necessarily induces forgetting.\n  Thus catastrophic forgetting arises not from the absence of multi-task solutions, but from irreversible loss of reconfiguration capacity under compositional learning dynamics. This establishes a trajectory-level capacity limit for continual learning.", "date": "2026-02-08", "categories": []}
{"id": "2508.16651", "title": "HiCL: Hippocampal-Inspired Continual Learning", "abstract": "We propose HiCL, a novel hippocampal-inspired dual-memory continual learning architecture designed to mitigate catastrophic forgetting by using elements inspired by the hippocampal circuitry. Our system encodes inputs through a grid-cell-like layer, followed by sparse pattern separation using a dentate gyrus-inspired module with top-k sparsity. Episodic memory traces are maintained in a CA3-like autoassociative memory. Task-specific processing is dynamically managed via a DG-gated mixture-of-experts mechanism, wherein inputs are routed to experts based on cosine similarity between their normalized sparse DG representations and learned task-specific DG prototypes computed through online exponential moving averages. This biologically grounded yet mathematically principled gating strategy enables differentiable, scalable task-routing without relying on a separate gating network, and enhances the model's adaptability and efficiency in learning multiple sequential tasks. Cortical outputs are consolidated using Elastic Weight Consolidation weighted by inter-task similarity. Crucially, we incorporate prioritized replay of stored patterns to reinforce essential past experiences. Evaluations on standard continual learning benchmarks demonstrate the effectiveness of our architecture in reducing task interference, achieving near state-of-the-art results in continual learning tasks at lower computational costs. Our code is available here https://github.com/kushalk173-sc/HiCL.", "date": "2025-08-19", "categories": []}
{"id": "2602.09681", "title": "Resilient Class-Incremental Learning: on the Interplay of Drifting, Unlabelled and Imbalanced Data Streams", "abstract": "In today's connected world, the generation of massive streaming data across diverse domains has become commonplace. In the presence of concept drift, class imbalance, label scarcity, and new class emergence, they jointly degrade representation stability, bias learning toward outdated distributions, and reduce the resilience and reliability of detection in dynamic environments. This paper proposes SCIL (Streaming Class-Incremental Learning) to address these challenges. The SCIL framework integrates an autoencoder (AE) with a multi-layer perceptron for multi-class prediction, uses a dual-loss strategy (classification and reconstruction) for prediction and new class detection, employs corrected pseudo-labels for online training, manages classes with queues, and applies oversampling to handle imbalance. The rationale behind the method's structure is elucidated through ablation studies and a comprehensive experimental evaluation is performed using both real-world and synthetic datasets that feature class imbalance, incremental classes, and concept drifts. Our results demonstrate that SCIL outperforms strong baselines and state-of-the-art methods. Based on our commitment to Open Science, we make our code and datasets available to the community.", "date": "2026-02-10", "categories": []}
{"id": "2602.09720", "title": "Continual Learning for non-stationary regression via Memory-Efficient Replay", "abstract": "Data streams are rarely static in dynamic environments like Industry 4.0. Instead, they constantly change, making traditional offline models outdated unless they can quickly adjust to the new data. This need can be adequately addressed by continual learning (CL), which allows systems to gradually acquire knowledge without incurring the prohibitive costs of retraining them from scratch. Most research on continual learning focuses on classification problems, while very few studies address regression tasks. We propose the first prototype-based generative replay framework designed for online task-free continual regression. Our approach defines an adaptive output-space discretization model, enabling prototype-based generative replay for continual regression without storing raw data. Evidence obtained from several benchmark datasets shows that our framework reduces forgetting and provides more stable performance than other state-of-the-art solutions.", "date": "2026-02-10", "categories": []}
{"id": "2602.06993", "title": "Attractor Patch Networks: Reducing Catastrophic Forgetting with Routed Low-Rank Patch Experts", "abstract": "Transformers achieve strong language modeling accuracy, yet their position-wise feed-forward networks (FFNs) are dense, globally shared, and typically updated end to end. These properties create two practical tensions. First, dense FFNs spend the same compute on every token regardless of context, and they allocate capacity uniformly even when language exhibits highly clustered context structure. Second, continual learning, in the sense of updating the model while serving a data stream, often produces interference because a small update touches broadly shared weights.\n  We propose Attractor Patch Networks (APN), a plug-compatible replacement for the Transformer FFN. APN is a bank of patch experts. A similarity router selects a small top-k set of patches for each token by matching the token representation to learned prototypes. Each selected patch emits a low-rank residual update conditioned on a compact code. The architecture yields conditional, context-specialized nonlinear transformations while preserving the standard Transformer interface.\n  This paper focuses on APN as an architectural primitive. We formalize APN, analyze its expressivity as a piecewise low-rank residual function class, and derive simple interference and stability arguments that make APN naturally compatible with continual learning. In experiments on character-level language modeling, APN achieves competitive perplexity (4.57 vs 4.32 PPL) while enabling dramatically better continual adaptation: when adapting to a shifted domain, APN achieves 2.6 times better retention (11.1 vs 29.4 PPL on the original domain) and 2.8 times better adaptation (6.4 vs 17.8 PPL on the new domain) compared to global fine-tuning of a dense FFN baseline.", "date": "2026-01-28", "categories": []}
{"id": "2602.07755", "title": "Learning to Continually Learn via Meta-learning Agentic Memory Designs", "abstract": "The statelessness of foundation models bottlenecks agentic systems' ability to continually learn, a core capability for long-horizon reasoning and adaptation. To address this limitation, agentic systems commonly incorporate memory modules to retain and reuse past experience, aiming for continual learning during test time. However, most existing memory designs are human-crafted and fixed, which limits their ability to adapt to the diversity and non-stationarity of real-world tasks. In this paper, we introduce ALMA (Automated meta-Learning of Memory designs for Agentic systems), a framework that meta-learns memory designs to replace hand-engineered memory designs, therefore minimizing human effort and enabling agentic systems to be continual learners across diverse domains. Our approach employs a Meta Agent that searches over memory designs expressed as executable code in an open-ended manner, theoretically allowing the discovery of arbitrary memory designs, including database schemas as well as their retrieval and update mechanisms. Extensive experiments across four sequential decision-making domains demonstrate that the learned memory designs enable more effective and efficient learning from experience than state-of-the-art human-crafted memory designs on all benchmarks. When developed and deployed safely, ALMA represents a step toward self-improving AI systems that learn to be adaptive, continual learners.", "date": "2026-02-08", "categories": []}
{"id": "2602.08036", "title": "TAAM:Inductive Graph-Class Incremental Learning with Task-Aware Adaptive Modulation", "abstract": "Graph Continual Learning (GCL) aims to solve the challenges of streaming graph data. However, current methods often depend on replay-based strategies, which raise concerns like memory limits and privacy issues, while also struggling to resolve the stability-plasticity dilemma. In this paper, we suggest that lightweight, task-specific modules can effectively guide the reasoning process of a fixed GNN backbone. Based on this idea, we propose Task-Aware Adaptive Modulation (TAAM). The key component of TAAM is its lightweight Neural Synapse Modulators (NSMs). For each new task, a dedicated NSM is trained and then frozen, acting as an \"expert module.\" These modules perform detailed, node-attentive adaptive modulation on the computational flow of a shared GNN backbone. This setup ensures that new knowledge is kept within compact, task-specific modules, naturally preventing catastrophic forgetting without using any data replay. Additionally, to address the important challenge of unknown task IDs in real-world scenarios, we propose and theoretically prove a novel method named Anchored Multi-hop Propagation (AMP). Notably, we find that existing GCL benchmarks have flaws that can cause data leakage and biased evaluations. Therefore, we conduct all experiments in a more rigorous inductive learning scenario. Extensive experiments show that TAAM comprehensively outperforms state-of-the-art methods across eight datasets. Code and Datasets are available at: https://github.com/1iuJT/TAAM_AAMAS2026.", "date": "2026-02-08", "categories": []}
{"id": "2506.00956", "title": "Continual-MEGA: A Large-scale Benchmark for Generalizable Continual Anomaly Detection", "abstract": "In this paper, we introduce a new benchmark for continual learning in anomaly detection, aimed at better reflecting real-world deployment scenarios. Our benchmark, Continual-MEGA, includes a large and diverse dataset that significantly expands existing evaluation settings by combining carefully curated existing datasets with our newly proposed dataset, ContinualAD. In addition to standard continual learning with expanded quantity, we propose a novel scenario that measures zero-shot generalization to unseen classes, those not observed during continual adaptation. This setting poses a new problem setting that continual adaptation also enhances zero-shot performance. We also present a unified baseline algorithm that improves robustness in few-shot detection and maintains strong generalization. Through extensive evaluations, we report three key findings: (1) existing methods show substantial room for improvement, particularly in pixel-level defect localization; (2) our proposed method consistently outperforms prior approaches; and (3) the newly introduced ContinualAD dataset enhances the performance of strong anomaly detection models. We release the benchmark and code in https://github.com/Continual-Mega/Continual-Mega.", "date": "2025-06-01", "categories": []}
{"id": "2602.06328", "title": "Adaptive and Balanced Re-initialization for Long-timescale Continual Test-time Domain Adaptation", "abstract": "Continual test-time domain adaptation (CTTA) aims to adjust models so that they can perform well over time across non-stationary environments. While previous methods have made considerable efforts to optimize the adaptation process, a crucial question remains: Can the model adapt to continually changing environments over a long time? In this work, we explore facilitating better CTTA in the long run using a re-initialization (or reset) based method. First, we observe that the long-term performance is associated with the trajectory pattern in label flip. Based on this observed correlation, we propose a simple yet effective policy, Adaptive-and-Balanced Re-initialization (ABR), towards preserving the model's long-term performance. In particular, ABR performs weight re-initialization using adaptive intervals. The adaptive interval is determined based on the change in label flip. The proposed method is validated on extensive CTTA benchmarks, achieving superior performance.", "date": "2026-02-06", "categories": []}
{"id": "2602.06586", "title": "Degradation of Feature Space in Continual Learning", "abstract": "Centralized training is the standard paradigm in deep learning, enabling models to learn from a unified dataset in a single location. In such setup, isotropic feature distributions naturally arise as a mean to support well-structured and generalizable representations. In contrast, continual learning operates on streaming and non-stationary data, and trains models incrementally, inherently facing the well-known plasticity-stability dilemma. In such settings, learning dynamics tends to yield increasingly anisotropic feature space. This arises a fundamental question: should isotropy be enforced to achieve a better balance between stability and plasticity, and thereby mitigate catastrophic forgetting? In this paper, we investigate whether promoting feature-space isotropy can enhance representation quality in continual learning. Through experiments using contrastive continual learning techniques on CIFAR-10 and CIFAR-100 data, we find that isotropic regularization fails to improve, and can in fact degrade, model accuracy in continual settings. Our results highlight essential differences in feature geometry between centralized and continual learning, suggesting that isotropy, while beneficial in centralized setups, may not constitute an appropriate inductive bias for non-stationary learning scenarios.", "date": "2026-02-06", "categories": []}
{"id": "2412.14865", "title": "Hierarchical Subspaces of Policies for Continual Offline Reinforcement Learning", "abstract": "We consider a Continual Reinforcement Learning setup, where a learning agent must continuously adapt to new tasks while retaining previously acquired skill sets, with a focus on the challenge of avoiding forgetting past gathered knowledge and ensuring scalability with the growing number of tasks. Such issues prevail in autonomous robotics and video game simulations, notably for navigation tasks prone to topological or kinematic changes. To address these issues, we introduce HiSPO, a novel hierarchical framework designed specifically for continual learning in navigation settings from offline data. Our method leverages distinct policy subspaces of neural networks to enable flexible and efficient adaptation to new tasks while preserving existing knowledge. We demonstrate, through a careful experimental study, the effectiveness of our method in both classical MuJoCo maze environments and complex video game-like navigation simulations, showcasing competitive performances and satisfying adaptability with respect to classical continual learning metrics, in particular regarding the memory usage and efficiency.", "date": "2024-12-19", "categories": []}
{"id": "2510.04660", "title": "An Attention-based Feature Memory Design for Energy-Efficient Continual Learning", "abstract": "Tabular data streams are increasingly prevalent in real-time decision-making across healthcare, finance, and the Internet of Things, often generated and processed on resource-constrained edge and mobile devices. Continual learning (CL) enables models to learn sequentially from such streams while retaining previously acquired knowledge. While recent CL advances have made significant progress in mitigating catastrophic forgetting, the energy and memory efficiency of CL for tabular data streams remains largely unexplored. To address this gap, we propose AttenMLP, which integrates attention-based feature replay with context retrieval and sliding buffer updates within a minibatch training framework for streaming tabular learning.\n  We evaluate AttenMLP against state-of-the-art (SOTA) tabular models on real-world concept drift benchmarks with temporal distribution shifts. Experimental results show that AttenMLP achieves accuracy comparable to strong baselines without replay, while substantially reducing energy consumption through tunable design choices. In particular, with the proposed attention-based feature memory design, AttenMLP costs a 0.062 decrease in final accuracy under the incremental concept drift dataset, while reducing energy usage up to 33.3\\% compared to TabPFNv2. Under the abrupt concept drift dataset, AttenMLP reduces 1.47\\% energy consumption compared to TabR, at the cost of a 0.038 decrease in final accuracy. Although ranking third in global efficiency, AttenMLP demonstrates energy-accuracy trade-offs across both abrupt and incremental concept drift scenarios compared to SOTA tabular models.", "date": "2025-10-06", "categories": []}
{"id": "2602.00166", "title": "Joint Continual Learning of Local Language Models and Cloud Offloading Decisions with Budget Constraints", "abstract": "Locally deployed Small Language Models (SLMs) must continually support diverse tasks under strict memory and computation constraints, making selective reliance on cloud Large Language Models (LLMs) unavoidable. Regulating cloud assistance during continual learning is challenging, as naive reward-based reinforcement learning often yields unstable offloading behavior and exacerbates catastrophic forgetting as task distributions shift. We propose DA-GRPO, a dual-advantage extension of Group Relative Policy Optimization that incorporates cloud-usage constraints directly into advantage computation, avoiding fixed reward shaping and external routing models. This design enables the local model to jointly learn task competence and collaboration behavior, allowing cloud requests to emerge naturally during post-training while respecting a prescribed assistance budget. Experiments on mathematical reasoning and code generation benchmarks show that DA-GRPO improves post-switch accuracy, substantially reduces forgetting, and maintains stable cloud usage compared to prior collaborative and routing-based approaches.", "date": "2026-01-29", "categories": []}
{"id": "2602.05085", "title": "Locas: Your Models are Principled Initializers of Locally-Supported Parametric Memories", "abstract": "In this paper, we aim to bridge test-time-training with a new type of parametric memory that can be flexibly offloaded from or merged into model parameters. We present Locas, a Locally-Supported parametric memory that shares the design of FFN blocks in modern transformers, allowing it to be flexibly permanentized into the model parameters while supporting efficient continual learning. We discuss two major variants of Locas: one with a conventional two-layer MLP design that has a clearer theoretical guarantee; the other one shares the same GLU-FFN structure with SOTA LLMs, and can be easily attached to existing models for both parameter-efficient and computation-efficient continual learning. Crucially, we show that proper initialization of such low-rank sideway-FFN-style memories -- performed in a principled way by reusing model parameters, activations and/or gradients -- is essential for fast convergence, improved generalization, and catastrophic forgetting prevention. We validate the proposed memory mechanism on the PG-19 whole-book language modeling and LoCoMo long-context dialogue question answering tasks. With only 0.02\\% additional parameters in the lowest case, Locas-GLU is capable of storing the information from past context while maintaining a much smaller context window. In addition, we also test the model's general capability loss after memorizing the whole book with Locas, through comparative MMLU evaluation. Results show the promising ability of Locas to permanentize past context into parametric knowledge with minimized catastrophic forgetting of the model's existing internal knowledge.", "date": "2026-02-04", "categories": []}
{"id": "2602.05271", "title": "Unlocking Prototype Potential: An Efficient Tuning Framework for Few-Shot Class-Incremental Learning", "abstract": "Few-shot class-incremental learning (FSCIL) seeks to continuously learn new classes from very limited samples while preserving previously acquired knowledge. Traditional methods often utilize a frozen pre-trained feature extractor to generate static class prototypes, which suffer from the inherent representation bias of the backbone. While recent prompt-based tuning methods attempt to adapt the backbone via minimal parameter updates, given the constraint of extreme data scarcity, the model's capacity to assimilate novel information and substantively enhance its global discriminative power is inherently limited. In this paper, we propose a novel shift in perspective: freezing the feature extractor while fine-tuning the prototypes. We argue that the primary challenge in FSCIL is not feature acquisition, but rather the optimization of decision regions within a static, high-quality feature space. To this end, we introduce an efficient prototype fine-tuning framework that evolves static centroids into dynamic, learnable components. The framework employs a dual-calibration method consisting of class-specific and task-aware offsets. These components function synergistically to improve the discriminative capacity of prototypes for ongoing incremental classes. Extensive results demonstrate that our method attains superior performance across multiple benchmarks while requiring minimal learnable parameters.", "date": "2026-02-05", "categories": []}
{"id": "2602.05454", "title": "Attention Retention for Continual Learning with Vision Transformers", "abstract": "Continual learning (CL) empowers AI systems to progressively acquire knowledge from non-stationary data streams. However, catastrophic forgetting remains a critical challenge. In this work, we identify attention drift in Vision Transformers as a primary source of catastrophic forgetting, where the attention to previously learned visual concepts shifts significantly after learning new tasks. Inspired by neuroscientific insights into the selective attention in the human visual system, we propose a novel attention-retaining framework to mitigate forgetting in CL. Our method constrains attention drift by explicitly modifying gradients during backpropagation through a two-step process: 1) extracting attention maps of the previous task using a layer-wise rollout mechanism and generating instance-adaptive binary masks, and 2) when learning a new task, applying these masks to zero out gradients associated with previous attention regions, thereby preventing disruption of learned visual concepts. For compatibility with modern optimizers, the gradient masking process is further enhanced by scaling parameter updates proportionally to maintain their relative magnitudes. Experiments and visualizations demonstrate the effectiveness of our method in mitigating catastrophic forgetting and preserving visual concepts. It achieves state-of-the-art performance and exhibits robust generalizability across diverse CL scenarios.", "date": "2026-02-05", "categories": []}
{"id": "2602.06043", "title": "Shared LoRA Subspaces for almost Strict Continual Learning", "abstract": "Adapting large pretrained models to new tasks efficiently and continually is crucial for real-world deployment but remains challenging due to catastrophic forgetting and the high cost of retraining. While parameter-efficient tuning methods like low rank adaptation (LoRA) reduce computational demands, they lack mechanisms for strict continual learning and knowledge integration, without relying on data replay, or multiple adapters. We propose Share, a novel approach to parameter efficient continual finetuning that learns and dynamically updates a single, shared low-rank subspace, enabling seamless adaptation across multiple tasks and modalities. Share constructs a foundational subspace that extracts core knowledge from past tasks and incrementally integrates new information by identifying essential subspace directions. Knowledge from each new task is incorporated into this evolving subspace, facilitating forward knowledge transfer, while minimizing catastrophic interference. This approach achieves up to 100x parameter reduction and 281x memory savings over traditional LoRA methods, maintaining performance comparable to jointly trained models. A single Share model can replace hundreds of task-specific LoRA adapters, supporting scalable, asynchronous continual learning. Experiments across image classification, natural language understanding, 3D pose estimation, and text-to-image generation validate its effectiveness, making Share a practical and scalable solution for lifelong learning in large-scale AI systems.", "date": "2026-02-05", "categories": []}
{"id": "2602.04252", "title": "ACIL: Active Class Incremental Learning for Image Classification", "abstract": "Continual learning (or class incremental learning) is a realistic learning scenario for computer vision systems, where deep neural networks are trained on episodic data, and the data from previous episodes are generally inaccessible to the model. Existing research in this domain has primarily focused on avoiding catastrophic forgetting, which occurs due to the continuously changing class distributions in each episode and the inaccessibility of the data from previous episodes. However, these methods assume that all the training samples in every episode are annotated; this not only incurs a huge annotation cost, but also results in a wastage of annotation effort, since most of the samples in a given episode will not be accessible to the model in subsequent episodes. Active learning algorithms identify the salient and informative samples from large amounts of unlabeled data and are instrumental in reducing the human annotation effort in inducing a deep neural network. In this paper, we propose ACIL, a novel active learning framework for class incremental learning settings. We exploit a criterion based on uncertainty and diversity to identify the exemplar samples that need to be annotated in each episode, and will be appended to the data in the next episode. Such a framework can drastically reduce annotation cost and can also avoid catastrophic forgetting. Our extensive empirical analyses on several vision datasets corroborate the promise and potential of our framework against relevant baselines.", "date": "2026-02-04", "categories": []}
{"id": "2602.04542", "title": "Continual Learning through Control Minimization", "abstract": "Catastrophic forgetting remains a fundamental challenge for neural networks when tasks are trained sequentially. In this work, we reformulate continual learning as a control problem where learning and preservation signals compete within neural activity dynamics. We convert regularization penalties into preservation signals that protect prior-task representations. Learning then proceeds by minimizing the control effort required to integrate new tasks while competing with the preservation of prior tasks. At equilibrium, the neural activities produce weight updates that implicitly encode the full prior-task curvature, a property we term the continual-natural gradient, requiring no explicit curvature storage. Experiments confirm that our learning framework recovers true prior-task curvature and enables task discrimination, outperforming existing methods on standard benchmarks without replay.", "date": "2026-02-04", "categories": []}
{"id": "2602.04868", "title": "CRoSS: A Continual Robotic Simulation Suite for Scalable Reinforcement Learning with High Task Diversity and Realistic Physics Simulation", "abstract": "Continual reinforcement learning (CRL) requires agents to learn from a sequence of tasks without forgetting previously acquired policies. In this work, we introduce a novel benchmark suite for CRL based on realistically simulated robots in the Gazebo simulator. Our Continual Robotic Simulation Suite (CRoSS) benchmarks rely on two robotic platforms: a two-wheeled differential-drive robot with lidar, camera and bumper sensor, and a robotic arm with seven joints. The former represent an agent in line-following and object-pushing scenarios, where variation of visual and structural parameters yields a large number of distinct tasks, whereas the latter is used in two goal-reaching scenarios with high-level cartesian hand position control (modeled after the Continual World benchmark), and low-level control based on joint angles. For the robotic arm benchmarks, we provide additional kinematics-only variants that bypass the need for physical simulation (as long as no sensor readings are required), and which can be run two orders of magnitude faster. CRoSS is designed to be easily extensible and enables controlled studies of continual reinforcement learning in robotic settings with high physical realism, and in particular allow the use of almost arbitrary simulated sensors. To ensure reproducibility and ease of use, we provide a containerized setup (Apptainer) that runs out-of-the-box, and report performances of standard RL algorithms, including Deep Q-Networks (DQN) and policy gradient methods. This highlights the suitability as a scalable and reproducible benchmark for CRL research.", "date": "2026-02-04", "categories": []}
{"id": "2602.04881", "title": "Contrastive Continual Learning for Model Adaptability in Internet of Things", "abstract": "Internet of Things (IoT) deployments operate in nonstationary, dynamic environments where factors such as sensor drift, evolving user behavior, and heterogeneous user privacy requirements can affect application utility. Continual learning (CL) addresses this by adapting models over time without catastrophic forgetting. Meanwhile, contrastive learning has emerged as a powerful representation-learning paradigm that improves robustness and sample efficiency in a self-supervised manner. This paper reviews the usage of \\emph{contrastive continual learning} (CCL) for IoT, connecting algorithmic design (replay, regularization, distillation, prompts) with IoT system realities (TinyML constraints, intermittent connectivity, privacy). We present a unifying problem formulation, derive common objectives that blend contrastive and distillation losses, propose an IoT-oriented reference architecture for on-device, edge, and cloud-based CCL, and provide guidance on evaluation protocols and metrics. Finally, we highlight open unique challenges with respect to the IoT domain, such as spanning tabular and streaming IoT data, concept drift, federated settings, and energy-aware training.", "date": "2026-02-04", "categories": []}
{"id": "2509.19664", "title": "MoTiC: Momentum Tightness and Contrast for Few-Shot Class-Incremental Learning", "abstract": "Few-Shot Class-Incremental Learning (FSCIL) must contend with the dual challenge of learning new classes from scarce samples while preserving old class knowledge. Existing methods use the frozen feature extractor and class-averaged prototypes to mitigate against catastrophic forgetting and overfitting. However, new-class prototypes suffer significant estimation bias due to extreme data scarcity, whereas base-class prototypes benefit from sufficient data. In this work, we theoretically demonstrate that aligning the new-class priors with old-class statistics via Bayesian analysis reduces variance and improves prototype accuracy. Furthermore, we propose large-scale contrastive learning to enforce cross-category feature tightness. To further enrich feature diversity and inject prior information for new-class prototypes, we integrate momentum self-supervision and virtual categories into the Momentum Tightness and Contrast framework (MoTiC), constructing a feature space with rich representations and enhanced interclass cohesion. Experiments on three FSCIL benchmarks produce state-of-the-art performances, particularly on the fine-grained task CUB-200, validating our method's ability to reduce estimation bias and improve incremental learning robustness.", "date": "2025-09-24", "categories": []}
{"id": "2602.02502", "title": "Sparse Adapter Fusion for Continual Learning in NLP", "abstract": "Continual learning in natural language processing plays a crucial role in adapting to evolving data and preventing catastrophic forgetting. Despite significant progress, existing methods still face challenges, such as inefficient parameter reuse across tasks, risking catastrophic forgetting when tasks are dissimilar, and the unnecessary introduction of new parameters for each task, which hampers knowledge sharing among similar tasks. To tackle these issues, we propose a Sparse Adapter Fusion Method (SAFM), which dynamically fuses old and new adapters to address these challenges. SAFM operates in two stages: the decision stage and the tuning stage. In the decision stage, SAFM determines whether to incorporate a new adapter, reuse an existing one, or add an empty adapter. The architecture search procedure, designed to prioritize reusing or adding empty adapters, minimizes parameter consumption and maximizes reuse. In the tuning stage, SAFM especially facilitates a layer-wise loss to encourage differentiation between adapters, effectively capturing knowledge within the same task. Experimental results consistently show that SAFM outperforms state-of-the-art (SOTA) methods, achieving comparable performance while utilizing less than 60% of the parameters.", "date": "2026-01-20", "categories": []}
{"id": "2602.02568", "title": "Mitigating Task-Order Sensitivity and Forgetting via Hierarchical Second-Order Consolidation", "abstract": "We introduce $\\textbf{Hierarchical Taylor Series-based Continual Learning (HTCL)}$, a framework that couples fast local adaptation with conservative, second-order global consolidation to address the high variance introduced by random task ordering. To address task-order effects, HTCL identifies the best intra-group task sequence and integrates the resulting local updates through a Hessian-regularized Taylor expansion, yielding a consolidation step with theoretical guarantees. The approach naturally extends to an $L$-level hierarchy, enabling multiscale knowledge integration in a manner not supported by conventional single-level CL systems. Across a wide range of datasets and replay and regularization baselines, HTCL acts as a model-agnostic consolidation layer that consistently enhances performance, yielding mean accuracy gains of $7\\%$ to $25\\%$ while reducing the standard deviation of final accuracy by up to $68\\%$ across random task permutations.", "date": "2026-01-31", "categories": []}
{"id": "2602.02767", "title": "Provable Effects of Data Replay in Continual Learning: A Feature Learning Perspective", "abstract": "Continual learning (CL) aims to train models on a sequence of tasks while retaining performance on previously learned ones. A core challenge in this setting is catastrophic forgetting, where new learning interferes with past knowledge. Among various mitigation strategies, data-replay methods, where past samples are periodically revisited, are considered simple yet effective, especially when memory constraints are relaxed. However, the theoretical effectiveness of full data replay, where all past data is accessible during training, remains largely unexplored. In this paper, we present a comprehensive theoretical framework for analyzing full data-replay training in continual learning from a feature learning perspective. Adopting a multi-view data model, we identify the signal-to-noise ratio (SNR) as a critical factor affecting forgetting. Focusing on task-incremental binary classification across $M$ tasks, our analysis verifies two key conclusions: (1) forgetting can still occur under full replay when the cumulative noise from later tasks dominates the signal from earlier ones; and (2) with sufficient signal accumulation, data replay can recover earlier tasks-even if their initial learning was poor. Notably, we uncover a novel insight into task ordering: prioritizing higher-signal tasks not only facilitates learning of lower-signal tasks but also helps prevent catastrophic forgetting. We validate our theoretical findings through synthetic and real-world experiments that visualize the interplay between signal learning and noise memorization across varying SNRs and task correlation regimes.", "date": "2026-02-02", "categories": []}
{"id": "2602.03355", "title": "PACE: Pretrained Audio Continual Learning", "abstract": "Audio is a fundamental modality for analyzing speech, music, and environmental sounds. Although pretrained audio models have significantly advanced audio understanding, they remain fragile in real-world settings where data distributions shift over time. In this work, we present the first systematic benchmark for audio continual learning (CL) with pretrained models (PTMs), together with a comprehensive analysis of its unique challenges. Unlike in vision, where parameter-efficient fine-tuning (PEFT) has proven effective for CL, directly transferring such strategies to audio leads to poor performance. This stems from a fundamental property of audio backbones: they focus on low-level spectral details rather than structured semantics, causing severe upstream-downstream misalignment. Through extensive empirical study, we identify analytic classifiers with first-session adaptation (FSA) as a promising direction, but also reveal two major limitations: representation saturation in coarse-grained scenarios and representation drift in fine-grained scenarios. To address these challenges, we propose PACE, a novel method that enhances FSA via a regularized analytic classifier and enables multi-session adaptation through adaptive subspace-orthogonal PEFT for improved semantic alignment. In addition, we introduce spectrogram-based boundary-aware perturbations to mitigate representation overlap and improve stability. Experiments on six diverse audio CL benchmarks demonstrate that PACE substantially outperforms state-of-the-art baselines, marking an important step toward robust and scalable audio continual learning with PTMs.", "date": "2026-02-03", "categories": []}
{"id": "2602.03445", "title": "CRL-VLA: Continual Vision-Language-Action Learning", "abstract": "Lifelong learning is critical for embodied agents in open-world environments, where reinforcement learning fine-tuning has emerged as an important paradigm to enable Vision-Language-Action (VLA) models to master dexterous manipulation through environmental interaction. Thus, Continual Reinforcement Learning (CRL) is a promising pathway for deploying VLA models in lifelong robotic scenarios, yet balancing stability (retaining old skills) and plasticity (learning new ones) remains a formidable challenge for existing methods. We introduce CRL-VLA, a framework for continual post-training of VLA models with rigorous theoretical bounds. We derive a unified performance bound linking the stability-plasticity trade-off to goal-conditioned advantage magnitude, scaled by policy divergence. CRL-VLA resolves this dilemma via asymmetric regulation: constraining advantage magnitudes on prior tasks while enabling controlled growth on new tasks. This is realized through a simple but effective dual-critic architecture with novel Goal-Conditioned Value Formulation (GCVF), where a frozen critic anchors semantic consistency and a trainable estimator drives adaptation. Experiments on the LIBERO benchmark demonstrate that CRL-VLA effectively harmonizes these conflicting objectives, outperforming baselines in both anti-forgetting and forward adaptation.", "date": "2026-02-03", "categories": []}
{"id": "2406.13653", "title": "Dual-Phase Continual Learning: Supervised Adaptation Meets Unsupervised Retention", "abstract": "Foundational Vision-Language Models (VLMs) excel across diverse tasks, but adapting them to new domains without forgetting prior knowledge remains a critical challenge. Continual Learning (CL) addresses this challenge by enabling models to learn sequentially from new data while mitigating the forgetting of prior information, typically under supervised settings involving label shift. Nonetheless, abrupt distribution shifts can still cause substantial forgetting, potentially nullifying the benefits of supervised updates, especially when storing or replaying past data is infeasible. In this work, we propose leveraging unlabeled testtime data in an unsupervised manner to reinforce prior task performance without requiring replay or stored examples. Unlike traditional Test Time Adaptation (TTA), which primarily focuses on domain shift or corruption, our method improves performance on earlier tasks by exploiting representative test samples encountered during deployment. We introduce a simple Teacher-Student framework with gradient-based sparse parameter updates, and show that it effectively mitigates forgetting in class-incremental CL for VLMs, offering a memory-free alternative to episodic replay with strong empirical results.", "date": "2024-06-19", "categories": []}
{"id": "2406.17381", "title": "Retrospective Feature Estimation for Continual Learning", "abstract": "The intrinsic capability to continuously learn a changing data stream is a desideratum of deep neural networks (DNNs). However, current DNNs suffer from catastrophic forgetting, which interferes with remembering past knowledge. To mitigate this issue, existing Continual Learning (CL) approaches often retain exemplars for replay, regularize learning, or allocate dedicated capacity for new tasks. This paper investigates an unexplored direction for CL called Retrospective Feature Estimation (RFE). RFE learns to reverse feature changes by aligning the features from the current trained DNN backward to the feature space of the old task, where performing predictions is easier. This retrospective process utilizes a chain of small feature mapping networks called retrospector modules. Empirical experiments on several CL benchmarks, including CIFAR10, CIFAR100, and Tiny ImageNet, demonstrate the effectiveness and potential of this novel CL direction compared to existing representative CL methods, motivating further research into retrospective mechanisms as a principled alternative for mitigating catastrophic forgetting in CL. Code is available at: https://github.com/mail-research/retrospective-feature-estimation.", "date": "2024-06-25", "categories": []}
{"id": "2505.13811", "title": "Context-Free Synthetic Data Mitigates Forgetting", "abstract": "Fine-tuning a language model often results in a degradation of its existing performance on other tasks, due to a shift in the model parameters; this phenomenon is often referred to as (catastrophic) forgetting. We are interested in mitigating this, in settings where we only have access to the model weights but no access to its training data/recipe. A natural approach is to penalize the KL divergence between the original model and the new one. Our main realization is that a simple process - which we term context-free generation - allows for an approximate unbiased estimation of this KL divergence. We show that augmenting a fine-tuning dataset with context-free generations mitigates forgetting, in two settings: (a) preserving the zero-shot performance of pretrained-only models, and (b) preserving the reasoning performance of thinking models. We show that contextual synthetic data, and even a portion of the pretraining data, are less effective. We also investigate the effect of choices like generation temperature, data ratios etc. We present our results for OLMo-1B for pretrained-only setting and R1-Distill-Llama-8B for the reasoning setting.", "date": "2025-05-20", "categories": []}
{"id": "2508.02426", "title": "Learning to Evolve: Bayesian-Guided Continual Knowledge Graph Embedding", "abstract": "As social media and the World Wide Web become hubs for information dissemination, effectively organizing and understanding the vast amounts of dynamically evolving Web content is crucial. Knowledge graphs (KGs) provide a powerful framework for structuring this information. However, the rapid emergence of new hot topics, user relationships, and events in social media renders traditional static knowledge graph embedding (KGE) models rapidly outdated. Continual Knowledge Graph Embedding (CKGE) aims to address this issue, but existing methods commonly suffer from catastrophic forgetting, whereby older, but still valuable, information is lost when learning new knowledge (such as new memes or trending events). This means the model cannot effectively learn the evolution of the data. We propose a novel CKGE framework, BAKE. Unlike existing methods, BAKE formulates CKGE as a sequential Bayesian inference problem and utilizes the Bayesian posterior update principle as a natural continual learning strategy. This principle is insensitive to data order and provides theoretical guarantees to preserve prior knowledge as much as possible. Specifically, we treat each batch of new data as a Bayesian update to the model's prior. By maintaining the posterior distribution, the model effectively preserves earlier knowledge even as it evolves over multiple snapshots. Furthermore, to constrain the evolution of knowledge across snapshots, we introduce a continual clustering method that maintains the compact cluster structure of entity embeddings through a regularization term, ensuring semantic consistency while allowing controlled adaptation to new knowledge. We conduct extensive experiments on multiple CKGE benchmarks, which demonstrate that BAKE achieves the top performance in the vast majority of cases compared to existing approaches.", "date": "2025-08-04", "categories": []}
{"id": "2511.04666", "title": "Forgetting is Everywhere", "abstract": "A fundamental challenge in developing general learning algorithms is their tendency to forget past knowledge when adapting to new data. Addressing this problem requires a principled understanding of forgetting; yet, despite decades of study, no unified definition has emerged that provides insights into the underlying dynamics of learning. We propose an algorithm- and task-agnostic theory that characterises forgetting as a lack of self-consistency in a learner's predictive distribution, manifesting as a loss of predictive information. Our theory naturally yields a general measure of an algorithm's propensity to forget and demonstrates that exact Bayesian inference allows for adaptation without forgetting. To validate the theory, we design a comprehensive set of experiments that span classification, regression, generative modelling, and reinforcement learning. We empirically demonstrate how forgetting is present across all deep learning settings and plays a significant role in determining learning efficiency. Together, these results establish a principled understanding of forgetting and lay the foundation for analysing and improving the information retention capabilities of general learning algorithms.", "date": "2025-11-06", "categories": []}
{"id": "2601.06758", "title": "A Backpropagation-Free Feedback-Hebbian Network for Continual Learning Dynamics", "abstract": "Feedback-rich neural architectures can regenerate earlier representations and inject temporal context, making them a natural setting for strictly local synaptic plasticity. Existing literature raises doubt about whether a minimal, backpropagation-free feedback-Hebbian system can already express interpretable continual-learning-relevant behaviors under controlled training schedules. In this work, we introduce a compact prediction-reconstruction architecture with a dedicated feedback pathway providing lightweight, locally trainable temporal context for continual adaptation. All synapses are updated by a unified local rule combining centered Hebbian covariance, Oja-style stabilization, and a local supervised drive where targets are available. With a simple two-pair association task, learning is characterized through layer-wise activity snapshots, connectivity trajectories (row and column means of learned weights), and a normalized retention index across phases. Under sequential A to B training, forward output connectivity exhibits a long-term depression (LTD)-like suppression of the earlier association, while feedback connectivity preserves an A-related trace during acquisition of B. Under an alternating sequence, both associations are concurrently maintained rather than sequentially suppressed. Architectural controls and rule-term ablations isolate the role of dedicated feedback in regeneration and co-maintenance, alongside the role of the local supervised term in output selectivity and unlearning. Together, the results show that a compact feedback pathway trained with local plasticity can support regeneration and continual-learning-relevant dynamics in a minimal, mechanistically transparent setting.", "date": "2026-01-11", "categories": []}
{"id": "2602.00144", "title": "Scalable Analytic Classifiers with Associative Drift Compensation for Class-Incremental Learning of Vision Transformers", "abstract": "Class-incremental learning (CIL) with Vision Transformers (ViTs) faces a major computational bottleneck during the classifier reconstruction phase, where most existing methods rely on costly iterative stochastic gradient descent (SGD). We observe that analytic Regularized Gaussian Discriminant Analysis (RGDA) provides a Bayes-optimal alternative with accuracy comparable to SGD-based classifiers; however, its quadratic inference complexity limits its use in large-scale CIL scenarios. To overcome this, we propose Low-Rank Factorized RGDA (LR-RGDA), a scalable classifier that combines RGDA's expressivity with the efficiency of linear classifiers. By exploiting the low-rank structure of the covariance via the Woodbury matrix identity, LR-RGDA decomposes the discriminant function into a global affine term refined by a low-rank quadratic perturbation, reducing the inference complexity from $\\mathcal{O}(Cd^2)$ to $\\mathcal{O}(d^2 + Crd^2)$, where $C$ is the class number, $d$ the feature dimension, and $r \\ll d$ the subspace rank. To mitigate representation drift caused by backbone updates, we further introduce Hopfield-based Distribution Compensator (HopDC), a training-free mechanism that uses modern continuous Hopfield Networks to recalibrate historical class statistics through associative memory dynamics on unlabeled anchors, accompanied by a theoretical bound on the estimation error. Extensive experiments on diverse CIL benchmarks demonstrate that our framework achieves state-of-the-art performance, providing a scalable solution for large-scale class-incremental learning with ViTs. Code: https://github.com/raoxuan98-hash/lr_rgda_hopdc.", "date": "2026-01-29", "categories": []}
{"id": "2602.00573", "title": "When Classes Evolve: A Benchmark and Framework for Stage-Aware Class-Incremental Learning", "abstract": "Class-Incremental Learning (CIL) aims to sequentially learn new classes while mitigating catastrophic forgetting of previously learned knowledge. Conventional CIL approaches implicitly assume that classes are morphologically static, focusing primarily on preserving previously learned representations as new classes are introduced. However, this assumption neglects intra-class evolution: a phenomenon wherein instances of the same semantic class undergo significant morphological transformations, such as a larva turning into a butterfly. Consequently, a model must both discriminate between classes and adapt to evolving appearances within a single class. To systematically address this challenge, we formalize Stage-Aware CIL (Stage-CIL), a paradigm in which each class is learned progressively through distinct morphological stages. To facilitate rigorous evaluation within this paradigm, we introduce the Stage-Bench, a 10-domain, 2-stages dataset and protocol that jointly measure inter- and intra-class forgetting. We further propose STAGE, a novel method that explicitly learns abstract and transferable evolution patterns within a fixed-size memory pool. By decoupling semantic identity from transformation dynamics, STAGE enables accurate prediction of future morphologies based on earlier representations. Extensive empirical evaluation demonstrates that STAGE consistently and substantially outperforms existing state-of-the-art approaches, highlighting its effectiveness in simultaneously addressing inter-class discrimination and intra-class morphological adaptation.", "date": "2026-01-31", "categories": []}
{"id": "2602.00722", "title": "Spectral Imbalance Causes Forgetting in Low-Rank Continual Adaptation", "abstract": "Parameter-efficient continual learning aims to adapt pre-trained models to sequential tasks without forgetting previously acquired knowledge. Most existing approaches treat continual learning as avoiding interference with past updates, rather than considering what properties make the current task-specific update naturally preserve previously acquired knowledge. From a knowledge-decomposition perspective, we observe that low-rank adaptations exhibit highly imbalanced singular value spectra: a few dominant components absorb most of the adaptation energy, thereby (i) more likely to disrupt previously acquired knowledge and (ii) making the update more vulnerable to interference from subsequent tasks. To enable explicit balance among components, we decouple the magnitude of the task update from its directional structure and formulate it as a constrained optimization problem on a restricted Stiefel manifold. We address this problem using a projected first-order method compatible with standard deep-learning optimizers used in vision-language models. Our method mitigates both backward and forward forgetting, consistently outperforming continual learning baselines. The implementation code is available at https://github.com/haodotgu/EBLoRA.", "date": "2026-01-31", "categories": []}
{"id": "2506.02883", "title": "A Continual Offline Reinforcement Learning Benchmark for Navigation Tasks", "abstract": "Autonomous agents operating in domains such as robotics or video game simulations must adapt to changing tasks without forgetting about the previous ones. This process called Continual Reinforcement Learning poses non-trivial difficulties, from preventing catastrophic forgetting to ensuring the scalability of the approaches considered. Building on recent advances, we introduce a benchmark providing a suite of video-game navigation scenarios, thus filling a gap in the literature and capturing key challenges : catastrophic forgetting, task adaptation, and memory efficiency. We define a set of various tasks and datasets, evaluation protocols, and metrics to assess the performance of algorithms, including state-of-the-art baselines. Our benchmark is designed not only to foster reproducible research and to accelerate progress in continual reinforcement learning for gaming, but also to provide a reproducible framework for production pipelines -- helping practitioners to identify and to apply effective approaches.", "date": "2025-06-03", "categories": []}
{"id": "2601.22475", "title": "Continual Policy Distillation from Distributed Reinforcement Learning Teachers", "abstract": "Continual Reinforcement Learning (CRL) aims to develop lifelong learning agents to continuously acquire knowledge across diverse tasks while mitigating catastrophic forgetting. This requires efficiently managing the stability-plasticity dilemma and leveraging prior experience to rapidly generalize to novel tasks. While various enhancement strategies for both aspects have been proposed, achieving scalable performance by directly applying RL to sequential task streams remains challenging. In this paper, we propose a novel teacher-student framework that decouples CRL into two independent processes: training single-task teacher models through distributed RL and continually distilling them into a central generalist model. This design is motivated by the observation that RL excels at solving single tasks, while policy distillation -- a relatively stable supervised learning process -- is well aligned with large foundation models and multi-task learning. Moreover, a mixture-of-experts (MoE) architecture and a replay-based approach are employed to enhance the plasticity and stability of the continual policy distillation process. Extensive experiments on the Meta-World benchmark demonstrate that our framework enables efficient continual RL, recovering over 85% of teacher performance while constraining task-wise forgetting to within 10%.", "date": "2026-01-30", "categories": []}
{"id": "2601.22573", "title": "DELNet: Continuous All-in-One Weather Removal via Dynamic Expert Library", "abstract": "All-in-one weather image restoration methods are valuable in practice but depend on pre-collected data and require retraining for unseen degradations, leading to high cost. We propose DELNet, a continual learning framework for weather image restoration. DELNet integrates a judging valve that measures task similarity to distinguish new from known tasks, and a dynamic expert library that stores experts trained on different degradations. For new tasks, the valve selects top-k experts for knowledge transfer while adding new experts to capture task-specific features; for known tasks, the corresponding experts are directly reused. This design enables continuous optimization without retraining existing models. Experiments on OTS, Rain100H, and Snow100K demonstrate that DELNet surpasses state-of-the-art continual learning methods, achieving PSNR gains of 16\\%, 11\\%, and 12\\%, respectively. These results highlight the effectiveness, robustness, and efficiency of DELNet, which reduces retraining cost and enables practical deployment in real-world scenarios.", "date": "2026-01-30", "categories": []}
{"id": "2601.22828", "title": "Decomposing and Composing: Towards Efficient Vision-Language Continual Learning via Rank-1 Expert Pool in a Single LoRA", "abstract": "Continual learning (CL) in vision-language models (VLMs) faces significant challenges in improving task adaptation and avoiding catastrophic forgetting. Existing methods usually have heavy inference burden or rely on external knowledge, while Low-Rank Adaptation (LoRA) has shown potential in reducing these issues by enabling parameter-efficient tuning. However, considering directly using LoRA to alleviate the catastrophic forgetting problem is non-trivial, we introduce a novel framework that restructures a single LoRA module as a decomposable Rank-1 Expert Pool. Our method learns to dynamically compose a sparse, task-specific update by selecting from this expert pool, guided by the semantics of the [CLS] token. In addition, we propose an Activation-Guided Orthogonal (AGO) loss that orthogonalizes critical parts of LoRA weights across tasks. This sparse composition and orthogonalization enable fewer parameter updates, resulting in domain-aware learning while minimizing inter-task interference and maintaining downstream task performance. Extensive experiments across multiple settings demonstrate state-of-the-art results in all metrics, surpassing zero-shot upper bounds in generalization. Notably, it reduces trainable parameters by 96.7% compared to the baseline method, eliminating reliance on external datasets or task-ID discriminators. The merged LoRAs retain less weights and incur no inference latency, making our method computationally lightweight.", "date": "2026-01-30", "categories": []}
{"id": "2309.16117", "title": "Low-redundancy Distillation for Continual Learning", "abstract": "Continual learning (CL) aims to learn new tasks without erasing previous knowledge. However, current CL methods primarily emphasize improving accuracy while often neglecting training efficiency, which consequently restricts their practical application. Drawing inspiration from the brain's contextual gating mechanism, which selectively filters neural information and continuously updates past memories, we propose Low-redundancy Distillation (LoRD), a novel CL method that enhances model performance while maintaining training efficiency. This is achieved by eliminating redundancy in three aspects of CL: student model redundancy, teacher model redundancy, and rehearsal sample redundancy. By compressing the learnable parameters of the student model and pruning the teacher model, LoRD facilitates the retention and optimization of prior knowledge, effectively decoupling task-specific knowledge without manually assigning isolated parameters for each task. Furthermore, we optimize the selection of rehearsal samples and refine rehearsal frequency to improve training efficiency. Through a meticulous design of distillation and rehearsal strategies, LoRD effectively balances training efficiency and model precision. Extensive experimentation across various benchmark datasets and environments demonstrates LoRD's superiority, achieving the highest accuracy with the lowest training FLOPs.", "date": "2023-09-28", "categories": []}
{"id": "2601.20870", "title": "STAER: Temporal Aligned Rehearsal for Continual Spiking Neural Network", "abstract": "Spiking Neural Networks (SNNs) are inherently suited for continuous learning due to their event-driven temporal dynamics; however, their application to Class-Incremental Learning (CIL) has been hindered by catastrophic forgetting and the temporal misalignment of spike patterns. In this work, we introduce Spiking Temporal Alignment with Experience Replay (STAER), a novel framework that explicitly preserves temporal structure to bridge the performance gap between SNNs and ANNs. Our approach integrates a differentiable Soft-DTW alignment loss to maintain spike timing fidelity and employs a temporal expansion and contraction mechanism on output logits to enforce robust representation learning. Implemented on a deep ResNet19 spiking backbone, STAER achieves state-of-the-art performance on Sequential-MNIST and Sequential-CIFAR10. Empirical results demonstrate that our method matches or outperforms strong ANN baselines (ER, DER++) while preserving biologically plausible dynamics. Ablation studies further confirm that explicit temporal alignment is critical for representational stability, positioning STAER as a scalable solution for spike-native lifelong learning. Code is available at https://github.com/matteogianferrari/staer.", "date": "2026-01-16", "categories": []}
{"id": "2601.20894", "title": "Is Parameter Isolation Better for Prompt-Based Continual Learning?", "abstract": "Prompt-based continual learning methods effectively mitigate catastrophic forgetting. However, most existing methods assign a fixed set of prompts to each task, completely isolating knowledge across tasks and resulting in suboptimal parameter utilization. To address this, we consider the practical needs of continual learning and propose a prompt-sharing framework. This framework constructs a global prompt pool and introduces a task-aware gated routing mechanism that sparsely activates a subset of prompts to achieve dynamic decoupling and collaborative optimization of task-specific feature representations. Furthermore, we introduce a history-aware modulator that leverages cumulative prompt activation statistics to protect frequently used prompts from excessive updates, thereby mitigating inefficient parameter usage and knowledge forgetting. Extensive analysis and empirical results demonstrate that our approach consistently outperforms existing static allocation strategies in effectiveness and efficiency.", "date": "2026-01-28", "categories": []}
{"id": "2601.21341", "title": "Dynamical Adapter Fusion: Constructing A Global Adapter for Pre-Trained Model-based Class-Incremental Learning", "abstract": "Class-Incremental Learning (CIL) requires models to continuously acquire new classes without forgetting previously learned ones. A dominant paradigm involves freezing a pre-trained model and training lightweight, task-specific adapters. However, maintaining task-specific parameters hinders knowledge transfer and incurs high retrieval costs, while naive parameter fusion often leads to destructive interference and catastrophic forgetting. To address these challenges, we propose Dynamical Adapter Fusion (DAF) to construct a single robust global adapter. Grounded in the PAC-Bayes theorem, we derive a fusion mechanism that explicitly integrates three components: the optimized task-specific adapter parameters, the previous global adapter parameters, and the initialization parameters. We utilize the Taylor expansion of the loss function to derive the optimal fusion coefficients, dynamically achieving the best balance between stability and plasticity. Furthermore, we propose a Robust Initialization strategy to effectively capture global knowledge patterns. Experiments on multiple CIL benchmarks demonstrate that DAF achieves state-of-the-art (SOTA) performance.", "date": "2026-01-29", "categories": []}
{"id": "2508.08920", "title": "Exploring Cross-Stage Adversarial Transferability in Class-Incremental Continual Learning", "abstract": "Class-incremental continual learning addresses catastrophic forgetting by enabling classification models to preserve knowledge of previously learned classes while acquiring new ones. However, the vulnerability of the models against adversarial attacks during this process has not been investigated sufficiently. In this paper, we present the first exploration of vulnerability to stage-transferred attacks, i.e., an adversarial example generated using the model in an earlier stage is used to attack the model in a later stage. Our findings reveal that continual learning methods are highly susceptible to these attacks, raising a serious security issue. We explain this phenomenon through model similarity between stages and gradual robustness degradation. Additionally, we find that existing adversarial training-based defense methods are not sufficiently effective to stage-transferred attacks. Codes are available at https://github.com/mcml-official/CSAT.", "date": "2025-08-12", "categories": []}
{"id": "2601.20852", "title": "C3Box: A CLIP-based Class-Incremental Learning Toolbox", "abstract": "Traditional machine learning systems are typically designed for static data distributions, which suffer from catastrophic forgetting when learning from evolving data streams. Class-Incremental Learning (CIL) addresses this challenge by enabling learning systems to continuously learn new classes while preserving prior knowledge. With the rise of pre-trained models (PTMs) such as CLIP, leveraging their strong generalization and semantic alignment capabilities has become a promising direction in CIL. However, existing CLIP-based CIL methods are often scattered across disparate codebases, rely on inconsistent configurations, hindering fair comparisons, reproducibility, and practical adoption. Therefore, we propose C3Box (CLIP-based Class-inCremental learning toolBOX), a modular and comprehensive Python toolbox. C3Box integrates representative traditional CIL methods, ViT-based CIL methods, and state-of-the-art CLIP-based CIL methods into a unified CLIP-based framework. By inheriting the streamlined design of PyCIL, C3Box provides a JSON-based configuration and standardized execution pipeline. This design enables reproducible experimentation with low engineering overhead and makes C3Box a reliable benchmark platform for continual learning research. Designed to be user-friendly, C3Box relies only on widely used open-source libraries and supports major operating systems. The code is available at https://github.com/LAMDA-CL/C3Box.", "date": "2026-01-28", "categories": []}
{"id": "2501.12121", "title": "Learning Dynamic Representations via An Optimally-Weighted Maximum Mean Discrepancy Optimization Framework for Continual Learning", "abstract": "Continual learning has emerged as a pivotal area of research, primarily due to its advantageous characteristic that allows models to persistently acquire and retain information. However, catastrophic forgetting can severely impair model performance. In this study, we address network forgetting by introducing a novel framework termed Optimally-Weighted Maximum Mean Discrepancy (OWMMD), which imposes penalties on representation alterations via a Multi-Level Feature Matching Mechanism (MLFMM). Furthermore, we propose an Adaptive Regularization Optimization (ARO) strategy to refine the adaptive weight vectors, which autonomously assess the significance of each feature layer throughout the optimization process, The proposed ARO approach can relieve the over-regularization problem and promote the future task learning. We conduct a comprehensive series of experiments, benchmarking our proposed method against several established baselines. The empirical findings indicate that our approach achieves state-of-the-art performance.", "date": "2025-01-21", "categories": []}
{"id": "2507.14056", "title": "Noradrenergic-inspired gain modulation attenuates the stability gap in joint training", "abstract": "Recent work in continual learning has highlighted the stability gap -- a temporary performance drop on previously learned tasks when new ones are introduced. This phenomenon reflects a mismatch between rapid adaptation and strong retention at task boundaries, underscoring the need for optimization mechanisms that balance plasticity and stability over abrupt distribution changes. While optimizers such as momentum-SGD and Adam introduce implicit multi-timescale behavior, they still exhibit pronounced stability gaps. Importantly, these gaps persist even under ideal joint training, making it crucial to study them in this setting to isolate their causes from other sources of forgetting. Motivated by how noradrenergic (neuromodulatory) bursts transiently increase neuronal gain under uncertainty, we introduce a dynamic gain scaling mechanism as a two-timescale optimization technique that balances adaptation and retention by modulating effective learning rates and flattening the local landscape through an effective reparameterization. Across domain- and class-incremental MNIST, CIFAR, and mini-ImageNet benchmarks under task-agnostic joint training, dynamic gain scaling effectively attenuates stability gaps while maintaining competitive accuracy, improving robustness at task transitions.", "date": "2025-07-18", "categories": []}
{"id": "2509.23593", "title": "Avoid Catastrophic Forgetting with Rank-1 Fisher from Diffusion Models", "abstract": "Catastrophic forgetting remains a central obstacle for continual learning in neural models. Popular approaches -- replay and elastic weight consolidation (EWC) -- have limitations: replay requires a strong generator and is prone to distributional drift, while EWC implicitly assumes a shared optimum across tasks and typically uses a diagonal Fisher approximation. In this work, we study the gradient geometry of diffusion models, which can already produce high-quality replay data. We provide theoretical and empirical evidence that, in the low signal-to-noise ratio (SNR) regime, per-sample gradients become strongly collinear, yielding an empirical Fisher that is effectively rank-1 and aligned with the mean gradient. Leveraging this structure, we propose a rank-1 variant of EWC that is as cheap as the diagonal approximation yet captures the dominant curvature direction. We pair this penalty with a replay-based approach to encourage parameter sharing across tasks while mitigating drift. On class-incremental image generation datasets (MNIST, FashionMNIST, CIFAR-10, ImageNet-1k), our method consistently improves average FID and reduces forgetting relative to replay-only and diagonal-EWC baselines. In particular, forgetting is nearly eliminated on MNIST and FashionMNIST and is more than halved on ImageNet-1k. These results suggest that diffusion models admit an approximately rank-1 Fisher. With a better Fisher estimate, EWC becomes a strong complement to replay: replay encourages parameter sharing across tasks, while EWC effectively constrains replay-induced drift.", "date": "2025-09-28", "categories": []}
{"id": "2601.19659", "title": "KeepLoRA: Continual Learning with Residual Gradient Adaptation", "abstract": "Continual learning for pre-trained vision-language models requires balancing three competing objectives: retaining pre-trained knowledge, preserving knowledge from a sequence of learned tasks, and maintaining the plasticity to acquire new knowledge. This paper presents a simple but effective approach called KeepLoRA to effectively balance these objectives. We first analyze the knowledge retention mechanism within the model parameter space and find that general knowledge is mainly encoded in the principal subspace, while task-specific knowledge is encoded in the residual subspace. Motivated by this finding, KeepLoRA learns new tasks by restricting LoRA parameter updates in the residual subspace to prevent interfering with previously learned capabilities. Specifically, we infuse knowledge for a new task by projecting its gradient onto a subspace orthogonal to both the principal subspace of pre-trained model and the dominant directions of previous task features. Our theoretical and empirical analyses confirm that KeepLoRA balances the three objectives and achieves state-of-the-art performance. The implementation code is available at https://github.com/MaolinLuo/KeepLoRA.", "date": "2026-01-27", "categories": []}
{"id": "2601.19766", "title": "The Effect of Architecture During Continual Learning", "abstract": "Continual learning is a challenge for models with static architecture, as they fail to adapt to when data distributions evolve across tasks. We introduce a mathematical framework that jointly models architecture and weights in a Sobolev space, enabling a rigorous investigation into the role of neural network architecture in continual learning and its effect on the forgetting loss. We derive necessary conditions for the continual learning solution and prove that learning only model weights is insufficient to mitigate catastrophic forgetting under distribution shifts. Consequently, we prove that by learning the architecture and weights simultaneously at each task, we can reduce catastrophic forgetting.\n  To learn weights and architecture simultaneously, we formulate continual learning as a bilevel optimization problem: the upper level selects an optimal architecture for a given task, while the lower level computes optimal weights via dynamic programming over all tasks. To solve the upper level problem, we introduce a derivative-free direct search algorithm to determine the optimal architecture. Once found, we must transfer knowledge from the current architecture to the optimal one. However, the optimal architecture will result in a weights parameter space different from the current architecture (i.e., dimensions of weights matrices will not match). To bridge the dimensionality gap, we develop a low-rank transfer mechanism to map knowledge across architectures of mismatched dimensions. Empirical studies across regression and classification problems, including feedforward, convolutional, and graph neural networks, demonstrate that learning the optimal architecture and weights simultaneously yields substantially improved performance (up to two orders of magnitude), reduced forgetting, and enhanced robustness to noise compared with static architecture approaches.", "date": "2026-01-27", "categories": []}
{"id": "2601.19788", "title": "Knowledge-Aware Evolution for Streaming Federated Continual Learning with Category Overlap and without Task Identifiers", "abstract": "Federated Continual Learning (FCL) leverages inter-client collaboration to balance new knowledge acquisition and prior knowledge retention in non-stationary data. However, existing batch-based FCL methods lack adaptability to streaming scenarios featuring category overlap between old and new data and absent task identifiers, leading to indistinguishability of old and new knowledge, uncertain task assignments for samples, and knowledge confusion.To address this, we propose streaming federated continual learning setting: per federated learning (FL) round, clients process streaming data with disjoint samples and potentially overlapping categories without task identifiers, necessitating sustained inference capability for all prior categories after each FL round.Next, we introduce FedKACE: 1) an adaptive inference model switching mechanism that enables unidirectional switching from local model to global model to achieve a trade-off between personalization and generalization; 2) a adaptive gradient-balanced replay scheme that reconciles new knowledge learning and old knowledge retention under overlapping-class scenarios; 3) a kernel spectral boundary buffer maintenance that preserves high-information and high-boundary-influence samples to optimize cross-round knowledge retention. Experiments across multiple scenarios and regret analysis demonstrate the effectiveness of FedKACE.", "date": "2026-01-27", "categories": []}
{"id": "2601.19897", "title": "Self-Distillation Enables Continual Learning", "abstract": "Continual learning, enabling models to acquire new skills and knowledge without degrading existing capabilities, remains a fundamental challenge for foundation models. While on-policy reinforcement learning can reduce forgetting, it requires explicit reward functions that are often unavailable. Learning from expert demonstrations, the primary alternative, is dominated by supervised fine-tuning (SFT), which is inherently off-policy. We introduce Self-Distillation Fine-Tuning (SDFT), a simple method that enables on-policy learning directly from demonstrations. SDFT leverages in-context learning by using a demonstration-conditioned model as its own teacher, generating on-policy training signals that preserve prior capabilities while acquiring new skills. Across skill learning and knowledge acquisition tasks, SDFT consistently outperforms SFT, achieving higher new-task accuracy while substantially reducing catastrophic forgetting. In sequential learning experiments, SDFT enables a single model to accumulate multiple skills over time without performance regression, establishing on-policy distillation as a practical path to continual learning from demonstrations.", "date": "2026-01-27", "categories": []}
{"id": "2504.00174", "title": "MetaCLBench: Meta Continual Learning Benchmark on Resource-Constrained Edge Devices", "abstract": "Meta-Continual Learning (Meta-CL) enables models to learn new classes from limited labelled samples, making it promising for IoT applications where manual labelling is costly. However, existing studies focus on accuracy while ignoring deployment viability on resource-constrained hardware. Thus, we present MetaCLBench, a benchmark framework that evaluates Meta-CL methods for both accuracy and deployment-critical metrics (memory footprint, latency, and energy consumption) on real IoT devices with RAM sizes ranging from 512 MB to 4 GB. We evaluate six Meta-CL methods across three architectures (CNN, YAMNet, ViT) and five datasets spanning image and audio modalities. Our evaluation reveals that, depending on the dataset, up to three of six methods cause out-of-memory failures on sub-1 GB devices, significantly narrowing viable deployment options. LifeLearner achieves near-oracle accuracy while consuming 2.54-7.43x less energy than the Oracle method. Notably, larger or more sophisticated architectures such as ViT and YAMNet do not necessarily yield better Meta-CL performance, with results varying across datasets and modalities, challenging conventional assumptions about model complexity. Finally, we provide practical deployment guidelines and will release our framework upon publication to enable fair evaluation across both accuracy and system-level metrics.", "date": "2025-03-31", "categories": []}
{"id": "2505.11998", "title": "Parameter Efficient Continual Learning with Dynamic Low-Rank Adaptation", "abstract": "Catastrophic forgetting has remained a critical challenge for deep neural networks in Continual Learning (CL) as it undermines consolidated knowledge when learning new tasks. Parameter efficient fine tuning CL techniques are gaining traction for their effectiveness in addressing catastrophic forgetting with a lightweight training schedule while avoiding degradation of consolidated knowledge in pre-trained models. However, low rank adapters (LoRA) in these approaches are highly sensitive to rank selection which can lead to sub-optimal resource allocation and performance. To this end, we introduce PEARL, a rehearsal-free CL framework that entails dynamic rank allocation for LoRA components during CL training. Specifically, PEARL leverages reference task weights and adaptively determines the rank of task-specific LoRA components based on the current tasks' proximity to reference task weights in parameter space. To demonstrate the versatility of PEARL, we evaluate it across three vision architectures (ResNet, Separable Convolutional Network and Vision Transformer) and a multitude of CL scenarios, and show that PEARL outperforms all considered baselines by a large margin.", "date": "2025-05-17", "categories": []}
{"id": "2505.14125", "title": "Contrastive Consolidation of Top-Down Modulations Achieves Sparsely Supervised Continual Learning", "abstract": "Biological brains learn continually from a stream of unlabeled data, while integrating specialized information from sparsely labeled examples without compromising their ability to generalize. Meanwhile, machine learning methods are susceptible to catastrophic forgetting in this natural learning setting, as supervised specialist fine-tuning degrades performance on the original task. We introduce task-modulated contrastive learning (TMCL), which takes inspiration from the biophysical machinery in the neocortex, using predictive coding principles to integrate top-down information continually and without supervision. We follow the idea that these principles build a view-invariant representation space, and that this can be implemented using a contrastive loss. Then, whenever labeled samples of a new class occur, new affine modulations are learned that improve separation of the new class from all others, without affecting feedforward weights. By co-opting the view-invariance learning mechanism, we then train feedforward weights to match the unmodulated representation of a data sample to its modulated counterparts. This introduces modulation invariance into the representation space, and, by also using past modulations, stabilizes it. Our experiments show improvements in both class-incremental and transfer learning over state-of-the-art unsupervised approaches, as well as over comparable supervised approaches, using as few as 1% of available labels. Taken together, our work suggests that top-down modulations play a crucial role in balancing stability and plasticity.", "date": "2025-05-20", "categories": []}
{"id": "2505.17902", "title": "Evolving Machine Learning in Non-Stationary Environments: A Unified Survey of Drift, Forgetting, and Adaptation", "abstract": "In an era defined by rapid data evolution, traditional Machine Learning (ML) models often struggle to adapt to dynamic environments. Evolving Machine Learning (EML) has emerged as a pivotal paradigm, enabling continuous learning and real-time adaptation to streaming data. While prior surveys have examined individual components of evolving learning - such as drift detection - there remains a lack of a unified analysis of its major challenges. This survey provides a comprehensive overview of EML, focusing on four core challenges: data drift, concept drift, catastrophic forgetting, and skewed learning. We systematically review over 100 studies, categorizing state-of-the-art methods across supervised, unsupervised, and semi-supervised learning. The survey further explores evaluation metrics, benchmark datasets, and real-world applications, offering a comparative perspective on the effectiveness and limitations of current approaches and proposing a taxonomy to organize them. In addition, we highlight the growing role of adaptive neural architectures, meta-learning, and ensemble strategies in managing evolving data complexities. By synthesizing insights from recent literature, this work not only maps the current landscape of EML but also identifies key research gaps and emerging opportunities. Our findings aim to guide researchers and practitioners in developing robust, ethical, and scalable EML systems for real-world deployment.", "date": "2025-05-23", "categories": []}
{"id": "2506.01327", "title": "Enhancing Federated Class-Incremental Learning via Spatial-Temporal Statistics Aggregation", "abstract": "Federated Class-Incremental Learning (FCIL) enables Class-Incremental Learning (CIL) from distributed data. Existing FCIL methods typically integrate old knowledge preservation into local client training. However, these methods cannot avoid spatial-temporal client drift caused by data heterogeneity and often incur significant computational and communication overhead, limiting practical deployment. To address these challenges simultaneously, we propose a novel approach, Spatial-Temporal Statistics Aggregation (STSA), which provides a unified framework to aggregate feature statistics both spatially (across clients) and temporally (across stages). The aggregated feature statistics are unaffected by data heterogeneity and can be used to update the classifier in closed form at each stage. Additionally, we introduce STSA-E, a communication-efficient variant with theoretical guarantees, achieving similar performance to STSA-E with much lower communication overhead. Extensive experiments on three widely used FCIL datasets, with varying degrees of data heterogeneity, show that our method outperforms state-of-the-art FCIL methods in terms of performance, flexibility, and both communication and computation efficiency. The code is available at https://github.com/Yuqin-G/STSA.", "date": "2025-06-02", "categories": []}
{"id": "2508.19005", "title": "Building Self-Evolving Agents via Experience-Driven Lifelong Learning: A Framework and Benchmark", "abstract": "As AI advances toward general intelligence, the focus is shifting from systems optimized for static tasks to creating open-ended agents that learn continuously. In this paper, we introduce Experience-driven Lifelong Learning (ELL), a framework for building self-evolving agents capable of continuous growth through real-world interaction. The framework is built on four core principles: (1) Experience Exploration: Agents learn through continuous, self-motivated interaction with dynamic environments, navigating interdependent tasks and generating rich experiential trajectories. (2) Long-term Memory: Agents preserve and structure historical knowledge, including personal experiences, domain expertise, and commonsense reasoning, into a persistent memory system. (3) Skill Learning: Agents autonomously improve by abstracting recurring patterns from experience into reusable skills, which are actively refined and validated for application in new tasks. (4) Knowledge Internalization: Agents internalize explicit and discrete experiences into implicit and intuitive capabilities as \"second nature\".\n  We also introduce StuLife, a benchmark dataset for ELL that simulates a student's holistic college journey, from enrollment to academic and personal development, across three core phases and ten detailed sub-scenarios. StuLife is designed around three key paradigm", "date": "2025-08-26", "categories": []}
{"id": "2601.12816", "title": "Fisher-Orthogonal Projected Natural Gradient Descent for Continual Learning", "abstract": "Continual learning aims to enable neural networks to acquire new knowledge on sequential tasks. However, the key challenge in such settings is to learn new tasks without catastrophically forgetting previously learned tasks. We propose the Fisher-Orthogonal Projected Natural Gradient Descent (FOPNG) optimizer, which enforces Fisher-orthogonal constraints on parameter updates to preserve old task performance while learning new tasks. Unlike existing methods that operate in Euclidean parameter space, FOPNG projects gradients onto the Fisher-orthogonal complement of previous task gradients. This approach unifies natural gradient descent with orthogonal gradient methods within an information-geometric framework. We provide theoretical analysis deriving the projected update, describe efficient and practical implementations using the diagonal Fisher, and demonstrate strong results on standard continual learning benchmarks such as Permuted-MNIST, Split-MNIST, Rotated-MNIST, Split-CIFAR10, and Split-CIFAR100. Our code is available at https://github.com/ishirgarg/FOPNG.", "date": "2026-01-19", "categories": []}
{"id": "2601.17076", "title": "E2PL: Effective and Efficient Prompt Learning for Incomplete Multi-view Multi-Label Class Incremental Learning", "abstract": "Multi-view multi-label classification (MvMLC) is indispensable for modern web applications aggregating information from diverse sources. However, real-world web-scale settings are rife with missing views and continuously emerging classes, which pose significant obstacles to robust learning. Prevailing methods are ill-equipped for this reality, as they either lack adaptability to new classes or incur exponential parameter growth when handling all possible missing-view patterns, severely limiting their scalability in web environments. To systematically address this gap, we formally introduce a novel task, termed \\emph{incomplete multi-view multi-label class incremental learning} (IMvMLCIL), which requires models to simultaneously address heterogeneous missing views and dynamic class expansion. To tackle this task, we propose \\textsf{E2PL}, an Effective and Efficient Prompt Learning framework for IMvMLCIL. \\textsf{E2PL} unifies two novel prompt designs: \\emph{task-tailored prompts} for class-incremental adaptation and \\emph{missing-aware prompts} for the flexible integration of arbitrary view-missing scenarios. To fundamentally address the exponential parameter explosion inherent in missing-aware prompts, we devise an \\emph{efficient prototype tensorization} module, which leverages atomic tensor decomposition to elegantly reduce the prompt parameter complexity from exponential to linear w.r.t. the number of views. We further incorporate a \\emph{dynamic contrastive learning} strategy explicitly model the complex dependencies among diverse missing-view patterns, thus enhancing the model's robustness. Extensive experiments on three benchmarks demonstrate that \\textsf{E2PL} consistently outperforms state-of-the-art methods in both effectiveness and efficiency. The codes and datasets are available at https://anonymous.4open.science/r/code-for-E2PL.", "date": "2026-01-23", "categories": []}
{"id": "2601.17905", "title": "Feature-Space Generative Models for One-Shot Class-Incremental Learning", "abstract": "Few-shot class-incremental learning (FSCIL) is a paradigm where a model, initially trained on a dataset of base classes, must adapt to an expanding problem space by recognizing novel classes with limited data. We focus on the challenging FSCIL setup where a model receives only a single sample (1-shot) for each novel class and no further training or model alterations are allowed after the base training phase. This makes generalization to novel classes particularly difficult. We propose a novel approach predicated on the hypothesis that base and novel class embeddings have structural similarity. We map the original embedding space into a residual space by subtracting the class prototype (i.e., the average class embedding) of input samples. Then, we leverage generative modeling with VAE or diffusion models to learn the multi-modal distribution of residuals over the base classes, and we use this as a valuable structural prior to improve recognition of novel classes. Our approach, Gen1S, consistently improves novel class recognition over the state of the art across multiple benchmarks and backbone architectures.", "date": "2026-01-25", "categories": []}
{"id": "2601.18255", "title": "Beyond Retention: Orchestrating Structural Safety and Plasticity in Continual Learning for LLMs", "abstract": "Continual learning in Large Language Models (LLMs) faces the critical challenge of balancing stability (retaining old knowledge) and plasticity (learning new tasks). While Experience Replay (ER) is a standard countermeasure against catastrophic forgetting, its impact across diverse capabilities remains underexplored. In this work, we uncover a critical dichotomy in ER's behavior: while it induces positive backward transfer on robust, unstructured tasks (e.g., boosting performance on previous NLP classification tasks through repeated rehearsal), it causes severe negative transfer on fragile, structured domains like code generation (e.g., a significant relative drop in coding accuracy). This reveals that ER trades structural integrity for broad consolidation. To address this dilemma, we propose \\textbf{Orthogonal Subspace Wake-up (OSW)}. OSW identifies essential parameter subspaces of previous tasks via a brief \"wake-up\" phase and enforces orthogonal updates for new tasks, providing a mathematically grounded \"safety guarantee\" for established knowledge structures. Empirical results across a diverse four-task sequence demonstrate that OSW uniquely succeeds in preserving fragile coding abilities where Replay fails, while simultaneously maintaining high plasticity for novel tasks. Our findings emphasize the necessity of evaluating structural safety alongside average retention in LLM continual learning.", "date": "2026-01-26", "categories": []}
{"id": "2601.18261", "title": "FGGM: Fisher-Guided Gradient Masking for Continual Learning", "abstract": "Catastrophic forgetting impairs the continuous learning of large language models. We propose Fisher-Guided Gradient Masking (FGGM), a framework that mitigates this by strategically selecting parameters for updates using diagonal Fisher Information. FGGM dynamically generates binary masks with adaptive thresholds, preserving critical parameters to balance stability and plasticity without requiring historical data. Unlike magnitude-based methods such as MIGU, our approach offers a mathematically principled parameter importance estimation. On the TRACE benchmark, FGGM shows a 9.6% relative improvement in retaining general capabilities over supervised fine-tuning (SFT) and a 4.4% improvement over MIGU on TRACE tasks. Additional analysis on code generation tasks confirms FGGM's superior performance and reduced forgetting, establishing it as an effective solution.", "date": "2026-01-26", "categories": []}
{"id": "2509.06100", "title": "Orthogonal Low-rank Adaptation in Lie Groups for Continual Learning of Large Language Models", "abstract": "Large language models (LLMs) suffer from catastrophic forgetting in sequential multi-task learning. Existing parameter regularization methods (e.g., O-LoRA, N-LoRA) mitigate interference via low-rank subspace orthogonality, but additive updates distort the intrinsic geometry of model parameters. We propose \\textbf{OLieRA}, a Lie group based fine-tuning framework that preserves parameter geometry through multiplicative updates while enforcing orthogonality across task subspaces. OLieRA achieves state-of-the-art performance on the Standard CL benchmark and remains highly competitive under large task sequences. It further inherits the replay-free and task-ID free inference properties of O-LoRA, establishing a principled paradigm for continual learning in LLMs.", "date": "2025-09-07", "categories": []}
{"id": "2601.16773", "title": "CASP: Few-Shot Class-Incremental Learning with CLS Token Attention Steering Prompts", "abstract": "Few-shot class-incremental learning (FSCIL) presents a core challenge in continual learning, requiring models to rapidly adapt to new classes with very limited samples while mitigating catastrophic forgetting. Recent prompt-based methods, which integrate pretrained backbones with task-specific prompts, have made notable progress. However, under extreme few-shot incremental settings, the model's ability to transfer and generalize becomes critical, and it is thus essential to leverage pretrained knowledge to learn feature representations that can be shared across future categories during the base session. Inspired by the mechanism of the CLS token, which is similar to human attention and progressively filters out task-irrelevant information, we propose the CLS Token Attention Steering Prompts (CASP). This approach introduces class-shared trainable bias parameters into the query, key, and value projections of the CLS token to explicitly modulate the self-attention weights. To further enhance generalization, we also design an attention perturbation strategy and perform Manifold Token Mixup in the shallow feature space, synthesizing potential new class features to improve generalization and reserve the representation capacity for upcoming tasks. Experiments on the CUB200, CIFAR100, and ImageNet-R datasets demonstrate that CASP outperforms state-of-the-art methods in both standard and fine-grained FSCIL settings without requiring fine-tuning during incremental phases and while significantly reducing the parameter overhead.", "date": "2026-01-23", "categories": []}
{"id": "2601.15643", "title": "Evolving Without Ending: Unifying Multimodal Incremental Learning for Continual Panoptic Perception", "abstract": "Continual learning (CL) is a great endeavour in developing intelligent perception AI systems. However, the pioneer research has predominantly focus on single-task CL, which restricts the potential in multi-task and multimodal scenarios. Beyond the well-known issue of catastrophic forgetting, the multi-task CL also brings semantic obfuscation across multimodal alignment, leading to severe model degradation during incremental training steps. In this paper, we extend CL to continual panoptic perception (CPP), integrating multimodal and multi-task CL to enhance comprehensive image perception through pixel-level, instance-level, and image-level joint interpretation. We formalize the CL task in multimodal scenarios and propose an end-to-end continual panoptic perception model. Concretely, CPP model features a collaborative cross-modal encoder (CCE) for multimodal embedding. We also propose a malleable knowledge inheritance module via contrastive feature distillation and instance distillation, addressing catastrophic forgetting from task-interactive boosting manner. Furthermore, we propose a cross-modal consistency constraint and develop CPP+, ensuring multimodal semantic alignment for model updating under multi-task incremental scenarios. Additionally, our proposed model incorporates an asymmetric pseudo-labeling manner, enabling model evolving without exemplar replay. Extensive experiments on multimodal datasets and diverse CL tasks demonstrate the superiority of the proposed model, particularly in fine-grained CL tasks.", "date": "2026-01-22", "categories": []}
{"id": "2507.05386", "title": "Reinforcement Fine-Tuning Naturally Mitigates Forgetting in Continual Post-Training", "abstract": "Continual post-training (CPT) is a popular and effective technique for adapting foundation models like multimodal large language models to specific and ever-evolving downstream tasks. While existing research has primarily concentrated on methods like data replay, model expansion, or parameter regularization, the fundamental role of the learning paradigm within CPT remains largely unexplored. This paper presents a comparative analysis of two core post-training paradigms: supervised fine-tuning (SFT) and reinforcement fine-tuning (RFT), investigating their respective impacts on knowledge retention during CPT. Our experiments are conducted on a benchmark comprising seven diverse multimodal tasks, utilizing Qwen2.5-VL-7B-Instruct as the base model for continual post-training. The investigation yields two significant findings: (1) When continuously learning on downstream tasks, SFT leads to catastrophic forgetting of previously learned tasks. In contrast, RFT inherently preserves prior knowledge and achieve performance comparable to multi-task training. (2) RFT successfully protects and even enhances the model's general knowledge on standard benchmarks (e.g., MMMU and MMLU-Pro). Conversely, SFT degrades general model capabilities severely. Further analysis reveals that this stability is not primarily due to explicit mechanisms like KL penalty or chain-of-thought reasoning. Instead, we identify an implicit regularization mechanism inherent to RFT as a key contributing factor. Our theoretical analysis suggests that RFT's gradient updates are naturally scaled by the reward variance, acting as a data-dependent regularizer that inherently protects previously acquired knowledge. Finally, we propose a rollout-based instance filtering algorithm to enhance the stability and efficiency of RFT. Our comprehensive study demonstrates the superiority of RFT as a robust paradigm for continual post-training.", "date": "2025-07-07", "categories": []}
{"id": "2510.19314", "title": "Continual Knowledge Adaptation for Reinforcement Learning", "abstract": "Reinforcement Learning enables agents to learn optimal behaviors through interactions with environments. However, real-world environments are typically non-stationary, requiring agents to continuously adapt to new tasks and changing conditions. Although Continual Reinforcement Learning facilitates learning across multiple tasks, existing methods often suffer from catastrophic forgetting and inefficient knowledge utilization. To address these challenges, we propose Continual Knowledge Adaptation for Reinforcement Learning (CKA-RL), which enables the accumulation and effective utilization of historical knowledge. Specifically, we introduce a Continual Knowledge Adaptation strategy, which involves maintaining a task-specific knowledge vector pool and dynamically using historical knowledge to adapt the agent to new tasks. This process mitigates catastrophic forgetting and enables efficient knowledge transfer across tasks by preserving and adapting critical model parameters. Additionally, we propose an Adaptive Knowledge Merging mechanism that combines similar knowledge vectors to address scalability challenges, reducing memory requirements while ensuring the retention of essential knowledge. Experiments on three benchmarks demonstrate that the proposed CKA-RL outperforms state-of-the-art methods, achieving an improvement of 4.20% in overall performance and 8.02% in forward transfer. The source code is available at https://github.com/Fhujinwu/CKA-RL.", "date": "2025-10-22", "categories": []}
{"id": "2511.09853", "title": "ConSurv: Multimodal Continual Learning for Survival Analysis", "abstract": "Survival prediction of cancers is crucial for clinical practice, as it informs mortality risks and influences treatment plans. However, a static model trained on a single dataset fails to adapt to the dynamically evolving clinical environment and continuous data streams, limiting its practical utility. While continual learning (CL) offers a solution to learn dynamically from new datasets, existing CL methods primarily focus on unimodal inputs and suffer from severe catastrophic forgetting in survival prediction. In real-world scenarios, multimodal inputs often provide comprehensive and complementary information, such as whole slide images and genomics; and neglecting inter-modal correlations negatively impacts the performance. To address the two challenges of catastrophic forgetting and complex inter-modal interactions between gigapixel whole slide images and genomics, we propose ConSurv, the first multimodal continual learning (MMCL) method for survival analysis. ConSurv incorporates two key components: Multi-staged Mixture of Experts (MS-MoE) and Feature Constrained Replay (FCR). MS-MoE captures both task-shared and task-specific knowledge at different learning stages of the network, including two modality encoders and the modality fusion component, learning inter-modal relationships. FCR further enhances learned knowledge and mitigates forgetting by restricting feature deviation of previous data at different levels, including encoder-level features of two modalities and the fusion-level representations. Additionally, we introduce a new benchmark integrating four datasets, Multimodal Survival Analysis Incremental Learning (MSAIL), for comprehensive evaluation in the CL setting. Extensive experiments demonstrate that ConSurv outperforms competing methods across multiple metrics.", "date": "2025-11-13", "categories": []}
{"id": "2601.12714", "title": "P2L-CA: An Effective Parameter Tuning Framework for Rehearsal-Free Multi-Label Class-Incremental Learning", "abstract": "Multi-label Class-Incremental Learning aims to continuously recognize novel categories in complex scenes where multiple objects co-occur. However, existing approaches often incur high computational costs due to full-parameter fine-tuning and substantial storage overhead from memory buffers, or they struggle to address feature confusion and domain discrepancies adequately. To overcome these limitations, we introduce P2L-CA, a parameter-efficient framework that integrates a Prompt-to-Label module with a Continuous Adapter module. The P2L module leverages class-specific prompts to disentangle multi-label representations while incorporating linguistic priors to enforce stable semantic-visual alignment. Meanwhile, the CA module employs lightweight adapters to mitigate domain gaps between pre-trained models and downstream tasks, thereby enhancing model plasticity. Extensive experiments across standard and challenging MLCIL settings on MS-COCO and PASCAL VOC show that P2L-CA not only achieves substantial improvements over state-of-the-art methods but also demonstrates strong generalization in CIL scenarios, all while requiring minimal trainable parameters and eliminating the need for memory buffers.", "date": "2026-01-19", "categories": []}
{"id": "2601.12931", "title": "Online Continual Learning for Time Series: a Natural Score-driven Approach", "abstract": "Online continual learning (OCL) methods adapt to changing environments without forgetting past knowledge. Similarly, online time series forecasting (OTSF) is a real-world problem where data evolve in time and success depends on both rapid adaptation and long-term memory. Indeed, time-varying and regime-switching forecasting models have been extensively studied, offering a strong justification for the use of OCL in these settings. Building on recent work that applies OCL to OTSF, this paper aims to strengthen the theoretical and practical connections between time series methods and OCL. First, we reframe neural network optimization as a parameter filtering problem, showing that natural gradient descent is a score-driven method and proving its information-theoretic optimality. Then, we show that using a Student's t likelihood in addition to natural gradient induces a bounded update, which improves robustness to outliers. Finally, we introduce Natural Score-driven Replay (NatSR), which combines our robust optimizer with a replay buffer and a dynamic scale heuristic that improves fast adaptation at regime drifts. Empirical results demonstrate that NatSR achieves stronger forecasting performance than more complex state-of-the-art methods.", "date": "2026-01-19", "categories": []}
{"id": "2601.13020", "title": "PASs-MoE: Mitigating Misaligned Co-drift among Router and Experts via Pathway Activation Subspaces for Continual Learning", "abstract": "Continual instruction tuning (CIT) requires multimodal large language models (MLLMs) to adapt to a stream of tasks without forgetting prior capabilities. A common strategy is to isolate updates by routing inputs to different LoRA experts. However, existing LoRA-based Mixture-of-Experts (MoE) methods often jointly update the router and experts in an indiscriminate way, causing the router's preferences to co-drift with experts' adaptation pathways and gradually deviate from early-stage input-expert specialization. We term this phenomenon Misaligned Co-drift, which blurs expert responsibilities and exacerbates forgetting.To address this, we introduce the pathway activation subspace (PASs), a LoRA-induced subspace that reflects which low-rank pathway directions an input activates in each expert, providing a capability-aligned coordinate system for routing and preservation. Based on PASs, we propose a fixed-capacity PASs-based MoE-LoRA method with two components: PAS-guided Reweighting, which calibrates routing using each expert's pathway activation signals, and PAS-aware Rank Stabilization, which selectively stabilizes rank directions important to previous tasks. Experiments on a CIT benchmark show that our approach consistently outperforms a range of conventional continual learning baselines and MoE-LoRA variants in both accuracy and anti-forgetting without adding parameters. Our code will be released upon acceptance.", "date": "2026-01-19", "categories": []}
{"id": "2601.14069", "title": "Unsupervised Video Class-Incremental Learning via Deep Embedded Clustering Management", "abstract": "Unsupervised video class incremental learning (uVCIL) represents an important learning paradigm for learning video information without forgetting, and without considering any data labels. Prior approaches have focused on supervised class-incremental learning, relying on using the knowledge of labels and task boundaries, which is costly, requires human annotation, or is simply not a realistic option. In this paper, we propose a simple yet effective approach to address the uVCIL. We first consider a deep feature extractor network, providing a set of representative video features during each task without assuming any class or task information. We then progressively build a series of deep clusters from the extracted features. During the successive task learning, the model updated from the previous task is used as an initial state in order to transfer knowledge to the current learning task. We perform in-depth evaluations on three standard video action recognition datasets, including UCF101, HMDB51, and Something-to-Something V2, by ignoring the labels from the supervised setting. Our approach significantly outperforms other baselines on all datasets.", "date": "2026-01-20", "categories": []}
{"id": "2501.07555", "title": "Dynamic Prototype Rehearsal for Continual ECG Arrhythmia Detection", "abstract": "Continual Learning (CL) methods aim to learn from a sequence of tasks while avoiding the challenge of forgetting previous knowledge. We present DREAM-CL, a novel CL method for ECG arrhythmia detection that introduces dynamic prototype rehearsal memory. DREAM-CL selects representative prototypes by clustering data based on learning behavior during each training session. Within each cluster, we apply a smooth sorting operation that ranks samples by training difficulty, compressing extreme values and removing outliers. The more challenging samples are then chosen as prototypes for the rehearsal memory, ensuring effective knowledge retention across sessions. We evaluate our method on time-incremental, class-incremental, and lead-incremental scenarios using two widely used ECG arrhythmia datasets, Chapman and PTB-XL. The results demonstrate that DREAM-CL outperforms the state-of-the-art in CL for ECG arrhythmia detection. Detailed ablation and sensitivity studies are performed to validate the different design choices of our method.", "date": "2025-01-13", "categories": []}
{"id": "2505.24149", "title": "RCCDA: Adaptive Model Updates in the Presence of Concept Drift under a Constrained Resource Budget", "abstract": "Machine learning (ML) algorithms deployed in real-world environments are often faced with the challenge of adapting models to concept drift, where the task data distributions are shifting over time. The problem becomes even more difficult when model performance must be maintained under adherence to strict resource constraints. Existing solutions often depend on drift-detection methods that produce high computational overhead for resource-constrained environments, and fail to provide strict guarantees on resource usage or theoretical performance assurances. To address these shortcomings, we propose RCCDA: a dynamic model update policy that optimizes ML training dynamics while ensuring compliance to predefined resource constraints, utilizing only past loss information and a tunable drift threshold. In developing our policy, we analytically characterize the evolution of model loss under concept drift with arbitrary training update decisions. Integrating these results into a Lyapunov drift-plus-penalty framework produces a lightweight greedy-optimal policy that provably limits update frequency and cost. Experimental results on four domain generalization datasets demonstrate that our policy outperforms baseline methods in inference accuracy while adhering to strict resource constraints under several schedules of concept drift, making our solution uniquely suited for real-time ML deployments.", "date": "2025-05-30", "categories": []}
{"id": "2511.20696", "title": "Prototype-Guided Non-Exemplar Continual Learning for Cross-subject EEG Decoding", "abstract": "Due to the significant variability in electroencephalo-gram (EEG) signals across individuals, knowledge acquired from previous subjects is often overwritten as new subjects are introduced in continual EEG decoding tasks. Existing methods mainly rely on storing historical data from seen subjects as replay buffers to mitigate forgetting, which is impractical under privacy or memory constraints. To address this issue, we propose a Prototype-guided Non-Exemplar Continual Learning (ProNECL) framework that preserves prior knowledge without accessing historical EEG samples. ProNECL summarizes subject-specific discriminative representations into class-level prototypes and incrementally aligns new subject representations with a global prototype memory through prototype-based feature regulariza-tion and cross-subject alignment. Experiments on the BCI Com-petition IV 2a and 2b datasets demonstrate that ProNECL effec-tively balances knowledge retention and adaptability, achieving superior performance in cross-subject continual EEG decoding tasks.", "date": "2025-11-24", "categories": []}
{"id": "2601.09974", "title": "SPRInG: Continual LLM Personalization via Selective Parametric Adaptation and Retrieval-Interpolated Generation", "abstract": "Personalizing Large Language Models typically relies on static retrieval or one-time adaptation, assuming user preferences remain invariant over time. However, real-world interactions are dynamic, where user interests continuously evolve, posing a challenge for models to adapt to preference drift without catastrophic forgetting. Standard continual learning approaches often struggle in this context, as they indiscriminately update on noisy interaction streams, failing to distinguish genuine preference shifts from transient contexts. To address this, we introduce SPRInG, a novel semi-parametric framework designed for effective continual personalization. During training, SPRInG employs drift-driven selective adaptation, which utilizes a likelihood-based scoring function to identify high-novelty interactions. This allows the model to selectively update the user-specific adapter on drift signals while preserving hard-to-learn residuals in a replay buffer. During inference, we apply strict relevance gating and fuse parametric knowledge with retrieved history via logit interpolation. Experiments on the long-form personalized generation benchmark demonstrate that SPRInG outperforms existing baselines, validating its robustness for real-world continual personalization.", "date": "2026-01-15", "categories": []}
{"id": "2502.11609", "title": "Exploiting Task Relationships in Continual Learning via Transferability-Aware Task Embeddings", "abstract": "Continual learning (CL) has been a critical topic in contemporary deep neural network applications, where higher levels of both forward and backward transfer are desirable for an effective CL performance. Existing CL strategies primarily focus on task models, either by regularizing model updates or by separating task-specific and shared components, while often overlooking the potential of leveraging inter-task relationships to enhance transfer. To address this gap, we propose a transferability-aware task embedding, termed H-embedding, and construct a hypernet framework under its guidance to learn task-conditioned model weights for CL tasks. Specifically, H-embedding is derived from an information theoretic measure of transferability and is designed to be online and easy to compute. Our method is also characterized by notable practicality, requiring only the storage of a low-dimensional task embedding per task and supporting efficient end-to-end training. Extensive evaluations on benchmarks including CIFAR-100, ImageNet-R, and DomainNet show that our framework performs prominently compared to various baseline and SOTA approaches, demonstrating strong potential in capturing and utilizing intrinsic task relationships. Our code is publicly available at https://github.com/viki760/Hembedding_Guided_Hypernet.", "date": "2025-02-17", "categories": []}
{"id": "2508.05065", "title": "Decoupling Continual Semantic Segmentation", "abstract": "Continual Semantic Segmentation (CSS) requires learning new classes without forgetting previously acquired knowledge, addressing the fundamental challenge of catastrophic forgetting in dense prediction tasks. However, existing CSS methods typically employ single-stage encoder-decoder architectures where segmentation masks and class labels are tightly coupled, leading to interference between old and new class learning and suboptimal retention-plasticity balance. We introduce DecoupleCSS, a novel two-stage framework for CSS. By decoupling class-aware detection from class-agnostic segmentation, DecoupleCSS enables more effective continual learning, preserving past knowledge while learning new classes. The first stage leverages pre-trained text and image encoders, adapted using LoRA, to encode class-specific information and generate location-aware prompts. In the second stage, the Segment Anything Model (SAM) is employed to produce precise segmentation masks, ensuring that segmentation knowledge is shared across both new and previous classes. This approach improves the balance between retention and adaptability in CSS, achieving state-of-the-art performance across a variety of challenging tasks. Our code is publicly available at: https://github.com/euyis1019/Decoupling-Continual-Semantic-Segmentation.", "date": "2025-08-07", "categories": []}
{"id": "2601.07858", "title": "Affect and Effect: Limitations of regularisation-based continual learning in EEG-based emotion classification", "abstract": "Generalisation to unseen subjects in EEG-based emotion classification remains a challenge due to high inter-and intra-subject variability. Continual learning (CL) poses a promising solution by learning from a sequence of tasks while mitigating catastrophic forgetting. Regularisation-based CL approaches, such as Elastic Weight Consolidation (EWC), Synaptic Intelligence (SI), and Memory Aware Synapses (MAS), are commonly used as baselines in EEG-based CL studies, yet their suitability for this problem remains underexplored. This study theoretically and empirically finds that regularisation-based CL methods show limited performance for EEG-based emotion classification on the DREAMER and SEED datasets. We identify a fundamental misalignment in the stability-plasticity trade-off, where regularisation-based methods prioritise mitigating catastrophic forgetting (backward transfer) over adapting to new subjects (forward transfer). We investigate this limitation under subject-incremental sequences and observe that: (1) the heuristics for estimating parameter importance become less reliable under noisy data and covariate shift, (2) gradients on parameters deemed important by these heuristics often interfere with gradient updates required for new subjects, moving optimisation away from the minimum, (3) importance values accumulated across tasks over-constrain the model, and (4) performance is sensitive to subject order. Forward transfer showed no statistically significant improvement over sequential fine-tuning (p > 0.05 across approaches and datasets). The high variability of EEG signals means past subjects provide limited value to future subjects. Regularisation-based continual learning approaches are therefore limited for robust generalisation to unseen subjects in EEG-based emotion classification.", "date": "2026-01-09", "categories": []}
{"id": "2601.08448", "title": "Divide and Conquer: Static-Dynamic Collaboration for Few-Shot Class-Incremental Learning", "abstract": "Few-shot class-incremental learning (FSCIL) aims to continuously recognize novel classes under limited data, which suffers from the key stability-plasticity dilemma: balancing the retention of old knowledge with the acquisition of new knowledge. To address this issue, we divide the task into two different stages and propose a framework termed Static-Dynamic Collaboration (SDC) to achieve a better trade-off between stability and plasticity. Specifically, our method divides the normal pipeline of FSCIL into Static Retaining Stage (SRS) and Dynamic Learning Stage (DLS), which harnesses old static and incremental dynamic class information, respectively. During SRS, we train an initial model with sufficient data in the base session and preserve the key part as static memory to retain fundamental old knowledge. During DLS, we introduce an extra dynamic projector jointly trained with the previous static memory. By employing both stages, our method achieves improved retention of old knowledge while continuously adapting to new classes. Extensive experiments on three public benchmarks and a real-world application dataset demonstrate that our method achieves state-of-the-art performance against other competitors.", "date": "2026-01-13", "categories": []}
{"id": "2601.08493", "title": "PKI: Prior Knowledge-Infused Neural Network for Few-Shot Class-Incremental Learning", "abstract": "Few-shot class-incremental learning (FSCIL) aims to continually adapt a model on a limited number of new-class examples, facing two well-known challenges: catastrophic forgetting and overfitting to new classes. Existing methods tend to freeze more parts of network components and finetune others with an extra memory during incremental sessions. These methods emphasize preserving prior knowledge to ensure proficiency in recognizing old classes, thereby mitigating catastrophic forgetting. Meanwhile, constraining fewer parameters can help in overcoming overfitting with the assistance of prior knowledge. Following previous methods, we retain more prior knowledge and propose a prior knowledge-infused neural network (PKI) to facilitate FSCIL. PKI consists of a backbone, an ensemble of projectors, a classifier, and an extra memory. In each incremental session, we build a new projector and add it to the ensemble. Subsequently, we finetune the new projector and the classifier jointly with other frozen network components, ensuring the rich prior knowledge is utilized effectively. By cascading projectors, PKI integrates prior knowledge accumulated from previous sessions and learns new knowledge flexibly, which helps to recognize old classes and efficiently learn new classes. Further, to reduce the resource consumption associated with keeping many projectors, we design two variants of the prior knowledge-infused neural network (PKIV-1 and PKIV-2) to trade off a balance between resource consumption and performance by reducing the number of projectors. Extensive experiments on three popular benchmarks demonstrate that our approach outperforms state-of-the-art methods.", "date": "2026-01-13", "categories": []}
{"id": "2601.08519", "title": "CD^2: Constrained Dataset Distillation for Few-Shot Class-Incremental Learning", "abstract": "Few-shot class-incremental learning (FSCIL) receives significant attention from the public to perform classification continuously with a few training samples, which suffers from the key catastrophic forgetting problem. Existing methods usually employ an external memory to store previous knowledge and treat it with incremental classes equally, which cannot properly preserve previous essential knowledge. To solve this problem and inspired by recent distillation works on knowledge transfer, we propose a framework termed \\textbf{C}onstrained \\textbf{D}ataset \\textbf{D}istillation (\\textbf{CD$^2$}) to facilitate FSCIL, which includes a dataset distillation module (\\textbf{DDM}) and a distillation constraint module~(\\textbf{DCM}). Specifically, the DDM synthesizes highly condensed samples guided by the classifier, forcing the model to learn compacted essential class-related clues from a few incremental samples. The DCM introduces a designed loss to constrain the previously learned class distribution, which can preserve distilled knowledge more sufficiently. Extensive experiments on three public datasets show the superiority of our method against other state-of-the-art competitors.", "date": "2026-01-13", "categories": []}
{"id": "2501.09705", "title": "Practical Continual Forgetting for Pre-trained Vision Models", "abstract": "For privacy and security concerns, the need to erase unwanted information from pre-trained vision models is becoming evident nowadays. In real-world scenarios, erasure requests originate at any time from both users and model owners, and these requests usually form a sequence. Therefore, under such a setting, selective information is expected to be continuously removed from a pre-trained model while maintaining the rest. We define this problem as continual forgetting and identify three key challenges. (i) For unwanted knowledge, efficient and effective deleting is crucial. (ii) For remaining knowledge, the impact brought by the forgetting procedure should be minimal. (iii) In real-world scenarios, the training samples may be scarce or partially missing during the process of forgetting. To address them, we first propose Group Sparse LoRA (GS-LoRA). Specifically, towards (i), we introduce Low-Rank Adaptation (LoRA) modules to fine-tune the Feed-Forward Network (FFN) layers in Transformer blocks for each forgetting task independently, and towards (ii), a simple group sparse regularization is adopted, enabling automatic selection of specific LoRA groups and zeroing out the others. To further extend GS-LoRA to more practical scenarios, we incorporate prototype information as additional supervision and introduce a more practical approach, GS-LoRA++. For each forgotten class, we move the logits away from its original prototype. For the remaining classes, we pull the logits closer to their respective prototypes. We conduct extensive experiments on face recognition, object detection, and image classification and demonstrate that our method manages to forget specific classes with minimal impact on other classes. Codes have been released on https://github.com/bjzhb666/GS-LoRA.", "date": "2025-01-16", "categories": []}
{"id": "2601.02708", "title": "CREAM: Continual Retrieval on Dynamic Streaming Corpora with Adaptive Soft Memory", "abstract": "Information retrieval (IR) in dynamic data streams is a crucial task, as shifts in data distribution degrade the performance of AI-powered IR systems. To mitigate this issue, memory-based continual learning has been widely adopted for IR. However, existing methods rely on a fixed set of queries with ground-truth documents, which limits generalization to unseen data, making them impractical for real-world applications. To enable more effective learning with unseen topics of a new corpus without ground-truth labels, we propose CREAM, a self-supervised framework for memory-based continual retrieval. CREAM captures the evolving semantics of streaming queries and documents into dynamically structured soft memory and leverages it to adapt to both seen and unseen topics in an unsupervised setting. We realize this through three key techniques: fine-grained similarity estimation, regularized cluster prototyping, and stratified coreset sampling. Experiments on two benchmark datasets demonstrate that CREAM exhibits superior adaptability and retrieval accuracy, outperforming the strongest method in a label-free setting by 27.79% in Success@5 and 44.5% in Recall@10 on average, and achieving performance comparable to or even exceeding that of supervised methods.", "date": "2026-01-06", "categories": []}
{"id": "2601.06392", "title": "Continual Quantum Architecture Search with Tensor-Train Encoding: Theory and Applications to Signal Processing", "abstract": "We introduce CL-QAS, a continual quantum architecture search framework that mitigates the challenges of costly amplitude encoding and catastrophic forgetting in variational quantum circuits. The method uses Tensor-Train encoding to efficiently compress high-dimensional stochastic signals into low-rank quantum feature representations. A bi-loop learning strategy separates circuit parameter optimization from architecture exploration, while an Elastic Weight Consolidation regularization ensures stability across sequential tasks. We derive theoretical upper bounds on approximation, generalization, and robustness under quantum noise, demonstrating that CL-QAS achieves controllable expressivity, sample-efficient generalization, and smooth convergence without barren plateaus. Empirical evaluations on electrocardiogram (ECG)-based signal classification and financial time-series forecasting confirm substantial improvements in accuracy, balanced accuracy, F1 score, and reward. CL-QAS maintains strong forward and backward transfer and exhibits bounded degradation under depolarizing and readout noise, highlighting its potential for adaptive, noise-resilient quantum learning on near-term devices.", "date": "2026-01-10", "categories": []}
{"id": "2601.06742", "title": "Federated Continual Learning for Privacy-Preserving Hospital Imaging Classification", "abstract": "Deep learning models for radiology interpretation increasingly rely on multi-institutional data, yet privacy regulations and distribution shift across hospitals limit central data pooling. Federated learning (FL) allows hospitals to collaboratively train models without sharing raw images, but current FL algorithms typically assume a static data distribution. In practice, hospitals experience continual evolution in case mix, annotation protocols, and imaging devices, which leads to catastrophic forgetting when models are updated sequentially. Federated continual learning (FCL) aims to reconcile these challenges but existing methods either ignore the stringent privacy constraints of healthcare or rely on replay buffers and public surrogate datasets that are difficult to justify in clinical settings. We study FCL for chest radiography classification in a setting where hospitals are clients that receive temporally evolving streams of cases and labels. We introduce DP-FedEPC (Differentially Private Federated Elastic Prototype Consolidation), a method that combines elastic weight consolidation (EWC), prototype-based rehearsal, and client-side differential privacy within a standard FedAvg framework. EWC constrains updates along parameters deemed important for previous tasks, while a memory of latent prototypes preserves class structure without storing raw images. Differentially private stochastic gradient descent (DP-SGD) at each client adds calibrated Gaussian noise to clipped gradients, providing formal privacy guarantees for individual radiographs.", "date": "2026-01-11", "categories": []}
{"id": "2601.07117", "title": "Few-shot Class-Incremental Learning via Generative Co-Memory Regularization", "abstract": "Few-shot class-incremental learning (FSCIL) aims to incrementally learn models from a small amount of novel data, which requires strong representation and adaptation ability of models learned under few-example supervision to avoid catastrophic forgetting on old classes and overfitting to novel classes. This work proposes a generative co-memory regularization approach to facilitate FSCIL. In the approach, the base learning leverages generative domain adaptation finetuning to finetune a pretrained generative encoder on a few examples of base classes by jointly incorporating a masked autoencoder (MAE) decoder for feature reconstruction and a fully-connected classifier for feature classification, which enables the model to efficiently capture general and adaptable representations. Using the finetuned encoder and learned classifier, we construct two class-wise memories: representation memory for storing the mean features for each class, and weight memory for storing the classifier weights. After that, the memory-regularized incremental learning is performed to train the classifier dynamically on the examples of few-shot classes in each incremental session by simultaneously optimizing feature classification and co-memory regularization. The memories are updated in a class-incremental manner and they collaboratively regularize the incremental learning. In this way, the learned models improve recognition accuracy, while mitigating catastrophic forgetting over old classes and overfitting to novel classes. Extensive experiments on popular benchmarks clearly demonstrate that our approach outperforms the state-of-the-arts.", "date": "2026-01-12", "categories": []}
{"id": "2601.07636", "title": "Beyond Sharpness: A Flatness Decomposition Framework for Efficient Continual Learning", "abstract": "Continual Learning (CL) aims to enable models to sequentially learn multiple tasks without forgetting previous knowledge. Recent studies have shown that optimizing towards flatter loss minima can improve model generalization. However, existing sharpness-aware methods for CL suffer from two key limitations: (1) they treat sharpness regularization as a unified signal without distinguishing the contributions of its components. and (2) they introduce substantial computational overhead that impedes practical deployment. To address these challenges, we propose FLAD, a novel optimization framework that decomposes sharpness-aware perturbations into gradient-aligned and stochastic-noise components, and show that retaining only the noise component promotes generalization. We further introduce a lightweight scheduling scheme that enables FLAD to maintain significant performance gains even under constrained training time. FLAD can be seamlessly integrated into various CL paradigms and consistently outperforms standard and sharpness-aware optimizers in diverse experimental settings, demonstrating its effectiveness and practicality in CL.", "date": "2026-01-12", "categories": []}
{"id": "2506.12409", "title": "Branch, or Layer? Zeroth-Order Optimization for Continual Learning of Vision-Language Models", "abstract": "Vision-Language Continual Learning (VLCL) has attracted significant research attention for its robust capabilities, and the adoption of Parameter-Efficient Fine-Tuning (PEFT) strategies is enabling these models to achieve competitive performance with substantially reduced resource consumption. However, dominated First-Order (FO) optimization is prone to trap models in suboptimal local minima, especially in limited exploration subspace within PEFT. To overcome this challenge, this paper pioneers a systematic exploration of adopting Zeroth-Order (ZO) optimization for PEFT-based VLCL. We first identify the incompatibility of naive full-ZO adoption in VLCL due to optimization process instability. We then investigate the application of ZO optimization from a modality branch-wise to a fine-grained layer-wise across various training units to identify an optimal strategy. Besides, a key theoretical insight reveals that vision modality exhibit higher variance than language counterparts in VLCL during the ZO optimization process, and we propose a modality-aware ZO strategy, which adopts gradient sign normalization in ZO and constrains vision modality perturbation to further improve performance. Benefiting from the adoption of ZO optimization, PEFT-based VLCL fulfills better ability to escape local minima during the optimization process, extensive experiments on four benchmarks demonstrate that our method achieves state-of-the-art results.", "date": "2025-06-14", "categories": []}
{"id": "2601.05623", "title": "Continual Learning of Achieving Forgetting-free and Positive Knowledge Transfer", "abstract": "Existing research on continual learning (CL) of a sequence of tasks focuses mainly on dealing with catastrophic forgetting (CF) to balance the learning plasticity of new tasks and the memory stability of old tasks. However, an ideal CL agent should not only be able to overcome CF, but also encourage positive forward and backward knowledge transfer (KT), i.e., using the learned knowledge from previous tasks for the new task learning (namely FKT), and improving the previous tasks' performance with the knowledge of the new task (namely BKT). To this end, this paper first models CL as an optimization problem in which each sequential learning task aims to achieve its optimal performance under the constraint that both FKT and BKT should be positive. It then proposes a novel Enhanced Task Continual Learning (ETCL) method, which achieves forgetting-free and positive KT. Furthermore, the bounds that can lead to negative FKT and BKT are estimated theoretically. Based on the bounds, a new strategy for online task similarity detection is also proposed to facilitate positive KT. To overcome CF, ETCL learns a set of task-specific binary masks to isolate a sparse sub-network for each task while preserving the performance of a dense network for the task. At the beginning of a new task learning, ETCL tries to align the new task's gradient with that of the sub-network of the previous most similar task to ensure positive FKT. By using a new bi-objective optimization strategy and an orthogonal gradient projection method, ETCL updates only the weights of previous similar tasks at the classification layer to achieve positive BKT. Extensive evaluations demonstrate that the proposed ETCL markedly outperforms strong baselines on dissimilar, similar, and mixed task sequences.", "date": "2026-01-09", "categories": []}
{"id": "2601.05874", "title": "Continual-learning for Modelling Low-Resource Languages from Large Language Models", "abstract": "Modelling a language model for a multi-lingual scenario includes several potential challenges, among which catastrophic forgetting is the major challenge. For example, small language models (SLM) built for low-resource languages by adapting large language models (LLMs) pose the challenge of catastrophic forgetting. This work proposes to employ a continual learning strategy using parts-of-speech (POS)-based code-switching along with a replay adapter strategy to mitigate the identified gap of catastrophic forgetting while training SLM from LLM. Experiments conducted on vision language tasks such as visual question answering and language modelling task exhibits the success of the proposed architecture.", "date": "2026-01-09", "categories": []}
{"id": "2601.05152", "title": "Safe Continual Reinforcement Learning Methods for Nonstationary Environments. Towards a Survey of the State of the Art", "abstract": "This work provides a state-of-the-art survey of continual safe online reinforcement learning (COSRL) methods. We discuss theoretical aspects, challenges, and open questions in building continual online safe reinforcement learning algorithms. We provide the taxonomy and the details of continual online safe reinforcement learning methods based on the type of safe learning mechanism that takes adaptation to nonstationarity into account. We categorize safety constraints formulation for online reinforcement learning algorithms, and finally, we discuss prospects for creating reliable, safe online learning algorithms.\n  Keywords: safe RL in nonstationary environments, safe continual reinforcement learning under nonstationarity, HM-MDP, NSMDP, POMDP, safe POMDP, constraints for continual learning, safe continual reinforcement learning review, safe continual reinforcement learning survey, safe continual reinforcement learning, safe online learning under distribution shift, safe continual online adaptation, safe reinforcement learning, safe exploration, safe adaptation, constrained Markov decision processes, safe reinforcement learning, partially observable Markov decision process, safe reinforcement learning and hidden Markov decision processes, Safe Online Reinforcement Learning, safe online reinforcement learning, safe online reinforcement learning, safe meta-learning, safe meta-reinforcement learning, safe context-based reinforcement learning, formulating safety constraints for continual learning", "date": "2026-01-08", "categories": []}
{"id": "2508.21296", "title": "MyGO: Memory Yielding Generative Offline-consolidation for Lifelong Learning Systems", "abstract": "Continual or Lifelong Learning aims to develop models capable of acquiring new knowledge from a sequence of tasks without catastrophically forgetting what has been learned before. Existing approaches often rely on storing samples from previous tasks (experience replay) or employing complex regularization terms to protect learned weights. However, these methods face challenges related to data privacy, storage limitations, and performance degradation when tasks are dissimilar. To address these challenges, we introduce MyGO (Memory Yielding Generative Offline-consolidation), a novel lifelong learning framework inspired by the biological wake-sleep cycle. During the \"wake\" phase, the system rapidly learns a new task and trains a compact generative model (Generative Memory, G-mem) to capture its data distribution. During the \"sleep\" phase, the system enters an offline state, using all learned G-mem models to generate pseudo-data (\"dreams\") and consolidate new and old knowledge into a core feature extractor via knowledge distillation. This approach obviates the need to store any raw data, retaining only compact generative models, which offers significant advantages in privacy and storage efficiency. We evaluate MyGO on computer vision (Split-MNIST) and natural language processing (Split-AG News) benchmarks, comparing it against a sequential fine-tuning baseline. The results demonstrate that MyGO significantly mitigates catastrophic forgetting and maintains high average accuracy across tasks, proving the framework's effectiveness and domain-generality.", "date": "2025-08-29", "categories": []}
{"id": "2601.02232", "title": "ELLA: Efficient Lifelong Learning for Adapters in Large Language Models", "abstract": "Large Language Models (LLMs) suffer severe catastrophic forgetting when adapted sequentially to new tasks in a continual learning (CL) setting. Existing approaches are fundamentally limited: replay-based methods are impractical and privacy-violating, while strict orthogonality-based methods collapse under scale: each new task is projected onto an orthogonal complement, progressively reducing the residual degrees of freedom and eliminating forward transfer by forbidding overlap in shared representations. In this work, we introduce ELLA, a training framework built on the principle of selective subspace de-correlation. Rather than forbidding all overlap, ELLA explicitly characterizes the structure of past updates and penalizes alignments along their high-energy, task-specific directions, while preserving freedom in the low-energy residual subspaces to enable transfer. Formally, this is realized via a lightweight regularizer on a single aggregated update matrix. We prove this mechanism corresponds to an anisotropic shrinkage operator that bounds interference, yielding a penalty that is both memory- and compute-constant regardless of task sequence length. ELLA requires no data replay, no architectural expansion, and negligible storage. Empirically, it achieves state-of-the-art CL performance on three popular benchmarks, with relative accuracy gains of up to $9.6\\%$ and a $35\\times$ smaller memory footprint. Further, ELLA scales robustly across architectures and actively enhances the model's zero-shot generalization performance on unseen tasks, establishing a principled and scalable solution for constructive lifelong LLM adaptation.", "date": "2026-01-05", "categories": []}
{"id": "2601.03509", "title": "Evolving Programmatic Skill Networks", "abstract": "We study continual skill acquisition in open-ended embodied environments where an agent must construct, refine, and reuse an expanding library of executable skills. We introduce the Programmatic Skill Network (PSN), a framework in which skills are executable symbolic programs forming a compositional network that evolves through experience. PSN defines three core mechanisms instantiated via large language models: (1)REFLECT for structured fault localization over skill compositions, (2) progressive optimization with maturity-aware update gating that stabilizes reliable skills while maintaining plasticity for uncertain ones, and (3) canonical structural refactoring under rollback validation that maintains network compactness. We further show that PSN's learning dynamics exhibit structural parallels to neural network training. Experiments on MineDojo and Crafter demonstrate robust skill reuse, rapid adaptation, and strong generalization across open-ended task distributions.\\footnote{We plan to open-source the code.", "date": "2026-01-07", "categories": []}
{"id": "2601.03658", "title": "Group and Exclusive Sparse Regularization-based Continual Learning of CNNs", "abstract": "We present a regularization-based approach for continual learning (CL) of fixed capacity convolutional neural networks (CNN) that does not suffer from the problem of catastrophic forgetting when learning multiple tasks sequentially. This method referred to as Group and Exclusive Sparsity based Continual Learning (GESCL) avoids forgetting of previous tasks by ensuring the stability of the CNN via a stability regularization term, which prevents filters detected as important for past tasks to deviate too much when learning a new task. On top of that, GESCL makes the network plastic via a plasticity regularization term that leverage the over-parameterization of CNNs to efficiently sparsify the network and tunes unimportant filters making them relevant for future tasks. Doing so, GESCL deals with significantly less parameters and computation compared to CL approaches that either dynamically expand the network or memorize past tasks' data. Experiments on popular CL vision benchmarks show that GESCL leads to significant improvements over state-of-the-art method in terms of overall CL performance, as measured by classification accuracy as well as in terms of avoiding catastrophic forgetting.", "date": "2026-01-07", "categories": []}
{"id": "2601.03858", "title": "What Does Loss Optimization Actually Teach, If Anything? Knowledge Dynamics in Continual Pre-training of LLMs", "abstract": "Continual Pre-Training (CPT) is widely used for acquiring and updating factual knowledge in LLMs. This practice treats loss as a proxy for knowledge learning, while offering no grounding into how it changes during training. We study CPT as a knowledge learning process rather than a solely optimization problem. We construct a controlled, distribution-matched benchmark of factual documents and interleave diagnostic probes directly into the CPT loop, enabling epoch-level measurement of knowledge acquisition dynamics and changes in Out-Of-Domain (OOD) general skills (e.g., math). We further analyze how CPT reshapes knowledge circuits during training. Across three instruction-tuned LLMs and multiple CPT strategies, optimization and learning systematically diverge as loss decreases monotonically while factual learning is unstable and non-monotonic. Acquired facts are rarely consolidated, learning is strongly conditioned on prior exposure, and OOD performance degrades from early epochs. Circuit analysis reveals rapid reconfiguration of knowledge pathways across epochs, providing an explanation for narrow acquisition windows and systematic forgetting. These results show that loss optimization is misaligned with learning progress in CPT and motivate evaluation of stopping criteria based on task-level learning dynamics.", "date": "2026-01-07", "categories": []}
{"id": "2507.12043", "title": "Information-Theoretic Generalization Bounds of Replay-based Continual Learning", "abstract": "Continual learning (CL) has emerged as a dominant paradigm for acquiring knowledge from sequential tasks while avoiding catastrophic forgetting. Although many CL methods have been proposed to show impressive empirical performance, the theoretical understanding of their generalization behavior remains limited, particularly for replay-based approaches. This paper establishes a unified theoretical framework for replay-based CL, deriving a series of information-theoretic generalization bounds that explicitly elucidate the impact of the memory buffer alongside the current task on generalization performance. Specifically, our hypothesis-based bounds capture the trade-off between the number of selected exemplars and the information dependency between the hypothesis and the memory buffer. Our prediction-based bounds yield tighter and computationally tractable upper bounds on the generalization error by leveraging low-dimensional variables. Theoretical analysis is general and broadly applicable to a wide range of learning algorithms, exemplified by stochastic gradient Langevin dynamics (SGLD) as a representative method. Comprehensive experimental evaluations demonstrate the effectiveness of our derived bounds in capturing the generalization dynamics in replay-based CL settings.", "date": "2025-07-16", "categories": []}
{"id": "2601.02500", "title": "GEM-Style Constraints for PEFT with Dual Gradient Projection in LoRA", "abstract": "Full fine-tuning of Large Language Models (LLMs) is computationally costly, motivating Continual Learning (CL) approaches that utilize parameter-efficient adapters. We revisit Gradient Episodic Memory (GEM) within the Low-Rank Adapter (LoRA) subspace and introduce I-GEM: a fixed-budget, GPU-resident dual projected-gradient approximation to GEM's quadratic projection. By constraining non-interference solely within the adapter parameters, I-GEM preserves GEM-like stability with orders-of-magnitude lower mean projection overhead. On a 3-task AG News split with induced domain drift, using GPT-2 (355M) and LoRA ($r=8$), I-GEM matches GEM's average accuracy (within $\\sim\\!0.04$ pts) and outperforms A-GEM by $\\sim\\!1.4$ pts. Crucially, it reduces projection time vs.\\ GEM by a factor of $\\sim\\!10^3$. These results suggest that applying GEM constraints in the LoRA subspace is a practical pathway for continual learning at the LLM scale.", "date": "2026-01-05", "categories": []}
{"id": "2601.01892", "title": "Forget Less by Learning from Parents Through Hierarchical Relationships", "abstract": "Custom Diffusion Models (CDMs) offer impressive capabilities for personalization in generative modeling, yet they remain vulnerable to catastrophic forgetting when learning new concepts sequentially. Existing approaches primarily focus on minimizing interference between concepts, often neglecting the potential for positive inter-concept interactions. In this work, we present Forget Less by Learning from Parents (FLLP), a novel framework that introduces a parent-child inter-concept learning mechanism in hyperbolic space to mitigate forgetting. By embedding concept representations within a Lorentzian manifold, naturally suited to modeling tree-like hierarchies, we define parent-child relationships in which previously learned concepts serve as guidance for adapting to new ones. Our method not only preserves prior knowledge but also supports continual integration of new concepts. We validate FLLP on three public datasets and one synthetic benchmark, showing consistent improvements in both robustness and generalization.", "date": "2026-01-05", "categories": []}
{"id": "2601.01926", "title": "MacVQA: Adaptive Memory Allocation and Global Noise Filtering for Continual Visual Question Answering", "abstract": "Visual Question Answering (VQA) requires models to reason over multimodal information, combining visual and textual data. With the development of continual learning, significant progress has been made in retaining knowledge and adapting to new information in the VQA domain. However, current methods often struggle with balancing knowledge retention, adaptation, and robust feature representation. To address these challenges, we propose a novel framework with adaptive memory allocation and global noise filtering called MacVQA for visual question answering. MacVQA fuses visual and question information while filtering noise to ensure robust representations, and employs prototype-based memory allocation to optimize feature quality and memory usage. These designs enable MacVQA to balance knowledge acquisition, retention, and compositional generalization in continual VQA learning. Experiments on ten continual VQA tasks show that MacVQA outperforms existing baselines, achieving 43.38% average accuracy and 2.32% average forgetting on standard tasks, and 42.53% average accuracy and 3.60% average forgetting on novel composition tasks.", "date": "2026-01-05", "categories": []}
{"id": "2601.01963", "title": "Forget Less by Learning Together through Concept Consolidation", "abstract": "Custom Diffusion Models (CDMs) have gained significant attention due to their remarkable ability to personalize generative processes. However, existing CDMs suffer from catastrophic forgetting when continuously learning new concepts. Most prior works attempt to mitigate this issue under the sequential learning setting with a fixed order of concept inflow and neglect inter-concept interactions. In this paper, we propose a novel framework - Forget Less by Learning Together (FL2T) - that enables concurrent and order-agnostic concept learning while addressing catastrophic forgetting. Specifically, we introduce a set-invariant inter-concept learning module where proxies guide feature selection across concepts, facilitating improved knowledge retention and transfer. By leveraging inter-concept guidance, our approach preserves old concepts while efficiently incorporating new ones. Extensive experiments, across three datasets, demonstrates that our method significantly improves concept retention and mitigates catastrophic forgetting, highlighting the effectiveness of inter-concept catalytic behavior in incremental concept learning of ten tasks with at least 2% gain on average CLIP Image Alignment scores.", "date": "2026-01-05", "categories": []}
{"id": "2504.04579", "title": "From Continual Learning to SGD and Back: Better Rates for Continual Linear Models", "abstract": "We study the common continual learning setup where an overparameterized model is sequentially fitted to a set of jointly realizable tasks. We analyze forgetting, defined as the loss on previously seen tasks, after $k$ iterations. For continual linear models, we prove that fitting a task is equivalent to a single stochastic gradient descent (SGD) step on a modified objective. We develop novel last-iterate SGD upper bounds in the realizable least squares setup and leverage them to derive new results for continual learning. Focusing on random orderings over $T$ tasks, we establish universal forgetting rates, whereas existing rates depend on problem dimensionality or complexity and become prohibitive in highly overparameterized regimes. In continual regression with replacement, we improve the best existing rate from $O((d-\\bar{r})/k)$ to $O(\\min(1/\\sqrt[4]{k}, \\sqrt{(d-\\bar{r})}/k, \\sqrt{T\\bar{r}}/k))$, where $d$ is the dimensionality and $\\bar{r}$ the average task rank. Furthermore, we establish the first rate for random task orderings without replacement. The resulting rate $O(\\min(1/\\sqrt[4]{T},\\, (d-\\bar{r})/T))$ shows that randomization alone, without task repetition, prevents catastrophic forgetting in sufficiently long task sequences. Finally, we prove a matching $O(1/\\sqrt[4]{k})$ forgetting rate for continual linear classification on separable data. Our universal rates extend to broader methods, such as block Kaczmarz and POCS, illuminating their loss convergence under i.i.d. and single-pass orderings.", "date": "2025-04-06", "categories": []}
{"id": "2512.22973", "title": "YOLO-IOD: Towards Real Time Incremental Object Detection", "abstract": "Current methods for incremental object detection (IOD) primarily rely on Faster R-CNN or DETR series detectors; however, these approaches do not accommodate the real-time YOLO detection frameworks. In this paper, we first identify three primary types of knowledge conflicts that contribute to catastrophic forgetting in YOLO-based incremental detectors: foreground-background confusion, parameter interference, and misaligned knowledge distillation. Subsequently, we introduce YOLO-IOD, a real-time Incremental Object Detection (IOD) framework that is constructed upon the pretrained YOLO-World model, facilitating incremental learning via a stage-wise parameter-efficient fine-tuning process. Specifically, YOLO-IOD encompasses three principal components: 1) Conflict-Aware Pseudo-Label Refinement (CPR), which mitigates the foreground-background confusion by leveraging the confidence levels of pseudo labels and identifying potential objects relevant to future tasks. 2) Importancebased Kernel Selection (IKS), which identifies and updates the pivotal convolution kernels pertinent to the current task during the current learning stage. 3) Cross-Stage Asymmetric Knowledge Distillation (CAKD), which addresses the misaligned knowledge distillation conflict by transmitting the features of the student target detector through the detection heads of both the previous and current teacher detectors, thereby facilitating asymmetric distillation between existing and newly introduced categories. We further introduce LoCo COCO, a more realistic benchmark that eliminates data leakage across stages. Experiments on both conventional and LoCo COCO benchmarks show that YOLO-IOD achieves superior performance with minimal forgetting.", "date": "2025-12-28", "categories": []}
{"id": "2601.00756", "title": "Memory Bank Compression for Continual Adaptation of Large Language Models", "abstract": "Large Language Models (LLMs) have become a mainstay for many everyday applications. However, as data evolve their knowledge quickly becomes outdated. Continual learning aims to update LLMs with new information without erasing previously acquired knowledge. Although methods such as full fine-tuning can incorporate new data, they are computationally expensive and prone to catastrophic forgetting, where prior knowledge is overwritten. Memory-augmented approaches address this by equipping LLMs with a memory bank, that is an external memory module which stores information for future use. However, these methods face a critical limitation, in particular, the memory bank constantly grows in the real-world scenario when large-scale data streams arrive. In this paper, we propose MBC, a model that compresses the memory bank through a codebook optimization strategy during online adaptation learning. To ensure stable learning, we also introduce an online resetting mechanism that prevents codebook collapse. In addition, we employ Key-Value Low-Rank Adaptation in the attention layers of the LLM, enabling efficient utilization of the compressed memory representations. Experiments with benchmark question-answering datasets demonstrate that MBC reduces the memory bank size to 0.3% when compared against the most competitive baseline, while maintaining high retention accuracy during online adaptation learning. Our code is publicly available at https://github.com/Thomkat/MBC.", "date": "2026-01-02", "categories": []}
{"id": "2508.07307", "title": "MCITlib: Multimodal Continual Instruction Tuning Library and Benchmark", "abstract": "Continual learning enables AI systems to acquire new knowledge while retaining previously learned information. While traditional unimodal methods have made progress, the rise of Multimodal Large Language Models (MLLMs) brings new challenges in Multimodal Continual Learning (MCL), where models are expected to address both catastrophic forgetting and cross-modal coordination. To advance research in this area, we present MCITlib, a comprehensive library for Multimodal Continual Instruction Tuning. MCITlib currently implements 8 representative algorithms and conducts evaluations on 3 benchmarks under 2 backbone models. The library will be continuously updated to support future developments in MCL. The codebase is released at https://github.com/Ghy0501/MCITlib.", "date": "2025-08-10", "categories": []}
{"id": "2512.24695", "title": "Nested Learning: The Illusion of Deep Learning Architectures", "abstract": "Despite the recent progresses, particularly in developing Language Models, there are fundamental challenges and unanswered questions about how such models can continually learn/memorize, self-improve, and find effective solutions. In this paper, we present a new learning paradigm, called Nested Learning (NL), that coherently represents a machine learning model with a set of nested, multi-level, and/or parallel optimization problems, each of which with its own context flow. Through the lenses of NL, existing deep learning methods learns from data through compressing their own context flow, and in-context learning naturally emerges in large models. NL suggests a philosophy to design more expressive learning algorithms with more levels, resulting in higher-order in-context learning and potentially unlocking effective continual learning capabilities. We advocate for NL by presenting three core contributions: (1) Expressive Optimizers: We show that known gradient-based optimizers, such as Adam, SGD with Momentum, etc., are in fact associative memory modules that aim to compress the gradients' information (by gradient descent). Building on this insight, we present other more expressive optimizers with deep memory and/or more powerful learning rules; (2) Self-Modifying Learning Module: Taking advantage of NL's insights on learning algorithms, we present a sequence model that learns how to modify itself by learning its own update algorithm; and (3) Continuum Memory System: We present a new formulation for memory system that generalizes the traditional viewpoint of long/short-term memory. Combining our self-modifying sequence model with the continuum memory system, we present a continual learning module, called Hope, showing promising results in language modeling, knowledge incorporation, and few-shot generalization tasks, continual learning, and long-context reasoning tasks.", "date": "2025-12-31", "categories": []}
{"id": "2512.25000", "title": "Bi-C2R: Bidirectional Continual Compatible Representation for Re-indexing Free Lifelong Person Re-identification", "abstract": "Lifelong person Re-IDentification (L-ReID) exploits sequentially collected data to continuously train and update a ReID model, focusing on the overall performance of all data. Its main challenge is to avoid the catastrophic forgetting problem of old knowledge while training on new data. Existing L-ReID methods typically re-extract new features for all historical gallery images for inference after each update, known as \"re-indexing\". However, historical gallery data typically suffers from direct saving due to the data privacy issue and the high re-indexing costs for large-scale gallery images. As a result, it inevitably leads to incompatible retrieval between query features extracted by the updated model and gallery features extracted by those before the update, greatly impairing the re-identification performance. To tackle the above issue, this paper focuses on a new task called Re-index Free Lifelong person Re-IDentification (RFL-ReID), which requires performing lifelong person re-identification without re-indexing historical gallery images. Therefore, RFL-ReID is more challenging than L-ReID, requiring continuous learning and balancing new and old knowledge in diverse streaming data, and making the features output by the new and old models compatible with each other. To this end, we propose a Bidirectional Continuous Compatible Representation (Bi-C2R) framework to continuously update the gallery features extracted by the old model to perform efficient L-ReID in a compatible manner. We verify our proposed Bi-C2R method through theoretical analysis and extensive experiments on multiple benchmarks, which demonstrate that the proposed method can achieve leading performance on both the introduced RFL-ReID task and the traditional L-ReID task.", "date": "2025-12-31", "categories": []}
{"id": "2410.01324", "title": "Fair Class-Incremental Learning using Sample Weighting", "abstract": "Model fairness is becoming important in class-incremental learning for Trustworthy AI. While accuracy has been a central focus in class-incremental learning, fairness has been relatively understudied. However, naively using all the samples of the current task for training results in unfair catastrophic forgetting for certain sensitive groups including classes. We theoretically analyze that forgetting occurs if the average gradient vector of the current task data is in an \"opposite direction\" compared to the average gradient vector of a sensitive group, which means their inner products are negative. We then propose a fair class-incremental learning framework that adjusts the training weights of current task samples to change the direction of the average gradient vector and thus reduce the forgetting of underperforming groups and achieve fairness. For various group fairness measures, we formulate optimization problems to minimize the overall losses of sensitive groups while minimizing the disparities among them. We also show the problems can be solved with linear programming and propose an efficient Fairness-aware Sample Weighting (FSW) algorithm. Experiments show that FSW achieves better accuracy-fairness tradeoff results than state-of-the-art approaches on real datasets.", "date": "2024-10-02", "categories": []}
{"id": "2506.06694", "title": "Breaking Data Silos: Towards Open and Scalable Mobility Foundation Models via Generative Continual Learning", "abstract": "Human mobility is a fundamental pillar of urban science and sustainability, providing critical insights into energy consumption, carbon emissions, and public health. However, the discovery of universal mobility laws is currently hindered by the ``data silo'' problem, where institutional boundaries and privacy regulations fragment the necessary large-scale datasets. In this paper, we propose MoveGCL, a transformative framework that facilitates collaborative and decentralized mobility science via generative continual learning. MoveGCL enables a distributed ecosystem of data holders to jointly evolve a foundation model without compromising individual privacy. The core of MoveGCL lies in its ability to replay synthetic trajectories derived from a generative teacher and utilize a mobility-pattern-aware Mixture-of-Experts (MoE) architecture. This allows the model to encapsulate the unique characteristics of diverse urban structures while mitigating the risk of knowledge erosion (catastrophic forgetting). With a specialized layer-wise progressive adaptation strategy, MoveGCL ensures stable convergence during the continuous integration of new urban domains. Our experiments on six global urban datasets demonstrate that MoveGCL achieves performance parity with joint training, a previously unattainable feat under siloed conditions. This work provides a scalable, privacy-preserving pathway toward Open Mobility Science, empowering researchers to address global sustainability challenges through cross-institutional AI collaboration. To facilitate reproducibility and future research, we have released the code and models at \\color{blue}{https://github.com/tsinghua-fib-lab/MoveGCL}.", "date": "2025-06-07", "categories": []}
{"id": "2511.07749", "title": "Class Incremental Medical Image Segmentation via Prototype-Guided Calibration and Dual-Aligned Distillation", "abstract": "Class incremental medical image segmentation (CIMIS) aims to preserve knowledge of previously learned classes while learning new ones without relying on old-class labels. However, existing methods 1) either adopt one-size-fits-all strategies that treat all spatial regions and feature channels equally, which may hinder the preservation of accurate old knowledge, 2) or focus solely on aligning local prototypes with global ones for old classes while overlooking their local representations in new data, leading to knowledge degradation. To mitigate the above issues, we propose Prototype-Guided Calibration Distillation (PGCD) and Dual-Aligned Prototype Distillation (DAPD) for CIMIS in this paper. Specifically, PGCD exploits prototype-to-feature similarity to calibrate class-specific distillation intensity in different spatial regions, effectively reinforcing reliable old knowledge and suppressing misleading information from old classes. Complementarily, DAPD aligns the local prototypes of old classes extracted from the current model with both global prototypes and local prototypes, further enhancing segmentation performance on old categories. Comprehensive evaluations on two widely used multi-organ segmentation benchmarks demonstrate that our method outperforms state-of-the-art methods, highlighting its robustness and generalization capabilities.", "date": "2025-11-11", "categories": []}
{"id": "2512.04888", "title": "ZeBROD: Zero-Retraining Based Recognition and Object Detection Framework", "abstract": "Object detection constitutes the primary task within the domain of computer vision. It is utilized in numerous domains. Nonetheless, object detection continues to encounter the issue of catastrophic forgetting. The model must be retrained whenever new products are introduced, utilizing not only the new products dataset but also the entirety of the previous dataset. The outcome is obvious: increasing model training expenses and significant time consumption. In numerous sectors, particularly retail checkout, the frequent introduction of new products presents a great challenge. This study introduces Zero-Retraining Based Recognition and Object Detection (ZeBROD), a methodology designed to address the issue of catastrophic forgetting by integrating YOLO11n for object localization with DeIT and Proxy Anchor Loss for feature extraction and metric learning. For classification, we utilize cosine similarity between the embedding features of the target product and those in the Qdrant vector database. In a case study conducted in a retail store with 140 products, the experimental results demonstrate that our proposed framework achieves encouraging accuracy, whether for detecting new or existing products. Furthermore, without retraining, the training duration difference is significant. We achieve almost 3 times the training time efficiency compared to classical object detection approaches. This efficiency escalates as additional new products are added to the product database. The average inference time is 580 ms per image containing multiple products, on an edge device, validating the proposed framework's feasibility for practical use.", "date": "2025-12-04", "categories": []}
{"id": "2512.23017", "title": "Merge before Forget: A Single LoRA Continual Learning via Continual Merging", "abstract": "Parameter-efficient continual learning has emerged as a promising approach for large language models (LLMs) to mitigate catastrophic forgetting while enabling adaptation to new tasks. Current Low-Rank Adaptation (LoRA) continual learning techniques often retain and freeze previously learned LoRAs or generate data representations to overcome forgetting, typically utilizing these to support new LoRAs learn new tasks. However, these methods not only ignore growing computational memory with tasks and limited storage space but also suffer from potential task interference due to the lack of effective LoRA merging mechanisms. In this paper, we propose a novel continual learning method that orthogonally initializes and sequentially merges LoRAs updates into a single unified LoRA. Our method leverages orthogonal basis extraction from previously learned LoRA to initialize the learning of new tasks, further exploits the intrinsic asymmetry property of LoRA components by using a time-aware scaling mechanism to balance new and old knowledge during continual merging. Our approach maintains constant memory complexity with respect to the number of tasks, minimizes interference between past and new tasks via orthogonal basis initialization, and improves performance over asymmetric LoRA merging via adaptive scaling. We provide theoretical analysis to justify our design and conduct extensive experiments across diverse continual learning benchmarks using various Llama models, demonstrating the effectiveness and efficiency of our method.", "date": "2025-12-28", "categories": []}
{"id": "2512.23419", "title": "The World Is Bigger! A Computationally-Embedded Perspective on the Big World Hypothesis", "abstract": "Continual learning is often motivated by the idea, known as the big world hypothesis, that \"the world is bigger\" than the agent. Recent problem formulations capture this idea by explicitly constraining an agent relative to the environment. These constraints lead to solutions in which the agent continually adapts to best use its limited capacity, rather than converging to a fixed solution. However, explicit constraints can be ad hoc, difficult to incorporate, and may limit the effectiveness of scaling up the agent's capacity. In this paper, we characterize a problem setting in which an agent, regardless of its capacity, is constrained by being embedded in the environment. In particular, we introduce a computationally-embedded perspective that represents an embedded agent as an automaton simulated within a universal (formal) computer. Such an automaton is always constrained; we prove that it is equivalent to an agent that interacts with a partially observable Markov decision process over a countably infinite state-space. We propose an objective for this setting, which we call interactivity, that measures an agent's ability to continually adapt its behaviour by learning new predictions. We then develop a model-based reinforcement learning algorithm for interactivity-seeking, and use it to construct a synthetic problem to evaluate continual learning capability. Our results show that deep nonlinear networks struggle to sustain interactivity, whereas deep linear networks sustain higher interactivity as capacity increases.", "date": "2025-12-29", "categories": []}
{"id": "2512.21515", "title": "Perplexity-Aware Data Scaling Law: Perplexity Landscapes Predict Performance for Continual Pre-training", "abstract": "Continual Pre-training (CPT) serves as a fundamental approach for adapting foundation models to domain-specific applications. Scaling laws for pre-training define a power-law relationship between dataset size and the test loss of an LLM. However, the marginal gains from simply increasing data for CPT diminish rapidly, yielding suboptimal data utilization and inefficient training. To address this challenge, we propose a novel perplexity-aware data scaling law to establish a predictive relationship between the perplexity landscape of domain-specific data and the test loss. Our approach leverages the perplexity derived from the pre-trained model on domain data as a proxy for estimating the knowledge gap, effectively quantifying the informational perplexity landscape of candidate training samples. By fitting this scaling law across diverse perplexity regimes, we enable adaptive selection of high-utility data subsets, prioritizing content that maximizes knowledge absorption while minimizing redundancy and noise. Extensive experiments demonstrate that our method consistently identifies near-optimal training subsets and achieves superior performance on both medical and general-domain benchmarks.", "date": "2025-12-25", "categories": []}
{"id": "2512.21743", "title": "Dynamic Feedback Engines: Layer-Wise Control for Self-Regulating Continual Learning", "abstract": "Continual learning aims to acquire new tasks while preserving performance on previously learned ones, but most methods struggle with catastrophic forgetting. Existing approaches typically treat all layers uniformly, often trading stability for plasticity or vice versa. However, different layers naturally exhibit varying levels of uncertainty (entropy) when classifying tasks. High-entropy layers tend to underfit by failing to capture task-specific patterns, while low-entropy layers risk overfitting by becoming overly confident and specialized. To address this imbalance, we propose an entropy-aware continual learning method that employs a dynamic feedback mechanism to regulate each layer based on its entropy. Specifically, our approach reduces entropy in high-entropy layers to mitigate underfitting and increases entropy in overly confident layers to alleviate overfitting. This adaptive regulation encourages the model to converge to wider local minima, which have been shown to improve generalization. Our method is general and can be seamlessly integrated with both replay- and regularization-based approaches. Experiments on various datasets demonstrate substantial performance gains over state-of-the-art continual learning baselines.", "date": "2025-12-25", "categories": []}
{"id": "2512.21845", "title": "Scalable Class-Incremental Learning Based on Parametric Neural Collapse", "abstract": "Incremental learning often encounter challenges such as overfitting to new data and catastrophic forgetting of old data. Existing methods can effectively extend the model for new tasks while freezing the parameters of the old model, but ignore the necessity of structural efficiency to lead to the feature difference between modules and the class misalignment due to evolving class distributions. To address these issues, we propose scalable class-incremental learning based on parametric neural collapse (SCL-PNC) that enables demand-driven, minimal-cost backbone expansion by adapt-layer and refines the static into a dynamic parametric Equiangular Tight Frame (ETF) framework according to incremental class. This method can efficiently handle the model expansion question with the increasing number of categories in real-world scenarios. Additionally, to counteract feature drift in serial expansion models, the parallel expansion framework is presented with a knowledge distillation algorithm to align features across expansion modules. Therefore, SCL-PNC can not only design a dynamic and extensible ETF classifier to address class misalignment due to evolving class distributions, but also ensure feature consistency by an adapt-layer with knowledge distillation between extended modules. By leveraging neural collapse, SCL-PNC induces the convergence of the incremental expansion model through a structured combination of the expandable backbone, adapt-layer, and the parametric ETF classifier. Experiments on standard benchmarks demonstrate the effectiveness and efficiency of our proposed method. Our code is available at https://github.com/zhangchuangxin71-cyber/dynamic_ ETF2. Keywords: Class incremental learning; Catastrophic forgetting; Neural collapse;Knowledge distillation; Expanded model.", "date": "2025-12-26", "categories": []}
{"id": "2512.22029", "title": "LibContinual: A Comprehensive Library towards Realistic Continual Learning", "abstract": "A fundamental challenge in Continual Learning (CL) is catastrophic forgetting, where adapting to new tasks degrades the performance on previous ones. While the field has evolved with diverse methods, this rapid surge in diverse methodologies has culminated in a fragmented research landscape. The lack of a unified framework, including inconsistent implementations, conflicting dependencies, and varying evaluation protocols, makes fair comparison and reproducible research increasingly difficult. To address this challenge, we propose LibContinual, a comprehensive and reproducible library designed to serve as a foundational platform for realistic CL. Built upon a high-cohesion, low-coupling modular architecture, LibContinual integrates 19 representative algorithms across five major methodological categories, providing a standardized execution environment. Meanwhile, leveraging this unified framework, we systematically identify and investigate three implicit assumptions prevalent in mainstream evaluation: (1) offline data accessibility, (2) unregulated memory resources, and (3) intra-task semantic homogeneity. We argue that these assumptions often overestimate the real-world applicability of CL methods. Through our comprehensive analysis using strict online CL settings, a novel unified memory budget protocol, and a proposed category-randomized setting, we reveal significant performance drops in many representative CL methods when subjected to these real-world constraints. Our study underscores the necessity of resource-aware and semantically robust CL strategies, and offers LibContinual as a foundational toolkit for future research in realistic continual learning. The source code is available from \\href{https://github.com/RL-VIG/LibContinual}{https://github.com/RL-VIG/LibContinual}.", "date": "2025-12-26", "categories": []}
{"id": "2512.20634", "title": "Real Time Detection and Quantitative Analysis of Spurious Forgetting in Continual Learning", "abstract": "Catastrophic forgetting remains a fundamental challenge in continual learning for large language models. Recent work revealed that performance degradation may stem from spurious forgetting caused by task alignment disruption rather than true knowledge loss. However, this work only qualitatively describes alignment, relies on post-hoc analysis, and lacks automatic distinction mechanisms.\n  We introduce the shallow versus deep alignment framework, providing the first quantitative characterization of alignment depth. We identify that current task alignment approaches suffer from shallow alignment - maintained only over the first few output tokens (approximately 3-5) - making models vulnerable to forgetting. This explains why spurious forgetting occurs, why it is reversible, and why fine-tuning attacks are effective.\n  We propose a comprehensive framework addressing all gaps: (1) quantitative metrics (0-1 scale) to measure alignment depth across token positions; (2) real-time detection methods for identifying shallow alignment during training; (3) specialized analysis tools for visualization and recovery prediction; and (4) adaptive mitigation strategies that automatically distinguish forgetting types and promote deep alignment. Extensive experiments on multiple datasets and model architectures (Qwen2.5-3B to Qwen2.5-32B) demonstrate 86.2-90.6% identification accuracy and show that promoting deep alignment improves robustness against forgetting by 3.3-7.1% over baselines.", "date": "2025-12-02", "categories": []}
{"id": "2505.08528", "title": "GradMix: Gradient-based Selective Mixup for Robust Data Augmentation in Class-Incremental Learning", "abstract": "In the context of continual learning, acquiring new knowledge while maintaining previous knowledge presents a significant challenge. Existing methods often use experience replay techniques that store a small portion of previous task data for training. In experience replay approaches, data augmentation has emerged as a promising strategy to further improve the model performance by mixing limited previous task data with sufficient current task data. However, we theoretically and empirically analyze that training with mixed samples from random sample pairs may harm the knowledge of previous tasks and cause greater catastrophic forgetting. We then propose GradMix, a robust data augmentation method specifically designed for mitigating catastrophic forgetting in class-incremental learning. GradMix performs gradient-based selective mixup using a class-based criterion that mixes only samples from helpful class pairs and not from detrimental class pairs for reducing catastrophic forgetting. Our experiments on various real datasets show that GradMix outperforms data augmentation baselines in accuracy by minimizing the forgetting of previous knowledge.", "date": "2025-05-13", "categories": []}
{"id": "2512.19725", "title": "Out-of-Distribution Detection for Continual Learning: Design Principles and Benchmarking", "abstract": "Recent years have witnessed significant progress in the development of machine learning models across a wide range of fields, fueled by increased computational resources, large-scale datasets, and the rise of deep learning architectures. From malware detection to enabling autonomous navigation, modern machine learning systems have demonstrated remarkable capabilities. However, as these models are deployed in ever-changing real-world scenarios, their ability to remain reliable and adaptive over time becomes increasingly important. For example, in the real world, new malware families are continuously developed, whereas autonomous driving cars are employed in many different cities and weather conditions. Models trained in fixed settings can not respond effectively to novel conditions encountered post-deployment. In fact, most machine learning models are still developed under the assumption that training and test data are independent and identically distributed (i.i.d.), i.e., sampled from the same underlying (unknown) distribution. While this assumption simplifies model development and evaluation, it does not hold in many real-world applications, where data changes over time and unexpected inputs frequently occur. Retraining models from scratch whenever new data appears is computationally expensive, time-consuming, and impractical in resource-constrained environments. These limitations underscore the need for Continual Learning (CL), which enables models to incrementally learn from evolving data streams without forgetting past knowledge, and Out-of-Distribution (OOD) detection, which allows systems to identify and respond to novel or anomalous inputs. Jointly addressing both challenges is critical to developing robust, efficient, and adaptive AI systems.", "date": "2025-12-16", "categories": []}
{"id": "2408.10566", "title": "Overcoming Growth-Induced Forgetting in Task-Agnostic Continual Learning", "abstract": "In continual learning (CL), model growth enhances adaptability to new data. However, when model growth is applied improperly, especially in task-agnostic CL, where the entire grown model is used for inference, it can lead to severe degradation of learned knowledge, a problem we term growth-induced forgetting. Most existing methods that adopt model growth to improve adaptability often overlook the forgetting issue, resulting in compromised knowledge retention, making them unsuitable for task-agnostic settings. To promote both adaptability and knowledge retention with model growth, we identify the key: gradient and parameter sparsity. Introducing SparseGrow, which increases gradient sparsity through layer expansion and gradient gating to enable focused updates on parameters while preserving critical parameters, thus inhibiting forgetting. Moreover, it promotes parameter sparsity with sparse initialization and training, aiming at better control of model plasticity, improving adaptability over new data. Extensive experiments across diverse datasets, task-agnostic settings, and a large number of tasks demonstrate the necessity of controlled layer expansion and validate the effectiveness of SparseGrow in achieving high adaptability while minimizing forgetting in continual learning. By enabling model growth with sparsified gradients and parameters, SparseGrow paves the way for building scalable lifelong learning systems capable of continual adaptation with better knowledge retention.", "date": "2024-08-20", "categories": []}
{"id": "2511.20732", "title": "Prompt-Aware Adaptive Elastic Weight Consolidation for Continual Learning in Medical Vision-Language Models", "abstract": "Medical AI systems face catastrophic forgetting when deployed in clinical settings, where models must learn new imaging protocols while retaining prior diagnostic capabilities. This challenge is particularly acute for medical vision-language models that must preserve complex cross-modal alignments between medical images and clinical terminology across diverse imaging modalities. We introduce Prompt- Aware Adaptive Elastic Weight Consolidation (PA-EWC), a novel continual learning approach that addresses catastrophic forgetting through prompt-guided parameter specialization. Our method systematically categorizes model parameters based on their functional roles in processing visual-descriptive, spatial-guided, and medical-semantic information, enabling targeted protection of critical knowledge while allowing adaptation to new clinical requirements. PA-EWC incorporates adaptive Fisher Information computation with gradient stability analysis and develops weighted complexity metrics based on medical terminology density. We evaluate our approach across five medical imaging datasets (Kvasir-SEG, ISIC 2018, CheXlocalize, BUSI, CAMUS) representing diverse modalities including endoscopy, dermoscopy, radiography, and ultrasound. Experimental results demonstrate that PA-EWC reduces catastrophic forgetting by up to 17.58% compared to baseline methods, with performance improvements of 4.30% on chest X-ray pathology localization and 6.06% on polyp segmentation.", "date": "2025-11-25", "categories": []}
{"id": "2512.18471", "title": "The Geometry of Abstraction: Continual Learning via Recursive Quotienting", "abstract": "Continual learning systems operating in fixed-dimensional spaces face a fundamental geometric barrier: the flat manifold problem. When experience is represented as a linear trajectory in Euclidean space, the geodesic distance between temporal events grows linearly with time, forcing the required covering number to diverge. In fixed-dimensional hardware, this volume expansion inevitably forces trajectory overlap, manifesting as catastrophic interference. In this work, we propose a geometric resolution to this paradox based on Recursive Metric Contraction. We formalize abstraction not as symbolic grouping, but as a topological deformation: a quotient map that collapses the metric tensor within validated temporal neighborhoods, effectively driving the diameter of local sub-manifolds to zero. We substantiate our framework with four rigorous results. First, the Bounded Capacity Theorem establishes that recursive quotient maps allow the embedding of arbitrarily long trajectories into bounded representational volumes, trading linear metric growth for logarithmic topological depth. Second, the Topological Collapse Separability Theorem, derived via Urysohn's Lemma, proves that recursive quotienting renders non-linearly separable temporal sequences linearly separable in the limit, bypassing the need for infinite-dimensional kernel projections. Third, the Parity-Partitioned Stability Theorem solves the catastrophic forgetting problem by proving that if the state space is partitioned into orthogonal flow and scaffold manifolds, the metric deformations of active learning do not disturb the stability of stored memories. Our analysis reveals that tokens in neural architectures are physically realizable as singularities or wormholes, regions of extreme positive curvature that bridge distant points in the temporal manifold.", "date": "2025-12-20", "categories": []}
{"id": "2512.18670", "title": "Demonstration-Guided Continual Reinforcement Learning in Dynamic Environments", "abstract": "Reinforcement learning (RL) excels in various applications but struggles in dynamic environments where the underlying Markov decision process evolves. Continual reinforcement learning (CRL) enables RL agents to continually learn and adapt to new tasks, but balancing stability (preserving prior knowledge) and plasticity (acquiring new knowledge) remains challenging. Existing methods primarily address the stability-plasticity dilemma through mechanisms where past knowledge influences optimization but rarely affects the agent's behavior directly, which may hinder effective knowledge reuse and efficient learning. In contrast, we propose demonstration-guided continual reinforcement learning (DGCRL), which stores prior knowledge in an external, self-evolving demonstration repository that directly guides RL exploration and adaptation. For each task, the agent dynamically selects the most relevant demonstration and follows a curriculum-based strategy to accelerate learning, gradually shifting from demonstration-guided exploration to fully self-exploration. Extensive experiments on 2D navigation and MuJoCo locomotion tasks demonstrate its superior average performance, enhanced knowledge transfer, mitigation of forgetting, and training efficiency. The additional sensitivity analysis and ablation study further validate its effectiveness.", "date": "2025-12-21", "categories": []}
{"id": "2512.18934", "title": "When Less is More: 8-bit Quantization Improves Continual Learning in Large Language Models", "abstract": "Catastrophic forgetting poses a fundamental challenge in continual learning, particularly when models are quantized for deployment efficiency. We systematically investigate the interplay between quantization precision (FP16, INT8, INT4) and replay buffer strategies in large language models, revealing unexpected dynamics. While FP16 achieves superior initial task performance (74.44% on NLU), we observe a striking inversion on subsequent tasks: quantized models outperform FP16 by 8-15% on final task forward accuracy, with INT4 achieving nearly double FP16's performance on Code generation (40% vs 20%). Critically, even minimal replay buffers (0.1%) dramatically improve retention - increasing NLU retention after Math training from 45% to 65% across all precision levels - with INT8 consistently achieving the optimal balance between learning plasticity and knowledge retention. We hypothesize that quantization-induced noise acts as implicit regularization, preventing the overfitting to new task gradients that plagues high-precision models. These findings challenge the conventional wisdom that higher precision is always preferable, suggesting instead that INT8 quantization offers both computational efficiency and superior continual learning dynamics. Our results provide practical guidelines for deploying compressed models in continual learning scenarios: small replay buffers (1-2%) suffice for NLU tasks, while Math and Code benefit from moderate buffers (5-10%), with quantized models requiring less replay than FP16 to achieve comparable retention. Code is available at https://github.com/Festyve/LessIsMore.", "date": "2025-12-22", "categories": []}
{"id": "2408.07588", "title": "Adjusting Model Size in Continual Gaussian Processes: How Big is Big Enough?", "abstract": "Many machine learning models require setting a parameter that controls their size before training, e.g. number of neurons in DNNs, or inducing points in GPs. Increasing capacity typically improves performance until all the information from the dataset is captured. After this point, computational cost keeps increasing, without improved performance. This leads to the question \"How big is big enough?\" We investigate this problem for Gaussian processes (single-layer neural networks) in continual learning. Here, data becomes available incrementally, and the final dataset size will therefore not be known before training, preventing the use of heuristics for setting a fixed model size. We develop a method to automatically adjust model size while maintaining near-optimal performance. Our experimental procedure follows the constraint that any hyperparameters must be set without seeing dataset properties, and we show that our method performs well across diverse datasets without the need to adjust its hyperparameter, showing it requires less tuning than others.", "date": "2024-08-14", "categories": []}
{"id": "2410.06800", "title": "Low-Rank Filtering and Smoothing for Sequential Deep Learning", "abstract": "Learning multiple tasks sequentially requires neural networks to balance retaining knowledge, yet being flexible enough to adapt to new tasks. Regularizing network parameters is a common approach, but it rarely incorporates prior knowledge about task relationships, and limits information flow to future tasks only. We propose a Bayesian framework that treats the network's parameters as the state space of a nonlinear Gaussian model, unlocking two key capabilities: (1) A principled way to encode domain knowledge about task relationships, allowing, e.g., control over which layers should adapt between tasks. (2) A novel application of Bayesian smoothing, allowing task-specific models to also incorporate knowledge from models learned later. This does not require direct access to their data, which is crucial, e.g., for privacy-critical applications. These capabilities rely on efficient filtering and smoothing operations, for which we propose diagonal plus low-rank approximations of the precision matrix in the Laplace approximation (LR-LGF). Empirical results demonstrate the efficiency of LR-LGF and the benefits of the unlocked capabilities.", "date": "2024-10-09", "categories": []}
{"id": "2505.20680", "title": "Continual Learning on CLIP via Incremental Prompt Tuning with Intrinsic Textual Anchors", "abstract": "Continual learning (CL) enables deep networks to acquire new knowledge while avoiding catastrophic forgetting. The powerful generalization ability of pre-trained models (PTMs), such as the Contrastive Language-Image Pre-training (CLIP) model, has inspired a range of CL methods targeting new and specialized tasks, providing rich multi-modal embeddings that support lightweight, incremental prompt tuning. Existing methods often rely on complex designs built upon specific assumptions, such as intricate regularization schemes for prompt pools, specialized routing mechanisms, or multi-stage incrementations, that introduce additional-and possibly unnecessary-complexity, underutilizing CLIP's intrinsic capabilities. In this paper, we propose a concise CL approach for CLIP based on incremental prompt tuning that fully exploits its multi-modal structure and the stability of textual representations. Our method, Textual Prototype-guided Prompt Tuning (TPPT), introduces textual prototypes not merely as static classifiers, as in existing methods, but as stable anchors to guide the learning of visual prompts, thereby shaping the embedding space (i.e., TPPT-V). We show that our bidirectional supervision strategy enables more effective learning of new knowledge while reducing forgetting. To further close the vision-language gap during CL, we jointly optimizes visual and textual prompts (i.e., TPPT-VT). We also introduce a relational diversity regularization on the textual anchors to prevent embedding space collapse and mitigate correlated forgetting. Extensive experiments and analyses demonstrate the effectiveness of our proposed approach, highlighting the benefits of leveraging CLIP's intrinsic guidance for continual adaptation.", "date": "2025-05-27", "categories": []}
{"id": "2509.25977", "title": "Data-Free Continual Learning of Server Models in Model-Heterogeneous Cloud-Device Collaboration", "abstract": "The rise of cloud-device collaborative computing has enabled intelligent services to be delivered across distributed edge devices while leveraging centralized cloud resources. In this paradigm, federated learning (FL) has become a key enabler for privacy-preserving model training without transferring raw data from edge devices to the cloud. However, with the continuous emergence of new data and increasing model diversity, traditional federated learning faces significant challenges, including inherent issues of data heterogeneity, model heterogeneity and catastrophic forgetting, along with new challenge of knowledge misalignment. In this study, we introduce FedDCL, a novel framework designed to enable data-free continual learning of the server model in a model-heterogeneous federated setting. We leverage pre-trained diffusion models to extract lightweight class-specific prototypes, which confer a threefold data-free advantage, enabling: (1) generation of synthetic data for the current task to augment training and counteract non-IID data distributions; (2) exemplar-free generative replay for retaining knowledge from previous tasks; and (3) data-free dynamic knowledge transfer from heterogeneous devices to the cloud server.Experimental results on various datasets demonstrate the effectiveness of FedDCL, showcasing its potential to enhance the generalizability and practical applicability of federated cloud-device collaboration in dynamic settings.", "date": "2025-09-30", "categories": []}
{"id": "2508.03283", "title": "Online Continual Graph Learning", "abstract": "Continual Learning (CL) aims to incrementally acquire new knowledge while mitigating catastrophic forgetting. Within this setting, Online Continual Learning (OCL) focuses on updating models promptly and incrementally from single or small batches of observations from a data stream. Extending OCL to graph-structured data is crucial, as many real-world networks evolve over time and require timely, online predictions. However, existing continual or streaming graph learning methods typically assume access to entire graph snapshots or multiple passes over tasks, violating the efficiency constraints of the online setting. To address this gap, we introduce the Online Continual Graph Learning (OCGL) setting, which formalizes node-level continual learning on evolving graphs under strict memory and computational budgets. OCGL defines how a model incrementally processes a stream of node-level information while maintaining anytime inference and respecting resource constraints. We further establish a comprehensive benchmark comprising seven datasets and nine CL strategies, suitably adapted to the OCGL setting, enabling a standardized evaluation setup. Finally, we present a minimalistic yet competitive baseline for OCGL, inspired by our benchmarking results, that achieves strong empirical performance with high efficiency.", "date": "2025-08-05", "categories": []}
{"id": "2512.00619", "title": "Neuroscience-Inspired Memory Replay for Continual Learning: A Comparative Study of Predictive Coding and Backpropagation-Based Strategies", "abstract": "Continual learning remains a fundamental challenge in artificial intelligence, with catastrophic forgetting posing a significant barrier to deploying neural networks in dynamic environments. Inspired by biological memory consolidation mechanisms, we propose a novel framework for generative replay that leverages predictive coding principles to mitigate forgetting. We present a comprehensive comparison between predictive coding-based and backpropagation-based generative replay strategies, evaluating their effectiveness on task retention and transfer efficiency across multiple benchmark datasets. Our experimental results demonstrate that predictive coding-based replay achieves superior retention performance (average 15.3% improvement) while maintaining competitive transfer efficiency, suggesting that biologically-inspired mechanisms can offer principled solutions to continual learning challenges. The proposed framework provides insights into the relationship between biological memory processes and artificial learning systems, opening new avenues for neuroscience-inspired AI research.", "date": "2025-11-29", "categories": []}
{"id": "2512.16871", "title": "Sequencing to Mitigate Catastrophic Forgetting in Continual Learning", "abstract": "To cope with real-world dynamics, an intelligent system needs to incrementally acquire, update, and exploit knowledge throughout its lifetime. This ability, known as Continual learning, provides a foundation for AI systems to develop themselves adaptively. Catastrophic forgetting is a major challenge to the progress of Continual Learning approaches, where learning a new task usually results in a dramatic performance drop on previously learned ones. Many approaches have emerged to counteract the impact of CF. Most of the proposed approaches can be categorized into five classes: replay-based, regularization-based, optimization-based, representation-based, and architecture-based. In this work, we approach the problem from a different angle, specifically by considering the optimal sequencing of tasks as they are presented to the model. We investigate the role of task sequencing in mitigating CF and propose a method for determining the optimal task order. The proposed method leverages zero-shot scoring algorithms inspired by neural architecture search (NAS). Results demonstrate that intelligent task sequencing can substantially reduce CF. Moreover, when combined with traditional continual learning strategies, sequencing offers enhanced performance and robustness against forgetting. Additionally, the presented approaches can find applications in other fields, such as curriculum learning.", "date": "2025-12-18", "categories": []}
{"id": "2305.16424", "title": "SketchOGD: Memory-Efficient Continual Learning", "abstract": "When machine learning models are trained continually on a sequence of tasks, they are often liable to forget what they learned on previous tasks--a phenomenon known as catastrophic forgetting. Proposed solutions to catastrophic forgetting tend to involve storing information about past tasks, meaning that memory usage is a chief consideration in determining their practicality. This paper develops a memory-efficient solution to catastrophic forgetting using the idea of matrix sketching, in the context of a simple continual learning algorithm known as orthogonal gradient descent (OGD). OGD finds weight updates that aim to preserve performance on prior datapoints, using gradients of the model on those datapoints. However, since the memory cost of storing prior model gradients grows with the runtime of the algorithm, OGD is ill-suited to continual learning over long time horizons. To address this problem, we propose SketchOGD. SketchOGD employs an online sketching algorithm to compress model gradients as they are encountered into a matrix of a fixed, user-determined size. In contrast to existing memory-efficient variants of OGD, SketchOGD runs online without the need for advance knowledge of the total number of tasks, is simple to implement, and is more amenable to analysis. We provide theoretical guarantees on the approximation error of the relevant sketches under a novel metric suited to the downstream task of OGD. Experimentally, we find that SketchOGD tends to outperform current state-of-the-art variants of OGD given a fixed memory budget.", "date": "2023-05-25", "categories": []}
{"id": "2402.08182", "title": "Variational Continual Test-Time Adaptation", "abstract": "Continual Test-Time Adaptation (CTTA) task investigates effective domain adaptation under the scenario of continuous domain shifts during testing time. Due to the utilization of solely unlabeled samples, there exists significant uncertainty in model updates, leading CTTA to encounter severe error accumulation issues. In this paper, we introduce VCoTTA, a variational Bayesian approach to measure uncertainties in CTTA. At the source stage, we transform a pretrained deterministic model into a Bayesian Neural Network (BNN) via a variational warm-up strategy, injecting uncertainties into the model. During the testing time, we employ a mean-teacher update strategy using variational inference for the student model and exponential moving average for the teacher model. Our novel approach updates the student model by combining priors from both the source and teacher models. The evidence lower bound is formulated as the cross-entropy between the student and teacher models, along with the Kullback-Leibler (KL) divergence of the prior mixture. Experimental results on three datasets demonstrate the method's effectiveness in mitigating error accumulation within the CTTA framework.", "date": "2024-02-13", "categories": []}
{"id": "2403.13522", "title": "REAL: Representation Enhanced Analytic Learning for Exemplar-free Class-incremental Learning", "abstract": "Exemplar-free class-incremental learning (EFCIL) aims to mitigate catastrophic forgetting in class-incremental learning (CIL) without available historical training samples as exemplars. Compared with its exemplar-based CIL counterpart that stores exemplars, EFCIL suffers more from forgetting issues. Recently, a new EFCIL branch named Analytic Continual Learning (ACL) introduces a gradient-free paradigm via Recursive Least-Square, achieving a forgetting-resistant classifier training with a frozen backbone during CIL. However, existing ACL suffers from ineffective representations and insufficient utilization of backbone knowledge. In this paper, we propose a representation-enhanced analytic learning (REAL) to address these problems. To enhance the representation, REAL constructs a dual-stream base pretraining followed by representation enhancing distillation process. The dual-stream base pretraining combines self-supervised contrastive learning for general features and supervised learning for class-specific knowledge, followed by the representation enhancing distillation to merge both streams, enhancing representations for subsequent CIL paradigm. To utilize more knowledge from the backbone, REAL presents a feature fusion buffer to multi-layer backbone features, providing informative features for the subsequent classifier training. Our method can be incorporated into existing ACL techniques and provides more competitive performance. Empirical results demonstrate that, REAL achieves state-of-the-art performance on CIFAR-100, ImageNet-100 and ImageNet-1k benchmarks, outperforming exemplar-free methods and rivaling exemplar-based approaches.", "date": "2024-03-20", "categories": []}
{"id": "2512.15267", "title": "Distillation-Guided Structural Transfer for Continual Learning Beyond Sparse Distributed Memory", "abstract": "Sparse neural systems are gaining traction for efficient continual learning due to their modularity and low interference. Architectures such as Sparse Distributed Memory Multi-Layer Perceptrons (SDMLP) construct task-specific subnetworks via Top-K activation and have shown resilience against catastrophic forgetting. However, their rigid modularity limits cross-task knowledge reuse and leads to performance degradation under high sparsity. We propose Selective Subnetwork Distillation (SSD), a structurally guided continual learning framework that treats distillation not as a regularizer but as a topology-aligned information conduit. SSD identifies neurons with high activation frequency and selectively distills knowledge within previous Top-K subnetworks and output logits, without requiring replay or task labels. This enables structural realignment while preserving sparse modularity. Experiments on Split CIFAR-10, CIFAR-100, and MNIST demonstrate that SSD improves accuracy, retention, and representation coverage, offering a structurally grounded solution for sparse continual learning.", "date": "2025-12-17", "categories": []}
{"id": "2512.15658", "title": "PPSEBM: An Energy-Based Model with Progressive Parameter Selection for Continual Learning", "abstract": "Continual learning remains a fundamental challenge in machine learning, requiring models to learn from a stream of tasks without forgetting previously acquired knowledge. A major obstacle in this setting is catastrophic forgetting, where performance on earlier tasks degrades as new tasks are learned. In this paper, we introduce PPSEBM, a novel framework that integrates an Energy-Based Model (EBM) with Progressive Parameter Selection (PPS) to effectively address catastrophic forgetting in continual learning for natural language processing tasks. In PPSEBM, progressive parameter selection allocates distinct, task-specific parameters for each new task, while the EBM generates representative pseudo-samples from prior tasks. These generated samples actively inform and guide the parameter selection process, enhancing the model's ability to retain past knowledge while adapting to new tasks. Experimental results on diverse NLP benchmarks demonstrate that PPSEBM outperforms state-of-the-art continual learning methods, offering a promising and robust solution to mitigate catastrophic forgetting.", "date": "2025-12-17", "categories": []}
{"id": "2406.04772", "title": "REP: Resource-Efficient Prompting for Rehearsal-Free Continual Learning", "abstract": "Recent rehearsal-free continual learning (CL) methods guided by prompts achieve strong performance on vision tasks with non-stationary data but remain resource-intensive, hindering real-world edge deployment. We introduce resource-efficient prompting (REP), which improves the computational and memory efficiency of prompt-based rehearsal-free continual learning methods while minimizing accuracy trade-offs. Our approach employs swift prompt selection to refine input data using a carefully provisioned model and introduces adaptive token merging (AToM) and adaptive layer dropping (ALD) for efficient prompt updates. AToM and ALD selectively skip data and model layers while preserving task-specific features during the learning of new tasks. Extensive experiments on multiple image classification datasets demonstrate REP's superior resource efficiency over state-of-the-art rehearsal-free CL methods.", "date": "2024-06-07", "categories": []}
{"id": "2512.13706", "title": "Mitigating Catastrophic Forgetting in Mathematical Reasoning Finetuning through Mixed Training", "abstract": "When finetuning large language models for specialized tasks such as mathematical reasoning, models exhibit catastrophic forgetting, losing previously learned capabilities. We investigate this by finetuning Flan-T5-Base (250M parameters) on the DeepMind Mathematics dataset and measuring forgetting on MultiNLI. Math-only training improves mathematical accuracy from 3.1\\% to 12.0\\% but causes NLI accuracy to collapse from 81.0\\% to 16.5\\%--a 64.5 percentage point drop occurring within the first 1,000 training steps. We propose mixed training strategies that interleave mathematical and NLI examples during training. Our results demonstrate that mixed training completely eliminates catastrophic forgetting while maintaining equivalent mathematical performance: the balanced 1:1 ratio achieves 12.0\\% math accuracy (matching math-only) while preserving 86.2\\% NLI accuracy. We systematically explore mixing ratios from 1:1 to 15:1, finding that even minimal NLI exposure (6.2\\%) provides effective regularization. These findings demonstrate that specialization need not require forgetting general capabilities, with implications for scaling to larger models where mixed training may confer additional benefits beyond forgetting prevention.", "date": "2025-12-05", "categories": []}
{"id": "2512.14311", "title": "Continual Learning at the Edge: An Agnostic IIoT Architecture", "abstract": "The exponential growth of Internet-connected devices has presented challenges to traditional centralized computing systems due to latency and bandwidth limitations. Edge computing has evolved to address these difficulties by bringing computations closer to the data source. Additionally, traditional machine learning algorithms are not suitable for edge-computing systems, where data usually arrives in a dynamic and continual way. However, incremental learning offers a good solution for these settings. We introduce a new approach that applies the incremental learning philosophy within an edge-computing scenario for the industrial sector with a specific purpose: real time quality control in a manufacturing system. Applying continual learning we reduce the impact of catastrophic forgetting and provide an efficient and effective solution.", "date": "2025-12-16", "categories": []}
{"id": "2505.18101", "title": "Dynamic Dual Buffer with Divide-and-Conquer Strategy for Online Continual Learning", "abstract": "Online Continual Learning (OCL) involves sequentially arriving data and is particularly challenged by catastrophic forgetting, which significantly impairs model performance. To address this issue, we introduce a novel framework, Online Dynamic Expandable Dual Memory (ODEDM), that integrates a short-term memory for fast memory and a long-term memory structured into sub-buffers anchored by cluster prototypes, enabling the storage of diverse and category-specific samples to mitigate forgetting. We propose a novel K-means-based strategy for prototype identification and an optimal transport-based mechanism to retain critical samples, prioritising those exhibiting high similarity to their corresponding prototypes. This design preserves semantically rich information. Additionally, we propose a Divide-and-Conquer (DAC) optimisation strategy that decomposes memory updates into subproblems, thereby reducing computational overhead. ODEDM functions as a plug-and-play module that can be seamlessly integrated with existing rehearsal-based approaches. Experimental results under both standard and imbalanced OCL settings show that ODEDM consistently achieves state-of-the-art performance across multiple datasets, delivering substantial improvements over the DER family as well as recent methods such as VR-MCL and POCL.", "date": "2025-05-23", "categories": []}
{"id": "2511.09791", "title": "PANDA -- Patch And Distribution-Aware Augmentation for Long-Tailed Exemplar-Free Continual Learning", "abstract": "Exemplar-Free Continual Learning (EFCL) restricts the storage of previous task data and is highly susceptible to catastrophic forgetting. While pre-trained models (PTMs) are increasingly leveraged for EFCL, existing methods often overlook the inherent imbalance of real-world data distributions. We discovered that real-world data streams commonly exhibit dual-level imbalances, dataset-level distributions combined with extreme or reversed skews within individual tasks, creating both intra-task and inter-task disparities that hinder effective learning and generalization. To address these challenges, we propose PANDA, a Patch-and-Distribution-Aware Augmentation framework that integrates seamlessly with existing PTM-based EFCL methods. PANDA amplifies low-frequency classes by using a CLIP encoder to identify representative regions and transplanting those into frequent-class samples within each task. Furthermore, PANDA incorporates an adaptive balancing strategy that leverages prior task distributions to smooth inter-task imbalances, reducing the overall gap between average samples across tasks and enabling fairer learning with frozen PTMs. Extensive experiments and ablation studies demonstrate PANDA's capability to work with existing PTM-based CL methods, improving accuracy and reducing catastrophic forgetting.", "date": "2025-11-12", "categories": []}
{"id": "2512.11867", "title": "On the Dangers of Bootstrapping Generation for Continual Learning and Beyond", "abstract": "The use of synthetically generated data for training models is becoming a common practice. While generated data can augment the training data, repeated training on synthetic data raises concerns about distribution drift and degradation of performance due to contamination of the dataset. We investigate the consequences of this bootstrapping process through the lens of continual learning, drawing a connection to Generative Experience Replay (GER) methods. We present a statistical analysis showing that synthetic data introduces significant bias and variance into training objectives, weakening the reliability of maximum likelihood estimation. We provide empirical evidence showing that popular generative models collapse under repeated training with synthetic data. We quantify this degradation and show that state-of-the-art GER methods fail to maintain alignment in the latent space. Our findings raise critical concerns about the use of synthetic data in continual learning.", "date": "2025-12-05", "categories": []}
{"id": "2512.13340", "title": "Link-Aware Energy-Frugal Continual Learning for Fault Detection in IoT Networks", "abstract": "The use of lightweight machine learning (ML) models in internet of things (IoT) networks enables resource constrained IoT devices to perform on-device inference for several critical applications. However, the inference accuracy deteriorates due to the non-stationarity in the IoT environment and limited initial training data. To counteract this, the deployed models can be updated occasionally with new observed data samples. However, this approach consumes additional energy, which is undesirable for energy constrained IoT devices. This letter introduces an event-driven communication framework that strategically integrates continual learning (CL) in IoT networks for energy-efficient fault detection. Our framework enables the IoT device and the edge server (ES) to collaboratively update the lightweight ML model by adapting to the wireless link conditions for communication and the available energy budget. Evaluation on real-world datasets show that the proposed approach can outperform both periodic sampling and non-adaptive CL in terms of inference recall; our proposed approach achieves up to a 42.8% improvement, even under tight energy and bandwidth constraint.", "date": "2025-12-15", "categories": []}
{"id": "2410.10868", "title": "Large Continual Instruction Assistant", "abstract": "Continual Instruction Tuning (CIT) is adopted to continually instruct Large Models to follow human intent data by data. It is observed that existing gradient update would heavily destroy the performance on previous datasets during CIT process. Instead, Exponential Moving Average (EMA), owns the ability to trace previous parameters, which can aid in decreasing forgetting. Nonetheless, its stable balance weight fails to deal with the ever-changing datasets, leading to the out-of-balance between plasticity and stability. In this paper, we propose a general continual instruction tuning framework to address the challenge. Starting from the trade-off prerequisite and EMA update, we propose the plasticity and stability ideal condition. Based on Taylor expansion in the loss function, we find the optimal balance weight can be automatically determined by the gradients and learned parameters. Therefore, we propose a stable-plasticity balanced coefficient to avoid knowledge interference. Based on the semantic similarity of the instructions, we can determine whether to retrain or expand the training parameters and allocate the most suitable parameters for the testing instances. Extensive experiments across multiple continual instruction tuning benchmarks demonstrate that our approach not only enhances anti-forgetting capabilities but also significantly improves overall continual tuning performance. Our code is available at https://github.com/JingyangQiao/CoIN.", "date": "2024-10-08", "categories": []}
{"id": "2507.07712", "title": "Class-wise Balancing Data Replay for Federated Class-Incremental Learning", "abstract": "Federated Class Incremental Learning (FCIL) aims to collaboratively process continuously increasing incoming tasks across multiple clients. Among various approaches, data replay has become a promising solution, which can alleviate forgetting by reintroducing representative samples from previous tasks. However, their performance is typically limited by class imbalance, both within the replay buffer due to limited global awareness and between replayed and newly arrived classes. To address this issue, we propose a class wise balancing data replay method for FCIL (FedCBDR), which employs a global coordination mechanism for class-level memory construction and reweights the learning objective to alleviate the aforementioned imbalances. Specifically, FedCBDR has two key components: 1) the global-perspective data replay module reconstructs global representations of prior task in a privacy-preserving manner, which then guides a class-aware and importance-sensitive sampling strategy to achieve balanced replay; 2) Subsequently, to handle class imbalance across tasks, the task aware temperature scaling module adaptively adjusts the temperature of logits at both class and instance levels based on task dynamics, which reduces the model's overconfidence in majority classes while enhancing its sensitivity to minority classes. Experimental results verified that FedCBDR achieves balanced class-wise sampling under heterogeneous data distributions and improves generalization under task imbalance between earlier and recent tasks, yielding a 2%-15% Top-1 accuracy improvement over six state-of-the-art methods.", "date": "2025-07-10", "categories": []}
{"id": "2512.10969", "title": "MoB: Mixture of Bidders", "abstract": "Mixture of Experts (MoE) architectures have demonstrated remarkable success in scaling neural networks, yet their application to continual learning remains fundamentally limited by a critical vulnerability: the learned gating network itself suffers from catastrophic forgetting. We introduce Mixture of Bidders (MoB), a novel framework that reconceptualizes expert routing as a decentralized economic mechanism. MoB replaces learned gating networks with Vickrey-Clarke-Groves (VCG) auctions, where experts compete for each data batch by bidding their true cost -- a principled combination of execution cost (predicted loss) and forgetting cost (Elastic Weight Consolidation penalty). This game-theoretic approach provides three key advantages: (1) {stateless routing that is immune to catastrophic forgetting, (2) \\textbf{truthful bidding} guaranteed by dominant-strategy incentive compatibility, and (3) emergent specialization without explicit task boundaries. On Split-MNIST benchmarks, MoB achieves 88.77% average accuracy compared to 19.54% for Gated MoE and 27.96% for Monolithic EWC, representing a 4.5 times improvement over the strongest baseline. We further extend MoB with autonomous self-monitoring experts that detect their own knowledge consolidation boundaries, eliminating the need for explicit task demarcation.", "date": "2025-11-30", "categories": []}
{"id": "2512.11243", "title": "Task-Aware Multi-Expert Architecture For Lifelong Deep Learning", "abstract": "Lifelong deep learning (LDL) trains neural networks to learn sequentially across tasks while preserving prior knowledge. We propose Task-Aware Multi-Expert (TAME), a continual learning algorithm that leverages task similarity to guide expert selection and knowledge transfer. TAME maintains a pool of pretrained neural networks and activates the most relevant expert for each new task. A shared dense layer integrates features from the chosen expert to generate predictions. To reduce catastrophic forgetting, TAME uses a replay buffer that stores representative samples and embeddings from previous tasks and reuses them during training. An attention mechanism further prioritizes the most relevant stored information for each prediction. Together, these components allow TAME to adapt flexibly while retaining important knowledge across evolving task sequences. Experiments on binary classification tasks derived from CIFAR-100 show that TAME improves accuracy on new tasks while sustaining performance on earlier ones, highlighting its effectiveness in balancing adaptation and retention in lifelong learning settings.", "date": "2025-12-12", "categories": []}
{"id": "2512.11317", "title": "Condensation-Concatenation Framework for Dynamic Graph Continual Learning", "abstract": "Dynamic graphs are prevalent in real-world scenarios, where continuous structural changes induce catastrophic forgetting in graph neural networks (GNNs). While continual learning has been extended to dynamic graphs, existing methods overlook the effects of topological changes on existing nodes. To address it, we propose a novel framework for continual learning on dynamic graphs, named Condensation-Concatenation-based Continual Learning (CCC). Specifically, CCC first condenses historical graph snapshots into compact semantic representations while aiming to preserve the original label distribution and topological properties. Then it concatenates these historical embeddings with current graph representations selectively. Moreover, we refine the forgetting measure (FM) to better adapt to dynamic graph scenarios by quantifying the predictive performance degradation of existing nodes caused by structural updates. CCC demonstrates superior performance over state-of-the-art baselines across four real-world datasets in extensive experiments.", "date": "2025-12-12", "categories": []}
{"id": "2512.11668", "title": "Bridging Streaming Continual Learning via In-Context Large Tabular Models", "abstract": "In streaming scenarios, models must learn continuously, adapting to concept drifts without erasing previously acquired knowledge. However, existing research communities address these challenges in isolation. Continual Learning (CL) focuses on long-term retention and mitigating catastrophic forgetting, often without strict real-time constraints. Stream Learning (SL) emphasizes rapid, efficient adaptation to high-frequency data streams, but typically neglects forgetting. Recent efforts have tried to combine these paradigms, yet no clear algorithmic overlap exists. We argue that large in-context tabular models (LTMs) provide a natural bridge for Streaming Continual Learning (SCL). In our view, unbounded streams should be summarized on-the-fly into compact sketches that can be consumed by LTMs. This recovers the classical SL motivation of compressing massive streams with fixed-size guarantees, while simultaneously aligning with the experience-replay desiderata of CL. To clarify this bridge, we show how the SL and CL communities implicitly adopt a divide-to-conquer strategy to manage the tension between plasticity (performing well on the current distribution) and stability (retaining past knowledge), while also imposing a minimal complexity constraint that motivates diversification (avoiding redundancy in what is stored) and retrieval (re-prioritizing past information when needed). Within this perspective, we propose structuring SCL with LTMs around two core principles of data selection for in-context learning: (1) distribution matching, which balances plasticity and stability, and (2) distribution compression, which controls memory size through diversification and retrieval mechanisms.", "date": "2025-12-12", "categories": []}
{"id": "2504.20566", "title": "Balanced Online Class-Incremental Learning via Dual Classifiers", "abstract": "Online class-incremental learning (OCIL) focuses on gradually learning new classes (called plasticity) from a stream of data in a single-pass, while concurrently preserving knowledge of previously learned classes (called stability). The primary challenge in OCIL lies in maintaining a good balance between the knowledge of old and new classes within the continually updated model. Most existing methods rely on explicit knowledge interaction through experience replay, and often employ exclusive training separation to address bias problems. Nevertheless, it still remains a big challenge to achieve a well-balanced learner, as these methods often exhibit either reduced plasticity or limited stability due to difficulties in continually integrating knowledge in the OCIL setting. In this paper, we propose a novel replay-based method, called Balanced Inclusive Separation for Online iNcremental learning (BISON), which can achieve both high plasticity and stability, thus ensuring more balanced performance in OCIL. Our BISON method proposes an inclusive training separation strategy using dual classifiers so that knowledge from both old and new classes can effectively be integrated into the model, while introducing implicit approaches for transferring knowledge across the two classifiers. Extensive experimental evaluations over three widely-used OCIL benchmark datasets demonstrate the superiority of BISON, showing more balanced yet better performance compared to state-of-the-art replay-based OCIL methods.", "date": "2025-04-29", "categories": []}
{"id": "2511.03270", "title": "SCALE: Upscaled Continual Learning of Large Language Models", "abstract": "We revisit continual pre-training for large language models and argue that progress now depends more on scaling the right structure than on scaling parameters alone. We introduce SCALE, a width upscaling architecture that inserts lightweight expansion into linear modules while freezing all pre-trained parameters. This preserves the residual and attention topologies and increases capacity without perturbing the base model's original functionality. SCALE is guided by two principles: Persistent Preservation, which maintains the base model's behavior via preservation-oriented initialization and freezing of the pre-trained weights, and Collaborative Adaptation, which selectively trains a subset of expansion components to acquire new knowledge with minimal interference. We instantiate these ideas as SCALE-Preserve (preservation-first), SCALE-Adapt (adaptation-first), and SCALE-Route, an optional routing extension that performs token-level routing between preservation and adaptation heads. On a controlled synthetic biography benchmark, SCALE mitigates the severe forgetting observed with depth expansion while still acquiring new knowledge. In continual pre-training on a Korean corpus, SCALE variants achieve less forgetting on English evaluations and competitive gains on Korean benchmarks, with these variants offering the best overall stability-plasticity trade-off. Accompanying analysis clarifies when preservation provably holds and why the interplay between preservation and adaptation stabilizes optimization compared to standard continual learning setups.", "date": "2025-11-05", "categories": []}
{"id": "2509.19698", "title": "A Unified Noise-Curvature View of Loss of Trainability", "abstract": "Loss of trainability refers to a phenomenon in continual learning where parameter updates no longer make progress on the optimization objective, so accuracy stalls or degrades as the learning problem changes over time. In this paper, we analyze loss of trainability through an optimization lens and find that the phenomenon is not reliably predicted by existing individual indicators such as Hessian rank, sharpness level, weight or gradient norms, gradient-to-parameter ratios, and unit-sign entropy. Motivated by our analysis, we introduce two complementary indicators: a batch-size-aware gradient-noise bound and a curvature volatility-controlled bound. We then combine these two indicators into a per-layer adaptive noise threshold on the effective step-size that anticipates trainability behavior. Using this insight, we propose a step-size scheduler that keeps each layer's effective parameter update below this bound, thereby avoiding loss of trainability. We demonstrate that our scheduler can improve the accuracy maintained by previously proposed approaches, such as concatenated ReLU (CReLU), Wasserstein regularizer, and L2 weight decay. Surprisingly, our scheduler produces adaptive step-size trajectories that, without tuning, mirror the manually engineered step-size decay schedules.", "date": "2025-09-24", "categories": []}
{"id": "2512.05802", "title": "Bring Your Dreams to Life: Continual Text-to-Video Customization", "abstract": "Customized text-to-video generation (CTVG) has recently witnessed great progress in generating tailored videos from user-specific text. However, most CTVG methods assume that personalized concepts remain static and do not expand incrementally over time. Additionally, they struggle with forgetting and concept neglect when continuously learning new concepts, including subjects and motions. To resolve the above challenges, we develop a novel Continual Customized Video Diffusion (CCVD) model, which can continuously learn new concepts to generate videos across various text-to-video generation tasks by tackling forgetting and concept neglect. To address catastrophic forgetting, we introduce a concept-specific attribute retention module and a task-aware concept aggregation strategy. They can capture the unique characteristics and identities of old concepts during training, while combining all subject and motion adapters of old concepts based on their relevance during testing. Besides, to tackle concept neglect, we develop a controllable conditional synthesis to enhance regional features and align video contexts with user conditions, by incorporating layer-specific region attention-guided noise estimation. Extensive experimental comparisons demonstrate that our CCVD outperforms existing CTVG baselines on both the DreamVideo and Wan 2.1 backbones. The code is available at https://github.com/JiahuaDong/CCVD.", "date": "2025-12-05", "categories": []}
{"id": "2512.08960", "title": "Resolving Conflicts in Lifelong Learning via Aligning Updates in Subspaces", "abstract": "Low-Rank Adaptation (LoRA) enables efficient Continual Learning but often suffers from catastrophic forgetting due to destructive interference between tasks. Our analysis reveals that this degradation is primarily driven by antagonistic directional updates where new task gradients directly oppose the historical weight trajectory. To address this, we propose PS-LoRA (Parameter Stability LoRA), a framework designed to resolve conflicts by aligning updates within the optimization subspace. Our approach employs a dual-regularization objective that penalizes conflicting directions and constrains magnitude deviations to ensure consistency with prior knowledge. Additionally, we implement a magnitude-based merging strategy to consolidate sequential adapters into a robust representation without retraining. Experiments on NLP and Vision benchmarks show that PS-LoRA outperforms state-of-the-art methods by preserving the stability of learned representations while efficiently adapting to new domains.", "date": "2025-11-28", "categories": []}
{"id": "2512.09441", "title": "Representation Calibration and Uncertainty Guidance for Class-Incremental Learning based on Vision Language Model", "abstract": "Class-incremental learning requires a learning system to continually learn knowledge of new classes and meanwhile try to preserve previously learned knowledge of old classes. As current state-of-the-art methods based on Vision-Language Models (VLMs) still suffer from the issue of differentiating classes across learning tasks. Here a novel VLM-based continual learning framework for image classification is proposed. In this framework, task-specific adapters are added to the pre-trained and frozen image encoder to learn new knowledge, and a novel cross-task representation calibration strategy based on a mixture of light-weight projectors is used to help better separate all learned classes in a unified feature space, alleviating class confusion across tasks. In addition, a novel inference strategy guided by prediction uncertainty is developed to more accurately select the most appropriate image feature for class prediction. Extensive experiments on multiple datasets under various settings demonstrate the superior performance of our method compared to existing ones.", "date": "2025-12-10", "categories": []}
{"id": "2512.07981", "title": "CIP-Net: Continual Interpretable Prototype-based Network", "abstract": "Continual learning constrains models to learn new tasks over time without forgetting what they have already learned. A key challenge in this setting is catastrophic forgetting, where learning new information causes the model to lose its performance on previous tasks. Recently, explainable AI has been proposed as a promising way to better understand and reduce forgetting. In particular, self-explainable models are useful because they generate explanations during prediction, which can help preserve knowledge. However, most existing explainable approaches use post-hoc explanations or require additional memory for each new task, resulting in limited scalability. In this work, we introduce CIP-Net, an exemplar-free self-explainable prototype-based model designed for continual learning. CIP-Net avoids storing past examples and maintains a simple architecture, while still providing useful explanations and strong performance. We demonstrate that CIPNet achieves state-of-the-art performances compared to previous exemplar-free and self-explainable methods in both task- and class-incremental settings, while bearing significantly lower memory-related overhead. This makes it a practical and interpretable solution for continual learning.", "date": "2025-12-08", "categories": []}
{"id": "2412.05012", "title": "SAMCL: Empowering SAM to Continually Learn from Dynamic Domains with Extreme Storage Efficiency", "abstract": "Segment Anything Model (SAM) struggles in open-world scenarios with diverse domains. In such settings, naive fine-tuning with a well-designed learning module is inadequate and often causes catastrophic forgetting issue when learning incrementally. To address this issue, we propose a novel continual learning (CL) method for SAM, termed SAMCL. Rather than relying on a fixed learning module, our method decomposes incremental knowledge into separate modules and trains a selector to choose the appropriate one during inference. However, this intuitive design introduces two key challenges: ensuring effective module learning and selection, and managing storage as tasks accumulate. To tackle these, we introduce two components: AugModule and Module Selector. AugModule reduces the storage of the popular LoRA learning module by sharing parameters across layers while maintaining accuracy. It also employs heatmaps-generated from point prompts-to further enhance domain adaptation with minimal additional cost. Module Selector leverages the observation that SAM's embeddings can effectively distinguish domains, enabling high selection accuracy by training on low-consumed embeddings instead of raw images. Experiments show that SAMCL outperforms state-of-the-art methods, achieving only 0.19% forgetting and at least 2.5% gain on unseen domains. Each AugModule requires just 0.233 MB, reducing storage by at least 24.3% over other fine-tuning approaches. The buffer storage for Module Selector is further reduced by up to 256$\\times$.", "date": "2024-12-06", "categories": []}
{"id": "2508.17169", "title": "ONG: Orthogonal Natural Gradient Descent", "abstract": "Orthogonal Gradient Descent (OGD) has emerged as a powerful method for continual learning. However, its Euclidean projections do not leverage the underlying information-geometric structure of the problem, which can lead to suboptimal convergence in learning tasks. To address this, we propose incorporating the natural gradient into OGD and present \\textbf{ONG (Orthogonal Natural Gradient Descent)}. ONG preconditions each new task-specific gradient with an efficient EKFAC approximation of the inverse Fisher information matrix, yielding updates that follow the steepest descent direction under a Riemannian metric. To preserve performance on previously learned tasks, ONG projects these natural gradients onto the orthogonal complement of prior tasks' natural gradients. We provide an initial theoretical justification for this procedure, introduce the Orthogonal Natural Gradient Descent (ONG) algorithm, and present preliminary results on the Permuted and Rotated MNIST benchmarks. Our preliminary results, however, indicate that a naive combination of natural gradients and orthogonal projections has potential issues. This finding has motivated continued future work focused on robustly reconciling these geometric perspectives to develop a continual learning method, establishing a more rigorous theoretical foundation with formal convergence guarantees, and extending empirical validation to large-scale continual learning benchmarks.", "date": "2025-08-24", "categories": []}
{"id": "2510.18874", "title": "Retaining by Doing: The Role of On-Policy Data in Mitigating Forgetting", "abstract": "Adapting language models (LMs) to new tasks via post-training carries the risk of degrading existing capabilities -- a phenomenon classically known as catastrophic forgetting. In this paper, toward identifying guidelines for mitigating this phenomenon, we systematically compare the forgetting patterns of two widely adopted post-training methods: supervised fine-tuning (SFT) and reinforcement learning (RL). Our experiments reveal a consistent trend across LM families (Llama, Qwen) and tasks (instruction following, general knowledge, and arithmetic reasoning): RL leads to less forgetting than SFT while achieving comparable or higher target task performance. To investigate the cause for this difference, we consider a simplified setting in which the LM is modeled as a mixture of two distributions, one corresponding to prior knowledge and the other to the target task. We identify that the mode-seeking nature of RL, which stems from its use of on-policy data, enables keeping prior knowledge intact when learning the target task. We then verify this insight by demonstrating that the use on-policy data underlies the robustness of RL to forgetting in practical settings, as opposed to other algorithmic choices such as the KL regularization or advantage estimation. Lastly, as a practical implication, our results highlight the potential of mitigating forgetting using approximately on-policy data, which can be substantially more efficient to obtain than fully on-policy data.", "date": "2025-10-21", "categories": []}
{"id": "2512.03125", "title": "Mitigating Intra- and Inter-modal Forgetting in Continual Learning of Unified Multimodal Models", "abstract": "Unified Multimodal Generative Models (UMGMs) unify visual understanding and image generation within a single autoregressive framework. However, their ability to continually learn new tasks is severely hindered by catastrophic forgetting, both within a modality (intra-modal) and across modalities (inter-modal). While intra-modal forgetting has been studied in prior continual learning (CL) work, inter-modal forgetting remains largely unexplored. In this paper, we identify and empirically validate this phenomenon in UMGMs and provide a theoretical explanation rooted in gradient conflict between modalities. To address both intra- and inter-modal forgetting, we propose Modality-Decoupled Experts (MoDE), a lightweight and scalable architecture that isolates modality-specific updates to mitigate the gradient conflict and leverages knowledge distillation to prevent catastrophic forgetting and preserve pre-trained capabilities. Unlike previous CL methods that remain modality-coupled and suffer from modality gradient conflict, MoDE explicitly decouples modalities to prevent interference. Experiments across diverse benchmarks demonstrate that MoDE significantly mitigates both inter- and intra-modal forgetting, outperforming prior CL baselines in unified multimodal generation settings. Codes will be publicly available: https://github.com/Christina200/MoDE-official.git", "date": "2025-12-02", "categories": []}
{"id": "2512.03882", "title": "Automatic Attack Discovery for Few-Shot Class-Incremental Learning via Large Language Models", "abstract": "Few-shot class incremental learning (FSCIL) is a more realistic and challenging paradigm in continual learning to incrementally learn unseen classes and overcome catastrophic forgetting on base classes with only a few training examples. Previous efforts have primarily centered around studying more effective FSCIL approaches. By contrast, less attention was devoted to thinking the security issues in contributing to FSCIL. This paper aims to provide a holistic study of the impact of attacks on FSCIL. We first derive insights by systematically exploring how human expert-designed attack methods (i.e., PGD, FGSM) affect FSCIL. We find that those methods either fail to attack base classes, or suffer from huge labor costs due to relying on huge expert knowledge. This highlights the need to craft a specialized attack method for FSCIL. Grounded in these insights, in this paper, we propose a simple yet effective ACraft method to automatically steer and discover optimal attack methods targeted at FSCIL by leveraging Large Language Models (LLMs) without human experts. Moreover, to improve the reasoning between LLMs and FSCIL, we introduce a novel Proximal Policy Optimization (PPO) based reinforcement learning to optimize learning, making LLMs generate better attack methods in the next generation by establishing positive feedback. Experiments on mainstream benchmarks show that our ACraft significantly degrades the performance of state-of-the-art FSCIL methods and dramatically beyond human expert-designed attack methods while maintaining the lowest costs of attack.", "date": "2025-12-03", "categories": []}
{"id": "2411.17766", "title": "Dual Prototypes for Adaptive Pre-Trained Model in Class-Incremental Learning", "abstract": "Class-incremental learning (CIL) aims to learn new classes while retaining previous knowledge. Although pre-trained model (PTM) based approaches show strong performance, directly fine-tuning PTMs on incremental task streams often causes renewed catastrophic forgetting. This paper proposes a Dual-Prototype Network with Task-wise Adaptation (DPTA) for PTM-based CIL. For each incremental learning task, an adapter module is built to fine-tune the PTM, where the center-adapt loss forces the representation to be more centrally clustered and class separable. The dual prototype network improves the prediction process by enabling test-time adapter selection, where the raw prototypes deduce several possible task indexes of test samples to select suitable adapter modules for PTM, and the augmented prototypes that could separate confusable classes are utilized to determine the final result. Experiments on multiple benchmarks show that DPTA consistently surpasses recent methods by 1\\% - 5\\%. Notably, on the VTAB dataset, it achieves approximately 3\\% improvement over state-of-the-art methods. The code is open-sourced in https://github.com/Yorkxzm/DPTA}", "date": "2024-11-26", "categories": []}
{"id": "2512.00694", "title": "Affordance-First Decomposition for Continual Learning in Video-Language Understanding", "abstract": "Continual learning for video--language understanding is increasingly important as models face non-stationary data, domains, and query styles, yet prevailing solutions blur what should stay stable versus what should adapt, rely on static routing/capacity, or require replaying past videos. We aim to explicitly specify where stability lives and where plasticity should be focused under realistic memory and privacy constraints. We introduce Affordance-First Decomposition (AFD): videos are mapped to slowly varying affordance tokens that form a shared, time-aligned substrate, while a lightweight, query-routed, conflict-aware scheduler concentrates adaptation and grows capacity only when needed. The substrate is stabilized via weak alignment and teacher consistency, and training uses question-only replay. AFD achieves state-of-the-art across protocols: 51.6% average accuracy with -1.8% forgetting on domain-incremental VideoQA, ViLCo R@1@0.5 of 29.6% (MQ) and 20.7% (NLQ) with 18.4% stAP@0.25 (VQ), and 39.5% accuracy with -1.6% forgetting on time-incremental iVQA. Overall, AFD offers an explicit, interpretable split between a stable interaction-centered substrate and targeted adaptation.", "date": "2025-11-30", "categories": []}
{"id": "2512.00940", "title": "Memory-Integrated Reconfigurable Adapters: A Unified Framework for Settings with Multiple Tasks", "abstract": "Organisms constantly pivot between tasks such as evading predators, foraging, traversing rugged terrain, and socializing, often within milliseconds. Remarkably, they preserve knowledge of once-learned environments sans catastrophic forgetting, a phenomenon neuroscientists hypothesize, is due to a singular neural circuitry dynamically overlayed by neuromodulatory agents such as dopamine and acetylcholine. In parallel, deep learning research addresses analogous challenges via domain generalization (DG) and continual learning (CL), yet these methods remain siloed, despite the brains ability to perform them seamlessly. In particular, prior work has not explored architectures involving associative memories (AMs), which are an integral part of biological systems, to jointly address these tasks. We propose Memory-Integrated Reconfigurable Adapters (MIRA), a unified framework that integrates Hopfield-style associative memory modules atop a shared backbone. Associative memory keys are learned post-hoc to index and retrieve an affine combination of stored adapter updates for any given task or domain on a per-sample basis. By varying only the task-specific objectives, we demonstrate that MIRA seamlessly accommodates domain shifts and sequential task exposures under one roof. Empirical evaluations on standard benchmarks confirm that our AM-augmented architecture significantly enhances adaptability and retention: in DG, MIRA achieves SoTA out-of-distribution accuracy, and in incremental learning settings, it outperforms architectures explicitly designed to handle catastrophic forgetting using generic CL algorithms. By unifying adapter-based modulation with biologically inspired associative memory, MIRA delivers rapid task switching and enduring knowledge retention in a single extensible architecture, charting a path toward more versatile and memory-augmented AI systems.", "date": "2025-11-30", "categories": []}
{"id": "2512.01818", "title": "Forget Less, Retain More: A Lightweight Regularizer for Rehearsal-Based Continual Learning", "abstract": "Deep neural networks suffer from catastrophic forgetting, where performance on previous tasks degrades after training on a new task. This issue arises due to the model's tendency to overwrite previously acquired knowledge with new information. We present a novel approach to address this challenge, focusing on the intersection of memory-based methods and regularization approaches. We formulate a regularization strategy, termed Information Maximization (IM) regularizer, for memory-based continual learning methods, which is based exclusively on the expected label distribution, thus making it class-agnostic. As a consequence, IM regularizer can be directly integrated into various rehearsal-based continual learning methods, reducing forgetting and favoring faster convergence. Our empirical validation shows that, across datasets and regardless of the number of tasks, our proposed regularization strategy consistently improves baseline performance at the expense of a minimal computational overhead. The lightweight nature of IM ensures that it remains a practical and scalable solution, making it applicable to real-world continual learning scenarios where efficiency is paramount. Finally, we demonstrate the data-agnostic nature of our regularizer by applying it to video data, which presents additional challenges due to its temporal structure and higher memory requirements. Despite the significant domain gap, our experiments show that IM regularizer also improves the performance of video continual learning methods.", "date": "2025-12-01", "categories": []}
{"id": "2512.01890", "title": "Elastic Weight Consolidation for Knowledge Graph Continual Learning: An Empirical Evaluation", "abstract": "Knowledge graphs (KGs) require continual updates as new information emerges, but neural embedding models suffer from catastrophic forgetting when learning new tasks sequentially. We evaluate Elastic Weight Consolidation (EWC), a regularization-based continual learning method, on KG link prediction using TransE embeddings on FB15k-237. Across multiple experiments with five random seeds, we find that EWC reduces catastrophic forgetting from 12.62% to 6.85%, a 45.7% reduction compared to naive sequential training. We observe that the task partitioning strategy affects the magnitude of forgetting: relation-based partitioning (grouping triples by relation type) exhibits 9.8 percentage points higher forgetting than randomly partitioned tasks (12.62% vs 2.81%), suggesting that task construction influences evaluation outcomes. While focused on a single embedding model and dataset, our results demonstrate that EWC effectively mitigates catastrophic forgetting in KG continual learning and highlight the importance of evaluation protocol design.", "date": "2025-12-01", "categories": []}
{"id": "2511.22367", "title": "SuRe: Surprise-Driven Prioritised Replay for Continual LLM Learning", "abstract": "Continual learning, one's ability to adapt to a sequence of tasks without forgetting previously acquired knowledge, remains a major challenge in machine learning and a key gap between artificial and human intelligence. While regularisation and replay perform well in vision, they lag behind multi-task learning for large language models (LLMs), especially at scale with many tasks. We revisit replay and argue that two failure modes drive this gap: selection (what to rehearse) and integration (how to consolidate new knowledge). To address selection, we propose Surprise-prioritised Replay (SuRe), a simple, architecture-agnostic rule that ranks and stores the most surprising (high Negative Log-Likelihood) sequences. SuRe achieves state-of-the-art performance in the Large Number of Tasks (LNT) setting and delivers the best overall average across both Standard CL and LNT benchmarks. To address integration, we add a dual-learner design with fast and slow LoRA adapters merged via an exponential moving average (EMA), enabling rapid adaptation while stabilising long-term knowledge. Combining SuRe with the dual learner yields further gains, including improvements of up to +5 accuracy points on LNT over prior SOTA. Ablation studies confirm that our proposed method remains robust under reduced replay frequency and small buffer size, demonstrating both effectiveness and sample efficiency. Taken together, our results establish replay as a strong baseline for continual LLM fine-tuning and demonstrate that surprise-based selection and slow-weight consolidation are complementary components for mitigating catastrophic forgetting.", "date": "2025-11-27", "categories": []}
{"id": "2511.22615", "title": "Stable-Drift: A Patient-Aware Latent Drift Replay Method for Stabilizing Representations in Continual Learning", "abstract": "When deep learning models are sequentially trained on new data, they tend to abruptly lose performance on previously learned tasks, a critical failure known as catastrophic forgetting. This challenge severely limits the deployment of AI in medical imaging, where models must continually adapt to data from new hospitals without compromising established diagnostic knowledge. To address this, we introduce a latent drift-guided replay method that identifies and replays samples with high representational instability. Specifically, our method quantifies this instability via latent drift, the change in a sample internal feature representation after naive domain adaptation. To ensure diversity and clinical relevance, we aggregate drift at the patient level, our memory buffer stores the per patient slices exhibiting the greatest multi-layer representation shift. Evaluated on a cross-hospital COVID-19 CT classification task using state-of-the-art CNN and Vision Transformer backbones, our method substantially reduces forgetting compared to naive fine-tuning and random replay. This work highlights latent drift as a practical and interpretable replay signal for advancing robust continual learning in real world medical settings.", "date": "2025-11-27", "categories": []}
{"id": "2511.23436", "title": "Towards Continuous Intelligence Growth: Self-Training, Continual Learning, and Dual-Scale Memory in SuperIntelliAgent", "abstract": "We introduce SuperIntelliAgent, an agentic learning framework that couples a trainable small diffusion model (the learner) with a frozen large language model (the verifier) to enable continual intelligence growth through self-supervised interaction. Unlike conventional supervised fine-tuning, SuperIntelliAgent learns autonomously without annotation: the learner generates candidate outputs, the verifier evaluates them through step-by-step reasoning, and their interaction produces chosen/rejected pairs for Direct Preference Optimization (DPO). This converts each input into a pseudo-training signal for continual improvement. The framework integrates dual-scale memory: short-term in-context memory that preserves reasoning traces across refinement cycles, and long-term memory that consolidates acquired knowledge through lightweight on-the-fly fine-tuning. A replay buffer retains samples that show verifiable progress and replays them as auxiliary supervision, reinforcing recent learning while forming adaptive curricula. SuperIntelliAgent is infrastructure-agnostic and can be plugged into existing agentic frameworks while turning ordinary inference loops into a lifelong optimization process. We posit that pairing a trainable learner with a reasoning-capable verifier forms a minimal reliable unit of growing intelligence, as paired feedback and partial-history replay yield richer learning curricula and stronger preference alignment. With a small number of automatically generated DPO pairs, the learner improves across all benchmarks, indicating that this mechanism provides a promising direction for continual intelligence accumulation and real-world deployment.", "date": "2025-11-28", "categories": []}
{"id": "2503.00515", "title": "Class-Independent Increment: An Efficient Approach for Multi-label Class-Incremental Learning", "abstract": "Current research on class-incremental learning primarily focuses on single-label classification tasks. However, real-world applications often involve multi-label scenarios, such as image retrieval and medical imaging. Therefore, this paper focuses on the challenging yet practical multi-label class-incremental learning (MLCIL) problem. In addition to the challenge of catastrophic forgetting, MLCIL encounters issues related to feature confusion, encompassing inter-session and intra-feature confusion. To address these problems, we propose a novel MLCIL approach called class-independent increment (CLIN). Specifically, in contrast to existing methods that extract image-level features, we propose a class-independent incremental network (CINet) to extract multiple class-level embeddings for multi-label samples. It learns and preserves the knowledge of different classes by constructing class-specific tokens. On this basis, we develop two novel loss functions, optimizing the learning of class-specific tokens and class-level embeddings, respectively. These losses aim to distinguish between new and old classes, further alleviating the problem of feature confusion. Extensive experiments on MS-COCO and PASCAL VOC datasets demonstrate the effectiveness of our method for improving recognition performance and mitigating forgetting on various MLCIL tasks.", "date": "2025-03-01", "categories": []}
{"id": "2511.21490", "title": "Merge and Bound: Direct Manipulations on Weights for Class Incremental Learning", "abstract": "We present a novel training approach, named Merge-and-Bound (M&B) for Class Incremental Learning (CIL), which directly manipulates model weights in the parameter space for optimization. Our algorithm involves two types of weight merging: inter-task weight merging and intra-task weight merging. Inter-task weight merging unifies previous models by averaging the weights of models from all previous stages. On the other hand, intra-task weight merging facilitates the learning of current task by combining the model parameters within current stage. For reliable weight merging, we also propose a bounded update technique that aims to optimize the target model with minimal cumulative updates and preserve knowledge from previous tasks; this strategy reveals that it is possible to effectively obtain new models near old ones, reducing catastrophic forgetting. M&B is seamlessly integrated into existing CIL methods without modifying architecture components or revising learning objectives. We extensively evaluate our algorithm on standard CIL benchmarks and demonstrate superior performance compared to state-of-the-art methods.", "date": "2025-11-26", "categories": []}
{"id": "2511.21652", "title": "Continual Error Correction on Low-Resource Devices", "abstract": "The proliferation of AI models in everyday devices has highlighted a critical challenge: prediction errors that degrade user experience. While existing solutions focus on error detection, they rarely provide efficient correction mechanisms, especially for resource-constrained devices. We present a novel system enabling users to correct AI misclassifications through few-shot learning, requiring minimal computational resources and storage. Our approach combines server-side foundation model training with on-device prototype-based classification, enabling efficient error correction through prototype updates rather than model retraining. The system consists of two key components: (1) a server-side pipeline that leverages knowledge distillation to transfer robust feature representations from foundation models to device-compatible architectures, and (2) a device-side mechanism that enables ultra-efficient error correction through prototype adaptation. We demonstrate our system's effectiveness on both image classification and object detection tasks, achieving over 50% error correction in one-shot scenarios on Food-101 and Flowers-102 datasets while maintaining minimal forgetting (less than 0.02%) and negligible computational overhead. Our implementation, validated through an Android demonstration app, proves the system's practicality in real-world scenarios.", "date": "2025-11-26", "categories": []}
{"id": "2511.19490", "title": "Generative Model-Aided Continual Learning for CSI Feedback in FDD mMIMO-OFDM Systems", "abstract": "Deep autoencoder (DAE) frameworks have demonstrated their effectiveness in reducing channel state information (CSI) feedback overhead in massive multiple-input multiple-output (mMIMO) orthogonal frequency division multiplexing (OFDM) systems. However, existing CSI feedback models struggle to adapt to dynamic environments caused by user mobility, requiring retraining when encountering new CSI distributions. Moreover, returning to previously encountered environments often leads to performance degradation due to catastrophic forgetting. Continual learning involves enabling models to incorporate new information while maintaining performance on previously learned tasks. To address these challenges, we propose a generative adversarial network (GAN)-based learning approach for CSI feedback. By using a GAN generator as a memory unit, our method preserves knowledge from past environments and ensures consistently high performance across diverse scenarios without forgetting. Simulation results show that the proposed approach enhances the generalization capability of the DAE framework while maintaining low memory overhead. Furthermore, it can be seamlessly integrated with other advanced CSI feedback models, highlighting its robustness and adaptability.", "date": "2025-11-23", "categories": []}
{"id": "2511.19491", "title": "OpenCML: End-to-End Framework of Open-world Machine Learning to Learn Unknown Classes Incrementally", "abstract": "Open-world machine learning is an emerging technique in artificial intelligence, where conventional machine learning models often follow closed-world assumptions, which can hinder their ability to retain previously learned knowledge for future tasks. However, automated intelligence systems must learn about novel classes and previously known tasks. The proposed model offers novel learning classes in an open and continuous learning environment. It consists of two different but connected tasks. First, it discovers unknown classes in the data and creates novel classes; next, it learns how to perform class incrementally for each new class. Together, they enable continual learning, allowing the system to expand its understanding of the data and improve over time. The proposed model also outperformed existing approaches in open-world learning. Furthermore, it demonstrated strong performance in continuous learning, achieving a highest average accuracy of 82.54% over four iterations and a minimum accuracy of 65.87%.", "date": "2025-11-23", "categories": []}
{"id": "2511.17668", "title": "MedPEFT-CL: Dual-Phase Parameter-Efficient Continual Learning with Medical Semantic Adapter and Bidirectional Memory Consolidation", "abstract": "Medical vision-language segmentation models suffer from catastrophic forgetting when adapting to new anatomical structures, requiring complete retraining that limits their clinical deployment. Although continual learning approaches have been studied for various applications, targeted research on continual learning approaches specifically designed for medical vision-language tasks remains underexplored. We propose MedPEFT-CL, a parameter-efficient continual learning framework that addresses both efficient learning of new tasks and preservation of previous knowledge through a dual-phase architecture based on CLIPSeg. Our dual-phase architecture features an adaptive learning phase that employs semantic similarity-based adapter allocation and parameter-efficient fine-tuning for medical tasks through prompt similarity analysis, and a knowledge consolidation phase employing bi-directional Fisher-memory coordination. This creates a reinforcing cycle: consolidation directs replay priorities while new tasks provide challenging samples that improve retention strategies. Our key contributions are: (1) a semantic-driven adapter allocation mechanism that enables efficient learning of new medical tasks, (2) a bi-modal LoRA adaptation that significantly reduces trainable parameters while maintaining cross-modal learning, and (3) bidirectional Fisher-memory coordination that prevents catastrophic forgetting from previous medical tasks. Extensive experiments across diverse medical datasets demonstrate superior forgetting mitigation and performance retention with minimal parameter overhead, making the framework effective for continual learning in medical vision-language scenarios.", "date": "2025-11-21", "categories": []}
{"id": "2511.17829", "title": "Unified Class and Domain Incremental Learning with Mixture of Experts for Indoor Localization", "abstract": "Indoor localization using machine learning has gained traction due to the growing demand for location-based services. However, its long-term reliability is hindered by hardware/software variations across mobile devices, which shift the model's input distribution to create domain shifts. Further, evolving indoor environments can introduce new locations over time, expanding the output space to create class shifts, making static machine learning models ineffective over time. To address these challenges, we propose a novel unified continual learning framework for indoor localization called MOELO that, for the first time, jointly addresses domain-incremental and class-incremental learning scenarios. MOELO enables a lightweight, robust, and adaptive localization solution that can be deployed on resource-limited mobile devices and is capable of continual learning in dynamic, heterogeneous real-world settings. This is made possible by a mixture-of-experts architecture, where experts are incrementally trained per region and selected through an equiangular tight frame based gating mechanism ensuring efficient routing, and low-latency inference, all within a compact model footprint. Experimental evaluations show that MOELO achieves improvements of up to 25.6x in mean localization error, 44.5x in worst-case localization error, and 21.5x lesser forgetting compared to state-of-the-art frameworks across diverse buildings, mobile devices, and learning scenarios.", "date": "2025-11-21", "categories": []}
{"id": "2511.17936", "title": "Mitigating Catastrophic Forgetting in Streaming Generative and Predictive Learning via Stateful Replay", "abstract": "Many deployed learning systems must update models on streaming data under memory constraints. The default strategy, sequential fine-tuning on each new phase, is architecture-agnostic but often suffers catastrophic forgetting when later phases correspond to different sub-populations or tasks. Replay with a finite buffer is a simple alternative, yet its behaviour across generative and predictive objectives is not well understood. We present a unified study of stateful replay for streaming autoencoding, time series forecasting, and classification. We view both sequential fine-tuning and replay as stochastic gradient methods for an ideal joint objective, and use a gradient alignment analysis to show when mixing current and historical samples should reduce forgetting. We then evaluate a single replay mechanism on six streaming scenarios built from Rotated MNIST, ElectricityLoadDiagrams 2011-2014, and Airlines delay data, using matched training budgets and three seeds. On heterogeneous multi task streams, replay reduces average forgetting by a factor of two to three, while on benign time based streams both methods perform similarly. These results position stateful replay as a strong and simple baseline for continual learning in streaming environments.", "date": "2025-11-22", "categories": []}
{"id": "2511.17973", "title": "Adversarial Pseudo-replay for Exemplar-free Class-incremental Learning", "abstract": "Exemplar-free class-incremental learning (EFCIL) aims to retain old knowledge acquired in the previous task while learning new classes, without storing the previous images due to storage constraints or privacy concerns. In EFCIL, the plasticity-stability dilemma, learning new tasks versus catastrophic forgetting, is a significant challenge, primarily due to the unavailability of images from earlier tasks. In this paper, we introduce adversarial pseudo-replay (APR), a method that perturbs the images of the new task with adversarial attack, to synthesize the pseudo-replay images online without storing any replay samples. During the new task training, the adversarial attack is conducted on the new task images with augmented old class mean prototypes as targets, and the resulting images are used for knowledge distillation to prevent semantic drift. Moreover, we calibrate the covariance matrices to compensate for the semantic drift after each task, by learning a transfer matrix on the pseudo-replay samples. Our method reconciles stability and plasticity, achieving state-of-the-art on challenging cold-start settings of the standard EFCIL benchmarks.", "date": "2025-11-22", "categories": []}
{"id": "2511.18468", "title": "SloMo-Fast: Slow-Momentum and Fast-Adaptive Teachers for Source-Free Continual Test-Time Adaptation", "abstract": "Continual Test-Time Adaptation (CTTA) is crucial for deploying models in real-world applications with unseen, evolving target domains. Existing CTTA methods, however, often rely on source data or prototypes, limiting their applicability in privacy-sensitive and resource-constrained settings. Additionally, these methods suffer from long-term forgetting, which degrades performance on previously encountered domains as target domains shift. To address these challenges, we propose SloMo-Fast, a source-free, dual-teacher CTTA framework designed for enhanced adaptability and generalization. It includes two complementary teachers: the Slow-Teacher, which exhibits slow forgetting and retains long-term knowledge of previously encountered domains to ensure robust generalization, and the Fast-Teacher rapidly adapts to new domains while accumulating and integrating knowledge across them. This framework preserves knowledge of past domains and adapts efficiently to new ones. We also introduce Cyclic Test-Time Adaptation (Cyclic-TTA), a novel CTTA benchmark that simulates recurring domain shifts. Our extensive experiments demonstrate that SloMo-Fast consistently outperforms state-of-the-art methods across Cyclic-TTA, as well as ten other CTTA settings, highlighting its ability to both adapt and generalize across evolving and revisited domains.", "date": "2025-11-23", "categories": []}
{"id": "2511.18516", "title": "Breaking Forgetting: Training-Free Few-Shot Class-Incremental Learning via Conditional Diffusion", "abstract": "Efforts to overcome catastrophic forgetting in Few-Shot Class-Incremental Learning (FSCIL) have primarily focused on developing more effective gradient-based optimization strategies. In contrast, little attention has been paid to the training cost explosion that inevitably arises as the number of novel classes increases, a consequence of relying on gradient learning even under extreme data scarcity. More critically, since FSCIL typically provides only a few samples for each new class, gradient-based updates not only induce severe catastrophic forgetting on base classes but also hinder adaptation to novel ones. This paper seeks to break this long-standing limitation by asking: Can we design a training-free FSCIL paradigm that entirely removes gradient optimization? We provide an affirmative answer by uncovering an intriguing connection between gradient-based optimization and the Conditional Diffusion process. Building on this observation, we propose a Conditional Diffusion-driven FSCIL (CD-FSCIL) framework that substitutes the conventional gradient update process with a diffusion-based generative transition, enabling training-free incremental adaptation while effectively mitigating forgetting. Furthermore, to enhance representation under few-shot constraints, we introduce a multimodal learning strategy that integrates visual features with natural language descriptions automatically generated by Large Language Models (LLMs). This synergy substantially alleviates the sample scarcity issue and improves generalization across novel classes. Extensive experiments on mainstream FSCIL benchmarks demonstrate that our method not only achieves state-of-the-art performance but also drastically reduces computational and memory overhead, marking a paradigm shift toward training-free continual adaptation.", "date": "2025-11-23", "categories": []}
{"id": "2511.18987", "title": "Dynamic Mixture of Experts Against Severe Distribution Shifts", "abstract": "The challenge of building neural networks that can continuously learn and adapt to evolving data streams is central to the fields of continual learning (CL) and reinforcement learning (RL). This lifelong learning problem is often framed in terms of the plasticity-stability dilemma, focusing on issues like loss of plasticity and catastrophic forgetting. Unlike neural networks, biological brains maintain plasticity through capacity growth, inspiring researchers to explore similar approaches in artificial networks, such as adding capacity dynamically. Prior solutions often lack parameter efficiency or depend on explicit task indices, but Mixture-of-Experts (MoE) architectures offer a promising alternative by specializing experts for distinct distributions. This paper aims to evaluate a DynamicMoE approach for continual and reinforcement learning environments and benchmark its effectiveness against existing network expansion methods.", "date": "2025-11-24", "categories": []}
{"id": "2511.19344", "title": "Annotation-Free Class-Incremental Learning", "abstract": "Despite significant progress in continual learning ranging from architectural novelty to clever strategies for mitigating catastrophic forgetting most existing methods rest on a strong but unrealistic assumption the availability of labeled data throughout the learning process. In real-world scenarios, however, data often arrives sequentially and without annotations, rendering conventional approaches impractical. In this work, we revisit the fundamental assumptions of continual learning and ask: Can current systems adapt when labels are absent and tasks emerge incrementally over time? To this end, we introduce Annotation-Free Class-Incremental Learning (AFCIL), a more realistic and challenging paradigm where unlabeled data arrives continuously, and the learner must incrementally acquire new classes without any supervision. To enable effective learning under AFCIL, we propose CrossWorld CL, a Cross Domain World Guided Continual Learning framework that incorporates external world knowledge as a stable auxiliary source. The method retrieves semantically related ImageNet classes for each downstream category, maps downstream and ImageNet features through a cross domain alignment strategy and finally introduce a novel replay strategy. This design lets the model uncover semantic structure without annotations while keeping earlier knowledge intact. Across four datasets, CrossWorld-CL surpasses CLIP baselines and existing continual and unlabeled learning methods, underscoring the benefit of world knowledge for annotation free continual learning.", "date": "2025-11-24", "categories": []}
{"id": "2509.10518", "title": "Holographic Knowledge Manifolds: A Novel Pipeline for Continual Learning Without Catastrophic Forgetting in Large Language Models", "abstract": "We introduce the Holographic Knowledge Manifold (HKM), a four-phase pipeline that achieves zero catastrophic forgetting in AI knowledge representation while maintaining minimal memory growth and high efficiency. Leveraging fractal quantization, probabilistic entanglement, and dynamic diffraction chipping, HKM compresses knowledge substrates by 3x with 67% storage savings, integrates holographically at 100%, and supports over 1,020 updates with 1% growth per increment. In experiments on combined WikiText and FB15k datasets (scaled to 2,997 nodes), we demonstrate industry-leading performance: 0% forgetting (infinite improvement over GEM baselines), 3x compression, and 53% training time reduction on consumer GPU hardware. Hypothetical cost analyses project $92.4M savings over 5 years at petabyte scale, with 21.2% energy reduction and 33% lower carbon footprint. This work hypothesizes a paradigm shift for public large language models (LLMs), enabling \"eternal\" adaptation without retraining. Future extensions to multimodal fusion and quantum hardware could further democratize scalable AI, potentially reducing fine-tuning costs by 60-80% for models like Llama-3 or Grok-4. Code, datasets, and full results are publicly available for reproducibility.", "date": "2025-09-03", "categories": []}
{"id": "2511.15375", "title": "Parameter Importance-Driven Continual Learning for Foundation Models", "abstract": "Domain-specific post-training often causes catastrophic forgetting, making foundation models lose their general reasoning ability and limiting their adaptability to dynamic real-world environments. Preserving general capabilities while acquiring downstream domain knowledge is a central challenge for large language and multimodal models. Traditional continual learning methods, such as regularization, replay and architectural isolation, suffer from poor downstream performance, reliance on inaccessible historical data, or additional parameter overhead. While recent parameter-efficient tuning (PET) methods can alleviate forgetting, their effectiveness strongly depends on the choice of parameters and update strategies. In this paper, we introduce PIECE, a Parameter Importance Estimation-based Continual Enhancement method that preserves general ability while efficiently learning domain knowledge without accessing prior training data or increasing model parameters. PIECE selectively updates only 0.1% of core parameters most relevant to new tasks, guided by two importance estimators: PIECE-F based on Fisher Information, and PIECE-S based on a second-order normalization that combines gradient and curvature information. Experiments across three language models and two multimodal models show that PIECE maintains general capabilities and achieves state-of-the-art continual learning performance across diverse downstream tasks. Our results highlight a practical path to scalable, domain-adaptive foundation models without catastrophic forgetting.", "date": "2025-11-19", "categories": []}
{"id": "2511.15597", "title": "Learning from Mistakes: Loss-Aware Memory Enhanced Continual Learning for LiDAR Place Recognition", "abstract": "LiDAR place recognition plays a crucial role in SLAM, robot navigation, and autonomous driving. However, existing LiDAR place recognition methods often struggle to adapt to new environments without forgetting previously learned knowledge, a challenge widely known as catastrophic forgetting. To address this issue, we propose KDF+, a novel continual learning framework for LiDAR place recognition that extends the KDF paradigm with a loss-aware sampling strategy and a rehearsal enhancement mechanism. The proposed sampling strategy estimates the learning difficulty of each sample via its loss value and selects samples for replay according to their estimated difficulty. Harder samples, which tend to encode more discriminative information, are sampled with higher probability while maintaining distributional coverage across the dataset. In addition, the rehearsal enhancement mechanism encourages memory samples to be further refined during new-task training by slightly reducing their loss relative to previous tasks, thereby reinforcing long-term knowledge retention. Extensive experiments across multiple benchmarks demonstrate that KDF+ consistently outperforms existing continual learning methods and can be seamlessly integrated into state-of-the-art continual learning for LiDAR place recognition frameworks to yield significant and stable performance gains. The code will be available at https://github.com/repo/KDF-plus.", "date": "2025-11-19", "categories": []}
{"id": "2511.15633", "title": "Hierarchical Semantic Tree Anchoring for CLIP-Based Class-Incremental Learning", "abstract": "Class-Incremental Learning (CIL) enables models to learn new classes continually while preserving past knowledge. Recently, vision-language models like CLIP offer transferable features via multi-modal pre-training, making them well-suited for CIL. However, real-world visual and linguistic concepts are inherently hierarchical: a textual concept like \"dog\" subsumes fine-grained categories such as \"Labrador\" and \"Golden Retriever,\" and each category entails its images. But existing CLIP-based CIL methods fail to explicitly capture this inherent hierarchy, leading to fine-grained class features drift during incremental updates and ultimately to catastrophic forgetting. To address this challenge, we propose HASTEN (Hierarchical Semantic Tree Anchoring) that anchors hierarchical information into CIL to reduce catastrophic forgetting. First, we employ an external knowledge graph as supervision to embed visual and textual features in hyperbolic space, effectively preserving hierarchical structure as data evolves. Second, to mitigate catastrophic forgetting, we project gradients onto the null space of the shared hyperbolic mapper, preventing interference with prior tasks. These two steps work synergistically to enable the model to resist forgetting by maintaining hierarchical relationships. Extensive experiments show that HASTEN consistently outperforms existing methods while providing a unified structured representation.", "date": "2025-11-19", "categories": []}
{"id": "2511.15652", "title": "Continual Reinforcement Learning for Cyber-Physical Systems: Lessons Learned and Open Challenges", "abstract": "Continual learning (CL) is a branch of machine learning that aims to enable agents to adapt and generalise previously learned abilities so that these can be reapplied to new tasks or environments. This is particularly useful in multi-task settings or in non-stationary environments, where the dynamics can change over time. This is particularly relevant in cyber-physical systems such as autonomous driving. However, despite recent advances in CL, successfully applying it to reinforcement learning (RL) is still an open problem.\n  This paper highlights open challenges in continual RL (CRL) based on experiments in an autonomous driving environment. In this environment, the agent must learn to successfully park in four different scenarios corresponding to parking spaces oriented at varying angles. The agent is successively trained in these four scenarios one after another, representing a CL environment, using Proximal Policy Optimisation (PPO). These experiments exposed a number of open challenges in CRL: finding suitable abstractions of the environment, oversensitivity to hyperparameters, catastrophic forgetting, and efficient use of neural network capacity.\n  Based on these identified challenges, we present open research questions that are important to be addressed for creating robust CRL systems. In addition, the identified challenges call into question the suitability of neural networks for CL. We also identify the need for interdisciplinary research, in particular between computer science and neuroscience.", "date": "2025-11-19", "categories": []}
{"id": "2510.06009", "title": "Continual Learning for Image Captioning through Improved Image-Text Alignment", "abstract": "Generating accurate and coherent image captions in a continual learning setting remains a major challenge due to catastrophic forgetting and the difficulty of aligning evolving visual concepts with language over time. In this work, we propose a novel multi-loss framework for continual image captioning that integrates semantic guidance through prompt-based continual learning and contrastive alignment. Built upon a pretrained ViT-GPT-2 backbone, our approach combines standard cross-entropy loss with three additional components: (1) a prompt-based cosine similarity loss that aligns image embeddings with synthetically constructed prompts encoding objects, attributes, and actions; (2) a CLIP-style loss that promotes alignment between image embeddings and target caption embedding; and (3) a language-guided contrastive loss that employs a triplet loss to enhance class-level discriminability between tasks. Notably, our approach introduces no additional overhead at inference time and requires no prompts during caption generation. We find that this approach mitigates catastrophic forgetting, while achieving better semantic caption alignment compared to state-of-the-art methods. The code can be found via the following link: https://github.com/Gepardius/Taetz_Bordelius_Continual_ImageCaptioning.", "date": "2025-10-07", "categories": []}
{"id": "2511.13880", "title": "AnaCP: Toward Upper-Bound Continual Learning via Analytic Contrastive Projection", "abstract": "This paper studies the problem of class-incremental learning (CIL), a core setting within continual learning where a model learns a sequence of tasks, each containing a distinct set of classes. Traditional CIL methods, which do not leverage pre-trained models (PTMs), suffer from catastrophic forgetting (CF) due to the need to incrementally learn both feature representations and the classifier. The integration of PTMs into CIL has recently led to efficient approaches that treat the PTM as a fixed feature extractor combined with analytic classifiers, achieving state-of-the-art performance. However, they still face a major limitation: the inability to continually adapt feature representations to best suit the CIL tasks, leading to suboptimal performance. To address this, we propose AnaCP (Analytic Contrastive Projection), a novel method that preserves the efficiency of analytic classifiers while enabling incremental feature adaptation without gradient-based training, thereby eliminating the CF caused by gradient updates. Our experiments show that AnaCP not only outperforms existing baselines but also achieves the accuracy level of joint training, which is regarded as the upper bound of CIL.", "date": "2025-11-17", "categories": []}
{"id": "2503.15414", "title": "Federated Continual 3D Segmentation With Single-round Communication", "abstract": "Federated learning seeks to foster collaboration among distributed clients while preserving the privacy of their local data. Traditionally, federated learning methods assume a fixed setting in which client data and learning objectives remain constant. However, in real-world scenarios, new clients may join, and existing clients may expand the segmentation label set as task requirements evolve. In such a dynamic federated analysis setup, the conventional federated communication strategy of model aggregation per communication round is suboptimal. As new clients join, this strategy requires retraining, linearly increasing communication and computation overhead. It also imposes requirements for synchronized communication, which is difficult to achieve among distributed clients. In this paper, we propose a federated continual learning strategy that employs a one-time model aggregation at the server through multi-model distillation. This approach builds and updates the global model while eliminating the need for frequent server communication. When integrating new data streams or onboarding new clients, this approach efficiently reuses previous client models, avoiding the need to retrain the global model across the entire federation. By minimizing communication load and bypassing the need to put unchanged clients online, our approach relaxes synchronization requirements among clients, providing an efficient and scalable federated analysis framework suited for real-world applications. Using multi-class 3D abdominal CT segmentation as an application task, we demonstrate the effectiveness of the proposed approach.", "date": "2025-03-19", "categories": []}
{"id": "2510.17160", "title": "Learning After Model Deployment", "abstract": "In classic supervised learning, once a model is deployed in an application, it is fixed. No updates will be made to it during the application. This is inappropriate for many dynamic and open environments, where unexpected samples from unseen classes may appear. In such an environment, the model should be able to detect these novel samples from unseen classes and learn them after they are labeled. We call this paradigm Autonomous Learning after Model Deployment (ALMD). The learning here is continuous and involves no human engineers. Labeling in this scenario is performed by human co-workers or other knowledgeable agents, which is similar to what humans do when they encounter an unfamiliar object and ask another person for its name. In ALMD, the detection of novel samples is dynamic and differs from traditional out-of-distribution (OOD) detection in that the set of in-distribution (ID) classes expands as new classes are learned during application, whereas ID classes is fixed in traditional OOD detection. Learning is also different from classic supervised learning because in ALMD, we learn the encountered new classes immediately and incrementally. It is difficult to retrain the model from scratch using all the past data from the ID classes and the novel samples from newly discovered classes, as this would be resource- and time-consuming. Apart from these two challenges, ALMD faces the data scarcity issue because instances of new classes often appear sporadically in real-life applications. To address these issues, we propose a novel method, PLDA, which performs dynamic OOD detection and incremental learning of new classes on the fly. Empirical evaluations will demonstrate the effectiveness of PLDA.", "date": "2025-10-20", "categories": []}
{"id": "2511.08322", "title": "Mitigating Negative Flips via Margin Preserving Training", "abstract": "Minimizing inconsistencies across successive versions of an AI system is as crucial as reducing the overall error. In image classification, such inconsistencies manifest as negative flips, where an updated model misclassifies test samples that were previously classified correctly. This issue becomes increasingly pronounced as the number of training classes grows over time, since adding new categories reduces the margin of each class and may introduce conflicting patterns that undermine their learning process, thereby degrading performance on the original subset. To mitigate negative flips, we propose a novel approach that preserves the margins of the original model while learning an improved one. Our method encourages a larger relative margin between the previously learned and newly introduced classes by introducing an explicit margin-calibration term on the logits. However, overly constraining the logit margin for the new classes can significantly degrade their accuracy compared to a new independently trained model. To address this, we integrate a double-source focal distillation loss with the previous model and a new independently trained model, learning an appropriate decision margin from both old and new data, even under a logit margin calibration. Extensive experiments on image classification benchmarks demonstrate that our approach consistently reduces the negative flip rate with high overall accuracy.", "date": "2025-11-11", "categories": []}
{"id": "2511.11676", "title": "Learning with Preserving for Continual Multitask Learning", "abstract": "Artificial intelligence systems in critical fields like autonomous driving and medical imaging analysis often continually learn new tasks using a shared stream of input data. For instance, after learning to detect traffic signs, a model may later need to learn to classify traffic lights or different types of vehicles using the same camera feed. This scenario introduces a challenging setting we term Continual Multitask Learning (CMTL), where a model sequentially learns new tasks on an underlying data distribution without forgetting previously learned abilities. Existing continual learning methods often fail in this setting because they learn fragmented, task-specific features that interfere with one another. To address this, we introduce Learning with Preserving (LwP), a novel framework that shifts the focus from preserving task outputs to maintaining the geometric structure of the shared representation space. The core of LwP is a Dynamically Weighted Distance Preservation (DWDP) loss that prevents representation drift by regularizing the pairwise distances between latent data representations. This mechanism of preserving the underlying geometric structure allows the model to retain implicit knowledge and support diverse tasks without requiring a replay buffer, making it suitable for privacy-conscious applications. Extensive evaluations on time-series and image benchmarks show that LwP not only mitigates catastrophic forgetting but also consistently outperforms state-of-the-art baselines in CMTL tasks. Notably, our method shows superior robustness to distribution shifts and is the only approach to surpass the strong single-task learning baseline, underscoring its effectiveness for real-world dynamic environments.", "date": "2025-11-11", "categories": []}
{"id": "2511.11707", "title": "FSC-Net: Fast-Slow Consolidation Networks for Continual Learning", "abstract": "Continual learning remains challenging due to catastrophic forgetting, where neural networks lose previously acquired knowledge when learning new tasks. Inspired by memory consolidation in neuroscience, we propose FSC-Net (Fast-Slow Consolidation Networks), a dual-network architecture that separates rapid task learning from gradual knowledge consolidation. Our method employs a fast network (NN1) for immediate adaptation to new tasks and a slow network (NN2) that consolidates knowledge through distillation and replay. Within the family of MLP-based NN1 variants we evaluated, consolidation effectiveness is driven more by methodology than architectural embellishments -- a simple MLP outperforms more complex similarity-gated variants by 1.2pp. Through systematic hyperparameter analysis, we observed empirically that pure replay without distillation during consolidation achieves superior performance, consistent with the hypothesis that distillation from the fast network introduces recency bias. On Split-MNIST (30 seeds), FSC-Net achieves 91.71% +/- 0.62% retention accuracy, a +4.27pp gain over the fast network alone (87.43% +/- 1.27%, paired t=23.585, p < 1e-10). On Split-CIFAR-10 (5 seeds), our method achieves 33.31% +/- 0.38% retention with an +8.20pp gain over the fast network alone (25.11% +/- 1.61%, paired t=9.75, p < 1e-3), demonstrating +8.20pp gain, though absolute performance (33.31%) remains modest and below random expectation, highlighting need for stronger backbones. Our results provide empirical evidence that the dual-timescale consolidation mechanism, rather than architectural complexity, is central to mitigating catastrophic forgetting in this setting.", "date": "2025-11-12", "categories": []}
{"id": "2511.12828", "title": "Catastrophic Forgetting in Kolmogorov-Arnold Networks", "abstract": "Catastrophic forgetting is a longstanding challenge in continual learning, where models lose knowledge from earlier tasks when learning new ones. While various mitigation strategies have been proposed for Multi-Layer Perceptrons (MLPs), recent architectural advances like Kolmogorov-Arnold Networks (KANs) have been suggested to offer intrinsic resistance to forgetting by leveraging localized spline-based activations. However, the practical behavior of KANs under continual learning remains unclear, and their limitations are not well understood. To address this, we present a comprehensive study of catastrophic forgetting in KANs and develop a theoretical framework that links forgetting to activation support overlap and intrinsic data dimension. We validate these analyses through systematic experiments on synthetic and vision tasks, measuring forgetting dynamics under varying model configurations and data complexity. Further, we introduce KAN-LoRA, a novel adapter design for parameter-efficient continual fine-tuning of language models, and evaluate its effectiveness in knowledge editing tasks. Our findings reveal that while KANs exhibit promising retention in low-dimensional algorithmic settings, they remain vulnerable to forgetting in high-dimensional domains such as image classification and language modeling. These results advance the understanding of KANs' strengths and limitations, offering practical insights for continual learning system design.", "date": "2025-11-16", "categories": []}
{"id": "2511.13351", "title": "Dual-LoRA and Quality-Enhanced Pseudo Replay for Multimodal Continual Food Learning", "abstract": "Food analysis has become increasingly critical for health-related tasks such as personalized nutrition and chronic disease prevention. However, existing large multimodal models (LMMs) in food analysis suffer from catastrophic forgetting when learning new tasks, requiring costly retraining from scratch. To address this, we propose a novel continual learning framework for multimodal food learning, integrating a Dual-LoRA architecture with Quality-Enhanced Pseudo Replay. We introduce two complementary low-rank adapters for each task: a specialized LoRA that learns task-specific knowledge with orthogonal constraints to previous tasks' subspaces, and a cooperative LoRA that consolidates shared knowledge across tasks via pseudo replay. To improve the reliability of replay data, our Quality-Enhanced Pseudo Replay strategy leverages self-consistency and semantic similarity to reduce hallucinations in generated samples. Experiments on the comprehensive Uni-Food dataset show superior performance in mitigating forgetting, representing the first effective continual learning approach for complex food tasks.", "date": "2025-11-17", "categories": []}
{"id": "2505.19943", "title": "Sparse Tuning Enhances Plasticity in PTM-based Continual Learning", "abstract": "Continual Learning with Pre-trained Models holds great promise for efficient adaptation across sequential tasks. However, most existing approaches freeze PTMs and rely on auxiliary modules like prompts or adapters, limiting model plasticity and leading to suboptimal generalization when facing significant distribution shifts. While full fine-tuning can improve adaptability, it risks disrupting crucial pre-trained knowledge. In this paper, we propose Mutual Information-guided Sparse Tuning (MIST), a plug-and-play method that selectively updates a small subset of PTM parameters, less than 5%, based on sensitivity to mutual information objectives. MIST enables effective task-specific adaptation while preserving generalization. To further reduce interference, we introduce strong sparsity regularization by randomly dropping gradients during tuning, resulting in fewer than 0.5% of parameters being updated per step. Applied before standard freeze-based methods, MIST consistently boosts performance across diverse continual learning benchmarks. Experiments show that integrating our method into multiple baselines yields significant performance gains. Our code is available at https://github.com/zhwhu/MIST.", "date": "2025-05-26", "categories": []}
{"id": "2511.10974", "title": "Preserving Cross-Modal Consistency for CLIP-based Class-Incremental Learning", "abstract": "Class-incremental learning (CIL) enables models to continuously learn new categories from sequential tasks without forgetting previously acquired knowledge. While recent advances in vision-language models such as CLIP have demonstrated strong generalization across domains, extending them to continual settings remains challenging. In particular, learning task-specific soft prompts for newly introduced classes often leads to severe classifier bias, as the text prototypes overfit to recent categories when prior data are unavailable. In this paper, we propose DMC, a simple yet effective two-stage framework for CLIP-based CIL that decouples the adaptation of the vision encoder and the optimization of textual soft prompts. Each stage is trained with the other frozen, allowing one modality to act as a stable semantic anchor for the other to preserve cross-modal alignment. Furthermore, current CLIP-based CIL approaches typically store class-wise Gaussian statistics for generative replay, yet they overlook the distributional drift that arises when the vision encoder is updated over time. To address this issue, we introduce DMC-OT, an enhanced version of DMC that incorporates an optimal-transport guided calibration strategy to align memory statistics across evolving encoders, along with a task-specific prompting design that enhances inter-task separability. Extensive experiments on CIFAR-100, Imagenet-R, CUB-200, and UCF-101 demonstrate that both DMC and DMC-OT achieve state-of-the-art performance, with DMC-OT further improving accuracy by an average of 1.80%.", "date": "2025-11-14", "categories": []}
{"id": "2511.11421", "title": "BOFA: Bridge-Layer Orthogonal Low-Rank Fusion for CLIP-Based Class-Incremental Learning", "abstract": "Class-Incremental Learning (CIL) aims to continually learn new categories without forgetting previously acquired knowledge. Vision-language models such as CLIP offer strong transferable representations via multi-modal supervision, making them promising for CIL. However, applying CLIP to CIL poses two major challenges: (1) adapting to downstream tasks often requires additional learnable modules, increasing model complexity and susceptibility to forgetting; and (2) while multi-modal representations offer complementary strengths, existing methods have yet to fully realize their potential in effectively integrating visual and textual modalities. To address these issues, we propose BOFA (Bridge-layer Orthogonal Fusion for Adaptation), a novel framework for CIL. BOFA confines all model adaptation exclusively to CLIP's existing cross-modal bridge-layer, thereby adding no extra parameters or inference cost. To prevent forgetting within this layer, it leverages Orthogonal Low-Rank Fusion, a mechanism that constrains parameter updates to a low-rank ``safe subspace\" mathematically constructed to be orthogonal to past task features. This ensures stable knowledge accumulation without data replay. Furthermore, BOFA employs a cross-modal hybrid prototype that synergizes stable textual prototypes with visual counterparts derived from our stably adapted bridge-layer, enhancing classification performance. Extensive experiments on standard benchmarks show that BOFA achieves superior accuracy and efficiency compared to existing methods.", "date": "2025-11-14", "categories": []}
{"id": "2409.17806", "title": "Caption, Create, Continue: Continual Learning with Pre-trained Generative Vision-Language Models", "abstract": "Continual learning (CL) enables models to adapt to evolving data streams without catastrophic forgetting, a fundamental requirement for real-world AI systems. However, the current methods often depend on large replay buffers or heavily annotated datasets which are impractical due to storage, privacy, and cost constraints. We propose CLTS (Continual Learning via Text-Image Synergy), a novel class-incremental framework that mitigates forgetting without storing real task data. CLTS leverages pre-trained vision-language models, BLIP (Bootstrapping Language-Image Pre-training) for caption generation and stable diffusion for sample generation. Each task is handled by a dedicated Task Head, while a Task Router learns to assign inputs to the correct Task Head using the generated data. On three benchmark datasets, CLTS improves average task accuracy by up to 54% and achieves 63 times better memory efficiency compared to four recent continual learning baselines, demonstrating improved retention and adaptability. CLTS introduces a novel perspective by integrating generative text-image augmentation for scalable continual learning.", "date": "2024-09-26", "categories": []}
{"id": "2511.09871", "title": "Expandable and Differentiable Dual Memories with Orthogonal Regularization for Exemplar-free Continual Learning", "abstract": "Continual learning methods used to force neural networks to process sequential tasks in isolation, preventing them from leveraging useful inter-task relationships and causing them to repeatedly relearn similar features or overly differentiate them. To address this problem, we propose a fully differentiable, exemplar-free expandable method composed of two complementary memories: One learns common features that can be used across all tasks, and the other combines the shared features to learn discriminative characteristics unique to each sample. Both memories are differentiable so that the network can autonomously learn latent representations for each sample. For each task, the memory adjustment module adaptively prunes critical slots and minimally expands capacity to accommodate new concepts, and orthogonal regularization enforces geometric separation between preserved and newly learned memory components to prevent interference. Experiments on CIFAR-10, CIFAR-100, and Tiny-ImageNet show that the proposed method outperforms 14 state-of-the-art methods for class-incremental learning, achieving final accuracies of 55.13\\%, 37.24\\%, and 30.11\\%, respectively. Additional analysis confirms that, through effective integration and utilization of knowledge, the proposed method can increase average performance across sequential tasks, and it produces feature extraction results closest to the upper bound, thus establishing a new milestone in continual learning.", "date": "2025-11-13", "categories": []}
{"id": "2511.09926", "title": "Compensating Distribution Drifts in Class-incremental Learning of Pre-trained Vision Transformers", "abstract": "Recent advances have shown that sequential fine-tuning (SeqFT) of pre-trained vision transformers (ViTs), followed by classifier refinement using approximate distributions of class features, can be an effective strategy for class-incremental learning (CIL). However, this approach is susceptible to distribution drift, caused by the sequential optimization of shared backbone parameters. This results in a mismatch between the distributions of the previously learned classes and that of the updater model, ultimately degrading the effectiveness of classifier performance over time. To address this issue, we introduce a latent space transition operator and propose Sequential Learning with Drift Compensation (SLDC). SLDC aims to align feature distributions across tasks to mitigate the impact of drift. First, we present a linear variant of SLDC, which learns a linear operator by solving a regularized least-squares problem that maps features before and after fine-tuning. Next, we extend this with a weakly nonlinear SLDC variant, which assumes that the ideal transition operator lies between purely linear and fully nonlinear transformations. This is implemented using learnable, weakly nonlinear mappings that balance flexibility and generalization. To further reduce representation drift, we apply knowledge distillation (KD) in both algorithmic variants. Extensive experiments on standard CIL benchmarks demonstrate that SLDC significantly improves the performance of SeqFT. Notably, by combining KD to address representation drift with SLDC to compensate distribution drift, SeqFT achieves performance comparable to joint training across all evaluated datasets. Code: https://github.com/raoxuan98-hash/sldc.git.", "date": "2025-11-13", "categories": []}
{"id": "2511.08634", "title": "CADIC: Continual Anomaly Detection Based on Incremental Coreset", "abstract": "The primary objective of Continual Anomaly Detection (CAD) is to learn the normal patterns of new tasks under dynamic data distribution assumptions while mitigating catastrophic forgetting. Existing embedding-based CAD approaches continuously update a memory bank with new embeddings to adapt to sequential tasks. However, these methods require constructing class-specific sub-memory banks for each task, which restricts their flexibility and scalability. To address this limitation, we propose a novel CAD framework where all tasks share a unified memory bank. During training, the method incrementally updates embeddings within a fixed-size coreset, enabling continuous knowledge acquisition from sequential tasks without task-specific memory fragmentation. In the inference phase, anomaly scores are computed via a nearest-neighbor matching mechanism, achieving state-of-the-art detection accuracy. We validate the method through comprehensive experiments on MVTec AD and Visa datasets. Results show that our approach outperforms existing baselines, achieving average image-level AUROC scores of 0.972 (MVTec AD) and 0.891 (Visa). Notably, on a real-world electronic paper dataset, it demonstrates 100% accuracy in anomaly sample detection, confirming its robustness in practical scenarios. The implementation will be open-sourced on GitHub.", "date": "2025-11-10", "categories": []}
{"id": "2511.09088", "title": "Improving Sustainability of Adversarial Examples in Class-Incremental Learning", "abstract": "Current adversarial examples (AEs) are typically designed for static models. However, with the wide application of Class-Incremental Learning (CIL), models are no longer static and need to be updated with new data distributed and labeled differently from the old ones. As a result, existing AEs often fail after CIL updates due to significant domain drift. In this paper, we propose SAE to enhance the sustainability of AEs against CIL. The core idea of SAE is to enhance the robustness of AE semantics against domain drift by making them more similar to the target class while distinguishing them from all other classes. Achieving this is challenging, as relying solely on the initial CIL model to optimize AE semantics often leads to overfitting. To resolve the problem, we propose a Semantic Correction Module. This module encourages the AE semantics to be generalized, based on a visual-language model capable of producing universal semantics. Additionally, it incorporates the CIL model to correct the optimization direction of the AE semantics, guiding them closer to the target class. To further reduce fluctuations in AE semantics, we propose a Filtering-and-Augmentation Module, which first identifies non-target examples with target-class semantics in the latent space and then augments them to foster more stable semantics. Comprehensive experiments demonstrate that SAE outperforms baselines by an average of 31.28% when updated with a 9-fold increase in the number of classes.", "date": "2025-11-12", "categories": []}
{"id": "2511.09167", "title": "Compact Memory for Continual Logistic Regression", "abstract": "Despite recent progress, continual learning still does not match the performance of batch training. To avoid catastrophic forgetting, we need to build compact memory of essential past knowledge, but no clear solution has yet emerged, even for shallow neural networks with just one or two layers. In this paper, we present a new method to build compact memory for logistic regression. Our method is based on a result by Khan and Swaroop [2021] who show the existence of optimal memory for such models. We formulate the search for the optimal memory as Hessian-matching and propose a probabilistic PCA method to estimate them. Our approach can drastically improve accuracy compared to Experience Replay. For instance, on Split-ImageNet, we get 60% accuracy compared to 30% obtained by replay with memory-size equivalent to 0.3% of the data size. Increasing the memory size to 2% further boosts the accuracy to 74%, closing the gap to the batch accuracy of 77.6% on this task. Our work opens a new direction for building compact memory that can also be useful in the future for continual deep learning.", "date": "2025-11-12", "categories": []}
{"id": "2511.09487", "title": "PDAC: Efficient Coreset Selection for Continual Learning via Probability Density Awareness", "abstract": "Rehearsal-based Continual Learning (CL) maintains a limited memory buffer to store replay samples for knowledge retention, making these approaches heavily reliant on the quality of the stored samples. Current Rehearsal-based CL methods typically construct the memory buffer by selecting a representative subset (referred to as coresets), aiming to approximate the training efficacy of the full dataset with minimal storage overhead. However, mainstream Coreset Selection (CS) methods generally formulate the CS problem as a bi-level optimization problem that relies on numerous inner and outer iterations to solve, leading to substantial computational cost thus limiting their practical efficiency. In this paper, we aim to provide a more efficient selection logic and scheme for coreset construction. To this end, we first analyze the Mean Squared Error (MSE) between the buffer-trained model and the Bayes-optimal model through the perspective of localized error decomposition to investigate the contribution of samples from different regions to MSE suppression. Further theoretical and experimental analyses demonstrate that samples with high probability density play a dominant role in error suppression. Inspired by this, we propose the Probability Density-Aware Coreset (PDAC) method. PDAC leverages the Projected Gaussian Mixture (PGM) model to estimate each sample's joint density, enabling efficient density-prioritized buffer selection. Finally, we introduce the streaming Expectation Maximization (EM) algorithm to enhance the adaptability of PGM parameters to streaming data, yielding Streaming PDAC (SPDAC) for streaming scenarios. Extensive comparative experiments show that our methods outperforms other baselines across various CL settings while ensuring favorable efficiency.", "date": "2025-11-12", "categories": []}
{"id": "2505.11816", "title": "Continuous Subspace Optimization for Continual Learning", "abstract": "Continual learning aims to learn multiple tasks sequentially while preserving prior knowledge, but faces the challenge of catastrophic forgetting when adapting to new tasks. Recently, approaches leveraging pre-trained models have gained increasing popularity in mitigating this issue, due to the strong generalization ability of foundation models. To adjust pre-trained models for new tasks, existing methods usually employ low-rank adaptation, which restricts parameter updates to a fixed low-rank subspace. However, constraining the optimization space inherently compromises the model's learning capacity, resulting in inferior performance. To address this limitation, we propose Continuous Subspace Optimization for Continual Learning (CoSO) to fine-tune the model in a series of subspaces rather than a single one. These sequential subspaces are dynamically determined through the singular value decomposition of the gradients. CoSO updates the model by projecting gradients onto these subspaces, ensuring memory-efficient optimization. To mitigate forgetting, the optimization subspace of each task is constrained to be orthogonal to the historical task subspace. During task learning, CoSO maintains a task-specific component that captures the critical update directions for the current task. Upon completing a task, this component is used to update the historical task subspace, laying the groundwork for subsequent learning. Extensive experiments on multiple datasets demonstrate that CoSO significantly outperforms state-of-the-art methods, especially in challenging scenarios with long task sequences.", "date": "2025-05-17", "categories": []}
{"id": "2508.01579", "title": "Harnessing Textual Semantic Priors for Knowledge Transfer and Refinement in CLIP-Driven Continual Learning", "abstract": "Continual learning (CL) aims to equip models with the ability to learn from a stream of tasks without forgetting previous knowledge. With the progress of vision-language models like Contrastive Language-Image Pre-training (CLIP), their promise for CL has attracted increasing attention due to their strong generalizability. However, the potential of rich textual semantic priors in CLIP in addressing the stability-plasticity dilemma remains underexplored. During backbone training, most approaches transfer past knowledge without considering semantic relevance, leading to interference from unrelated tasks that disrupt the balance between stability and plasticity. Besides, while text-based classifiers provide strong generalization, they suffer from limited plasticity due to the inherent modality gap in CLIP. Visual classifiers help bridge this gap, but their prototypes lack rich and precise semantics. To address these challenges, we propose Semantic-Enriched Continual Adaptation (SECA), a unified framework that harnesses the anti-forgetting and structured nature of textual priors to guide semantic-aware knowledge transfer in the backbone and reinforce the semantic structure of the visual classifier. Specifically, a Semantic-Guided Adaptive Knowledge Transfer (SG-AKT) module is proposed to assess new images' relevance to diverse historical visual knowledge via textual cues, and aggregate relevant knowledge in an instance-adaptive manner as distillation signals. Moreover, a Semantic-Enhanced Visual Prototype Refinement (SE-VPR) module is introduced to refine visual prototypes using inter-class semantic relations captured in class-wise textual embeddings. Extensive experiments on multiple benchmarks validate the effectiveness of our approach.", "date": "2025-08-03", "categories": []}
{"id": "2509.21606", "title": "Task-Agnostic Federated Continual Learning via Replay-Free Gradient Projection", "abstract": "Federated continual learning (FCL) enables distributed client devices to learn from streaming data across diverse and evolving tasks. A major challenge to continual learning, catastrophic forgetting, is exacerbated in decentralized settings by the data heterogeneity, constrained communication and privacy concerns. We propose Federated gradient Projection-based Continual Learning with Task Identity Prediction (FedProTIP), a novel FCL framework that mitigates forgetting by projecting client updates onto the orthogonal complement of the subspace spanned by previously learned representations of the global model. This projection reduces interference with earlier tasks and preserves performance across the task sequence. To further address the challenge of task-agnostic inference, we incorporate a lightweight mechanism that leverages core bases from prior tasks to predict task identity and dynamically adjust the global model's outputs. Extensive experiments across standard FCL benchmarks demonstrate that FedProTIP significantly outperforms state-of-the-art methods in average accuracy, particularly in settings where task identities are a priori unknown.", "date": "2025-09-25", "categories": []}
{"id": "2511.08226", "title": "The Online Patch Redundancy Eliminator (OPRE): A novel approach to online agnostic continual learning using dataset compression", "abstract": "In order to achieve Continual Learning (CL), the problem of catastrophic forgetting, one that has plagued neural networks since their inception, must be overcome. The evaluation of continual learning methods relies on splitting a known homogeneous dataset and learning the associated tasks one after the other. We argue that most CL methods introduce a priori information about the data to come and cannot be considered agnostic. We exemplify this point with the case of methods relying on pretrained feature extractors, which are still used in CL. After showing that pretrained feature extractors imply a loss of generality with respect to the data that can be learned by the model, we then discuss other kinds of a priori information introduced in other CL methods. We then present the Online Patch Redundancy Eliminator (OPRE), an online dataset compression algorithm, which, along with the training of a classifier at test time, yields performance on CIFAR-10 and CIFAR-100 superior to a number of other state-of-the-art online continual learning methods. Additionally, OPRE requires only minimal and interpretable hypothesis on the data to come. We suggest that online dataset compression could well be necessary to achieve fully agnostic CL.", "date": "2025-11-11", "categories": []}
{"id": "2411.07959", "title": "On the Convergence of Continual Federated Learning Using Incrementally Aggregated Gradients", "abstract": "The holy grail of machine learning is to enable Continual Federated Learning (CFL) to enhance the efficiency, privacy, and scalability of AI systems while learning from streaming data. The primary challenge of a CFL system is to overcome global catastrophic forgetting, wherein the accuracy of the global model trained on new tasks declines on the old tasks. In this work, we propose Continual Federated Learning with Aggregated Gradients (C-FLAG), a novel replay-memory based federated strategy consisting of edge-based gradient updates on memory and aggregated gradients on the current data. We provide convergence analysis of the C-FLAG approach which addresses forgetting and bias while converging at a rate of $O(1/\\sqrt{T})$ over $T$ communication rounds. We formulate an optimization sub-problem that minimizes catastrophic forgetting, translating CFL into an iterative algorithm with adaptive learning rates that ensure seamless learning across tasks. We empirically show that C-FLAG outperforms several state-of-the-art baselines on both task and class-incremental settings with respect to metrics such as accuracy and forgetting.", "date": "2024-11-12", "categories": []}
{"id": "2511.05566", "title": "Efficient Online Continual Learning in Sensor-Based Human Activity Recognition", "abstract": "Machine learning models for sensor-based human activity recognition (HAR) are expected to adapt post-deployment to recognize new activities and different ways of performing existing ones. To address this need, Online Continual Learning (OCL) mechanisms have been proposed, allowing models to update their knowledge incrementally as new data become available while preserving previously acquired information. However, existing OCL approaches for sensor-based HAR are computationally intensive and require extensive labeled samples to represent new changes. Recently, pre-trained model-based (PTM-based) OCL approaches have shown significant improvements in performance and efficiency for computer vision applications. These methods achieve strong generalization capabilities by pre-training complex models on large datasets, followed by fine-tuning on downstream tasks for continual learning. However, applying PTM-based OCL approaches to sensor-based HAR poses significant challenges due to the inherent heterogeneity of HAR datasets and the scarcity of labeled data in post-deployment scenarios. This paper introduces PTRN-HAR, the first successful application of PTM-based OCL to sensor-based HAR. Unlike prior PTM-based OCL approaches, PTRN-HAR pre-trains the feature extractor using contrastive loss with a limited amount of data. This extractor is then frozen during the streaming stage. Furthermore, it replaces the conventional dense classification layer with a relation module network. Our design not only significantly reduces the resource consumption required for model training while maintaining high performance, but also improves data efficiency by reducing the amount of labeled data needed for effective continual learning, as demonstrated through experiments on three public datasets, outperforming the state-of-the-art. The code can be found here: https://anonymous.4open.science/r/PTRN-HAR-AF60/", "date": "2025-11-04", "categories": []}
{"id": "2511.06237", "title": "Mixtures of SubExperts for Large Language Continual Learning", "abstract": "Adapting Large Language Models (LLMs) to a continuous stream of tasks is a critical yet challenging endeavor. While Parameter-Efficient Fine-Tuning (PEFT) methods have become a standard for this, they face a fundamental dilemma in continual learning. Reusing a single set of PEFT parameters for new tasks often leads to catastrophic forgetting of prior knowledge. Conversely, allocating distinct parameters for each task prevents forgetting but results in a linear growth of the model's size and fails to facilitate knowledge transfer between related tasks. To overcome these limitations, we propose a novel adaptive PEFT method referred to as \\textit{Mixtures of SubExperts (MoSEs)}, a novel continual learning framework designed for minimal forgetting and efficient scalability. MoSEs integrate a sparse Mixture of SubExperts into the transformer layers, governed by a task-specific routing mechanism. This architecture allows the model to isolate and protect knowledge within dedicated SubExperts, thereby minimizing parameter interference and catastrophic forgetting. Crucially, the router can adaptively select and combine previously learned sparse parameters for new tasks, enabling effective knowledge transfer while ensuring that the model's capacity grows sublinearly. We evaluate MoSEs on the comprehensive TRACE benchmark datasets. Our experiments demonstrate that MoSEs significantly outperform conventional continual learning approaches in both knowledge retention and scalability to new tasks, achieving state-of-the-art performance with substantial memory and computational savings.", "date": "2025-11-09", "categories": []}
{"id": "2511.06723", "title": "Multi-Modal Continual Learning via Cross-Modality Adapters and Representation Alignment with Knowledge Preservation", "abstract": "Continual learning is essential for adapting models to new tasks while retaining previously acquired knowledge. While existing approaches predominantly focus on uni-modal data, multi-modal learning offers substantial benefits by utilizing diverse sensory inputs, akin to human perception. However, multi-modal continual learning presents additional challenges, as the model must effectively integrate new information from various modalities while preventing catastrophic forgetting. In this work, we propose a pre-trained model-based framework for multi-modal continual learning. Our framework includes a novel cross-modality adapter with a mixture-of-experts structure to facilitate effective integration of multi-modal information across tasks. We also introduce a representation alignment loss that fosters learning of robust multi-modal representations, and regularize relationships between learned representations to preserve knowledge from previous tasks. Experiments on several multi-modal datasets demonstrate that our approach consistently outperforms baselines in both class-incremental and domain-incremental learning, achieving higher accuracy and reduced forgetting.", "date": "2025-11-10", "categories": []}
{"id": "2511.05420", "title": "ProDER: A Continual Learning Approach for Fault Prediction in Evolving Smart Grids", "abstract": "As smart grids evolve to meet growing energy demands and modern operational challenges, the ability to accurately predict faults becomes increasingly critical. However, existing AI-based fault prediction models struggle to ensure reliability in evolving environments where they are required to adapt to new fault types and operational zones. In this paper, we propose a continual learning (CL) framework in the smart grid context to evolve the model together with the environment. We design four realistic evaluation scenarios grounded in class-incremental and domain-incremental learning to emulate evolving grid conditions. We further introduce Prototype-based Dark Experience Replay (ProDER), a unified replay-based approach that integrates prototype-based feature regularization, logit distillation, and a prototype-guided replay memory. ProDER achieves the best performance among tested CL techniques, with only a 0.045 accuracy drop for fault type prediction and 0.015 for fault zone prediction. These results demonstrate the practicality of CL for scalable, real-world fault prediction in smart grids.", "date": "2025-11-07", "categories": []}
{"id": "2503.18469", "title": "CFReID: Continual Few-shot Person Re-Identification", "abstract": "Real-world surveillance systems are dynamically evolving, requiring a person Re-identification model to continuously handle newly incoming data from various domains. To cope with these dynamics, Lifelong ReID (LReID) has been proposed to learn and accumulate knowledge across multiple domains incrementally. However, LReID models need to be trained on large-scale labeled data for each unseen domain, which are typically inaccessible due to privacy and cost concerns. In this paper, we propose a new paradigm called Continual Few-shot ReID (CFReID), which requires models to be incrementally trained using few-shot data and tested on all seen domains. Under few-shot conditions, CFREID faces two core challenges: 1) learning knowledge from few-shot data of unseen domain, and 2) avoiding catastrophic forgetting of seen domains. To tackle these two challenges, we propose a Stable Distribution Alignment (SDA) framework from feature distribution perspective. Specifically, our SDA is composed of two modules, i.e., Meta Distribution Alignment (MDA) and Prototype-based Few-shot Adaptation (PFA). To support the study of CFReID, we establish an evaluation benchmark for CFReID on five publicly available ReID datasets. Extensive experiments demonstrate that our SDA can enhance the few-shot learning and anti-forgetting capabilities under few-shot conditions. Notably, our approach, using only 5\\% of the data, i.e., 32 IDs, significantly outperforms LReID's state-of-the-art performance, which requires 700 to 1,000 IDs.", "date": "2025-03-24", "categories": []}
{"id": "2403.05066", "title": "Reset & Distill: A Recipe for Overcoming Negative Transfer in Continual Reinforcement Learning", "abstract": "We argue that the negative transfer problem occurring when the new task to learn arrives is an important problem that needs not be overlooked when developing effective Continual Reinforcement Learning (CRL) algorithms. Through comprehensive experimental validation, we demonstrate that such issue frequently exists in CRL and cannot be effectively addressed by several recent work on either mitigating plasticity loss of RL agents or enhancing the positive transfer in CRL scenario. To that end, we develop Reset & Distill (R&D), a simple yet highly effective baseline method, to overcome the negative transfer problem in CRL. R&D combines a strategy of resetting the agent's online actor and critic networks to learn a new task and an offline learning step for distilling the knowledge from the online actor and previous expert's action probabilities. We carried out extensive experiments on long sequence of Meta World tasks and show that our simple baseline method consistently outperforms recent approaches, achieving significantly higher success rates across a range of tasks. Our findings highlight the importance of considering negative transfer in CRL and emphasize the need for robust strategies like R&D to mitigate its detrimental effects.", "date": "2024-03-08", "categories": []}
{"id": "2511.02025", "title": "Path-Coordinated Continual Learning with Neural Tangent Kernel-Justified Plasticity: A Theoretical Framework with Near State-of-the-Art Performance", "abstract": "Catastrophic forgetting is one of the fundamental issues of continual learning because neural networks forget the tasks learned previously when trained on new tasks. The proposed framework is a new path-coordinated framework of continual learning that unites the Neural Tangent Kernel (NTK) theory of principled plasticity bounds, statistical validation by Wilson confidence intervals, and evaluation of path quality by the use of multiple metrics. Experimental evaluation shows an average accuracy of 66.7% at the cost of 23.4% catastrophic forgetting on Split-CIFAR10, a huge improvement over the baseline and competitive performance achieved, which is very close to state-of-the-art results. Further, it is found out that NTK condition numbers are predictive indicators of learning capacity limits, showing the existence of a critical threshold at condition number $>10^{11}$. It is interesting to note that the proposed strategy shows a tendency of lowering forgetting as the sequence of tasks progresses (27% to 18%), which is a system stabilization. The framework validates 80% of discovered paths with a rigorous statistical guarantee and maintains 90-97% retention on intermediate tasks. The core capacity limits of the continual learning environment are determined in the analysis, and actionable insights to enhance the adaptive regularization are offered.", "date": "2025-11-03", "categories": []}
{"id": "2207.13853", "title": "ORFit: One-Pass Learning via Bridging Orthogonal Gradient Descent and Recursive Least-Squares", "abstract": "While large machine learning models have shown remarkable performance in various domains, their training typically requires iterating for many passes over the training data. However, due to computational and memory constraints and potential privacy concerns, storing and accessing all the data is impractical in many real-world scenarios where the data arrives in a stream. In this paper, we investigate the problem of one-pass learning, in which a model is trained on sequentially arriving data without retraining on previous datapoints. Motivated by the demonstrated effectiveness of overparameterized models and the phenomenon of benign overfitting, we propose Orthogonal Recursive Fitting (ORFit), an algorithm for one-pass learning which seeks to perfectly fit each new datapoint while minimally altering the predictions on previous datapoints. ORFit updates the parameters in a direction orthogonal to past gradients, similar to orthogonal gradient descent (OGD) in continual learning. We show that, interestingly, ORFit's update leads to an operation similar to the recursive least-squares (RLS) algorithm in adaptive filtering but with significantly improved memory and computational efficiency, i.e., linear, instead of quadratic, in the number of parameters. To further reduce memory usage, we leverage the structure of the streaming data via an incremental principal component analysis (IPCA). We show that using the principal components is minimax optimal, i.e., it minimizes the worst-case forgetting of previous predictions for unknown future updates. Further, we prove that, for overparameterized linear models, the parameter vector obtained by ORFit matches what the standard multi-pass stochastic gradient descent (SGD) would converge to. Finally, we extend our results to the nonlinear setting for highly overparameterized models, relevant for deep learning.", "date": "2022-07-28", "categories": []}
{"id": "2510.00365", "title": "Continual Learning with Query-Only Attention", "abstract": "Continual learning involves learning from a stream of data without repetition of data points, a scenario that is inherently complex due to distributional shift across tasks. We propose a query-only attention mechanism that discards keys and values, yet preserves the core inductive bias of transformer architectures. In continual learning scenarios, this simplified mechanism significantly mitigates both loss of plasticity and catastrophic forgetting, outperforming baselines such as selective re-initialization. We establish a conceptual link between query-only attention, full transformer attention, and model agnostic meta-learning, framing them as instances of meta-learning. We further provide intuition for why query-based models and attention networks help preserve plasticity in continual settings. Finally, through preliminary Hessian spectrum analysis, we observe that models maintaining higher curvature rank across tasks tend to retain plasticity. Our findings suggest that full attention may not be essential for capturing the benefits of meta-learning in continual learning.", "date": "2025-10-01", "categories": []}
{"id": "2510.22405", "title": "Knowledge-guided Continual Learning for Behavioral Analytics Systems", "abstract": "User behavior on online platforms is evolving, reflecting real-world changes in how people post, whether it's helpful messages or hate speech. Models that learn to capture this content can experience a decrease in performance over time due to data drift, which can lead to ineffective behavioral analytics systems. However, fine-tuning such a model over time with new data can be detrimental due to catastrophic forgetting. Replay-based approaches in continual learning offer a simple yet efficient method to update such models, minimizing forgetting by maintaining a buffer of important training instances from past learned tasks. However, the main limitation of this approach is the fixed size of the buffer. External knowledge bases can be utilized to overcome this limitation through data augmentation. We propose a novel augmentation-based approach to incorporate external knowledge in the replay-based continual learning framework. We evaluate several strategies with three datasets from prior studies related to deviant behavior classification to assess the integration of external knowledge in continual learning and demonstrate that augmentation helps outperform baseline replay-based approaches.", "date": "2025-10-25", "categories": []}
{"id": "2511.01093", "title": "Continual Learning, Not Training: Online Adaptation For Agents", "abstract": "Continual Learning (CL) methods have traditionally focused on mitigating catastrophic forgetting through gradient-based retraining, an approach ill-suited for deployed agents that must adapt in real time. We introduce our Adaptive Teaching and Learning System (ATLAS), a dual-agent architecture that decouples reasoning (Teacher) from execution (Student) and incorporates a persistent learning memory that stores distilled guidance from experience. This informs the orchestration layer, enabling the system to dynamically adjust its operational strategies, such as supervision level or initial plan selection, at inference time. In doing so, ATLAS achieves gradient-free continual learning, shifting the locus of adaptation from model parameters to system-level orchestration. We formulate this as a system-centric paradigm for continual learning, where the objective is adaptive efficiency: maximizing task success while minimizing computational cost through inference-time orchestration rather than parameter updates. Evaluated on Microsoft's ExCyTIn-Bench, an open-source benchmark simulating complex cyberthreat investigation, ATLAS achieves 54.1% success with GPT-5-mini as its Student, outperforming the larger GPT-5 (High) by 13% while reducing cost by 86%. Cross-incident validation demonstrates generalization: frozen pamphlets from Incident #5 improve accuracy from 28% to 41% with zero retraining, while shifting output composition from verbose exploration to structured reasoning. Together, these findings establish gradient-free continual learning as a viable path toward adaptive, deployable AI systems and provide causally annotated traces valuable for training explicit world models.", "date": "2025-11-02", "categories": []}
{"id": "2511.01847", "title": "Bridging Lifelong and Multi-Task Representation Learning via Algorithm and Complexity Measure", "abstract": "In lifelong learning, a learner faces a sequence of tasks with shared structure and aims to identify and leverage it to accelerate learning. We study the setting where such structure is captured by a common representation of data. Unlike multi-task learning or learning-to-learn, where tasks are available upfront to learn the representation, lifelong learning requires the learner to make use of its existing knowledge while continually gathering partial information in an online fashion. In this paper, we consider a generalized framework of lifelong representation learning. We propose a simple algorithm that uses multi-task empirical risk minimization as a subroutine and establish a sample complexity bound based on a new notion we introduce--the task-eluder dimension. Our result applies to a wide range of learning problems involving general function classes. As concrete examples, we instantiate our result on classification and regression tasks under noise.", "date": "2025-11-03", "categories": []}
{"id": "2510.26311", "title": "Model Inversion with Layer-Specific Modeling and Alignment for Data-Free Continual Learning", "abstract": "Continual learning (CL) aims to incrementally train a model on a sequence of tasks while retaining performance on prior ones. However, storing and replaying data is often infeasible due to privacy or security constraints and impractical for arbitrary pre-trained models. Data-free CL seeks to update models without access to previous data. Beyond regularization, we employ model inversion to synthesize data from the trained model, enabling replay without storing samples. Yet, model inversion in predictive models faces two challenges: (1) generating inputs solely from compressed output labels causes drift between synthetic and real data, and replaying such data can erode prior knowledge; (2) inversion is computationally expensive since each step backpropagates through the full model. These issues are amplified in large pre-trained models such as CLIP. To improve efficiency, we propose Per-layer Model Inversion (PMI), inspired by faster convergence in single-layer optimization. PMI provides strong initialization for full-model inversion, substantially reducing iterations. To mitigate feature shift, we model class-wise features via Gaussian distributions and contrastive model, ensuring alignment between synthetic and real features. Combining PMI and feature modeling, our approach enables continual learning of new classes by generating pseudo-images from semantic-aware projected features, achieving strong effectiveness and compatibility across multiple CL settings.", "date": "2025-10-30", "categories": []}
{"id": "2403.09066", "title": "Hyperparameters in Continual Learning: A Reality Check", "abstract": "Continual learning (CL) aims to train a model on a sequence of tasks (i.e., a CL scenario) while balancing the trade-off between plasticity (learning new tasks) and stability (retaining prior knowledge). The dominantly adopted conventional evaluation protocol for CL algorithms selects the best hyperparameters (e.g., learning rate, mini-batch size, regularization strengths, etc.) within a given scenario and then evaluates the algorithms using these hyperparameters in the same scenario. However, this protocol has significant shortcomings: it overestimates the CL capacity of algorithms and relies on unrealistic hyperparameter tuning, which is not feasible for real-world applications. From the fundamental principles of evaluation in machine learning, we argue that the evaluation of CL algorithms should focus on assessing the generalizability of their CL capacity to unseen scenarios. Based on this, we propose the Generalizable Two-phase Evaluation Protocol (GTEP) consisting of hyperparameter tuning and evaluation phases. Both phases share the same scenario configuration (e.g., number of tasks) but are generated from different datasets. Hyperparameters of CL algorithms are tuned in the first phase and applied in the second phase to evaluate the algorithms. We apply this protocol to class-incremental learning, both with and without pretrained models. Across more than 8,000 experiments, our results show that most state-of-the-art algorithms fail to replicate their reported performance, highlighting that their CL capacity has been significantly overestimated in the conventional evaluation protocol. Our implementation can be found in https://github.com/csm9493/GTEP.", "date": "2024-03-14", "categories": []}
{"id": "2503.17061", "title": "Replay4NCL: An Efficient Memory Replay-based Methodology for Neuromorphic Continual Learning in Embedded AI Systems", "abstract": "Neuromorphic Continual Learning (NCL) paradigm leverages Spiking Neural Networks (SNNs) to enable continual learning (CL) capabilities for AI systems to adapt to dynamically changing environments. Currently, the state-of-the-art employ a memory replay-based method to maintain the old knowledge. However, this technique relies on long timesteps and compression-decompression steps, thereby incurring significant latency and energy overheads, which are not suitable for tightly-constrained embedded AI systems (e.g., mobile agents/robotics). To address this, we propose Replay4NCL, a novel efficient memory replay-based methodology for enabling NCL in embedded AI systems. Specifically, Replay4NCL compresses the latent data (old knowledge), then replays them during the NCL training phase with small timesteps, to minimize the processing latency and energy consumption. To compensate the information loss from reduced spikes, we adjust the neuron threshold potential and learning rate settings. Experimental results on the class-incremental scenario with the Spiking Heidelberg Digits (SHD) dataset show that Replay4NCL can preserve old knowledge with Top-1 accuracy of 90.43% compared to 86.22% from the state-of-the-art, while effectively learning new tasks, achieving 4.88x latency speed-up, 20% latent memory saving, and 36.43% energy saving. These results highlight the potential of our Replay4NCL methodology to further advances NCL capabilities for embedded AI systems.", "date": "2025-03-21", "categories": []}
{"id": "2411.11668", "title": "Efficient and Robust Continual Graph Learning for Graph Classification in Biology", "abstract": "Graph classification is essential for understanding complex biological systems, where molecular structures and interactions are naturally represented as graphs. Traditional graph neural networks (GNNs) perform well on static tasks but struggle in dynamic settings due to catastrophic forgetting. We present Perturbed and Sparsified Continual Graph Learning (PSCGL), a robust and efficient continual graph learning framework for graph data classification, specifically targeting biological datasets. We introduce a perturbed sampling strategy to identify critical data points that contribute to model learning and a motif-based graph sparsification technique to reduce storage needs while maintaining performance. Additionally, our PSCGL framework inherently defends against graph backdoor attacks, which is crucial for applications in sensitive biological contexts. Extensive experiments on biological datasets demonstrate that PSCGL not only retains knowledge across tasks but also enhances the efficiency and robustness of graph classification models in biology.", "date": "2024-11-18", "categories": []}
{"id": "2503.18672", "title": "CalFuse: Multi-Modal Continual Learning via Feature Calibration and Parameter Fusion", "abstract": "With the proliferation of multi-modal data in large-scale visual recognition systems, enabling models to continuously acquire knowledge from evolving data streams while preserving prior information has become increasingly critical. Class-Continual Learning (CCL) addresses this challenge by incrementally incorporating new class knowledge without revisiting historical data, making it essential for real-world big data applications. While traditional CCL methods rely solely on visual features, recent advances in Vision-Language Models (VLMs) such as CLIP demonstrate significant potential for CCL by leveraging pre-trained multi-modal knowledge. However, existing approaches face challenges in mitigating catastrophic forgetting while maintaining the cross-modal generalization capabilities of VLMs. To address these limitations, we propose CalFuse, a framework that synergizes feature Calibration with parameter Fusion to enable effective multi-modal knowledge integration in continual learning scenarios. CalFuse introduces a dynamic feature calibration mechanism that adaptively balances original CLIP visual representations with task-specific features, preserving the model's intrinsic cross-modal generalization while adapting to new classes. Concurrently, a QR decomposition-based parameter fusion strategy progressively integrates newly acquired knowledge with historical task parameters, maintaining equilibrium between learning new class representations and retaining prior knowledge across sequential tasks. Extensive experiments on benchmark datasets validate the effectiveness of our approach in large-scale multi-modal continual learning settings, demonstrating superior performance over state-of-the-art methods in both average accuracy and final task retention.", "date": "2025-03-24", "categories": []}
{"id": "2510.23756", "title": "Explaining Robustness to Catastrophic Forgetting Through Incremental Concept Formation", "abstract": "Catastrophic forgetting remains a central challenge in continual learning, where models are required to integrate new knowledge over time without losing what they have previously learned. In prior work, we introduced Cobweb/4V, a hierarchical concept formation model that exhibited robustness to catastrophic forgetting in visual domains. Motivated by this robustness, we examine three hypotheses regarding the factors that contribute to such stability: (1) adaptive structural reorganization enhances knowledge retention, (2) sparse and selective updates reduce interference, and (3) information-theoretic learning based on sufficiency statistics provides advantages over gradient-based backpropagation. To test these hypotheses, we compare Cobweb/4V with neural baselines, including CobwebNN, a neural implementation of the Cobweb framework introduced in this work. Experiments on datasets of varying complexity (MNIST, Fashion-MNIST, MedMNIST, and CIFAR-10) show that adaptive restructuring enhances learning plasticity, sparse updates help mitigate interference, and the information-theoretic learning process preserves prior knowledge without revisiting past data. Together, these findings provide insight into mechanisms that can mitigate catastrophic forgetting and highlight the potential of concept-based, information-theoretic approaches for building stable and adaptive continual learning systems.", "date": "2025-10-27", "categories": []}
{"id": "2505.15424", "title": "Gated Integration of Low-Rank Adaptation for Continual Learning of Large Language Models", "abstract": "Continual learning (CL), which requires the model to learn multiple tasks sequentially, is crucial for large language models (LLMs). Recently, low-rank adaptation~(LoRA), one of the most representative parameter-efficient fine-tuning (PEFT) methods, has gained increasing attention in CL of LLMs. However, most existing CL methods based on LoRA typically expand a new LoRA branch to learn each new task and force the new and old LoRA branches to influence old tasks equally, potentially leading to forgetting. In this work, we propose a new method, called gated integration of low-rank adaptation (GainLoRA), for CL of LLMs. GainLoRA expands a new LoRA branch for each new task and introduces gating modules to integrate the new and old LoRA branches. Furthermore, GainLoRA leverages the new gating module to minimize the influence from the new LoRA branch to old tasks, effectively mitigating forgetting and improving the model's overall performance. Experimental results on CL benchmarks demonstrate that GainLoRA outperforms existing state-of-the-art methods.", "date": "2025-05-21", "categories": []}
{"id": "2506.06501", "title": "Optimal Rates in Continual Linear Regression via Increasing Regularization", "abstract": "We study realizable continual linear regression under random task orderings, a common setting for developing continual learning theory. In this setup, the worst-case expected loss after $k$ learning iterations admits a lower bound of $\\Omega(1/k)$. However, prior work using an unregularized scheme has only established an upper bound of $O(1/k^{1/4})$, leaving a significant gap. Our paper proves that this gap can be narrowed, or even closed, using two frequently used regularization schemes: (1) explicit isotropic $\\ell_2$ regularization, and (2) implicit regularization via finite step budgets. We show that these approaches, which are used in practice to mitigate forgetting, reduce to stochastic gradient descent (SGD) on carefully defined surrogate losses. Through this lens, we identify a fixed regularization strength that yields a near-optimal rate of $O(\\log k / k)$. Moreover, formalizing and analyzing a generalized variant of SGD for time-varying functions, we derive an increasing regularization strength schedule that provably achieves an optimal rate of $O(1/k)$. This suggests that schedules that increase the regularization coefficient or decrease the number of steps per task are beneficial, at least in the worst case.", "date": "2025-06-06", "categories": []}
{"id": "2510.21836", "title": "COLA: Continual Learning via Autoencoder Retrieval of Adapters", "abstract": "Learning a set of tasks over time, also known as continual learning (CL), is one of the most challenging problems in artificial intelligence due to catastrophic forgetting. Large language models (LLMs) are often impractical to frequent re-training and continual learning , due to high cost of computational resources for training. Moreover, LLM are not suitable for continual learning as updating these models over time for acquiring new knowledge leads to overwrites existing knowledge leading to common phenomenon know as \\textit{catastrophic forgetting}. In this paper, we aim to address these concerns using a novel framework , COLA that employs an autoencoder to learn capture low-dimensional embeddings of the weights associated with various tasks. Our approach facilitates the transfer of knowledge to new tasks while preventing catastrophic forgetting, all without using data replay or a substantial set of task-specific parameters. Our approach, COLA, makes the LLM efficiently learn new tasks with minimal training, insignificant performance degradation on previous tasks, and eliminates the need for retaining earlier training data. Empirical evaluation on different datasets ranging from task oriented dialouge system to intent classsfication datasets showcases that our method not only overcomes catastrophic forgetting but also achieves significant reduction in parameter usage and memory size, across multiple tasks and outperforming the existing state of the art methods across multiple datasets.", "date": "2025-10-22", "categories": []}
{"id": "2510.23198", "title": "PTPP-Aware Adaptation Scaling Laws: Predicting Domain-Adaptation Performance at Unseen Pre-Training Budgets", "abstract": "Continual pre-training (CPT) for domain adaptation must balance target-domain gains with stability on the base domain. Existing CPT scaling laws typically assume a fixed pre-training budget, which limits their ability to forecast adaptation outcomes for models trained at different tokens-per-parameter (PTPP). We present \\emph{PTPP-aware} adaptation scaling laws that make the pre-training budget an explicit variable, enabling accurate \\emph{prediction} of adaptation loss at unseen \\ptpp. On a multilingual setup (English/Arabic $\\rightarrow$ French), PTPP-aware formulations trained on early stages (\\ptpp{}=\\{15,31\\}) predict target loss at \\ptpp{}=279 and outperform a PTPP-agnostic \\dcpt{} transfer baseline on metrics (Huber-on-log, MAE$_\\mathrm{rel}$, calibration slope); full diagnostics (RMSE, MAPE) are in the appendix. Beyond forecasting, we show a practical use case: planning replay ratios and adaptation token budgets that satisfy target and forgetting constraints under compute limits.", "date": "2025-10-27", "categories": []}
{"id": "2510.21175", "title": "Memory-Free Continual Learning with Null Space Adaptation for Zero-Shot Vision-Language Models", "abstract": "Pre-trained vision-language models (VLMs), such as CLIP, have demonstrated remarkable zero-shot generalization, enabling deployment in a wide range of real-world tasks without additional task-specific training. However, in real deployment scenarios with evolving environments or emerging classes, these models inevitably face distributional shifts and novel tasks. In such contexts, static zero-shot capabilities are insufficient, and there is a growing need for continual learning methods that allow models to adapt over time while avoiding catastrophic forgetting. We introduce NuSA-CL (Null Space Adaptation for Continual Learning), a lightweight memory-free continual learning framework designed to address this challenge. NuSA-CL employs low-rank adaptation and constrains task-specific weight updates to lie within an approximate null space of the model's current parameters. This strategy minimizes interference with previously acquired knowledge, effectively preserving the zero-shot capabilities of the original model. Unlike methods relying on replay buffers or costly distillation, NuSA-CL imposes minimal computational and memory overhead, making it practical for deployment in resource-constrained, real-world continual learning environments. Experiments show that our framework not only effectively preserves zero-shot transfer capabilities but also achieves highly competitive performance on continual learning benchmarks. These results position NuSA-CL as a practical and scalable solution for continually evolving zero-shot VLMs in real-world applications.", "date": "2025-10-24", "categories": []}
{"id": "2510.21188", "title": "PLAN: Proactive Low-Rank Allocation for Continual Learning", "abstract": "Continual learning (CL) requires models to continuously adapt to new tasks without forgetting past knowledge. In this work, we propose \\underline{P}roactive \\underline{L}ow-rank \\underline{A}llocatio\\underline{N} (PLAN), a framework that extends Low-Rank Adaptation (LoRA) to enable efficient and interference-aware fine-tuning of large pre-trained models in CL settings. PLAN proactively manages the allocation of task-specific subspaces by introducing orthogonal basis vectors for each task and optimizing them through a perturbation-based strategy that minimizes conflicts with previously learned parameters. Furthermore, PLAN incorporates a novel selection mechanism that identifies and assigns basis vectors with minimal sensitivity to interference, reducing the risk of degrading past knowledge while maintaining efficient adaptation to new tasks. Empirical results on standard CL benchmarks demonstrate that PLAN consistently outperforms existing methods, establishing a new state-of-the-art for continual learning with foundation models.", "date": "2025-10-24", "categories": []}
{"id": "2510.21367", "title": "Randomized Neural Network with Adaptive Forward Regularization for Online Task-free Class Incremental Learning", "abstract": "Class incremental learning (CIL) requires an agent to learn distinct tasks consecutively with knowledge retention against forgetting. Problems impeding the practical applications of CIL methods are twofold: (1) non-i.i.d batch streams and no boundary prompts to update, known as the harsher online task-free CIL (OTCIL) scenario; (2) CIL methods suffer from memory loss in learning long task streams, as shown in Fig. 1 (a). To achieve efficient decision-making and decrease cumulative regrets during the OTCIL process, a randomized neural network (Randomized NN) with forward regularization (-F) is proposed to resist forgetting and enhance learning performance. This general framework integrates unsupervised knowledge into recursive convex optimization, has no learning dissipation, and can outperform the canonical ridge style (-R) in OTCIL. Based on this framework, we derive the algorithm of the ensemble deep random vector functional link network (edRVFL) with adjustable forward regularization (-kF), where k mediates the intensity of the intervention. edRVFL-kF generates one-pass closed-form incremental updates and variable learning rates, effectively avoiding past replay and catastrophic forgetting while achieving superior performance. Moreover, to curb unstable penalties caused by non-i.i.d and mitigate intractable tuning of -kF in OTCIL, we improve it to the plug-and-play edRVFL-kF-Bayes, enabling all hard ks in multiple sub-learners to be self-adaptively determined based on Bayesian learning. Experiments were conducted on 2 image datasets including 6 metrics, dynamic performance, ablation tests, and compatibility, which distinctly validates the efficacy of our OTCIL frameworks with -kF-Bayes and -kF styles.", "date": "2025-10-24", "categories": []}
{"id": "2410.07812", "title": "Temporal-Difference Variational Continual Learning", "abstract": "Machine Learning models in real-world applications must continuously learn new tasks to adapt to shifts in the data-generating distribution. Yet, for Continual Learning (CL), models often struggle to balance learning new tasks (plasticity) with retaining previous knowledge (memory stability). Consequently, they are susceptible to Catastrophic Forgetting, which degrades performance and undermines the reliability of deployed systems. In the Bayesian CL literature, variational methods tackle this challenge by employing a learning objective that recursively updates the posterior distribution while constraining it to stay close to its previous estimate. Nonetheless, we argue that these methods may be ineffective due to compounding approximation errors over successive recursions. To mitigate this, we propose new learning objectives that integrate the regularization effects of multiple previous posterior estimations, preventing individual errors from dominating future posterior updates and compounding over time. We reveal insightful connections between these objectives and Temporal-Difference methods, a popular learning mechanism in Reinforcement Learning and Neuroscience. Experiments on challenging CL benchmarks show that our approach effectively mitigates Catastrophic Forgetting, outperforming strong Variational CL methods.", "date": "2024-10-10", "categories": []}
{"id": "2505.22389", "title": "Train with Perturbation, Infer after Merging: A Two-Stage Framework for Continual Learning", "abstract": "Continual Learning (CL) aims to enable models to continuously acquire new knowledge from a sequence of tasks with avoiding the forgetting of learned information. However, existing CL methods only rely on the parameters of the most recent task for inference, which makes them susceptible to catastrophic forgetting. Inspired by the recent success of model merging techniques, we propose \\textbf{Perturb-and-Merge (P\\&M)}, a novel continual learning framework that integrates model merging into the CL paradigm to mitigate forgetting. Specifically, after training on each task, P\\&M constructs a new model by forming a convex combination of the previous model and the newly trained task-specific model. Through theoretical analysis, We minimize the total loss increase across all tasks and derive a closed-form solution for the merging coefficient under mild assumptions. To further improve the performance of the merged model, we observe that the degradation introduced during merging can be alleviated by a regularization term composed of the task vector and the Hessian matrix of the loss function. Interestingly, we show that this term can be efficiently approximated using second-order symmetric finite differences, and a stochastic perturbation strategy along the task vector direction is accordingly devised which incurs no additional forward or backward passes while providing an effective approximation of the regularization term. Finally, we combine P\\&M with LoRA, a parameter-efficient fine-tuning method, to reduce memory overhead. Our proposed approach achieves state-of-the-art performance on several continual learning benchmark datasets. The code is available at https://github.com/qhmiao/P-M-for-Continual-Learning.", "date": "2025-05-28", "categories": []}
{"id": "2510.19941", "title": "Are Greedy Task Orderings Better Than Random in Continual Linear Regression?", "abstract": "We analyze task orderings in continual learning for linear regression, assuming joint realizability of training data. We focus on orderings that greedily maximize dissimilarity between consecutive tasks, a concept briefly explored in prior work but still surrounded by open questions. Using tools from the Kaczmarz method literature, we formalize such orderings and develop geometric and algebraic intuitions around them. Empirically, we demonstrate that greedy orderings converge faster than random ones in terms of the average loss across tasks, both for linear regression with random data and for linear probing on CIFAR-100 classification tasks. Analytically, in a high-rank regression setting, we prove a loss bound for greedy orderings analogous to that of random ones. However, under general rank, we establish a repetition-dependent separation. Specifically, while prior work showed that for random orderings, with or without replacement, the average loss after $k$ iterations is bounded by $\\mathcal{O}(1/\\sqrt{k})$, we prove that single-pass greedy orderings may fail catastrophically, whereas those allowing repetition converge at rate $\\mathcal{O}(1/\\sqrt[3]{k})$. Overall, we reveal nuances within and between greedy and random orderings.", "date": "2025-10-22", "categories": []}
{"id": "2510.20479", "title": "RECALL: REpresentation-aligned Catastrophic-forgetting ALLeviation via Hierarchical Model Merging", "abstract": "We unveil that internal representations in large language models (LLMs) serve as reliable proxies of learned knowledge, and propose RECALL, a novel representation-aware model merging framework for continual learning without access to historical data. RECALL computes inter-model similarity from layer-wise hidden representations over clustered typical samples, and performs adaptive, hierarchical parameter fusion to align knowledge across models. This design enables the preservation of domain-general features in shallow layers while allowing task-specific adaptation in deeper layers. Unlike prior methods that require task labels or incur performance trade-offs, RECALL achieves seamless multi-domain integration and strong resistance to catastrophic forgetting. Extensive experiments across five NLP tasks and multiple continual learning scenarios show that RECALL outperforms baselines in both knowledge retention and generalization, providing a scalable and data-free solution for evolving LLMs.", "date": "2025-10-23", "categories": []}
{"id": "2510.20709", "title": "Separating the what and how of compositional computation to enable reuse and continual learning", "abstract": "The ability to continually learn, retain and deploy skills to accomplish goals is a key feature of intelligent and efficient behavior. However, the neural mechanisms facilitating the continual learning and flexible (re-)composition of skills remain elusive. Here, we study continual learning and the compositional reuse of learned computations in recurrent neural network (RNN) models using a novel two-system approach: one system that infers what computation to perform, and one that implements how to perform it. We focus on a set of compositional cognitive tasks commonly studied in neuroscience. To construct the what system, we first show that a large family of tasks can be systematically described by a probabilistic generative model, where compositionality stems from a shared underlying vocabulary of discrete task epochs. The shared epoch structure makes these tasks inherently compositional. We first show that this compositionality can be systematically described by a probabilistic generative model. Furthermore, We develop an unsupervised online learning approach that can learn this model on a single-trial basis, building its vocabulary incrementally as it is exposed to new tasks, and inferring the latent epoch structure as a time-varying computational context within a trial. We implement the how system as an RNN whose low-rank components are composed according to the context inferred by the what system. Contextual inference facilitates the creation, learning, and reuse of low-rank RNN components as new tasks are introduced sequentially, enabling continual learning without catastrophic forgetting. Using an example task set, we demonstrate the efficacy and competitive performance of this two-system learning framework, its potential for forward and backward transfer, as well as fast compositional generalization to unseen tasks.", "date": "2025-10-23", "categories": []}
{"id": "2503.07663", "title": "Merge then Realign: Simple and Effective Modality-Incremental Continual Learning for Multimodal LLMs", "abstract": "Recent advances in Multimodal Large Language Models (MLLMs) have enhanced their versatility as they integrate a growing number of modalities. Considering the heavy cost of training MLLMs, it is efficient to reuse the existing ones and extend them to more modalities through Modality-incremental Continual Learning (MCL). The exploration of MCL is in its early stages. In this work, we dive into the causes of performance degradation in MCL. We uncover that it suffers not only from forgetting as in traditional continual learning, but also from misalignment between the modality-agnostic and modality-specific components. To this end, we propose an elegantly simple MCL paradigm called \"MErge then ReAlign\" (MERA) to address both forgetting and misalignment. MERA avoids introducing heavy model budgets or modifying model architectures, hence is easy to deploy and highly reusable in the MLLM community. Extensive experiments demonstrate the impressive performance of MERA, holding an average of 99.84\\% Backward Relative Gain when extending to four modalities, achieving nearly lossless MCL performance. Our findings underscore the misalignment issue in MCL. More broadly, our work showcases how to adjust different components of MLLMs during continual learning.", "date": "2025-03-08", "categories": []}
{"id": "2505.11883", "title": "MINGLE: Mixture of Null-Space Gated Low-Rank Experts for Test-Time Continual Model Merging", "abstract": "Continual model merging integrates independently fine-tuned models sequentially without access to the original training data, offering a scalable and efficient solution for continual learning. However, existing methods face two critical challenges: parameter interference among tasks, which leads to catastrophic forgetting, and limited adaptability to evolving test distributions. To address these issues, we introduce the task of Test-Time Continual Model Merging (TTCMM), which leverages a small set of unlabeled test samples during inference to alleviate parameter conflicts and handle distribution shifts. We propose MINGLE, a novel framework for TTCMM. MINGLE employs a mixture-of-experts architecture with parameter-efficient, low-rank experts, which enhances adaptability to evolving test distributions while dynamically merging models to mitigate conflicts. To further reduce forgetting, we propose Null-Space Constrained Gating, which restricts gating updates to subspaces orthogonal to prior task representations, thereby suppressing activations on old tasks and preserving past knowledge. We further introduce an Adaptive Relaxation Strategy that adjusts constraint strength dynamically based on interference signals observed during test-time adaptation, striking a balance between stability and adaptability. Extensive experiments on standard continual merging benchmarks demonstrate that MINGLE achieves robust generalization, significantly reduces forgetting, and consistently surpasses previous state-of-the-art methods by 7-9% on average across diverse task orders. Our code is available at: https://github.com/zihuanqiu/MINGLE", "date": "2025-05-17", "categories": []}
{"id": "2502.13935", "title": "Foundations of a Developmental Design Paradigm for Integrated Continual Learning, Deliberative Behavior, and Comprehensibility", "abstract": "Inherent limitations of contemporary machine learning systems in crucial areas -- importantly in continual learning, information reuse, comprehensibility, and integration with deliberate behavior -- are receiving increasing attention. To address these challenges, we introduce a system design, fueled by a novel learning approach conceptually grounded in principles of evolutionary developmental biology, that overcomes key limitations of current methods. Our design comprises three core components: The Modeller, a gradient-free learning mechanism inherently capable of continual learning and structural adaptation; a planner for goal-directed action over learned models; and a behavior encapsulation mechanism that can decompose complex behaviors into a hierarchical structure. We demonstrate proof-of-principle operation in a simple test environment. Additionally, we extend our modeling framework to higher-dimensional network-structured spaces, using MNIST for a shape detection task. Our framework shows promise in overcoming multiple major limitations of contemporary machine learning systems simultaneously and in an organic manner.", "date": "2025-02-19", "categories": []}
{"id": "2510.16077", "title": "Continual Knowledge Consolidation LORA for Domain Incremental Learning", "abstract": "Domain Incremental Learning (DIL) is a continual learning sub-branch that aims to address never-ending arrivals of new domains without catastrophic forgetting problems. Despite the advent of parameter-efficient fine-tuning (PEFT) approaches, existing works create task-specific LoRAs overlooking shared knowledge across tasks. Inaccurate selection of task-specific LORAs during inference results in significant drops in accuracy, while existing works rely on linear or prototype-based classifiers, which have suboptimal generalization powers. Our paper proposes continual knowledge consolidation low rank adaptation (CONEC-LoRA) addressing the DIL problems. CONEC-LoRA is developed from consolidations between task-shared LORA to extract common knowledge and task-specific LORA to embrace domain-specific knowledge. Unlike existing approaches, CONEC-LoRA integrates the concept of a stochastic classifier whose parameters are sampled from a distribution, thus enhancing the likelihood of correct classifications. Last but not least, an auxiliary network is deployed to optimally predict the task-specific LoRAs for inferences and implements the concept of a different-depth network structure in which every layer is connected with a local classifier to take advantage of intermediate representations. This module integrates the ball-generator loss and transformation module to address the synthetic sample bias problem. Our rigorous experiments demonstrate the advantage of CONEC-LoRA over prior arts in 4 popular benchmark problems with over 5% margins.", "date": "2025-10-17", "categories": []}
{"id": "2510.16089", "title": "STABLE: Gated Continual Learning for Large Language Models", "abstract": "Large language models (LLMs) increasingly require mechanisms for continual adaptation without full retraining. However, sequential updates can lead to catastrophic forgetting, where new edits degrade previously acquired knowledge. This work presents STABLE, a gated continual self editing framework that constrains forgetting during sequential updates using parameter efficient fine tuning via Low Rank Adaptation (LoRA; see arXiv:2106.09685). Each candidate edit is evaluated against a stability budget using one of three metrics: (i) Exact Match (EM) drop, capturing factual accuracy loss; (ii) bits increase, reflecting reduced model confidence; and (iii) KL divergence, quantifying distributional drift between the base and adapted models. If a threshold is exceeded, the LoRA update is rescaled through a clipping procedure or rejected. Experiments on the Qwen-2.5-7B model show that gating effectively mitigates forgetting while preserving adaptability. EM based gating achieved the highest cumulative performance in short continual learning sequences. Our results show that different gating strategies can achieve comparable distribution shift (measured by KL divergence) while producing different accuracy outcomes, highlighting the importance of gating design in continual adaptation. This approach offers a principled method for continual model editing, enabling LLMs to integrate new knowledge while maintaining reliability. Code: https://github.com/Bhoy1/STABLE", "date": "2025-10-17", "categories": []}
{"id": "2510.16914", "title": "Domain Generalizable Continual Learning", "abstract": "To adapt effectively to dynamic real-world environments, intelligent systems must continually acquire new skills while generalizing them to diverse, unseen scenarios. Here, we introduce a novel and realistic setting named domain generalizable continual learning (DGCL): a model learns sequential tasks with each involving a single domain, aiming to perform well across all encountered tasks and domains. This setting poses unique challenges in acquiring, retaining, and leveraging both semantic- and domain-relevant information for robust generalization. Although state-of-the-art continual learning (CL) methods have employed pre-trained models (PTMs) to enhance task-specific generalization, they typically assume identical training and testing domains for each task and therefore perform poorly in DGCL. To this end, we propose adaptive Domain Transformation (DoT), an innovative PTMs-based approach tailored to DGCL. Inspired by the distributed-plus-hub theory of the human brain, DoT disentangles semantic- and domain-relevant information in representation learning, and adaptively transforms task representations across various domains for output alignment, ensuring balanced and generalized predictions. DoT serves as a plug-in strategy that greatly facilitates state-of-the-art CL baselines under both full parameter tuning and parameter-efficient tuning paradigms in DGCL, validated by extensive experiments. Also, DoT is shown to accumulate domain-generalizable knowledge from DGCL, and ensure resource efficiency with a lightweight implementation.", "date": "2025-10-19", "categories": []}
{"id": "2510.17200", "title": "EndoCIL: A Class-Incremental Learning Framework for Endoscopic Image Classification", "abstract": "Class-incremental learning (CIL) for endoscopic image analysis is crucial for real-world clinical applications, where diagnostic models should continuously adapt to evolving clinical data while retaining performance on previously learned ones. However, existing replay-based CIL methods fail to effectively mitigate catastrophic forgetting due to severe domain discrepancies and class imbalance inherent in endoscopic imaging. To tackle these challenges, we propose EndoCIL, a novel and unified CIL framework specifically tailored for endoscopic image diagnosis. EndoCIL incorporates three key components: Maximum Mean Discrepancy Based Replay (MDBR), employing a distribution-aligned greedy strategy to select diverse and representative exemplars, Prior Regularized Class Balanced Loss (PRCBL), designed to alleviate both inter-phase and intra-phase class imbalance by integrating prior class distributions and balance weights into the loss function, and Calibration of Fully-Connected Gradients (CFG), which adjusts the classifier gradients to mitigate bias toward new classes. Extensive experiments conducted on four public endoscopic datasets demonstrate that EndoCIL generally outperforms state-of-the-art CIL methods across varying buffer sizes and evaluation metrics. The proposed framework effectively balances stability and plasticity in lifelong endoscopic diagnosis, showing promising potential for clinical scalability and deployment.", "date": "2025-10-20", "categories": []}
{"id": "2510.15103", "title": "Continual Learning via Sparse Memory Finetuning", "abstract": "Modern language models are powerful, but typically static after deployment. A major obstacle to building models that continually learn over time is catastrophic forgetting, where updating on new data erases previously acquired capabilities. Motivated by the intuition that mitigating forgetting is challenging because trainable parameters are shared across all tasks, we investigate whether sparse parameter updates can enable learning without catastrophic forgetting. We introduce sparse memory finetuning, leveraging memory layer models (Berges et al., 2024), which are sparsely updated by design. By updating only the memory slots that are highly activated by a new piece of knowledge relative to usage on pretraining data, we reduce interference between new knowledge and the model's existing capabilities. We evaluate learning and forgetting compared to full finetuning and parameter-efficient finetuning with LoRA on two question answering tasks. We find that sparse memory finetuning learns new knowledge while exhibiting substantially less forgetting: while NaturalQuestions F1 drops by 89% after full finetuning on new facts and 71% with LoRA, sparse memory finetuning yields only an 11% drop with the same level of new knowledge acquisition. Our results suggest sparsity in memory layers offers a promising path toward continual learning in large language models.", "date": "2025-10-16", "categories": []}
{"id": "2505.22358", "title": "Adaptive Budget Allocation for Orthogonal-Subspace Adapter Tuning in LLMs Continual Learning", "abstract": "Large language models (LLMs) often suffer from catastrophic forgetting in continual learning (CL) scenarios, where performance on previously learned tasks degrades severely while training on sequentially arriving tasks. Although pioneering CL approaches using orthogonal subspaces can mitigate task interference, they typically employ fixed budget allocation, neglecting the varying complexity across tasks and layers. Besides, recent budget-adaptive tuning methods for LLMs often adopt multi-stage paradigms that decouple optimization and budget allocation. Such decoupling results in potential misalignment, which hinders those approaches' practical application in CL scenarios. To address these limitations, we propose OA-Adapter, a novel parameter-efficient approach for continual learning in LLMs that unifies dynamic budget adaptation with orthogonal subspace learning in an end-to-end training stage. Specifically, OA-Adapter introduces a dynamic bottleneck dimension adaptation mechanism that simultaneously allocates an efficient parameter budget and optimizes task objectives without misalignment.To effectively preserve previously acquired knowledge while coordinating with the dynamic budget allocation, orthogonal constraints are applied specifically between the parameter subspace of the current task and the dynamically allocated parameter subspaces of historical tasks. Experimental results on continual learning benchmarks demonstrate that OA-Adapter outperforms state-of-the-art methods in both accuracy and parameter efficiency. OA-Adapter achieves higher average accuracy while using 58.5% fewer parameters on the standard CL benchmark, and maintains its advantages on two larger benchmarks comprising 15 tasks.", "date": "2025-05-28", "categories": []}
{"id": "2507.08922", "title": "The Bayesian Approach to Continual Learning: An Overview", "abstract": "Continual learning is an online paradigm where a learner continually accumulates knowledge from different tasks encountered over sequential time steps. Importantly, the learner is required to extend and update its knowledge without forgetting about the learning experience acquired from the past, and while avoiding the need to retrain from scratch. Given its sequential nature and its resemblance to the way humans think, continual learning offers an opportunity to address several challenges which currently stand in the way of widening the range of applicability of deep models to further real-world problems. The continual need to update the learner with data arriving sequentially strikes inherent congruence between continual learning and Bayesian inference which provides a principal platform to keep updating the prior beliefs of a model given new data, without completely forgetting the knowledge acquired from the old data. This survey inspects different settings of Bayesian continual learning, namely task-incremental learning and class-incremental learning. We begin by discussing definitions of continual learning along with its Bayesian setting, as well as the links with related fields, such as domain adaptation, transfer learning and meta-learning. Afterwards, we introduce a taxonomy offering a comprehensive categorization of algorithms belonging to the Bayesian continual learning paradigm. Meanwhile, we analyze the state-of-the-art while zooming in on some of the most prominent Bayesian continual learning algorithms to date. Furthermore, we shed some light on links between continual learning and developmental psychology, and correspondingly introduce analogies between both fields. We follow that with a discussion of current challenges, and finally conclude with potential areas for future research on Bayesian continual learning.", "date": "2025-07-11", "categories": []}
{"id": "2509.18133", "title": "Self-Evolving LLMs via Continual Instruction Tuning", "abstract": "In real-world industrial settings, large language models (LLMs) must learn continually to keep pace with diverse and evolving tasks, requiring self-evolution to refine knowledge under dynamic data distributions. However, existing continual learning (CL) approaches, such as replay and parameter isolation, often suffer from catastrophic forgetting: training on new tasks degrades performance on earlier ones by overfitting to the new distribution and weakening generalization.We propose MoE-CL, a parameter-efficient adversarial mixture-of-experts framework for industrial-scale, self-evolving continual instruction tuning of LLMs. MoE-CL uses a dual-expert design: (1) a dedicated LoRA expert per task to preserve task-specific knowledge via parameter independence, mitigating forgetting; and (2) a shared LoRA expert to enable cross-task transfer. To prevent transferring task-irrelevant noise through the shared pathway, we integrate a task-aware discriminator within a GAN. The discriminator encourages the shared expert to pass only task-aligned information during sequential training. Through adversarial learning, the shared expert acquires generalized representations that mimic the discriminator, while dedicated experts retain task-specific details, balancing knowledge retention and cross-task generalization and thereby supporting self-evolution.Extensive experiments on the public MTL5 benchmark and an industrial Tencent3 benchmark validate the effectiveness of MoE-CL for continual instruction tuning. In real-world A/B testing for content compliance review on the Tencent Video platform, MoE-CL reduced manual review costs by 15.3%. These results demonstrate that MoE-CL is practical for large-scale industrial deployment where continual adaptation and stable transfer are critical.", "date": "2025-09-14", "categories": []}
{"id": "2510.13008", "title": "CurLL: A Developmental Framework to Evaluate Continual Learning in Language Models", "abstract": "We introduce a comprehensive continual learning dataset and benchmark (CurlL) grounded in human developmental trajectories from ages 5-10, enabling systematic and fine-grained assessment of models' ability to progressively acquire new skills. CurlL spans five developmental stages (0-4) covering ages 5-10, supported by a skill graph that breaks down broad skills into smaller abilities, concrete goals, and measurable indicators, while also capturing which abilities build on others. We generate a 23.4B-token synthetic dataset with controlled skill progression, vocabulary complexity, and format diversity, comprising paragraphs, comprehension-based QA (CQA), skill-testing QA (CSQA), and instruction-response (IR) pairs. Stage-wise token counts range from 2.12B to 6.78B tokens, supporting precise analysis of forgetting, forward transfer, and backward transfer. Using a 135M-parameter transformer trained under independent, joint, and sequential (continual) setups, we show trade-offs in skill retention and transfer efficiency. By mirroring human learning patterns and providing fine-grained control over skill dependencies, this work advances continual learning evaluations for language models.", "date": "2025-10-14", "categories": []}
{"id": "2510.13567", "title": "DOLFIN: Balancing Stability and Plasticity in Federated Continual Learning", "abstract": "Federated continual learning (FCL) enables models to learn new tasks across multiple distributed clients, protecting privacy and without forgetting previously acquired knowledge. However, current methods face challenges balancing performance, privacy preservation, and communication efficiency. We introduce a Distributed Online LoRA for Federated INcremental learning method DOLFIN, a novel approach combining Vision Transformers with low-rank adapters designed to efficiently and stably learn new tasks in federated environments. Our method leverages LoRA for minimal communication overhead and incorporates DualGradient Projection Memory (DualGPM) to prevent forgetting. Evaluated on CIFAR-100, ImageNet-R, ImageNet-A, and CUB-200 under two Dirichlet heterogeneity settings, DOLFIN consistently surpasses six strong baselines in final average accuracy while matching their memory footprint. Orthogonal low-rank adapters offer an effective and scalable solution for privacy-preserving continual learning in federated settings.", "date": "2025-10-15", "categories": []}
{"id": "2211.13969", "title": "Unsupervised Continual Semantic Adaptation through Neural Rendering", "abstract": "An increasing amount of applications rely on data-driven models that are deployed for perception tasks across a sequence of scenes. Due to the mismatch between training and deployment data, adapting the model on the new scenes is often crucial to obtain good performance. In this work, we study continual multi-scene adaptation for the task of semantic segmentation, assuming that no ground-truth labels are available during deployment and that performance on the previous scenes should be maintained. We propose training a Semantic-NeRF network for each scene by fusing the predictions of a segmentation model and then using the view-consistent rendered semantic labels as pseudo-labels to adapt the model. Through joint training with the segmentation model, the Semantic-NeRF model effectively enables 2D-3D knowledge transfer. Furthermore, due to its compact size, it can be stored in a long-term memory and subsequently used to render data from arbitrary viewpoints to reduce forgetting. We evaluate our approach on ScanNet, where we outperform both a voxel-based baseline and a state-of-the-art unsupervised domain adaptation method.", "date": "2022-11-25", "categories": []}
{"id": "2305.14782", "title": "IBCL: Zero-shot Model Generation under Stability-Plasticity Trade-offs", "abstract": "Algorithms that balance the stability-plasticity trade-off are well studied in the Continual Learning literature. However, only a few focus on obtaining models for specified trade-off preferences. When solving the problem of continual learning under specific trade-offs (CLuST), state-of-the-art techniques leverage rehearsal-based learning, which requires retraining when a model corresponding to a new trade-off preference is requested. This is inefficient, since there potentially exists a significant number of different trade-offs, and a large number of models may be requested. As a response, we propose Imprecise Bayesian Continual Learning (IBCL), an algorithm that tackles CLuST efficiently. IBCL replaces retraining with a constant-time convex combination. Given a new task, IBCL (1) updates the knowledge base as a convex hull of model parameter distributions, and (2) generates one Pareto-optimal model per given trade-off via convex combination without additional training. That is, obtaining models corresponding to specified trade-offs via IBCL is zero-shot. Experiments whose baselines are current CLuST algorithms show that IBCL improves classification by at most 44% on average per task accuracy, and by 45% on peak per task accuracy while maintaining a near-zero to positive backward transfer, with memory overheads converging to constants. In addition, its training overhead, measured by the number of batch updates, remains constant at every task, regardless of the number of preferences requested. IBCL also improves multi-objective reinforcement learning tasks by maintaining the same Pareto front hypervolume, while significantly reducing the training cost. Details can be found at: https://github.com/ibcl-anon/ibcl.", "date": "2023-05-24", "categories": []}
{"id": "2501.08737", "title": "Resource-Constrained Federated Continual Learning: What Does Matter?", "abstract": "Federated Continual Learning (FCL) aims to enable sequentially privacy-preserving model training on streams of incoming data that vary in edge devices by preserving previous knowledge while adapting to new data. Current FCL literature focuses on restricted data privacy and access to previously seen data while imposing no constraints on the training overhead. This is unreasonable for FCL applications in real-world scenarios, where edge devices are primarily constrained by resources such as storage, computational budget, and label rate. We revisit this problem with a large-scale benchmark and analyze the performance of state-of-the-art FCL approaches under different resource-constrained settings. Various typical FCL techniques and six datasets in two incremental learning scenarios (Class-IL and Domain-IL) are involved in our experiments. Through extensive experiments amounting to a total of over 1,000+ GPU hours, we find that, under limited resource-constrained settings, existing FCL approaches, with no exception, fail to achieve the expected performance. Our conclusions are consistent in the sensitivity analysis. This suggests that most existing FCL methods are particularly too resource-dependent for real-world deployment. Moreover, we study the performance of typical FCL techniques with resource constraints and shed light on future research directions in FCL.", "date": "2025-01-15", "categories": []}
{"id": "2508.16463", "title": "Modular Embedding Recomposition for Incremental Learning", "abstract": "The advent of pre-trained Vision-Language Models (VLMs) has significantly transformed Continual Learning (CL), mainly due to their zero-shot classification abilities. Such proficiency makes VLMs well-suited for real-world applications, enabling robust performance on novel unseen classes without requiring adaptation. However, fine-tuning remains essential when downstream tasks deviate significantly from the pre-training domain. Prior CL approaches primarily focus on preserving the zero-shot capabilities of VLMs during incremental fine-tuning on a downstream task. We take a step further by devising an approach that transforms preservation into enhancement of the zero-shot capabilities of VLMs. Our approach, named MoDular Embedding Recomposition (MoDER), introduces a modular framework that trains multiple textual experts, each specialized in a single seen class, and stores them in a foundational hub. At inference time, for each unseen class, we query the hub and compose the retrieved experts to synthesize a refined prototype that improves classification. We show the effectiveness of our method across two popular zero-shot incremental protocols, Class-IL and MTIL, comprising a total of 14 datasets. The codebase is available at https://github.com/aimagelab/mammoth.", "date": "2025-08-22", "categories": []}
{"id": "2510.11842", "title": "Balancing Synthetic Data and Replay for Enhancing Task-Specific Capabilities", "abstract": "Adapting language models to new tasks through continued pretraining faces a fundamental trade-off: models must learn new capabilities while avoiding catastrophic forgetting of existing knowledge. While prior work has studied synthetic data generation techniques, the optimal replay ratios for balancing task performance and knowledge retention under computational constraints remain poorly understood. We present a comprehensive empirical study investigating the interplay between replay ratio configuration and computational budget when adapting language models to new tasks. Using the bAbI reasoning tasks as our target objective, we apply synthetic data generation and systematically evaluate different total token budgets and replay ratio configurations. We analyze their effects on both task mastery and general knowledge retention. Our experiments reveal an optimal configuration that balances task-specific performance with general knowledge retention. Based on our findings, we provide empirically-grounded guidelines for selecting replay ratios based on computational budget, enabling practitioners to achieve strong task adaptation with significantly reduced training costs.", "date": "2025-10-13", "categories": []}
{"id": "2510.12107", "title": "DRL: Discriminative Representation Learning with Parallel Adapters for Class Incremental Learning", "abstract": "With the excellent representation capabilities of Pre-Trained Models (PTMs), remarkable progress has been made in non-rehearsal Class-Incremental Learning (CIL) research. However, it remains an extremely challenging task due to three conundrums: increasingly large model complexity, non-smooth representation shift during incremental learning and inconsistency between stage-wise sub-problem optimization and global inference. In this work, we propose the Discriminative Representation Learning (DRL) framework to specifically address these challenges. To conduct incremental learning effectively and yet efficiently, the DRL's network, called Incremental Parallel Adapter (IPA) network, is built upon a PTM and increasingly augments the model by learning a lightweight adapter with a small amount of parameter learning overhead in each incremental stage. The adapter is responsible for adapting the model to new classes, it can inherit and propagate the representation capability from the current model through parallel connection between them by a transfer gate. As a result, this design guarantees a smooth representation shift between different incremental stages. Furthermore, to alleviate inconsistency and enable comparable feature representations across incremental stages, we design the Decoupled Anchor Supervision (DAS). It decouples constraints of positive and negative samples by respectively comparing them with the virtual anchor. This decoupling promotes discriminative representation learning and aligns the feature spaces learned at different stages, thereby narrowing the gap between stage-wise local optimization over a subset of data and global inference across all classes. Extensive experiments on six benchmarks reveal that our DRL consistently outperforms other state-of-the-art methods throughout the entire CIL period while maintaining high efficiency in both training and inference phases.", "date": "2025-10-14", "categories": []}
{"id": "2510.12150", "title": "Class-aware Domain Knowledge Fusion and Fission for Continual Test-Time Adaptation", "abstract": "Continual Test-Time Adaptation (CTTA) aims to quickly fine-tune the model during the test phase so that it can adapt to multiple unknown downstream domain distributions without pre-acquiring downstream domain data. To this end, existing advanced CTTA methods mainly reduce the catastrophic forgetting of historical knowledge caused by irregular switching of downstream domain data by restoring the initial model or reusing historical models. However, these methods are usually accompanied by serious insufficient learning of new knowledge and interference from potentially harmful historical knowledge, resulting in severe performance degradation. To this end, we propose a class-aware domain Knowledge Fusion and Fission method for continual test-time adaptation, called KFF, which adaptively expands and merges class-aware domain knowledge in old and new domains according to the test-time data from different domains, where discriminative historical knowledge can be dynamically accumulated. Specifically, considering the huge domain gap within streaming data, a domain Knowledge FIssion (KFI) module is designed to adaptively separate new domain knowledge from a paired class-aware domain prompt pool, alleviating the impact of negative knowledge brought by old domains that are distinct from the current domain. Besides, to avoid the cumulative computation and storage overheads from continuously fissioning new knowledge, a domain Knowledge FUsion (KFU) module is further designed to merge the fissioned new knowledge into the existing knowledge pool with minimal cost, where a greedy knowledge dynamic merging strategy is designed to improve the compatibility of new and old knowledge while keeping the computational efficiency. Extensive experiments on the ImageNet-C dataset verify the effectiveness of our proposed method against other methods.", "date": "2025-10-14", "categories": []}
{"id": "2510.07648", "title": "Continual Learning for Adaptive AI Systems", "abstract": "Continual learning the ability of a neural network to learn multiple sequential tasks without catastrophic forgetting remains a central challenge in developing adaptive artificial intelligence systems. While deep learning models achieve state-of-the-art performance across domains, they remain limited by overfitting and forgetting. This paper introduces Cluster-Aware Replay (CAR), a hybrid continual learning framework that integrates a small, class-balanced replay buffer with a regularization term based on Inter-Cluster Fitness (ICF) in the feature space. The ICF loss penalizes overlapping feature representations between new and previously learned tasks, encouraging geometric separation in the latent space and reducing interference. Using the standard five-task Split CIFAR-10 benchmark with a ResNet-18 backbone, initial experiments demonstrate that CAR better preserves earlier task performance compared to fine-tuning alone. These findings are preliminary but highlight feature-space regularization as a promising direction for mitigating catastrophic forgetting.", "date": "2025-10-09", "categories": []}
{"id": "2510.10071", "title": "ADEPT: Continual Pretraining via Adaptive Expansion and Dynamic Decoupled Tuning", "abstract": "Conventional continual pretraining (CPT) for large language model (LLM) domain adaptation often suffers from catastrophic forgetting and limited domain capacity. Existing strategies adopt layer expansion, introducing additional trainable parameters to accommodate new knowledge. However, the uniform expansion and updates still entangle general and domain learning, undermining its effectiveness. Our pilot studies reveal that LLMs exhibit functional specialization, where layers and units differentially encode general-critical capabilities, suggesting that parameter expansion and optimization should be function-aware. We then propose ADEPT, Adaptive Expansion and Dynamic Decoupled Tuning for continual pretraining, a two-stage framework for domain-adaptive CPT. ADEPT first performs General-Competence Guided Selective Layer Expansion, duplicating layers least critical for the general domain to increase representational capacity while minimizing interference with general knowledge. It then applies Adaptive Unit-Wise Decoupled Tuning, disentangling parameter units within expanded layers according to their general-domain importance and assigning asymmetric learning rates to balance knowledge injection and retention. Experiments on mathematical and medical benchmarks show that ADEPT outperforms full-parameter CPT by up to 5.76% on the general domain and 5.58% on the target domain with only 15% of parameters tuned and less than 50% training time. Ablation studies, theoretical analysis, and extended investigations further demonstrate the necessity of targeted expansion and decoupled optimization, providing new principles for efficient and robust domain-adaptive CPT. Our code is open-sourced at https://github.com/PuppyKnightUniversity/ADEPT", "date": "2025-10-11", "categories": []}
{"id": "2510.11444", "title": "GenCNER: A Generative Framework for Continual Named Entity Recognition", "abstract": "Traditional named entity recognition (NER) aims to identify text mentions into pre-defined entity types. Continual Named Entity Recognition (CNER) is introduced since entity categories are continuously increasing in various real-world scenarios. However, existing continual learning (CL) methods for NER face challenges of catastrophic forgetting and semantic shift of non-entity type. In this paper, we propose GenCNER, a simple but effective Generative framework for CNER to mitigate the above drawbacks. Specifically, we skillfully convert the CNER task into sustained entity triplet sequence generation problem and utilize a powerful pre-trained seq2seq model to solve it. Additionally, we design a type-specific confidence-based pseudo labeling strategy along with knowledge distillation (KD) to preserve learned knowledge and alleviate the impact of label noise at the triplet level. Experimental results on two benchmark datasets show that our framework outperforms previous state-of-the-art methods in multiple CNER settings, and achieves the smallest gap compared with non-CL results.", "date": "2025-10-13", "categories": []}
{"id": "2403.19979", "title": "Continual Adapter Tuning with Semantic Shift Compensation for Class-Incremental Learning", "abstract": "Class-incremental learning (CIL) aims to enable models to continuously learn new classes while overcoming catastrophic forgetting. The introduction of pre-trained models has brought new tuning paradigms to CIL. In this paper, we revisit different parameter-efficient tuning (PET) methods within the context of continual learning. We observe that adapter tuning demonstrates superiority over prompt-based methods, even without parameter expansion in each learning session. Motivated by this, we propose incrementally tuning the shared adapter without imposing parameter update constraints, enhancing the learning capacity of the backbone. Additionally, we employ feature sampling from stored prototypes to retrain a unified classifier, further improving its performance. We estimate the semantic shift of old prototypes without access to past samples and update stored prototypes session by session. Our proposed method eliminates model expansion and avoids retaining any image samples. It surpasses previous pre-trained model-based CIL methods and demonstrates remarkable continual learning capabilities. Experimental results on five CIL benchmarks validate the effectiveness of our approach, achieving state-of-the-art (SOTA) performance.", "date": "2024-03-29", "categories": []}
{"id": "2510.09181", "title": "On the Implicit Adversariality of Catastrophic Forgetting in Deep Continual Learning", "abstract": "Continual learning seeks the human-like ability to accumulate new skills in machine intelligence. Its central challenge is catastrophic forgetting, whose underlying cause has not been fully understood for deep networks. In this paper, we demystify catastrophic forgetting by revealing that the new-task training is implicitly an adversarial attack against the old-task knowledge. Specifically, the new-task gradients automatically and accurately align with the sharp directions of the old-task loss landscape, rapidly increasing the old-task loss. This adversarial alignment is intriguingly counter-intuitive because the sharp directions are too sparsely distributed to align with by chance. To understand it, we theoretically show that it arises from training's low-rank bias, which, through forward and backward propagation, confines the two directions into the same low-dimensional subspace, facilitating alignment. Gradient projection (GP) methods, a representative family of forgetting-mitigating methods, reduce adversarial alignment caused by forward propagation, but cannot address the alignment due to backward propagation. We propose backGP to address it, which reduces forgetting by 10.8% and improves accuracy by 12.7% on average over GP methods.", "date": "2025-10-10", "categories": []}
{"id": "2506.02011", "title": "OASIS: Online Sample Selection for Continual Visual Instruction Tuning", "abstract": "In continual instruction tuning (CIT) scenarios, where new instruction tuning data continuously arrive in an online streaming manner, training delays from large-scale data significantly hinder real-time adaptation. Data selection can mitigate this overhead, but existing strategies often rely on pretrained reference models, which are impractical in CIT setups since future data are unknown. Recent reference model-free online sample selection methods address this, but typically select a fixed number of samples per batch (e.g., top-k), making them vulnerable to distribution shifts where informativeness varies across batches. To address these limitations, we propose OASIS, an adaptive online sample selection approach for CIT that (1) selects informative samples by estimating each sample's informativeness relative to all previously seen data, beyond batch-level constraints, and (2) minimizes informative redundancy of selected samples through iterative selection score updates. Experiments on various large foundation models show that OASIS, using only 25 percent of the data, achieves comparable performance to full-data training and outperforms the state-of-the-art sampling methods.", "date": "2025-05-27", "categories": []}
{"id": "2510.08152", "title": "DACIP-RC: Domain Adaptive Continual Instruction Pre-Training via Reading Comprehension on Business Conversations", "abstract": "The rapid advancements in Large Language Models (LLMs) have enabled their adoption in real-world industrial scenarios for various natural language processing tasks. However, the high inference cost of large-scale LLMs makes their deployment impractical, necessitating the use of smaller models. Despite their efficiency, smaller LLMs lack robust zero-shot instruction-following capabilities across diverse domains, limiting their adaptability to dynamic user requirements. Traditional fine-tuning approaches exacerbate this issue by inducing catastrophic forgetting, reducing the model's generalization ability for unseen tasks. In this paper, we propose Domain Adaptive Continual Instruction Pre-Training via Reading Comprehension (DACIP-RC), a continual pre-training technique that enhances smaller LLMs' domain adaptability for business conversational tasks. Unlike conventional pre-training approaches that rely on next-token prediction, DACIP-RC generates diverse task instructions and responses via reading comprehension on conversation transcripts, enabling better instruction generalization. Our empirical evaluations demonstrate that DACIP-RC significantly improves zero-shot generalization across a wide range of business conversational tasks, including meeting summarization, action item generation, and call purpose identification. To the best of our knowledge, this is the first work to apply instruction pre-training on business conversational data, providing insights into how industries can leverage proprietary datasets for domain adaptation.", "date": "2025-10-09", "categories": []}
{"id": "2510.06819", "title": "The Unreasonable Effectiveness of Randomized Representations in Online Continual Graph Learning", "abstract": "Catastrophic forgetting is one of the main obstacles for Online Continual Graph Learning (OCGL), where nodes arrive one by one, distribution drifts may occur at any time and offline training on task-specific subgraphs is not feasible. In this work, we explore a surprisingly simple yet highly effective approach for OCGL: we use a fixed, randomly initialized encoder to generate robust and expressive node embeddings by aggregating neighborhood information, training online only a lightweight classifier. By freezing the encoder, we eliminate drifts of the representation parameters, a key source of forgetting, obtaining embeddings that are both expressive and stable. When evaluated across several OCGL benchmarks, despite its simplicity and lack of memory buffer, this approach yields consistent gains over state-of-the-art methods, with surprising improvements of up to 30% and performance often approaching that of the joint offline-training upper bound. These results suggest that in OCGL, catastrophic forgetting can be minimized without complex replay or regularization by embracing architectural simplicity and stability.", "date": "2025-10-08", "categories": []}
{"id": "2510.06842", "title": "Continual Action Quality Assessment via Adaptive Manifold-Aligned Graph Regularization", "abstract": "Action Quality Assessment (AQA) quantifies human actions in videos, supporting applications in sports scoring, rehabilitation, and skill evaluation. A major challenge lies in the non-stationary nature of quality distributions in real-world scenarios, which limits the generalization ability of conventional methods. We introduce Continual AQA (CAQA), which equips AQA with Continual Learning (CL) capabilities to handle evolving distributions while mitigating catastrophic forgetting. Although parameter-efficient fine-tuning of pretrained models has shown promise in CL for image classification, we find it insufficient for CAQA. Our empirical and theoretical analyses reveal two insights: (i) Full-Parameter Fine-Tuning (FPFT) is necessary for effective representation learning; yet (ii) uncontrolled FPFT induces overfitting and feature manifold shift, thereby aggravating forgetting. To address this, we propose Adaptive Manifold-Aligned Graph Regularization (MAGR++), which couples backbone fine-tuning that stabilizes shallow layers while adapting deeper ones with a two-step feature rectification pipeline: a manifold projector to translate deviated historical features into the current representation space, and a graph regularizer to align local and global distributions. We construct four CAQA benchmarks from three datasets with tailored evaluation protocols and strong baselines, enabling systematic cross-dataset comparison. Extensive experiments show that MAGR++ achieves state-of-the-art performance, with average correlation gains of 3.6% offline and 12.2% online over the strongest baseline, confirming its robustness and effectiveness. Our code is available at https://github.com/ZhouKanglei/MAGRPP.", "date": "2025-10-08", "categories": []}
{"id": "2509.23027", "title": "Understanding Catastrophic Interference: On the Identifibility of Latent Representations", "abstract": "Catastrophic interference, also known as catastrophic forgetting, is a fundamental challenge in machine learning, where a trained learning model progressively loses performance on previously learned tasks when adapting to new ones. In this paper, we aim to better understand and model the catastrophic interference problem from a latent representation learning point of view, and propose a novel theoretical framework that formulates catastrophic interference as an identification problem. Our analysis demonstrates that the forgetting phenomenon can be quantified by the distance between partial-task aware (PTA) and all-task aware (ATA) setups. Building upon recent advances in identifiability theory, we prove that this distance can be minimized through identification of shared latent variables between these setups. When learning, we propose our method \\ourmeos with two-stage training strategy: First, we employ maximum likelihood estimation to learn the latent representations from both PTA and ATA configurations. Subsequently, we optimize the KL divergence to identify and learn the shared latent variables. Through theoretical guarantee and empirical validations, we establish that identifying and learning these shared representations can effectively mitigate catastrophic interference in machine learning systems. Our approach provides both theoretical guarantees and practical performance improvements across both synthetic and benchmark datasets.", "date": "2025-09-27", "categories": []}
{"id": "2405.18925", "title": "Federated Continual Learning Goes Online: Uncertainty-Aware Memory Management for Vision Tasks and Beyond", "abstract": "Given the ability to model more realistic and dynamic problems, Federated Continual Learning (FCL) has been increasingly investigated recently. A well-known problem encountered in this setting is the so-called catastrophic forgetting, for which the learning model is inclined to focus on more recent tasks while forgetting the previously learned knowledge. The majority of the current approaches in FCL propose generative-based solutions to solve said problem. However, this setting requires multiple training epochs over the data, implying an offline setting where datasets are stored locally and remain unchanged over time. Furthermore, the proposed solutions are tailored for vision tasks solely. To overcome these limitations, we propose a new approach to deal with different modalities in the online scenario where new data arrive in streams of mini-batches that can only be processed once. To solve catastrophic forgetting, we propose an uncertainty-aware memory-based approach. Specifically, we suggest using an estimator based on the Bregman Information (BI) to compute the model's variance at the sample level. Through measures of predictive uncertainty, we retrieve samples with specific characteristics, and - by retraining the model on such samples - we demonstrate the potential of this approach to reduce the forgetting effect in realistic settings while maintaining data confidentiality and competitive communication efficiency compared to state-of-the-art approaches.", "date": "2024-05-29", "categories": []}
{"id": "2411.08224", "title": "Joint Diffusion models in Continual Learning", "abstract": "In this work, we introduce JDCL - a new method for continual learning with generative rehearsal based on joint diffusion models. Neural networks suffer from catastrophic forgetting defined as abrupt loss in the model's performance when retrained with additional data coming from a different distribution. Generative-replay-based continual learning methods try to mitigate this issue by retraining a model with a combination of new and rehearsal data sampled from a generative model. In this work, we propose to extend this idea by combining a continually trained classifier with a diffusion-based generative model into a single - jointly optimized neural network. We show that such shared parametrization, combined with the knowledge distillation technique allows for stable adaptation to new tasks without catastrophic forgetting. We evaluate our approach on several benchmarks, where it outperforms recent state-of-the-art generative replay techniques. Additionally, we extend our method to the semi-supervised continual learning setup, where it outperforms competing buffer-based replay techniques, and evaluate, in a self-supervised manner, the quality of trained representations.", "date": "2024-11-12", "categories": []}
{"id": "2506.00037", "title": "Query Drift Compensation: Enabling Compatibility in Continual Learning of Retrieval Embedding Models", "abstract": "Text embedding models enable semantic search, powering several NLP applications like Retrieval Augmented Generation by efficient information retrieval (IR). However, text embedding models are commonly studied in scenarios where the training data is static, thus limiting its applications to dynamic scenarios where new training data emerges over time. IR methods generally encode a huge corpus of documents to low-dimensional embeddings and store them in a database index. During retrieval, a semantic search over the corpus is performed and the document whose embedding is most similar to the query embedding is returned. When updating an embedding model with new training data, using the already indexed corpus is suboptimal due to the non-compatibility issue, since the model which was used to obtain the embeddings of the corpus has changed. While re-indexing of old corpus documents using the updated model enables compatibility, it requires much higher computation and time. Thus, it is critical to study how the already indexed corpus can still be effectively used without the need of re-indexing. In this work, we establish a continual learning benchmark with large-scale datasets and continually train dense retrieval embedding models on query-document pairs from new datasets in each task and observe forgetting on old tasks due to significant drift of embeddings. We employ embedding distillation on both query and document embeddings to maintain stability and propose a novel query drift compensation method during retrieval to project new model query embeddings to the old embedding space. This enables compatibility with previously indexed corpus embeddings extracted using the old model and thus reduces the forgetting. We show that the proposed method significantly improves performance without any re-indexing. Code is available at https://github.com/dipamgoswami/QDC.", "date": "2025-05-27", "categories": []}
{"id": "2509.16929", "title": "K-DeCore: Facilitating Knowledge Transfer in Continual Structured Knowledge Reasoning via Knowledge Decoupling", "abstract": "Continual Structured Knowledge Reasoning (CSKR) focuses on training models to handle sequential tasks, where each task involves translating natural language questions into structured queries grounded in structured knowledge. Existing general continual learning approaches face significant challenges when applied to this task, including poor generalization to heterogeneous structured knowledge and inefficient reasoning due to parameter growth as tasks increase. To address these limitations, we propose a novel CSKR framework, \\textsc{K-DeCore}, which operates with a fixed number of tunable parameters. Unlike prior methods, \\textsc{K-DeCore} introduces a knowledge decoupling mechanism that disentangles the reasoning process into task-specific and task-agnostic stages, effectively bridging the gaps across diverse tasks. Building on this foundation, \\textsc{K-DeCore} integrates a dual-perspective memory consolidation mechanism for distinct stages and introduces a structure-guided pseudo-data synthesis strategy to further enhance the model's generalization capabilities. Extensive experiments on four benchmark datasets demonstrate the superiority of \\textsc{K-DeCore} over existing continual learning methods across multiple metrics, leveraging various backbone large language models.", "date": "2025-09-21", "categories": []}
{"id": "2503.10439", "title": "EFC++: Elastic Feature Consolidation with Prototype Re-balancing for Cold Start Exemplar-free Incremental Learning", "abstract": "Exemplar-free Class Incremental Learning (EFCIL) aims to learn from a sequence of tasks without having access to previous task data. In this paper, we consider the challenging Cold Start scenario in which insufficient data is available in the first task to learn a high-quality backbone. This is especially challenging for EFCIL since it requires high plasticity, resulting in feature drift which is difficult to compensate for in the exemplar-free setting. To address this problem, we propose an effective approach to consolidate feature representations by regularizing drift in directions highly relevant to previous tasks while employing prototypes to reduce task-recency bias. Our approach, which we call Elastic Feature Consolidation++ (EFC++) exploits a tractable second-order approximation of feature drift based on a proposed Empirical Feature Matrix (EFM). The EFM induces a pseudo-metric in feature space which we use to regularize feature drift in important directions and to update Gaussian prototypes. In addition, we introduce a post-training prototype re-balancing phase that updates classifiers to compensate for feature drift. Experimental results on CIFAR-100, Tiny-ImageNet, ImageNet-Subset, ImageNet-1K and DomainNet demonstrate that EFC++ is better able to learn new tasks by maintaining model plasticity and significantly outperforms the state-of-the-art.", "date": "2025-03-13", "categories": []}
{"id": "2510.02631", "title": "Deep Generative Continual Learning using Functional LoRA: FunLoRA", "abstract": "Continual adaptation of deep generative models holds tremendous potential and critical importance, given their rapid and expanding usage in text and vision based applications. Incremental training, however, remains highly challenging due to catastrophic forgetting phenomenon, which makes it difficult for neural networks to effectively incorporate new knowledge. A common strategy consists in retraining the generative model on its own synthetic data in order to mitigate forgetting. Yet, such an approach faces two major limitations: (i) the continually increasing training time eventually becomes intractable, and (ii) reliance on synthetic data inevitably leads to long-term performance degradation, since synthetic samples lack the richness of real training data. In this paper, we attenuate these issues by designing a novel and more expressive conditioning mechanism for generative models based on low rank adaptation (LoRA), that exclusively employs rank 1 matrices, whose reparametrized matrix rank is functionally increased using carefully selected functions -- and dubbed functional LoRA: FunLoRA. Using this dynamic conditioning, the generative model is guaranteed to avoid catastrophic forgetting and needs only to be trained on data from the current task. Extensive experiments using flow-matching based models trained from scratch, showcase that our proposed parameter-efficient fine-tuning (PEFT) method surpasses prior state-of-the-art results based on diffusion models, reaching higher classification accuracy scores, while only requiring a fraction of the memory cost and sampling time.", "date": "2025-10-03", "categories": []}
{"id": "2501.12635", "title": "Multiple Queries with Multiple Keys: A Precise Prompt Matching Paradigm for Prompt-based Continual Learning", "abstract": "Continual learning requires machine learning models to continuously acquire new knowledge in dynamic environments while avoiding the forgetting of previous knowledge. Prompt-based continual learning methods effectively address the issue of catastrophic forgetting through prompt expansion and selection. However, existing approaches often suffer from low accuracy in prompt selection, which can result in the model receiving biased knowledge and making biased predictions. To address this issue, we propose the Multiple Queries with Multiple Keys (MQMK) prompt matching paradigm for precise prompt selection. The goal of MQMK is to select the prompts whose training data distribution most closely matches that of the test sample. Specifically, Multiple Queries enable precise breadth search by introducing task-specific knowledge, while Multiple Keys perform deep search by representing the feature distribution of training samples at a fine-grained level. Each query is designed to perform local matching with a designated task to reduce interference across queries. Experiments show that MQMK enhances the prompt matching rate by over 30\\% in challenging scenarios and achieves state-of-the-art performance on three widely adopted continual learning benchmarks. The code is available at https://github.com/DunweiTu/MQMK.", "date": "2025-01-22", "categories": []}
{"id": "2510.01649", "title": "Source-Free Cross-Domain Continual Learning", "abstract": "Although existing cross-domain continual learning approaches successfully address many streaming tasks having domain shifts, they call for a fully labeled source domain hindering their feasibility in the privacy constrained environments. This paper goes one step ahead with the problem of source-free cross-domain continual learning where the use of source-domain samples are completely prohibited. We propose the idea of rehearsal-free frequency-aware dynamic prompt collaborations (REFEREE) to cope with the absence of labeled source-domain samples in realm of cross-domain continual learning. REFEREE is built upon a synergy between a source-pre-trained model and a large-scale vision-language model, thus overcoming the problem of sub-optimal generalizations when relying only on a source pre-trained model. The domain shift problem between the source domain and the target domain is handled by a frequency-aware prompting technique encouraging low-frequency components while suppressing high-frequency components. This strategy generates frequency-aware augmented samples, robust against noisy pseudo labels. The noisy pseudo-label problem is further addressed with the uncertainty-aware weighting strategy where the mean and covariance matrix are weighted by prediction uncertainties, thus mitigating the adverse effects of the noisy pseudo label. Besides, the issue of catastrophic forgetting (CF) is overcome by kernel linear discriminant analysis (KLDA) where the backbone network is frozen while the classification is performed using the linear discriminant analysis approach guided by the random kernel method. Our rigorous numerical studies confirm the advantage of our approach where it beats prior arts having access to source domain samples with significant margins.", "date": "2025-10-02", "categories": []}
{"id": "2506.05453", "title": "MLLM-CL: Continual Learning for Multimodal Large Language Models", "abstract": "Recent Multimodal Large Language Models (MLLMs) excel in vision-language understanding but face challenges in adapting to dynamic real-world scenarios that require continuous integration of new knowledge and skills. While continual learning (CL) offers a potential solution, existing benchmarks and methods suffer from critical limitations. In this paper, we introduce MLLM-CL, a novel benchmark encompassing domain and ability continual learning, where the former focuses on independently and identically distributed (IID) evaluation across evolving mainstream domains, whereas the latter evaluates on non-IID scenarios with new model abilities. Methodologically, we propose preventing catastrophic interference through parameter isolation and an MLLM-based routing mechanism. Extensive experiments demonstrate that our approach can integrate domain-specific knowledge and functional abilities with minimal forgetting, significantly outperforming existing methods. Our benchmark and code are available at https://github.com/bjzhb666/MLLM-CL.", "date": "2025-06-05", "categories": []}
{"id": "2510.00467", "title": "Rehearsal-free and Task-free Online Continual Learning With Contrastive Prompt", "abstract": "The main challenge of continual learning is \\textit{catastrophic forgetting}. Because of processing data in one pass, online continual learning (OCL) is one of the most difficult continual learning scenarios. To address catastrophic forgetting in OCL, some existing studies use a rehearsal buffer to store samples and replay them in the later learning process, other studies do not store samples but assume a sequence of learning tasks so that the task identities can be explored. However, storing samples may raise data security or privacy concerns and it is not always possible to identify the boundaries between learning tasks in one pass of data processing. It motivates us to investigate rehearsal-free and task-free OCL (F2OCL). By integrating prompt learning with an NCM classifier, this study has effectively tackled catastrophic forgetting without storing samples and without usage of task boundaries or identities. The extensive experimental results on two benchmarks have demonstrated the effectiveness of the proposed method.", "date": "2025-10-01", "categories": []}
{"id": "2510.00475", "title": "Diagnosing Shortcut-Induced Rigidity in Continual Learning: The Einstellung Rigidity Index (ERI)", "abstract": "Deep neural networks frequently exploit shortcut features, defined as incidental correlations between inputs and labels without causal meaning. Shortcut features undermine robustness and reduce reliability under distribution shifts. In continual learning (CL), the consequences of shortcut exploitation can persist and intensify: weights inherited from earlier tasks bias representation reuse toward whatever features most easily satisfied prior labels, mirroring the cognitive Einstellung effect, a phenomenon where past habits block optimal solutions. Whereas catastrophic forgetting erodes past skills, shortcut-induced rigidity throttles the acquisition of new ones. We introduce the Einstellung Rigidity Index (ERI), a compact diagnostic that disentangles genuine transfer from cue-inflated performance using three interpretable facets: (i) Adaptation Delay (AD), (ii) Performance Deficit (PD), and (iii) Relative Suboptimal Feature Reliance (SFR_rel). On a two-phase CIFAR-100 CL benchmark with a deliberately spurious magenta patch in Phase 2, we evaluate Naive fine-tuning (SGD), online Elastic Weight Consolidation (EWC_on), Dark Experience Replay (DER++), Gradient Projection Memory (GPM), and Deep Generative Replay (DGR). Across these continual learning methods, we observe that CL methods reach accuracy thresholds earlier than a Scratch-T2 baseline (negative AD) but achieve slightly lower final accuracy on patched shortcut classes (positive PD). Masking the patch improves accuracy for CL methods while slightly reducing Scratch-T2, yielding negative SFR_rel. This pattern indicates the patch acted as a distractor for CL models in this setting rather than a helpful shortcut.", "date": "2025-10-01", "categories": []}
{"id": "2505.13997", "title": "StPR: Spatiotemporal Preservation and Routing for Exemplar-Free Video Class-Incremental Learning", "abstract": "Video Class-Incremental Learning (VCIL) seeks to develop models that continuously learn new action categories over time without forgetting previously acquired knowledge. Unlike traditional Class-Incremental Learning (CIL), VCIL introduces the added complexity of spatiotemporal structures, making it particularly challenging to mitigate catastrophic forgetting while effectively capturing both frame-shared semantics and temporal dynamics. Existing approaches either rely on exemplar rehearsal, raising concerns over memory and privacy, or adapt static image-based methods that neglect temporal modeling. To address these limitations, we propose Spatiotemporal Preservation and Routing (StPR), a unified and exemplar-free VCIL framework that explicitly disentangles and preserves spatiotemporal information. First, we introduce Frame-Shared Semantics Distillation (FSSD), which identifies semantically stable and meaningful channels by jointly considering semantic sensitivity and classification contribution. These important semantic channels are selectively regularized to maintain prior knowledge while allowing for adaptation. Second, we design a Temporal Decomposition-based Mixture-of-Experts (TD-MoE), which dynamically routes task-specific experts based on their temporal dynamics, enabling inference without task ID or stored exemplars. Together, StPR effectively leverages spatial semantics and temporal dynamics, achieving a unified, exemplar-free VCIL framework. Extensive experiments on UCF101, HMDB51, and Kinetics400 show that our method outperforms existing baselines while offering improved interpretability and efficiency in VCIL. Code is available in the supplementary materials.", "date": "2025-05-20", "categories": []}
{"id": "2410.20098", "title": "Self-Normalized Resets for Plasticity in Continual Learning", "abstract": "Plasticity Loss is an increasingly important phenomenon that refers to the empirical observation that as a neural network is continually trained on a sequence of changing tasks, its ability to adapt to a new task diminishes over time. We introduce Self-Normalized Resets (SNR), a simple adaptive algorithm that mitigates plasticity loss by resetting a neuron's weights when evidence suggests its firing rate has effectively dropped to zero. Across a battery of continual learning problems and network architectures, we demonstrate that SNR consistently attains superior performance compared to its competitor algorithms. We also demonstrate that SNR is robust to its sole hyperparameter, its rejection percentile threshold, while competitor algorithms show significant sensitivity. SNR's threshold-based reset mechanism is motivated by a simple hypothesis test that we derive. Seen through the lens of this hypothesis test, competing reset proposals yield suboptimal error rates in correctly detecting inactive neurons, potentially explaining our experimental observations. We also conduct a theoretical investigation of the optimization landscape for the problem of learning a single ReLU. We show that even when initialized adversarially, an idealized version of SNR learns the target ReLU, while regularization based approaches can fail to learn.", "date": "2024-10-26", "categories": []}
{"id": "2503.23402", "title": "Beyond Synthetic Replays: Turning Diffusion Features into Few-Shot Class-Incremental Learning Knowledge", "abstract": "Few-shot class-incremental learning (FSCIL) is challenging due to extremely limited training data while requiring models to acquire new knowledge without catastrophic forgetting. Recent works have explored generative models, particularly Stable Diffusion (SD), to address these challenges. However, existing approaches use SD mainly as a replay generator, whereas we demonstrate that SD's rich multi-scale representations can serve as a unified backbone. Motivated by this observation, we introduce Diffusion-FSCIL, which extracts four synergistic feature types from SD by capturing real image characteristics through inversion, providing semantic diversity via class-conditioned synthesis, enhancing generalization through controlled noise injection, and enabling replay without image storage through generative features. Unlike conventional approaches requiring synthetic buffers and separate classification backbones, our unified framework operates entirely in the latent space with only lightweight networks ($\\approx$6M parameters). Extensive experiments on CUB-200, miniImageNet, and CIFAR-100 demonstrate state-of-the-art performance, with comprehensive ablations confirming the necessity of each feature type. Furthermore, we confirm that our streamlined variant maintains competitive accuracy while substantially improving efficiency, establishing the viability of generative models as practical and effective backbones for FSCIL.", "date": "2025-03-30", "categories": []}
{"id": "2509.22764", "title": "In-Context Learning can Perform Continual Learning Like Humans", "abstract": "Large language models (LLMs) can adapt to new tasks via in-context learning (ICL) without parameter updates, making them powerful learning engines for fast adaptation. While extensive research has examined ICL as a few-shot learner, whether it can achieve long-term retention and cross-task knowledge accumulation when multitasks arrive sequentially remains underexplored. Motivated by human memory studies, we investigate the retention characteristics of ICL in multitask settings and extend it to in-context continual learning (ICCL), where continual learning ability emerges through task scheduling and prompt rearrangement. Experiments on Markov-Chain benchmarks demonstrate that, for specific large-language models, ICCL benefits from distributed practice (DP) in a manner analogous to humans, consistently revealing a spacing \"sweet spot\" for retention. Beyond retention performance, we propose a human-retention similarity metric to quantify how closely a continual-learning (CL) method aligns with human retention dynamics. Using this metric, we show that linear-attention models such as MAMBA and RWKV exhibit particularly human-like retention patterns, despite their retention performance lagging behind that of Transformer-based LLMs. Overall, our results establish ICCL as both cognitively plausible and practically effective, providing an inference-only CL paradigm that mitigates catastrophic forgetting and addresses the stability-plasticity dilemma in conventional CL methods.", "date": "2025-09-26", "categories": []}
{"id": "2509.23683", "title": "Decentralized Dynamic Cooperation of Personalized Models for Federated Continual Learning", "abstract": "Federated continual learning (FCL) has garnered increasing attention for its ability to support distributed computation in environments with evolving data distributions. However, the emergence of new tasks introduces both temporal and cross-client shifts, making catastrophic forgetting a critical challenge. Most existing works aggregate knowledge from clients into a global model, which may not enhance client performance since irrelevant knowledge could introduce interference, especially in heterogeneous scenarios. Additionally, directly applying decentralized approaches to FCL suffers from ineffective group formation caused by task changes. To address these challenges, we propose a decentralized dynamic cooperation framework for FCL, where clients establish dynamic cooperative learning coalitions to balance the acquisition of new knowledge and the retention of prior learning, thereby obtaining personalized models. To maximize model performance, each client engages in selective cooperation, dynamically allying with others who offer meaningful performance gains. This results in non-overlapping, variable coalitions at each stage of the task. Moreover, we use coalitional affinity game to simulate coalition relationships between clients. By assessing both client gradient coherence and model similarity, we quantify the client benefits derived from cooperation. We also propose a merge-blocking algorithm and a dynamic cooperative evolution algorithm to achieve cooperative and dynamic equilibrium. Comprehensive experiments demonstrate the superiority of our method compared to various baselines. Code is available at: https://github.com/ydn3229/DCFCL.", "date": "2025-09-28", "categories": []}
{"id": "2509.23893", "title": "Dynamic Orthogonal Continual Fine-tuning for Mitigating Catastrophic Forgettings", "abstract": "Catastrophic forgetting remains a critical challenge in continual learning for large language models (LLMs), where models struggle to retain performance on historical tasks when fine-tuning on new sequential data without access to past datasets. In this paper, we first reveal that the drift of functional directions during the fine-tuning process is a key reason why existing regularization-based methods fail in long-term LLM continual learning. To address this, we propose Dynamic Orthogonal Continual (DOC) fine-tuning, a novel approach that tracks the drift of these functional directions and dynamically updates them during the fine-tuning process. Furthermore, by adjusting the gradients of new task parameters to be orthogonal to the tracked historical function directions, our method mitigates interference between new and old tasks. Extensive experiments on various LLM continual learning benchmarks demonstrate that this approach outperforms prior methods, effectively reducing catastrophic forgetting and providing a robust tool for continuous LLM fine-tuning. Our code is available at https://github.com/meloxxxxxx/DOC.", "date": "2025-09-28", "categories": []}
{"id": "2509.23906", "title": "EWC-Guided Diffusion Replay for Exemplar-Free Continual Learning in Medical Imaging", "abstract": "Medical imaging foundation models must adapt over time, yet full retraining is often blocked by privacy constraints and cost. We present a continual learning framework that avoids storing patient exemplars by pairing class conditional diffusion replay with Elastic Weight Consolidation. Using a compact Vision Transformer backbone, we evaluate across eight MedMNIST v2 tasks and CheXpert. On CheXpert our approach attains 0.851 AUROC, reduces forgetting by more than 30\\% relative to DER\\texttt{++}, and approaches joint training at 0.869 AUROC, while remaining efficient and privacy preserving. Analyses connect forgetting to two measurable factors: fidelity of replay and Fisher weighted parameter drift, highlighting the complementary roles of replay diffusion and synaptic stability. The results indicate a practical route for scalable, privacy aware continual adaptation of clinical imaging models.", "date": "2025-09-28", "categories": []}
{"id": "2509.24547", "title": "LEAF: A Robust Expert-Based Framework for Few-Shot Continual Event Detection", "abstract": "Few-shot Continual Event Detection (FCED) poses the dual challenges of learning from limited data and mitigating catastrophic forgetting across sequential tasks. Existing approaches often suffer from severe forgetting due to the full fine-tuning of a shared base model, which leads to knowledge interference between tasks. Moreover, they frequently rely on data augmentation strategies that can introduce unnatural or semantically distorted inputs. To address these limitations, we propose LEAF, a novel and robust expert-based framework for FCED. LEAF integrates a specialized mixture of experts architecture into the base model, where each expert is parameterized with low-rank adaptation (LoRA) matrices. A semantic-aware expert selection mechanism dynamically routes instances to the most relevant experts, enabling expert specialization and reducing knowledge interference. To improve generalization in limited-data settings, LEAF incorporates a contrastive learning objective guided by label descriptions, which capture high-level semantic information about event types. Furthermore, to prevent overfitting on the memory buffer, our framework employs a knowledge distillation strategy that transfers knowledge from previous models to the current one. Extensive experiments on multiple FCED benchmarks demonstrate that LEAF consistently achieves state-of-the-art performance.", "date": "2025-09-29", "categories": []}
{"id": "2505.18697", "title": "Can LLMs Alleviate Catastrophic Forgetting in Graph Continual Learning? A Systematic Study", "abstract": "Nowadays, real-world data, including graph-structure data, often arrives in a streaming manner, which means that learning systems need to continuously acquire new knowledge without forgetting previously learned information. Although substantial existing works attempt to address catastrophic forgetting in graph machine learning, they are all based on training from scratch with streaming data. With the rise of pretrained models, an increasing number of studies have leveraged their strong generalization ability for continual learning. Therefore, in this work, we attempt to answer whether large language models (LLMs) can mitigate catastrophic forgetting in Graph Continual Learning (GCL). We first point out that current experimental setups for GCL have significant flaws, as the evaluation stage may lead to task ID leakage. Then, we evaluate the performance of LLMs in more realistic scenarios and find that even minor modifications can lead to outstanding results. Finally, based on extensive experiments, we propose a simple-yet-effective method, Simple Graph Continual Learning (SimGCL), that surpasses the previous state-of-the-art GNN-based baseline by around 20% under the rehearsal-free constraint. To facilitate reproducibility, we have developed an easy-to-use benchmark LLM4GCL for training and evaluating existing GCL methods. The code is available at: https://github.com/ZhixunLEE/LLM4GCL.", "date": "2025-05-24", "categories": []}
{"id": "2509.21898", "title": "Closing the Oracle Gap: Increment Vector Transformation for Class Incremental Learning", "abstract": "Class Incremental Learning (CIL) aims to sequentially acquire knowledge of new classes without forgetting previously learned ones. Despite recent progress, current CIL methods still exhibit significant performance gaps compared to their oracle counterparts-models trained with full access to historical data. Inspired by recent insights on Linear Mode Connectivity (LMC), we revisit the geometric properties of oracle solutions in CIL and uncover a fundamental observation: these oracle solutions typically maintain low-loss linear connections to the optimum of previous tasks. Motivated by this finding, we propose Increment Vector Transformation (IVT), a novel plug-and-play framework designed to mitigate catastrophic forgetting during training. Rather than directly following CIL updates, IVT periodically teleports the model parameters to transformed solutions that preserve linear connectivity to previous task optimum. By maintaining low-loss along these connecting paths, IVT effectively ensures stable performance on previously learned tasks. The transformation is efficiently approximated using diagonal Fisher Information Matrices, making IVT suitable for both exemplar-free and exemplar-based scenarios, and compatible with various initialization strategies. Extensive experiments on CIFAR-100, FGVCAircraft, ImageNet-Subset, and ImageNet-Full demonstrate that IVT consistently enhances the performance of strong CIL baselines. Specifically, on CIFAR-100, IVT improves the last accuracy of the PASS baseline by +5.12% and reduces forgetting by 2.54%. For the CLIP-pre-trained SLCA baseline on FGVCAircraft, IVT yields gains of +14.93% in average accuracy and +21.95% in last accuracy. The code will be released.", "date": "2025-09-26", "categories": []}
{"id": "2509.22645", "title": "Hierarchical Representation Matching for CLIP-based Class-Incremental Learning", "abstract": "Class-Incremental Learning (CIL) aims to endow models with the ability to continuously adapt to evolving data streams. Recent advances in pre-trained vision-language models (e.g., CLIP) provide a powerful foundation for this task. However, existing approaches often rely on simplistic templates, such as \"a photo of a [CLASS]\", which overlook the hierarchical nature of visual concepts. For example, recognizing \"cat\" versus \"car\" depends on coarse-grained cues, while distinguishing \"cat\" from \"lion\" requires fine-grained details. Similarly, the current feature mapping in CLIP relies solely on the representation from the last layer, neglecting the hierarchical information contained in earlier layers. In this work, we introduce HiErarchical Representation MAtchiNg (HERMAN) for CLIP-based CIL. Our approach leverages LLMs to recursively generate discriminative textual descriptors, thereby augmenting the semantic space with explicit hierarchical cues. These descriptors are matched to different levels of the semantic hierarchy and adaptively routed based on task-specific requirements, enabling precise discrimination while alleviating catastrophic forgetting in incremental tasks. Extensive experiments on multiple benchmarks demonstrate that our method consistently achieves state-of-the-art performance.", "date": "2025-09-26", "categories": []}
{"id": "2505.23412", "title": "Buffer-free Class-Incremental Learning with Out-of-Distribution Detection", "abstract": "Class-incremental learning (CIL) poses significant challenges in open-world scenarios, where models must not only learn new classes over time without forgetting previous ones but also handle inputs from unknown classes that a closed-set model would misclassify. Recent works address both issues by (i)~training multi-head models using the task-incremental learning framework, and (ii) predicting the task identity employing out-of-distribution (OOD) detectors. While effective, the latter mainly relies on joint training with a memory buffer of past data, raising concerns around privacy, scalability, and increased training time. In this paper, we present an in-depth analysis of post-hoc OOD detection methods and investigate their potential to eliminate the need for a memory buffer. We uncover that these methods, when applied appropriately at inference time, can serve as a strong substitute for buffer-based OOD detection. We show that this buffer-free approach achieves comparable or superior performance to buffer-based methods both in terms of class-incremental learning and the rejection of unknown samples. Experimental results on CIFAR-10, CIFAR-100 and Tiny ImageNet datasets support our findings, offering new insights into the design of efficient and privacy-preserving CIL systems for open-world settings.", "date": "2025-05-29", "categories": []}
{"id": "2506.18789", "title": "Shift Happens: Mixture of Experts based Continual Adaptation in Federated Learning", "abstract": "Federated Learning (FL) enables collaborative model training across decentralized clients without sharing raw data, yet faces significant challenges in real-world settings where client data distributions evolve dynamically over time. This paper tackles the critical problem of covariate and label shifts in streaming FL environments, where non-stationary data distributions degrade model performance and necessitate a middleware layer that adapts FL to distributional shifts. We introduce ShiftEx, a shift-aware mixture of experts framework that dynamically creates and trains specialized global models in response to detected distribution shifts using Maximum Mean Discrepancy for covariate shifts. The framework employs a latent memory mechanism for expert reuse and implements facility location-based optimization to jointly minimize covariate mismatch, expert creation costs, and label imbalance. Through theoretical analysis and comprehensive experiments on benchmark datasets, we demonstrate 5.5-12.9 percentage point accuracy improvements and 22-95 % faster adaptation compared to state-of-the-art FL baselines across diverse shift scenarios. The proposed approach offers a scalable, privacy-preserving middleware solution for FL systems operating in non-stationary, real-world conditions while minimizing communication and computational overhead.", "date": "2025-06-23", "categories": []}
{"id": "2509.20269", "title": "Predictive Coding-based Deep Neural Network Fine-tuning for Computationally Efficient Domain Adaptation", "abstract": "As deep neural networks are increasingly deployed in dynamic, real-world environments, relying on a single static model is often insufficient. Changes in input data distributions caused by sensor drift or lighting variations necessitate continual model adaptation. In this paper, we propose a hybrid training methodology that enables efficient on-device domain adaptation by combining the strengths of Backpropagation and Predictive Coding. The method begins with a deep neural network trained offline using Backpropagation to achieve high initial performance. Subsequently, Predictive Coding is employed for online adaptation, allowing the model to recover accuracy lost due to shifts in the input data distribution. This approach leverages the robustness of Backpropagation for initial representation learning and the computational efficiency of Predictive Coding for continual learning, making it particularly well-suited for resource-constrained edge devices or future neuromorphic accelerators. Experimental results on the MNIST and CIFAR-10 datasets demonstrate that this hybrid strategy enables effective adaptation with a reduced computational overhead, offering a promising solution for maintaining model performance in dynamic environments.", "date": "2025-09-24", "categories": []}
{"id": "2509.21161", "title": "DATS: Distance-Aware Temperature Scaling for Calibrated Class-Incremental Learning", "abstract": "Continual Learning (CL) is recently gaining increasing attention for its ability to enable a single model to learn incrementally from a sequence of new classes. In this scenario, it is important to keep consistent predictive performance across all the classes and prevent the so-called Catastrophic Forgetting (CF). However, in safety-critical applications, predictive performance alone is insufficient. Predictive models should also be able to reliably communicate their uncertainty in a calibrated manner - that is, with confidence scores aligned to the true frequencies of target events. Existing approaches in CL address calibration primarily from a data-centric perspective, relying on a single temperature shared across all tasks. Such solutions overlook task-specific differences, leading to large fluctuations in calibration error across tasks. For this reason, we argue that a more principled approach should adapt the temperature according to the distance to the current task. However, the unavailability of the task information at test time/during deployment poses a major challenge to achieve the intended objective. For this, we propose Distance-Aware Temperature Scaling (DATS), which combines prototype-based distance estimation with distance-aware calibration to infer task proximity and assign adaptive temperatures without prior task information. Through extensive empirical evaluation on both standard benchmarks and real-world, imbalanced datasets taken from the biomedical domain, our approach demonstrates to be stable, reliable and consistent in reducing calibration error across tasks compared to state-of-the-art approaches.", "date": "2025-09-25", "categories": []}
{"id": "2509.16738", "title": "Mixture of Noise for Pre-Trained Model-Based Class-Incremental Learning", "abstract": "Class Incremental Learning (CIL) aims to continuously learn new categories while retaining the knowledge of old ones. Pre-trained models (PTMs) show promising capabilities in CIL. However, existing approaches that apply lightweight fine-tuning to backbones still induce parameter drift, thereby compromising the generalization capability of pre-trained models. Parameter drift can be conceptualized as a form of noise that obscures critical patterns learned for previous tasks. However, recent researches have shown that noise is not always harmful. For example, the large number of visual patterns learned from pre-training can be easily abused by a single task, and introducing appropriate noise can suppress some low-correlation features, thus leaving a margin for future tasks. To this end, we propose learning beneficial noise for CIL guided by information theory and propose Mixture of Noise (Min), aiming to mitigate the degradation of backbone generalization from adapting new tasks. Specifically, task-specific noise is learned from high-dimension features of new tasks. Then, a set of weights is adjusted dynamically for optimal mixture of different task noise. Finally, Min embeds the beneficial noise into the intermediate features to mask the response of inefficient patterns. Extensive experiments on six benchmark datasets demonstrate that Min achieves state-of-the-art performance in most incremental settings, with particularly outstanding results in 50-steps incremental settings. This shows the significant potential for beneficial noise in continual learning. Code is available at https://github.com/ASCIIJK/MiN-NeurIPS2025.", "date": "2025-09-20", "categories": []}
{"id": "2509.19819", "title": "Adaptive Model Ensemble for Continual Learning", "abstract": "Model ensemble is an effective strategy in continual learning, which alleviates catastrophic forgetting by interpolating model parameters, achieving knowledge fusion learned from different tasks. However, existing model ensemble methods usually encounter the knowledge conflict issue at task and layer levels, causing compromised learning performance in both old and new tasks. To solve this issue, we propose meta-weight-ensembler that adaptively fuses knowledge of different tasks for continual learning. Concretely, we employ a mixing coefficient generator trained via meta-learning to generate appropriate mixing coefficients for model ensemble to address the task-level knowledge conflict. The mixing coefficient is individually generated for each layer to address the layer-level knowledge conflict. In this way, we learn the prior knowledge about adaptively accumulating knowledge of different tasks in a fused model, achieving efficient learning in both old and new tasks. Meta-weight-ensembler can be flexibly combined with existing continual learning methods to boost their ability of alleviating catastrophic forgetting. Experiments on multiple continual learning datasets show that meta-weight-ensembler effectively alleviates catastrophic forgetting and achieves state-of-the-art performance.", "date": "2025-09-24", "categories": []}
{"id": "2509.18504", "title": "Hyperbolic Coarse-to-Fine Few-Shot Class-Incremental Learning", "abstract": "In the field of machine learning, hyperbolic space demonstrates superior representation capabilities for hierarchical data compared to conventional Euclidean space. This work focuses on the Coarse-To-Fine Few-Shot Class-Incremental Learning (C2FSCIL) task. Our study follows the Knowe approach, which contrastively learns coarse class labels and subsequently normalizes and freezes the classifier weights of learned fine classes in the embedding space. To better interpret the \"coarse-to-fine\" paradigm, we propose embedding the feature extractor into hyperbolic space. Specifically, we employ the Poincar\\'e ball model of hyperbolic space, enabling the feature extractor to transform input images into feature vectors within the Poincar\\'e ball instead of Euclidean space. We further introduce hyperbolic contrastive loss and hyperbolic fully-connected layers to facilitate model optimization and classification in hyperbolic space. Additionally, to enhance performance under few-shot conditions, we implement maximum entropy distribution in hyperbolic space to estimate the probability distribution of fine-class feature vectors. This allows generation of augmented features from the distribution to mitigate overfitting during training with limited samples. Experiments on C2FSCIL benchmarks show that our method effectively improves both coarse and fine class accuracies.", "date": "2025-09-23", "categories": []}
{"id": "2406.09384", "title": "Reflecting on the State of Rehearsal-free Continual Learning with Pretrained Models", "abstract": "With the advent and recent ubiquity of foundation models, continual learning (CL) has recently shifted from continual training from scratch to the continual adaptation of pretrained models, seeing particular success on rehearsal-free CL benchmarks (RFCL). To achieve this, most proposed methods adapt and restructure parameter-efficient finetuning techniques (PEFT) to suit the continual nature of the problem. Based most often on input-conditional query-mechanisms or regularizations on top of prompt- or adapter-based PEFT, these PEFT-style RFCL (P-RFCL) approaches report peak performances; often convincingly outperforming existing CL techniques. However, on the other end, critical studies have recently highlighted competitive results by training on just the first task or via simple non-parametric baselines. Consequently, questions arise about the relationship between methodological choices in P-RFCL and their reported high benchmark scores. In this work, we tackle these questions to better understand the true drivers behind strong P-RFCL performances, their placement w.r.t. recent first-task adaptation studies, and their relation to preceding CL standards such as EWC or SI. In particular, we show: (1) P-RFCL techniques relying on input-conditional query mechanisms work not because, but rather despite them by collapsing towards standard PEFT shortcut solutions. (2) Indeed, we show how most often, P-RFCL techniques can be matched by a simple and lightweight PEFT baseline. (3) Using this baseline, we identify the implicit bound on tunable parameters when deriving RFCL approaches from PEFT methods as a potential denominator behind P-RFCL efficacy. Finally, we (4) better disentangle continual versus first-task adaptation, and (5) motivate standard RFCL techniques s.a. EWC or SI in light of recent P-RFCL methods.", "date": "2024-06-13", "categories": []}
{"id": "2503.14963", "title": "Continual Multimodal Contrastive Learning", "abstract": "Multimodal Contrastive Learning (MCL) advances in aligning different modalities and generating multimodal representations in a joint space. By leveraging contrastive learning across diverse modalities, large-scale multimodal data enhances representational quality. However, a critical yet often overlooked challenge remains: multimodal data is rarely collected in a single process, and training from scratch is computationally expensive. Instead, emergent multimodal data can be used to optimize existing models gradually, i.e., models are trained on a sequence of modality pair data. We define this problem as Continual Multimodal Contrastive Learning (CMCL), an underexplored yet crucial research direction at the intersection of multimodal and continual learning. In this paper, we formulate CMCL through two specialized principles of stability and plasticity. We theoretically derive a novel optimization-based method, which projects updated gradients from dual sides onto subspaces where any gradient is prevented from interfering with the previously learned knowledge. Two upper bounds provide theoretical insights on both stability and plasticity in our solution. Beyond our theoretical contributions, we conduct experiments on multiple datasets by comparing our method against advanced continual learning baselines. The empirical results further support our claims and demonstrate the efficacy of our method. Our codes are available at https://github.com/Xiaohao-Liu/CMCL.", "date": "2025-03-19", "categories": []}
{"id": "2509.14958", "title": "Seeing 3D Through 2D Lenses: 3D Few-Shot Class-Incremental Learning via Cross-Modal Geometric Rectification", "abstract": "The rapid growth of 3D digital content necessitates expandable recognition systems for open-world scenarios. However, existing 3D class-incremental learning methods struggle under extreme data scarcity due to geometric misalignment and texture bias. While recent approaches integrate 3D data with 2D foundation models (e.g., CLIP), they suffer from semantic blurring caused by texture-biased projections and indiscriminate fusion of geometric-textural cues, leading to unstable decision prototypes and catastrophic forgetting. To address these issues, we propose Cross-Modal Geometric Rectification (CMGR), a framework that enhances 3D geometric fidelity by leveraging CLIP's hierarchical spatial semantics. Specifically, we introduce a Structure-Aware Geometric Rectification module that hierarchically aligns 3D part structures with CLIP's intermediate spatial priors through attention-driven geometric fusion. Additionally, a Texture Amplification Module synthesizes minimal yet discriminative textures to suppress noise and reinforce cross-modal consistency. To further stabilize incremental prototypes, we employ a Base-Novel Discriminator that isolates geometric variations. Extensive experiments demonstrate that our method significantly improves 3D few-shot class-incremental learning, achieving superior geometric coherence and robustness to texture bias across cross-domain and within-domain settings.", "date": "2025-09-18", "categories": []}
{"id": "2509.16882", "title": "Dynamic Expert Specialization: Towards Catastrophic Forgetting-Free Multi-Domain MoE Adaptation", "abstract": "Mixture-of-Experts (MoE) models offer immense capacity via sparsely gated expert subnetworks, yet adapting them to multiple domains without catastrophic forgetting remains an open challenge. Existing approaches either incur prohibitive computation, suffer cross-domain interference, or require separate runs per domain. We propose DES-MoE, a dynamic expert specialization framework for multi-domain adaptation of Mixture-of-Experts models. DES-MoE addresses catastrophic forgetting through three innovations: (1) an adaptive router balancing pre-trained knowledge retention and task-specific updates via distillation, (2) real-time expert-domain correlation mapping to isolate domain-specific gradients, and (3) a three-phase adaptive fine-tuning schedule that progressively freezes non-specialized parameters. Evaluated on six domains (math, code, law, etc.), DES-MoE matches single-domain ESFT performance while training one unified model, reduces forgetting by 89% compared to full fine-tuning as domains scale from 2 to 6, and achieves 68% faster convergence than conventional methods. Our work establishes dynamic expert isolation as a scalable paradigm for multi-task MoE adaptation.", "date": "2025-09-21", "categories": []}
{"id": "2509.17348", "title": "AIMMerging: Adaptive Iterative Model Merging Using Training Trajectories for Language Model Continual Learning", "abstract": "Continual learning (CL) is essential for deploying large language models (LLMs) in dynamic real-world environments without the need for costly retraining. Recent model merging-based methods have attracted significant attention, but they still struggle to effectively manage the trade-off between learning new knowledge and preventing forgetting, a challenge largely stemming from suboptimal number of merges and merging frequency. In this paper, we introduce Adaptive Iterative Model Merging (AimMerging), a novel CL framework that utilizes learning and forgetting signals from the training trajectory to dynamically monitor the model's training status. Guided by dynamic monitoring, the training trajectory-guided merge controller adaptively determines the timing and frequency of iterative fusion, while the rehearsal-based knowledge fusion module computes the merging weights and executes the fusion. Comprehensive experiments on three CL benchmarks with various model sizes (from 770M to 13B) demonstrate that AimMerging achieves significant performance improvements over existing state-of-the-art methods, with an average relative improvement of 80% and 59% on FWT and BWT, respectively. The source code is provided for reproducibility.", "date": "2025-09-22", "categories": []}
{"id": "2509.17439", "title": "SPICED: A Synaptic Homeostasis-Inspired Framework for Unsupervised Continual EEG Decoding", "abstract": "Human brain achieves dynamic stability-plasticity balance through synaptic homeostasis. Inspired by this biological principle, we propose SPICED: a neuromorphic framework that integrates the synaptic homeostasis mechanism for unsupervised continual EEG decoding, particularly addressing practical scenarios where new individuals with inter-individual variability emerge continually. SPICED comprises a novel synaptic network that enables dynamic expansion during continual adaptation through three bio-inspired neural mechanisms: (1) critical memory reactivation; (2) synaptic consolidation and (3) synaptic renormalization. The interplay within synaptic homeostasis dynamically strengthens task-discriminative memory traces and weakens detrimental memories. By integrating these mechanisms with continual learning system, SPICED preferentially replays task-discriminative memory traces that exhibit strong associations with newly emerging individuals, thereby achieving robust adaptations. Meanwhile, SPICED effectively mitigates catastrophic forgetting by suppressing the replay prioritization of detrimental memories during long-term continual learning. Validated on three EEG datasets, SPICED show its effectiveness.", "date": "2025-09-22", "categories": []}
{"id": "2509.17530", "title": "An Unlearning Framework for Continual Learning", "abstract": "Growing concerns surrounding AI safety and data privacy have driven the development of Machine Unlearning as a potential solution. However, current machine unlearning algorithms are designed to complement the offline training paradigm. The emergence of the Continual Learning (CL) paradigm promises incremental model updates, enabling models to learn new tasks sequentially. Naturally, some of those tasks may need to be unlearned to address safety or privacy concerns that might arise. We find that applying conventional unlearning algorithms in continual learning environments creates two critical problems: performance degradation on retained tasks and task relapse, where previously unlearned tasks resurface during subsequent learning. Furthermore, most unlearning algorithms require data to operate, which conflicts with CL's philosophy of discarding past data. A clear need arises for unlearning algorithms that are data-free and mindful of future learning. To that end, we propose UnCLe, an Unlearning framework for Continual Learning. UnCLe employs a hypernetwork that learns to generate task-specific network parameters, using task embeddings. Tasks are unlearned by aligning the corresponding generated network parameters with noise, without requiring any data. Empirical evaluations on several vision data sets demonstrate UnCLe's ability to sequentially perform multiple learning and unlearning operations with minimal disruption to previously acquired knowledge.", "date": "2025-09-22", "categories": []}
{"id": "2509.15347", "title": "Global Pre-fixing, Local Adjusting: A Simple yet Effective Contrastive Strategy for Continual Learning", "abstract": "Continual learning (CL) involves acquiring and accumulating knowledge from evolving tasks while alleviating catastrophic forgetting. Recently, leveraging contrastive loss to construct more transferable and less forgetful representations has been a promising direction in CL. Despite advancements, their performance is still limited due to confusion arising from both inter-task and intra-task features. To address the problem, we propose a simple yet effective contrastive strategy named \\textbf{G}lobal \\textbf{P}re-fixing, \\textbf{L}ocal \\textbf{A}djusting for \\textbf{S}upervised \\textbf{C}ontrastive learning (GPLASC). Specifically, to avoid task-level confusion, we divide the entire unit hypersphere of representations into non-overlapping regions, with the centers of the regions forming an inter-task pre-fixed \\textbf{E}quiangular \\textbf{T}ight \\textbf{F}rame (ETF). Meanwhile, for individual tasks, our method helps regulate the feature structure and form intra-task adjustable ETFs within their respective allocated regions. As a result, our method \\textit{simultaneously} ensures discriminative feature structures both between tasks and within tasks and can be seamlessly integrated into any existing contrastive continual learning framework. Extensive experiments validate its effectiveness.", "date": "2025-09-18", "categories": []}
{"id": "2509.15785", "title": "CBPNet: A Continual Backpropagation Prompt Network for Alleviating Plasticity Loss on Edge Devices", "abstract": "To meet the demands of applications like robotics and autonomous driving that require real-time responses to dynamic environments, efficient continual learning methods suitable for edge devices have attracted increasing attention. In this transition, using frozen pretrained models with prompts has become a mainstream strategy to combat catastrophic forgetting. However, this approach introduces a new critical bottleneck: plasticity loss, where the model's ability to learn new knowledge diminishes due to the frozen backbone and the limited capacity of prompt parameters. We argue that the reduction in plasticity stems from a lack of update vitality in underutilized parameters during the training process. To this end, we propose the Continual Backpropagation Prompt Network (CBPNet), an effective and parameter efficient framework designed to restore the model's learning vitality. We innovatively integrate an Efficient CBP Block that counteracts plasticity decay by adaptively reinitializing these underutilized parameters. Experimental results on edge devices demonstrate CBPNet's effectiveness across multiple benchmarks. On Split CIFAR-100, it improves average accuracy by over 1% against a strong baseline, and on the more challenging Split ImageNet-R, it achieves a state of the art accuracy of 69.41%. This is accomplished by training additional parameters that constitute less than 0.2% of the backbone's size, validating our approach.", "date": "2025-09-19", "categories": []}
{"id": "2509.16011", "title": "Towards Robust Visual Continual Learning with Multi-Prototype Supervision", "abstract": "Language-guided supervision, which utilizes a frozen semantic target from a Pretrained Language Model (PLM), has emerged as a promising paradigm for visual Continual Learning (CL). However, relying on a single target introduces two critical limitations: 1) semantic ambiguity, where a polysemous category name results in conflicting visual representations, and 2) intra-class visual diversity, where a single prototype fails to capture the rich variety of visual appearances within a class. To this end, we propose MuproCL, a novel framework that replaces the single target with multiple, context-aware prototypes. Specifically, we employ a lightweight LLM agent to perform category disambiguation and visual-modal expansion to generate a robust set of semantic prototypes. A LogSumExp aggregation mechanism allows the vision model to adaptively align with the most relevant prototype for a given image. Extensive experiments across various CL baselines demonstrate that MuproCL consistently enhances performance and robustness, establishing a more effective path for language-guided continual learning.", "date": "2025-09-19", "categories": []}
{"id": "2509.13211", "title": "HAM: Hierarchical Adapter Merging for Scalable Continual Learning", "abstract": "Continual learning is an essential capability of human cognition, yet it poses significant challenges for current deep learning models. The primary issue is that new knowledge can interfere with previously learned information, causing the model to forget earlier knowledge in favor of the new, a phenomenon known as catastrophic forgetting. Although large pre-trained models can partially mitigate forgetting by leveraging their existing knowledge and over-parameterization, they often struggle when confronted with novel data distributions. Parameter-Efficient Fine-Tuning (PEFT) methods, such as LoRA, enable efficient adaptation to new knowledge. However, they still face challenges in scaling to dynamic learning scenarios and long sequences of tasks, as maintaining one adapter per task introduces complexity and increases the potential for interference. In this paper, we introduce Hierarchical Adapters Merging (HAM), a novel framework that dynamically combines adapters from different tasks during training. This approach enables HAM to scale effectively, allowing it to manage more tasks than competing baselines with improved efficiency. To achieve this, HAM maintains a fixed set of groups that hierarchically consolidate new adapters. For each task, HAM trains a low-rank adapter along with an importance scalar, then dynamically groups tasks based on adapter similarity. Within each group, adapters are pruned, scaled and merge, facilitating transfer learning between related tasks. Extensive experiments on three vision benchmarks show that HAM significantly outperforms state-of-the-art methods, particularly as the number of tasks increases.", "date": "2025-09-16", "categories": []}
{"id": "2509.13672", "title": "CL$^2$GEC: A Multi-Discipline Benchmark for Continual Learning in Chinese Literature Grammatical Error Correction", "abstract": "The growing demand for automated writing assistance in diverse academic domains highlights the need for robust Chinese Grammatical Error Correction (CGEC) systems that can adapt across disciplines. However, existing CGEC research largely lacks dedicated benchmarks for multi-disciplinary academic writing, overlooking continual learning (CL) as a promising solution to handle domain-specific linguistic variation and prevent catastrophic forgetting. To fill this crucial gap, we introduce CL$^2$GEC, the first Continual Learning benchmark for Chinese Literature Grammatical Error Correction, designed to evaluate adaptive CGEC across multiple academic fields. Our benchmark includes 10,000 human-annotated sentences spanning 10 disciplines, each exhibiting distinct linguistic styles and error patterns. CL$^2$GEC focuses on evaluating grammatical error correction in a continual learning setting, simulating sequential exposure to diverse academic disciplines to reflect real-world editorial dynamics. We evaluate large language models under sequential tuning, parameter-efficient adaptation, and four representative CL algorithms, using both standard GEC metrics and continual learning metrics adapted to task-level variation. Experimental results reveal that regularization-based methods mitigate forgetting more effectively than replay-based or naive sequential approaches. Our benchmark provides a rigorous foundation for future research in adaptive grammatical error correction across diverse academic domains.", "date": "2025-09-17", "categories": []}
{"id": "2509.13974", "title": "Personalization on a Budget: Minimally-Labeled Continual Learning for Resource-Efficient Seizure Detection", "abstract": "Objective: Epilepsy, a prevalent neurological disease, demands careful diagnosis and continuous care. Seizure detection remains challenging, as current clinical practice relies on expert analysis of electroencephalography, which is a time-consuming process and requires specialized knowledge. Addressing this challenge, this paper explores automated epileptic seizure detection using deep learning, focusing on personalized continual learning models that adapt to each patient's unique electroencephalography signal features, which evolve over time. Methods: In this context, our approach addresses the challenge of integrating new data into existing models without catastrophic forgetting, a common issue in static deep learning models. We propose EpiSMART, a continual learning framework for seizure detection that uses a size-constrained replay buffer and an informed sample selection strategy to incrementally adapt to patient-specific electroencephalography signals. By selectively retaining high-entropy and seizure-predicted samples, our method preserves critical past information while maintaining high performance with minimal memory and computational requirements. Results: Validation on the CHB-MIT dataset, shows that EpiSMART achieves a 21% improvement in the F1 score over a trained baseline without updates in all other patients. On average, EpiSMART requires only 6.46 minutes of labeled data and 6.28 updates per day, making it suitable for real-time deployment in wearable systems. Conclusion:EpiSMART enables robust and personalized seizure detection under realistic and resource-constrained conditions by effectively integrating new data into existing models without degrading past knowledge. Significance: This framework advances automated seizure detection by providing a continual learning approach that supports patient-specific adaptation and practical deployment in wearable healthcare systems.", "date": "2025-09-17", "categories": []}
{"id": "2509.12727", "title": "Unbiased Online Curvature Approximation for Regularized Graph Continual Learning", "abstract": "Graph continual learning (GCL) aims to learn from a continuous sequence of graph-based tasks. Regularization methods are vital for preventing catastrophic forgetting in GCL, particularly in the challenging replay-free, class-incremental setting, where each task consists of a set of unique classes. In this work, we first establish a general regularization framework for GCL based on the curved parameter space induced by the Fisher information matrix (FIM). We show that the dominant Elastic Weight Consolidation (EWC) and its variants are a special case within this framework, using a diagonal approximation of the empirical FIM based on parameters from previous tasks. To overcome their limitations, we propose a new unbiased online curvature approximation of the full FIM based on the model's current learning state. Our method directly estimates the regularization term in an online manner without explicitly evaluating and storing the FIM itself. This enables the model to better capture the loss landscape during learning new tasks while retaining the knowledge learned from previous tasks. Extensive experiments on three graph datasets demonstrate that our method significantly outperforms existing regularization-based methods, achieving a superior trade-off between stability (retaining old knowledge) and plasticity (acquiring new knowledge).", "date": "2025-09-16", "categories": []}
{"id": "2410.02995", "title": "Task-agnostic Lifelong Robot Learning with Retrieval-based Weighted Local Adaptation", "abstract": "A fundamental objective in intelligent robotics is to move towards lifelong learning robot that can learn and adapt to unseen scenarios over time. However, continually learning new tasks would introduce catastrophic forgetting problems due to data distribution shifts. To mitigate this, we store a subset of data from previous tasks and utilize it in two manners: leveraging experience replay to retain learned skills and applying a novel Retrieval-based Local Adaptation technique to restore relevant knowledge. Since a lifelong learning robot must operate in task-free scenarios, where task IDs and even boundaries are not available, our method performs effectively without relying on such information. We also incorporate a selective weighting mechanism to focus on the most \"forgotten\" skill segment, ensuring effective knowledge restoration. Experimental results across diverse manipulation tasks demonstrate that our framework provides a scalable paradigm for lifelong learning, enhancing robot performance in open-ended, task-free scenarios.", "date": "2024-10-03", "categories": []}
{"id": "2412.10869", "title": "TinySubNets: An efficient and low capacity continual learning strategy", "abstract": "Continual Learning (CL) is a highly relevant setting gaining traction in recent machine learning research. Among CL works, architectural and hybrid strategies are particularly effective due to their potential to adapt the model architecture as new tasks are presented. However, many existing solutions do not efficiently exploit model sparsity, and are prone to capacity saturation due to their inefficient use of available weights, which limits the number of learnable tasks. In this paper, we propose TinySubNets (TSN), a novel architectural CL strategy that addresses the issues through the unique combination of pruning with different sparsity levels, adaptive quantization, and weight sharing. Pruning identifies a subset of weights that preserve model performance, making less relevant weights available for future tasks. Adaptive quantization allows a single weight to be separated into multiple parts which can be assigned to different tasks. Weight sharing between tasks boosts the exploitation of capacity and task similarity, allowing for the identification of a better trade-off between model accuracy and capacity. These features allow TSN to efficiently leverage the available capacity, enhance knowledge transfer, and reduce computational resource consumption. Experimental results involving common benchmark CL datasets and scenarios show that our proposed strategy achieves better results in terms of accuracy than existing state-of-the-art CL strategies. Moreover, our strategy is shown to provide a significantly improved model capacity exploitation. Code released at: https://github.com/lifelonglab/tinysubnets.", "date": "2024-12-14", "categories": []}
{"id": "2507.21109", "title": "Task-Focused Consolidation with Spaced Recall: Making Neural Networks Learn like College Students", "abstract": "Deep neural networks often suffer from a critical limitation known as catastrophic forgetting, where performance on past tasks degrades after learning new ones. This paper introduces a novel continual learning approach inspired by human learning strategies like Active Recall, Deliberate Practice, and Spaced Repetition, named Task-Focused Consolidation with Spaced Recall (TFC-SR). TFC-SR enhances the standard experience replay framework with a mechanism we term the Active Recall Probe. It is a periodic, task-aware evaluation of the model's memory that stabilizes the representations of past knowledge. We test TFC-SR on the Split MNIST and the Split CIFAR-100 benchmarks against leading regularization-based and replay-based baselines. Our results show that TFC-SR performs significantly better than these methods. For instance, on the Split CIFAR-100, it achieves a final accuracy of 13.17% compared to Standard Experience Replay's 7.40%. We demonstrate that this advantage comes from the stabilizing effect of the probe itself, and not from the difference in replay volume. Additionally, we analyze the trade-off between memory size and performance and show that while TFC-SR performs better in memory-constrained environments, higher replay volume is still more effective when available memory is abundant. We conclude that TFC-SR is a robust and efficient approach, highlighting the importance of integrating active memory retrieval mechanisms into continual learning systems.", "date": "2025-07-10", "categories": []}
{"id": "2509.10529", "title": "Mitigating Catastrophic Forgetting and Mode Collapse in Text-to-Image Diffusion via Latent Replay", "abstract": "Continual learning -- the ability to acquire knowledge incrementally without forgetting previous skills -- is fundamental to natural intelligence. While the human brain excels at this, artificial neural networks struggle with \"catastrophic forgetting,\" where learning new tasks erases previously acquired knowledge. This challenge is particularly severe for text-to-image diffusion models, which generate images from textual prompts. Additionally, these models face \"mode collapse,\" where their outputs become increasingly repetitive over time. To address these challenges, we apply Latent Replay, a neuroscience-inspired approach, to diffusion models. Traditional replay methods mitigate forgetting by storing and revisiting past examples, typically requiring large collections of images. Latent Replay instead retains only compact, high-level feature representations extracted from the model's internal architecture. This mirrors the hippocampal process of storing neural activity patterns rather than raw sensory inputs, reducing memory usage while preserving critical information. Through experiments with five sequentially learned visual concepts, we demonstrate that Latent Replay significantly outperforms existing methods in maintaining model versatility. After learning all concepts, our approach retained 77.59% Image Alignment (IA) on the earliest concept, 14% higher than baseline methods, while maintaining diverse outputs. Surprisingly, random selection of stored latent examples outperforms similarity-based strategies. Our findings suggest that Latent Replay enables efficient continual learning for generative AI models, paving the way for personalized text-to-image models that evolve with user needs without excessive computational costs.", "date": "2025-09-04", "categories": []}
{"id": "2509.11285", "title": "Efficient Single-Step Framework for Incremental Class Learning in Neural Networks", "abstract": "Incremental learning remains a critical challenge in machine learning, as models often struggle with catastrophic forgetting -the tendency to lose previously acquired knowledge when learning new information. These challenges are even more pronounced in resource-limited settings. Many existing Class Incremental Learning (CIL) methods achieve high accuracy by continually adapting their feature representations; however, they often require substantial computational resources and complex, iterative training procedures. This work introduces CIFNet (Class Incremental and Frugal Network), a novel CIL approach that addresses these limitations by offering a highly efficient and sustainable solution. CIFNet's key innovation lies in its novel integration of several existing, yet separately explored, components: a pre-trained and frozen feature extractor, a compressed data buffer, and an efficient non-iterative one-layer neural network for classification. A pre-trained and frozen feature extractor eliminates computationally expensive fine-tuning of the backbone. This, combined with a compressed buffer for efficient memory use, enables CIFNet to perform efficient class-incremental learning through a single-step optimization process on fixed features, minimizing computational overhead and training time without requiring multiple weight updates. Experiments on benchmark datasets confirm that CIFNet effectively mitigates catastrophic forgetting at the classifier level, achieving high accuracy comparable to that of existing state-of-the-art methods, while substantially improving training efficiency and sustainability. CIFNet represents a significant advancement in making class-incremental learning more accessible and pragmatic in environments with limited resources, especially when strong pre-trained feature extractors are available.", "date": "2025-09-14", "categories": []}
{"id": "2509.11414", "title": "Continually Adding New Languages to Multilingual Language Models", "abstract": "Multilingual language models are trained on a fixed set of languages, and to support new languages, the models need to be retrained from scratch. This is an expensive endeavor and is often infeasible, as model developers tend not to release their pre-training data. Naive approaches, such as continued pretraining, suffer from catastrophic forgetting; however, mitigation strategies like experience replay cannot be applied due to the lack of original pretraining data. In this work, we investigate the problem of continually adding new languages to a multilingual model, assuming access to pretraining data in only the target languages. We explore multiple approaches to address this problem and propose Layer-Selective LoRA (LayRA), which adds Low-Rank Adapters (LoRA) to selected initial and final layers while keeping the rest of the model frozen. LayRA builds on two insights: (1) LoRA reduces forgetting, and (2) multilingual models encode inputs in the source language in the initial layers, reason in English in intermediate layers, and translate back to the source language in final layers. We experiment with adding multiple combinations of Galician, Swahili, and Urdu to pretrained language models and evaluate each method on diverse multilingual tasks. We find that LayRA provides the overall best tradeoff between preserving models' capabilities in previously supported languages, while being competitive with existing approaches such as LoRA in learning new languages. We also demonstrate that using model arithmetic, the adapted models can be equipped with strong instruction following abilities without access to any instruction tuning data in the target languages.", "date": "2025-09-14", "categories": []}
{"id": "2505.19455", "title": "MM-Prompt: Cross-Modal Prompt Tuning for Continual Visual Question Answering", "abstract": "Continual Visual Question Answering (CVQA) based on pre-trained models(PTMs) has achieved promising progress by leveraging prompt tuning to enable continual multi-modal learning. However, most existing methods adopt cross-modal prompt isolation, constructing visual and textual prompts separately, which exacerbates modality imbalance and leads to degraded performance over time. To tackle this issue, we propose MM-Prompt, a novel framework incorporating cross-modal prompt query and cross-modal prompt recovery. The former enables balanced prompt selection by incorporating cross-modal signals during query formation, while the latter promotes joint prompt reconstruction through iterative cross-modal interactions, guided by an alignment loss to prevent representational drift. Extensive experiments show that MM-Prompt surpasses prior approaches in accuracy and knowledge retention, while maintaining balanced modality engagement throughout continual learning.", "date": "2025-05-26", "categories": []}
{"id": "2403.05175", "title": "Continual Learning and Catastrophic Forgetting", "abstract": "This book chapter delves into the dynamics of continual learning, which is the process of incrementally learning from a non-stationary stream of data. Although continual learning is a natural skill for the human brain, it is very challenging for artificial neural networks. An important reason is that, when learning something new, these networks tend to quickly and drastically forget what they had learned before, a phenomenon known as catastrophic forgetting. Especially in the last decade, continual learning has become an extensively studied topic in deep learning. This book chapter reviews the insights that this field has generated.", "date": "2024-03-08", "categories": []}
{"id": "2503.02844", "title": "Beyond Cosine Decay: On the effectiveness of Infinite Learning Rate Schedule for Continual Pre-training", "abstract": "The ever-growing availability of unlabeled data presents both opportunities and challenges for training artificial intelligence systems. While self-supervised learning (SSL) has emerged as a powerful paradigm for extracting meaningful representations from vast amounts of unlabeled data, existing methods still struggle to adapt to the non-stationary, non-IID nature of real-world data streams without forgetting previously learned knowledge. Recent works have adopted a repeated cosine annealing schedule for large-scale continual pre-training; however, these schedules (1) inherently cause forgetting during the re-warming phase and (2) have not been systematically compared to existing continual SSL methods. In this work, we systematically compare the widely used cosine schedule with the recently proposed infinite learning rate schedule and empirically find the latter to be a more effective alternative. Our extensive empirical evaluation across diverse image and language datasets demonstrates that the infinite learning rate schedule consistently enhances continual pre-training performance compared to a repeated cosine decay without being restricted to a fixed iteration budget. For instance, in a small-scale MAE pre-training setup, it outperforms several strong baselines from the literature. We then scale up our experiments to larger MAE pre-training and autoregressive language model pre-training. Our results show that the infinite learning rate schedule remains effective at scale, surpassing repeated cosine decay for both MAE pre-training and zero-shot LM benchmarks.", "date": "2025-03-04", "categories": []}
{"id": "2409.18061", "title": "Optimal Protocols for Continual Learning via Statistical Physics and Control Theory", "abstract": "Artificial neural networks often struggle with catastrophic forgetting when learning multiple tasks sequentially, as training on new tasks degrades the performance on previously learned tasks. Recent theoretical work has addressed this issue by analysing learning curves in synthetic frameworks under predefined training protocols. However, these protocols relied on heuristics and lacked a solid theoretical foundation assessing their optimality. In this paper, we fill this gap by combining exact equations for training dynamics, derived using statistical physics techniques, with optimal control methods. We apply this approach to teacher-student models for continual learning and multi-task problems, obtaining a theory for task-selection protocols maximising performance while minimising forgetting. Our theoretical analysis offers non-trivial yet interpretable strategies for mitigating catastrophic forgetting, shedding light on how optimal learning protocols modulate established effects, such as the influence of task similarity on forgetting. Finally, we validate our theoretical findings with experiments on real-world data.", "date": "2024-09-26", "categories": []}
{"id": "2504.18471", "title": "Action Flow Matching for Continual Robot Learning", "abstract": "Continual learning in robotics seeks systems that can constantly adapt to changing environments and tasks, mirroring human adaptability. A key challenge is refining dynamics models, essential for planning and control, while addressing issues such as safe adaptation, catastrophic forgetting, outlier management, data efficiency, and balancing exploration with exploitation -- all within task and onboard resource constraints. Towards this goal, we introduce a generative framework leveraging flow matching for online robot dynamics model alignment. Rather than executing actions based on a misaligned model, our approach refines planned actions to better match with those the robot would take if its model was well aligned. We find that by transforming the actions themselves rather than exploring with a misaligned model -- as is traditionally done -- the robot collects informative data more efficiently, thereby accelerating learning. Moreover, we validate that the method can handle an evolving and possibly imperfect model while reducing, if desired, the dependency on replay buffers or legacy model snapshots. We validate our approach using two platforms: an unmanned ground vehicle and a quadrotor. The results highlight the method's adaptability and efficiency, with a record 34.2\\% higher task success rate, demonstrating its potential towards enabling continual robot learning. Code: https://github.com/AlejandroMllo/action_flow_matching.", "date": "2025-04-25", "categories": []}
{"id": "2508.19597", "title": "Complementary Learning System Empowers Online Continual Learning of Vehicle Motion Forecasting in Smart Cities", "abstract": "Artificial intelligence underpins most smart city services, yet deep neural network (DNN) that forecasts vehicle motion still struggle with catastrophic forgetting, the loss of earlier knowledge when models are updated. Conventional fixes enlarge the training set or replay past data, but these strategies incur high data collection costs, sample inefficiently and fail to balance long- and short-term experience, leaving them short of human-like continual learning. Here we introduce Dual-LS, a task-free, online continual learning paradigm for DNN-based motion forecasting that is inspired by the complementary learning system of the human brain. Dual-LS pairs two synergistic memory rehearsal replay mechanisms to accelerate experience retrieval while dynamically coordinating long-term and short-term knowledge representations. Tests on naturalistic data spanning three countries, over 772,000 vehicles and cumulative testing mileage of 11,187 km show that Dual-LS mitigates catastrophic forgetting by up to 74.31\\% and reduces computational resource demand by up to 94.02\\%, markedly boosting predictive stability in vehicle motion forecasting without inflating data requirements. Meanwhile, it endows DNN-based vehicle motion forecasting with computation efficient and human-like continual learning adaptability fit for smart cities.", "date": "2025-08-27", "categories": []}
{"id": "2509.06219", "title": "MCIGLE: Multimodal Exemplar-Free Class-Incremental Graph Learning", "abstract": "Exemplar-free class-incremental learning enables models to learn new classes over time without storing data from old ones. As multimodal graph-structured data becomes increasingly prevalent, existing methods struggle with challenges like catastrophic forgetting, distribution bias, memory limits, and weak generalization. We propose MCIGLE, a novel framework that addresses these issues by extracting and aligning multimodal graph features and applying Concatenated Recursive Least Squares for effective knowledge retention. Through multi-channel processing, MCIGLE balances accuracy and memory preservation. Experiments on public datasets validate its effectiveness and generalizability.", "date": "2025-09-07", "categories": []}
{"id": "2509.04815", "title": "An Arbitration Control for an Ensemble of Diversified DQN variants in Continual Reinforcement Learning", "abstract": "Deep reinforcement learning (RL) models, despite their efficiency in learning an optimal policy in static environments, easily loses previously learned knowledge (i.e., catastrophic forgetting). It leads RL models to poor performance in continual reinforcement learning (CRL) scenarios. To address this, we present an arbitration control mechanism over an ensemble of RL agents. It is motivated by and closely aligned with how humans make decisions in a CRL context using an arbitration control of multiple RL agents in parallel as observed in the prefrontal cortex. We integrated two key ideas into our model: (1) an ensemble of RLs (i.e., DQN variants) explicitly trained to have diverse value functions and (2) an arbitration control that prioritizes agents with higher reliability (i.e., less error) in recent trials. We propose a framework for CRL, an Arbitration Control for an Ensemble of Diversified DQN variants (ACED-DQN). We demonstrate significant performance improvements in both static and continual environments, supported by empirical evidence showing the effectiveness of arbitration control over diversified DQNs during training. In this work, we introduced a framework that enables RL agents to continuously learn, with inspiration from the human brain.", "date": "2025-09-05", "categories": []}
{"id": "2401.11666", "title": "P2DT: Mitigating Forgetting in task-incremental Learning with progressive prompt Decision Transformer", "abstract": "Catastrophic forgetting poses a substantial challenge for managing intelligent agents controlled by a large model, causing performance degradation when these agents face new tasks. In our work, we propose a novel solution - the Progressive Prompt Decision Transformer (P2DT). This method enhances a transformer-based model by dynamically appending decision tokens during new task training, thus fostering task-specific policies. Our approach mitigates forgetting in continual and offline reinforcement learning scenarios. Moreover, P2DT leverages trajectories collected via traditional reinforcement learning from all tasks and generates new task-specific tokens during training, thereby retaining knowledge from previous studies. Preliminary results demonstrate that our model effectively alleviates catastrophic forgetting and scales well with increasing task environments.", "date": "2024-01-22", "categories": []}
{"id": "2401.11667", "title": "INCPrompt: Task-Aware incremental Prompting for Rehearsal-Free Class-incremental Learning", "abstract": "This paper introduces INCPrompt, an innovative continual learning solution that effectively addresses catastrophic forgetting. INCPrompt's key innovation lies in its use of adaptive key-learner and task-aware prompts that capture task-relevant information. This unique combination encapsulates general knowledge across tasks and encodes task-specific knowledge. Our comprehensive evaluation across multiple continual learning benchmarks demonstrates INCPrompt's superiority over existing algorithms, showing its effectiveness in mitigating catastrophic forgetting while maintaining high performance. These results highlight the significant impact of task-aware incremental prompting on continual learning performance.", "date": "2024-01-22", "categories": []}
{"id": "2506.19608", "title": "ChordPrompt: Orchestrating Cross-Modal Prompt Synergy for Multi-Domain Incremental Learning in CLIP", "abstract": "Continual learning (CL) empowers pre-trained vision-language models to adapt effectively to novel or previously underrepresented data distributions without comprehensive retraining, enhancing their adaptability and efficiency. While vision-language models like CLIP show great promise, they struggle to maintain performance across domains in incremental learning scenarios. Existing prompt learning methods face two main limitations: 1) they primarily focus on class-incremental learning scenarios, lacking specific strategies for multi-domain task incremental learning; 2) most current approaches employ single-modal prompts, neglecting the potential benefits of cross-modal information exchange. To address these challenges, we propose the \\ChordPrompt framework, which facilitates a harmonious interplay between visual and textual prompts. \\ChordPrompt introduces cross-modal prompts to leverage interactions between visual and textual information. Our approach also employs domain-adaptive text prompts to select appropriate prompts for continual adaptation across multiple domains. Comprehensive experiments on multi-domain incremental learning benchmarks demonstrate that \\ChordPrompt outperforms state-of-the-art methods in zero-shot generalization and downstream task performance.", "date": "2025-06-24", "categories": []}
{"id": "2404.18311", "title": "Towards Incremental Learning in Large Language Models: A Critical Review", "abstract": "Incremental learning is the ability of systems to acquire knowledge over time, enabling their adaptation and generalization to novel tasks. It is a critical ability for intelligent, real-world systems, especially when data changes frequently or is limited. This review provides a comprehensive analysis of incremental learning in Large Language Models. It synthesizes the state-of-the-art incremental learning paradigms, including continual learning, meta-learning, parameter-efficient learning, and mixture-of-experts learning. We demonstrate their utility for incremental learning by describing specific achievements from these related topics and their critical factors. An important finding is that many of these approaches do not update the core model, and none of them update incrementally in real-time. The paper highlights current problems and challenges for future research in the field. By consolidating the latest relevant research developments, this review offers a comprehensive understanding of incremental learning and its implications for designing and developing LLM-based learning systems.", "date": "2024-04-28", "categories": []}
{"id": "2509.00047", "title": "Teaching AI to Remember: Insights from Brain-Inspired Replay in Continual Learning", "abstract": "Artificial neural networks (ANNs) continue to face challenges in continual learning, particularly due to catastrophic forgetting, the loss of previously learned knowledge when acquiring new tasks. Inspired by memory consolidation in the human brain, we investigate the internal replay mechanism proposed by~\\citep{brain_inspired_replay1}, which reactivates latent representations of prior experiences during learning. As internal replay was identified as the most influential component among the brain-inspired mechanisms in their framework, it serves as the central focus of our in-depth investigation. Using the CIFAR-100 dataset in a class-incremental setting, we evaluate the effectiveness of internal replay, both in isolation and in combination with Synaptic Intelligence (SI). Our experiments show that internal replay significantly mitigates forgetting, especially when paired with SI, but at the cost of reduced initial task accuracy, highlighting a trade-off between memory stability and learning plasticity. Further analyses using log-likelihood distributions, reconstruction errors, silhouette scores, and UMAP projections reveal that internal replay increases representational overlap in latent space, potentially limiting task-specific differentiation. These results underscore the limitations of current brain-inspired methods and suggest future directions for balancing retention and adaptability in continual learning systems.", "date": "2025-08-24", "categories": []}
{"id": "2509.00735", "title": "Task-Aware Adaptive Modulation: A Replay-Free and Resource-Efficient Approach For Continual Graph Learning", "abstract": "Continual Graph Learning(CGL)focuses on acquiring new knowledge while retaining previously learned information, essential for real-world graph applications. Current methods grapple with two main issues:1) The Stability-Plasticity Dilemma: Replay-based methods often create an imbalance between the Dilemma, while incurring significant storage costs.2) The Resource-Heavy Pre-training: Leading replay-free methods critically depend on extensively pre-trained backbones, this reliance imposes a substantial resource burden.In this paper, we argue that the key to overcoming these challenges lies not in replaying data or fine-tuning the entire network, but in dynamically modulating the internal computational flow of a frozen backbone. We posit that lightweight, task-specific modules can effectively steer a GNN's reasoning process. Motivated by this insight, we propose Task-Aware Adaptive Modulation(TAAM), a replay-free, resource-efficient approach that charts a new path for navigating the stability-plasticity dilemma. TAAM's core is its Neural Synapse Modulators(NSM), which are trained and then frozen for each task to store expert knowledge. A pivotal prototype-guided strategy governs these modulators: 1) For training, it initializes a new NSM by deep-copying from a similar past modulator to boost knowledge transfer. 2) For inference, it selects the most relevant frozen NSM for each task. These NSMs insert into a frozen GNN backbone to perform fine-grained, node-attentive modulation of its internal flow-different from the static perturbations of prior methods. Extensive experiments show that TAAM comprehensively outperforms state-of-the-art methods across six GCIL benchmark datasets. The code will be released upon acceptance of the paper.", "date": "2025-08-31", "categories": []}
{"id": "2509.01213", "title": "Mitigating Catastrophic Forgetting in Continual Learning through Model Growth", "abstract": "Catastrophic forgetting is a significant challenge in continual learning, in which a model loses prior knowledge when it is fine-tuned on new tasks. This problem is particularly critical for large language models (LLMs) undergoing continual learning, as retaining performance across diverse domains is important for their general utility. In this paper, we explore model growth, a promising strategy that leverages smaller models to expedite and structure the training of larger ones for mitigating the catastrophic forgetting problem. Although growth-based pretraining, particularly via transformer stacking, has shown promise in accelerating convergence, its impact on forgetting remains under-explored. Therefore, we evaluate whether growth-based models can retain previously learned capabilities more effectively across a sequence of fine-tuning tasks involving domain knowledge, reasoning, reading comprehension, and bias. Our findings show that both models -- one trained with growth (Stack LLM) and one without (LLM) -- exhibit improvements in domain knowledge. However, reasoning and reading comprehension degrade over time, indicating signs of catastrophic forgetting. Stack LLM consistently shows less degradation, especially in reading comprehension, suggesting enhanced retention capabilities. Interestingly, in bias evaluation, the baseline LLM becomes progressively more neutral with continued fine-tuning, while Stack LLM maintains a steady bias ratio around 60--61\\%. These results indicate that growth-based pretraining may deliver modest improvements in resisting catastrophic forgetting, though trade-offs remain in handling social biases.", "date": "2025-09-01", "categories": []}
{"id": "2509.01533", "title": "Forward-Only Continual Learning", "abstract": "Catastrophic forgetting remains a central challenge in continual learning (CL) with pre-trained models. While existing approaches typically freeze the backbone and fine-tune a small number of parameters to mitigate forgetting, they still rely on iterative error backpropagation and gradient-based optimization, which can be computationally intensive and less suitable for resource-constrained environments. To address this, we propose FoRo, a forward-only, gradient-free continual learning method. FoRo consists of a lightweight prompt tuning strategy and a novel knowledge encoding mechanism, both designed without modifying the pre-trained model. Specifically, prompt embeddings are inserted at the input layer and optimized using the Covariance Matrix Adaptation Evolution Strategy (CMA-ES), which mitigates distribution shifts and extracts high-quality task representations. Subsequently, task-specific knowledge is encoded into a knowledge encoding matrix via nonlinear random projection and recursive least squares, enabling incremental updates to the classifier without revisiting prior data. Experiments show that FoRo significantly reduces average forgetting and improves accuracy. Thanks to forward-only learning, FoRo reduces memory usage and run time while maintaining high knowledge retention across long task sequences. These results suggest that FoRo could serve as a promising direction for exploring continual learning with pre-trained models, especially in real-world multimedia applications where both efficiency and effectiveness are critical.", "date": "2025-09-01", "categories": []}
{"id": "2509.01730", "title": "BM-CL: Bias Mitigation through the lens of Continual Learning", "abstract": "Biases in machine learning pose significant challenges, particularly when models amplify disparities that affect disadvantaged groups. Traditional bias mitigation techniques often lead to a {\\itshape leveling-down effect}, whereby improving outcomes of disadvantaged groups comes at the expense of reduced performance for advantaged groups. This study introduces Bias Mitigation through Continual Learning (BM-CL), a novel framework that leverages the principles of continual learning to address this trade-off. We postulate that mitigating bias is conceptually similar to domain-incremental continual learning, where the model must adjust to changing fairness conditions, improving outcomes for disadvantaged groups without forgetting the knowledge that benefits advantaged groups. Drawing inspiration from techniques such as Learning without Forgetting and Elastic Weight Consolidation, we reinterpret bias mitigation as a continual learning problem. This perspective allows models to incrementally balance fairness objectives, enhancing outcomes for disadvantaged groups while preserving performance for advantaged groups. Experiments on synthetic and real-world image datasets, characterized by diverse sources of bias, demonstrate that the proposed framework mitigates biases while minimizing the loss of original knowledge. Our approach bridges the fields of fairness and continual learning, offering a promising pathway for developing machine learning systems that are both equitable and effective.", "date": "2025-09-01", "categories": []}
{"id": "2508.18860", "title": "C-Flat++: Towards a More Efficient and Powerful Framework for Continual Learning", "abstract": "Balancing sensitivity to new tasks and stability for retaining past knowledge is crucial in continual learning (CL). Recently, sharpness-aware minimization has proven effective in transfer learning and has also been adopted in continual learning (CL) to improve memory retention and learning efficiency. However, relying on zeroth-order sharpness alone may favor sharper minima over flatter ones in certain settings, leading to less robust and potentially suboptimal solutions. In this paper, we propose \\textbf{C}ontinual \\textbf{Flat}ness (\\textbf{C-Flat}), a method that promotes flatter loss landscapes tailored for CL. C-Flat offers plug-and-play compatibility, enabling easy integration with minimal modifications to the code pipeline. Besides, we present a general framework that integrates C-Flat into all major CL paradigms and conduct comprehensive comparisons with loss-minima optimizers and flat-minima-based CL methods. Our results show that C-Flat consistently improves performance across a wide range of settings. In addition, we introduce C-Flat++, an efficient yet effective framework that leverages selective flatness-driven promotion, significantly reducing the update cost required by C-Flat. Extensive experiments across multiple CL methods, datasets, and scenarios demonstrate the effectiveness and efficiency of our proposed approaches. Code is available at https://github.com/WanNaa/C-Flat.", "date": "2025-08-26", "categories": []}
{"id": "2508.21240", "title": "Class Incremental Continual Learning with Self-Organizing Maps and Variational Autoencoders Using Synthetic Replay", "abstract": "This work introduces a novel generative continual learning framework based on self-organizing maps (SOMs) and variational autoencoders (VAEs) to enable memory-efficient replay, eliminating the need to store raw data samples or task labels. For high-dimensional input spaces, such as of CIFAR-10 and CIFAR-100, we design a scheme where the SOM operates over the latent space learned by a VAE, whereas, for lower-dimensional inputs, such as those found in MNIST and FashionMNIST, the SOM operates in a standalone fashion. Our method stores a running mean, variance, and covariance for each SOM unit, from which synthetic samples are then generated during future learning iterations. For the VAE-based method, generated samples are then fed through the decoder to then be used in subsequent replay. Experimental results on standard class-incremental benchmarks show that our approach performs competitively with state-of-the-art memory-based methods and outperforms memory-free methods, notably improving over best state-of-the-art single class incremental performance on CIFAR-10 and CIFAR-100 by nearly $10$\\% and $7$\\%, respectively. Our methodology further facilitates easy visualization of the learning process and can also be utilized as a generative model post-training. Results show our method's capability as a scalable, task-label-free, and memory-efficient solution for continual learning.", "date": "2025-08-28", "categories": []}
{"id": "2508.21424", "title": "Unsupervised Incremental Learning Using Confidence-Based Pseudo-Labels", "abstract": "Deep learning models have achieved state-of-the-art performance in many computer vision tasks. However, in real-world scenarios, novel classes that were unseen during training often emerge, requiring models to acquire new knowledge incrementally. Class-Incremental Learning (CIL) methods enable a model to learn novel classes while retaining knowledge of previous classes. However, these methods make the strong assumption that the incremental dataset is fully labeled, which is unrealistic in practice. In this work, we propose an unsupervised Incremental Learning method using Confidence-based Pseudo-labels (ICPL), which replaces human annotations with pseudo-labels, enabling incremental learning from unlabeled datasets. We integrate these pseudo-labels into various CIL methods with confidence-based selection and evaluate performance degradation on CIFAR100 and ImageNet100. Then, we compare our approach to popular Class Incremental Novel Category Discovery (class-iNCD) methods addressing similar challenges. Additionally, we apply our method to fine-grained datasets to demonstrate its real-world practicality and measure its computational complexity to validate its suitability for resource-constrained environments. ICPL achieves competitive results compared to supervised methods and outperforms state-of-the-art class-iNCD methods by more than 5% in final accuracy.", "date": "2025-08-29", "categories": []}
{"id": "2508.21773", "title": "Unsupervised Video Continual Learning via Non-Parametric Deep Embedded Clustering", "abstract": "We propose a realistic scenario for the unsupervised video learning where neither task boundaries nor labels are provided when learning a succession of tasks. We also provide a non-parametric learning solution for the under-explored problem of unsupervised video continual learning. Videos represent a complex and rich spatio-temporal media information, widely used in many applications, but which have not been sufficiently explored in unsupervised continual learning. Prior studies have only focused on supervised continual learning, relying on the knowledge of labels and task boundaries, while having labeled data is costly and not practical. To address this gap, we study the unsupervised video continual learning (uVCL). uVCL raises more challenges due to the additional computational and memory requirements of processing videos when compared to images. We introduce a general benchmark experimental protocol for uVCL by considering the learning of unstructured video data categories during each task. We propose to use the Kernel Density Estimation (KDE) of deep embedded video features extracted by unsupervised video transformer networks as a non-parametric probabilistic representation of the data. We introduce a novelty detection criterion for the incoming new task data, dynamically enabling the expansion of memory clusters, aiming to capture new knowledge when learning a succession of tasks. We leverage the use of transfer learning from the previous tasks as an initial state for the knowledge transfer to the current learning task. We found that the proposed methodology substantially enhances the performance of the model when successively learning many tasks. We perform in-depth evaluations on three standard video action recognition datasets, including UCF101, HMDB51, and Something-to-Something V2, without using any labels or class boundaries.", "date": "2025-08-29", "categories": []}
{"id": "2412.17009", "title": "Expert Routing with Synthetic Data for Continual Learning", "abstract": "In many real-world settings, regulations and economic incentives permit the sharing of models but not data across institutional boundaries. In such scenarios, practitioners might hope to adapt models to new domains, without losing performance on previous domains (so-called catastrophic forgetting). While any single model may struggle to achieve this goal, learning an ensemble of domain-specific experts offers the potential to adapt more closely to each individual institution. However, a core challenge in this context is determining which expert to deploy at test time. In this paper, we propose Generate to Discriminate (G2D), a domain-incremental continual learning method that leverages synthetic data to train a domain-discriminator that routes samples at inference time to the appropriate expert. Surprisingly, we find that leveraging synthetic data in this capacity is more effective than using the samples to \\textit{directly} train the downstream classifier (the more common approach to leveraging synthetic data in the lifelong learning literature). We observe that G2D outperforms competitive domain-incremental learning methods on tasks in both vision and language modalities, providing a new perspective on the use of synthetic data in the lifelong learning literature.", "date": "2024-12-22", "categories": []}
{"id": "2508.19571", "title": "Escaping Stability-Plasticity Dilemma in Online Continual Learning for Motion Forecasting via Synergetic Memory Rehearsal", "abstract": "Deep neural networks (DNN) have achieved remarkable success in motion forecasting. However, most DNN-based methods suffer from catastrophic forgetting and fail to maintain their performance in previously learned scenarios after adapting to new data. Recent continual learning (CL) studies aim to mitigate this phenomenon by enhancing memory stability of DNN, i.e., the ability to retain learned knowledge. Yet, excessive emphasis on the memory stability often impairs learning plasticity, i.e., the capacity of DNN to acquire new information effectively. To address such stability-plasticity dilemma, this study proposes a novel CL method, synergetic memory rehearsal (SyReM), for DNN-based motion forecasting. SyReM maintains a compact memory buffer to represent learned knowledge. To ensure memory stability, it employs an inequality constraint that limits increments in the average loss over the memory buffer. Synergistically, a selective memory rehearsal mechanism is designed to enhance learning plasticity by selecting samples from the memory buffer that are most similar to recently observed data. This selection is based on an online-measured cosine similarity of loss gradients, ensuring targeted memory rehearsal. Since replayed samples originate from learned scenarios, this memory rehearsal mechanism avoids compromising memory stability. We validate SyReM under an online CL paradigm where training samples from diverse scenarios arrive as a one-pass stream. Experiments on 11 naturalistic driving datasets from INTERACTION demonstrate that, compared to non-CL and CL baselines, SyReM significantly mitigates catastrophic forgetting in past scenarios while improving forecasting accuracy in new ones. The implementation is publicly available at https://github.com/BIT-Jack/SyReM.", "date": "2025-08-27", "categories": []}
{"id": "2407.19183", "title": "Graph Memory Learning: Imitating Lifelong Remembering and Forgetting of Brain Networks", "abstract": "Graph data in real-world scenarios undergo rapid and frequent changes, making it challenging for existing graph models to effectively handle the continuous influx of new data and accommodate data withdrawal requests. The approach to frequently retraining graph models is resource intensive and impractical. To address this pressing challenge, this paper introduces a new concept of graph memory learning. Its core idea is to enable a graph model to selectively remember new knowledge but forget old knowledge. Building on this approach, the paper presents a novel graph memory learning framework - Brain-inspired Graph Memory Learning (BGML), inspired by brain network dynamics and function-structure coupling strategies. BGML incorporates a multi-granular hierarchical progressive learning mechanism rooted in feature graph grain learning to mitigate potential conflict between memorization and forgetting in graph memory learning. This mechanism allows for a comprehensive and multi-level perception of local details within evolving graphs. In addition, to tackle the issue of unreliable structures in newly added incremental information, the paper introduces an information self-assessment ownership mechanism. This mechanism not only facilitates the propagation of incremental information within the model but also effectively preserves the integrity of past experiences. We design five types of graph memory learning tasks: regular, memory, unlearning, data-incremental, and class-incremental to evaluate BGML. Its excellent performance is confirmed through extensive experiments on multiple real-world node classification datasets.", "date": "2024-07-27", "categories": []}
{"id": "2410.05849", "title": "ModalPrompt: Towards Efficient Multimodal Continual Instruction Tuning with Dual-Modality Guided Prompt", "abstract": "Large Multimodal Models (LMMs) exhibit remarkable multi-tasking ability by learning mixed instruction datasets. However, novel tasks would be encountered sequentially in dynamic world, which urges for equipping LMMs with multimodal continual instruction learning (MCIT) ability especially for diverse and challenging generative tasks. Existing MCIT methods do not fully exploit the unique attribute of LMMs and often gain performance at the expense of efficiency. In this paper, we propose a novel prompt learning framework for MCIT to effectively alleviate forgetting of previous knowledge while managing computational complexity with natural image-text supervision. Concretely, we learn prompts for each task and exploit efficient prompt fusion for knowledge transfer and prompt selection for complexity management with dual-modality guidance. Extensive experiments demonstrate that our approach achieves substantial +14.26% performance gain on MCIT benchmarks with remarkable $\\times$ 1.42 inference speed free from growing computation. Code is available at https://github.com/AuroraZengfh/ModalPrompt.", "date": "2024-10-08", "categories": []}
{"id": "2504.05214", "title": "Post-Training Language Models for Continual Relation Extraction", "abstract": "Real-world data, such as news articles, social media posts, and chatbot conversations, is inherently dynamic and non-stationary, presenting significant challenges for constructing real-time structured representations through knowledge graphs (KGs). Relation Extraction (RE), a fundamental component of KG creation, often struggles to adapt to evolving data when traditional models rely on static, outdated datasets. Continual Relation Extraction (CRE) methods tackle this issue by incrementally learning new relations while preserving previously acquired knowledge. This study investigates the application of pre-trained language models (PLMs), specifically large language models (LLMs), to CRE, with a focus on leveraging memory replay to address catastrophic forgetting. We evaluate decoder-only models (eg, Mistral-7B and Llama2-7B) and encoder-decoder models (eg, Flan-T5 Base) on the TACRED and FewRel datasets. Task-incremental fine-tuning of LLMs demonstrates superior performance over earlier approaches using encoder-only models like BERT on TACRED, excelling in seen-task accuracy and overall performance (measured by whole and average accuracy), particularly with the Mistral and Flan-T5 models. Results on FewRel are similarly promising, achieving second place in whole and average accuracy metrics. This work underscores critical factors in knowledge transfer, language model architecture, and KG completeness, advancing CRE with LLMs and memory replay for dynamic, real-time relation extraction.", "date": "2025-04-07", "categories": []}
{"id": "2506.13045", "title": "Continual Learning for Generative AI: From LLMs to MLLMs and Beyond", "abstract": "The rapid advancement of generative models has empowered modern AI systems to comprehend and produce highly sophisticated content, even achieving human-level performance in specific domains. However, these models are fundamentally constrained by \\emph{catastrophic forgetting}, \\ie~a persistent challenge where models experience performance degradation on previously learned tasks when adapting to new tasks. To address this practical limitation, numerous approaches have been proposed to enhance the adaptability and scalability of generative AI in real-world applications. In this work, we present a comprehensive survey of continual learning methods for mainstream generative AI models, encompassing large language models, multimodal large language models, vision-language-action models, and diffusion models. Drawing inspiration from the memory mechanisms of the human brain, we systematically categorize these approaches into three paradigms: architecture-based, regularization-based, and replay-based methods, while elucidating their underlying methodologies and motivations. We further analyze continual learning setups for different generative models, including training objectives, benchmarks, and core backbones, thereby providing deeper insights into the field. The project page of this paper is available at https://github.com/Ghy0501/Awesome-Continual-Learning-in-Generative-Models.", "date": "2025-06-16", "categories": []}
{"id": "2508.16632", "title": "Adaptive Variance-Penalized Continual Learning with Fisher Regularization", "abstract": "The persistent challenge of catastrophic forgetting in neural networks has motivated extensive research in continual learning . This work presents a novel continual learning framework that integrates Fisher-weighted asymmetric regularization of parameter variances within a variational learning paradigm. Our method dynamically modulates regularization intensity according to parameter uncertainty, achieving enhanced stability and performance. Comprehensive evaluations on standard continual learning benchmarks including SplitMNIST, PermutedMNIST, and SplitFashionMNIST demonstrate substantial improvements over existing approaches such as Variational Continual Learning and Elastic Weight Consolidation . The asymmetric variance penalty mechanism proves particularly effective in maintaining knowledge across sequential tasks while improving model accuracy. Experimental results show our approach not only boosts immediate task performance but also significantly mitigates knowledge degradation over time, effectively addressing the fundamental challenge of catastrophic forgetting in neural networks", "date": "2025-08-15", "categories": []}
{"id": "2501.05017", "title": "Continuous Knowledge-Preserving Decomposition with Adaptive Layer Selection for Few-Shot Class-Incremental Learning", "abstract": "Few-Shot Class-Incremental Learning (FSCIL) faces a critical challenge: balancing the retention of prior knowledge with the acquisition of new classes. Existing methods either freeze the backbone to prevent catastrophic forgetting, sacrificing plasticity, or add new modules, incurring high costs. These approaches treat pretrained models as black boxes, overlooking two key opportunities to exploit their internal capacity: reusing redundant representational space within layers and selectively adapting layers based on their sensitivity to forgetting. We propose CKPD-FSCIL, a unified framework that unlocks the underutilized capacity of pretrained weights, achieving a superior stability-plasticity balance with zero inference overhead. Our design integrates two continuously adapting mechanisms: At the weight level, a Continuous Knowledge-Preserving Decomposition mechanism uses feature covariance to split each weight matrix into a frozen subspace that safeguards prior knowledge and a learnable, redundant subspace for new tasks. At the layer level, a Continuous Adaptive Layer Selection mechanism leverages an Adapter Sensitivity Ratio to automatically select layers with the highest redundant capacity and lowest forgetting risk for adaptation. By targeting only safe, high-potential subspaces and layers, CKPD-FSCIL enables efficient adaptation. After each session, the learned adapters are merged back into the original weights, ensuring zero additional parameters or FLOPs during inference. Extensive experiments on multiple FSCIL benchmarks demonstrate that our method consistently outperforms state-of-the-art approaches in both adaptability and knowledge retention. The code is available at https://github.com/xiaojieli0903/CKPD-FSCIL.", "date": "2025-01-09", "categories": []}
{"id": "2503.24259", "title": "Advances in Continual Graph Learning for Anti-Money Laundering Systems: A Comprehensive Review", "abstract": "Financial institutions are required by regulation to report suspicious financial transactions related to money laundering. Therefore, they need to constantly monitor vast amounts of incoming and outgoing transactions. A particular challenge in detecting money laundering is that money launderers continuously adapt their tactics to evade detection. Hence, detection methods need constant fine-tuning. Traditional machine learning models suffer from catastrophic forgetting when fine-tuning the model on new data, thereby limiting their effectiveness in dynamic environments. Continual learning methods may address this issue and enhance current anti-money laundering (AML) practices, by allowing models to incorporate new information while retaining prior knowledge. Research on continual graph learning for AML, however, is still scarce. In this review, we critically evaluate state-of-the-art continual graph learning approaches for AML applications. We categorise methods into replay-based, regularization-based, and architecture-based strategies within the graph neural network (GNN) framework, and we provide in-depth experimental evaluations on both synthetic and real-world AML data sets that showcase the effect of the different hyperparameters. Our analysis demonstrates that continual learning improves model adaptability and robustness in the face of extreme class imbalances and evolving fraud patterns. Finally, we outline key challenges and propose directions for future research.", "date": "2025-03-31", "categories": []}
{"id": "2505.11936", "title": "CCD: Continual Consistency Diffusion for Lifelong Generative Modeling", "abstract": "While diffusion-based models have shown remarkable generative capabilities in static settings, their extension to continual learning (CL) scenarios remains fundamentally constrained by Generative Catastrophic Forgetting (GCF). We observe that even with a rehearsal buffer, new generative skills often overwrite previous ones, degrading performance on earlier tasks. Although some initial efforts have explored this space, most rely on heuristics borrowed from continual classification methods or use trained diffusion models as ad hoc replay generators, lacking a principled, unified solution to mitigating GCF and often conducting experiments under fragmented and inconsistent settings. To address this gap, we introduce the Continual Diffusion Generation (CDG), a structured pipeline that redefines how diffusion models are implemented under CL and enables systematic evaluation of GCF. Beyond the empirical pipeline, we propose the first theoretical foundation for CDG, grounded in a cross-task analysis of diffusion-specific generative dynamics. Our theoretical investigation identifies three fundamental consistency principles essential for preserving knowledge in the rehearsal buffer over time: inter-task knowledge consistency, unconditional knowledge consistency, and prior knowledge consistency. These criteria expose the latent mechanisms through which generative forgetting manifests across sequential tasks. Motivated by these insights, we further propose \\textit{Continual Consistency Diffusion} (CCD), a principled training framework that enforces these consistency objectives via hierarchical loss functions: $\\mathcal{L}_{IKC}$, $\\mathcal{L}_{UKC}$, and $\\mathcal{L}_{PKC}$. Extensive experiments show that CCD achieves SOTA performance across various benchmarks, especially improving generative metrics in overlapping-task scenarios.", "date": "2025-05-17", "categories": []}
{"id": "2409.13792", "title": "Continual Learning for Multimodal Data Fusion of a Soft Gripper", "abstract": "Continual learning (CL) refers to the ability of an algorithm to continuously and incrementally acquire new knowledge from its environment while retaining previously learned information. A model trained on one data modality often fails when tested with a different modality. A straightforward approach might be to fuse the two modalities by concatenating their features and training the model on the fused data. However, this requires retraining the model from scratch each time it encounters a new domain. In this paper, we introduce a continual learning algorithm capable of incrementally learning different data modalities by leveraging both class-incremental and domain-incremental learning scenarios in an artificial environment where labeled data is scarce, yet non-iid (independent and identical distribution) unlabeled data from the environment is plentiful. The proposed algorithm is efficient and only requires storing prototypes for each class. We evaluate the algorithm's effectiveness on a challenging custom multimodal dataset comprising of tactile data from a soft pneumatic gripper, and visual data from non-stationary images of objects extracted from video sequences. Additionally, we conduct an ablation study on the custom dataset and the Core50 dataset to highlight the contributions of different components of the algorithm. To further demonstrate the robustness of the algorithm, we perform a real-time experiment for object classification using the soft gripper and an external independent camera setup, all synchronized with the Robot Operating System (ROS) framework.", "date": "2024-09-20", "categories": []}
{"id": "2508.15494", "title": "High-dimensional Asymptotics of Generalization Performance in Continual Ridge Regression", "abstract": "Continual learning is motivated by the need to adapt to real-world dynamics in tasks and data distribution while mitigating catastrophic forgetting. Despite significant advances in continual learning techniques, the theoretical understanding of their generalization performance lags behind. This paper examines the theoretical properties of continual ridge regression in high-dimensional linear models, where the dimension is proportional to the sample size in each task. Using random matrix theory, we derive exact expressions of the asymptotic prediction risk, thereby enabling the characterization of three evaluation metrics of generalization performance in continual learning: average risk, backward transfer, and forward transfer. Furthermore, we present the theoretical risk curves to illustrate the trends in these evaluation metrics throughout the continual learning process. Our analysis reveals several intriguing phenomena in the risk curves, demonstrating how model specifications influence the generalization performance. Simulation studies are conducted to validate our theoretical findings.", "date": "2025-08-21", "categories": []}
{"id": "2504.13691", "title": "MEGA: Second-Order Gradient Alignment for Catastrophic Forgetting Mitigation in GFSCIL", "abstract": "Graph Few-Shot Class-Incremental Learning (GFSCIL) enables models to continually learn from limited samples of novel tasks after initial training on a large base dataset. Existing GFSCIL approaches typically utilize Prototypical Networks (PNs) for metric-based class representations and fine-tune the model during the incremental learning stage. However, these PN-based methods oversimplify learning via novel query set fine-tuning and fail to integrate Graph Continual Learning (GCL) techniques due to architectural constraints. To address these challenges, we propose a more rigorous and practical setting for GFSCIL that excludes query sets during the incremental training phase. Building on this foundation, we introduce Model-Agnostic Meta Graph Continual Learning (MEGA), aimed at effectively alleviating catastrophic forgetting for GFSCIL. Specifically, by calculating the incremental second-order gradient during the meta-training stage, we endow the model to learn high-quality priors that enhance incremental learning by aligning its behaviors across both the meta-training and incremental learning stages. Extensive experiments on four mainstream graph datasets demonstrate that MEGA achieves state-of-the-art results and enhances the effectiveness of various GCL methods in GFSCIL. We believe that our proposed MEGA serves as a model-agnostic GFSCIL paradigm, paving the way for future research.", "date": "2025-04-18", "categories": []}
{"id": "2508.00212", "title": "Reinitializing weights vs units for maintaining plasticity in neural networks", "abstract": "Loss of plasticity is a phenomenon in which a neural network loses its ability to learn when trained for an extended time on non-stationary data. It is a crucial problem to overcome when designing systems that learn continually. An effective technique for preventing loss of plasticity is reinitializing parts of the network. In this paper, we compare two different reinitialization schemes: reinitializing units vs reinitializing weights. We propose a new algorithm, which we name \\textit{selective weight reinitialization}, for reinitializing the least useful weights in a network. We compare our algorithm to continual backpropagation and ReDo, two previously proposed algorithms that reinitialize units in the network. Through our experiments in continual supervised learning problems, we identify two settings when reinitializing weights is more effective at maintaining plasticity than reinitializing units: (1) when the network has a small number of units and (2) when the network includes layer normalization. Conversely, reinitializing weights and units are equally effective at maintaining plasticity when the network is of sufficient size and does not include layer normalization. We found that reinitializing weights maintains plasticity in a wider variety of settings than reinitializing units.", "date": "2025-07-31", "categories": []}
{"id": "2508.14081", "title": "Toward Lifelong Learning in Equilibrium Propagation: Sleep-like and Awake Rehearsal for Enhanced Stability", "abstract": "Recurrent neural networks (RNNs) trained using Equilibrium Propagation (EP), a biologically plausible training algorithm, have demonstrated strong performance in various tasks such as image classification and reinforcement learning. However, these networks face a critical challenge in continuous learning: catastrophic forgetting, where previously acquired knowledge is overwritten when new tasks are learned. This limitation contrasts with the human brain's ability to retain and integrate both old and new knowledge, aided by processes like memory consolidation during sleep through the replay of learned information. To address this challenge in RNNs, here we propose a sleep-like replay consolidation (SRC) algorithm for EP-trained RNNs. We found that SRC significantly improves RNN's resilience to catastrophic forgetting in continuous learning scenarios. In class-incremental learning with SRC implemented after each new task training, the EP-trained multilayer RNN model (MRNN-EP) performed significantly better compared to feedforward networks incorporating several well-established regularization techniques. The MRNN-EP performed on par with MRNN trained using Backpropagation Through Time (BPTT) when both were equipped with SRC on MNIST data and surpassed BPTT-based models on the Fashion MNIST, Kuzushiji-MNIST, CIFAR10, and ImageNet datasets. Combining SRC with rehearsal, also known as \"awake replay\", further boosted the network's ability to retain long-term knowledge while continuing to learn new tasks. Our study reveals the applicability of sleep-like replay techniques to RNNs and highlights the potential for integrating human-like learning behaviors into artificial neural networks (ANNs).", "date": "2025-08-12", "categories": []}
{"id": "2504.13822", "title": "Parameter-Efficient Continual Fine-Tuning: A Survey", "abstract": "The emergence of large pre-trained networks has revolutionized the AI field, unlocking new possibilities and achieving unprecedented performance. However, these models inherit a fundamental limitation from traditional Machine Learning approaches: their strong dependence on the \\textit{i.i.d.} assumption hinders their adaptability to dynamic learning scenarios. We believe the next breakthrough in AI lies in enabling efficient adaptation to evolving environments -- such as the real world -- where new data and tasks arrive sequentially. This challenge defines the field of Continual Learning (CL), a Machine Learning paradigm focused on developing lifelong learning neural models. One alternative to efficiently adapt these large-scale models is known Parameter-Efficient Fine-Tuning (PEFT). These methods tackle the issue of adapting the model to a particular data or scenario by performing small and efficient modifications, achieving similar performance to full fine-tuning. However, these techniques still lack the ability to adjust the model to multiple tasks continually, as they suffer from the issue of Catastrophic Forgetting. In this survey, we first provide an overview of CL algorithms and PEFT methods before reviewing the state-of-the-art on Parameter-Efficient Continual Fine-Tuning (PECFT). We examine various approaches, discuss evaluation metrics, and explore potential future research directions. Our goal is to highlight the synergy between CL and Parameter-Efficient Fine-Tuning, guide researchers in this field, and pave the way for novel future research directions.", "date": "2025-04-18", "categories": []}
{"id": "2508.14024", "title": "UNICON: UNIfied CONtinual Learning for Medical Foundational Models", "abstract": "Foundational models are trained on extensive datasets to capture the general trends of a domain. However, in medical imaging, the scarcity of data makes pre-training for every domain, modality, or task challenging. Continual learning offers a solution by fine-tuning a model sequentially on different domains or tasks, enabling it to integrate new knowledge without requiring large datasets for each training phase. In this paper, we propose UNIfied CONtinual Learning for Medical Foundational Models (UNICON), a framework that enables the seamless adaptation of foundation models to diverse domains, tasks, and modalities. Unlike conventional adaptation methods that treat these changes in isolation, UNICON provides a unified, perpetually expandable framework. Through careful integration, we show that foundation models can dynamically expand across imaging modalities, anatomical regions, and clinical objectives without catastrophic forgetting or task interference. Empirically, we validate our approach by adapting a chest CT foundation model initially trained for classification to a prognosis and segmentation task. Our results show improved performance across both additional tasks. Furthermore, we continually incorporated PET scans and achieved a 5\\% improvement in Dice score compared to respective baselines. These findings establish that foundation models are not inherently constrained to their initial training scope but can evolve, paving the way toward generalist AI models for medical imaging.", "date": "2025-08-19", "categories": []}
{"id": "2406.09370", "title": "Data-dependent and Oracle Bounds on Forgetting in Continual Learning", "abstract": "In continual learning, knowledge must be preserved and re-used between tasks, maintaining good transfer to future tasks and minimizing forgetting of previously learned ones. While several practical algorithms have been devised for this setting, there have been few theoretical works aiming to quantify and bound the degree of Forgetting in general settings. For \\emph{exemplar-free} methods, we provide both data-dependent upper bounds that apply \\emph{regardless of model and algorithm choice}, and oracle bounds for Gibbs posteriors. We derive an algorithm based on our bounds and demonstrate empirically that our approach yields tight and practical bounds on forgetting for several continual learning problems and algorithms.", "date": "2024-06-13", "categories": []}
{"id": "2412.02509", "title": "FCL-ViT: Task-Aware Attention Tuning for Continual Learning", "abstract": "Continual Learning (CL) involves adapting the prior Deep Neural Network (DNN) knowledge to new tasks, without forgetting the old ones. However, modern CL techniques focus on provisioning memory capabilities to existing DNN models rather than designing new ones that are able to adapt according to the task at hand. This paper presents the novel Feedback Continual Learning Vision Transformer (FCL-ViT) that uses a feedback mechanism to generate real-time dynamic attention features tailored to the current task. The FCL-ViT operates in two Phases. In phase 1, the generic image features are produced and determine where the Transformer should attend on the current image. In phase 2, task-specific image features are generated that leverage dynamic attention. To this end, Tunable self-Attention Blocks (TABs) and Task Specific Blocks (TSBs) are introduced that operate in both phases and are responsible for tuning the TABs attention, respectively. The FCL-ViT surpasses state-of-the-art performance on Continual Learning compared to benchmark methods, while retaining a small number of trainable DNN parameters.", "date": "2024-12-03", "categories": []}
{"id": "2503.10003", "title": "Does Prior Data Matter? Exploring Joint Training in the Context of Few-Shot Class-Incremental Learning", "abstract": "Class-incremental learning (CIL) aims to adapt to continuously emerging new classes while preserving knowledge of previously learned ones. Few-shot class-incremental learning (FSCIL) presents a greater challenge that requires the model to learn new classes from only a limited number of samples per class. While incremental learning typically assumes restricted access to past data, it often remains available in many real-world scenarios. This raises a practical question: should one retrain the model on the full dataset (i.e., joint training), or continue updating it solely with new data? In CIL, joint training is considered an ideal benchmark that provides a reference for evaluating the trade-offs between performance and computational cost. However, in FSCIL, joint training becomes less reliable due to severe imbalance between base and incremental classes. This results in the absence of a practical baseline, making it unclear which strategy is preferable for practitioners. To this end, we revisit joint training in the context of FSCIL by incorporating imbalance mitigation techniques, and suggest a new imbalance-aware joint training benchmark for FSCIL. We then conduct extensive comparisons between this benchmark and FSCIL methods to analyze which approach is most suitable when prior data is accessible. Our analysis offers realistic insights and guidance for selecting training strategies in real-world FSCIL scenarios. Code is available at: https://github.com/shiwonkim/Joint_FSCIL", "date": "2025-03-13", "categories": []}
{"id": "2506.17169", "title": "Continual Learning with Columnar Spiking Neural Networks", "abstract": "Continual learning is a key feature of biological neural systems, but artificial neural networks often suffer from catastrophic forgetting. Instead of backpropagation, biologically plausible learning algorithms may enable stable continual learning. This study proposes columnar-organized spiking neural networks (SNNs) with local learning rules for continual learning and catastrophic forgetting. Using CoLaNET (Columnar Layered Network), we show that its microcolumns adapt most efficiently to new tasks when they lack shared structure with prior learning. We demonstrate how CoLaNET hyperparameters govern the trade-off between retaining old knowledge (stability) and acquiring new information (plasticity). We evaluate CoLaNET on two benchmarks: Permuted MNIST (ten sequential pixel-permuted tasks) and a two-task MNIST/EMNIST setup. Our model learns ten sequential tasks effectively, maintaining 92% accuracy on each. It shows low forgetting, with only 4% performance degradation on the first task after training on nine subsequent tasks.", "date": "2025-06-20", "categories": []}
{"id": "2508.04792", "title": "Federated Continual Recommendation", "abstract": "The increasing emphasis on privacy in recommendation systems has led to the adoption of Federated Learning (FL) as a privacy-preserving solution, enabling collaborative training without sharing user data. While Federated Recommendation (FedRec) effectively protects privacy, existing methods struggle with non-stationary data streams, failing to maintain consistent recommendation quality over time. On the other hand, Continual Learning Recommendation (CLRec) methods address evolving user preferences but typically assume centralized data access, making them incompatible with FL constraints. To bridge this gap, we introduce Federated Continual Recommendation (FCRec), a novel task that integrates FedRec and CLRec, requiring models to learn from streaming data while preserving privacy. As a solution, we propose F3CRec, a framework designed to balance knowledge retention and adaptation under the strict constraints of FCRec. F3CRec introduces two key components: Adaptive Replay Memory on the client side, which selectively retains past preferences based on user-specific shifts, and Item-wise Temporal Mean on the server side, which integrates new knowledge while preserving prior information. Extensive experiments demonstrate that F3CRec outperforms existing approaches in maintaining recommendation quality over time in a federated environment.", "date": "2025-08-06", "categories": []}
{"id": "2508.11673", "title": "Contrastive Regularization over LoRA for Multimodal Biomedical Image Incremental Learning", "abstract": "Multimodal Biomedical Image Incremental Learning (MBIIL) is essential for handling diverse tasks and modalities in the biomedical domain, as training separate models for each modality or task significantly increases inference costs. Existing incremental learning methods focus on task expansion within a single modality, whereas MBIIL seeks to train a unified model incrementally across modalities. The MBIIL faces two challenges: I) How to preserve previously learned knowledge during incremental updates? II) How to effectively leverage knowledge acquired from existing modalities to support new modalities? To address these challenges, we propose MSLoRA-CR, a method that fine-tunes Modality-Specific LoRA modules while incorporating Contrastive Regularization to enhance intra-modality knowledge sharing and promote inter-modality knowledge differentiation. Our approach builds upon a large vision-language model (LVLM), keeping the pretrained model frozen while incrementally adapting new LoRA modules for each modality or task. Experiments on the incremental learning of biomedical images demonstrate that MSLoRA-CR outperforms both the state-of-the-art (SOTA) approach of training separate models for each modality and the general incremental learning method (incrementally fine-tuning LoRA). Specifically, MSLoRA-CR achieves a 1.88% improvement in overall performance compared to unconstrained incremental learning methods while maintaining computational efficiency. Our code is publicly available at https://github.com/VentusAislant/MSLoRA_CR.", "date": "2025-08-08", "categories": []}
{"id": "2508.12031", "title": "Learning Wisdom from Errors: Promoting LLM's Continual Relation Learning through Exploiting Error Cases", "abstract": "Continual Relation Extraction (CRE) aims to continually learn new emerging relations while avoiding catastrophic forgetting. Existing CRE methods mainly use memory replay and contrastive learning to mitigate catastrophic forgetting. However, these methods do not attach importance to the error cases that can reveal the model's cognitive biases more effectively. To address this issue, we propose an instruction-based continual contrastive tuning approach for Large Language Models (LLMs) in CRE. Different from existing CRE methods that typically handle the training and memory data in a unified manner, this approach splits the training and memory data of each task into two parts respectively based on the correctness of the initial responses and treats them differently through dual-task fine-tuning. In addition, leveraging the advantages of LLM's instruction-following ability, we propose a novel instruction-based contrastive tuning strategy for LLM to continuously correct current cognitive biases with the guidance of previous data in an instruction-tuning manner, which mitigates the gap between old and new relations in a more suitable way for LLMs. We experimentally evaluate our model on TACRED and FewRel, and the results show that our model achieves new state-of-the-art CRE performance with significant improvements, demonstrating the importance of specializing in exploiting error cases.", "date": "2025-08-16", "categories": []}
{"id": "2508.12643", "title": "Learn Faster and Remember More: Balancing Exploration and Exploitation for Continual Test-time Adaptation", "abstract": "Continual Test-Time Adaptation (CTTA) aims to adapt a source pre-trained model to continually changing target domains during inference. As a fundamental principle, an ideal CTTA method should rapidly adapt to new domains (exploration) while retaining and exploiting knowledge from previously encountered domains to handle similar domains in the future. Despite significant advances, balancing exploration and exploitation in CTTA is still challenging: 1) Existing methods focus on adjusting predictions based on deep-layer outputs of neural networks. However, domain shifts typically affect shallow features, which are inefficient to be adjusted from deep predictions, leading to dilatory exploration; 2) A single model inevitably forgets knowledge of previous domains during the exploration, making it incapable of exploiting historical knowledge to handle similar future domains. To address these challenges, this paper proposes a mean teacher framework that strikes an appropriate Balance between Exploration and Exploitation (BEE) during the CTTA process. For the former challenge, we introduce a Multi-level Consistency Regularization (MCR) loss that aligns the intermediate features of the student and teacher models, accelerating adaptation to the current domain. For the latter challenge, we employ a Complementary Anchor Replay (CAR) mechanism to reuse historical checkpoints (anchors), recovering complementary knowledge for diverse domains. Experiments show that our method significantly outperforms state-of-the-art methods on several benchmarks, demonstrating its effectiveness for CTTA tasks.", "date": "2025-08-18", "categories": []}
{"id": "2508.13005", "title": "Empirical Evidences for the Effects of Feature Diversity in Open Set Recognition and Continual Learning", "abstract": "Open set recognition (OSR) and continual learning are two critical challenges in machine learning, focusing respectively on detecting novel classes at inference time and updating models to incorporate the new classes. While many recent approaches have addressed these problems, particularly OSR, by heuristically promoting feature diversity, few studies have directly examined the role that feature diversity plays in tackling them. In this work, we provide empirical evidence that enhancing feature diversity improves the recognition of open set samples. Moreover, increased feature diversity also facilitates both the retention of previously learned data and the integration of new data in continual learning. We hope our findings can inspire further research into both practical methods and theoretical understanding in these domains.", "date": "2025-08-18", "categories": []}
{"id": "2508.13006", "title": "Monte Carlo Functional Regularisation for Continual Learning", "abstract": "Continual learning (CL) is crucial for the adaptation of neural network models to new environments. Although outperforming weight-space regularisation approaches, the functional regularisation-based CL methods suffer from high computational costs and large linear approximation errors. In this work, we present a new functional regularisation CL framework, called MCFRCL, which approximates model prediction distributions by Monte Carlo (MC) sampling. Moreover, three continuous distributions are leveraged to capture the statistical characteristics of the MC samples via moment-based methods. Additionally, both the Wasserstein distance and the Kullback-Leibler (KL) distance are employed to construct the regularisation function. The proposed MCFRCL is evaluated against multiple benchmark methods on the MNIST and CIFAR datasets, with simulation results highlighting its effectiveness in both prediction accuracy and training efficiency.", "date": "2025-08-18", "categories": []}
{"id": "2508.11173", "title": "Exploring the Tradeoff Between Diversity and Discrimination for Continuous Category Discovery", "abstract": "Continuous category discovery (CCD) aims to automatically discover novel categories in continuously arriving unlabeled data. This is a challenging problem considering that there is no number of categories and labels in the newly arrived data, while also needing to mitigate catastrophic forgetting. Most CCD methods cannot handle the contradiction between novel class discovery and classification well. They are also prone to accumulate errors in the process of gradually discovering novel classes. Moreover, most of them use knowledge distillation and data replay to prevent forgetting, occupying more storage space. To address these limitations, we propose Independence-based Diversity and Orthogonality-based Discrimination (IDOD). IDOD mainly includes independent enrichment of diversity module, joint discovery of novelty module, and continuous increment by orthogonality module. In independent enrichment, the backbone is trained separately using contrastive loss to avoid it focusing only on features for classification. Joint discovery transforms multi-stage novel class discovery into single-stage, reducing error accumulation impact. Continuous increment by orthogonality module generates mutually orthogonal prototypes for classification and prevents forgetting with lower space overhead via representative representation replay. Experimental results show that on challenging fine-grained datasets, our method outperforms the state-of-the-art methods.", "date": "2025-08-15", "categories": []}
{"id": "2503.08064", "title": "Continual Learning for Multiple Modalities", "abstract": "Continual learning aims to learn knowledge of tasks observed in sequential time steps while mitigating the forgetting of previously learned knowledge. Existing methods were designed to learn a single modality (e.g., image) over time, which limits their applicability in scenarios involving multiple modalities. In this work, we propose a novel continual learning framework that accommodates multiple modalities (image, video, audio, depth, and text). We train a model to align various modalities with text, leveraging its rich semantic information. However, this increases the risk of forgetting previously learned knowledge, exacerbated by the differing input traits across tasks. To alleviate the overwriting of previous knowledge of modalities, we propose a framework that consolidates intra-modal knowledge while incorporating relevant inter-modal information. This is achieved by self-regulating shifts in learned representations to gradually integrating novel knowledge into the information retained across modalities. Simultaneously, it mitigates inter-modal interference by selectively integrating knowledge from previously encountered modalities based on their mutual relevance. Furthermore, we introduce a strategy to re-align modality embeddings, effectively addressing biased alignment between modalities. We evaluate the proposed method in a wide range of continual learning scenarios using multiple datasets with different modalities. Extensive experiments demonstrate that ours outperforms existing methods in the scenarios, regardless of whether the identity of the modality is given.", "date": "2025-03-11", "categories": []}
{"id": "2505.02469", "title": "Efficient Continual Learning in Keyword Spotting using Binary Neural Networks", "abstract": "Keyword spotting (KWS) is an essential function that enables interaction with ubiquitous smart devices. However, in resource-limited devices, KWS models are often static and can thus not adapt to new scenarios, such as added keywords. To overcome this problem, we propose a Continual Learning (CL) approach for KWS built on Binary Neural Networks (BNNs). The framework leverages the reduced computation and memory requirements of BNNs while incorporating techniques that enable the seamless integration of new keywords over time. This study evaluates seven CL techniques on a 16-class use case, reporting an accuracy exceeding 95% for a single additional keyword and up to 86% for four additional classes. Sensitivity to the amount of training samples in the CL phase, and differences in computational complexities are being evaluated. These evaluations demonstrate that batch-based algorithms are more sensitive to the CL dataset size, and that differences between the computational complexities are insignificant. These findings highlight the potential of developing an effective and computationally efficient technique for continuously integrating new keywords in KWS applications that is compatible with resource-constrained devices.", "date": "2025-05-05", "categories": []}
{"id": "2508.03189", "title": "Unifying Locality of KANs and Feature Drift Compensation Projection for Data-free Replay based Continual Face Forgery Detection", "abstract": "The rapid advancements in face forgery techniques necessitate that detectors continuously adapt to new forgery methods, thus situating face forgery detection within a continual learning paradigm. However, when detectors learn new forgery types, their performance on previous types often degrades rapidly, a phenomenon known as catastrophic forgetting. Kolmogorov-Arnold Networks (KANs) utilize locally plastic splines as their activation functions, enabling them to learn new tasks by modifying only local regions of the functions while leaving other areas unaffected. Therefore, they are naturally suitable for addressing catastrophic forgetting. However, KANs have two significant limitations: 1) the splines are ineffective for modeling high-dimensional images, while alternative activation functions that are suitable for images lack the essential property of locality; 2) in continual learning, when features from different domains overlap, the mapping of different domains to distinct curve regions always collapses due to repeated modifications of the same regions. In this paper, we propose a KAN-based Continual Face Forgery Detection (KAN-CFD) framework, which includes a Domain-Group KAN Detector (DG-KD) and a data-free replay Feature Separation strategy via KAN Drift Compensation Projection (FS-KDCP). DG-KD enables KANs to fit high-dimensional image inputs while preserving locality and local plasticity. FS-KDCP avoids the overlap of the KAN input spaces without using data from prior tasks. Experimental results demonstrate that the proposed method achieves superior performance while notably reducing forgetting.", "date": "2025-08-05", "categories": []}
{"id": "2508.08052", "title": "On Understanding of the Dynamics of Model Capacity in Continual Learning", "abstract": "The stability-plasticity dilemma, closely related to a neural network's (NN) capacity-its ability to represent tasks-is a fundamental challenge in continual learning (CL). Within this context, we introduce CL's effective model capacity (CLEMC) that characterizes the dynamic behavior of the stability-plasticity balance point. We develop a difference equation to model the evolution of the interplay between the NN, task data, and optimization procedure. We then leverage CLEMC to demonstrate that the effective capacity-and, by extension, the stability-plasticity balance point is inherently non-stationary. We show that regardless of the NN architecture or optimization method, a NN's ability to represent new tasks diminishes when incoming task distributions differ from previous ones. We conduct extensive experiments to support our theoretical findings, spanning a range of architectures-from small feedforward network and convolutional networks to medium-sized graph neural networks and transformer-based large language models with millions of parameters.", "date": "2025-08-11", "categories": []}
{"id": "2508.10432", "title": "CRISP: Contrastive Residual Injection and Semantic Prompting for Continual Video Instance Segmentation", "abstract": "Continual video instance segmentation demands both the plasticity to absorb new object categories and the stability to retain previously learned ones, all while preserving temporal consistency across frames. In this work, we introduce Contrastive Residual Injection and Semantic Prompting (CRISP), an earlier attempt tailored to address the instance-wise, category-wise, and task-wise confusion in continual video instance segmentation. For instance-wise learning, we model instance tracking and construct instance correlation loss, which emphasizes the correlation with the prior query space while strengthening the specificity of the current task query. For category-wise learning, we build an adaptive residual semantic prompt (ARSP) learning framework, which constructs a learnable semantic residual prompt pool generated by category text and uses an adjustive query-prompt matching mechanism to build a mapping relationship between the query of the current task and the semantic residual prompt. Meanwhile, a semantic consistency loss based on the contrastive learning is introduced to maintain semantic coherence between object queries and residual prompts during incremental training. For task-wise learning, to ensure the correlation at the inter-task level within the query space, we introduce a concise yet powerful initialization strategy for incremental prompts. Extensive experiments on YouTube-VIS-2019 and YouTube-VIS-2021 datasets demonstrate that CRISP significantly outperforms existing continual segmentation methods in the long-term continual video instance segmentation task, avoiding catastrophic forgetting and effectively improving segmentation and classification performance. The code is available at https://github.com/01upup10/CRISP.", "date": "2025-08-14", "categories": []}
{"id": "2508.10896", "title": "ESSENTIAL: Episodic and Semantic Memory Integration for Video Class-Incremental Learning", "abstract": "In this work, we tackle the problem of video classincremental learning (VCIL). Many existing VCIL methods mitigate catastrophic forgetting by rehearsal training with a few temporally dense samples stored in episodic memory, which is memory-inefficient. Alternatively, some methods store temporally sparse samples, sacrificing essential temporal information and thereby resulting in inferior performance. To address this trade-off between memory-efficiency and performance, we propose EpiSodic and SEmaNTIc memory integrAtion for video class-incremental Learning (ESSENTIAL). ESSENTIAL consists of episodic memory for storing temporally sparse features and semantic memory for storing general knowledge represented by learnable prompts. We introduce a novel memory retrieval (MR) module that integrates episodic memory and semantic prompts through cross-attention, enabling the retrieval of temporally dense features from temporally sparse features. We rigorously validate ESSENTIAL on diverse datasets: UCF-101, HMDB51, and Something-Something-V2 from the TCD benchmark and UCF-101, ActivityNet, and Kinetics-400 from the vCLIMB benchmark. Remarkably, with significantly reduced memory, ESSENTIAL achieves favorable performance on the benchmarks.", "date": "2025-08-14", "categories": []}
{"id": "2506.16884", "title": "The Importance of Being Lazy: Scaling Limits of Continual Learning", "abstract": "Despite recent efforts, neural networks still struggle to learn in non-stationary environments, and our understanding of catastrophic forgetting (CF) is far from complete. In this work, we perform a systematic study on the impact of model scale and the degree of feature learning in continual learning. We reconcile existing contradictory observations on scale in the literature, by differentiating between lazy and rich training regimes through a variable parameterization of the architecture. We show that increasing model width is only beneficial when it reduces the amount of feature learning, yielding more laziness. Using the framework of dynamical mean field theory, we then study the infinite width dynamics of the model in the feature learning regime and characterize CF, extending prior theoretical results limited to the lazy regime. We study the intricate relationship between feature learning, task non-stationarity, and forgetting, finding that high feature learning is only beneficial with highly similar tasks. We identify a transition modulated by task similarity where the model exits an effectively lazy regime with low forgetting to enter a rich regime with significant forgetting. Finally, our findings reveal that neural networks achieve optimal performance at a critical level of feature learning, which depends on task non-stationarity and transfers across model scales. This work provides a unified perspective on the role of scale and feature learning in continual learning.", "date": "2025-06-20", "categories": []}
{"id": "2508.09489", "title": "Large-Small Model Collaborative Framework for Federated Continual Learning", "abstract": "Continual learning (CL) for Foundation Models (FMs) is an essential yet underexplored challenge, especially in Federated Continual Learning (FCL), where each client learns from a private, evolving task stream under strict data and communication constraints. Despite their powerful generalization abilities, FMs often exhibit suboptimal performance on local downstream tasks, as they are unable to utilize private local data. Furthermore, enabling FMs to learn new tasks without forgetting prior knowledge is inherently a challenging problem, primarily due to their immense parameter count and high model complexity. In contrast, small models can be trained locally under resource-constrained conditions and benefit from more mature CL techniques. To bridge the gap between small models and FMs, we propose the first collaborative framework in FCL, where lightweight local models act as a dynamic bridge, continually adapting to new tasks while enhancing the utility of the large model. Two novel components are also included: Small Model Continual Fine-tuning is for preventing small models from temporal forgetting; One-by-One Distillation performs personalized fusion of heterogeneous local knowledge on the server. Experimental results demonstrate its superior performance, even when clients utilize heterogeneous small models.", "date": "2025-08-13", "categories": []}
{"id": "2508.09510", "title": "Enhancing Memory Recall in LLMs with Gauss-Tin: A Hybrid Instructional and Gaussian Replay Approach", "abstract": "Despite the significant advancements in Large Language Models (LLMs), catastrophic forgetting remains a substantial challenge, where models lose previously acquired knowledge upon learning new information. Continual learning (CL) strategies have emerged as a potential solution to this problem, with replay-based techniques demonstrating superior performance in preserving learned knowledge. In this context, we introduce Gauss-Tin, a novel approach that integrates the replay strategy with a Gaussian mixture model to enhance the quality of sample selection during training, supplemented by instructional guidance to facilitate the generation of past learning. This method aims to improve LLMs' retention capabilities by strategically reinforcing important past learnings while accommodating new information. Our experimental results indicate a promising 6\\% improvement in retention metrics over traditional methods, suggesting that Gauss-Tin is an effective strategy for mitigating catastrophic forgetting in LLMs. This study underscores the potential of hybrid models in enhancing the robustness and adaptability of LLMs in dynamic learning environments.", "date": "2025-08-13", "categories": []}
{"id": "2508.09785", "title": "DSS-Prompt: Dynamic-Static Synergistic Prompting for Few-Shot Class-Incremental Learning", "abstract": "Learning from large-scale pre-trained models with strong generalization ability has shown remarkable success in a wide range of downstream tasks recently, but it is still underexplored in the challenging few-shot class-incremental learning (FSCIL) task. It aims to continually learn new concepts from limited training samples without forgetting the old ones at the same time. In this paper, we introduce DSS-Prompt, a simple yet effective approach that transforms the pre-trained Vision Transformer with minimal modifications in the way of prompts into a strong FSCIL classifier. Concretely, we synergistically utilize two complementary types of prompts in each Transformer block: static prompts to bridge the domain gap between the pre-training and downstream datasets, thus enabling better adaption; and dynamic prompts to capture instance-aware semantics, thus enabling easy transfer from base to novel classes. Specially, to generate dynamic prompts, we leverage a pre-trained multi-modal model to extract input-related diverse semantics, thereby generating complementary input-aware prompts, and then adaptively adjust their importance across different layers. In this way, on top of the prompted visual embeddings, a simple prototype classifier can beat state-of-the-arts without further training on the incremental tasks. We conduct extensive experiments on four benchmarks to validate the effectiveness of our DSS-Prompt and show that it consistently achieves better performance than existing approaches on all datasets and can alleviate the catastrophic forgetting issue as well.", "date": "2025-08-13", "categories": []}
{"id": "2508.09974", "title": "Dynamic Mixture-of-Experts for Incremental Graph Learning", "abstract": "Graph incremental learning is a learning paradigm that aims to adapt trained models to continuously incremented graphs and data over time without the need for retraining on the full dataset. However, regular graph machine learning methods suffer from catastrophic forgetting when applied to incremental learning settings, where previously learned knowledge is overridden by new knowledge. Previous approaches have tried to address this by treating the previously trained model as an inseparable unit and using techniques to maintain old behaviors while learning new knowledge. These approaches, however, do not account for the fact that previously acquired knowledge at different timestamps contributes differently to learning new tasks. Some prior patterns can be transferred to help learn new data, while others may deviate from the new data distribution and be detrimental. To address this, we propose a dynamic mixture-of-experts (DyMoE) approach for incremental learning. Specifically, a DyMoE GNN layer adds new expert networks specialized in modeling the incoming data blocks. We design a customized regularization loss that utilizes data sequence information so existing experts can maintain their ability to solve old tasks while helping the new expert learn the new data effectively. As the number of data blocks grows over time, the computational cost of the full mixture-of-experts (MoE) model increases. To address this, we introduce a sparse MoE approach, where only the top-$k$ most relevant experts make predictions, significantly reducing the computation time. Our model achieved 4.92\\% relative accuracy increase compared to the best baselines on class incremental learning, showing the model's exceptional power.", "date": "2025-08-13", "categories": []}
{"id": "2503.07979", "title": "Achieving More with Less: Additive Prompt Tuning for Rehearsal-Free Class-Incremental Learning", "abstract": "Class-incremental learning (CIL) enables models to learn new classes progressively while preserving knowledge of previously learned ones. Recent advances in this field have shifted towards parameter-efficient fine-tuning techniques, with many approaches building upon the framework that maintains a pool of learnable prompts. Although effective, these methods introduce substantial computational overhead, primarily due to prompt pool querying and increased input sequence lengths from prompt concatenation. In this work, we present a novel prompt-based approach that addresses this limitation. Our method trains a single set of shared prompts across all tasks and, rather than concatenating prompts to the input, directly modifies the CLS token's attention computation by adding the prompts to it. This simple and lightweight design not only significantly reduces computational complexity-both in terms of inference costs and the number of trainable parameters-but also eliminates the need to optimize prompt lengths for different downstream tasks, offering a more efficient yet powerful solution for rehearsal-free class-incremental learning. Extensive experiments across a diverse range of CIL benchmarks demonstrate the effectiveness of our approach, highlighting its potential to establish a new prompt-based CIL paradigm. Furthermore, experiments on general recognition benchmarks beyond the CIL setting also show strong performance, positioning our method as a promising candidate for a general parameter-efficient fine-tuning approach.", "date": "2025-03-11", "categories": []}
{"id": "2508.08612", "title": "Hierarchical Visual Prompt Learning for Continual Video Instance Segmentation", "abstract": "Video instance segmentation (VIS) has gained significant attention for its capability in tracking and segmenting object instances across video frames. However, most of the existing VIS approaches unrealistically assume that the categories of object instances remain fixed over time. Moreover, they experience catastrophic forgetting of old classes when required to continuously learn object instances belonging to new categories. To resolve these challenges, we develop a novel Hierarchical Visual Prompt Learning (HVPL) model that overcomes catastrophic forgetting of previous categories from both frame-level and video-level perspectives. Specifically, to mitigate forgetting at the frame level, we devise a task-specific frame prompt and an orthogonal gradient correction (OGC) module. The OGC module helps the frame prompt encode task-specific global instance information for new classes in each individual frame by projecting its gradients onto the orthogonal feature space of old classes. Furthermore, to address forgetting at the video level, we design a task-specific video prompt and a video context decoder. This decoder first embeds structural inter-class relationships across frames into the frame prompt features, and then propagates task-specific global video contexts from the frame prompt features to the video prompt. Through rigorous comparisons, our HVPL model proves to be more effective than baseline approaches. The code is available at https://github.com/JiahuaDong/HVPL.", "date": "2025-08-12", "categories": []}
{"id": "2507.04106", "title": "Addressing The Devastating Effects Of Single-Task Data Poisoning In Exemplar-Free Continual Learning", "abstract": "Our research addresses the overlooked security concerns related to data poisoning in continual learning (CL). Data poisoning - the intentional manipulation of training data to affect the predictions of machine learning models - was recently shown to be a threat to CL training stability. While existing literature predominantly addresses scenario-dependent attacks, we propose to focus on a more simple and realistic single-task poison (STP) threats. In contrast to previously proposed poisoning settings, in STP adversaries lack knowledge and access to the model, as well as to both previous and future tasks. During an attack, they only have access to the current task within the data stream. Our study demonstrates that even within these stringent conditions, adversaries can compromise model performance using standard image corruptions. We show that STP attacks are able to strongly disrupt the whole continual training process: decreasing both the stability (its performance on past tasks) and plasticity (capacity to adapt to new tasks) of the algorithm. Finally, we propose a high-level defense framework for CL along with a poison task detection method based on task vectors. The code is available at https://github.com/stapaw/STP.git .", "date": "2025-07-05", "categories": []}
{"id": "2508.01713", "title": "Dynamic Robot-Assisted Surgery with Hierarchical Class-Incremental Semantic Segmentation", "abstract": "Robot-assisted surgeries rely on accurate and real-time scene understanding to safely guide surgical instruments. However, segmentation models trained on static datasets face key limitations when deployed in these dynamic and evolving surgical environments. Class-incremental semantic segmentation (CISS) allows models to continually adapt to new classes while avoiding catastrophic forgetting of prior knowledge, without training on previous data. In this work, we build upon the recently introduced Taxonomy-Oriented Poincar\\'e-regularized Incremental Class Segmentation (TOPICS) approach and propose an enhanced variant, termed TOPICS+, specifically tailored for robust segmentation of surgical scenes. Concretely, we incorporate the Dice loss into the hierarchical loss formulation to handle strong class imbalances, introduce hierarchical pseudo-labeling, and design tailored label taxonomies for robotic surgery environments. We also propose six novel CISS benchmarks designed for robotic surgery environments including multiple incremental steps and several semantic categories to emulate realistic class-incremental settings in surgical environments. In addition, we introduce a refined set of labels with more than 144 classes on the Syn-Mediverse synthetic dataset, hosted online as an evaluation benchmark. We make the code and trained models publicly available at http://topics.cs.uni-freiburg.de.", "date": "2025-08-03", "categories": []}
{"id": "2508.04368", "title": "Continual Multiple Instance Learning for Hematologic Disease Diagnosis", "abstract": "The dynamic environment of laboratories and clinics, with streams of data arriving on a daily basis, requires regular updates of trained machine learning models for consistent performance. Continual learning is supposed to help train models without catastrophic forgetting. However, state-of-the-art methods are ineffective for multiple instance learning (MIL), which is often used in single-cell-based hematologic disease diagnosis (e.g., leukemia detection). Here, we propose the first continual learning method tailored specifically to MIL. Our method is rehearsal-based over a selection of single instances from various bags. We use a combination of the instance attention score and distance from the bag mean and class mean vectors to carefully select which samples and instances to store in exemplary sets from previous tasks, preserving the diversity of the data. Using the real-world input of one month of data from a leukemia laboratory, we study the effectiveness of our approach in a class incremental scenario, comparing it to well-known continual learning methods. We show that our method considerably outperforms state-of-the-art methods, providing the first continual learning approach for MIL. This enables the adaptation of models to shifting data distributions over time, such as those caused by changes in disease occurrence or underlying genetic alterations.", "date": "2025-08-06", "categories": []}
{"id": "2508.07248", "title": "Prompt Tuning for Few-Shot Continual Learning Named Entity Recognition", "abstract": "Knowledge distillation has been successfully applied to Continual Learning Named Entity Recognition (CLNER) tasks, by using a teacher model trained on old-class data to distill old-class entities present in new-class data as a form of regularization, thereby avoiding catastrophic forgetting. However, in Few-Shot CLNER (FS-CLNER) tasks, the scarcity of new-class entities makes it difficult for the trained model to generalize during inference. More critically, the lack of old-class entity information hinders the distillation of old knowledge, causing the model to fall into what we refer to as the Few-Shot Distillation Dilemma. In this work, we address the above challenges through a prompt tuning paradigm and memory demonstration template strategy. Specifically, we designed an expandable Anchor words-oriented Prompt Tuning (APT) paradigm to bridge the gap between pre-training and fine-tuning, thereby enhancing performance in few-shot scenarios. Additionally, we incorporated Memory Demonstration Templates (MDT) into each training instance to provide replay samples from previous tasks, which not only avoids the Few-Shot Distillation Dilemma but also promotes in-context learning. Experiments show that our approach achieves competitive performances on FS-CLNER.", "date": "2025-08-10", "categories": []}
{"id": "2508.07419", "title": "Statistical Theory of Multi-stage Newton Iteration Algorithm for Online Continual Learning", "abstract": "We focus on the critical challenge of handling non-stationary data streams in online continual learning environments, where constrained storage capacity prevents complete retention of historical data, leading to catastrophic forgetting during sequential task training. To more effectively analyze and address the problem of catastrophic forgetting in continual learning, we propose a novel continual learning framework from a statistical perspective. Our approach incorporates random effects across all model parameters and allows the dimension of parameters to diverge to infinity, offering a general formulation for continual learning problems. To efficiently process streaming data, we develop a Multi-step Newton Iteration algorithm that significantly reduces computational costs in certain scenarios by alleviating the burden of matrix inversion. Theoretically, we derive the asymptotic normality of the estimator, enabling subsequent statistical inference. Comprehensive validation through synthetic data experiments and two real datasets analyses demonstrates the effectiveness of our proposed method.", "date": "2025-08-10", "categories": []}
{"id": "2508.07738", "title": "Separation and Collaboration: Two-Level Routing Grouped Mixture-of-Experts for Multi-Domain Continual Learning", "abstract": "Multi-Domain Continual Learning (MDCL) acquires knowledge from sequential tasks with shifting class sets and distribution. Despite the Parameter-Efficient Fine-Tuning (PEFT) methods can adapt for this dual heterogeneity, they still suffer from catastrophic forgetting and forward forgetting. To address these challenges, we propose a Two-Level Routing Grouped Mixture-of-Experts (TRGE) method. Firstly, TRGE dynamically expands the pre-trained CLIP model, assigning specific expert group for each task to mitigate catastrophic forgetting. With the number of experts continually grows in this process, TRGE maintains the static experts count within the group and introduces the intra-group router to alleviate routing overfitting caused by the increasing routing complexity. Meanwhile, we design an inter-group routing policy based on task identifiers and task prototype distance, which dynamically selects relevant expert groups and combines their outputs to enhance inter-task collaboration. Secondly, to get the correct task identifiers, we leverage Multimodal Large Language Models (MLLMs) which own powerful multimodal comprehension capabilities to generate semantic task descriptions and recognize the correct task identifier. Finally, to mitigate forward forgetting, we dynamically fuse outputs for unseen samples from the frozen CLIP model and TRGE adapter based on training progress, leveraging both pre-trained and learned knowledge. Through extensive experiments across various settings, our method outperforms other advanced methods with fewer trainable parameters.", "date": "2025-08-11", "categories": []}
{"id": "2508.08165", "title": "Integrating Task-Specific and Universal Adapters for Pre-Trained Model-based Class-Incremental Learning", "abstract": "Class-Incremental Learning (CIL) requires a learning system to continually learn new classes without forgetting. Existing pre-trained model-based CIL methods often freeze the pre-trained network and adapt to incremental tasks using additional lightweight modules such as adapters. However, incorrect module selection during inference hurts performance, and task-specific modules often overlook shared general knowledge, leading to errors on distinguishing between similar classes across tasks. To address the aforementioned challenges, we propose integrating Task-Specific and Universal Adapters (TUNA) in this paper. Specifically, we train task-specific adapters to capture the most crucial features relevant to their respective tasks and introduce an entropy-based selection mechanism to choose the most suitable adapter. Furthermore, we leverage an adapter fusion strategy to construct a universal adapter, which encodes the most discriminative features shared across tasks. We combine task-specific and universal adapter predictions to harness both specialized and general knowledge during inference. Extensive experiments on various benchmark datasets demonstrate the state-of-the-art performance of our approach. Code is available at: https://github.com/LAMDA-CL/ICCV2025-TUNA", "date": "2025-08-11", "categories": []}
{"id": "2404.02113", "title": "Position: Lifetime tuning is incompatible with continual reinforcement learning", "abstract": "In continual RL we want agents capable of never-ending learning, and yet our evaluation methodologies do not reflect this. The standard practice in RL is to assume unfettered access to the deployment environment for the full lifetime of the agent. For example, agent designers select the best performing hyperparameters in Atari by testing each for 200 million frames and then reporting results on 200 million frames. In this position paper, we argue and demonstrate the pitfalls of this inappropriate empirical methodology: lifetime tuning. We provide empirical evidence to support our position by testing DQN and SAC across several of continuing and non-stationary environments with two main findings: (1) lifetime tuning does not allow us to identify algorithms that work well for continual learning -- all algorithms equally succeed; (2) recently developed continual RL algorithms outperform standard non-continual algorithms when tuning is limited to a fraction of the agent's lifetime. The goal of this paper is to provide an explanation for why recent progress in continual RL has been mixed and motivate the development of empirical practices that better match the goals of continual RL.", "date": "2024-04-02", "categories": []}
{"id": "2508.05001", "title": "CRAM: Large-scale Video Continual Learning with Bootstrapped Compression", "abstract": "Continual learning (CL) promises to allow neural networks to learn from continuous streams of inputs, instead of IID (independent and identically distributed) sampling, which requires random access to a full dataset. This would allow for much smaller storage requirements and self-sufficiency of deployed systems that cope with natural distribution shifts, similarly to biological learning. We focus on video CL employing a rehearsal-based approach, which reinforces past samples from a memory buffer. We posit that part of the reason why practical video CL is challenging is the high memory requirements of video, further exacerbated by long-videos and continual streams, which are at odds with the common rehearsal-buffer size constraints. To address this, we propose to use compressed vision, i.e. store video codes (embeddings) instead of raw inputs, and train a video classifier by IID sampling from this rolling buffer. Training a video compressor online (so not depending on any pre-trained networks) means that it is also subject to catastrophic forgetting. We propose a scheme to deal with this forgetting by refreshing video codes, which requires careful decompression with a previous version of the network and recompression with a new one. We name our method Continually Refreshed Amodal Memory (CRAM). We expand current video CL benchmarks to large-scale settings, namely EpicKitchens-100 and Kinetics-700, storing thousands of relatively long videos in under 2 GB, and demonstrate empirically that our video CL method outperforms prior art with a significantly reduced memory footprint.", "date": "2025-08-07", "categories": []}
{"id": "2508.06280", "title": "A Study on Regularization-Based Continual Learning Methods for Indic ASR", "abstract": "Indias linguistic diversity poses significant challenges for developing inclusive Automatic Speech Recognition (ASR) systems. Traditional multilingual models, which require simultaneous access to all language data, are impractical due to the sequential arrival of data and privacy constraints. Continual Learning (CL) offers a solution by enabling models to learn new languages sequentially without catastrophically forgetting previously learned knowledge. This paper investigates CL for ASR on Indian languages using a subset of the IndicSUPERB benchmark. We employ a Conformer-based hybrid RNN-T/CTC model, initially pretrained on Hindi, which is then incrementally trained on eight additional Indian languages, for a total sequence of nine languages. We evaluate three prominent regularization- and distillation-based CL strategies: Elastic Weight Consolidation (EWC), Memory Aware Synapses (MAS), and Learning without Forgetting (LwF), selected for their suitability in no-replay, privacy-conscious scenarios. Performance is analyzed using Word Error Rate (WER) for both RNN-T and CTC paths on clean and noisy data, as well as knowledge retention via Backward Transfer. We also explore the impact of varying the number of training epochs (1, 2, 5, and 10) per task. Results, compared against naive fine-tuning, demonstrate CLs effectiveness in mitigating forgetting, making it a promising approach for scalable ASR in diverse Indian languages under realistic constraints. The code is available at: https://github.com/FrozenWolf-Cyber/Indic-CL-ASR", "date": "2025-08-08", "categories": []}
{"id": "2508.05094", "title": "Sculpting Margin Penalty: Intra-Task Adapter Merging and Classifier Calibration for Few-Shot Class-Incremental Learning", "abstract": "Real-world applications often face data privacy constraints and high acquisition costs, making the assumption of sufficient training data in incremental tasks unrealistic and leading to significant performance degradation in class-incremental learning. Forward-compatible learning, which prospectively prepares for future tasks during base task training, has emerged as a promising solution for Few-Shot Class-Incremental Learning (FSCIL). However, existing methods still struggle to balance base-class discriminability and new-class generalization. Moreover, limited access to original data during incremental tasks often results in ambiguous inter-class decision boundaries. To address these challenges, we propose SMP (Sculpting Margin Penalty), a novel FSCIL method that strategically integrates margin penalties at different stages within the parameter-efficient fine-tuning paradigm. Specifically, we introduce the Margin-aware Intra-task Adapter Merging (MIAM) mechanism for base task learning. MIAM trains two sets of low-rank adapters with distinct classification losses: one with a margin penalty to enhance base-class discriminability, and the other without margin constraints to promote generalization to future new classes. These adapters are then adaptively merged to improve forward compatibility. For incremental tasks, we propose a Margin Penalty-based Classifier Calibration (MPCC) strategy to refine decision boundaries by fine-tuning classifiers on all seen classes' embeddings with a margin penalty. Extensive experiments on CIFAR100, ImageNet-R, and CUB200 demonstrate that SMP achieves state-of-the-art performance in FSCIL while maintaining a better balance between base and new classes.", "date": "2025-08-07", "categories": []}
{"id": "2508.05316", "title": "Divide-and-Conquer for Enhancing Unlabeled Learning, Stability, and Plasticity in Semi-supervised Continual Learning", "abstract": "Semi-supervised continual learning (SSCL) seeks to leverage both labeled and unlabeled data in a sequential learning setup, aiming to reduce annotation costs while managing continual data arrival. SSCL introduces complex challenges, including ensuring effective unlabeled learning (UL), while balancing memory stability (MS) and learning plasticity (LP). Previous SSCL efforts have typically focused on isolated aspects of the three, while this work presents USP, a divide-and-conquer framework designed to synergistically enhance these three aspects: (1) Feature Space Reservation (FSR) strategy for LP, which constructs reserved feature locations for future classes by shaping old classes into an equiangular tight frame; (2) Divide-and-Conquer Pseudo-labeling (DCP) approach for UL, which assigns reliable pseudo-labels across both high- and low-confidence unlabeled data; and (3) Class-mean-anchored Unlabeled Distillation (CUD) for MS, which reuses DCP's outputs to anchor unlabeled data to stable class means for distillation to prevent forgetting. Comprehensive evaluations show USP outperforms prior SSCL methods, with gains up to 5.94% in the last accuracy, validating its effectiveness. The code is available at https://github.com/NJUyued/USP4SSCL.", "date": "2025-08-07", "categories": []}
{"id": "2411.17475", "title": "COBRA: A Continual Learning Approach to Vision-Brain Understanding", "abstract": "Vision-Brain Understanding (VBU) aims to extract visual information perceived by humans from brain activity recorded through functional Magnetic Resonance Imaging (fMRI). Despite notable advancements in recent years, existing studies in VBU continue to face the challenge of catastrophic forgetting, where models lose knowledge from prior subjects as they adapt to new ones. Addressing continual learning in this field is, therefore, essential. This paper introduces a novel framework called Continual Learning for Vision-Brain (COBRA) to address continual learning in VBU. Our approach includes three novel modules: a Subject Commonality (SC) module, a Prompt-based Subject Specific (PSS) module, and a transformer-based module for fMRI, denoted as MRIFormer module. The SC module captures shared vision-brain patterns across subjects, preserving this knowledge as the model encounters new subjects, thereby reducing the impact of catastrophic forgetting. On the other hand, the PSS module learns unique vision-brain patterns specific to each subject. Finally, the MRIFormer module contains a transformer encoder and decoder that learns the fMRI features for VBU from common and specific patterns. In a continual learning setup, COBRA is trained in new PSS and MRIFormer modules for new subjects, leaving the modules of previous subjects unaffected. As a result, COBRA effectively addresses catastrophic forgetting and achieves state-of-the-art performance in both continual learning and vision-brain reconstruction tasks, surpassing previous methods.", "date": "2024-11-25", "categories": []}
{"id": "2507.09031", "title": "Confounder-Free Continual Learning via Recursive Feature Normalization", "abstract": "Confounders are extraneous variables that affect both the input and the target, resulting in spurious correlations and biased predictions. There are recent advances in dealing with or removing confounders in traditional models, such as metadata normalization (MDN), where the distribution of the learned features is adjusted based on the study confounders. However, in the context of continual learning, where a model learns continuously from new data over time without forgetting, learning feature representations that are invariant to confounders remains a significant challenge. To remove their influence from intermediate feature representations, we introduce the Recursive MDN (R-MDN) layer, which can be integrated into any deep learning architecture, including vision transformers, and at any model stage. R-MDN performs statistical regression via the recursive least squares algorithm to maintain and continually update an internal model state with respect to changing distributions of data and confounding variables. Our experiments demonstrate that R-MDN promotes equitable predictions across population groups, both within static learning and across different stages of continual learning, by reducing catastrophic forgetting caused by confounder effects changing over time.", "date": "2025-07-11", "categories": []}
{"id": "2508.04267", "title": "Revisiting Continual Semantic Segmentation with Pre-trained Vision Models", "abstract": "Continual Semantic Segmentation (CSS) seeks to incrementally learn to segment novel classes while preserving knowledge of previously encountered ones. Recent advancements in CSS have been largely driven by the adoption of Pre-trained Vision Models (PVMs) as backbones. Among existing strategies, Direct Fine-Tuning (DFT), which sequentially fine-tunes the model across classes, remains the most straightforward approach. Prior work often regards DFT as a performance lower bound due to its presumed vulnerability to severe catastrophic forgetting, leading to the development of numerous complex mitigation techniques. However, we contend that this prevailing assumption is flawed. In this paper, we systematically revisit forgetting in DFT across two standard benchmarks, Pascal VOC 2012 and ADE20K, under eight CSS settings using two representative PVM backbones: ResNet101 and Swin-B. Through a detailed probing analysis, our findings reveal that existing methods significantly underestimate the inherent anti-forgetting capabilities of PVMs. Even under DFT, PVMs retain previously learned knowledge with minimal forgetting. Further investigation of the feature space indicates that the observed forgetting primarily arises from the classifier's drift away from the PVM, rather than from degradation of the backbone representations. Based on this insight, we propose DFT*, a simple yet effective enhancement to DFT that incorporates strategies such as freezing the PVM backbone and previously learned classifiers, as well as pre-allocating future classifiers. Extensive experiments show that DFT* consistently achieves competitive or superior performance compared to sixteen state-of-the-art CSS methods, while requiring substantially fewer trainable parameters and less training time.", "date": "2025-08-06", "categories": []}
{"id": "2508.04676", "title": "GeRe: Towards Efficient Anti-Forgetting in Continual Learning of LLM via General Samples Replay", "abstract": "The continual learning capability of large language models (LLMs) is crucial for advancing artificial general intelligence. However, continual fine-tuning LLMs across various domains often suffers from catastrophic forgetting, characterized by: 1) significant forgetting of their general capabilities, and 2) sharp performance declines in previously learned tasks. To simultaneously address both issues in a simple yet stable manner, we propose General Sample Replay (GeRe), a framework that use usual pretraining texts for efficient anti-forgetting. Beyond revisiting the most prevalent replay-based practices under GeRe, we further leverage neural states to introduce a enhanced activation states constrained optimization method using threshold-based margin (TM) loss, which maintains activation state consistency during replay learning. We are the first to validate that a small, fixed set of pre-collected general replay samples is sufficient to resolve both concerns--retaining general capabilities while promoting overall performance across sequential tasks. Indeed, the former can inherently facilitate the latter. Through controlled experiments, we systematically compare TM with different replay strategies under the GeRe framework, including vanilla label fitting, logit imitation via KL divergence and feature imitation via L1/L2 losses. Results demonstrate that TM consistently improves performance and exhibits better robustness. Our work paves the way for efficient replay of LLMs for the future. Our code and data are available at https://github.com/Qznan/GeRe.", "date": "2025-08-06", "categories": []}
{"id": "2411.17471", "title": "Learning New Concepts, Remembering the Old: Continual Learning for Multimodal Concept Bottleneck Models", "abstract": "Concept Bottleneck Models (CBMs) enhance the interpretability of AI systems, particularly by bridging visual input with human-understandable concepts, effectively acting as a form of multimodal interpretability model. However, existing CBMs typically assume static datasets, which fundamentally limits their adaptability to real-world, continuously evolving multimodal data streams. To address this, we define a novel continual learning task for CBMs: simultaneously handling concept-incremental and class-incremental learning. This task requires models to continuously acquire new concepts (often representing cross-modal attributes) and classes while robustly preserving previously learned knowledge. To tackle this challenging problem, we propose CONceptual Continual Incremental Learning (CONCIL), a novel framework that fundamentally re-imagines concept and decision layer updates as linear regression problems. This reformulation eliminates the need for gradient-based optimization, thereby effectively preventing catastrophic forgetting. Crucially, CONCIL relies solely on recursive matrix operations, rendering it highly computationally efficient and well-suited for real-time and large-scale multimodal data applications. Experimental results compellingly demonstrate that CONCIL achieves \"absolute knowledge memory\" and significantly surpasses the performance of traditional CBM methods in both concept- and class-incremental settings, thus establishing a new paradigm for continual learning in CBMs, particularly valuable for dynamic multimodal understanding.", "date": "2024-11-25", "categories": []}
{"id": "2502.09152", "title": "Vertical Federated Continual Learning via Evolving Prototype Knowledge", "abstract": "Vertical Federated Learning (VFL) has garnered significant attention as a privacy-preserving machine learning framework for sample-aligned feature federation. However, traditional VFL approaches do not address the challenges of class and feature continual learning, resulting in catastrophic forgetting of knowledge from previous tasks. To address the above challenge, we propose a novel vertical federated continual learning method, named Vertical Federated Continual Learning via Evolving Prototype Knowledge (V-LETO), which primarily facilitates the transfer of knowledge from previous tasks through the evolution of prototypes. Specifically, we propose an evolving prototype knowledge method, enabling the global model to retain both previous and current task knowledge. Furthermore, we introduce a model optimization technique that mitigates the forgetting of previous task knowledge by restricting updates to specific parameters of the local model, thereby enhancing overall performance. Extensive experiments conducted in both CIL and FIL settings demonstrate that our method, V-LETO, outperforms the other state-of-the-art methods. For example, our method outperforms the state-of-the-art method by 10.39% and 35.15% for CIL and FIL tasks, respectively. Our code is available at https://anonymous.4open.science/r/V-LETO-0108/README.md.", "date": "2025-02-13", "categories": []}
{"id": "2508.03259", "title": "Exploring Stability-Plasticity Trade-offs for Continual Named Entity Recognition", "abstract": "Continual Named Entity Recognition (CNER) is an evolving field that focuses on sequentially updating an existing model to incorporate new entity types. Previous CNER methods primarily utilize Knowledge Distillation (KD) to preserve prior knowledge and overcome catastrophic forgetting, strictly ensuring that the representations of old and new models remain consistent. Consequently, they often impart the model with excessive stability (i.e., retention of old knowledge) but limited plasticity (i.e., acquisition of new knowledge). To address this issue, we propose a Stability-Plasticity Trade-off (SPT) method for CNER that balances these aspects from both representation and weight perspectives. From the representation perspective, we introduce a pooling operation into the original KD, permitting a level of plasticity by consolidating representation dimensions. From the weight perspective, we dynamically merge the weights of old and new models, strengthening old knowledge while maintaining new knowledge. During this fusion, we implement a weight-guided selective mechanism to prioritize significant weights. Moreover, we develop a confidence-based pseudo-labeling approach for the current non-entity type, which predicts entity types using the old model to handle the semantic shift of the non-entity type, a challenge specific to CNER that has largely been ignored by previous methods. Extensive experiments across ten CNER settings on three benchmark datasets demonstrate that our SPT method surpasses previous CNER approaches, highlighting its effectiveness in achieving a suitable stability-plasticity trade-off.", "date": "2025-08-05", "categories": []}
{"id": "2508.03375", "title": "GaitAdapt: Continual Learning for Evolving Gait Recognition", "abstract": "Current gait recognition methodologies generally necessitate retraining when encountering new datasets. Nevertheless, retrained models frequently encounter difficulties in preserving knowledge from previous datasets, leading to a significant decline in performance on earlier test sets. To tackle these challenges, we present a continual gait recognition task, termed GaitAdapt, which supports the progressive enhancement of gait recognition capabilities over time and is systematically categorized according to various evaluation scenarios. Additionally, we propose GaitAdapter, a non-replay continual learning approach for gait recognition. This approach integrates the GaitPartition Adaptive Knowledge (GPAK) module, employing graph neural networks to aggregate common gait patterns from current data into a repository constructed from graph vectors. Subsequently, this repository is used to improve the discriminability of gait features in new tasks, thereby enhancing the model's ability to effectively recognize gait patterns. We also introduce a Euclidean Distance Stability Method (EDSN) based on negative pairs, which ensures that newly added gait samples from different classes maintain similar relative spatial distributions across both previous and current gait tasks, thereby alleviating the impact of task changes on the distinguishability of original domain features. Extensive evaluations demonstrate that GaitAdapter effectively retains gait knowledge acquired from diverse tasks, exhibiting markedly superior discriminative capability compared to alternative methods.", "date": "2025-08-05", "categories": []}
{"id": "2508.03571", "title": "Tackling Distribution Shift in LLM via KILO: Knowledge-Instructed Learning for Continual Adaptation", "abstract": "Large Language Models (LLMs) often suffer from performance degradation when faced with domain shifts, primarily due to catastrophic forgetting. In this work, we propose KILO (Knowledge-Instructed Learning for Continual Adaptation), a novel continual learning framework that integrates dynamic knowledge graphs with instruction tuning. By leveraging retrieved domain-specific knowledge as guidance during training, KILO enhances both adaptability to new domains and retention of previously acquired knowledge. We pretrain our model on WikiText-103 and evaluate sequential adaptation across four diverse target domains: BioASQ, SciQ, TweetEval, and MIND. Our experiments demonstrate that KILO consistently outperforms strong baselines, including continual fine-tuning, ERNIE 2.0, and CPT, in terms of backward transfer, forward transfer, F1 score, retention rate, and training efficiency. These results highlight the effectiveness of combining structured knowledge retrieval and instruction prompting to overcome domain shift challenges in continual learning scenarios.", "date": "2025-08-05", "categories": []}
{"id": "2312.09481", "title": "Continual Adversarial Defense", "abstract": "In response to the rapidly evolving nature of adversarial attacks against visual classifiers, numerous defenses have been proposed to generalize against as many known attacks as possible. However, designing a defense method that generalizes to all types of attacks is unrealistic, as the environment in which the defense system operates is dynamic. Over time, new attacks inevitably emerge that exploit the vulnerabilities of existing defenses and bypass them. Therefore, we propose a continual defense strategy under a practical threat model and, for the first time, introduce the Continual Adversarial Defense (CAD) framework. CAD continuously collects adversarial data online and adapts to evolving attack sequences, while adhering to four practical principles: (1) continual adaptation to new attacks without catastrophic forgetting, (2) few-shot adaptation, (3) memory-efficient adaptation, and (4) high classification accuracy on both clean and adversarial data. We explore and integrate cutting-edge techniques from continual learning, few-shot learning, and ensemble learning to fulfill the principles. Extensive experiments validate the effectiveness of our approach against multi-stage adversarial attacks and demonstrate significant improvements over a wide range of baseline methods. We further observe that CAD's defense performance tends to saturate as the number of attacks increases, indicating its potential as a persistent defense once adapted to a sufficiently diverse set of attacks. Our research sheds light on a brand-new paradigm for continual defense adaptation against dynamic and evolving attacks.", "date": "2023-12-15", "categories": []}
{"id": "2506.21260", "title": "DuET: Dual Incremental Object Detection via Exemplar-Free Task Arithmetic", "abstract": "Real-world object detection systems, such as those in autonomous driving and surveillance, must continuously learn new object categories and simultaneously adapt to changing environmental conditions. Existing approaches, Class Incremental Object Detection (CIOD) and Domain Incremental Object Detection (DIOD) only address one aspect of this challenge. CIOD struggles in unseen domains, while DIOD suffers from catastrophic forgetting when learning new classes, limiting their real-world applicability. To overcome these limitations, we introduce Dual Incremental Object Detection (DuIOD), a more practical setting that simultaneously handles class and domain shifts in an exemplar-free manner. We propose DuET, a Task Arithmetic-based model merging framework that enables stable incremental learning while mitigating sign conflicts through a novel Directional Consistency Loss. Unlike prior methods, DuET is detector-agnostic, allowing models like YOLO11 and RT-DETR to function as real-time incremental object detectors. To comprehensively evaluate both retention and adaptation, we introduce the Retention-Adaptability Index (RAI), which combines the Average Retention Index (Avg RI) for catastrophic forgetting and the Average Generalization Index for domain adaptability into a common ground. Extensive experiments on the Pascal Series and Diverse Weather Series demonstrate DuET's effectiveness, achieving a +13.12% RAI improvement while preserving 89.3% Avg RI on the Pascal Series (4 tasks), as well as a +11.39% RAI improvement with 88.57% Avg RI on the Diverse Weather Series (3 tasks), outperforming existing methods.", "date": "2025-06-26", "categories": []}
{"id": "2508.01167", "title": "T2S: Tokenized Skill Scaling for Lifelong Imitation Learning", "abstract": "The main challenge in lifelong imitation learning lies in the balance between mitigating catastrophic forgetting of previous skills while maintaining sufficient capacity for acquiring new ones. However, current approaches typically address these aspects in isolation, overlooking their internal correlation in lifelong skill acquisition. We address this limitation with a unified framework named Tokenized Skill Scaling (T2S). Specifically, by tokenizing the model parameters, the linear parameter mapping of the traditional transformer is transformed into cross-attention between input and learnable tokens, thereby enhancing model scalability through the easy extension of new tokens. Additionally, we introduce language-guided skill scaling to transfer knowledge across tasks efficiently and avoid linearly growing parameters. Extensive experiments across diverse tasks demonstrate that T2S: 1) effectively prevents catastrophic forgetting (achieving an average NBT of 1.0% across the three LIBERO task suites), 2) excels in new skill scaling with minimal increases in trainable parameters (needing only 8.0% trainable tokens in an average of lifelong tasks), and 3) enables efficient knowledge transfer between tasks (achieving an average FWT of 77.7% across the three LIBERO task suites), offering a promising solution for lifelong imitation learning.", "date": "2025-08-02", "categories": []}
{"id": "2508.01598", "title": "Drift-aware Collaborative Assistance Mixture of Experts for Heterogeneous Multistream Learning", "abstract": "Learning from multiple data streams in real-world scenarios is fundamentally challenging due to intrinsic heterogeneity and unpredictable concept drifts. Existing methods typically assume homogeneous streams and employ static architectures with indiscriminate knowledge fusion, limiting generalizability in complex dynamic environments. To tackle this gap, we propose CAMEL, a dynamic \\textbf{C}ollaborative \\textbf{A}ssistance \\textbf{M}ixture of \\textbf{E}xperts \\textbf{L}earning framework. It addresses heterogeneity by assigning each stream an independent system with a dedicated feature extractor and task-specific head. Meanwhile, a dynamic pool of specialized private experts captures stream-specific idiosyncratic patterns. Crucially, collaboration across these heterogeneous streams is enabled by a dedicated assistance expert. This expert employs a multi-head attention mechanism to distill and integrate relevant context autonomously from all other concurrent streams. It facilitates targeted knowledge transfer while inherently mitigating negative transfer from irrelevant sources. Furthermore, we propose an Autonomous Expert Tuner (AET) strategy, which dynamically manages expert lifecycles in response to drift. It instantiates new experts for emerging concepts (freezing prior ones to prevent catastrophic forgetting) and prunes obsolete ones. This expert-level plasticity provides a robust and efficient mechanism for online model capacity adaptation. Extensive experiments demonstrate CAMEL's superior generalizability across diverse multistreams and exceptional resilience against complex concept drifts.", "date": "2025-08-03", "categories": []}
{"id": "2508.01908", "title": "Revisiting Replay and Gradient Alignment for Continual Pre-Training of Large Language Models", "abstract": "Training large language models (LLMs) typically involves pre-training on massive corpora, only to restart the process entirely when new data becomes available. A more efficient and resource-conserving approach would be continual pre-training, where models are updated with new data rather than retraining from scratch. However, the introduction of new data often causes distribution shifts, leading to performance degradation on previously learned tasks. In this paper, we take a deeper look at two popular proposals for addressing this distribution shift within the continual learning literature: experience replay and gradient alignment. We consider continual pre-training of models within the Llama family of architectures at a large scale across languages with 100 billion tokens of training data in each language, finding that both replay and gradient alignment lead to more stable learning without forgetting. This conclusion holds both as we vary the model scale and as we vary the number and diversity of tasks. Moreover, we are the first to demonstrate the effectiveness of gradient alignment techniques in the context of LLM pre-training and propose an efficient implementation of meta-experience replay (MER) that imbues experience replay with the benefits of gradient alignment despite negligible compute and memory overhead. Our scaling analysis across model sizes and replay rates indicates that small rates of replaying old examples are definitely a more valuable use of compute than investing in model size, but that it is more compute efficient to scale the size of the model than invest in high rates of replaying old examples.", "date": "2025-08-03", "categories": []}
{"id": "2508.00260", "title": "Instruction-Grounded Visual Projectors for Continual Learning of Generative Vision-Language Models", "abstract": "Continual learning enables pre-trained generative vision-language models (VLMs) to incorporate knowledge from new tasks without retraining data from previous ones. Recent methods update a visual projector to translate visual information for new tasks, connecting pre-trained vision encoders with large language models. However, such adjustments may cause the models to prioritize visual inputs over language instructions, particularly learning tasks with repetitive types of textual instructions. To address the neglect of language instructions, we propose a novel framework that grounds the translation of visual information on instructions for language models. We introduce a mixture of visual projectors, each serving as a specialized visual-to-language translation expert based on the given instruction context to adapt to new tasks. To avoid using experts for irrelevant instruction contexts, we propose an expert recommendation strategy that reuses experts for tasks similar to those previously learned. Additionally, we introduce expert pruning to alleviate interference from the use of experts that cumulatively activated in previous tasks. Extensive experiments on diverse vision-language tasks demonstrate that our method outperforms existing continual learning approaches by generating instruction-following responses.", "date": "2025-08-01", "categories": []}
{"id": "2311.18266", "title": "Prompt-Based Exemplar Super-Compression and Regeneration for Class-Incremental Learning", "abstract": "Replay-based methods in class-incremental learning (CIL) have attained remarkable success. Despite their effectiveness, the inherent memory restriction results in saving a limited number of exemplars with poor diversity. In this paper, we introduce PESCR, a novel approach that substantially increases the quantity and enhances the diversity of exemplars based on a pre-trained general-purpose diffusion model, without fine-tuning it on target datasets or storing it in the memory buffer. Images are compressed into visual and textual prompts, which are saved instead of the original images, decreasing memory consumption by a factor of 24. In subsequent phases, diverse exemplars are regenerated by the diffusion model. We further propose partial compression and diffusion-based data augmentation to minimize the domain gap between generated exemplars and real images. PESCR significantly improves CIL performance across multiple benchmarks, e.g., 3.2% above the previous state-of-the-art on ImageNet-100.", "date": "2023-11-30", "categories": []}
{"id": "2502.06210", "title": "Achieving Deep Continual Learning via Evolution", "abstract": "Deep neural networks, despite their remarkable success, remain fundamentally limited in their ability to perform Continual Learning (CL). While most current methods aim to enhance the capabilities of a single model, Inspired by the collective learning mechanisms of human populations, we introduce Evolving Continual Learning (ECL), a framework that maintains and evolves a diverse population of neural network models. ECL continually searches for an optimal architecture for each introduced incremental task. This tailored model is trained on the corresponding task and archived as a specialized expert, contributing to a growing collection of skills. This approach inherently resolves the core CL challenges: stability is achieved through the isolation of expert models, while plasticity is greatly enhanced by evolving unique, task-specific architectures. Experimental results demonstrate that ECL significantly outperforms state-of-the-art individual-level CL methods. By shifting the focus from individual adaptation to collective evolution, ECL presents a novel path toward AI systems capable of CL.", "date": "2025-02-10", "categories": []}
{"id": "2507.23237", "title": "Ambiguity-Guided Learnable Distribution Calibration for Semi-Supervised Few-Shot Class-Incremental Learning", "abstract": "Few-Shot Class-Incremental Learning (FSCIL) focuses on models learning new concepts from limited data while retaining knowledge of previous classes. Recently, many studies have started to leverage unlabeled samples to assist models in learning from few-shot samples, giving rise to the field of Semi-supervised Few-shot Class-Incremental Learning (Semi-FSCIL). However, these studies often assume that the source of unlabeled data is only confined to novel classes of the current session, which presents a narrow perspective and cannot align well with practical scenarios. To better reflect real-world scenarios, we redefine Semi-FSCIL as Generalized Semi-FSCIL (GSemi-FSCIL) by incorporating both base and all the ever-seen novel classes in the unlabeled set. This change in the composition of unlabeled samples poses a new challenge for existing methods, as they struggle to distinguish between unlabeled samples from base and novel classes. To address this issue, we propose an Ambiguity-guided Learnable Distribution Calibration (ALDC) strategy. ALDC dynamically uses abundant base samples to correct biased feature distributions for few-shot novel classes. Experiments on three benchmark datasets show that our method outperforms existing works, setting new state-of-the-art results.", "date": "2025-07-31", "categories": []}
{"id": "2507.23311", "title": "Forgetting of task-specific knowledge in model merging-based continual learning", "abstract": "This paper investigates the linear merging of models in the context of continual learning (CL). Using controlled visual cues in computer vision experiments, we demonstrate that merging largely preserves or enhances shared knowledge, while unshared task-specific knowledge rapidly degrades. We further find that merging models from an incremental training process consistently outperforms merging models trained in parallel.", "date": "2025-07-31", "categories": []}
{"id": "2507.22553", "title": "RainbowPrompt: Diversity-Enhanced Prompt-Evolving for Continual Learning", "abstract": "Prompt-based continual learning provides a rehearsal-free solution by tuning small sets of parameters while keeping pre-trained models frozen. To meet the complex demands of sequential tasks, it is crucial to integrate task-specific knowledge within prompts effectively. However, existing works rely on either fixed learned prompts (i.e., prompts whose representations remain unchanged during new task learning) or on prompts generated from an entangled task-shared space, limiting the representational diversity of the integrated prompt. To address this issue, we propose a novel prompt-evolving mechanism to adaptively aggregate base prompts (i.e., task-specific prompts) into a unified prompt while ensuring diversity. By transforming and aligning base prompts, both previously learned and newly introduced, our approach continuously evolves accumulated knowledge to facilitate learning new tasks. We further introduce a learnable probabilistic gate that adaptively determines which layers to activate during the evolution process. We validate our method on image classification and video action recognition tasks in class-incremental learning, achieving average gains of 9.07% and 7.40% over existing methods across all scenarios.", "date": "2025-07-30", "categories": []}
{"id": "2407.08279", "title": "Continually Learn to Map Visual Concepts to Large Language Models in Resource-constrained Environments", "abstract": "Learning continually from a stream of non-i.i.d. data is an open challenge in deep learning, even more so when working in resource-constrained environments such as embedded devices. Visual models that are continually updated through supervised learning are often prone to overfitting, catastrophic forgetting, and biased representations. On the other hand, large language models contain knowledge about multiple concepts and their relations, which can foster a more robust, informed and coherent learning process. This work proposes Continual Visual Mapping (CVM), an approach that continually ground vision representations to a knowledge space extracted from a fixed Language model. Specifically, CVM continually trains a small and efficient visual model to map its representations into a conceptual space established by a fixed Large Language Model. Due to their smaller nature, CVM can be used when directly adapting large visual pre-trained models is unfeasible due to computational or data constraints. CVM overcome state-of-the-art continual learning methods on five benchmarks and offers a promising avenue for addressing generalization capabilities in continual learning, even in computationally constrained devices.", "date": "2024-07-11", "categories": []}
{"id": "2409.13864", "title": "Persistent Backdoor Attacks in Continual Learning", "abstract": "Backdoor attacks pose a significant threat to neural networks, enabling adversaries to manipulate model outputs on specific inputs, often with devastating consequences, especially in critical applications. While backdoor attacks have been studied in various contexts, little attention has been given to their practicality and persistence in continual learning, particularly in understanding how the continual updates to model parameters, as new data distributions are learned and integrated, impact the effectiveness of these attacks over time. To address this gap, we introduce two persistent backdoor attacks-Blind Task Backdoor and Latent Task Backdoor-each leveraging minimal adversarial influence. Our blind task backdoor subtly alters the loss computation without direct control over the training process, while the latent task backdoor influences only a single task's training, with all other tasks trained benignly. We evaluate these attacks under various configurations, demonstrating their efficacy with static, dynamic, physical, and semantic triggers. Our results show that both attacks consistently achieve high success rates across different continual learning algorithms, while effectively evading state-of-the-art defenses, such as SentiNet and I-BAU.", "date": "2024-09-20", "categories": []}
{"id": "2507.13568", "title": "LoRA-Loop: Closing the Synthetic Replay Cycle for Continual VLM Learning", "abstract": "Continual learning for vision-language models has achieved remarkable performance through synthetic replay, where samples are generated using Stable Diffusion to regularize during finetuning and retain knowledge. However, real-world downstream applications often exhibit domain-specific nuances and fine-grained semantics not captured by generators, causing synthetic-replay methods to produce misaligned samples that misguide finetuning and undermine retention of prior knowledge. In this work, we propose a LoRA-enhanced synthetic-replay framework that injects task-specific low-rank adapters into a frozen Stable Diffusion model, efficiently capturing each new task's unique visual and semantic patterns. Specifically, we introduce a two-stage, confidence-based sample selection: we first rank real task data by post-finetuning VLM confidence to focus LoRA finetuning on the most representative examples, then generate synthetic samples and again select them by confidence for distillation. Our approach integrates seamlessly with existing replay pipelines-simply swap in the adapted generator to boost replay fidelity. Extensive experiments on the Multi-domain Task Incremental Learning (MTIL) benchmark show that our method outperforms previous synthetic-replay techniques, achieving an optimal balance among plasticity, stability, and zero-shot capability. These results demonstrate the effectiveness of generator adaptation via LoRA for robust continual learning in VLMs.", "date": "2025-07-17", "categories": []}
{"id": "2507.21479", "title": "Capacity-Constrained Continual Learning", "abstract": "Any agents we can possibly build are subject to capacity constraints, as memory and compute resources are inherently finite. However, comparatively little attention has been dedicated to understanding how agents with limited capacity should allocate their resources for optimal performance. The goal of this paper is to shed some light on this question by studying a simple yet relevant continual learning problem: the capacity-constrained linear-quadratic-Gaussian (LQG) sequential prediction problem. We derive a solution to this problem under appropriate technical conditions. Moreover, for problems that can be decomposed into a set of sub-problems, we also demonstrate how to optimally allocate capacity across these sub-problems in the steady state. We view the results of this paper as a first step in the systematic theoretical study of learning under capacity constraints.", "date": "2025-07-29", "categories": []}
{"id": "2507.21588", "title": "Progressive Homeostatic and Plastic Prompt Tuning for Audio-Visual Multi-Task Incremental Learning", "abstract": "Audio-visual multi-task incremental learning aims to continuously learn from multiple audio-visual tasks without the need for joint training on all tasks. The challenge of the problem is how to preserve the old task knowledge while facilitating the learning of new task with previous experiences. To address these challenges, we introduce a three-stage Progressive Homeostatic and Plastic audio-visual prompt (PHP) method. In the shallow phase, we design the task-shared modality aggregating adapter to foster cross-task and cross-modal audio-visual representation learning to enhance shared understanding between tasks. In the middle phase, we propose the task-specific modality-shared dynamic generating adapter, which constructs prompts that are tailored to individual tasks while remaining general across modalities, which balances the models ability to retain knowledge against forgetting with its potential for versatile multi-task transferability. In the deep phase, we introduce the task-specific modality-independent prompts to further refine the understand ability by targeting individual information for each task and modality. By incorporating these three phases, PHP retains task-specific prompts while adapting shared parameters for new tasks to effectively balance knowledge sharing and specificity. Our method achieves SOTA performance in different orders of four tasks (AVE, AVVP, AVS and AVQA). Our code can be available at https://github.com/ENJOY-Yin-jiong/PHP.", "date": "2025-07-29", "categories": []}
{"id": "2502.04469", "title": "Ask and Remember: A Questions-Only Replay Strategy for Continual Visual Question Answering", "abstract": "Continual Learning in Visual Question Answering (VQACL) requires models to acquire new visual-linguistic skills (plasticity) while preserving previously learned knowledge (stability). The inherent multimodality of VQACL exacerbates this challenge, as models must balance stability across visual and textual domains while adapting to novel objects and reasoning tasks. Existing methods, primarily designed for unimodal settings, often fall short in addressing this dual requirement. In this work, we present QUestion-only replay with Attention Distillation (QUAD), a novel approach for VQACL that leverages only past task questions for regularization. By eliminating the need to store visual data, QUAD not only reduces memory overhead, but also alleviates privacy concerns. Our method introduces a Question-only Replay mechanism that selectively reuses prior task questions to counteract overfitting to the answer space of the current task, addressing the problem out of answer set. Complementing this, we propose Attention Consistency Distillation to enforce both intra-modal and inter-modal attention consistency across tasks, preserving essential visual-linguistic associations. Extensive experiments on VQAv2 and NExT-QA demonstrate that QUAD significantly outperforms state-of-the-art methods, achieving robust performance in continual VQA. Code is available at: https://github.com/IemProg/QUAD.", "date": "2025-02-06", "categories": []}
{"id": "2502.20974", "title": "Improving Open-world Continual Learning under the Constraints of Scarce Labeled Data", "abstract": "Open-world continual learning (OWCL) adapts to sequential tasks with open samples, learning knowledge incrementally while preventing forgetting. However, existing OWCL still requires a large amount of labeled data for training, which is often impractical in real-world applications. Given that new categories/entities typically come with limited annotations and are in small quantities, a more realistic situation is OWCL with scarce labeled data, i.e., few-shot training samples. Hence, this paper investigates the problem of open-world few-shot continual learning (OFCL), challenging in (i) learning unbounded tasks without forgetting previous knowledge and avoiding overfitting, (ii) constructing compact decision boundaries for open detection with limited labeled data, and (iii) transferring knowledge about knowns and unknowns and even update the unknowns to knowns once the labels of open samples are learned. In response, we propose a novel OFCL framework that integrates three key components: (1) an instance-wise token augmentation (ITA) that represents and enriches sample representations with additional knowledge, (2) a margin-based open boundary (MOB) that supports open detection with new tasks emerge over time, and (3) an adaptive knowledge space (AKS) that endows unknowns with knowledge for the updating from unknowns to knowns. Finally, extensive experiments show that the proposed OFCL framework outperforms all baselines remarkably with practical importance and reproducibility. The source code is released at https://github.com/liyj1201/OFCL.", "date": "2025-02-28", "categories": []}
{"id": "2503.04258", "title": "TAIL: Text-Audio Incremental Learning", "abstract": "Many studies combine text and audio to capture multi-modal information but they overlook the model's generalization ability on new datasets. Introducing new datasets may affect the feature space of the original dataset, leading to catastrophic forgetting. Meanwhile, large model parameters can significantly impact training performance. To address these limitations, we introduce a novel task called Text-Audio Incremental Learning (TAIL) task for text-audio retrieval, and propose a new method, PTAT, Prompt Tuning for Audio-Text incremental learning. This method utilizes prompt tuning to optimize the model parameters while incorporating an audio-text similarity and feature distillation module to effectively mitigate catastrophic forgetting. We benchmark our method and previous incremental learning methods on AudioCaps, Clotho, BBC Sound Effects and Audioset datasets, and our method outperforms previous methods significantly, particularly demonstrating stronger resistance to forgetting on older datasets. Compared to the full-parameters Finetune (Sequential) method, our model only requires 2.42\\% of its parameters, achieving 4.46\\% higher performance.", "date": "2025-03-06", "categories": []}
{"id": "2504.00707", "title": "Interleaved Multitask Learning with Energy Modulated Learning Progress", "abstract": "As humans learn new skills and apply their existing knowledge while maintaining previously learned information, \"continual learning\" in machine learning aims to incorporate new data while retaining and utilizing past knowledge. However, existing machine learning methods often does not mimic human learning where tasks are intermixed due to individual preferences and environmental conditions. Humans typically switch between tasks instead of completely mastering one task before proceeding to the next. To explore how human-like task switching can enhance learning efficiency, we propose a multi task learning architecture that alternates tasks based on task-agnostic measures such as \"learning progress\" and \"neural computational energy expenditure\". To evaluate the efficacy of our method, we run several systematic experiments by using a set of effect-prediction tasks executed by a simulated manipulator robot. The experiments show that our approach surpasses random interleaved and sequential task learning in terms of average learning accuracy. Moreover, by including energy expenditure in the task switching logic, our approach can still perform favorably while reducing neural energy expenditure.", "date": "2025-04-01", "categories": []}
{"id": "2507.19839", "title": "GNSP: Gradient Null Space Projection for Preserving Cross-Modal Alignment in VLMs Continual Learning", "abstract": "Contrastive Language-Image Pretraining has demonstrated remarkable zero-shot generalization by aligning visual and textual modalities in a shared embedding space. However, when continuously fine-tuned on diverse tasks, CLIP suffers from catastrophic forgetting and degradation of its embedding alignment, undermining its zero-shot capabilities. In this work, we propose Gradient Null Space Projection (GNSP), an efficient continual learning method that projects task-specific gradients onto the null space of previously learned knowledge. This orthogonal projection mathematically prevents interference with previous tasks without relying on rehearsal or architectural modification. Furthermore, to preserve the inherent generalization property of CLIP, we introduce knowledge distillation and combine it with a modality alignment preservation loss inspired by CLIP pre-training to stabilize the structure of the multimodal embedding space during fine-tuning. On the MTIL benchmark consisting of 11 tasks, our method achieved SOTA performance on both the Average and Last key metrics. More importantly, experiments show that our method successfully maintains the original modality gap and cross-modal retrieval performance of CLIP, confirming its effectiveness in maintaining a robust visual-language space throughout the continual learning process.", "date": "2025-07-26", "categories": []}
{"id": "2507.19887", "title": "CLoRA: Parameter-Efficient Continual Learning with Low-Rank Adaptation", "abstract": "In the past, continual learning (CL) was mostly concerned with the problem of catastrophic forgetting in neural networks, that arises when incrementally learning a sequence of tasks. Current CL methods function within the confines of limited data access, without any restrictions imposed on computational resources. However, in real-world scenarios, the latter takes precedence as deployed systems are often computationally constrained. A major drawback of most CL methods is the need to retrain the entire model for each new task. The computational demands of retraining large models can be prohibitive, limiting the applicability of CL in environments with limited resources. Through CLoRA, we explore the applicability of Low-Rank Adaptation (LoRA), a parameter-efficient fine-tuning method for class-incremental semantic segmentation. CLoRA leverages a small set of parameters of the model and uses the same set for learning across all tasks. Results demonstrate the efficacy of CLoRA, achieving performance on par with and exceeding the baseline methods. We further evaluate CLoRA using NetScore, underscoring the need to factor in resource efficiency and evaluate CL methods beyond task performance. CLoRA significantly reduces the hardware requirements for training, making it well-suited for CL in resource-constrained environments after deployment.", "date": "2025-07-26", "categories": []}
{"id": "2507.20057", "title": "What Can Grokking Teach Us About Learning Under Nonstationarity?", "abstract": "In continual learning problems, it is often necessary to overwrite components of a neural network's learned representation in response to changes in the data stream; however, neural networks often exhibit \\primacy bias, whereby early training data hinders the network's ability to generalize on later tasks. While feature-learning dynamics of nonstationary learning problems are not well studied, the emergence of feature-learning dynamics is known to drive the phenomenon of grokking, wherein neural networks initially memorize their training data and only later exhibit perfect generalization. This work conjectures that the same feature-learning dynamics which facilitate generalization in grokking also underlie the ability to overwrite previous learned features as well, and methods which accelerate grokking by facilitating feature-learning dynamics are promising candidates for addressing primacy bias in non-stationary learning problems. We then propose a straightforward method to induce feature-learning dynamics as needed throughout training by increasing the effective learning rate, i.e. the ratio between parameter and update norms. We show that this approach both facilitates feature-learning and improves generalization in a variety of settings, including grokking, warm-starting neural network training, and reinforcement learning tasks.", "date": "2025-07-26", "categories": []}
{"id": "2507.16039", "title": "Reactivation: Empirical NTK Dynamics Under Task Shifts", "abstract": "The Neural Tangent Kernel (NTK) offers a powerful tool to study the functional dynamics of neural networks. In the so-called lazy, or kernel regime, the NTK remains static during training and the network function is linear in the static neural tangents feature space. The evolution of the NTK during training is necessary for feature learning, a key driver of deep learning success. The study of the NTK dynamics has led to several critical discoveries in recent years, in generalization and scaling behaviours. However, this body of work has been limited to the single task setting, where the data distribution is assumed constant over time. In this work, we present a comprehensive empirical analysis of NTK dynamics in continual learning, where the data distribution shifts over time. Our findings highlight continual learning as a rich and underutilized testbed for probing the dynamics of neural training. At the same time, they challenge the validity of static-kernel approximations in theoretical treatments of continual learning, even at large scale.", "date": "2025-07-21", "categories": []}
{"id": "2507.18996", "title": "Adapting to Fragmented and Evolving Data: A Fisher Information Perspective", "abstract": "Modern machine learning systems operating in dynamic environments often face \\textit{sequential covariate shift} (SCS), where input distributions evolve over time while the conditional distribution remains stable. We introduce FADE (Fisher-based Adaptation to Dynamic Environments), a lightweight and theoretically grounded framework for robust learning under SCS. FADE employs a shift-aware regularization mechanism anchored in Fisher information geometry, guiding adaptation by modulating parameter updates based on sensitivity and stability. To detect significant distribution changes, we propose a Cramer-Rao-informed shift signal that integrates KL divergence with temporal Fisher dynamics. Unlike prior methods requiring task boundaries, target supervision, or experience replay, FADE operates online with fixed memory and no access to target labels. Evaluated on seven benchmarks spanning vision, language, and tabular data, FADE achieves up to 19\\% higher accuracy under severe shifts, outperforming methods such as TENT and DIW. FADE also generalizes naturally to federated learning by treating heterogeneous clients as temporally fragmented environments, enabling scalable and stable adaptation in decentralized settings. Theoretical analysis guarantees bounded regret and parameter consistency, while empirical results demonstrate FADE's robustness across modalities and shift intensities.", "date": "2025-07-25", "categories": []}
{"id": "2503.08510", "title": "External Knowledge Injection for CLIP-Based Class-Incremental Learning", "abstract": "Class-Incremental Learning (CIL) enables learning systems to continuously adapt to evolving data streams. With the advancement of pre-training, leveraging pre-trained vision-language models (e.g., CLIP) offers a promising starting point for CIL. However, CLIP makes decisions by matching visual embeddings to class names, overlooking the rich contextual information conveyed through language. For instance, the concept of ``cat'' can be decomposed into features like tail, fur, and face for recognition. Besides, since the model is continually updated, these detailed features are overwritten in CIL, requiring external knowledge for compensation. In this paper, we introduce ExterNal knowledGe INjEction (ENGINE) for CLIP-based CIL. To enhance knowledge transfer from outside the dataset, we propose a dual-branch injection tuning framework that encodes informative knowledge from both visual and textual modalities. The visual branch is enhanced with data augmentation to enrich the visual features, while the textual branch leverages GPT-4 to rewrite discriminative descriptors. In addition to this on-the-fly knowledge injection, we also implement post-tuning knowledge by re-ranking the prediction results during inference. With the injected knowledge, the model can better capture informative features for downstream tasks as data evolves. Extensive experiments demonstrate the state-of-the-art performance of ENGINE. Code is available at: https://github.com/LAMDA-CL/ICCV25-ENGINE", "date": "2025-03-11", "categories": []}
{"id": "2507.17953", "title": "Clo-HDnn: A 4.66 TFLOPS/W and 3.78 TOPS/W Continual On-Device Learning Accelerator with Energy-efficient Hyperdimensional Computing via Progressive Search", "abstract": "Clo-HDnn is an on-device learning (ODL) accelerator designed for emerging continual learning (CL) tasks. Clo-HDnn integrates hyperdimensional computing (HDC) along with low-cost Kronecker HD Encoder and weight clustering feature extraction (WCFE) to optimize accuracy and efficiency. Clo-HDnn adopts gradient-free CL to efficiently update and store the learned knowledge in the form of class hypervectors. Its dual-mode operation enables bypassing costly feature extraction for simpler datasets, while progressive search reduces complexity by up to 61% by encoding and comparing only partial query hypervectors. Achieving 4.66 TFLOPS/W (FE) and 3.78 TOPS/W (classifier), Clo-HDnn delivers 7.77x and 4.85x higher energy efficiency compared to SOTA ODL accelerators.", "date": "2025-07-23", "categories": []}
{"id": "2507.18313", "title": "Regression-aware Continual Learning for Android Malware Detection", "abstract": "Malware evolves rapidly, forcing machine learning (ML)-based detectors to adapt continuously. With antivirus vendors processing hundreds of thousands of new samples daily, datasets can grow to billions of examples, making full retraining impractical. Continual learning (CL) has emerged as a scalable alternative, enabling incremental updates without full data access while mitigating catastrophic forgetting. In this work, we analyze a critical yet overlooked issue in this context: security regression. Unlike forgetting, which manifests as a general performance drop on previously seen data, security regression captures harmful prediction changes at the sample level, such as a malware sample that was once correctly detected but evades detection after a model update. Although often overlooked, regressions pose serious risks in security-critical applications, as the silent reintroduction of previously detected threats in the system may undermine users' trust in the whole updating process. To address this issue, we formalize and quantify security regression in CL-based malware detectors and propose a regression-aware penalty to mitigate it. Specifically, we adapt Positive Congruent Training (PCT) to the CL setting, preserving prior predictive behavior in a model-agnostic manner. Experiments on the ELSA, Tesseract, and AZ-Class datasets show that our method effectively reduces regression across different CL scenarios while maintaining strong detection performance over time.", "date": "2025-07-24", "categories": []}
{"id": "2507.17368", "title": "ViRN: Variational Inference and Distribution Trilateration for Long-Tailed Continual Representation Learning", "abstract": "Continual learning (CL) with long-tailed data distributions remains a critical challenge for real-world AI systems, where models must sequentially adapt to new classes while retaining knowledge of old ones, despite severe class imbalance. Existing methods struggle to balance stability and plasticity, often collapsing under extreme sample scarcity. To address this, we propose ViRN, a novel CL framework that integrates variational inference (VI) with distributional trilateration for robust long-tailed learning. First, we model class-conditional distributions via a Variational Autoencoder to mitigate bias toward head classes. Second, we reconstruct tail-class distributions via Wasserstein distance-based neighborhood retrieval and geometric fusion, enabling sample-efficient alignment of tail-class representations. Evaluated on six long-tailed classification benchmarks, including speech (e.g., rare acoustic events, accents) and image tasks, ViRN achieves a 10.24% average accuracy gain over state-of-the-art methods.", "date": "2025-07-23", "categories": []}
{"id": "2507.17382", "title": "Continual Generalized Category Discovery: Learning and Forgetting from a Bayesian Perspective", "abstract": "Continual Generalized Category Discovery (C-GCD) faces a critical challenge: incrementally learning new classes from unlabeled data streams while preserving knowledge of old classes. Existing methods struggle with catastrophic forgetting, especially when unlabeled data mixes known and novel categories. We address this by analyzing C-GCD's forgetting dynamics through a Bayesian lens, revealing that covariance misalignment between old and new classes drives performance degradation. Building on this insight, we propose Variational Bayes C-GCD (VB-CGCD), a novel framework that integrates variational inference with covariance-aware nearest-class-mean classification. VB-CGCD adaptively aligns class distributions while suppressing pseudo-label noise via stochastic variational updates. Experiments show VB-CGCD surpasses prior art by +15.21% with the overall accuracy in the final session on standard benchmarks. We also introduce a new challenging benchmark with only 10% labeled data and extended online phases, VB-CGCD achieves a 67.86% final accuracy, significantly higher than state-of-the-art (38.55%), demonstrating its robust applicability across diverse scenarios. Code is available at: https://github.com/daihao42/VB-CGCD", "date": "2025-07-23", "categories": []}
{"id": "2507.17482", "title": "LTLZinc: a Benchmarking Framework for Continual Learning and Neuro-Symbolic Temporal Reasoning", "abstract": "Neuro-symbolic artificial intelligence aims to combine neural architectures with symbolic approaches that can represent knowledge in a human-interpretable formalism. Continual learning concerns with agents that expand their knowledge over time, improving their skills while avoiding to forget previously learned concepts. Most of the existing approaches for neuro-symbolic artificial intelligence are applied to static scenarios only, and the challenging setting where reasoning along the temporal dimension is necessary has been seldom explored. In this work we introduce LTLZinc, a benchmarking framework that can be used to generate datasets covering a variety of different problems, against which neuro-symbolic and continual learning methods can be evaluated along the temporal and constraint-driven dimensions. Our framework generates expressive temporal reasoning and continual learning tasks from a linear temporal logic specification over MiniZinc constraints, and arbitrary image classification datasets. Fine-grained annotations allow multiple neural and neuro-symbolic training settings on the same generated datasets. Experiments on six neuro-symbolic sequence classification and four class-continual learning tasks generated by LTLZinc, demonstrate the challenging nature of temporal learning and reasoning, and highlight limitations of current state-of-the-art methods. We release the LTLZinc generator and ten ready-to-use tasks to the neuro-symbolic and continual learning communities, in the hope of fostering research towards unified temporal learning and reasoning frameworks.", "date": "2025-07-23", "categories": []}
{"id": "2507.17687", "title": "Towards Effective Open-set Graph Class-incremental Learning", "abstract": "Graph class-incremental learning (GCIL) allows graph neural networks (GNNs) to adapt to evolving graph analytical tasks by incrementally learning new class knowledge while retaining knowledge of old classes. Existing GCIL methods primarily focus on a closed-set assumption, where all test samples are presumed to belong to previously known classes. Such an assumption restricts their applicability in real-world scenarios, where unknown classes naturally emerge during inference, and are absent during training. In this paper, we explore a more challenging open-set graph class-incremental learning scenario with two intertwined challenges: catastrophic forgetting of old classes, which impairs the detection of unknown classes, and inadequate open-set recognition, which destabilizes the retention of learned knowledge. To address the above problems, a novel OGCIL framework is proposed, which utilizes pseudo-sample embedding generation to effectively mitigate catastrophic forgetting and enable robust detection of unknown classes. To be specific, a prototypical conditional variational autoencoder is designed to synthesize node embeddings for old classes, enabling knowledge replay without storing raw graph data. To handle unknown classes, we employ a mixing-based strategy to generate out-of-distribution (OOD) samples from pseudo in-distribution and current node embeddings. A novel prototypical hypersphere classification loss is further proposed, which anchors in-distribution embeddings to their respective class prototypes, while repelling OOD embeddings away. Instead of assigning all unknown samples into one cluster, our proposed objective function explicitly models them as outliers through prototype-aware rejection regions, ensuring a robust open-set recognition. Extensive experiments on five benchmarks demonstrate the effectiveness of OGCIL over existing GCIL and open-set GNN methods.", "date": "2025-07-23", "categories": []}
{"id": "2310.07497", "title": "Energy-Efficient and Real-Time Sensing for Federated Continual Learning via Sample-Driven Control", "abstract": "An intelligent Real-Time Sensing (RTS) system must continuously acquire, update, integrate, and apply knowledge to adapt to real-world dynamics. Managing distributed intelligence in this context requires Federated Continual Learning (FCL). However, effectively capturing the diverse characteristics of RTS data in FCL systems poses significant challenges, including severely impacting computational and communication resources, escalating energy costs, and ultimately degrading overall system performance. To overcome these challenges, we investigate how the data distribution shift from ideal to practical RTS scenarios affects Artificial Intelligence (AI) model performance by leveraging the \\textit{generalization gap} concept. In this way, we can analyze how sampling time in RTS correlates with the decline in AI performance, computation cost, and communication efficiency. Based on this observation, we develop a novel Sample-driven Control for Federated Continual Learning (SCFL) technique, specifically designed for mobile edge networks with RTS capabilities. In particular, SCFL is an optimization problem that harnesses the sampling process to concurrently minimize the generalization gap and improve overall accuracy while upholding the energy efficiency of the FCL framework. To solve the highly complex and time-varying optimization problem, we introduce a new soft actor-critic algorithm with explicit and implicit constraints (A2C-EI). Our empirical experiments reveal that we can achieve higher efficiency compared to other DRL baselines. Notably, SCFL can significantly reduce energy consumption up to $85\\%$ while maintaining FL convergence and timely data transmission.", "date": "2023-10-11", "categories": []}
{"id": "2407.07668", "title": "How to Leverage Predictive Uncertainty Estimates for Reducing Catastrophic Forgetting in Online Continual Learning", "abstract": "Many real-world applications require machine-learning models to be able to deal with non-stationary data distributions and thus learn autonomously over an extended period of time, often in an online setting. One of the main challenges in this scenario is the so-called catastrophic forgetting (CF) for which the learning model tends to focus on the most recent tasks while experiencing predictive degradation on older ones. In the online setting, the most effective solutions employ a fixed-size memory buffer to store old samples used for replay when training on new tasks. Many approaches have been presented to tackle this problem. However, it is not clear how predictive uncertainty information for memory management can be leveraged in the most effective manner and conflicting strategies are proposed to populate the memory. Are the easiest-to-forget or the easiest-to-remember samples more effective in combating CF? Starting from the intuition that predictive uncertainty provides an idea of the samples' location in the decision space, this work presents an in-depth analysis of different uncertainty estimates and strategies for populating the memory. The investigation provides a better understanding of the characteristics data points should have for alleviating CF. Then, we propose an alternative method for estimating predictive uncertainty via the generalised variance induced by the negative log-likelihood. Finally, we demonstrate that the use of predictive uncertainty measures helps in reducing CF in different settings.", "date": "2024-07-10", "categories": []}
{"id": "2410.09218", "title": "Continual Learning with Neuromorphic Computing: Foundations, Methods, and Emerging Applications", "abstract": "The challenging deployment of compute- and memory-intensive methods from Deep Neural Network (DNN)-based Continual Learning (CL) underscores the critical need for a paradigm shift towards more efficient approaches. Neuromorphic Continual Learning (NCL) appears as an emerging solution, by leveraging the principles of Spiking Neural Networks (SNNs) which enable efficient CL algorithms executed in dynamically-changed environments with resource-constrained computing systems. Motivated by the need for a holistic study of NCL, in this survey, we first provide a detailed background on CL, encompassing the desiderata, settings, metrics, scenario taxonomy, Online Continual Learning (OCL) paradigm, recent DNN-based methods to address catastrophic forgetting (CF). Then, we analyze these methods considering CL desiderata, computational and memory costs, as well as network complexity, hence emphasizing the need for energy-efficient CL. Afterward, we provide background of low-power neuromorphic systems including encoding techniques, neuronal dynamics, network architectures, learning rules, hardware processors, software and hardware frameworks, datasets, benchmarks, and evaluation metrics. Then, this survey comprehensively reviews and analyzes state-of-the-art in NCL. The key ideas, implementation frameworks, and performance assessments are also provided. This survey covers several hybrid approaches that combine supervised and unsupervised learning paradigms. It also covers optimization techniques including SNN operations reduction, weight quantization, and knowledge distillation. Then, this survey discusses the progress of real-world NCL applications. Finally, this paper provides a future perspective on the open research challenges for NCL, since the purpose of this study is to be useful for the wider neuromorphic AI research community and to inspire future research in bio-plausible OCL.", "date": "2024-10-11", "categories": []}
{"id": "2410.20444", "title": "Vector Quantization Prompting for Continual Learning", "abstract": "Continual learning requires to overcome catastrophic forgetting when training a single model on a sequence of tasks. Recent top-performing approaches are prompt-based methods that utilize a set of learnable parameters (i.e., prompts) to encode task knowledge, from which appropriate ones are selected to guide the fixed pre-trained model in generating features tailored to a certain task. However, existing methods rely on predicting prompt identities for prompt selection, where the identity prediction process cannot be optimized with task loss. This limitation leads to sub-optimal prompt selection and inadequate adaptation of pre-trained features for a specific task. Previous efforts have tried to address this by directly generating prompts from input queries instead of selecting from a set of candidates. However, these prompts are continuous, which lack sufficient abstraction for task knowledge representation, making them less effective for continual learning. To address these challenges, we propose VQ-Prompt, a prompt-based continual learning method that incorporates Vector Quantization (VQ) into end-to-end training of a set of discrete prompts. In this way, VQ-Prompt can optimize the prompt selection process with task loss and meanwhile achieve effective abstraction of task knowledge for continual learning. Extensive experiments show that VQ-Prompt outperforms state-of-the-art continual learning methods across a variety of benchmarks under the challenging class-incremental setting. The code is available at \\href{https://github.com/jiaolifengmi/VQ-Prompt}{this https URL}.", "date": "2024-10-27", "categories": []}
{"id": "2502.03350", "title": "Optimal Task Order for Continual Learning of Multiple Tasks", "abstract": "Continual learning of multiple tasks remains a major challenge for neural networks. Here, we investigate how task order influences continual learning and propose a strategy for optimizing it. Leveraging a linear teacher-student model with latent factors, we derive an analytical expression relating task similarity and ordering to learning performance. Our analysis reveals two principles that hold under a wide parameter range: (1) tasks should be arranged from the least representative to the most typical, and (2) adjacent tasks should be dissimilar. We validate these rules on both synthetic data and real-world image classification datasets (Fashion-MNIST, CIFAR-10, CIFAR-100), demonstrating consistent performance improvements in both multilayer perceptrons and convolutional neural networks. Our work thus presents a generalizable framework for task-order optimization in task-incremental continual learning.", "date": "2025-02-05", "categories": []}
{"id": "2503.12897", "title": "Federated Continual Instruction Tuning", "abstract": "A vast amount of instruction tuning data is crucial for the impressive performance of Large Multimodal Models (LMMs), but the associated computational costs and data collection demands during supervised fine-tuning make it impractical for most researchers. Federated learning (FL) has the potential to leverage all distributed data and training resources to reduce the overhead of joint training. However, most existing methods assume a fixed number of tasks, while in real-world scenarios, clients continuously encounter new knowledge and often struggle to retain old tasks due to memory constraints. In this work, we introduce the Federated Continual Instruction Tuning (FCIT) benchmark to model this real-world challenge. Our benchmark includes two realistic scenarios, encompassing four different settings and twelve carefully curated instruction tuning datasets. To address the challenges posed by FCIT, we propose dynamic knowledge organization to effectively integrate updates from different tasks during training and subspace selective activation to allocate task-specific output during inference. Extensive experimental results demonstrate that our proposed method significantly enhances model performance across varying levels of data heterogeneity and catastrophic forgetting. Code and dataset are released at https://github.com/Ghy0501/FCIT.", "date": "2025-03-17", "categories": []}
{"id": "2507.09183", "title": "Revisiting Pool-based Prompt Learning for Few-shot Class-incremental Learning", "abstract": "Few-Shot Class-Incremental Learning (FSCIL) faces dual challenges of data scarcity and incremental learning in real-world scenarios. While pool-based prompting methods have demonstrated success in traditional incremental learning, their effectiveness in FSCIL settings remains unexplored. This paper presents the first study of current prompt pool methods in FSCIL tasks, revealing an unanticipated performance degradation in incremental sessions. Through comprehensive analysis, we identify that this phenomenon stems from token-dimension saturation: with limited data, excessive prompts compete for task-relevant information, leading to model overfitting. Based on this finding, we propose LGSP-Prompt (Local-Global Spatial Prompting), which innovatively shifts pool-based prompt learning from the token dimension to the spatial dimension. LGSP-Prompt generates spatial prompts by synergistically combining local spatial features and global frequency-domain representations to highlight key patterns in input images. We construct two spatial prompt pools enabling dynamic prompt selection to maintain acquired knowledge while effectively learning novel sessions. Extensive experiments demonstrate that our approach achieves state-of-the-art performance across multiple FSCIL benchmarks, showing significant advantages in both base knowledge preservation and incremental learning. Our implementation is available at https://github.com/Jywsuperman/LGSP.", "date": "2025-07-12", "categories": []}
{"id": "2507.10637", "title": "A Simple Baseline for Stable and Plastic Neural Networks", "abstract": "Continual learning in computer vision requires that models adapt to a continuous stream of tasks without forgetting prior knowledge, yet existing approaches often tip the balance heavily toward either plasticity or stability. We introduce RDBP, a simple, low-overhead baseline that unites two complementary mechanisms: ReLUDown, a lightweight activation modification that preserves feature sensitivity while preventing neuron dormancy, and Decreasing Backpropagation, a biologically inspired gradient-scheduling scheme that progressively shields early layers from catastrophic updates. Evaluated on the Continual ImageNet benchmark, RDBP matches or exceeds the plasticity and stability of state-of-the-art methods while reducing computational cost. RDBP thus provides both a practical solution for real-world continual learning and a clear benchmark against which future continual learning strategies can be measured.", "date": "2025-07-14", "categories": []}
{"id": "2507.13739", "title": "Can Synthetic Images Conquer Forgetting? Beyond Unexplored Doubts in Few-Shot Class-Incremental Learning", "abstract": "Few-shot class-incremental learning (FSCIL) is challenging due to extremely limited training data; while aiming to reduce catastrophic forgetting and learn new information. We propose Diffusion-FSCIL, a novel approach that employs a text-to-image diffusion model as a frozen backbone. Our conjecture is that FSCIL can be tackled using a large generative model's capabilities benefiting from 1) generation ability via large-scale pre-training; 2) multi-scale representation; 3) representational flexibility through the text encoder. To maximize the representation capability, we propose to extract multiple complementary diffusion features to play roles as latent replay with slight support from feature distillation for preventing generative biases. Our framework realizes efficiency through 1) using a frozen backbone; 2) minimal trainable components; 3) batch processing of multiple feature extractions. Extensive experiments on CUB-200, \\emph{mini}ImageNet, and CIFAR-100 show that Diffusion-FSCIL surpasses state-of-the-art methods, preserving performance on previously learned classes and adapting effectively to new ones.", "date": "2025-07-18", "categories": []}
{"id": "2507.14050", "title": "Foundation Models as Class-Incremental Learners for Dermatological Image Classification", "abstract": "Class-Incremental Learning (CIL) aims to learn new classes over time without forgetting previously acquired knowledge. The emergence of foundation models (FM) pretrained on large datasets presents new opportunities for CIL by offering rich, transferable representations. However, their potential for enabling incremental learning in dermatology remains largely unexplored. In this paper, we systematically evaluate frozen FMs pretrained on large-scale skin lesion datasets for CIL in dermatological disease classification. We propose a simple yet effective approach where the backbone remains frozen, and a lightweight MLP is trained incrementally for each task. This setup achieves state-of-the-art performance without forgetting, outperforming regularization, replay, and architecture based methods. To further explore the capabilities of frozen FMs, we examine zero training scenarios using nearest mean classifiers with prototypes derived from their embeddings. Through extensive ablation studies, we demonstrate that this prototype based variant can also achieve competitive results. Our findings highlight the strength of frozen FMs for continual learning in dermatology and support their broader adoption in real world medical applications. Our code and datasets are available here.", "date": "2025-07-18", "categories": []}
{"id": "2507.13107", "title": "R^2MoE: Redundancy-Removal Mixture of Experts for Lifelong Concept Learning", "abstract": "Enabling large-scale generative models to continuously learn new visual concepts is essential for personalizing pre-trained models to meet individual user preferences. Existing approaches for continual visual concept learning are constrained by two fundamental challenges: catastrophic forgetting and parameter expansion. In this paper, we propose Redundancy-Removal Mixture of Experts (R^2MoE), a parameter-efficient framework for lifelong visual concept learning that effectively learns new concepts while incurring minimal parameter overhead. Our framework includes three key innovative contributions: First, we propose a mixture-of-experts framework with a routing distillation mechanism that enables experts to acquire concept-specific knowledge while preserving the gating network's routing capability, thereby effectively mitigating catastrophic forgetting. Second, we propose a strategy for eliminating redundant layer-wise experts that reduces the number of expert parameters by fully utilizing previously learned experts. Third, we employ a hierarchical local attention-guided inference approach to mitigate interference between generated visual concepts. Extensive experiments have demonstrated that our method generates images with superior conceptual fidelity compared to the state-of-the-art (SOTA) method, achieving an impressive 87.8\\% reduction in forgetting rates and 63.3\\% fewer parameters on the CustomConcept 101 dataset. Our code is available at {https://github.com/learninginvision/R2MoE}", "date": "2025-07-17", "categories": []}
{"id": "2502.07059", "title": "Federated Continual Learning: Concepts, Challenges, and Solutions", "abstract": "Federated Continual Learning (FCL) has emerged as a robust solution for collaborative model training in dynamic environments, where data samples are continuously generated and distributed across multiple devices. This survey provides a comprehensive review of FCL, focusing on key challenges such as heterogeneity, model stability, communication overhead, and privacy preservation. We explore various forms of heterogeneity and their impact on model performance. Solutions to non-IID data, resource-constrained platforms, and personalized learning are reviewed in an effort to show the complexities of handling heterogeneous data distributions. Next, we review techniques for ensuring model stability and avoiding catastrophic forgetting, which are critical in non-stationary environments. Privacy-preserving techniques are another aspect of FCL that have been reviewed in this work. This survey has integrated insights from federated learning and continual learning to present strategies for improving the efficacy and scalability of FCL systems, making it applicable to a wide range of real-world scenarios.", "date": "2025-02-10", "categories": []}
{"id": "2507.12297", "title": "RegCL: Continual Adaptation of Segment Anything Model via Model Merging", "abstract": "To address the performance limitations of the Segment Anything Model (SAM) in specific domains, existing works primarily adopt adapter-based one-step adaptation paradigms. However, some of these methods are specific developed for specific domains. If used on other domains may lead to performance degradation. This issue of catastrophic forgetting severely limits the model's scalability. To address this issue, this paper proposes RegCL, a novel non-replay continual learning (CL) framework designed for efficient multi-domain knowledge integration through model merging. Specifically, RegCL incorporates the model merging algorithm into the continual learning paradigm by merging the parameters of SAM's adaptation modules (e.g., LoRA modules) trained on different domains. The merging process is guided by weight optimization, which minimizes prediction discrepancies between the merged model and each of the domain-specific models. RegCL effectively consolidates multi-domain knowledge while maintaining parameter efficiency, i.e., the model size remains constant regardless of the number of tasks, and no historical data storage is required. Experimental results demonstrate that RegCL achieves favorable continual learning performance across multiple downstream datasets, validating its effectiveness in dynamic scenarios.", "date": "2025-07-16", "categories": []}
{"id": "2507.12305", "title": "PROL : Rehearsal Free Continual Learning in Streaming Data via Prompt Online Learning", "abstract": "The data privacy constraint in online continual learning (OCL), where the data can be seen only once, complicates the catastrophic forgetting problem in streaming data. A common approach applied by the current SOTAs in OCL is with the use of memory saving exemplars or features from previous classes to be replayed in the current task. On the other hand, the prompt-based approach performs excellently in continual learning but with the cost of a growing number of trainable parameters. The first approach may not be applicable in practice due to data openness policy, while the second approach has the issue of throughput associated with the streaming data. In this study, we propose a novel prompt-based method for online continual learning that includes 4 main components: (1) single light-weight prompt generator as a general knowledge, (2) trainable scaler-and-shifter as specific knowledge, (3) pre-trained model (PTM) generalization preserving, and (4) hard-soft updates mechanism. Our proposed method achieves significantly higher performance than the current SOTAs in CIFAR100, ImageNet-R, ImageNet-A, and CUB dataset. Our complexity analysis shows that our method requires a relatively smaller number of parameters and achieves moderate training time, inference time, and throughput. For further study, the source code of our method is available at https://github.com/anwarmaxsum/PROL.", "date": "2025-07-16", "categories": []}
{"id": "2504.08161", "title": "Rethinking the Foundations for Continual Reinforcement Learning", "abstract": "In the traditional view of reinforcement learning, the agent's goal is to find an optimal policy that maximizes its expected sum of rewards. Once the agent finds this policy, the learning ends. This view contrasts with \\emph{continual reinforcement learning}, where learning does not end, and agents are expected to continually learn and adapt indefinitely. Despite the clear distinction between these two paradigms of learning, much of the progress in continual reinforcement learning has been shaped by foundations rooted in the traditional view of reinforcement learning. In this paper, we first examine whether the foundations of traditional reinforcement learning are suitable for the continual reinforcement learning paradigm. We identify four key pillars of the traditional reinforcement learning foundations that are antithetical to the goals of continual learning: the Markov decision process formalism, the focus on atemporal artifacts, the expected sum of rewards as an evaluation metric, and episodic benchmark environments that embrace the other three foundations. We then propose a new formalism that sheds the first and the third foundations and replaces them with the history process as a mathematical formalism and a new definition of deviation regret, adapted for continual learning, as an evaluation metric. Finally, we discuss possible approaches to shed the other two foundations.", "date": "2025-04-10", "categories": []}
{"id": "2507.10434", "title": "CLA: Latent Alignment for Online Continual Self-Supervised Learning", "abstract": "Self-supervised learning (SSL) is able to build latent representations that generalize well to unseen data. However, only a few SSL techniques exist for the online CL setting, where data arrives in small minibatches, the model must comply with a fixed computational budget, and task boundaries are absent. We introduce Continual Latent Alignment (CLA), a novel SSL strategy for Online CL that aligns the representations learned by the current model with past representations to mitigate forgetting. We found that our CLA is able to speed up the convergence of the training process in the online scenario, outperforming state-of-the-art approaches under the same computational budget. Surprisingly, we also discovered that using CLA as a pretraining protocol in the early stages of pretraining leads to a better final performance when compared to a full i.i.d. pretraining.", "date": "2025-07-14", "categories": []}
{"id": "2507.11393", "title": "A Neural Network Model of Complementary Learning Systems: Pattern Separation and Completion for Continual Learning", "abstract": "Learning new information without forgetting prior knowledge is central to human intelligence. In contrast, neural network models suffer from catastrophic forgetting: a significant degradation in performance on previously learned tasks when acquiring new information. The Complementary Learning Systems (CLS) theory offers an explanation for this human ability, proposing that the brain has distinct systems for pattern separation (encoding distinct memories) and pattern completion (retrieving complete memories from partial cues). To capture these complementary functions, we leverage the representational generalization capabilities of variational autoencoders (VAEs) and the robust memory storage properties of Modern Hopfield networks (MHNs), combining them into a neurally plausible continual learning model. We evaluate this model on the Split-MNIST task, a popular continual learning benchmark, and achieve close to state-of-the-art accuracy (~90%), substantially reducing forgetting. Representational analyses empirically confirm the functional dissociation: the VAE underwrites pattern completion, while the MHN drives pattern separation. By capturing pattern separation and completion in scalable architectures, our work provides a functional template for modeling memory consolidation, generalization, and continual learning in both biological and artificial systems.", "date": "2025-07-15", "categories": []}
{"id": "2507.09118", "title": "Mind the Gap: Preserving and Compensating for the Modality Gap in CLIP-Based Continual Learning", "abstract": "Continual learning aims to enable models to learn sequentially from continuously incoming data while retaining performance on previously learned tasks. With the Contrastive Language-Image Pre-trained model (CLIP) exhibiting strong capabilities across various downstream tasks, there has been growing interest in leveraging CLIP for continual learning in such scenarios. Most existing works overlook the inherent modality gap in CLIP, a key factor in its generalization and adaptability. In this paper, we analyze the variations in the modality gap during the fine-tuning of vision-language pre-trained models. Our observations reveal that the modality gap effectively reflects the extent to which pre-trained knowledge is preserved. Based on these insights, we propose a simple yet effective method, MG-CLIP, that improves CLIP's performance in class-incremental learning. Our approach leverages modality gap preservation to mitigate forgetting and modality gap compensation to enhance the capacity for new data, introducing a novel modality-gap-based perspective for continual learning. Extensive experiments on multiple benchmarks demonstrate that our method outperforms existing approaches without requiring additional replay data. Our code is available at https://github.com/linlany/MindtheGap.", "date": "2025-07-12", "categories": []}
{"id": "2507.09177", "title": "Continual Reinforcement Learning by Planning with Online World Models", "abstract": "Continual reinforcement learning (CRL) refers to a naturalistic setting where an agent needs to endlessly evolve, by trial and error, to solve multiple tasks that are presented sequentially. One of the largest obstacles to CRL is that the agent may forget how to solve previous tasks when learning a new task, known as catastrophic forgetting. In this paper, we propose to address this challenge by planning with online world models. Specifically, we learn a Follow-The-Leader shallow model online to capture the world dynamics, in which we plan using model predictive control to solve a set of tasks specified by any reward functions. The online world model is immune to forgetting by construction with a proven regret bound of $\\mathcal{O}(\\sqrt{K^2D\\log(T)})$ under mild assumptions. The planner searches actions solely based on the latest online model, thus forming a FTL Online Agent (OA) that updates incrementally. To assess OA, we further design Continual Bench, a dedicated environment for CRL, and compare with several strong baselines under the same model-planning algorithmic framework. The empirical results show that OA learns continuously to solve new tasks while not forgetting old skills, outperforming agents built on deep world models with various continual learning techniques.", "date": "2025-07-12", "categories": []}
{"id": "2507.09471", "title": "CKAA: Cross-subspace Knowledge Alignment and Aggregation for Robust Continual Learning", "abstract": "Continual Learning (CL) empowers AI models to continuously learn from sequential task streams. Recently, parameter-efficient fine-tuning (PEFT)-based CL methods have garnered increasing attention due to their superior performance. They typically allocate a unique sub-module for learning each task, with a task recognizer to select the appropriate sub-modules for testing images. However, due to the feature subspace misalignment from independently trained sub-modules, these methods tend to produce ambiguous decisions under misleading task-ids. To address this, we propose Cross-subspace Knowledge Alignment and Aggregation (CKAA), a novel framework that enhances model robustness against misleading task-ids through two key innovations: (1) Dual-level Knowledge Alignment (DKA): By aligning intra-class feature distributions across different subspaces and learning a robust global classifier through a feature simulation process, DKA enables the model to distinguish features from both correct and incorrect subspaces during training. (2) Task-Confidence-guided Mixture of Adapters (TC-MoA): A robust inference scheme that adaptively aggregates task-specific knowledge from relevant sub-modules based on task-confidence scores, avoiding overconfidence in misleading task-id predictions. Extensive experiments demonstrate that CKAA outperforms existing PEFT-based CL methods.", "date": "2025-07-13", "categories": []}
{"id": "2507.10485", "title": "Overcoming catastrophic forgetting in neural networks", "abstract": "Catastrophic forgetting is the primary challenge that hinders continual learning, which refers to a neural network ability to sequentially learn multiple tasks while retaining previously acquired knowledge. Elastic Weight Consolidation, a regularization-based approach inspired by synaptic consolidation in biological neural systems, has been used to overcome this problem. In this study prior research is replicated and extended by evaluating EWC in supervised learning settings using the PermutedMNIST and RotatedMNIST benchmarks. Through systematic comparisons with L2 regularization and stochastic gradient descent (SGD) without regularization, we analyze how different approaches balance knowledge retention and adaptability. Our results confirm what was shown in previous research, showing that EWC significantly reduces forgetting compared to naive training while slightly compromising learning efficiency on new tasks. Moreover, we investigate the impact of dropout regularization and varying hyperparameters, offering insights into the generalization of EWC across diverse learning scenarios. These results underscore EWC's potential as a viable solution for lifelong learning in neural networks.", "date": "2025-07-14", "categories": []}
{"id": "2411.09250", "title": "Embedding Space Allocation with Angle-Norm Joint Classifiers for Few-Shot Class-Incremental Learning", "abstract": "Few-shot class-incremental learning (FSCIL) aims to continually learn new classes from only a few samples without forgetting previous ones, requiring intelligent agents to adapt to dynamic environments. FSCIL combines the characteristics and challenges of class-incremental learning and few-shot learning: (i) Current classes occupy the entire feature space, which is detrimental to learning new classes. (ii) The small number of samples in incremental rounds is insufficient for fully training. In existing mainstream virtual class methods, for addressing the challenge (i), they attempt to use virtual classes as placeholders. However, new classes may not necessarily align with the virtual classes. For the challenge (ii), they replace trainable fully connected layers with Nearest Class Mean (NCM) classifiers based on cosine similarity, but NCM classifiers do not account for sample imbalance issues. To address these issues in previous methods, we propose the class-center guided embedding Space Allocation with Angle-Norm joint classifiers (SAAN) learning framework, which provides balanced space for all classes and leverages norm differences caused by sample imbalance to enhance classification criteria. Specifically, for challenge (i), SAAN divides the feature space into multiple subspaces and allocates a dedicated subspace for each session by guiding samples with the pre-set category centers. For challenge (ii), SAAN establishes a norm distribution for each class and generates angle-norm joint logits. Experiments demonstrate that SAAN can achieve state-of-the-art performance and it can be directly embedded into other SOTA methods as a plug-in, further enhancing their performance.", "date": "2024-11-14", "categories": []}
{"id": "2507.08736", "title": "Catastrophic Forgetting Mitigation Through Plateau Phase Activity Profiling", "abstract": "Catastrophic forgetting in deep neural networks occurs when learning new tasks degrades performance on previously learned tasks due to knowledge overwriting. Among the approaches to mitigate this issue, regularization techniques aim to identify and constrain \"important\" parameters to preserve previous knowledge. In the highly nonconvex optimization landscape of deep learning, we propose a novel perspective: tracking parameters during the final training plateau is more effective than monitoring them throughout the entire training process. We argue that parameters that exhibit higher activity (movement and variability) during this plateau reveal directions in the loss landscape that are relatively flat, making them suitable for adaptation to new tasks while preserving knowledge from previous ones. Our comprehensive experiments demonstrate that this approach achieves superior performance in balancing catastrophic forgetting mitigation with strong performance on newly learned tasks.", "date": "2025-07-11", "categories": []}
{"id": "2402.04129", "title": "OVOR: OnePrompt with Virtual Outlier Regularization for Rehearsal-Free Class-Incremental Learning", "abstract": "Recent works have shown that by using large pre-trained models along with learnable prompts, rehearsal-free methods for class-incremental learning (CIL) settings can achieve superior performance to prominent rehearsal-based ones. Rehearsal-free CIL methods struggle with distinguishing classes from different tasks, as those are not trained together. In this work we propose a regularization method based on virtual outliers to tighten decision boundaries of the classifier, such that confusion of classes among different tasks is mitigated. Recent prompt-based methods often require a pool of task-specific prompts, in order to prevent overwriting knowledge of previous tasks with that of the new task, leading to extra computation in querying and composing an appropriate prompt from the pool. This additional cost can be eliminated, without sacrificing accuracy, as we reveal in the paper. We illustrate that a simplified prompt-based method can achieve results comparable to previous state-of-the-art (SOTA) methods equipped with a prompt pool, using much less learnable parameters and lower inference cost. Our regularization method has demonstrated its compatibility with different prompt-based methods, boosting those previous SOTA rehearsal-free CIL methods' accuracy on the ImageNet-R and CIFAR-100 benchmarks. Our source code is available at https://github.com/jpmorganchase/ovor.", "date": "2024-02-06", "categories": []}
{"id": "2411.15469", "title": "Mamba-CL: Optimizing Selective State Space Model in Null Space for Continual Learning", "abstract": "Continual Learning (CL) aims to equip AI models with the ability to learn a sequence of tasks over time, without forgetting previously learned knowledge. Recently, State Space Models (SSMs), particularly the Mamba model, have achieved notable success in computer vision. Building on the strengths of SSMs, this study explores leveraging the Mamba model for CL. Therefore, we introduce Mamba-CL, a framework that continuously fine-tunes the core SSMs of the large-scale Mamba foundation model by updating parameters orthogonal to the feature subspace of previous tasks. This approach theoretically guarantees the consistency objective aiming to preserves consistent output for each SSM module across both previous and current tasks, so as to overcome catastrophic forgetting issue. Specifically, we achieve this goal by deducing the overall consistency constraints on four key time-invariant parameters in the Mamba model, streamlining its recurrent state-space structure and non-linear discretization process in SSM. In practice, we apply the null-space projection to efficiently implement the orthogonality within Mamba model. Extensive experiments on four class-incremental benchmarks demonstrate the effectiveness of Mamba-CL for anti-forgetting, achieving superior performances to state-of-the-art methods. Code is available in the supplementary materials.", "date": "2024-11-23", "categories": []}
{"id": "2507.07831", "title": "Rethinking Query-based Transformer for Continual Image Segmentation", "abstract": "Class-incremental/Continual image segmentation (CIS) aims to train an image segmenter in stages, where the set of available categories differs at each stage. To leverage the built-in objectness of query-based transformers, which mitigates catastrophic forgetting of mask proposals, current methods often decouple mask generation from the continual learning process. This study, however, identifies two key issues with decoupled frameworks: loss of plasticity and heavy reliance on input data order. To address these, we conduct an in-depth investigation of the built-in objectness and find that highly aggregated image features provide a shortcut for queries to generate masks through simple feature alignment. Based on this, we propose SimCIS, a simple yet powerful baseline for CIS. Its core idea is to directly select image features for query assignment, ensuring \"perfect alignment\" to preserve objectness, while simultaneously allowing queries to select new classes to promote plasticity. To further combat catastrophic forgetting of categories, we introduce cross-stage consistency in selection and an innovative \"visual query\"-based replay mechanism. Experiments demonstrate that SimCIS consistently outperforms state-of-the-art methods across various segmentation tasks, settings, splits, and input data orders. All models and codes will be made publicly available at https://github.com/SooLab/SimCIS.", "date": "2025-07-10", "categories": []}
{"id": "2310.11482", "title": "Enhancing Plasticity for First Session Adaptation Continual Learning", "abstract": "The integration of large pre-trained models (PTMs) into Class-Incremental Learning (CIL) has facilitated the development of computationally efficient strategies such as First-Session Adaptation (FSA), which fine-tunes the model solely on the first task while keeping it frozen for subsequent tasks. Although effective in homogeneous task sequences, these approaches struggle when faced with the heterogeneity of real-world task distributions. We introduce Plasticity-Enhanced Test-Time Adaptation in Class-Incremental Learning (PLASTIC), a method that reinstates plasticity in CIL while preserving model stability. PLASTIC leverages Test-Time Adaptation (TTA) by dynamically fine-tuning LayerNorm parameters on unlabeled test data, enabling adaptability to evolving tasks and improving robustness against data corruption. To prevent TTA-induced model divergence and maintain stable learning across tasks, we introduce a teacher-student distillation framework, ensuring that adaptation remains controlled and generalizable. Extensive experiments across multiple benchmarks demonstrate that PLASTIC consistently outperforms both conventional and state-of-the-art PTM-based CIL approaches, while also exhibiting inherent robustness to data corruptions. Code is available at: https://github.com/IemProg/PLASTIC.", "date": "2023-10-17", "categories": []}
{"id": "2507.07100", "title": "Addressing Imbalanced Domain-Incremental Learning through Dual-Balance Collaborative Experts", "abstract": "Domain-Incremental Learning (DIL) focuses on continual learning in non-stationary environments, requiring models to adjust to evolving domains while preserving historical knowledge. DIL faces two critical challenges in the context of imbalanced data: intra-domain class imbalance and cross-domain class distribution shifts. These challenges significantly hinder model performance, as intra-domain imbalance leads to underfitting of few-shot classes, while cross-domain shifts require maintaining well-learned many-shot classes and transferring knowledge to improve few-shot class performance in old domains. To overcome these challenges, we introduce the Dual-Balance Collaborative Experts (DCE) framework. DCE employs a frequency-aware expert group, where each expert is guided by specialized loss functions to learn features for specific frequency groups, effectively addressing intra-domain class imbalance. Subsequently, a dynamic expert selector is learned by synthesizing pseudo-features through balanced Gaussian sampling from historical class statistics. This mechanism navigates the trade-off between preserving many-shot knowledge of previous domains and leveraging new data to improve few-shot class performance in earlier tasks. Extensive experimental results on four benchmark datasets demonstrate DCE's state-of-the-art performance.", "date": "2025-07-09", "categories": []}
{"id": "2503.13575", "title": "Analytic Subspace Routing: How Recursive Least Squares Works in Continual Learning of Large Language Model", "abstract": "Large Language Models (LLMs) possess encompassing capabilities that can process diverse language-related tasks. However, finetuning on LLMs will diminish this general skills and continual finetuning will further cause severe degradation on accumulated knowledge. Recently, Continual Learning (CL) in Large Language Models (LLMs) arises which aims to continually adapt the LLMs to new tasks while maintaining previously learned knowledge and inheriting general skills. Existing techniques either leverage previous data to replay, leading to extra computational costs, or utilize a single parameter-efficient module to learn the downstream task, constraining new knowledge absorption with interference between different tasks. Toward these issues, this paper proposes Analytic Subspace Routing(ASR) to address these challenges. For each task, we isolate the learning within a subspace of deep layers' features via low-rank adaptation, eliminating knowledge interference between different tasks. Additionally, we propose an analytic routing mechanism to properly utilize knowledge learned in different subspaces. Our approach employs Recursive Least Squares to train a multi-task router model, allowing the router to dynamically adapt to incoming data without requiring access to historical data. Also, the router effectively assigns the current task to an appropriate subspace and has a non-forgetting property of previously learned tasks with a solid theoretical guarantee. Experimental results demonstrate that our method achieves near-perfect retention of prior knowledge while seamlessly integrating new information, effectively overcoming the core limitations of existing methods. Our code will be released after acceptance.", "date": "2025-03-17", "categories": []}
{"id": "2507.02395", "title": "Continual Multiple Instance Learning with Enhanced Localization for Histopathological Whole Slide Image Analysis", "abstract": "Multiple instance learning (MIL) significantly reduced annotation costs via bag-level weak labels for large-scale images, such as histopathological whole slide images (WSIs). However, its adaptability to continual tasks with minimal forgetting has been rarely explored, especially on instance classification for localization. Weakly incremental learning for semantic segmentation has been studied for continual localization, but it focused on natural images, leveraging global relationships among hundreds of small patches (e.g., $16 \\times 16$) using pre-trained models. This approach seems infeasible for MIL localization due to enormous amounts ($\\sim 10^5$) of large patches (e.g., $256 \\times 256$) and no available global relationships such as cancer cells. To address these challenges, we propose Continual Multiple Instance Learning with Enhanced Localization (CoMEL), an MIL framework for both localization and adaptability with minimal forgetting. CoMEL consists of (1) Grouped Double Attention Transformer (GDAT) for efficient instance encoding, (2) Bag Prototypes-based Pseudo-Labeling (BPPL) for reliable instance pseudo-labeling, and (3) Orthogonal Weighted Low-Rank Adaptation (OWLoRA) to mitigate forgetting in both bag and instance classification. Extensive experiments on three public WSI datasets demonstrate superior performance of CoMEL, outperforming the prior arts by up to $11.00\\%$ in bag-level accuracy and up to $23.4\\%$ in localization accuracy under the continual MIL setup.", "date": "2025-07-03", "categories": []}
{"id": "2507.05939", "title": "Remember Past, Anticipate Future: Learning Continual Multimodal Misinformation Detectors", "abstract": "Nowadays, misinformation articles, especially multimodal ones, are widely spread on social media platforms and cause serious negative effects. To control their propagation, Multimodal Misinformation Detection (MMD) becomes an active topic in the community to automatically identify misinformation. Previous MMD methods focus on supervising detectors by collecting offline data. However, in real-world scenarios, new events always continually emerge, making MMD models trained on offline data consistently outdated and ineffective. To address this issue, training MMD models under online data streams is an alternative, inducing an emerging task named continual MMD. Unfortunately, it is hindered by two major challenges. First, training on new data consistently decreases the detection performance on past data, named past knowledge forgetting. Second, the social environment constantly evolves over time, affecting the generalization on future data. To alleviate these challenges, we propose to remember past knowledge by isolating interference between event-specific parameters with a Dirichlet process-based mixture-of-expert structure, and anticipate future environmental distributions by learning a continuous-time dynamics model. Accordingly, we induce a new continual MMD method DAEDCMD. Extensive experiments demonstrate that DAEDCMD can consistently and significantly outperform the compared methods, including six MMD baselines and three continual learning methods.", "date": "2025-07-08", "categories": []}
{"id": "2303.06572", "title": "Continual Visual Reinforcement Learning with A Life-Long World Model", "abstract": "Learning physical dynamics in a series of non-stationary environments is a challenging but essential task for model-based reinforcement learning (MBRL) with visual inputs. It requires the agent to consistently adapt to novel tasks without forgetting previous knowledge. In this paper, we present a new continual learning approach for visual dynamics modeling and explore its efficacy in visual control. The key assumption is that an ideal world model can provide a non-forgetting environment simulator, which enables the agent to optimize the policy in a multi-task learning manner based on the imagined trajectories from the world model. To this end, we first introduce the life-long world model, which learns task-specific latent dynamics using a mixture of Gaussians and incorporates generative experience replay to mitigate catastrophic forgetting. Then, we further address the value estimation challenge for previous tasks with the exploratory-conservative behavior learning approach. Our model remarkably outperforms the straightforward combinations of existing continual learning and visual RL algorithms on DeepMind Control Suite and Meta-World benchmarks with continual visual control tasks.", "date": "2023-03-12", "categories": []}
{"id": "2406.02480", "title": "Fairness Evolution in Continual Learning for Medical Imaging", "abstract": "Deep Learning has advanced significantly in medical applications, aiding disease diagnosis in Chest X-ray images. However, expanding model capabilities with new data remains a challenge, which Continual Learning (CL) aims to address. Previous studies have evaluated CL strategies based on classification performance; however, in sensitive domains such as healthcare, it is crucial to assess performance across socially salient groups to detect potential biases. This study examines how bias evolves across tasks using domain-specific fairness metrics and how different CL strategies impact this evolution. Our results show that Learning without Forgetting and Pseudo-Label achieve optimal classification performance, but Pseudo-Label is less biased.", "date": "2024-04-10", "categories": []}
{"id": "2406.09935", "title": "Predicting the Susceptibility of Examples to Catastrophic Forgetting", "abstract": "Catastrophic forgetting - the tendency of neural networks to forget previously learned data when learning new information - remains a central challenge in continual learning. In this work, we adopt a behavioral approach, observing a connection between learning speed and forgetting: examples learned more quickly are less prone to forgetting. Focusing on replay-based continual learning, we show that the composition of the replay buffer - specifically, whether it contains quickly or slowly learned examples - has a significant effect on forgetting. Motivated by this insight, we introduce Speed-Based Sampling (SBS), a simple yet general strategy that selects replay examples based on their learning speed. SBS integrates easily into existing buffer-based methods and improves performance across a wide range of competitive continual learning benchmarks, advancing state-of-the-art results. Our findings underscore the value of accounting for the forgetting dynamics when designing continual learning algorithms.", "date": "2024-06-14", "categories": []}
{"id": "2411.15235", "title": "CODE-CL: Conceptor-Based Gradient Projection for Deep Continual Learning", "abstract": "Continual learning (CL) - the ability to progressively acquire and integrate new concepts - is essential to intelligent systems to adapt to dynamic environments. However, deep neural networks struggle with catastrophic forgetting (CF) when learning tasks sequentially, as training for new tasks often overwrites previously learned knowledge. To address this, recent approaches constrain updates to orthogonal subspaces using gradient projection, effectively preserving important gradient directions for previous tasks. While effective in reducing forgetting, these approaches inadvertently hinder forward knowledge transfer (FWT), particularly when tasks are highly correlated. In this work, we propose Conceptor-based gradient projection for Deep Continual Learning (CODE-CL), a novel method that leverages conceptor matrix representations, a form of regularized reconstruction, to adaptively handle highly correlated tasks. CODE-CL mitigates CF by projecting gradients onto pseudo-orthogonal subspaces of previous task feature spaces while simultaneously promoting FWT. It achieves this by learning a linear combination of shared basis directions, allowing efficient balance between stability and plasticity and transfer of knowledge between overlapping input feature representations. Extensive experiments on continual learning benchmarks validate CODE-CL's efficacy, demonstrating superior performance, reduced forgetting, and improved FWT as compared to state-of-the-art methods.", "date": "2024-11-21", "categories": []}
{"id": "2503.20612", "title": "IAP: Improving Continual Learning of Vision-Language Models via Instance-Aware Prompting", "abstract": "Recent pre-trained vision-language models (PT-VLMs) often face a Multi-Domain Task Incremental Learning (MTIL) scenario in practice, where several classes and domains of multi-modal tasks are incrementally arrived. Without access to previously seen tasks and unseen tasks, memory-constrained MTIL suffers from forward and backward forgetting. To alleviate the above challenges, parameter-efficient fine-tuning techniques (PEFT), such as prompt tuning, are employed to adapt the PT-VLM to the diverse incrementally learned tasks. To achieve effective new task adaptation, existing methods only consider the effect of PEFT strategy selection, but neglect the influence of PEFT parameter setting (e.g., prompting). In this paper, we tackle the challenge of optimizing prompt designs for diverse tasks in MTIL and propose an Instance-Aware Prompting (IAP) framework. Specifically, our Instance-Aware Gated Prompting (IA-GP) strategy enhances adaptation to new tasks while mitigating forgetting by adaptively assigning prompts across transformer layers at the instance level. Our Instance-Aware Class-Distribution-Driven Prompting (IA-CDDP) improves the task adaptation process by determining an accurate task-label-related confidence score for each instance. Experimental evaluations across 11 datasets, using three performance metrics, demonstrate the effectiveness of our proposed method. The source codes are available at https://github.com/FerdinandZJU/IAP.", "date": "2025-03-26", "categories": []}
{"id": "2504.20282", "title": "FedCCL: Federated Clustered Continual Learning Framework for Privacy-focused Energy Forecasting", "abstract": "Privacy-preserving distributed model training is crucial for modern machine learning applications, yet existing Federated Learning approaches struggle with heterogeneous data distributions and varying computational capabilities. Traditional solutions either treat all participants uniformly or require costly dynamic clustering during training, leading to reduced efficiency and delayed model specialization. We present FedCCL (Federated Clustered Continual Learning), a framework specifically designed for environments with static organizational characteristics but dynamic client availability. By combining static pre-training clustering with an adapted asynchronous FedAvg algorithm, FedCCL enables new clients to immediately profit from specialized models without prior exposure to their data distribution, while maintaining reduced coordination overhead and resilience to client disconnections. Our approach implements an asynchronous Federated Learning protocol with a three-tier model topology - global, cluster-specific, and local models - that efficiently manages knowledge sharing across heterogeneous participants. Evaluation using photovoltaic installations across central Europe demonstrates that FedCCL's location-based clustering achieves an energy prediction error of 3.93% (+-0.21%), while maintaining data privacy and showing that the framework maintains stability for population-independent deployments, with 0.14 percentage point degradation in performance for new installations. The results demonstrate that FedCCL offers an effective framework for privacy-preserving distributed learning, maintaining high accuracy and adaptability even with dynamic participant populations.", "date": "2025-04-28", "categories": []}
{"id": "2507.02503", "title": "Continual Gradient Low-Rank Projection Fine-Tuning for LLMs", "abstract": "Continual fine-tuning of Large Language Models (LLMs) is hampered by the trade-off between efficiency and expressiveness. Low-Rank Adaptation (LoRA) offers efficiency but constrains the model's ability to learn new tasks and transfer knowledge due to its low-rank nature and reliance on explicit parameter constraints. We propose GORP (Gradient LOw Rank Projection) for Continual Learning, a novel training strategy that overcomes these limitations by synergistically combining full and low-rank parameters and jointly updating within a unified low-rank gradient subspace. GORP expands the optimization space while preserving efficiency and mitigating catastrophic forgetting. Extensive experiments on continual learning benchmarks demonstrate GORP's superior performance compared to existing state-of-the-art approaches. Code is available at https://github.com/Wcxwcxw/GORP.", "date": "2025-07-03", "categories": []}
{"id": "2507.03045", "title": "Optimisation Is Not What You Need", "abstract": "The Artificial Intelligence field has focused on developing optimisation methods to solve multiple problems, specifically problems that we thought to be only solvable through cognition. The obtained results have been outstanding, being able to even surpass the Turing Test. However, we have found that these optimisation methods share some fundamental flaws that impede them to become a true artificial cognition. Specifically, the field have identified catastrophic forgetting as a fundamental problem to develop such cognition. This paper formally proves that this problem is inherent to optimisation methods, and as such it will always limit approaches that try to solve the Artificial General Intelligence problem as an optimisation problem. Additionally, it addresses the problem of overfitting and discuss about other smaller problems that optimisation methods pose. Finally, it empirically shows how world-modelling methods avoid suffering from either problem. As a conclusion, the field of Artificial Intelligence needs to look outside the machine learning field to find methods capable of developing an artificial cognition.", "date": "2025-07-03", "categories": []}
{"id": "2507.04020", "title": "Exploring Kolmogorov-Arnold Network Expansions in Vision Transformers for Mitigating Catastrophic Forgetting in Continual Learning", "abstract": "Continual learning (CL), the ability of a model to learn new tasks without forgetting previously acquired knowledge, remains a critical challenge in artificial intelligence, particularly for vision transformers (ViTs) utilizing Multilayer Perceptrons (MLPs) for global representation learning. Catastrophic forgetting, where new information overwrites prior knowledge, is especially problematic in these models. This research proposes replacing MLPs in ViTs with Kolmogorov-Arnold Network (KANs) to address this issue. KANs leverage local plasticity through spline-based activations, ensuring that only a subset of parameters is updated per sample, thereby preserving previously learned knowledge. The study investigates the efficacy of KAN-based ViTs in CL scenarios across benchmark datasets (MNIST, CIFAR100), focusing on their ability to retain accuracy on earlier tasks while adapting to new ones. Experimental results demonstrate that KAN-based ViTs significantly mitigate catastrophic forgetting, outperforming traditional MLP-based ViTs in knowledge retention and task adaptation. This novel integration of KANs into ViTs represents a promising step toward more robust and adaptable models for dynamic environments.", "date": "2025-07-05", "categories": []}
{"id": "2412.11074", "title": "Adapter-Enhanced Semantic Prompting for Continual Learning", "abstract": "Continual learning (CL) enables models to adapt to evolving data streams. A major challenge of CL is catastrophic forgetting, where new knowledge will overwrite previously acquired knowledge. Traditional methods usually retain the past data for replay or add additional branches in the model to learn new knowledge, which has high memory requirements. In this paper, we propose a novel lightweight CL framework, Adapter-Enhanced Semantic Prompting (AESP), which integrates prompt tuning and adapter techniques. Specifically, we design semantic-guided prompts to enhance the generalization ability of visual features and utilize adapters to efficiently fuse the semantic information, aiming to learn more adaptive features for the continual learning task. Furthermore, to choose the right task prompt for feature adaptation, we have developed a novel matching mechanism for prompt selection. Extensive experiments on three CL datasets demonstrate that our approach achieves favorable performance across multiple metrics, showing its potential for advancing CL.", "date": "2024-12-15", "categories": []}
{"id": "2507.02103", "title": "What Neuroscience Can Teach AI About Learning in Continuously Changing Environments", "abstract": "Modern AI models, such as large language models, are usually trained once on a huge corpus of data, potentially fine-tuned for a specific task, and then deployed with fixed parameters. Their training is costly, slow, and gradual, requiring billions of repetitions. In stark contrast, animals continuously adapt to the ever-changing contingencies in their environments. This is particularly important for social species, where behavioral policies and reward outcomes may frequently change in interaction with peers. The underlying computational processes are often marked by rapid shifts in an animal's behaviour and rather sudden transitions in neuronal population activity. Such computational capacities are of growing importance for AI systems operating in the real world, like those guiding robots or autonomous vehicles, or for agentic AI interacting with humans online. Can AI learn from neuroscience? This Perspective explores this question, integrating the literature on continual and in-context learning in AI with the neuroscience of learning on behavioral tasks with shifting rules, reward probabilities, or outcomes. We will outline an agenda for how specifically insights from neuroscience may inform current developments in AI in this area, and - vice versa - what neuroscience may learn from AI, contributing to the evolving field of NeuroAI.", "date": "2025-07-02", "categories": []}
{"id": "2507.01559", "title": "How Weight Resampling and Optimizers Shape the Dynamics of Continual Learning and Forgetting in Neural Networks", "abstract": "Recent work in continual learning has highlighted the beneficial effect of resampling weights in the last layer of a neural network (``zapping\"). Although empirical results demonstrate the effectiveness of this approach, the underlying mechanisms that drive these improvements remain unclear. In this work, we investigate in detail the pattern of learning and forgetting that take place inside a convolutional neural network when trained in challenging settings such as continual learning and few-shot transfer learning, with handwritten characters and natural images. Our experiments show that models that have undergone zapping during training more quickly recover from the shock of transferring to a new domain. Furthermore, to better observe the effect of continual learning in a multi-task setting we measure how each individual task is affected. This shows that, not only zapping, but the choice of optimizer can also deeply affect the dynamics of learning and forgetting, causing complex patterns of synergy/interference between tasks to emerge when the model learns sequentially at transfer time.", "date": "2025-07-02", "categories": []}
{"id": "2411.13949", "title": "SMoLoRA: Exploring and Defying Dual Catastrophic Forgetting in Continual Visual Instruction Tuning", "abstract": "Visual instruction tuning (VIT) enables multimodal large language models (MLLMs) to effectively handle a wide range of vision tasks by framing them as language-based instructions. Building on this, continual visual instruction tuning (CVIT) extends the capability of MLLMs to incrementally learn new tasks, accommodating evolving functionalities. While prior work has advanced CVIT through the development of new benchmarks and approaches to mitigate catastrophic forgetting, these efforts largely follow traditional continual learning paradigms, neglecting the unique challenges specific to CVIT. We identify a dual form of catastrophic forgetting in CVIT, where MLLMs not only forget previously learned visual understanding but also experience a decline in instruction following abilities as they acquire new tasks. To address this, we introduce the Separable Mixture of Low-Rank Adaptation (SMoLoRA) framework, which employs separable routing through two distinct modules-one for visual understanding and another for instruction following. This dual-routing design enables specialized adaptation in both domains, preventing forgetting while improving performance. Furthermore, we propose a new CVIT benchmark that goes beyond existing benchmarks by additionally evaluating a model's ability to generalize to unseen tasks and handle diverse instructions across various tasks. Extensive experiments demonstrate that SMoLoRA outperforms existing methods in mitigating dual forgetting, improving generalization to unseen tasks, and ensuring robustness in following diverse instructions. Code is available at https://github.com/Minato-Zackie/SMoLoRA.", "date": "2024-11-21", "categories": []}
{"id": "2503.06385", "title": "A Good Start Matters: Enhancing Continual Learning with Data-Driven Weight Initialization", "abstract": "To adapt to real-world data streams, continual learning (CL) systems must rapidly learn new concepts while preserving and utilizing prior knowledge. When it comes to adding new information to continually-trained deep neural networks (DNNs), classifier weights for newly encountered categories are typically initialized randomly, leading to high initial training loss (spikes) and instability. Consequently, achieving optimal convergence and accuracy requires prolonged training, increasing computational costs. Inspired by Neural Collapse (NC), we propose a weight initialization strategy to improve learning efficiency in CL. In DNNs trained with mean-squared-error, NC gives rise to a Least-Square (LS) classifier in the last layer, whose weights can be analytically derived from learned features. We leverage this LS formulation to initialize classifier weights in a data-driven manner, aligning them with the feature distribution rather than using random initialization. Our method mitigates initial loss spikes and accelerates adaptation to new tasks. We evaluate our approach in large-scale CL settings, demonstrating faster adaptation and improved CL performance.", "date": "2025-03-09", "categories": []}
{"id": "2507.00014", "title": "SWE-Bench-CL: Continual Learning for Coding Agents", "abstract": "Large Language Models (LLMs) have achieved impressive results on static code-generation benchmarks, but real-world software development unfolds as a continuous stream of evolving issues, fixes, and feature requests. We introduce SWE-Bench-CL, a novel continual learning benchmark built on the human-verified SWE-Bench Verified dataset introduced by OpenAI and Princeton-NLP in 2024. By organizing GitHub issues into chronologically ordered sequences that reflect natural repository evolution, SWE-Bench-CL enables direct evaluation of an agent's ability to accumulate experience, transfer knowledge across tasks, and resist catastrophic forgetting. We complement the dataset with (i) a preliminary analysis of inter-task structural similarity and contextual sensitivity, (ii) an interactive LangGraph-based evaluation framework augmented with a FAISS-backed semantic memory module, and (iii) a suite of specialized continual learning metrics -- including average accuracy, forgetting, forward/backward transfer, tool-use efficiency, and a generalized Composite Continual Learning Score and CL-F-beta score -- to capture the stability-plasticity trade-off. We outline a rigorous experimental protocol comparing memory-enabled and memory-disabled agents across diverse Python repositories. All code and data are publicly available at https://github.com/thomasjoshi/agents-never-forget, providing the community with a reproducible platform for developing more adaptive and robust AI agents in software engineering.", "date": "2025-06-13", "categories": []}
{"id": "2507.00042", "title": "Catastrophic Forgetting Mitigation via Discrepancy-Weighted Experience Replay", "abstract": "Continually adapting edge models in cloud-edge collaborative object detection for traffic monitoring suffers from catastrophic forgetting, where models lose previously learned knowledge when adapting to new data distributions. This is especially problematic in dynamic traffic environments characterised by periodic variations (e.g., day/night, peak hours), where past knowledge remains valuable. Existing approaches like experience replay and visual prompts offer some mitigation, but struggle to effectively prioritize and leverage historical data for optimal knowledge retention and adaptation. Specifically, simply storing and replaying all historical data can be inefficient, while treating all historical experiences as equally important overlooks their varying relevance to the current domain. This paper proposes ER-EMU, an edge model update algorithm based on adaptive experience replay, to address these limitations. ER-EMU utilizes a limited-size experience buffer managed using a First-In-First-Out (FIFO) principle, and a novel Domain Distance Metric-based Experience Selection (DDM-ES) algorithm. DDM-ES employs the multi-kernel maximum mean discrepancy (MK-MMD) to quantify the dissimilarity between target domains, prioritizing the selection of historical data that is most dissimilar to the current target domain. This ensures training diversity and facilitates the retention of knowledge from a wider range of past experiences, while also preventing overfitting to the new domain. The experience buffer is also updated using a simple random sampling strategy to maintain a balanced representation of previous domains. Experiments on the Bellevue traffic video dataset, involving repeated day/night cycles, demonstrate that ER-EMU consistently improves the performance of several state-of-the-art cloud-edge collaborative object detection frameworks.", "date": "2025-06-23", "categories": []}
{"id": "2507.00469", "title": "Bisecle: Binding and Separation in Continual Learning for Video Language Understanding", "abstract": "Frontier vision-language models (VLMs) have made remarkable improvements in video understanding tasks. However, real-world videos typically exist as continuously evolving data streams (e.g., dynamic scenes captured by wearable glasses), necessitating models to continually adapt to shifting data distributions and novel scenarios. Considering the prohibitive computational costs of fine-tuning models on new tasks, usually, a small subset of parameters is updated while the bulk of the model remains frozen. This poses new challenges to existing continual learning frameworks in the context of large multimodal foundation models, i.e., catastrophic forgetting and update conflict. While the foundation models struggle with parameter-efficient continual learning, the hippocampus in the human brain has evolved highly efficient mechanisms for memory formation and consolidation. Inspired by the rapid Binding and pattern separation mechanisms in the hippocampus, in this work, we propose Bisecle for video-language continual learning, where a multi-directional supervision module is used to capture more cross-modal relationships and a contrastive prompt learning scheme is designed to isolate task-specific knowledge to facilitate efficient memory storage. Binding and separation processes further strengthen the ability of VLMs to retain complex experiences, enabling robust and efficient continual learning in video understanding tasks. We perform a thorough evaluation of the proposed Bisecle, demonstrating its ability to mitigate forgetting and enhance cross-task generalization on several VideoQA benchmarks.", "date": "2025-07-01", "categories": []}
{"id": "2507.00534", "title": "NIRANTAR: Continual Learning with New Languages and Domains on Real-world Speech Data", "abstract": "We introduce Nirantar, a comprehensive framework for evaluating continual learning (CL) in multilingual and multi-domain ASR. Designed to reflect real-world CL challenges, Nirantar leverages data collected incrementally across 22 languages and 208 districts in India through natural episodes. This enables evaluation across Language-Incremental (LIL), Domain-Incremental (DIL), and the novel Language-Incremental Domain-Incremental Learning (LIDIL) scenarios. Unlike prior work that relies on simulated episodes, Nirantar presents dynamic, non-uniform language and domain shifts, making it an ideal testbed for CL research. With 3250 hours of human-transcribed speech, including 1720 hours newly introduced in this work, our framework enables systematic benchmarking of CL methods. We evaluate existing approaches and demonstrate that no single method performs consistently well, underscoring the need for more robust CL strategies.", "date": "2025-07-01", "categories": []}
{"id": "2406.12593", "title": "PromptDSI: Prompt-based Rehearsal-free Continual Learning for Document Retrieval", "abstract": "Differentiable Search Index (DSI) utilizes pre-trained language models to perform indexing and document retrieval via end-to-end learning without relying on external indexes. However, DSI requires full re-training to index new documents, causing significant computational inefficiencies. Continual learning (CL) offers a solution by enabling the model to incrementally update without full re-training. Existing CL solutions in document retrieval rely on memory buffers or generative models for rehearsal, which is infeasible when accessing previous training data is restricted due to privacy concerns. To this end, we introduce PromptDSI, a prompt-based, rehearsal-free continual learning approach for document retrieval. PromptDSI follows the Prompt-based Continual Learning (PCL) framework, using learnable prompts to efficiently index new documents without accessing previous documents or queries. To improve retrieval latency, we remove the initial forward pass of PCL, which otherwise greatly increases training and inference time, with a negligible trade-off in performance. Additionally, we introduce a novel topic-aware prompt pool that employs neural topic embeddings as fixed keys, eliminating the instability of prompt key optimization while maintaining competitive performance with existing PCL prompt pools. In a challenging rehearsal-free continual learning setup, we demonstrate that PromptDSI variants outperform rehearsal-based baselines, match the strong cache-based baseline in mitigating forgetting, and significantly improving retrieval performance on new corpora.", "date": "2024-06-18", "categories": []}
{"id": "2407.21252", "title": "Lifelong Person Search", "abstract": "Person search is the task to localize a query person in gallery datasets of scene images. Existing methods have been mainly developed to handle a single target dataset only, however diverse datasets are continuously given in practical applications of person search. In such cases, they suffer from the catastrophic knowledge forgetting in the old datasets when trained on new datasets. In this paper, we first introduce a novel problem of lifelong person search (LPS) where the model is incrementally trained on the new datasets while preserving the knowledge learned in the old datasets. We propose an end-to-end LPS framework that facilitates the knowledge distillation to enforce the consistency learning between the old and new models by utilizing the prototype features of the foreground persons as well as the hard background proposals in the old domains. Moreover, we also devise the rehearsal-based instance matching to further improve the discrimination ability in the old domains by using the unlabeled person instances additionally. Experimental results demonstrate that the proposed method achieves significantly superior performance of both the detection and re-identification to preserve the knowledge learned in the old domains compared with the existing methods.", "date": "2024-07-31", "categories": []}
{"id": "2409.08516", "title": "AWF: Adaptive Weight Fusion for Enhanced Class Incremental Semantic Segmentation", "abstract": "Class Incremental Semantic Segmentation (CISS) aims to mitigate catastrophic forgetting by maintaining a balance between previously learned and newly introduced knowledge. Existing methods, primarily based on regularization techniques like knowledge distillation, help preserve old knowledge but often face challenges in effectively integrating new knowledge, resulting in limited overall improvement. Endpoints Weight Fusion (EWF) method, while simple, effectively addresses some of these limitations by dynamically fusing the model weights from previous steps with those from the current step, using a fusion parameter alpha determined by the relative number of previously known classes and newly introduced classes. However, the simplicity of the alpha calculation may limit its ability to fully capture the complexities of different task scenarios, potentially leading to suboptimal fusion outcomes. In this paper, we propose an enhanced approach called Adaptive Weight Fusion (AWF), which introduces an alternating training strategy for the fusion parameter, allowing for more flexible and adaptive weight integration. AWF achieves superior performance by better balancing the retention of old knowledge with the learning of new classes, significantly improving results on benchmark CISS tasks compared to the original EWF. And our experiment code will be released on Github.", "date": "2024-09-13", "categories": []}
{"id": "2409.18860", "title": "LW2G: Learning Whether to Grow for Prompt-based Continual Learning", "abstract": "Recent Prompt-based Continual learning (PCL) has achieved remarkable performance with pre-trained models. These approaches expand a prompt pool by adding a new set of prompts while learning and select the correct set during inference. Previous studies have revealed that learning task-wised prompt sets individually and low selection accuracy pose challenges to the performance of PCL. In this paper, we propose a plug-in method, $\\textbf{L}$earning $\\textbf{W}$hether $\\textbf{t}$o $\\textbf{G}$row $\\textbf{(LW2G)}$, which leverages the disparities between tasks to form an effective and efficient prompt sets pool, thereby achieving intra-task knowledge sharing and cooperation and avoiding the unbounded increase in the cost of the prompt pool. Specifically, a shared set is utilized when several tasks share certain commonalities, and a new set is added when there are significant differences between the new and previous tasks. To achieve this, we develop a metric called Hinder Forward Capability (HFC) to measure the hindrance imposed on learning new tasks by surgically modifying the original gradient onto the orthogonal complement of the old feature space. With HFC, an automated scheme, Dynamic Growing Approach, adaptively learns whether to grow with a dynamic threshold. Furthermore, we design a gradient-based constraint to ensure consistency between the updating prompts and pre-trained knowledge. Extensive experiments show the effectiveness of our method. Code is available at https://github.com/RAIAN08/LW2G.", "date": "2024-09-27", "categories": []}
{"id": "2407.06136", "title": "Mamba-FSCIL: Dynamic Adaptation with Selective State Space Model for Few-Shot Class-Incremental Learning", "abstract": "Few-shot class-incremental learning (FSCIL) aims to incrementally learn novel classes from limited examples while preserving knowledge of previously learned classes. Existing methods face a critical dilemma: static architectures rely on a fixed parameter space to learn from data that arrive sequentially, prone to overfitting to the current session, while dynamic architectures require the expansion of the parameter space continually, leading to increased complexity. In this study, we explore the potential of Selective State Space Models (SSMs) for FSCIL. Mamba leverages its input-dependent parameters to dynamically adjust its processing patterns and generate content-aware scan patterns within a fixed architecture. This enables it to configure distinct processing for base and novel classes, effectively preserving existing knowledge while adapting to new ones. To leverage Mamba's potential for FSCIL, we design two key modules: First, we propose a dual selective SSM projector that dynamically adjusts the projection parameters based on the intermediate features for dynamic adaptation. The dual-design structurally decouples base and novel class processing with a frozen base branch, employing a frozen base branch to maintain robust base-class features and a dynamic incremental branch that adaptively learns distinctive feature shifts for novel classes. Second, we develop a class-sensitive selective scan mechanism to guide dynamic adaptation of the incremental branch. It minimizes the disruption to base-class representations caused by training on novel data, and meanwhile, forces the selective scan to perform in distinct patterns between base and novel classes. Extensive experiments on miniImageNet, CUB-200, and CIFAR-100 demonstrate that Mamba-FSCIL achieves state-of-the-art performance. The code is available at https://github.com/xiaojieli0903/Mamba-FSCIL.", "date": "2024-07-08", "categories": []}
{"id": "2506.21899", "title": "Advancements and Challenges in Continual Reinforcement Learning: A Comprehensive Review", "abstract": "The diversity of tasks and dynamic nature of reinforcement learning (RL) require RL agents to be able to learn sequentially and continuously, a learning paradigm known as continuous reinforcement learning. This survey reviews how continual learning transforms RL agents into dynamic continual learners. This enables RL agents to acquire and retain useful and reusable knowledge seamlessly. The paper delves into fundamental aspects of continual reinforcement learning, exploring key concepts, significant challenges, and novel methodologies. Special emphasis is placed on recent advancements in continual reinforcement learning within robotics, along with a succinct overview of evaluation environments utilized in prominent research, facilitating accessibility for newcomers to the field. The review concludes with a discussion on limitations and promising future directions, providing valuable insights for researchers and practitioners alike.", "date": "2025-06-27", "categories": []}
{"id": "2210.09394", "title": "Review learning: Real world validation of privacy preserving continual learning across medical institutions", "abstract": "When a deep learning model is trained sequentially on different datasets, it often forgets the knowledge learned from previous data, a problem known as catastrophic forgetting. This damages the model's performance on diverse datasets, which is critical in privacy-preserving deep learning (PPDL) applications based on transfer learning (TL). To overcome this, we introduce \"review learning\" (RevL), a low cost continual learning algorithm for diagnosis prediction using electronic health records (EHR) within a PPDL framework. RevL generates data samples from the model which are used to review knowledge from previous datasets. Six simulated institutional experiments and one real-world experiment involving three medical institutions were conducted to validate RevL, using three binary classification EHR data. In the real-world experiment with data from 106,508 patients, the mean global area under the receiver operating curve was 0.710 for RevL and 0.655 for TL. These results demonstrate RevL's ability to retain previously learned knowledge and its effectiveness in real-world PPDL scenarios. Our work establishes a realistic pipeline for PPDL research based on model transfers across institutions and highlights the practicality of continual learning in real-world medical settings using private EHR data.", "date": "2022-10-17", "categories": []}
{"id": "2307.04345", "title": "Continual Learning as Computationally Constrained Reinforcement Learning", "abstract": "An agent that efficiently accumulates knowledge to develop increasingly sophisticated skills over a long lifetime could advance the frontier of artificial intelligence capabilities. The design of such agents, which remains a long-standing challenge of artificial intelligence, is addressed by the subject of continual learning. This monograph clarifies and formalizes concepts of continual learning, introducing a framework and set of tools to stimulate further research.", "date": "2023-07-10", "categories": []}
{"id": "2403.14684", "title": "Self-Regulated Neurogenesis for Online Data-Incremental Learning", "abstract": "Neural networks often struggle with catastrophic forgetting when learning sequences of tasks or data streams, unlike humans who can continuously learn and consolidate new concepts even in the absence of explicit cues. Online data-incremental learning seeks to emulate this capability by processing each sample only once, without having access to task or stream cues at any point in time since this is more realistic compared to offline setups, where all data from novel class(es) is assumed to be readily available. However, existing methods typically rely on storing the subsets of data in memory or expanding the initial model architecture, resulting in significant computational overhead. Drawing inspiration from 'self-regulated neurogenesis'-brain's mechanism for creating specialized regions or circuits for distinct functions-we propose a novel approach SERENA which encodes each concept in a specialized network path called 'concept cell', integrated into a single over-parameterized network. Once a concept is learned, its corresponding concept cell is frozen, effectively preventing the forgetting of previously acquired information. Furthermore, we introduce two new continual learning scenarios that more closely reflect real-world conditions, characterized by gradually changing sample sizes. Experimental results show that our method not only establishes new state-of-the-art results across ten benchmarks but also remarkably surpasses offline supervised batch learning performance. The code is available at https://github.com/muratonuryildirim/serena.", "date": "2024-03-13", "categories": []}
{"id": "2506.21312", "title": "Continual Self-Supervised Learning with Masked Autoencoders in Remote Sensing", "abstract": "The development of continual learning (CL) methods, which aim to learn new tasks in a sequential manner from the training data acquired continuously, has gained great attention in remote sensing (RS). The existing CL methods in RS, while learning new tasks, enhance robustness towards catastrophic forgetting. This is achieved by using a large number of labeled training samples, which is costly and not always feasible to gather in RS. To address this problem, we propose a novel continual self-supervised learning method in the context of masked autoencoders (denoted as CoSMAE). The proposed CoSMAE consists of two components: i) data mixup; and ii) model mixup knowledge distillation. Data mixup is associated with retaining information on previous data distributions by interpolating images from the current task with those from the previous tasks. Model mixup knowledge distillation is associated with distilling knowledge from past models and the current model simultaneously by interpolating their model weights to form a teacher for the knowledge distillation. The two components complement each other to regularize the MAE at the data and model levels to facilitate better generalization across tasks and reduce the risk of catastrophic forgetting. Experimental results show that CoSMAE achieves significant improvements of up to 4.94% over state-of-the-art CL methods applied to MAE. Our code is publicly available at: https://git.tu-berlin.de/rsim/CoSMAE.", "date": "2025-06-26", "categories": []}
{"id": "2503.21227", "title": "LLaVA-CMoE: Towards Continual Mixture of Experts for Large Vision-Language Models", "abstract": "Mixture of Experts (MoE) architectures have recently advanced the scalability and adaptability of large language models (LLMs) for continual multimodal learning. However, efficiently extending these models to accommodate sequential tasks remains challenging. As new tasks arrive, naive model expansion leads to rapid parameter growth, while modifying shared routing components often causes catastrophic forgetting, undermining previously learned knowledge. To address these issues, we propose LLaVA-CMoE, a continual learning framework for LLMs that requires no replay data of previous tasks and ensures both parameter efficiency and robust knowledge retention. Our approach introduces a Probe-Guided Knowledge Extension mechanism, which uses probe experts to dynamically determine when and where new experts should be added, enabling adaptive and minimal parameter expansion tailored to task complexity. Furthermore, we present a Probabilistic Task Locator that assigns each task a dedicated, lightweight router. To handle the practical issue that task labels are unknown during inference, we leverage a VAE-based reconstruction strategy to identify the most suitable router by matching input distributions, allowing automatic and accurate expert allocation. This design mitigates routing conflicts and catastrophic forgetting, enabling robust continual learning without explicit task labels. Extensive experiments on the CoIN benchmark, covering eight diverse VQA tasks, demonstrate that LLaVA-CMoE delivers strong continual learning performance with a compact model size, significantly reducing forgetting and parameter overhead compared to prior methods. These results showcase the effectiveness and scalability of our approach for parameter-efficient continual learning in large language models. Our code will be open-sourced soon.", "date": "2025-03-27", "categories": []}
{"id": "2506.19022", "title": "Orthogonal Projection Subspace to Aggregate Online Prior-knowledge for Continual Test-time Adaptation", "abstract": "Continual Test Time Adaptation (CTTA) is a task that requires a source pre-trained model to continually adapt to new scenarios with changing target distributions. Existing CTTA methods primarily focus on mitigating the challenges of catastrophic forgetting and error accumulation. Though there have been emerging methods based on forgetting adaptation with parameter-efficient fine-tuning, they still struggle to balance competitive performance and efficient model adaptation, particularly in complex tasks like semantic segmentation. In this paper, to tackle the above issues, we propose a novel pipeline, Orthogonal Projection Subspace to aggregate online Prior-knowledge, dubbed OoPk. Specifically, we first project a tuning subspace orthogonally which allows the model to adapt to new domains while preserving the knowledge integrity of the pre-trained source model to alleviate catastrophic forgetting. Then, we elaborate an online prior-knowledge aggregation strategy that employs an aggressive yet efficient image masking strategy to mimic potential target dynamism, enhancing the student model's domain adaptability. This further gradually ameliorates the teacher model's knowledge, ensuring high-quality pseudo labels and reducing error accumulation. We demonstrate our method with extensive experiments that surpass previous CTTA methods and achieve competitive performances across various continual TTA benchmarks in semantic segmentation tasks.", "date": "2025-06-23", "categories": []}
{"id": "2506.19692", "title": "Leveraging Lightweight Generators for Memory Efficient Continual Learning", "abstract": "Catastrophic forgetting can be trivially alleviated by keeping all data from previous tasks in memory. Therefore, minimizing the memory footprint while maximizing the amount of relevant information is crucial to the challenge of continual learning. This paper aims to decrease required memory for memory-based continuous learning algorithms. We explore the options of extracting a minimal amount of information, while maximally alleviating forgetting. We propose the usage of lightweight generators based on Singular Value Decomposition to enhance existing continual learning methods, such as A-GEM and Experience Replay. These generators need a minimal amount of memory while being maximally effective. They require no training time, just a single linear-time fitting step, and can capture a distribution effectively from a small number of data samples. Depending on the dataset and network architecture, our results show a significant increase in average accuracy compared to the original methods. Our method shows great potential in minimizing the memory footprint of memory-based continual learning algorithms.", "date": "2025-06-24", "categories": []}
{"id": "2506.17848", "title": "Pathway-based Progressive Inference (PaPI) for Energy-Efficient Continual Learning", "abstract": "Continual learning systems face the dual challenge of preventing catastrophic forgetting while maintaining energy efficiency, particularly in resource-constrained environments. This paper introduces Pathway-based Progressive Inference (PaPI), a novel theoretical framework that addresses these challenges through a mathematically rigorous approach to pathway selection and adaptation. We formulate continual learning as an energy-constrained optimization problem and provide formal convergence guarantees for our pathway routing mechanisms. Our theoretical analysis demonstrates that PaPI achieves an $\\mathcal{O}(K)$ improvement in the stability-plasticity trade-off compared to monolithic architectures, where $K$ is the number of pathways. We derive tight bounds on forgetting rates using Fisher Information Matrix analysis and prove that PaPI's energy consumption scales with the number of active parameters rather than the total model size. Comparative theoretical analysis shows that PaPI provides stronger guarantees against catastrophic forgetting than Elastic Weight Consolidation (EWC) while maintaining better energy efficiency than both EWC and Gradient Episodic Memory (GEM). Our experimental validation confirms these theoretical advantages across multiple benchmarks, demonstrating PaPI's effectiveness for continual learning in energy-constrained settings. Our codes are available at https://github.com/zser092/PAPI_FILES.", "date": "2025-06-21", "categories": []}
{"id": "2506.12034", "title": "Human-like Forgetting Curves in Deep Neural Networks", "abstract": "This study bridges cognitive science and neural network design by examining whether artificial models exhibit human-like forgetting curves. Drawing upon Ebbinghaus' seminal work on memory decay and principles of spaced repetition, we propose a quantitative framework to measure information retention in neural networks. Our approach computes the recall probability by evaluating the similarity between a network's current hidden state and previously stored prototype representations. This retention metric facilitates the scheduling of review sessions, thereby mitigating catastrophic forgetting during deployment and enhancing training efficiency by prompting targeted reviews. Our experiments with Multi-Layer Perceptrons reveal human-like forgetting curves, with knowledge becoming increasingly robust through scheduled reviews. This alignment between neural network forgetting curves and established human memory models identifies neural networks as an architecture that naturally emulates human memory decay and can inform state-of-the-art continual learning algorithms.", "date": "2025-05-22", "categories": []}
{"id": "2506.15720", "title": "Tripartite Weight-Space Ensemble for Few-Shot Class-Incremental Learning", "abstract": "Few-shot class incremental learning (FSCIL) enables the continual learning of new concepts with only a few training examples. In FSCIL, the model undergoes substantial updates, making it prone to forgetting previous concepts and overfitting to the limited new examples. Most recent trend is typically to disentangle the learning of the representation from the classification head of the model. A well-generalized feature extractor on the base classes (many examples and many classes) is learned, and then fixed during incremental learning. Arguing that the fixed feature extractor restricts the model's adaptability to new classes, we introduce a novel FSCIL method to effectively address catastrophic forgetting and overfitting issues. Our method enables to seamlessly update the entire model with a few examples. We mainly propose a tripartite weight-space ensemble (Tri-WE). Tri-WE interpolates the base, immediately previous, and current models in weight-space, especially for the classification heads of the models. Then, it collaboratively maintains knowledge from the base and previous models. In addition, we recognize the challenges of distilling generalized representations from the previous model from scarce data. Hence, we suggest a regularization loss term using amplified data knowledge distillation. Simply intermixing the few-shot data, we can produce richer data enabling the distillation of critical knowledge from the previous model. Consequently, we attain state-of-the-art results on the miniImageNet, CUB200, and CIFAR100 datasets.", "date": "2025-06-04", "categories": []}
{"id": "2506.16574", "title": "Weight Factorization and Centralization for Continual Learning in Speech Recognition", "abstract": "Modern neural network based speech recognition models are required to continually absorb new data without re-training the whole system, especially in downstream applications using foundation models, having no access to the original training data. Continually training the models in a rehearsal-free, multilingual, and language agnostic condition, likely leads to catastrophic forgetting, when a seemingly insignificant disruption to the weights can destructively harm the quality of the models. Inspired by the ability of human brains to learn and consolidate knowledge through the waking-sleeping cycle, we propose a continual learning approach with two distinct phases: factorization and centralization, learning and merging knowledge accordingly. Our experiments on a sequence of varied code-switching datasets showed that the centralization stage can effectively prevent catastrophic forgetting by accumulating the knowledge in multiple scattering low-rank adapters.", "date": "2025-06-19", "categories": []}
{"id": "2412.09441", "title": "MOS: Model Surgery for Pre-Trained Model-Based Class-Incremental Learning", "abstract": "Class-Incremental Learning (CIL) requires models to continually acquire knowledge of new classes without forgetting old ones. Despite Pre-trained Models (PTMs) have shown excellent performance in CIL, catastrophic forgetting still occurs as the model learns new concepts. Existing work seeks to utilize lightweight components to adjust the PTM, while the forgetting phenomenon still comes from {\\em parameter and retrieval} levels. Specifically, iterative updates of the model result in parameter drift, while mistakenly retrieving irrelevant modules leads to the mismatch during inference. To this end, we propose MOdel Surgery (MOS) to rescue the model from forgetting previous knowledge. By training task-specific adapters, we continually adjust the PTM to downstream tasks. To mitigate parameter-level forgetting, we present an adapter merging approach to learn task-specific adapters, which aims to bridge the gap between different components while reserve task-specific information. Besides, to address retrieval-level forgetting, we introduce a training-free self-refined adapter retrieval mechanism during inference, which leverages the model's inherent ability for better adapter retrieval. By jointly rectifying the model with those steps, MOS can robustly resist catastrophic forgetting in the learning process. Extensive experiments on seven benchmark datasets validate MOS's state-of-the-art performance. Code is available at: https://github.com/sun-hailong/AAAI25-MOS", "date": "2024-12-12", "categories": []}
{"id": "2506.14811", "title": "Self-Composing Policies for Scalable Continual Reinforcement Learning", "abstract": "This work introduces a growable and modular neural network architecture that naturally avoids catastrophic forgetting and interference in continual reinforcement learning. The structure of each module allows the selective combination of previous policies along with its internal policy, accelerating the learning process on the current task. Unlike previous growing neural network approaches, we show that the number of parameters of the proposed approach grows linearly with respect to the number of tasks, and does not sacrifice plasticity to scale. Experiments conducted in benchmark continuous control and visual problems reveal that the proposed approach achieves greater knowledge transfer and performance than alternative methods.", "date": "2025-06-04", "categories": []}
{"id": "2506.15566", "title": "Task-Agnostic Experts Composition for Continual Learning", "abstract": "Compositionality is one of the fundamental abilities of the human reasoning process, that allows to decompose a complex problem into simpler elements. Such property is crucial also for neural networks, especially when aiming for a more efficient and sustainable AI framework. We propose a compositional approach by ensembling zero-shot a set of expert models, assessing our methodology using a challenging benchmark, designed to test compositionality capabilities. We show that our Expert Composition method is able to achieve a much higher accuracy than baseline algorithms while requiring less computational resources, hence being more efficient.", "date": "2025-06-18", "categories": []}
{"id": "2506.12031", "title": "Improving Generalization in Heterogeneous Federated Continual Learning via Spatio-Temporal Gradient Matching with Prototypical Coreset", "abstract": "Federated Continual Learning (FCL) has recently emerged as a crucial research area, as data from distributed clients typically arrives as a stream, requiring sequential learning. This paper explores a more practical and challenging FCL setting, where clients may have unrelated or even conflicting data and tasks. In this scenario, statistical heterogeneity and data noise can create spurious correlations, leading to biased feature learning and catastrophic forgetting. Existing FCL approaches often use generative replay to create pseudo-datasets of previous tasks. However, generative replay itself suffers from catastrophic forgetting and task divergence among clients, leading to overfitting in FCL. Existing FCL approaches often use generative replay to create pseudo-datasets of previous tasks. However, generative replay itself suffers from catastrophic forgetting and task divergence among clients, leading to overfitting in FCL. To address these challenges, we propose a novel approach called Spatio-Temporal grAdient Matching with network-free Prototype (STAMP). Our contributions are threefold: 1) We develop a model-agnostic method to determine subset of samples that effectively form prototypes when using a prototypical network, making it resilient to continual learning challenges; 2) We introduce a spatio-temporal gradient matching approach, applied at both the client-side (temporal) and server-side (spatial), to mitigate catastrophic forgetting and data heterogeneity; 3) We leverage prototypes to approximate task-wise gradients, improving gradient matching on the client-side. Extensive experiments demonstrate our method's superiority over existing baselines.", "date": "2025-05-22", "categories": []}
{"id": "2506.12351", "title": "EKPC: Elastic Knowledge Preservation and Compensation for Class-Incremental Learning", "abstract": "Class-Incremental Learning (CIL) aims to enable AI models to continuously learn from sequentially arriving data of different classes over time while retaining previously acquired knowledge. Recently, Parameter-Efficient Fine-Tuning (PEFT) methods, like prompt pool-based approaches and adapter tuning, have shown great attraction in CIL. However, these methods either introduce additional parameters that increase memory usage, or rely on rigid regularization techniques which reduce forgetting but compromise model flexibility. To overcome these limitations, we propose the Elastic Knowledge Preservation and Compensation (EKPC) method, integrating Importance-aware Parameter Regularization (IPR) and Trainable Semantic Drift Compensation (TSDC) for CIL. Specifically, the IPR method assesses the sensitivity of network parameters to prior tasks using a novel parameter-importance algorithm. It then selectively constrains updates within the shared adapter according to these importance values, thereby preserving previously acquired knowledge while maintaining the model's flexibility. However, it still exhibits slight semantic differences in previous knowledge to accommodate new incremental tasks, leading to decision boundaries confusion in classifier. To eliminate this confusion, TSDC trains a unified classifier by compensating prototypes with trainable semantic drift. Extensive experiments on five CIL benchmarks demonstrate the effectiveness of the proposed method, showing superior performances to existing state-of-the-art methods.", "date": "2025-06-14", "categories": []}
{"id": "2402.06434", "title": "Where is the Truth? The Risk of Getting Confounded in a Continual World", "abstract": "A dataset is confounded if it is most easily solved via a spurious correlation, which fails to generalize to new data. In this work, we show that, in a continual learning setting where confounders may vary in time across tasks, the challenge of mitigating the effect of confounders far exceeds the standard forgetting problem normally considered. In particular, we provide a formal description of such continual confounders and identify that, in general, spurious correlations are easily ignored when training for all tasks jointly, but it is harder to avoid confounding when they are considered sequentially. These descriptions serve as a basis for constructing a novel CLEVR-based continually confounded dataset, which we term the ConCon dataset. Our evaluations demonstrate that standard continual learning methods fail to ignore the dataset's confounders. Overall, our work highlights the challenges of confounding factors, particularly in continual learning settings, and demonstrates the need for developing continual learning methods to robustly tackle these.", "date": "2024-02-09", "categories": []}
{"id": "2506.08666", "title": "LLaVA-c: Continual Improved Visual Instruction Tuning", "abstract": "Multimodal models like LLaVA-1.5 achieve state-of-the-art visual understanding through visual instruction tuning on multitask datasets, enabling strong instruction-following and multimodal performance. However, multitask learning faces challenges such as task balancing, requiring careful adjustment of data proportions, and expansion costs, where new tasks risk catastrophic forgetting and need costly retraining. Continual learning provides a promising alternative to acquiring new knowledge incrementally while preserving existing capabilities. However, current methods prioritize task-specific performance, neglecting base model degradation from overfitting to specific instructions, which undermines general capabilities. In this work, we propose a simple but effective method with two modifications on LLaVA-1.5: spectral-aware consolidation for improved task balance and unsupervised inquiry regularization to prevent base model degradation. We evaluate both general and task-specific performance across continual pretraining and fine-tuning. Experiments demonstrate that LLaVA-c consistently enhances standard benchmark performance and preserves general capabilities. For the first time, we show that task-by-task continual learning can achieve results that match or surpass multitask joint learning. The code will be publicly released.", "date": "2025-06-10", "categories": []}
{"id": "2506.11038", "title": "MoTE: Mixture of Task-specific Experts for Pre-Trained ModelBased Class-incremental Learning", "abstract": "Class-incremental learning (CIL) requires deep learning models to continuously acquire new knowledge from streaming data while preserving previously learned information. Recently, CIL based on pre-trained models (PTMs) has achieved remarkable success. However, prompt-based approaches suffer from prompt overwriting, while adapter-based methods face challenges such as dimensional misalignment between tasks. While the idea of expert fusion in Mixture of Experts (MoE) can help address dimensional inconsistency, both expert and routing parameters are prone to being overwritten in dynamic environments, making MoE challenging to apply directly in CIL. To tackle these issues, we propose a mixture of task-specific experts (MoTE) framework that effectively mitigates the miscalibration caused by inconsistent output dimensions across tasks. Inspired by the weighted feature fusion and sparse activation mechanisms in MoE, we introduce task-aware expert filtering and reliable expert joint inference during the inference phase, mimicking the behavior of routing layers without inducing catastrophic forgetting. Extensive experiments demonstrate the superiority of our method without requiring an exemplar set. Furthermore, the number of tasks in MoTE scales linearly with the number of adapters. Building on this, we further explore the trade-off between adapter expansion and model performance and propose the Adapter-Limited MoTE. The code is available at https://github.com/Franklilinjie/MoTE.", "date": "2025-05-21", "categories": []}
{"id": "2506.11672", "title": "Dynamic Mixture of Curriculum LoRA Experts for Continual Multimodal Instruction Tuning", "abstract": "Continual multimodal instruction tuning is crucial for adapting Multimodal Large Language Models (MLLMs) to evolving tasks. However, most existing methods adopt a fixed architecture, struggling with adapting to new tasks due to static model capacity. We propose to evolve the architecture under parameter budgets for dynamic task adaptation, which remains unexplored and imposes two challenges: 1) task architecture conflict, where different tasks require varying layer-wise adaptations, and 2) modality imbalance, where different tasks rely unevenly on modalities, leading to unbalanced updates. To address these challenges, we propose a novel Dynamic Mixture of Curriculum LoRA Experts (D-MoLE) method, which automatically evolves MLLM's architecture with controlled parameter budgets to continually adapt to new tasks while retaining previously learned knowledge. Specifically, we propose a dynamic layer-wise expert allocator, which automatically allocates LoRA experts across layers to resolve architecture conflicts, and routes instructions layer-wisely to facilitate knowledge sharing among experts. Then, we propose a gradient-based inter-modal continual curriculum, which adjusts the update ratio of each module in MLLM based on the difficulty of each modality within the task to alleviate the modality imbalance problem. Extensive experiments show that D-MoLE significantly outperforms state-of-the-art baselines, achieving a 15% average improvement over the best baseline. To the best of our knowledge, this is the first study of continual learning for MLLMs from an architectural perspective.", "date": "2025-06-13", "categories": []}
{"id": "2502.02797", "title": "Upweighting Easy Samples in Fine-Tuning Mitigates Forgetting", "abstract": "Fine-tuning a pre-trained model on a downstream task often degrades its original capabilities, a phenomenon known as \"catastrophic forgetting\". This is especially an issue when one does not have access to the data and recipe used to develop the pre-trained model. Under this constraint, most existing methods for mitigating forgetting are inapplicable. To address this challenge, we propose a sample weighting scheme for the fine-tuning data solely based on the pre-trained model's losses. Specifically, we upweight the easy samples on which the pre-trained model's loss is low and vice versa to limit the drift from the pre-trained model. Our approach is orthogonal and yet complementary to existing methods; while such methods mostly operate on parameter or gradient space, we concentrate on the sample space. We theoretically analyze the impact of fine-tuning with our method in a linear setting, showing that it stalls learning in a certain subspace which inhibits overfitting to the target task. We empirically demonstrate the efficacy of our method on both language and vision tasks. As an example, when fine-tuning Gemma 2 2B on MetaMathQA, our method results in only a $0.8\\%$ drop in accuracy on GSM8K (another math dataset) compared to standard fine-tuning, while preserving $5.4\\%$ more accuracy on the pre-training datasets. Our code is publicly available at https://github.com/sanyalsunny111/FLOW_finetuning .", "date": "2025-02-05", "categories": []}
{"id": "2505.20970", "title": "Measuring Representational Shifts in Continual Learning: A Linear Transformation Perspective", "abstract": "In continual learning scenarios, catastrophic forgetting of previously learned tasks is a critical issue, making it essential to effectively measure such forgetting. Recently, there has been growing interest in focusing on representation forgetting, the forgetting measured at the hidden layer. In this paper, we provide the first theoretical analysis of representation forgetting and use this analysis to better understand the behavior of continual learning. First, we introduce a new metric called representation discrepancy, which measures the difference between representation spaces constructed by two snapshots of a model trained through continual learning. We demonstrate that our proposed metric serves as an effective surrogate for the representation forgetting while remaining analytically tractable. Second, through mathematical analysis of our metric, we derive several key findings about the dynamics of representation forgetting: the forgetting occurs more rapidly to a higher degree as the layer index increases, while increasing the width of the network slows down the forgetting process. Third, we support our theoretical findings through experiments on real image datasets, including Split-CIFAR100 and ImageNet1K.", "date": "2025-05-27", "categories": []}
{"id": "2506.09999", "title": "Leveraging Pre-Trained Models for Multimodal Class-Incremental Learning under Adaptive Fusion", "abstract": "Unlike traditional Multimodal Class-Incremental Learning (MCIL) methods that focus only on vision and text, this paper explores MCIL across vision, audio and text modalities, addressing challenges in integrating complementary information and mitigating catastrophic forgetting. To tackle these issues, we propose an MCIL method based on multimodal pre-trained models. Firstly, a Multimodal Incremental Feature Extractor (MIFE) based on Mixture-of-Experts (MoE) structure is introduced to achieve effective incremental fine-tuning for AudioCLIP. Secondly, to enhance feature discriminability and generalization, we propose an Adaptive Audio-Visual Fusion Module (AAVFM) that includes a masking threshold mechanism and a dynamic feature fusion mechanism, along with a strategy to enhance text diversity. Thirdly, a novel multimodal class-incremental contrastive training loss is proposed to optimize cross-modal alignment in MCIL. Finally, two MCIL-specific evaluation metrics are introduced for comprehensive assessment. Extensive experiments on three multimodal datasets validate the effectiveness of our method.", "date": "2025-02-07", "categories": []}
{"id": "2506.10355", "title": "TreeLoRA: Efficient Continual Learning via Layer-Wise LoRAs Guided by a Hierarchical Gradient-Similarity Tree", "abstract": "Many real-world applications collect data in a streaming environment, where learning tasks are encountered sequentially. This necessitates continual learning (CL) to update models online, enabling adaptation to new tasks while preserving past knowledge to prevent catastrophic forgetting. Nowadays, with the flourish of large pre-trained models (LPMs), efficiency has become increasingly critical for CL, due to their substantial computational demands and growing parameter sizes. In this paper, we introduce TreeLoRA (K-D Tree of Low-Rank Adapters), a novel approach that constructs layer-wise adapters by leveraging hierarchical gradient similarity to enable efficient CL, particularly for LPMs. To reduce the computational burden of task similarity estimation, we employ bandit techniques to develop an algorithm based on lower confidence bounds to efficiently explore the task structure. Furthermore, we use sparse gradient updates to facilitate parameter optimization, making the approach better suited for LPMs. Theoretical analysis is provided to justify the rationale behind our approach, and experiments on both vision transformers (ViTs) and large language models (LLMs) demonstrate the effectiveness and efficiency of our approach across various domains, including vision and natural language processing tasks.", "date": "2025-06-12", "categories": []}
{"id": "2506.10489", "title": "Class-Incremental Learning for Honey Botanical Origin Classification with Hyperspectral Images: A Study with Continual Backpropagation", "abstract": "Honey is an important commodity in the global market. Honey types of different botanical origins provide diversified flavors and health benefits, thus having different market values. Developing accurate and effective botanical origin-distinguishing techniques is crucial to protect consumers' interests. However, it is impractical to collect all the varieties of honey products at once to train a model for botanical origin differentiation. Therefore, researchers developed class-incremental learning (CIL) techniques to address this challenge. This study examined and compared multiple CIL algorithms on a real-world honey hyperspectral imaging dataset. A novel technique is also proposed to improve the performance of class-incremental learning algorithms by combining with a continual backpropagation (CB) algorithm. The CB method addresses the issue of loss-of-plasticity by reinitializing a proportion of less-used hidden neurons to inject variability into neural networks. Experiments showed that CB improved the performance of most CIL methods by 1-7\\%.", "date": "2025-06-12", "categories": []}
{"id": "2506.10710", "title": "Continual Hyperbolic Learning of Instances and Classes", "abstract": "Continual learning has traditionally focused on classifying either instances or classes, but real-world applications, such as robotics and self-driving cars, require models to handle both simultaneously. To mirror real-life scenarios, we introduce the task of continual learning of instances and classes, at the same time. This task challenges models to adapt to multiple levels of granularity over time, which requires balancing fine-grained instance recognition with coarse-grained class generalization. In this paper, we identify that classes and instances naturally form a hierarchical structure. To model these hierarchical relationships, we propose HyperCLIC, a continual learning algorithm that leverages hyperbolic space, which is uniquely suited for hierarchical data due to its ability to represent tree-like structures with low distortion and compact embeddings. Our framework incorporates hyperbolic classification and distillation objectives, enabling the continual embedding of hierarchical relations. To evaluate performance across multiple granularities, we introduce continual hierarchical metrics. We validate our approach on EgoObjects, the only dataset that captures the complexity of hierarchical object recognition in dynamic real-world environments. Empirical results show that HyperCLIC operates effectively at multiple granularities with improved hierarchical generalization.", "date": "2025-06-12", "categories": []}
{"id": "2505.22370", "title": "SplitLoRA: Balancing Stability and Plasticity in Continual Learning Through Gradient Space Splitting", "abstract": "Continual Learning requires a model to learn multiple tasks in sequence while maintaining both stability:preserving knowledge from previously learned tasks, and plasticity:effectively learning new tasks. Gradient projection has emerged as an effective and popular paradigm in CL, where it partitions the gradient space of previously learned tasks into two orthogonal subspaces: a primary subspace and a minor subspace. New tasks are learned effectively within the minor subspace, thereby reducing interference with previously acquired knowledge. However, existing Gradient Projection methods struggle to achieve an optimal balance between plasticity and stability, as it is hard to appropriately partition the gradient space. In this work, we consider a continual learning paradigm based on Low-Rank Adaptation, which has gained considerable attention due to its efficiency and wide applicability, and propose a novel approach for continual learning, called SplitLoRA. We first provide a theoretical analysis of how subspace partitioning affects model stability and plasticity. Informed by this analysis, we then introduce an effective method that derives the optimal partition of the gradient space for previously learned tasks. This approach effectively balances stability and plasticity in continual learning. Experimental results on multiple datasets demonstrate that the proposed method achieves state-of-the-art performance.", "date": "2025-05-28", "categories": []}
{"id": "2506.09347", "title": "ErrorEraser: Unlearning Data Bias for Improved Continual Learning", "abstract": "Continual Learning (CL) primarily aims to retain knowledge to prevent catastrophic forgetting and transfer knowledge to facilitate learning new tasks. Unlike traditional methods, we propose a novel perspective: CL not only needs to prevent forgetting, but also requires intentional forgetting.This arises from existing CL methods ignoring biases in real-world data, leading the model to learn spurious correlations that transfer and amplify across tasks. From feature extraction and prediction results, we find that data biases simultaneously reduce CL's ability to retain and transfer knowledge. To address this, we propose ErrorEraser, a universal plugin that removes erroneous memories caused by biases in CL, enhancing performance in both new and old tasks. ErrorEraser consists of two modules: Error Identification and Error Erasure. The former learns the probability density distribution of task data in the feature space without prior knowledge, enabling accurate identification of potentially biased samples. The latter ensures only erroneous knowledge is erased by shifting the decision space of representative outlier samples. Additionally, an incremental feature distribution learning strategy is designed to reduce the resource overhead during error identification in downstream tasks. Extensive experimental results show that ErrorEraser significantly mitigates the negative impact of data biases, achieving higher accuracy and lower forgetting rates across three types of CL methods. The code is available at https://github.com/diadai/ErrorEraser.", "date": "2025-06-11", "categories": []}
{"id": "2410.18074", "title": "UnCLe: Benchmarking Unsupervised Continual Learning for Depth Completion", "abstract": "We propose UnCLe, the first standardized benchmark for Unsupervised Continual Learning of a multimodal 3D reconstruction task: Depth completion aims to infer a dense depth map from a pair of synchronized RGB image and sparse depth map. We benchmark depth completion models under the practical scenario of unsupervised learning over continuous streams of data. While unsupervised learning of depth boasts the possibility continual learning of novel data distributions over time, existing methods are typically trained on a static, or stationary, dataset. However, when adapting to novel nonstationary distributions, they ``catastrophically forget'' previously learned information. UnCLe simulates these non-stationary distributions by adapting depth completion models to sequences of datasets containing diverse scenes captured from distinct domains using different visual and range sensors. We adopt representative methods from continual learning paradigms and translate them to enable unsupervised continual learning of depth completion. We benchmark these models across indoor and outdoor environments, and investigate the degree of catastrophic forgetting through standard quantitative metrics. We find that unsupervised continual learning of depth completion is an open problem, and we invite researchers to leverage UnCLe as a development platform.", "date": "2024-10-23", "categories": []}
{"id": "2506.08562", "title": "Hierarchical Neural Collapse Detection Transformer for Class Incremental Object Detection", "abstract": "Recently, object detection models have witnessed notable performance improvements, particularly with transformer-based models. However, new objects frequently appear in the real world, requiring detection models to continually learn without suffering from catastrophic forgetting. Although Incremental Object Detection (IOD) has emerged to address this challenge, these existing models are still not practical due to their limited performance and prolonged inference time. In this paper, we introduce a novel framework for IOD, called Hier-DETR: Hierarchical Neural Collapse Detection Transformer, ensuring both efficiency and competitive performance by leveraging Neural Collapse for imbalance dataset and Hierarchical relation of classes' labels.", "date": "2025-06-10", "categories": []}
{"id": "2408.01076", "title": "Exploiting the Semantic Knowledge of Pre-trained Text-Encoders for Continual Learning", "abstract": "Deep neural networks (DNNs) excel on fixed datasets but struggle with incremental and shifting data in real-world scenarios. Continual learning addresses this challenge by allowing models to learn from new data while retaining previously learned knowledge. Existing methods mainly rely on visual features, often neglecting the rich semantic information encoded in text. The semantic knowledge available in the label information of the images, offers important semantic information that can be related with previously acquired knowledge of semantic classes. Consequently, effectively leveraging this information throughout continual learning is expected to be beneficial. To address this, we propose integrating semantic guidance within and across tasks by capturing semantic similarity using text embeddings. We start from a pre-trained CLIP model, employ the \\emph{Semantically-guided Representation Learning (SG-RL)} module for a soft-assignment towards all current task classes, and use the Semantically-guided Knowledge Distillation (SG-KD) module for enhanced knowledge transfer. Experimental results demonstrate the superiority of our method on general and fine-grained datasets. Our code can be found in https://github.com/aprilsveryown/semantically-guided-continual-learning.", "date": "2024-08-02", "categories": []}
{"id": "2411.11932", "title": "Unveiling and Addressing Pseudo Forgetting in Large Language Models", "abstract": "Although substantial efforts have been made to mitigate catastrophic forgetting in continual learning, the intrinsic mechanisms are not well understood. In this work, we demonstrate the existence of \"pseudo forgetting\": the performance degradation on previous tasks is not attributed to a loss of capabilities, but rather to the failure of the instructions to activate the appropriate model abilities. We show that the model's performance on previous tasks can be restored through two simple interventions: (1) providing partial external correct rationale, and (2) appending semantically meaningless suffixes to the original instructions, to guide the generation of correct rationales. Through empirical analysis of the internal mechanisms governing rationale generation, we reveal that models exhibiting pseudo forgetting show reduced instruction dependence during rationale generation, leading to suboptimal activation of their inherent capabilities. Based on this insight, we propose Rationale-Guidance Difficulty based Replay (RGD-R) framework that dynamically allocates replay data based on the model's ability to correctly leverage the intrinsic capabilities. Experimental results demonstrate that RGD-R effectively mitigates pseudo forgetting while maintaining model plasticity.", "date": "2024-11-18", "categories": []}
{"id": "2502.18762", "title": "From Offline to Online Memory-Free and Task-Free Continual Learning via Fine-Grained Hypergradients", "abstract": "Continual Learning (CL) aims to learn from a non-stationary data stream where the underlying distribution changes over time. While recent advances have produced efficient memory-free methods in the offline CL (offCL) setting, where tasks are known in advance and data can be revisited, online CL (onCL) remains dominated by memory-based approaches. The transition from offCL to onCL is challenging, as many offline methods rely on (1) prior knowledge of task boundaries and (2) sophisticated scheduling or optimization schemes, both of which are unavailable when data arrives sequentially and can be seen only once. In this paper, we investigate the adaptation of state-of-the-art memory-free offCL methods to the online setting. We first show that augmenting these methods with lightweight prototypes significantly improves performance, albeit at the cost of increased Gradient Imbalance, resulting in a biased learning towards earlier tasks. To address this issue, we introduce Fine-Grained Hypergradients, an online mechanism for rebalancing gradient updates during training. Our experiments demonstrate that the synergy between prototype memory and hypergradient reweighting substantially enhances the performance of memory-free methods in onCL and surpasses onCL baselines. Code will be released upon acceptance.", "date": "2025-02-26", "categories": []}
{"id": "2505.16809", "title": "Hypergraph Tversky-Aware Domain Incremental Learning for Brain Tumor Segmentation with Missing Modalities", "abstract": "Existing methods for multimodal MRI segmentation with missing modalities typically assume that all MRI modalities are available during training. However, in clinical practice, some modalities may be missing due to the sequential nature of MRI acquisition, leading to performance degradation. Furthermore, retraining models to accommodate newly available modalities can be inefficient and may cause overfitting, potentially compromising previously learned knowledge. To address these challenges, we propose Replay-based Hypergraph Domain Incremental Learning (ReHyDIL) for brain tumor segmentation with missing modalities. ReHyDIL leverages Domain Incremental Learning (DIL) to enable the segmentation model to learn from newly acquired MRI modalities without forgetting previously learned information. To enhance segmentation performance across diverse patient scenarios, we introduce the Cross-Patient Hypergraph Segmentation Network (CHSNet), which utilizes hypergraphs to capture high-order associations between patients. Additionally, we incorporate Tversky-Aware Contrastive (TAC) loss to effectively mitigate information imbalance both across and within different modalities. Extensive experiments on the BraTS2019 dataset demonstrate that ReHyDIL outperforms state-of-the-art methods, achieving an improvement of over 2% in the Dice Similarity Coefficient across various tumor regions. Our code is available at https://github.com/reeive/ReHyDIL.", "date": "2025-05-22", "categories": []}
{"id": "2506.07414", "title": "DPFormer: Dynamic Prompt Transformer for Continual Learning", "abstract": "In continual learning, solving the catastrophic forgetting problem may make the models fall into the stability-plasticity dilemma. Moreover, inter-task confusion will also occur due to the lack of knowledge exchanges between different tasks. In order to solve the aforementioned problems, we propose a novel dynamic prompt transformer (DPFormer) with prompt schemes. The prompt schemes help the DPFormer memorize learned knowledge of previous classes and tasks, and keep on learning new knowledge from new classes and tasks under a single network structure with a nearly fixed number of model parameters. Moreover, they also provide discrepant information to represent different tasks to solve the inter-task confusion problem. Based on prompt schemes, a unified classification module with the binary cross entropy loss, the knowledge distillation loss and the auxiliary loss is proposed to train the whole model in an end-to-end trainable manner. Compared with state-of-the-art methods, our method achieves the best performance in the CIFAR-100, ImageNet100 and ImageNet1K datasets under different class-incremental settings in continual learning. The source code will be available at our GitHub after acceptance.", "date": "2025-06-09", "categories": []}
{"id": "2501.01045", "title": "ZeroFlow: Overcoming Catastrophic Forgetting is Easier than You Think", "abstract": "Backpropagation provides a generalized configuration for overcoming catastrophic forgetting. Optimizers such as SGD and Adam are commonly used for weight updates in continual learning and continual pre-training. However, access to gradient information is not always feasible in practice due to black-box APIs, hardware constraints, or non-differentiable systems, a challenge we refer to as the gradient bans. To bridge this gap, we introduce ZeroFlow, the first benchmark designed to evaluate gradient-free optimization algorithms for overcoming forgetting. ZeroFlow examines a suite of forward pass-based methods across various algorithms, forgetting scenarios, and datasets. Our results show that forward passes alone can be sufficient to mitigate forgetting. We uncover novel optimization principles that highlight the potential of forward pass-based methods in mitigating forgetting, managing task conflicts, and reducing memory demands. Additionally, we propose new enhancements that further improve forgetting resistance using only forward passes. This work provides essential tools and insights to advance the development of forward-pass-based methods for continual learning.", "date": "2025-01-02", "categories": []}
{"id": "2503.04256", "title": "Knowledge Retention for Continual Model-Based Reinforcement Learning", "abstract": "We propose DRAGO, a novel approach for continual model-based reinforcement learning aimed at improving the incremental development of world models across a sequence of tasks that differ in their reward functions but not the state space or dynamics. DRAGO comprises two key components: Synthetic Experience Rehearsal, which leverages generative models to create synthetic experiences from past tasks, allowing the agent to reinforce previously learned dynamics without storing data, and Regaining Memories Through Exploration, which introduces an intrinsic reward mechanism to guide the agent toward revisiting relevant states from prior tasks. Together, these components enable the agent to maintain a comprehensive and continually developing world model, facilitating more effective learning and adaptation across diverse environments. Empirical evaluations demonstrate that DRAGO is able to preserve knowledge across tasks, achieving superior performance in various continual learning scenarios.", "date": "2025-03-06", "categories": []}
{"id": "2504.02107", "title": "TiC-LM: A Web-Scale Benchmark for Time-Continual LLM Pretraining", "abstract": "Large Language Models (LLMs) trained on historical web data inevitably become outdated. We investigate evaluation strategies and update methods for LLMs as new data becomes available. We introduce a web-scale dataset for time-continual pretraining of LLMs derived from 114 dumps of Common Crawl (CC) - orders of magnitude larger than previous continual language modeling benchmarks. We also design time-stratified evaluations across both general CC data and specific domains (Wikipedia, StackExchange, and code documentation) to assess how well various continual learning methods adapt to new data while retaining past knowledge. Our findings demonstrate that, on general CC data, autoregressive meta-schedules combined with a fixed-ratio replay of older data can achieve comparable held-out loss to re-training from scratch, while requiring significantly less computation (2.6x). However, the optimal balance between incorporating new data and replaying old data differs as replay is crucial to avoid forgetting on generic web data but less so on specific domains.", "date": "2025-04-02", "categories": []}
{"id": "2506.05702", "title": "Action-Adaptive Continual Learning: Enabling Policy Generalization under Dynamic Action Spaces", "abstract": "Continual Learning (CL) is a powerful tool that enables agents to learn a sequence of tasks, accumulating knowledge learned in the past and using it for problem-solving or future task learning. However, existing CL methods often assume that the agent's capabilities remain static within dynamic environments, which doesn't reflect real-world scenarios where capabilities dynamically change. This paper introduces a new and realistic problem: Continual Learning with Dynamic Capabilities (CL-DC), posing a significant challenge for CL agents: How can policy generalization across different action spaces be achieved? Inspired by the cortical functions, we propose an Action-Adaptive Continual Learning framework (AACL) to address this challenge. Our framework decouples the agent's policy from the specific action space by building an action representation space. For a new action space, the encoder-decoder of action representations is adaptively fine-tuned to maintain a balance between stability and plasticity. Furthermore, we release a benchmark based on three environments to validate the effectiveness of methods for CL-DC. Experimental results demonstrate that our framework outperforms popular methods by generalizing the policy across action spaces.", "date": "2025-06-06", "categories": []}
{"id": "2401.15098", "title": "Multi-granularity Knowledge Transfer for Continual Reinforcement Learning", "abstract": "Continual reinforcement learning (CRL) empowers RL agents with the ability to learn a sequence of tasks, accumulating knowledge learned in the past and using the knowledge for problemsolving or future task learning. However, existing methods often focus on transferring fine-grained knowledge across similar tasks, which neglects the multi-granularity structure of human cognitive control, resulting in insufficient knowledge transfer across diverse tasks. To enhance coarse-grained knowledge transfer, we propose a novel framework called MT-Core (as shorthand for Multi-granularity knowledge Transfer for Continual reinforcement learning). MT-Core has a key characteristic of multi-granularity policy learning: 1) a coarsegrained policy formulation for utilizing the powerful reasoning ability of the large language model (LLM) to set goals, and 2) a fine-grained policy learning through RL which is oriented by the goals. We also construct a new policy library (knowledge base) to store policies that can be retrieved for multi-granularity knowledge transfer. Experimental results demonstrate the superiority of the proposed MT-Core in handling diverse CRL tasks versus popular baselines.", "date": "2024-01-25", "categories": []}
{"id": "2404.14829", "title": "Revisiting Neural Networks for Continual Learning: An Architectural Perspective", "abstract": "Efforts to overcome catastrophic forgetting have primarily centered around developing more effective Continual Learning (CL) methods. In contrast, less attention was devoted to analyzing the role of network architecture design (e.g., network depth, width, and components) in contributing to CL. This paper seeks to bridge this gap between network architecture design and CL, and to present a holistic study on the impact of network architectures on CL. This work considers architecture design at the network scaling level, i.e., width and depth, and also at the network components, i.e., skip connections, global pooling layers, and down-sampling. In both cases, we first derive insights through systematically exploring how architectural designs affect CL. Then, grounded in these insights, we craft a specialized search space for CL and further propose a simple yet effective ArchCraft method to steer a CL-friendly architecture, namely, this method recrafts AlexNet/ResNet into AlexAC/ResAC. Experimental validation across various CL settings and scenarios demonstrates that improved architectures are parameter-efficient, achieving state-of-the-art performance of CL while being 86%, 61%, and 97% more compact in terms of parameters than the naive CL architecture in Task IL and Class IL. Code is available at https://github.com/byyx666/ArchCraft.", "date": "2024-04-23", "categories": []}
{"id": "2506.03951", "title": "Rethinking the Stability-Plasticity Trade-off in Continual Learning from an Architectural Perspective", "abstract": "The quest for Continual Learning (CL) seeks to empower neural networks with the ability to learn and adapt incrementally. Central to this pursuit is addressing the stability-plasticity dilemma, which involves striking a balance between two conflicting objectives: preserving previously learned knowledge and acquiring new knowledge. While numerous CL methods aim to achieve this trade-off, they often overlook the impact of network architecture on stability and plasticity, restricting the trade-off to the parameter level. In this paper, we delve into the conflict between stability and plasticity at the architectural level. We reveal that under an equal parameter constraint, deeper networks exhibit better plasticity, while wider networks are characterized by superior stability. To address this architectural-level dilemma, we introduce a novel framework denoted Dual-Arch, which serves as a plug-in component for CL. This framework leverages the complementary strengths of two distinct and independent networks: one dedicated to plasticity and the other to stability. Each network is designed with a specialized and lightweight architecture, tailored to its respective objective. Extensive experiments demonstrate that Dual-Arch enhances the performance of existing CL methods while being up to 87% more compact in terms of parameters. Code: https://github.com/byyx666/Dual-Arch.", "date": "2025-06-04", "categories": []}
{"id": "2506.04377", "title": "Replay Can Provably Increase Forgetting", "abstract": "Continual learning seeks to enable machine learning systems to solve an increasing corpus of tasks sequentially. A critical challenge for continual learning is forgetting, where the performance on previously learned tasks decreases as new tasks are introduced. One of the commonly used techniques to mitigate forgetting, sample replay, has been shown empirically to reduce forgetting by retaining some examples from old tasks and including them in new training episodes. In this work, we provide a theoretical analysis of sample replay in an over-parameterized continual linear regression setting, where each task is given by a linear subspace and with enough replay samples, one would be able to eliminate forgetting. Our analysis focuses on sample replay and highlights the role of the replayed samples and the relationship between task subspaces. Surprisingly, we find that, even in a noiseless setting, forgetting can be non-monotonic with respect to the number of replay samples. We present tasks where replay can be harmful with respect to worst-case settings, and also in distributional settings where replay of randomly selected samples increases forgetting in expectation. We also give empirical evidence that harmful replay is not limited to training with linear models by showing similar behavior for a neural networks equipped with SGD. Through experiments on a commonly used benchmark, we provide additional evidence that, even in seemingly benign scenarios, performance of the replay heavily depends on the choice of replay samples and the relationship between tasks.", "date": "2025-06-04", "categories": []}
{"id": "2506.04595", "title": "Hierarchical-Task-Aware Multi-modal Mixture of Incremental LoRA Experts for Embodied Continual Learning", "abstract": "Previous continual learning setups for embodied intelligence focused on executing low-level actions based on human commands, neglecting the ability to learn high-level planning and multi-level knowledge. To address these issues, we propose the Hierarchical Embodied Continual Learning Setups (HEC) that divide the agent's continual learning process into two layers: high-level instructions and low-level actions, and define five embodied continual learning sub-setups. Building on these setups, we introduce the Task-aware Mixture of Incremental LoRA Experts (Task-aware MoILE) method. This approach achieves task recognition by clustering visual-text embeddings and uses both a task-level router and a token-level router to select the appropriate LoRA experts. To effectively address the issue of catastrophic forgetting, we apply Singular Value Decomposition (SVD) to the LoRA parameters obtained from prior tasks, preserving key components while orthogonally training the remaining parts. The experimental results show that our method stands out in reducing the forgetting of old tasks compared to other methods, effectively supporting agents in retaining prior knowledge while continuously learning new tasks.", "date": "2025-06-05", "categories": []}
{"id": "2506.03189", "title": "Continual Learning in Vision-Language Models via Aligned Model Merging", "abstract": "Continual learning is conventionally tackled through sequential fine-tuning, a process that, while enabling adaptation, inherently favors plasticity over the stability needed to retain prior knowledge. While existing approaches attempt to mitigate catastrophic forgetting, a bias towards recent tasks persists as they build upon this sequential nature. In this work we present a new perspective based on model merging to maintain stability while still retaining plasticity. Rather than just sequentially updating the model weights, we propose merging newly trained task parameters with previously learned ones, promoting a better balance. To maximize the effectiveness of the merging process, we propose a simple mechanism that promotes learning aligned weights with previous ones, thereby avoiding interference when merging. We evaluate this approach on large Vision-Language Models (VLMs), and demonstrate its effectiveness in reducing forgetting, increasing robustness to various task orders and similarities, and improving generalization.", "date": "2025-05-30", "categories": []}
{"id": "2506.03320", "title": "The Future of Continual Learning in the Era of Foundation Models: Three Key Directions", "abstract": "Continual learning--the ability to acquire, retain, and refine knowledge over time--has always been fundamental to intelligence, both human and artificial. Historically, different AI paradigms have acknowledged this need, albeit with varying priorities: early expert and production systems focused on incremental knowledge consolidation, while reinforcement learning emphasised dynamic adaptation. With the rise of deep learning, deep continual learning has primarily focused on learning robust and reusable representations over time to solve sequences of increasingly complex tasks. However, the emergence of Large Language Models (LLMs) and foundation models has raised the question: Do we still need continual learning when centralised, monolithic models can tackle diverse tasks with access to internet-scale knowledge? We argue that continual learning remains essential for three key reasons: (i) continual pre-training is still necessary to ensure foundation models remain up to date, mitigating knowledge staleness and distribution shifts while integrating new information; (ii) continual fine-tuning enables models to specialise and personalise, adapting to domain-specific tasks, user preferences, and real-world constraints without full retraining, avoiding the need for computationally expensive long context-windows; (iii) continual compositionality offers a scalable and modular approach to intelligence, enabling the orchestration of foundation models and agents to be dynamically composed, recombined, and adapted. While continual pre-training and fine-tuning are explored as niche research directions, we argue it is continual compositionality that will mark the rebirth of continual learning. The future of AI will not be defined by a single static model but by an ecosystem of continually evolving and interacting models, making continual learning more relevant than ever.", "date": "2025-06-03", "categories": []}
{"id": "2412.18926", "title": "Exemplar-condensed Federated Class-incremental Learning", "abstract": "We propose Exemplar-Condensed federated class-incremental learning (ECoral) to distil the training characteristics of real images from streaming data into informative rehearsal exemplars. The proposed method eliminates the limitations of exemplar selection in replay-based approaches for mitigating catastrophic forgetting in federated continual learning (FCL). The limitations particularly related to the heterogeneity of information density of each summarized data. Our approach maintains the consistency of training gradients and the relationship to past tasks for the summarized exemplars to represent the streaming data compared to the original images effectively. Additionally, our approach reduces the information-level heterogeneity of the summarized data by inter-client sharing of the disentanglement generative model. Extensive experiments show that our ECoral outperforms several state-of-the-art methods and can be seamlessly integrated with many existing approaches to enhance performance.", "date": "2024-12-25", "categories": []}
{"id": "2506.01496", "title": "Continual Speech Learning with Fused Speech Features", "abstract": "Rapid growth in speech data demands adaptive models, as traditional static methods fail to keep pace with dynamic and diverse speech information. We introduce continuous speech learning, a new set-up targeting at bridging the adaptation gap in current speech models. We use the encoder-decoder Whisper model to standardize speech tasks into a generative format. We integrate a learnable gated-fusion layer on the top of the encoder to dynamically select task-specific features for downstream tasks. Our approach improves accuracy significantly over traditional methods in six speech processing tasks, demonstrating gains in adapting to new speech tasks without full retraining.", "date": "2025-06-02", "categories": []}
{"id": "2506.01965", "title": "TaskVAE: Task-Specific Variational Autoencoders for Exemplar Generation in Continual Learning for Human Activity Recognition", "abstract": "As machine learning based systems become more integrated into daily life, they unlock new opportunities but face the challenge of adapting to dynamic data environments. Various forms of data shift-gradual, abrupt, or cyclic-threaten model accuracy, making continual adaptation essential. Continual Learning (CL) enables models to learn from evolving data streams while minimizing forgetting of prior knowledge. Among CL strategies, replay-based methods have proven effective, but their success relies on balancing memory constraints and retaining old class accuracy while learning new classes. This paper presents TaskVAE, a framework for replay-based CL in class-incremental settings. TaskVAE employs task-specific Variational Autoencoders (VAEs) to generate synthetic exemplars from previous tasks, which are then used to train the classifier alongside new task data. In contrast to traditional methods that require prior knowledge of the total class count or rely on a single VAE for all tasks, TaskVAE adapts flexibly to increasing tasks without such constraints. We focus on Human Activity Recognition (HAR) using IMU sensor-equipped devices. Unlike previous HAR studies that combine data across all users, our approach focuses on individual user data, better reflecting real-world scenarios where a person progressively learns new activities. Extensive experiments on 5 different HAR datasets show that TaskVAE outperforms experience replay methods, particularly with limited data, and exhibits robust performance as dataset size increases. Additionally, memory footprint of TaskVAE is minimal, being equivalent to only 60 samples per task, while still being able to generate an unlimited number of synthetic samples. The contributions lie in balancing memory constraints, task-specific generation, and long-term stability, making it a reliable solution for real-world applications in domains like HAR.", "date": "2025-05-10", "categories": []}
{"id": "2506.02041", "title": "Enhancing Multimodal Continual Instruction Tuning with BranchLoRA", "abstract": "Multimodal Continual Instruction Tuning (MCIT) aims to finetune Multimodal Large Language Models (MLLMs) to continually align with human intent across sequential tasks. Existing approaches often rely on the Mixture-of-Experts (MoE) LoRA framework to preserve previous instruction alignments. However, these methods are prone to Catastrophic Forgetting (CF), as they aggregate all LoRA blocks via simple summation, which compromises performance over time. In this paper, we identify a critical parameter inefficiency in the MoELoRA framework within the MCIT context. Based on this insight, we propose BranchLoRA, an asymmetric framework to enhance both efficiency and performance. To mitigate CF, we introduce a flexible tuning-freezing mechanism within BranchLoRA, enabling branches to specialize in intra-task knowledge while fostering inter-task collaboration. Moreover, we incrementally incorporate task-specific routers to ensure an optimal branch distribution over time, rather than favoring the most recent task. To streamline inference, we introduce a task selector that automatically routes test inputs to the appropriate router without requiring task identity. Extensive experiments on the latest MCIT benchmark demonstrate that BranchLoRA significantly outperforms MoELoRA and maintains its superiority across various MLLM sizes.", "date": "2025-05-31", "categories": []}
{"id": "2506.02065", "title": "EWGN: Elastic Weight Generation and Context Switching in Deep Learning", "abstract": "The ability to learn and retain a wide variety of tasks is a hallmark of human intelligence that has inspired research in artificial general intelligence. Continual learning approaches provide a significant step towards achieving this goal. It has been known that task variability and context switching are challenging for learning in neural networks. Catastrophic forgetting refers to the poor performance on retention of a previously learned task when a new task is being learned. Switching between different task contexts can be a useful approach to mitigate the same by preventing the interference between the varying task weights of the network. This paper introduces Elastic Weight Generative Networks (EWGN) as an idea for context switching between two different tasks. The proposed EWGN architecture uses an additional network that generates the weights of the primary network dynamically while consolidating the weights learned. The weight generation is input-dependent and thus enables context switching. Using standard computer vision datasets, namely MNIST and fashion-MNIST, we analyse the retention of previously learned task representations in Fully Connected Networks, Convolutional Neural Networks, and EWGN architectures with Stochastic Gradient Descent and Elastic Weight Consolidation learning algorithms. Understanding dynamic weight generation and context-switching ability can be useful in enabling continual learning for improved performance.", "date": "2025-06-01", "categories": []}
{"id": "2506.00205", "title": "Unlocking the Power of Rehearsal in Continual Learning: A Theoretical Perspective", "abstract": "Rehearsal-based methods have shown superior performance in addressing catastrophic forgetting in continual learning (CL) by storing and training on a subset of past data alongside new data in current task. While such a concurrent rehearsal strategy is widely used, it remains unclear if this approach is always optimal. Inspired by human learning, where sequentially revisiting tasks helps mitigate forgetting, we explore whether sequential rehearsal can offer greater benefits for CL compared to standard concurrent rehearsal. To address this question, we conduct a theoretical analysis of rehearsal-based CL in overparameterized linear models, comparing two strategies: 1) Concurrent Rehearsal, where past and new data are trained together, and 2) Sequential Rehearsal, where new data is trained first, followed by revisiting past data sequentially. By explicitly characterizing forgetting and generalization error, we show that sequential rehearsal performs better when tasks are less similar. These insights further motivate a novel Hybrid Rehearsal method, which trains similar tasks concurrently and revisits dissimilar tasks sequentially. We characterize its forgetting and generalization performance, and our experiments with deep neural networks further confirm that the hybrid approach outperforms standard concurrent rehearsal. This work provides the first comprehensive theoretical analysis of rehearsal-based CL.", "date": "2025-05-30", "categories": []}
{"id": "2506.00477", "title": "Flashbacks to Harmonize Stability and Plasticity in Continual Learning", "abstract": "We introduce Flashback Learning (FL), a novel method designed to harmonize the stability and plasticity of models in Continual Learning (CL). Unlike prior approaches that primarily focus on regularizing model updates to preserve old information while learning new concepts, FL explicitly balances this trade-off through a bidirectional form of regularization. This approach effectively guides the model to swiftly incorporate new knowledge while actively retaining its old knowledge. FL operates through a two-phase training process and can be seamlessly integrated into various CL methods, including replay, parameter regularization, distillation, and dynamic architecture techniques. In designing FL, we use two distinct knowledge bases: one to enhance plasticity and another to improve stability. FL ensures a more balanced model by utilizing both knowledge bases to regularize model updates. Theoretically, we analyze how the FL mechanism enhances the stability-plasticity balance. Empirically, FL demonstrates tangible improvements over baseline methods within the same training budget. By integrating FL into at least one representative baseline from each CL category, we observed an average accuracy improvement of up to 4.91% in Class-Incremental and 3.51% in Task-Incremental settings on standard image classification benchmarks. Additionally, measurements of the stability-to-plasticity ratio confirm that FL effectively enhances this balance. FL also outperforms state-of-the-art CL methods on more challenging datasets like ImageNet.", "date": "2025-05-31", "categories": []}
{"id": "2506.00592", "title": "Mitigating Plasticity Loss in Continual Reinforcement Learning by Reducing Churn", "abstract": "Plasticity, or the ability of an agent to adapt to new tasks, environments, or distributions, is crucial for continual learning. In this paper, we study the loss of plasticity in deep continual RL from the lens of churn: network output variability for out-of-batch data induced by mini-batch training. We demonstrate that (1) the loss of plasticity is accompanied by the exacerbation of churn due to the gradual rank decrease of the Neural Tangent Kernel (NTK) matrix; (2) reducing churn helps prevent rank collapse and adjusts the step size of regular RL gradients adaptively. Moreover, we introduce Continual Churn Approximated Reduction (C-CHAIN) and demonstrate it improves learning performance and outperforms baselines in a diverse range of continual learning environments on OpenAI Gym Control, ProcGen, DeepMind Control Suite, and MinAtar benchmarks.", "date": "2025-05-31", "categories": []}
{"id": "2506.00816", "title": "L3A: Label-Augmented Analytic Adaptation for Multi-Label Class Incremental Learning", "abstract": "Class-incremental learning (CIL) enables models to learn new classes continually without forgetting previously acquired knowledge. Multi-label CIL (MLCIL) extends CIL to a real-world scenario where each sample may belong to multiple classes, introducing several challenges: label absence, which leads to incomplete historical information due to missing labels, and class imbalance, which results in the model bias toward majority classes. To address these challenges, we propose Label-Augmented Analytic Adaptation (L3A), an exemplar-free approach without storing past samples. L3A integrates two key modules. The pseudo-label (PL) module implements label augmentation by generating pseudo-labels for current phase samples, addressing the label absence problem. The weighted analytic classifier (WAC) derives a closed-form solution for neural networks. It introduces sample-specific weights to adaptively balance the class contribution and mitigate class imbalance. Experiments on MS-COCO and PASCAL VOC datasets demonstrate that L3A outperforms existing methods in MLCIL tasks. Our code is available at https://github.com/scut-zx/L3A.", "date": "2025-06-01", "categories": []}
{"id": "2506.01545", "title": "Class Incremental Learning for Algorithm Selection", "abstract": "Algorithm selection is commonly used to predict the best solver from a portfolio per per-instance. In many real scenarios, instances arrive in a stream: new instances become available over time, while the number of class labels can also grow as new data distributions arrive downstream. As a result, the classification model needs to be periodically updated to reflect additional solvers without catastrophic forgetting of past data. In machine-learning (ML), this is referred to as Class Incremental Learning (CIL). While commonly addressed in ML settings, its relevance to algorithm-selection in optimisation has not been previously studied. Using a bin-packing dataset, we benchmark 8 continual learning methods with respect to their ability to withstand catastrophic forgetting. We find that rehearsal-based methods significantly outperform other CIL methods. While there is evidence of forgetting, the loss is small at around 7%. Hence, these methods appear to be a viable approach to continual learning in streaming optimisation scenarios.", "date": "2025-06-02", "categories": []}
{"id": "2502.17510", "title": "Recurrent Knowledge Identification and Fusion for Language Model Continual Learning", "abstract": "Continual learning (CL) is crucial for deploying large language models (LLMs) in dynamic real-world environments without costly retraining. While recent model ensemble and model merging methods guided by parameter importance have gained popularity, they often struggle to balance knowledge transfer and forgetting, mainly due to the reliance on static importance estimates during sequential training. In this paper, we present Recurrent-KIF, a novel CL framework for Recurrent Knowledge Identification and Fusion, which enables dynamic estimation of parameter importance distributions to enhance knowledge transfer. Inspired by human continual learning, Recurrent-KIF employs an inner loop that rapidly adapts to new tasks while identifying important parameters, coupled with an outer loop that globally manages the fusion of new and historical knowledge through redundant knowledge pruning and key knowledge merging. These inner-outer loops iteratively perform multiple rounds of fusion, allowing Recurrent-KIF to leverage intermediate training information and adaptively adjust fusion strategies based on evolving importance distributions. Extensive experiments on two CL benchmarks with various model sizes (from 770M to 13B) demonstrate that Recurrent-KIF effectively mitigates catastrophic forgetting and enhances knowledge transfer.", "date": "2025-02-22", "categories": []}
{"id": "2505.24816", "title": "CL-LoRA: Continual Low-Rank Adaptation for Rehearsal-Free Class-Incremental Learning", "abstract": "Class-Incremental Learning (CIL) aims to learn new classes sequentially while retaining the knowledge of previously learned classes. Recently, pre-trained models (PTMs) combined with parameter-efficient fine-tuning (PEFT) have shown remarkable performance in rehearsal-free CIL without requiring exemplars from previous tasks. However, existing adapter-based methods, which incorporate lightweight learnable modules into PTMs for CIL, create new adapters for each new task, leading to both parameter redundancy and failure to leverage shared knowledge across tasks. In this work, we propose ContinuaL Low-Rank Adaptation (CL-LoRA), which introduces a novel dual-adapter architecture combining \\textbf{task-shared adapters} to learn cross-task knowledge and \\textbf{task-specific adapters} to capture unique features of each new task. Specifically, the shared adapters utilize random orthogonal matrices and leverage knowledge distillation with gradient reassignment to preserve essential shared knowledge. In addition, we introduce learnable block-wise weights for task-specific adapters, which mitigate inter-task interference while maintaining the model's plasticity. We demonstrate CL-LoRA consistently achieves promising performance under multiple benchmarks with reduced training and inference computation, establishing a more efficient and scalable paradigm for continual learning with pre-trained models.", "date": "2025-05-30", "categories": []}
{"id": "2403.07404", "title": "Improving Continual Learning Performance and Efficiency with Auxiliary Classifiers", "abstract": "Continual learning is crucial for applying machine learning in challenging, dynamic, and often resource-constrained environments. However, catastrophic forgetting - overwriting previously learned knowledge when new information is acquired - remains a major challenge. In this work, we examine the intermediate representations in neural network layers during continual learning and find that such representations are less prone to forgetting, highlighting their potential to accelerate computation. Motivated by these findings, we propose to use auxiliary classifiers(ACs) to enhance performance and demonstrate that integrating ACs into various continual learning methods consistently improves accuracy across diverse evaluation settings, yielding an average 10% relative gain. We also leverage the ACs to reduce the average cost of the inference by 10-60% without compromising accuracy, enabling the model to return the predictions before computing all the layers. Our approach provides a scalable and efficient solution for continual learning.", "date": "2024-03-12", "categories": []}
{"id": "2502.05540", "title": "Demystifying Catastrophic Forgetting in Two-Stage Incremental Object Detector", "abstract": "Catastrophic forgetting is a critical chanllenge for incremental object detection (IOD). Most existing methods treat the detector monolithically, relying on instance replay or knowledge distillation without analyzing component-specific forgetting. Through dissection of Faster R-CNN, we reveal a key insight: Catastrophic forgetting is predominantly localized to the RoI Head classifier, while regressors retain robustness across incremental stages. This finding challenges conventional assumptions, motivating us to develop a framework termed NSGP-RePRE. Regional Prototype Replay (RePRE) mitigates classifier forgetting via replay of two types of prototypes: coarse prototypes represent class-wise semantic centers of RoI features, while fine-grained prototypes model intra-class variations. Null Space Gradient Projection (NSGP) is further introduced to eliminate prototype-feature misalignment by updating the feature extractor in directions orthogonal to subspace of old inputs via gradient projection, aligning RePRE with incremental learning dynamics. Our simple yet effective design allows NSGP-RePRE to achieve state-of-the-art performance on the Pascal VOC and MS COCO datasets under various settings. Our work not only advances IOD methodology but also provide pivotal insights for catastrophic forgetting mitigation in IOD. Code is available at \\href{https://github.com/fanrena/NSGP-RePRE}{https://github.com/fanrena/NSGP-RePRE} .", "date": "2025-02-08", "categories": []}
{"id": "2503.12941", "title": "HiDe-LLaVA: Hierarchical Decoupling for Continual Instruction Tuning of Multimodal Large Language Model", "abstract": "Instruction tuning is widely used to improve a pre-trained Multimodal Large Language Model (MLLM) by training it on curated task-specific datasets, enabling better comprehension of human instructions. However, it is infeasible to collect all possible instruction datasets simultaneously in real-world scenarios. Thus, enabling MLLM with continual instruction tuning is essential for maintaining their adaptability. However, existing methods often trade off memory efficiency for performance gains, significantly compromising overall efficiency. In this paper, we propose a task-specific expansion and task-general fusion framework based on the variations in Centered Kernel Alignment (CKA) similarity across different model layers when trained on diverse datasets. Furthermore, we analyze the information leakage present in the existing benchmark and propose a new and more challenging benchmark to rationally evaluate the performance of different methods. Comprehensive experiments showcase a significant performance improvement of our method compared to existing state-of-the-art methods. Code and dataset are released at https://github.com/Ghy0501/HiDe-LLaVA.", "date": "2025-03-17", "categories": []}
{"id": "2504.02666", "title": "BECAME: BayEsian Continual Learning with Adaptive Model MErging", "abstract": "Continual Learning (CL) strives to learn incrementally across tasks while mitigating catastrophic forgetting. A key challenge in CL is balancing stability (retaining prior knowledge) and plasticity (learning new tasks). While representative gradient projection methods ensure stability, they often limit plasticity. Model merging techniques offer promising solutions, but prior methods typically rely on empirical assumptions and carefully selected hyperparameters. In this paper, we explore the potential of model merging to enhance the stability-plasticity trade-off, providing theoretical insights that underscore its benefits. Specifically, we reformulate the merging mechanism using Bayesian continual learning principles and derive a closed-form solution for the optimal merging coefficient that adapts to the diverse characteristics of tasks. To validate our approach, we introduce a two-stage framework named BECAME, which synergizes the expertise of gradient projection and adaptive merging. Extensive experiments show that our approach outperforms state-of-the-art CL methods and existing merging strategies.", "date": "2025-04-03", "categories": []}
{"id": "2505.22018", "title": "Improving Continual Pre-training Through Seamless Data Packing", "abstract": "Continual pre-training has demonstrated significant potential in enhancing model performance, particularly in domain-specific scenarios. The most common approach for packing data before continual pre-training involves concatenating input texts and splitting them into fixed-length sequences. While straightforward and efficient, this method often leads to excessive truncation and context discontinuity, which can hinder model performance. To address these issues, we explore the potential of data engineering to enhance continual pre-training, particularly its impact on model performance and efficiency. We propose Seamless Packing (SP), a novel data packing strategy aimed at preserving contextual information more effectively and enhancing model performance. Our approach employs a sliding window technique in the first stage that synchronizes overlapping tokens across consecutive sequences, ensuring better continuity and contextual coherence. In the second stage, we adopt a First-Fit-Decreasing algorithm to pack shorter texts into bins slightly larger than the target sequence length, thereby minimizing padding and truncation. Empirical evaluations across various model architectures and corpus domains demonstrate the effectiveness of our method, outperforming baseline method in 99% of all settings. Code is available at https://github.com/Infernus-WIND/Seamless-Packing.", "date": "2025-05-28", "categories": []}
{"id": "2505.23271", "title": "LADA: Scalable Label-Specific CLIP Adapter for Continual Learning", "abstract": "Continual learning with vision-language models like CLIP offers a pathway toward scalable machine learning systems by leveraging its transferable representations. Existing CLIP-based methods adapt the pre-trained image encoder by adding multiple sets of learnable parameters, with each task using a partial set of parameters. This requires selecting the expected parameters for input images during inference, which is prone to error that degrades performance. To address this problem, we introduce LADA (Label-specific ADApter). Instead of partitioning parameters across tasks, LADA appends lightweight, label-specific memory units to the frozen CLIP image encoder, enabling discriminative feature generation by aggregating task-agnostic knowledge. To prevent catastrophic forgetting, LADA employs feature distillation for seen classes, preventing their features from being interfered with by new classes. Positioned after the image encoder, LADA prevents gradient flow to the frozen CLIP parameters, ensuring efficient training. Extensive results show that LADA achieves state-of-the-art performance in continual learning settings. The implementation code is available at https://github.com/MaolinLuo/LADA.", "date": "2025-05-29", "categories": []}
{"id": "2505.23744", "title": "Boosting Domain Incremental Learning: Selecting the Optimal Parameters is All You Need", "abstract": "Deep neural networks (DNNs) often underperform in real-world, dynamic settings where data distributions change over time. Domain Incremental Learning (DIL) offers a solution by enabling continual model adaptation, with Parameter-Isolation DIL (PIDIL) emerging as a promising paradigm to reduce knowledge conflicts. However, existing PIDIL methods struggle with parameter selection accuracy, especially as the number of domains and corresponding classes grows. To address this, we propose SOYO, a lightweight framework that improves domain selection in PIDIL. SOYO introduces a Gaussian Mixture Compressor (GMC) and Domain Feature Resampler (DFR) to store and balance prior domain data efficiently, while a Multi-level Domain Feature Fusion Network (MDFN) enhances domain feature extraction. Our framework supports multiple Parameter-Efficient Fine-Tuning (PEFT) methods and is validated across tasks such as image classification, object detection, and speech enhancement. Experimental results on six benchmarks demonstrate SOYO's consistent superiority over existing baselines, showcasing its robustness and adaptability in complex, evolving environments. The codes will be released in https://github.com/qwangcv/SOYO.", "date": "2025-05-29", "categories": []}
{"id": "2505.20135", "title": "Data-Distill-Net: A Data Distillation Approach Tailored for Reply-based Continual Learning", "abstract": "Replay-based continual learning (CL) methods assume that models trained on a small subset can also effectively minimize the empirical risk of the complete dataset. These methods maintain a memory buffer that stores a sampled subset of data from previous tasks to consolidate past knowledge. However, this assumption is not guaranteed in practice due to the limited capacity of the memory buffer and the heuristic criteria used for buffer data selection. To address this issue, we propose a new dataset distillation framework tailored for CL, which maintains a learnable memory buffer to distill the global information from the current task data and accumulated knowledge preserved in the previous memory buffer. Moreover, to avoid the computational overhead and overfitting risks associated with parameterizing the entire buffer during distillation, we introduce a lightweight distillation module that can achieve global information distillation solely by generating learnable soft labels for the memory buffer data. Extensive experiments show that, our method can achieve competitive results and effectively mitigates forgetting across various datasets. The source code will be publicly available.", "date": "2025-05-26", "categories": []}
{"id": "2505.21942", "title": "Continual Learning Beyond Experience Rehearsal and Full Model Surrogates", "abstract": "Continual learning (CL) has remained a significant challenge for deep neural networks as learning new tasks erases previously acquired knowledge, either partially or completely. Existing solutions often rely on experience rehearsal or full model surrogates to mitigate CF. While effective, these approaches introduce substantial memory and computational overhead, limiting their scalability and applicability in real-world scenarios. To address this, we propose SPARC, a scalable CL approach that eliminates the need for experience rehearsal and full-model surrogates. By effectively combining task-specific working memories and task-agnostic semantic memory for cross-task knowledge consolidation, SPARC results in a remarkable parameter efficiency, using only 6% of the parameters required by full-model surrogates. Despite its lightweight design, SPARC achieves superior performance on Seq-TinyImageNet and matches rehearsal-based methods on various CL benchmarks. Additionally, weight re-normalization in the classification layer mitigates task-specific biases, establishing SPARC as a practical and scalable solution for CL under stringent efficiency constraints.", "date": "2025-05-28", "categories": []}
{"id": "2505.22408", "title": "Frugal Incremental Generative Modeling using Variational Autoencoders", "abstract": "Continual or incremental learning holds tremendous potential in deep learning with different challenges including catastrophic forgetting. The advent of powerful foundation and generative models has propelled this paradigm even further, making it one of the most viable solution to train these models. However, one of the persisting issues lies in the increasing volume of data particularly with replay-based methods. This growth introduces challenges with scalability since continuously expanding data becomes increasingly demanding as the number of tasks grows. In this paper, we attenuate this issue by devising a novel replay-free incremental learning model based on Variational Autoencoders (VAEs). The main contribution of this work includes (i) a novel incremental generative modelling, built upon a well designed multi-modal latent space, and also (ii) an orthogonality criterion that mitigates catastrophic forgetting of the learned VAEs. The proposed method considers two variants of these VAEs: static and dynamic with no (or at most a controlled) growth in the number of parameters. Extensive experiments show that our method is (at least) an order of magnitude more ``memory-frugal'' compared to the closely related works while achieving SOTA accuracy scores.", "date": "2025-05-28", "categories": []}
{"id": "2403.10056", "title": "Don't Half-listen: Capturing Key-part Information in Continual Instruction Tuning", "abstract": "Instruction tuning for large language models (LLMs) can drive them to produce results consistent with human goals in specific downstream tasks. However, the process of continual instruction tuning (CIT) for LLMs may bring about the catastrophic forgetting (CF) problem, where previously learned abilities are degraded. Recent methods try to alleviate the CF problem by modifying models or replaying data, which may only remember the surface-level pattern of instructions and get confused on held-out tasks. In this paper, we propose a novel continual instruction tuning method based on Key-part Information Gain (KPIG). Our method computes the information gain on masked parts to dynamically replay data and refine the training objective, which enables LLMs to capture task-aware information relevant to the correct response and alleviate overfitting to general descriptions in instructions. In addition, we propose two metrics, P-score and V-score, to measure the generalization and instruction-following abilities of LLMs. Experiments demonstrate our method achieves superior performance on both seen and held-out tasks.", "date": "2024-03-15", "categories": []}
{"id": "2503.07114", "title": "Sequential Function-Space Variational Inference via Gaussian Mixture Approximation", "abstract": "Continual learning in neural networks aims to learn new tasks without forgetting old tasks. Sequential function-space variational inference (SFSVI) uses a Gaussian variational distribution to approximate the distribution of the outputs of the neural network corresponding to a finite number of selected inducing points. Since the posterior distribution of a neural network is multi-modal, a Gaussian distribution could only match one mode of the posterior distribution, and a Gaussian mixture distribution could be used to better approximate the posterior distribution. We propose an SFSVI method based on a Gaussian mixture variational distribution. We also compare different types of variational inference methods with a fixed pre-trained feature extractor (where continual learning is performed on the final layer) and without a fixed pre-trained feature extractor (where continual learning is performed on all layers). We find that in terms of final average accuracy, likelihood-focused Gaussian mixture SFSVI outperforms other sequential variational inference methods, especially in the latter case.", "date": "2025-03-10", "categories": []}
{"id": "2505.20933", "title": "Information-Theoretic Complementary Prompts for Improved Continual Text Classification", "abstract": "Continual Text Classification (CTC) aims to continuously classify new text data over time while minimizing catastrophic forgetting of previously acquired knowledge. However, existing methods often focus on task-specific knowledge, overlooking the importance of shared, task-agnostic knowledge. Inspired by the complementary learning systems theory, which posits that humans learn continually through the interaction of two systems -- the hippocampus, responsible for forming distinct representations of specific experiences, and the neocortex, which extracts more general and transferable representations from past experiences -- we introduce Information-Theoretic Complementary Prompts (InfoComp), a novel approach for CTC. InfoComp explicitly learns two distinct prompt spaces: P(rivate)-Prompt and S(hared)-Prompt. These respectively encode task-specific and task-invariant knowledge, enabling models to sequentially learn classification tasks without relying on data replay. To promote more informative prompt learning, InfoComp uses an information-theoretic framework that maximizes mutual information between different parameters (or encoded representations). Within this framework, we design two novel loss functions: (1) to strengthen the accumulation of task-specific knowledge in P-Prompt, effectively mitigating catastrophic forgetting, and (2) to enhance the retention of task-invariant knowledge in S-Prompt, improving forward knowledge transfer. Extensive experiments on diverse CTC benchmarks show that our approach outperforms previous state-of-the-art methods.", "date": "2025-05-27", "categories": []}
{"id": "2406.02447", "title": "Federated Class-Incremental Learning with Hierarchical Generative Prototypes", "abstract": "Federated Learning (FL) aims at unburdening the training of deep models by distributing computation across multiple devices (clients) while safeguarding data privacy. On top of that, Federated Continual Learning (FCL) also accounts for data distribution evolving over time, mirroring the dynamic nature of real-world environments. While previous studies have identified Catastrophic Forgetting and Client Drift as primary causes of performance degradation in FCL, we shed light on the importance of Incremental Bias and Federated Bias, which cause models to prioritize classes that are recently introduced or locally predominant, respectively. Our proposal constrains both biases in the last layer by efficiently finetuning a pre-trained backbone using learnable prompts, resulting in clients that produce less biased representations and more biased classifiers. Therefore, instead of solely relying on parameter aggregation, we leverage generative prototypes to effectively balance the predictions of the global model. Our method significantly improves the current State Of The Art, providing an average increase of +7.8% in accuracy.", "date": "2024-06-04", "categories": []}
{"id": "2412.00776", "title": "Learning Mamba as a Continual Learner: Meta-learning Selective State Space Models for Efficient Continual Learning", "abstract": "Continual learning (CL) aims to efficiently learn from a non-stationary data stream, without storing or recomputing all seen samples. CL enables prediction on new tasks by incorporating sequential training samples. Building on this connection between CL and sequential modeling, meta-continual learning (MCL) aims to meta-learn an efficient continual learner as a sequence prediction model, with advanced sequence models like Transformers being natural choices. However, despite decent performance, Transformers rely on a linearly growing cache to store all past representations, conflicting with CL's objective of not storing all seen samples and limiting efficiency. In this paper, we focus on meta-learning sequence-prediction-based continual learners without retaining all past representations. While attention-free models with fixed-size hidden states (e.g., Linear Transformers) align with CL's essential goal and efficiency needs, they have shown limited effectiveness in MCL in previous literature. Given Mamba's strong sequence modeling performance and attention-free nature, we explore a key question: Can attention-free models like Mamba perform well on MCL? By formulating Mamba and the SSM for MCL tasks, we propose MambaCL, a meta-learned continual learner. To enhance MambaCL's training, we introduce selectivity regularization, leveraging the connection between Mamba and Transformers to guide its behavior over sequences. Furthermore, we study how Mamba and other models perform across various MCL scenarios through extensive and well-designed experiments. Our results highlight the promising performance and strong generalization of Mamba and attention-free models in MCL, demonstrating its potential for efficient continual learning and adaptation.", "date": "2024-12-01", "categories": []}
{"id": "2505.18568", "title": "Learning without Isolation: Pathway Protection for Continual Learning", "abstract": "Deep networks are prone to catastrophic forgetting during sequential task learning, i.e., losing the knowledge about old tasks upon learning new tasks. To this end, continual learning(CL) has emerged, whose existing methods focus mostly on regulating or protecting the parameters associated with the previous tasks. However, parameter protection is often impractical, since the size of parameters for storing the old-task knowledge increases linearly with the number of tasks, otherwise it is hard to preserve the parameters related to the old-task knowledge. In this work, we bring a dual opinion from neuroscience and physics to CL: in the whole networks, the pathways matter more than the parameters when concerning the knowledge acquired from the old tasks. Following this opinion, we propose a novel CL framework, learning without isolation(LwI), where model fusion is formulated as graph matching and the pathways occupied by the old tasks are protected without being isolated. Thanks to the sparsity of activation channels in a deep network, LwI can adaptively allocate available pathways for a new task, realizing pathway protection and addressing catastrophic forgetting in a parameter-efficient manner. Experiments on popular benchmark datasets demonstrate the superiority of the proposed LwI.", "date": "2025-05-24", "categories": []}
{"id": "2505.19565", "title": "Few-Shot Class-Incremental Learning For Efficient SAR Automatic Target Recognition", "abstract": "Synthetic aperture radar automatic target recognition (SAR-ATR) systems have rapidly evolved to tackle incremental recognition challenges in operational settings. Data scarcity remains a major hurdle that conventional SAR-ATR techniques struggle to address. To cope with this challenge, we propose a few-shot class-incremental learning (FSCIL) framework based on a dual-branch architecture that focuses on local feature extraction and leverages the discrete Fourier transform and global filters to capture long-term spatial dependencies. This incorporates a lightweight cross-attention mechanism that fuses domain-specific features with global dependencies to ensure robust feature interaction, while maintaining computational efficiency by introducing minimal scale-shift parameters. The framework combines focal loss for class distinction under imbalance and center loss for compact intra-class distributions to enhance class separation boundaries. Experimental results on the MSTAR benchmark dataset demonstrate that the proposed framework consistently outperforms state-of-the-art methods in FSCIL SAR-ATR, attesting to its effectiveness in real-world scenarios.", "date": "2025-05-26", "categories": []}
{"id": "2505.19680", "title": "Cut out and Replay: A Simple yet Versatile Strategy for Multi-Label Online Continual Learning", "abstract": "Multi-Label Online Continual Learning (MOCL) requires models to learn continuously from endless multi-label data streams, facing complex challenges including persistent catastrophic forgetting, potential missing labels, and uncontrollable imbalanced class distributions. While existing MOCL methods attempt to address these challenges through various techniques, \\textit{they all overlook label-specific region identifying and feature learning} - a fundamental solution rooted in multi-label learning but challenging to achieve in the online setting with incremental and partial supervision. To this end, we first leverage the inherent structural information of input data to evaluate and verify the innate localization capability of different pre-trained models. Then, we propose CUTER (CUT-out-and-Experience-Replay), a simple yet versatile strategy that provides fine-grained supervision signals by further identifying, strengthening and cutting out label-specific regions for efficient experience replay. It not only enables models to simultaneously address catastrophic forgetting, missing labels, and class imbalance challenges, but also serves as an orthogonal solution that seamlessly integrates with existing approaches. Extensive experiments on multiple multi-label image benchmarks demonstrate the superiority of our proposed method. The code is available at \\href{https://github.com/wxr99/Cut-Replay}{https://github.com/wxr99/Cut-Replay}", "date": "2025-05-26", "categories": []}
{"id": "2503.04655", "title": "CLDyB: Towards Dynamic Benchmarking for Continual Learning with Pre-trained Models", "abstract": "The advent of the foundation model era has sparked significant research interest in leveraging pre-trained representations for continual learning (CL), yielding a series of top-performing CL methods on standard evaluation benchmarks. Nonetheless, there are growing concerns regarding potential data contamination during the pre-training stage. Furthermore, standard evaluation benchmarks, which are typically static, fail to capture the complexities of real-world CL scenarios, resulting in saturated performance. To address these issues, we describe CL on dynamic benchmarks (CLDyB), a general computational framework based on Markov decision processes for evaluating CL methods reliably. CLDyB dynamically identifies inherently difficult and algorithm-dependent tasks for the given CL methods, and determines challenging task orders using Monte Carlo tree search. Leveraging CLDyB, we first conduct a joint evaluation of multiple state-of-the-art CL methods, leading to a set of commonly challenging and generalizable task sequences where existing CL methods tend to perform poorly. We then conduct separate evaluations of individual CL methods using CLDyB, discovering their respective strengths and weaknesses. The source code and generated task sequences are publicly accessible at https://github.com/szc12153/CLDyB.", "date": "2025-03-06", "categories": []}
{"id": "2505.17496", "title": "Analyzing Mitigation Strategies for Catastrophic Forgetting in End-to-End Training of Spoken Language Models", "abstract": "End-to-end training of Spoken Language Models (SLMs) commonly involves adapting pre-trained text-based Large Language Models (LLMs) to the speech modality through multi-stage training on diverse tasks such as ASR, TTS and spoken question answering (SQA). Although this multi-stage continual learning equips LLMs with both speech understanding and generation capabilities, the substantial differences in task and data distributions across stages can lead to catastrophic forgetting, where previously acquired knowledge is lost. This paper investigates catastrophic forgetting and evaluates three mitigation strategies-model merging, discounting the LoRA scaling factor, and experience replay to balance knowledge retention with new learning. Results show that experience replay is the most effective, with further gains achieved by combining it with other methods. These findings provide insights for developing more robust and efficient SLM training pipelines.", "date": "2025-05-23", "categories": []}
{"id": "2505.17664", "title": "What is the role of memorization in Continual Learning?", "abstract": "Memorization impacts the performance of deep learning algorithms. Prior works have studied memorization primarily in the context of generalization and privacy. This work studies the memorization effect on incremental learning scenarios. Forgetting prevention and memorization seem similar. However, one should discuss their differences. We designed extensive experiments to evaluate the impact of memorization on continual learning. We clarified that learning examples with high memorization scores are forgotten faster than regular samples. Our findings also indicated that memorization is necessary to achieve the highest performance. However, at low memory regimes, forgetting regular samples is more important. We showed that the importance of a high-memorization score sample rises with an increase in the buffer size. We introduced a memorization proxy and employed it in the buffer policy problem to showcase how memorization could be used during incremental training. We demonstrated that including samples with a higher proxy memorization score is beneficial when the buffer size is large.", "date": "2025-05-23", "categories": []}
{"id": "2505.16860", "title": "GCAL: Adapting Graph Models to Evolving Domain Shifts", "abstract": "This paper addresses the challenge of graph domain adaptation on evolving, multiple out-of-distribution (OOD) graphs. Conventional graph domain adaptation methods are confined to single-step adaptation, making them ineffective in handling continuous domain shifts and prone to catastrophic forgetting. This paper introduces the Graph Continual Adaptive Learning (GCAL) method, designed to enhance model sustainability and adaptability across various graph domains. GCAL employs a bilevel optimization strategy. The \"adapt\" phase uses an information maximization approach to fine-tune the model with new graph domains while re-adapting past memories to mitigate forgetting. Concurrently, the \"generate memory\" phase, guided by a theoretical lower bound derived from information bottleneck theory, involves a variational memory graph generation module to condense original graphs into memories. Extensive experimental evaluations demonstrate that GCAL substantially outperforms existing methods in terms of adaptability and knowledge retention.", "date": "2025-05-22", "categories": []}
{"id": "2505.15178", "title": "A Unified Gradient-based Framework for Task-agnostic Continual Learning-Unlearning", "abstract": "Recent advancements in deep models have highlighted the need for intelligent systems that combine continual learning (CL) for knowledge acquisition with machine unlearning (MU) for data removal, forming the Continual Learning-Unlearning (CLU) paradigm. While existing work treats CL and MU as separate processes, we reveal their intrinsic connection through a unified optimization framework based on Kullback-Leibler divergence minimization. This framework decomposes gradient updates for approximate CLU into four components: learning new knowledge, unlearning targeted data, preserving existing knowledge, and modulation via weight saliency. A critical challenge lies in balancing knowledge update and retention during sequential learning-unlearning cycles. To resolve this stability-plasticity dilemma, we introduce a remain-preserved manifold constraint to induce a remaining Hessian compensation for CLU iterations. A fast-slow weight adaptation mechanism is designed to efficiently approximate the second-order optimization direction, combined with adaptive weighting coefficients and a balanced weight saliency mask, proposing a unified implementation framework for gradient-based CLU. Furthermore, we pioneer task-agnostic CLU scenarios that support fine-grained unlearning at the cross-task category and random sample levels beyond the traditional task-aware setups. Experiments demonstrate that the proposed UG-CLU framework effectively coordinates incremental learning, precise unlearning, and knowledge stability across multiple datasets and model architectures, providing a theoretical foundation and methodological support for dynamic, compliant intelligent systems.", "date": "2025-05-21", "categories": []}
{"id": "2405.18069", "title": "An Empirical Analysis of Forgetting in Pre-trained Models with Incremental Low-Rank Updates", "abstract": "Broad, open source availability of large pretrained foundation models on the internet through platforms such as HuggingFace has taken the world of practical deep learning by storm. A classical pipeline for neural network training now typically consists of finetuning these pretrained network on a small target dataset instead of training from scratch. In the case of large models this can be done even on modest hardware using a low rank training technique known as Low-Rank Adaptation (LoRA). While Low Rank training has already been studied in the continual learning setting, existing works often consider storing the learned adapter along with the existing model but rarely attempt to modify the weights of the pretrained model by merging the LoRA with the existing weights after finishing the training of each task. In this article we investigate this setting and study the impact of LoRA rank on the forgetting of the pretraining foundation task and on the plasticity and forgetting of subsequent ones. We observe that this rank has an important impact on forgetting of both the pretraining and downstream tasks. We also observe that vision transformers finetuned in that way exhibit a sort of ``contextual'' forgetting, a behaviour that we do not observe for residual networks and that we believe has not been observed yet in previous continual learning works.", "date": "2024-05-28", "categories": []}
{"id": "2407.13911", "title": "Continual Distillation Learning: Knowledge Distillation in Prompt-based Continual Learning", "abstract": "We introduce the problem of continual distillation learning (CDL) in order to use knowledge distillation (KD) to improve prompt-based continual learning (CL) models. The CDL problem is valuable to study since the use of a larger vision transformer (ViT) leads to better performance in prompt-based continual learning. The distillation of knowledge from a large ViT to a small ViT improves the inference efficiency for prompt-based CL models. We empirically found that existing KD methods such as logit distillation and feature distillation cannot effectively improve the student model in the CDL setup. To address this issue, we introduce a novel method named Knowledge Distillation based on Prompts (KDP), in which globally accessible prompts specifically designed for knowledge distillation are inserted into the frozen ViT backbone of the student model. We demonstrate that our KDP method effectively enhances the distillation performance in comparison to existing KD methods in the CDL setup.", "date": "2024-07-18", "categories": []}
{"id": "2411.14318", "title": "Velocitune: A Velocity-based Dynamic Domain Reweighting Method for Continual Pre-training", "abstract": "It is well-known that a diverse corpus is critical for training large language models, which are typically constructed from a mixture of various domains. In general, previous efforts resort to sampling training data from different domains with static proportions, as well as adjusting data proportions during training. However, few methods have addressed the complexities of domain-adaptive continual pre-training. To fill this gap, we propose Velocitune, a novel framework dynamically assesses learning velocity and adjusts data proportions accordingly, favoring slower-learning domains while shunning faster-learning ones, which is guided by a scaling law to indicate the desired learning goal for each domain with less associated cost. To evaluate the effectiveness of Velocitune, we conduct experiments in a reasoning-focused dataset with CodeLlama, as well as in a corpus specialised for system command generation with Llama3 and Mistral. Velocitune achieves performance gains in both math and code reasoning tasks and command-line generation benchmarks. Further analysis reveals that key factors driving Velocitune's effectiveness include target loss prediction and data ordering.", "date": "2024-11-21", "categories": []}
{"id": "2503.07035", "title": "Universal Incremental Learning: Mitigating Confusion from Inter- and Intra-task Distribution Randomness", "abstract": "Incremental learning (IL) aims to overcome catastrophic forgetting of previous tasks while learning new ones. Existing IL methods make strong assumptions that the incoming task type will either only increases new classes or domains (i.e. Class IL, Domain IL), or increase by a static scale in a class- and domain-agnostic manner (i.e. Versatile IL (VIL)), which greatly limit their applicability in the unpredictable and dynamic wild. In this work, we investigate $\\textbf{Universal Incremental Learning (UIL)}$, where a model neither knows which new classes or domains will increase along sequential tasks, nor the scale of the increments within each task. This uncertainty prevents the model from confidently learning knowledge from all task distributions and symmetrically focusing on the diverse knowledge within each task distribution. Consequently, UIL presents a more general and realistic IL scenario, making the model face confusion arising from inter-task and intra-task distribution randomness. To $\\textbf{Mi}$tigate both $\\textbf{Co}$nfusion, we propose a simple yet effective framework for UIL, named $\\textbf{MiCo}$. At the inter-task distribution level, we employ a multi-objective learning scheme to enforce accurate and deterministic predictions, and its effectiveness is further enhanced by a direction recalibration module that reduces conflicting gradients. Moreover, at the intra-task distribution level, we introduce a magnitude recalibration module to alleviate asymmetrical optimization towards imbalanced class distribution. Extensive experiments on three benchmarks demonstrate the effectiveness of our method, outperforming existing state-of-the-art methods in both the UIL scenario and the VIL scenario. Our code will be available at $\\href{https://github.com/rolsheng/UIL}{here}$.", "date": "2025-03-10", "categories": []}
{"id": "2505.12043", "title": "MoL for LLMs: Dual-Loss Optimization to Enhance Domain Expertise While Preserving General Capabilities", "abstract": "Although large language models (LLMs) perform well in general tasks, domain-specific applications suffer from hallucinations and accuracy limitations. Continual Pre-Training (CPT) approaches encounter two key issues: (1) domain-biased data degrades general language skills, and (2) improper corpus-mixture ratios limit effective adaptation. To address these, we propose a novel framework, Mixture of Losses (MoL), which decouples optimization objectives for domain-specific and general corpora. Specifically, cross-entropy (CE) loss is applied to domain-corpus to ensure knowledge acquisition, while Kullback-Leibler (KL) divergence aligns general-corpus training with the base model's foundational capabilities. This dual-loss architecture preserves universal skills while enhancing domain expertise, avoiding catastrophic forgetting. Empirically, we validate that a 1:1 domain-to-general corpus ratio optimally balances training and overfitting without the need for extensive tuning or resource-intensive experiments. Furthermore, our experiments demonstrate significant performance gains compared to traditional CPT approaches, which often suffer from degradation in general language capabilities; our model achieves 27.9% higher accuracy on the Math-500 benchmark in the non-think reasoning mode, and an impressive 83.3% improvement on the challenging AIME25 subset in the think mode, underscoring the effectiveness of our approach.", "date": "2025-05-17", "categories": []}
{"id": "2410.00645", "title": "LoRanPAC: Low-rank Random Features and Pre-trained Models for Bridging Theory and Practice in Continual Learning", "abstract": "The goal of continual learning (CL) is to train a model that can solve multiple tasks presented sequentially. Recent CL approaches have achieved strong performance by leveraging large pre-trained models that generalize well to downstream tasks. However, such methods lack theoretical guarantees, making them prone to unexpected failures. Conversely, principled CL approaches often fail to achieve competitive performance. In this work, we aim to bridge this gap between theory and practice by designing a simple CL method that is theoretically sound and highly performant. Specifically, we lift pre-trained features into a higher dimensional space and formulate an over-parametrized minimum-norm least-squares problem. We find that the lifted features are highly ill-conditioned, potentially leading to large training errors (numerical instability) and increased generalization errors. We address these challenges by continually truncating the singular value decomposition of the lifted features. Our approach, termed LoRanPAC, is stable with respect to the choice of hyperparameters, can handle hundreds of tasks, and outperforms state-of-the-art CL methods on multiple datasets. Importantly, our method satisfies a recurrence relation throughout its continual learning process, which allows us to prove it maintains small training and test errors by appropriately truncating a fraction of SVD factors. This results in a stable continual learning method with strong empirical performance and theoretical guarantees. Code available: https://github.com/liangzu/loranpac.", "date": "2024-10-01", "categories": []}
{"id": "2503.05423", "title": "Semantic Shift Estimation via Dual-Projection and Classifier Reconstruction for Exemplar-Free Class-Incremental Learning", "abstract": "Exemplar-Free Class-Incremental Learning (EFCIL) aims to sequentially learn from distinct categories without retaining exemplars but easily suffers from catastrophic forgetting of learned knowledge. While existing EFCIL methods leverage knowledge distillation to alleviate forgetting, they still face two critical challenges: semantic shift and decision bias. Specifically, the embeddings of old tasks shift in the embedding space after learning new tasks, and the classifier becomes biased towards new tasks due to training solely with new data, hindering the balance between old and new knowledge. To address these issues, we propose the Dual-Projection Shift Estimation and Classifier Reconstruction (DPCR) approach for EFCIL. DPCR effectively estimates semantic shift through a dual-projection, which combines a learnable transformation with a row-space projection to capture both task-wise and category-wise shifts. Furthermore, to mitigate decision bias, DPCR employs ridge regression to reformulate a classifier reconstruction process. This reconstruction exploits previous in covariance and prototype of each class after calibration with estimated shift, thereby reducing decision bias. Extensive experiments demonstrate that, on various datasets, DPCR effectively balances old and new tasks, outperforming state-of-the-art EFCIL methods. Our codes are available at https://github.com/RHe502/ICML25-DPCR.", "date": "2025-03-07", "categories": []}
{"id": "2505.08586", "title": "PrePrompt: Predictive prompting for class incremental learning", "abstract": "Class Incremental Learning (CIL) based on pre-trained models offers a promising direction for open-world continual learning. Existing methods typically rely on correlation-based strategies, where an image's classification feature is used as a query to retrieve the most related key prompts and select the corresponding value prompts for training. However, these approaches face an inherent limitation: fitting the entire feature space of all tasks with only a few trainable prompts is fundamentally challenging. We propose Predictive Prompting (PrePrompt), a novel CIL framework that circumvents correlation-based limitations by leveraging pre-trained models' natural classification ability to predict task-specific prompts. Specifically, PrePrompt decomposes CIL into a two-stage prediction framework: task-specific prompt prediction followed by label prediction. While theoretically appealing, this framework risks bias toward recent classes due to missing historical data for older classifier calibration. PrePrompt then mitigates this by incorporating feature translation, dynamically balancing stability and plasticity. Experiments across multiple benchmarks demonstrate PrePrompt's superiority over state-of-the-art prompt-based CIL methods. Code available at \\href{github.com/libo-huang/preprompt}{github.com/libo-huang/preprompt}.", "date": "2025-05-13", "categories": []}
{"id": "2505.10575", "title": "Robust Emotion Recognition via Bi-Level Self-Supervised Continual Learning", "abstract": "Emotion recognition through physiological signals such as electroencephalogram (EEG) has become an essential aspect of affective computing and provides an objective way to capture human emotions. However, physiological data characterized by cross-subject variability and noisy labels hinder the performance of emotion recognition models. Existing domain adaptation and continual learning methods struggle to address these issues, especially under realistic conditions where data is continuously streamed and unlabeled. To overcome these limitations, we propose a novel bi-level self-supervised continual learning framework, SSOCL, based on a dynamic memory buffer. This bi-level architecture iteratively refines the dynamic buffer and pseudo-label assignments to effectively retain representative samples, enabling generalization from continuous, unlabeled physiological data streams for emotion recognition. The assigned pseudo-labels are subsequently leveraged for accurate emotion prediction. Key components of the framework, including a fast adaptation module and a cluster-mapping module, enable robust learning and effective handling of evolving data streams. Experimental validation on two mainstream EEG tasks demonstrates the framework's ability to adapt to continuous data streams while maintaining strong generalization across subjects, outperforming existing approaches.", "date": "2025-05-13", "categories": []}
{"id": "2505.12245", "title": "AFCL: Analytic Federated Continual Learning for Spatio-Temporal Invariance of Non-IID Data", "abstract": "Federated Continual Learning (FCL) enables distributed clients to collaboratively train a global model from online task streams in dynamic real-world scenarios. However, existing FCL methods face challenges of both spatial data heterogeneity among distributed clients and temporal data heterogeneity across online tasks. Such data heterogeneity significantly degrades the model performance with severe spatial-temporal catastrophic forgetting of local and past knowledge. In this paper, we identify that the root cause of this issue lies in the inherent vulnerability and sensitivity of gradients to non-IID data. To fundamentally address this issue, we propose a gradient-free method, named Analytic Federated Continual Learning (AFCL), by deriving analytical (i.e., closed-form) solutions from frozen extracted features. In local training, our AFCL enables single-epoch learning with only a lightweight forward-propagation process for each client. In global aggregation, the server can recursively and efficiently update the global model with single-round aggregation. Theoretical analyses validate that our AFCL achieves spatio-temporal invariance of non-IID data. This ideal property implies that, regardless of how heterogeneous the data are distributed across local clients and online tasks, the aggregated model of our AFCL remains invariant and identical to that of centralized joint learning. Extensive experiments show the consistent superiority of our AFCL over state-of-the-art baselines across various benchmark datasets and settings.", "date": "2025-05-18", "categories": []}
{"id": "2505.12512", "title": "Scalable Strategies for Continual Learning with Replay", "abstract": "Future deep learning models will be distinguished by systems that perpetually learn through interaction, imagination, and cooperation, blurring the line between training and inference. This makes continual learning a critical challenge, as methods that efficiently maximize bidirectional transfer across learning trajectories will be essential. Replay is on track to play a foundational role in continual learning, allowing models to directly reconcile new information with past knowledge. In practice, however, replay is quite unscalable, doubling the cost of continual learning when applied naively. Moreover, the continual learning literature has not fully synchronized with the multi-task fine-tuning literature, having not fully integrated highly scalable techniques like model merging and low rank adaptation into a replay-enabled toolset that can produce a unified model in the face of many sequential tasks. In this paper, we begin by applying and analyzing low rank adaptation in a continual learning setting. Next, we introduce consolidation, a phasic approach to replay which leads to up to 55\\% less replay samples being needed for a given performance target. Then, we propose sequential merging, an offshoot of task arithmetic which is tailored to the continual learning setting and is shown to work well in combination with replay. Finally, we demonstrate that the developed strategies can operate synergistically, resulting in a highly scalable toolset that outperforms standalone variants.", "date": "2025-05-18", "categories": []}
{"id": "2503.24088", "title": "A Plasticity-Aware Method for Continual Self-Supervised Learning in Remote Sensing", "abstract": "Continual self-supervised learning (CSSL) methods have gained increasing attention in remote sensing (RS) due to their capability to learn new tasks sequentially from continuous streams of unlabeled data.\n  Existing CSSL methods, while learning new tasks, focus on preventing catastrophic forgetting. To this end, most of them use regularization strategies to retain knowledge of previous tasks. This reduces the model's ability to adapt to the data of new tasks (i.e., learning plasticity), which can degrade performance. To address this problem, in this paper, we propose a novel CSSL method that aims to learn tasks sequentially, while achieving high learning plasticity. To this end, the proposed method uses a knowledge distillation strategy with an integrated decoupling mechanism. The decoupling is achieved by first dividing the feature dimensions into task-common and task-specific parts. Then, the task-common features are forced to be correlated to ensure memory stability while the task-specific features are forced to be de-correlated facilitating the learning of new features. Experimental results show the effectiveness of the proposed method compared to CaSSLe, which is a widely used CSSL framework, with improvements of up to 1.12% in average accuracy and 2.33% in intransigence in a task-incremental scenario, and 1.24% in average accuracy and 2.01% in intransigence in a class-incremental scenario.", "date": "2025-03-31", "categories": []}
{"id": "2505.05926", "title": "Autoencoder-Based Hybrid Replay for Class-Incremental Learning", "abstract": "In class-incremental learning (CIL), effective incremental learning strategies are essential to mitigate task confusion and catastrophic forgetting, especially as the number of tasks $t$ increases. Current exemplar replay strategies impose $\\mathcal{O}(t)$ memory/compute complexities. We propose an autoencoder-based hybrid replay (AHR) strategy that leverages our new hybrid autoencoder (HAE) to function as a compressor to alleviate the requirement for large memory, achieving $\\mathcal{O}(0.1 t)$ at the worst case with the computing complexity of $\\mathcal{O}(t)$ while accomplishing state-of-the-art performance. The decoder later recovers the exemplar data stored in the latent space, rather than in raw format. Additionally, HAE is designed for both discriminative and generative modeling, enabling classification and replay capabilities, respectively. HAE adopts the charged particle system energy minimization equations and repulsive force algorithm for the incremental embedding and distribution of new class centroids in its latent space. Our results demonstrate that AHR consistently outperforms recent baselines across multiple benchmarks while operating with the same memory/compute budgets. The source code is included in the supplementary material and will be open-sourced upon publication.", "date": "2025-05-09", "categories": []}
{"id": "2505.07450", "title": "Prototype Augmented Hypernetworks for Continual Learning", "abstract": "Continual learning (CL) aims to learn a sequence of tasks without forgetting prior knowledge, but gradient updates for a new task often overwrite the weights learned earlier, causing catastrophic forgetting (CF). We propose Prototype-Augmented Hypernetworks (PAH), a framework where a single hypernetwork, conditioned on learnable task prototypes, dynamically generates task-specific classifier heads on demand. To mitigate forgetting, PAH combines cross-entropy with dual distillation losses, one to align logits and another to align prototypes, ensuring stable feature representations across tasks. Evaluations on Split-CIFAR100 and TinyImageNet demonstrate that PAH achieves state-of-the-art performance, reaching 74.5 % and 63.7 % accuracy with only 1.7 % and 4.4 % forgetting, respectively, surpassing prior methods without storing samples or heads.", "date": "2025-05-12", "categories": []}
{"id": "2505.10649", "title": "Advancing Multiple Instance Learning with Continual Learning for Whole Slide Imaging", "abstract": "Advances in medical imaging and deep learning have propelled progress in whole slide image (WSI) analysis, with multiple instance learning (MIL) showing promise for efficient and accurate diagnostics. However, conventional MIL models often lack adaptability to evolving datasets, as they rely on static training that cannot incorporate new information without extensive retraining. Applying continual learning (CL) to MIL models is a possible solution, but often sees limited improvements. In this paper, we analyze CL in the context of attention MIL models and find that the model forgetting is mainly concentrated in the attention layers of the MIL model. Using the results of this analysis we propose two components for improving CL on MIL: Attention Knowledge Distillation (AKD) and the Pseudo-Bag Memory Pool (PMP). AKD mitigates catastrophic forgetting by focusing on retaining attention layer knowledge between learning sessions, while PMP reduces the memory footprint by selectively storing only the most informative patches, or ``pseudo-bags'' from WSIs. Experimental evaluations demonstrate that our method significantly improves both accuracy and memory efficiency on diverse WSI datasets, outperforming current state-of-the-art CL methods. This work provides a foundation for CL in large-scale, weakly annotated clinical datasets, paving the way for more adaptable and resilient diagnostic models.", "date": "2025-05-15", "categories": []}
{"id": "2505.10677", "title": "A Conformal Predictive Measure for Assessing Catastrophic Forgetting", "abstract": "This work introduces a novel methodology for assessing catastrophic forgetting (CF) in continual learning. We propose a new conformal prediction (CP)-based metric, termed the Conformal Prediction Confidence Factor (CPCF), to quantify and evaluate CF effectively. Our framework leverages adaptive CP to estimate forgetting by monitoring the model's confidence on previously learned tasks. This approach provides a dynamic and practical solution for monitoring and measuring CF of previous tasks as new ones are introduced, offering greater suitability for real-world applications. Experimental results on four benchmark datasets demonstrate a strong correlation between CPCF and the accuracy of previous tasks, validating the reliability and interpretability of the proposed metric. Our results highlight the potential of CPCF as a robust and effective tool for assessing and understanding CF in dynamic learning environments.", "date": "2025-05-15", "categories": []}
{"id": "2505.10941", "title": "Privacy-Aware Lifelong Learning", "abstract": "Lifelong learning algorithms enable models to incrementally acquire new knowledge without forgetting previously learned information. Contrarily, the field of machine unlearning focuses on explicitly forgetting certain previous knowledge from pretrained models when requested, in order to comply with data privacy regulations on the right-to-be-forgotten. Enabling efficient lifelong learning with the capability to selectively unlearn sensitive information from models presents a critical and largely unaddressed challenge with contradicting objectives. We address this problem from the perspective of simultaneously preventing catastrophic forgetting and allowing forward knowledge transfer during task-incremental learning, while ensuring exact task unlearning and minimizing memory requirements, based on a single neural network model to be adapted. Our proposed solution, privacy-aware lifelong learning (PALL), involves optimization of task-specific sparse subnetworks with parameter sharing within a single architecture. We additionally utilize an episodic memory rehearsal mechanism to facilitate exact unlearning without performance degradations. We empirically demonstrate the scalability of PALL across various architectures in image classification, and provide a state-of-the-art solution that uniquely integrates lifelong learning and privacy-aware unlearning mechanisms for responsible AI applications.", "date": "2025-05-16", "categories": []}
{"id": "2412.13779", "title": "Rehearsal-Free Continual Federated Learning with Synergistic Synaptic Intelligence", "abstract": "Continual Federated Learning (CFL) allows distributed devices to collaboratively learn novel concepts from continuously shifting training data while avoiding knowledge forgetting of previously seen tasks. To tackle this challenge, most current CFL approaches rely on extensive rehearsal of previous data. Despite effectiveness, rehearsal comes at a cost to memory, and it may also violate data privacy. Considering these, we seek to apply regularization techniques to CFL by considering their cost-efficient properties that do not require sample caching or rehearsal. Specifically, we first apply traditional regularization techniques to CFL and observe that existing regularization techniques, especially synaptic intelligence, can achieve promising results under homogeneous data distribution but fail when the data is heterogeneous. Based on this observation, we propose a simple yet effective regularization algorithm for CFL named FedSSI, which tailors the synaptic intelligence for the CFL with heterogeneous data settings. FedSSI can not only reduce computational overhead without rehearsal but also address the data heterogeneity issue. Extensive experiments show that FedSSI achieves superior performance compared to state-of-the-art methods.", "date": "2024-12-18", "categories": []}
{"id": "2505.09925", "title": "Reinforced Interactive Continual Learning via Real-time Noisy Human Feedback", "abstract": "This paper introduces an interactive continual learning paradigm where AI models dynamically learn new skills from real-time human feedback while retaining prior knowledge. This paradigm distinctively addresses two major limitations of traditional continual learning: (1) dynamic model updates using streaming, real-time human-annotated data, rather than static datasets with fixed labels, and (2) the assumption of clean labels, by explicitly handling the noisy feedback common in real-world interactions. To tackle these problems, we propose RiCL, a Reinforced interactive Continual Learning framework leveraging Large Language Models (LLMs) to learn new skills effectively from dynamic feedback. RiCL incorporates three key components: a temporal consistency-aware purifier to automatically discern clean from noisy samples in data streams; an interaction-aware direct preference optimization strategy to align model behavior with human intent by reconciling AI-generated and human-provided feedback; and a noise-resistant contrastive learning module that captures robust representations by exploiting inherent data relationships, thus avoiding reliance on potentially unreliable labels. Extensive experiments on two benchmark datasets (FewRel and TACRED), contaminated with realistic noise patterns, demonstrate that our RiCL approach substantially outperforms existing combinations of state-of-the-art online continual learning and noisy-label learning methods.", "date": "2025-05-15", "categories": []}
{"id": "2505.09952", "title": "Task-Core Memory Management and Consolidation for Long-term Continual Learning", "abstract": "In this paper, we focus on a long-term continual learning (CL) task, where a model learns sequentially from a stream of vast tasks over time, acquiring new knowledge while retaining previously learned information in a manner akin to human learning. Unlike traditional CL settings, long-term CL involves handling a significantly larger number of tasks, which exacerbates the issue of catastrophic forgetting. Our work seeks to address two critical questions: 1) How do existing CL methods perform in the context of long-term CL? and 2) How can we mitigate the catastrophic forgetting that arises from prolonged sequential updates? To tackle these challenges, we propose a novel framework inspired by human memory mechanisms for long-term continual learning (Long-CL). Specifically, we introduce a task-core memory management strategy to efficiently index crucial memories and adaptively update them as learning progresses. Additionally, we develop a long-term memory consolidation mechanism that selectively retains hard and discriminative samples, ensuring robust knowledge retention. To facilitate research in this area, we construct and release two multi-modal and textual benchmarks, MMLongCL-Bench and TextLongCL-Bench, providing a valuable resource for evaluating long-term CL approaches. Experimental results show that Long-CL outperforms the previous state-of-the-art by 7.4\\% and 6.5\\% AP on the two benchmarks, respectively, demonstrating the effectiveness of our approach.", "date": "2025-05-15", "categories": []}
{"id": "2505.10040", "title": "Instance-Prototype Affinity Learning for Non-Exemplar Continual Graph Learning", "abstract": "Graph Neural Networks (GNN) endure catastrophic forgetting, undermining their capacity to preserve previously acquired knowledge amid the assimilation of novel information. Rehearsal-based techniques revisit historical examples, adopted as a principal strategy to alleviate this phenomenon. However, memory explosion and privacy infringements impose significant constraints on their utility. Non-Exemplar methods circumvent the prior issues through Prototype Replay (PR), yet feature drift presents new challenges. In this paper, our empirical findings reveal that Prototype Contrastive Learning (PCL) exhibits less pronounced drift than conventional PR. Drawing upon PCL, we propose Instance-Prototype Affinity Learning (IPAL), a novel paradigm for Non-Exemplar Continual Graph Learning (NECGL). Exploiting graph structural information, we formulate Topology-Integrated Gaussian Prototypes (TIGP), guiding feature distributions towards high-impact nodes to augment the model's capacity for assimilating new knowledge. Instance-Prototype Affinity Distillation (IPAD) safeguards task memory by regularizing discontinuities in class relationships. Moreover, we embed a Decision Boundary Perception (DBP) mechanism within PCL, fostering greater inter-class discriminability. Evaluations on four node classification benchmark datasets demonstrate that our method outperforms existing state-of-the-art methods, achieving a better trade-off between plasticity and stability.", "date": "2025-05-15", "categories": []}
{"id": "2505.09003", "title": "Continual Reinforcement Learning via Autoencoder-Driven Task and New Environment Recognition", "abstract": "Continual learning for reinforcement learning agents remains a significant challenge, particularly in preserving and leveraging existing information without an external signal to indicate changes in tasks or environments. In this study, we explore the effectiveness of autoencoders in detecting new tasks and matching observed environments to previously encountered ones. Our approach integrates policy optimization with familiarity autoencoders within an end-to-end continual learning system. This system can recognize and learn new tasks or environments while preserving knowledge from earlier experiences and can selectively retrieve relevant knowledge when re-encountering a known environment. Initial results demonstrate successful continual learning without external signals to indicate task changes or reencounters, showing promise for this methodology.", "date": "2025-05-13", "categories": []}
{"id": "2505.09486", "title": "Preserving Plasticity in Continual Learning with Adaptive Linearity Injection", "abstract": "Loss of plasticity in deep neural networks is the gradual reduction in a model's capacity to incrementally learn and has been identified as a key obstacle to learning in non-stationary problem settings. Recent work has shown that deep linear networks tend to be resilient towards loss of plasticity. Motivated by this observation, we propose Adaptive Linearization (AdaLin), a general approach that dynamically adapts each neuron's activation function to mitigate plasticity loss. Unlike prior methods that rely on regularization or periodic resets, AdaLin equips every neuron with a learnable parameter and a gating mechanism that injects linearity into the activation function based on its gradient flow. This adaptive modulation ensures sufficient gradient signal and sustains continual learning without introducing additional hyperparameters or requiring explicit task boundaries. When used with conventional activation functions like ReLU, Tanh, and GeLU, we demonstrate that AdaLin can significantly improve performance on standard benchmarks, including Random Label and Permuted MNIST, Random Label and Shuffled CIFAR-10, and Class-Split CIFAR-100. Furthermore, its efficacy is shown in more complex scenarios, such as class-incremental learning on CIFAR-100 with a ResNet-18 backbone, and in mitigating plasticity loss in off-policy reinforcement learning agents. We perform a systematic set of ablations that show that neuron-level adaptation is crucial for good performance and analyze a number of metrics in the network that might be correlated to loss of plasticity.", "date": "2025-05-14", "categories": []}
{"id": "2505.08327", "title": "Low-Complexity Inference in Continual Learning via Compressed Knowledge Transfer", "abstract": "Continual learning (CL) aims to train models that can learn a sequence of tasks without forgetting previously acquired knowledge. A core challenge in CL is balancing stability -- preserving performance on old tasks -- and plasticity -- adapting to new ones. Recently, large pre-trained models have been widely adopted in CL for their ability to support both, offering strong generalization for new tasks and resilience against forgetting. However, their high computational cost at inference time limits their practicality in real-world applications, especially those requiring low latency or energy efficiency. To address this issue, we explore model compression techniques, including pruning and knowledge distillation (KD), and propose two efficient frameworks tailored for class-incremental learning (CIL), a challenging CL setting where task identities are unavailable during inference. The pruning-based framework includes pre- and post-pruning strategies that apply compression at different training stages. The KD-based framework adopts a teacher-student architecture, where a large pre-trained teacher transfers downstream-relevant knowledge to a compact student. Extensive experiments on multiple CIL benchmarks demonstrate that the proposed frameworks achieve a better trade-off between accuracy and inference complexity, consistently outperforming strong baselines. We further analyze the trade-offs between the two frameworks in terms of accuracy and efficiency, offering insights into their use across different scenarios.", "date": "2025-05-13", "categories": []}
{"id": "2404.13576", "title": "I2CANSAY:Inter-Class Analogical Augmentation and Intra-Class Significance Analysis for Non-Exemplar Online Task-Free Continual Learning", "abstract": "Online task-free continual learning (OTFCL) is a more challenging variant of continual learning which emphasizes the gradual shift of task boundaries and learns in an online mode. Existing methods rely on a memory buffer composed of old samples to prevent forgetting. However,the use of memory buffers not only raises privacy concerns but also hinders the efficient learning of new samples. To address this problem, we propose a novel framework called I2CANSAY that gets rid of the dependence on memory buffers and efficiently learns the knowledge of new data from one-shot samples. Concretely, our framework comprises two main modules. Firstly, the Inter-Class Analogical Augmentation (ICAN) module generates diverse pseudo-features for old classes based on the inter-class analogy of feature distributions for different new classes, serving as a substitute for the memory buffer. Secondly, the Intra-Class Significance Analysis (ISAY) module analyzes the significance of attributes for each class via its distribution standard deviation, and generates the importance vector as a correction bias for the linear classifier, thereby enhancing the capability of learning from new samples. We run our experiments on four popular image classification datasets: CoRe50, CIFAR-10, CIFAR-100, and CUB-200, our approach outperforms the prior state-of-the-art by a large margin.", "date": "2024-04-21", "categories": []}
{"id": "2505.06603", "title": "ReplayCAD: Generative Diffusion Replay for Continual Anomaly Detection", "abstract": "Continual Anomaly Detection (CAD) enables anomaly detection models in learning new classes while preserving knowledge of historical classes. CAD faces two key challenges: catastrophic forgetting and segmentation of small anomalous regions. Existing CAD methods store image distributions or patch features to mitigate catastrophic forgetting, but they fail to preserve pixel-level detailed features for accurate segmentation. To overcome this limitation, we propose ReplayCAD, a novel diffusion-driven generative replay framework that replay high-quality historical data, thus effectively preserving pixel-level detailed features. Specifically, we compress historical data by searching for a class semantic embedding in the conditional space of the pre-trained diffusion model, which can guide the model to replay data with fine-grained pixel details, thus improving the segmentation performance. However, relying solely on semantic features results in limited spatial diversity. Hence, we further use spatial features to guide data compression, achieving precise control of sample space, thereby generating more diverse data. Our method achieves state-of-the-art performance in both classification and segmentation, with notable improvements in segmentation: 11.5% on VisA and 8.1% on MVTec. Our source code is available at https://github.com/HULEI7/ReplayCAD.", "date": "2025-05-10", "categories": []}
{"id": "2505.07690", "title": "Beyond CLIP Generalization: Against Forward&Backward Forgetting Adapter for Continual Learning of Vision-Language Models", "abstract": "This study aims to address the problem of multi-domain task incremental learning~(MTIL), which requires that vision-language models~(VLMs) continuously acquire new knowledge while maintaining their inherent zero-shot recognition capability. Existing paradigms delegate the testing of unseen-domain samples to the original CLIP, which only prevents the degradation of the model's zero-shot capability but fails to enhance the generalization of the VLM further. To this end, we propose a novel MTIL framework, named AFA, which comprises two core modules: (1) an against forward-forgetting adapter that learns task-invariant information for each dataset in the incremental tasks to enhance the zero-shot recognition ability of VLMs; (2) an against backward-forgetting adapter that strengthens the few-shot learning capability of VLMs while supporting incremental learning. Extensive experiments demonstrate that the AFA method significantly outperforms existing state-of-the-art approaches, especially in few-shot MTIL tasks, and surpasses the inherent zero-shot performance of CLIP in terms of transferability. The code is provided in the Supplementary Material.", "date": "2025-05-12", "categories": []}
{"id": "2505.04787", "title": "Replay to Remember (R2R): An Efficient Uncertainty-driven Unsupervised Continual Learning Framework Using Generative Replay", "abstract": "Continual Learning entails progressively acquiring knowledge from new data while retaining previously acquired knowledge, thereby mitigating ``Catastrophic Forgetting'' in neural networks. Our work presents a novel uncertainty-driven Unsupervised Continual Learning framework using Generative Replay, namely ``Replay to Remember (R2R)''. The proposed R2R architecture efficiently uses unlabelled and synthetic labelled data in a balanced proportion using a cluster-level uncertainty-driven feedback mechanism and a VLM-powered generative replay module. Unlike traditional memory-buffer methods that depend on pretrained models and pseudo-labels, our R2R framework operates without any prior training. It leverages visual features from unlabeled data and adapts continuously using clustering-based uncertainty estimation coupled with dynamic thresholding. Concurrently, a generative replay mechanism along with DeepSeek-R1 powered CLIP VLM produces labelled synthetic data representative of past experiences, resembling biological visual thinking that replays memory to remember and act in new, unseen tasks. Extensive experimental analyses are carried out in CIFAR-10, CIFAR-100, CINIC-10, SVHN and TinyImageNet datasets. Our proposed R2R approach improves knowledge retention, achieving a state-of-the-art performance of 98.13%, 73.06%, 93.41%, 95.18%, 59.74%, respectively, surpassing state-of-the-art performance by over 4.36%.", "date": "2025-05-07", "categories": []}
{"id": "2501.13584", "title": "Towards Robust Incremental Learning under Ambiguous Supervision", "abstract": "  Traditional Incremental Learning (IL) targets to handle sequential\nfully-supervised learning problems where novel classes emerge from time to\ntime. However, due to inherent annotation uncertainty and ambiguity, collecting\nhigh-quality annotated data in a dynamic learning system can be extremely\nexpensive. To mitigate this problem, we propose a novel weakly-supervised\nlearning paradigm called Incremental Partial Label Learning (IPLL), where the\nsequentially arrived data relate to a set of candidate labels rather than the\nground truth. Technically, we develop the Prototype-Guided Disambiguation and\nReplay Algorithm (PGDR) which leverages the class prototypes as a proxy to\nmitigate two intertwined challenges in IPLL, i.e., label ambiguity and\ncatastrophic forgetting. To handle the former, PGDR encapsulates a\nmomentum-based pseudo-labeling algorithm along with prototype-guided\ninitialization, resulting in a balanced perception of classes. To alleviate\nforgetting, we develop a memory replay technique that collects\nwell-disambiguated samples while maintaining representativeness and diversity.\nBy jointly distilling knowledge from curated memory data, our framework\nexhibits a great disambiguation ability for samples of new tasks and achieves\nless forgetting of knowledge. Extensive experiments demonstrate that PGDR\nachieves superior\n", "date": "2025-01-23", "categories": []}
{"id": "2505.04575", "title": "Componential Prompt-Knowledge Alignment for Domain Incremental Learning", "abstract": "  Domain Incremental Learning (DIL) aims to learn from non-stationary data\nstreams across domains while retaining and utilizing past knowledge. Although\nprompt-based methods effectively store multi-domain knowledge in prompt\nparameters and obtain advanced performance through cross-domain prompt fusion,\nwe reveal an intrinsic limitation: component-wise misalignment between\ndomain-specific prompts leads to conflicting knowledge integration and degraded\npredictions. This arises from the random positioning of knowledge components\nwithin prompts, where irrelevant component fusion introduces interference.To\naddress this, we propose Componential Prompt-Knowledge Alignment (KA-Prompt), a\nnovel prompt-based DIL method that introduces component-aware prompt-knowledge\nalignment during training, significantly improving both the learning and\ninference capacity of the model. KA-Prompt operates in two phases: (1) Initial\nComponential Structure Configuring, where a set of old prompts containing\nknowledge relevant to the new domain are mined via greedy search, which is then\nexploited to initialize new prompts to achieve reusable knowledge transfer and\nestablish intrinsic alignment between new and old prompts. (2) Online Alignment\nPreservation, which dynamically identifies the target old prompts and applies\nadaptive componential consistency constraints as new prompts evolve. Extensive\nexperiments on DIL benchmarks demonstrate the effectiveness of our KA-Prompt.\nOur source code is available at\nhttps://github.com/zhoujiahuan1991/ICML2025-KA-Prompt\n", "date": "2025-05-07", "categories": []}
{"id": "2405.15444", "title": "HINT: Hypernetwork Approach to Training Weight Interval Regions in\n  Continual Learning", "abstract": "  Recently, a new Continual Learning (CL) paradigm was presented to control\ncatastrophic forgetting, called Interval Continual Learning (InterContiNet),\nwhich relies on enforcing interval constraints on the neural network parameter\nspace. Unfortunately, InterContiNet training is challenging due to the high\ndimensionality of the weight space, making intervals difficult to manage. To\naddress this issue, we introduce HINT, a technique that employs interval\narithmetic within the embedding space and utilizes a hypernetwork to map these\nintervals to the target network parameter space. We train interval embeddings\nfor consecutive tasks and train a hypernetwork to transform these embeddings\ninto weights of the target network. An embedding for a given task is trained\nalong with the hypernetwork, preserving the response of the target network for\nthe previous task embeddings. Interval arithmetic works with a more manageable,\nlower-dimensional embedding space rather than directly preparing intervals in a\nhigh-dimensional weight space. Our model allows faster and more efficient\ntraining. Furthermore, HINT maintains the guarantee of not forgetting. At the\nend of training, we can choose one universal embedding to produce a single\nnetwork dedicated to all tasks. In such a framework, hypernetwork is used only\nfor training and, finally, we can utilize one set of weights. HINT obtains\nsignificantly better results than InterContiNet and gives SOTA results on\nseveral benchmarks.\n", "date": "2024-05-24", "categories": []}
{"id": "2412.13840", "title": "Unleashing the Power of Continual Learning on Non-Centralized Devices: A\n  Survey", "abstract": "  Non-Centralized Continual Learning (NCCL) has become an emerging paradigm for\nenabling distributed devices such as vehicles and servers to handle streaming\ndata from a joint non-stationary environment. To achieve high reliability and\nscalability in deploying this paradigm in distributed systems, it is essential\nto conquer challenges stemming from both spatial and temporal dimensions,\nmanifesting as distribution shifts, catastrophic forgetting, heterogeneity, and\nprivacy issues. This survey focuses on a comprehensive examination of the\ndevelopment of the non-centralized continual learning algorithms and the\nreal-world deployment across distributed devices. We begin with an introduction\nto the background and fundamentals of non-centralized learning and continual\nlearning. Then, we review existing solutions from three levels to represent how\nexisting techniques alleviate the catastrophic forgetting and distribution\nshift. Additionally, we delve into the various types of heterogeneity issues,\nsecurity, and privacy attributes, as well as real-world applications across\nthree prevalent scenarios. Furthermore, we establish a large-scale benchmark to\nrevisit this problem and analyze the performance of the state-of-the-art NCCL\napproaches. Finally, we discuss the important challenges and future research\ndirections in NCCL.\n", "date": "2024-12-18", "categories": []}
{"id": "2504.08982", "title": "Adaptive Additive Parameter Updates of Vision Transformers for Few-Shot\n  Continual Learning", "abstract": "  Integrating new class information without losing previously acquired\nknowledge remains a central challenge in artificial intelligence, often\nreferred to as catastrophic forgetting. Few-shot class incremental learning\n(FSCIL) addresses this by first training a model on a robust dataset of base\nclasses and then incrementally adapting it in successive sessions using only a\nfew labeled examples per novel class. However, this approach is prone to\noverfitting on the limited new data, which can compromise overall performance\nand exacerbate forgetting. In this work, we propose a simple yet effective\nnovel FSCIL framework that leverages a frozen Vision Transformer (ViT) backbone\naugmented with parameter-efficient additive updates. Our approach freezes the\npre-trained ViT parameters and selectively injects trainable weights into the\nself-attention modules via an additive update mechanism. This design updates\nonly a small subset of parameters to accommodate new classes without\nsacrificing the representations learned during the base session. By fine-tuning\na limited number of parameters, our method preserves the generalizable features\nin the frozen ViT while reducing the risk of overfitting. Furthermore, as most\nparameters remain fixed, the model avoids overwriting previously learned\nknowledge when small novel data batches are introduced. Extensive experiments\non benchmark datasets demonstrate that our approach yields state-of-the-art\nperformance compared to baseline FSCIL methods.\n", "date": "2025-04-11", "categories": []}
{"id": "2505.02486", "title": "SEFE: Superficial and Essential Forgetting Eliminator for Multimodal\n  Continual Instruction Tuning", "abstract": "  Multimodal Continual Instruction Tuning (MCIT) aims to enable Multimodal\nLarge Language Models (MLLMs) to incrementally learn new tasks without\ncatastrophic forgetting. In this paper, we explore forgetting in this context,\ncategorizing it into superficial forgetting and essential forgetting.\nSuperficial forgetting refers to cases where the model's knowledge may not be\ngenuinely lost, but its responses to previous tasks deviate from expected\nformats due to the influence of subsequent tasks' answer styles, making the\nresults unusable. By contrast, essential forgetting refers to situations where\nthe model provides correctly formatted but factually inaccurate answers,\nindicating a true loss of knowledge. Assessing essential forgetting\nnecessitates addressing superficial forgetting first, as severe superficial\nforgetting can obscure the model's knowledge state. Hence, we first introduce\nthe Answer Style Diversification (ASD) paradigm, which defines a standardized\nprocess for transforming data styles across different tasks, unifying their\ntraining sets into similarly diversified styles to prevent superficial\nforgetting caused by style shifts. Building on this, we propose RegLoRA to\nmitigate essential forgetting. RegLoRA stabilizes key parameters where prior\nknowledge is primarily stored by applying regularization, enabling the model to\nretain existing competencies. Experimental results demonstrate that our overall\nmethod, SEFE, achieves state-of-the-art performance.\n", "date": "2025-05-05", "categories": []}
{"id": "2505.00365", "title": "SacFL: Self-Adaptive Federated Continual Learning for\n  Resource-Constrained End Devices", "abstract": "  The proliferation of end devices has led to a distributed computing paradigm,\nwherein on-device machine learning models continuously process diverse data\ngenerated by these devices. The dynamic nature of this data, characterized by\ncontinuous changes or data drift, poses significant challenges for on-device\nmodels. To address this issue, continual learning (CL) is proposed, enabling\nmachine learning models to incrementally update their knowledge and mitigate\ncatastrophic forgetting. However, the traditional centralized approach to CL is\nunsuitable for end devices due to privacy and data volume concerns. In this\ncontext, federated continual learning (FCL) emerges as a promising solution,\npreserving user data locally while enhancing models through collaborative\nupdates. Aiming at the challenges of limited storage resources for CL, poor\nautonomy in task shift detection, and difficulty in coping with new adversarial\ntasks in FCL scenario, we propose a novel FCL framework named SacFL. SacFL\nemploys an Encoder-Decoder architecture to separate task-robust and\ntask-sensitive components, significantly reducing storage demands by retaining\nlightweight task-sensitive components for resource-constrained end devices.\nMoreover, $\\rm{SacFL}$ leverages contrastive learning to introduce an\nautonomous data shift detection mechanism, enabling it to discern whether a new\ntask has emerged and whether it is a benign task. This capability ultimately\nallows the device to autonomously trigger CL or attack defense strategy without\nadditional information, which is more practical for end devices. Comprehensive\nexperiments conducted on multiple text and image datasets, such as Cifar100 and\nTHUCNews, have validated the effectiveness of $\\rm{SacFL}$ in both\nclass-incremental and domain-incremental scenarios. Furthermore, a demo system\nhas been developed to verify its practicality.\n", "date": "2025-05-01", "categories": []}
{"id": "2504.21239", "title": "Memorization and Knowledge Injection in Gated LLMs", "abstract": "  Large Language Models (LLMs) currently struggle to sequentially add new\nmemories and integrate new knowledge. These limitations contrast with the human\nability to continuously learn from new experiences and acquire knowledge\nthroughout life. Most existing approaches add memories either through large\ncontext windows or external memory buffers (e.g., Retrieval-Augmented\nGeneration), and studies on knowledge injection rarely test scenarios\nresembling everyday life events. In this work, we introduce a continual\nlearning framework, Memory Embedded in Gated LLMs (MEGa), which injects event\nmemories directly into the weights of LLMs. Each memory is stored in a\ndedicated set of gated low-rank weights. During inference, a gating mechanism\nactivates relevant memory weights by matching query embeddings to stored memory\nembeddings. This enables the model to both recall entire memories and answer\nrelated questions. On two datasets - fictional characters and Wikipedia events\n- MEGa outperforms baseline approaches in mitigating catastrophic forgetting.\nOur model draws inspiration from the complementary memory system of the human\nbrain.\n", "date": "2025-04-30", "categories": []}
{"id": "2504.21565", "title": "Towards proactive self-adaptive AI for non-stationary environments with\n  dataset shifts", "abstract": "  Artificial Intelligence (AI) models deployed in production frequently face\nchallenges in maintaining their performance in non-stationary environments.\nThis issue is particularly noticeable in medical settings, where temporal\ndataset shifts often occur. These shifts arise when the distributions of\ntraining data differ from those of the data encountered during deployment over\ntime. Further, new labeled data to continuously retrain AI is not typically\navailable in a timely manner due to data access limitations. To address these\nchallenges, we propose a proactive self-adaptive AI approach, or pro-adaptive,\nwhere we model the temporal trajectory of AI parameters, allowing us to\nshort-term forecast parameter values. To this end, we use polynomial spline\nbases, within an extensible Functional Data Analysis framework. We validate our\nmethodology with a logistic regression model addressing prior probability\nshift, covariate shift, and concept shift. This validation is conducted on both\na controlled simulated dataset and a publicly available real-world COVID-19\ndataset from Mexico, with various shifts occurring between 2020 and 2024. Our\nresults indicate that this approach enhances the performance of AI against\nshifts compared to baseline stable models trained at different time distances\nfrom the present, without requiring updated training data. This work lays the\nfoundation for pro-adaptive AI research against dynamic, non-stationary\nenvironments, being compatible with data protection, in resilient AI production\nenvironments for health.\n", "date": "2025-04-30", "categories": []}
{"id": "2504.20109", "title": "Personalized Artificial General Intelligence (AGI) via\n  Neuroscience-Inspired Continuous Learning Systems", "abstract": "  Artificial Intelligence has made remarkable advancements in recent years,\nprimarily driven by increasingly large deep learning models. However, achieving\ntrue Artificial General Intelligence (AGI) demands fundamentally new\narchitectures rather than merely scaling up existing models. Current approaches\nlargely depend on expanding model parameters, which improves task-specific\nperformance but falls short in enabling continuous, adaptable, and generalized\nlearning. Achieving AGI capable of continuous learning and personalization on\nresource-constrained edge devices is an even bigger challenge.\n  This paper reviews the state of continual learning and neuroscience-inspired\nAI, and proposes a novel architecture for Personalized AGI that integrates\nbrain-like learning mechanisms for edge deployment. We review literature on\ncontinuous lifelong learning, catastrophic forgetting, and edge AI, and discuss\nkey neuroscience principles of human learning, including Synaptic Pruning,\nHebbian plasticity, Sparse Coding, and Dual Memory Systems, as inspirations for\nAI systems. Building on these insights, we outline an AI architecture that\nfeatures complementary fast-and-slow learning modules, synaptic\nself-optimization, and memory-efficient model updates to support on-device\nlifelong adaptation.\n  Conceptual diagrams of the proposed architecture and learning processes are\nprovided. We address challenges such as catastrophic forgetting, memory\nefficiency, and system scalability, and present application scenarios for\nmobile AI assistants and embodied AI systems like humanoid robots. We conclude\nwith key takeaways and future research directions toward truly continual,\npersonalized AGI on the edge. While the architecture is theoretical, it\nsynthesizes diverse findings and offers a roadmap for future implementation.\n", "date": "2025-04-27", "categories": []}
{"id": "2504.20797", "title": "Partitioned Memory Storage Inspired Few-Shot Class-Incremental learning", "abstract": "  Current mainstream deep learning techniques exhibit an over-reliance on\nextensive training data and a lack of adaptability to the dynamic world,\nmarking a considerable disparity from human intelligence. To bridge this gap,\nFew-Shot Class-Incremental Learning (FSCIL) has emerged, focusing on continuous\nlearning of new categories with limited samples without forgetting old\nknowledge. Existing FSCIL studies typically use a single model to learn\nknowledge across all sessions, inevitably leading to the stability-plasticity\ndilemma. Unlike machines, humans store varied knowledge in different cerebral\ncortices. Inspired by this characteristic, our paper aims to develop a method\nthat learns independent models for each session. It can inherently prevent\ncatastrophic forgetting. During the testing stage, our method integrates\nUncertainty Quantification (UQ) for model deployment. Our method provides a\nfresh viewpoint for FSCIL and demonstrates the state-of-the-art performance on\nCIFAR-100 and mini-ImageNet datasets.\n", "date": "2025-04-29", "categories": []}
{"id": "2307.00183", "title": "Long-Tailed Continual Learning For Visual Food Recognition", "abstract": "  Deep learning-based food recognition has made significant progress in\npredicting food types from eating occasion images. However, two key challenges\nhinder real-world deployment: (1) continuously learning new food classes\nwithout forgetting previously learned ones, and (2) handling the long-tailed\ndistribution of food images, where a few common classes and many more rare\nclasses. To address these, food recognition methods should focus on long-tailed\ncontinual learning. In this work, We introduce a dataset that encompasses 186\nAmerican foods along with comprehensive annotations. We also introduce three\nnew benchmark datasets, VFN186-LT, VFN186-INSULIN and VFN186-T2D, which reflect\nreal-world food consumption for healthy populations, insulin takers and\nindividuals with type 2 diabetes without taking insulin. We propose a novel\nend-to-end framework that improves the generalization ability for instance-rare\nfood classes using a knowledge distillation-based predictor to avoid\nmisalignment of representation during continual learning. Additionally, we\nintroduce an augmentation technique by integrating class-activation-map (CAM)\nand CutMix to improve generalization on instance-rare food classes. Our method,\nevaluated on Food101-LT, VFN-LT, VFN186-LT, VFN186-INSULIN, and VFN186-T2DM,\nshows significant improvements over existing methods. An ablation study\nhighlights further performance enhancements, demonstrating its potential for\nreal-world food recognition applications.\n", "date": "2023-07-01", "categories": []}
{"id": "2409.14983", "title": "Dynamic Integration of Task-Specific Adapters for Class Incremental\n  Learning", "abstract": "  Non-exemplar class Incremental Learning (NECIL) enables models to\ncontinuously acquire new classes without retraining from scratch and storing\nold task exemplars, addressing privacy and storage issues. However, the absence\nof data from earlier tasks exacerbates the challenge of catastrophic forgetting\nin NECIL. In this paper, we propose a novel framework called Dynamic\nIntegration of task-specific Adapters (DIA), which comprises two key\ncomponents: Task-Specific Adapter Integration (TSAI) and Patch-Level Model\nAlignment. TSAI boosts compositionality through a patch-level adapter\nintegration strategy, which provides a more flexible compositional solution\nwhile maintaining low computation costs. Patch-Level Model Alignment maintains\nfeature consistency and accurate decision boundaries via two specialized\nmechanisms: Patch-Level Distillation Loss (PDL) and Patch-Level Feature\nReconstruction method (PFR). Specifically, the PDL preserves feature-level\nconsistency between successive models by implementing a distillation loss based\non the contributions of patch tokens to new class learning. The PFR facilitates\naccurate classifier alignment by reconstructing old class features from\nprevious tasks that adapt to new task knowledge. Extensive experiments validate\nthe effectiveness of our DIA, revealing significant improvements on benchmark\ndatasets in the NECIL setting, maintaining an optimal balance between\ncomputational complexity and accuracy.\n", "date": "2024-09-23", "categories": []}
{"id": "2504.12712", "title": "Convergence and Implicit Bias of Gradient Descent on Continual Linear\n  Classification", "abstract": "  We study continual learning on multiple linear classification tasks by\nsequentially running gradient descent (GD) for a fixed budget of iterations per\ntask. When all tasks are jointly linearly separable and are presented in a\ncyclic/random order, we show the directional convergence of the trained linear\nclassifier to the joint (offline) max-margin solution. This is surprising\nbecause GD training on a single task is implicitly biased towards the\nindividual max-margin solution for the task, and the direction of the joint\nmax-margin solution can be largely different from these individual solutions.\nAdditionally, when tasks are given in a cyclic order, we present a\nnon-asymptotic analysis on cycle-averaged forgetting, revealing that (1)\nalignment between tasks is indeed closely tied to catastrophic forgetting and\nbackward knowledge transfer and (2) the amount of forgetting vanishes to zero\nas the cycle repeats. Lastly, we analyze the case where the tasks are no longer\njointly separable and show that the model trained in a cyclic order converges\nto the unique minimum of the joint loss function.\n", "date": "2025-04-17", "categories": []}
{"id": "2504.19638", "title": "LODAP: On-Device Incremental Learning Via Lightweight Operations and\n  Data Pruning", "abstract": "  Incremental learning that learns new classes over time after the model's\ndeployment is becoming increasingly crucial, particularly for industrial edge\nsystems, where it is difficult to communicate with a remote server to conduct\ncomputation-intensive learning. As more classes are expected to learn after\ntheir execution for edge devices. In this paper, we propose LODAP, a new\non-device incremental learning framework for edge systems. The key part of\nLODAP is a new module, namely Efficient Incremental Module (EIM). EIM is\ncomposed of normal convolutions and lightweight operations. During incremental\nlearning, EIM exploits some lightweight operations, called adapters, to\neffectively and efficiently learn features for new classes so that it can\nimprove the accuracy of incremental learning while reducing model complexity as\nwell as training overhead. The efficiency of LODAP is further enhanced by a\ndata pruning strategy that significantly reduces the training data, thereby\nlowering the training overhead. We conducted extensive experiments on the\nCIFAR-100 and Tiny- ImageNet datasets. Experimental results show that LODAP\nimproves the accuracy by up to 4.32\\% over existing methods while reducing\naround 50\\% of model complexity. In addition, evaluations on real edge systems\ndemonstrate its applicability for on-device machine learning. The code is\navailable at https://github.com/duanbiqing/LODAP.\n", "date": "2025-04-28", "categories": []}
{"id": "2504.17963", "title": "Mathematics of Continual Learning", "abstract": "  Continual learning is an emerging subject in machine learning that aims to\nsolve multiple tasks presented sequentially to the learner without forgetting\npreviously learned tasks. Recently, many deep learning based approaches have\nbeen proposed for continual learning, however the mathematical foundations\nbehind existing continual learning methods remain underdeveloped. On the other\nhand, adaptive filtering is a classic subject in signal processing with a rich\nhistory of mathematically principled methods. However, its role in\nunderstanding the foundations of continual learning has been underappreciated.\nIn this tutorial, we review the basic principles behind both continual learning\nand adaptive filtering, and present a comparative analysis that highlights\nmultiple connections between them. These connections allow us to enhance the\nmathematical foundations of continual learning based on existing results for\nadaptive filtering, extend adaptive filtering insights using existing continual\nlearning methods, and discuss a few research directions for continual learning\nsuggested by the historical developments in adaptive filtering.\n", "date": "2025-04-24", "categories": []}
{"id": "2504.18059", "title": "POET: Prompt Offset Tuning for Continual Human Action Adaptation", "abstract": "  As extended reality (XR) is redefining how users interact with computing\ndevices, research in human action recognition is gaining prominence. Typically,\nmodels deployed on immersive computing devices are static and limited to their\ndefault set of classes. The goal of our research is to provide users and\ndevelopers with the capability to personalize their experience by adding new\naction classes to their device models continually. Importantly, a user should\nbe able to add new classes in a low-shot and efficient manner, while this\nprocess should not require storing or replaying any of user's sensitive\ntraining data. We formalize this problem as privacy-aware few-shot continual\naction recognition. Towards this end, we propose POET: Prompt-Offset Tuning.\nWhile existing prompt tuning approaches have shown great promise for continual\nlearning of image, text, and video modalities; they demand access to\nextensively pretrained transformers. Breaking away from this assumption, POET\ndemonstrates the efficacy of prompt tuning a significantly lightweight\nbackbone, pretrained exclusively on the base class data. We propose a novel\nspatio-temporal learnable prompt offset tuning approach, and are the first to\napply such prompt tuning to Graph Neural Networks. We contribute two new\nbenchmarks for our new problem setting in human action recognition: (i) NTU\nRGB+D dataset for activity recognition, and (ii) SHREC-2017 dataset for hand\ngesture recognition. We find that POET consistently outperforms comprehensive\nbenchmarks. Source code at\nhttps://github.com/humansensinglab/POET-continual-action-recognition.\n", "date": "2025-04-25", "categories": []}
{"id": "2504.18437", "title": "Enhancing Pre-Trained Model-Based Class-Incremental Learning through\n  Neural Collapse", "abstract": "  Class-Incremental Learning (CIL) is a critical capability for real-world\napplications, enabling learning systems to adapt to new tasks while retaining\nknowledge from previous ones. Recent advancements in pre-trained models (PTMs)\nhave significantly advanced the field of CIL, demonstrating superior\nperformance over traditional methods. However, understanding how features\nevolve and are distributed across incremental tasks remains an open challenge.\nIn this paper, we propose a novel approach to modeling feature evolution in\nPTM-based CIL through the lens of neural collapse (NC), a striking phenomenon\nobserved in the final phase of training, which leads to a well-separated,\nequiangular feature space. We explore the connection between NC and CIL\neffectiveness, showing that aligning feature distributions with the NC geometry\nenhances the ability to capture the dynamic behavior of continual learning.\nBased on this insight, we introduce Neural Collapse-inspired Pre-Trained\nModel-based CIL (NCPTM-CIL), a method that dynamically adjusts the feature\nspace to conform to the elegant NC structure, thereby enhancing the continual\nlearning process. Extensive experiments demonstrate that NCPTM-CIL outperforms\nstate-of-the-art methods across four benchmark datasets. Notably, when\ninitialized with ViT-B/16-IN1K, NCPTM-CIL surpasses the runner-up method by\n6.73% on VTAB, 1.25% on CIFAR-100, and 2.5% on OmniBenchmark.\n", "date": "2025-04-25", "categories": []}
{"id": "2403.11743", "title": "PARMESAN: Parameter-Free Memory Search and Transduction for Dense\n  Prediction Tasks", "abstract": "  This work addresses flexibility in deep learning by means of transductive\nreasoning. For adaptation to new data and tasks, e.g., in continual learning,\nexisting methods typically involve tuning learnable parameters or complete\nre-training from scratch, rendering such approaches unflexible in practice. We\nargue that the notion of separating computation from memory by the means of\ntransduction can act as a stepping stone for solving these issues. We therefore\npropose PARMESAN (parameter-free memory search and transduction), a scalable\nmethod which leverages a memory module for solving dense prediction tasks. At\ninference, hidden representations in memory are being searched to find\ncorresponding patterns. In contrast to other methods that rely on continuous\ntraining of learnable parameters, PARMESAN learns via memory consolidation\nsimply by modifying stored contents. Our method is compatible with commonly\nused architectures and canonically transfers to 1D, 2D, and 3D grid-based data.\nThe capabilities of our approach are demonstrated at the complex task of\ncontinual learning. PARMESAN learns by 3-4 orders of magnitude faster than\nestablished baselines while being on par in terms of predictive performance,\nhardware-efficiency, and knowledge retention.\n", "date": "2024-03-18", "categories": []}
{"id": "2504.17780", "title": "Replay to Remember: Retaining Domain Knowledge in Streaming Language\n  Models", "abstract": "  Continual learning in large language models (LLMs) typically encounters the\ncritical challenge of catastrophic forgetting, where previously acquired\nknowledge deteriorates upon exposure to new data. While techniques like replay\nbuffers and parameter-efficient tuning (e.g., Low-Rank Adaptation or LoRA) have\nbeen proposed, few studies investigate real-time domain adaptation under strict\ncomputational and data-stream constraints. In this paper, we demonstrate a\nlightweight method combining LoRA and a minimal replay mechanism in a realistic\nstreaming setting across three diverse knowledge domains: medical question\nanswering, genetics, and law. Using perplexity, semantic similarity, and\nGPT-based human-like evaluation metrics, we quantify the model's adaptation,\nforgetting, and recovery over time. Our experiments reveal that while\ncatastrophic forgetting naturally occurs, even minimal replay significantly\nstabilizes and partially restores domain-specific knowledge. This study\ncontributes practical insights for deploying adaptable LLMs in\nresource-constrained, real-world scenarios.\n", "date": "2025-04-24", "categories": []}
{"id": "2504.16501", "title": "Dynamic Time-aware Continual User Representation Learning", "abstract": "  Traditional user modeling (UM) approaches have primarily focused on designing\nmodels for a single specific task, but they face limitations in generalization\nand adaptability across various tasks. Recognizing these challenges, recent\nstudies have shifted towards continual learning (CL)-based universal user\nrepresentation learning aiming to develop a single model capable of handling\nmultiple tasks. Despite advancements, existing methods are in fact evaluated\nunder an unrealistic scenario that does not consider the passage of time as\ntasks progress, which overlooks newly emerged items that may change the item\ndistribution of previous tasks. In this paper, we introduce a practical\nevaluation scenario on which CL-based universal user representation learning\napproaches should be evaluated, which takes into account the passage of time as\ntasks progress. Then, we propose a novel framework Dynamic Time-aware continual\nuser representation learner, named DITTO, designed to alleviate catastrophic\nforgetting despite continuous shifts in item distribution, while also allowing\nthe knowledge acquired from previous tasks to adapt to the current shifted item\ndistribution. Through our extensive experiments, we demonstrate the superiority\nof DITTO over state-of-the-art methods under a practical evaluation scenario.\nOur source code is available at\nhttps://github.com/seungyoon-Choi/DITTO_official.\n", "date": "2025-04-23", "categories": []}
{"id": "2504.16763", "title": "Noise-Tolerant Coreset-Based Class Incremental Continual Learning", "abstract": "  Many applications of computer vision require the ability to adapt to novel\ndata distributions after deployment. Adaptation requires algorithms capable of\ncontinual learning (CL). Continual learners must be plastic to adapt to novel\ntasks while minimizing forgetting of previous tasks.However, CL opens up\navenues for noise to enter the training pipeline and disrupt the CL. This work\nfocuses on label noise and instance noise in the context of class-incremental\nlearning (CIL), where new classes are added to a classifier over time, and\nthere is no access to external data from past classes. We aim to understand the\nsensitivity of CL methods that work by replaying items from a memory\nconstructed using the idea of Coresets. We derive a new bound for the\nrobustness of such a method to uncorrelated instance noise under a general\nadditive noise threat model, revealing several insights. Putting the theory\ninto practice, we create two continual learning algorithms to construct\nnoise-tolerant replay buffers. We empirically compare the effectiveness of\nprior memory-based continual learners and the proposed algorithms under label\nand uncorrelated instance noise on five diverse datasets. We show that existing\nmemory-based CL are not robust whereas the proposed methods exhibit significant\nimprovements in maximizing classification accuracy and minimizing forgetting in\nthe noisy CIL setting.\n", "date": "2025-04-23", "categories": []}
{"id": "2503.17436", "title": "On-Device Federated Continual Learning on RISC-V-based Ultra-Low-Power\n  SoC for Intelligent Nano-Drone Swarms", "abstract": "  RISC-V-based architectures are paving the way for efficient On-Device\nLearning (ODL) in smart edge devices. When applied across multiple nodes, ODL\nenables the creation of intelligent sensor networks that preserve data privacy.\nHowever, developing ODL-capable, battery-operated embedded platforms presents\nsignificant challenges due to constrained computational resources and limited\ndevice lifetime, besides intrinsic learning issues such as catastrophic\nforgetting. We face these challenges by proposing a regularization-based\nOn-Device Federated Continual Learning algorithm tailored for multiple\nnano-drones performing face recognition tasks. We demonstrate our approach on a\nRISC-V-based 10-core ultra-low-power SoC, optimizing the ODL computational\nrequirements. We improve the classification accuracy by 24% over naive\nfine-tuning, requiring 178 ms per local epoch and 10.5 s per global epoch,\ndemonstrating the effectiveness of the architecture for this task.\n", "date": "2025-03-21", "categories": []}
{"id": "2504.15328", "title": "Bayesian Federated Learning for Continual Training", "abstract": "  Bayesian Federated Learning (BFL) enables uncertainty quantification and\nrobust adaptation in distributed learning. In contrast to the frequentist\napproach, it estimates the posterior distribution of a global model, offering\ninsights into model reliability. However, current BFL methods neglect continual\nlearning challenges in dynamic environments where data distributions shift over\ntime. We propose a continual BFL framework applied to human sensing with radar\ndata collected over several days. Using Stochastic Gradient Langevin Dynamics\n(SGLD), our approach sequentially updates the model, leveraging past posteriors\nto construct the prior for the new tasks. We assess the accuracy, the expected\ncalibration error (ECE) and the convergence speed of our approach against\nseveral baselines. Results highlight the effectiveness of continual Bayesian\nupdates in preserving knowledge and adapting to evolving data.\n", "date": "2025-04-21", "categories": []}
{"id": "2504.15561", "title": "SPECI: Skill Prompts based Hierarchical Continual Imitation Learning for\n  Robot Manipulation", "abstract": "  Real-world robot manipulation in dynamic unstructured environments requires\nlifelong adaptability to evolving objects, scenes and tasks. Traditional\nimitation learning relies on static training paradigms, which are ill-suited\nfor lifelong adaptation. Although Continual Imitation Learnin (CIL) enables\nincremental task adaptation while preserving learned knowledge, current CIL\nmethods primarily overlook the intrinsic skill characteristics of robot\nmanipulation or depend on manually defined and rigid skills, leading to\nsuboptimal cross-task knowledge transfer. To address these issues, we propose\nSkill Prompts-based HiErarchical Continual Imitation Learning (SPECI), a novel\nend-to-end hierarchical CIL policy architecture for robot manipulation. The\nSPECI framework consists of a multimodal perception and fusion module for\nheterogeneous sensory information encoding, a high-level skill inference module\nfor dynamic skill extraction and selection, and a low-level action execution\nmodule for precise action generation. To enable efficient knowledge transfer on\nboth skill and task levels, SPECI performs continual implicit skill acquisition\nand reuse via an expandable skill codebook and an attention-driven skill\nselection mechanism. Furthermore, we introduce mode approximation to augment\nthe last two modules with task-specific and task-sharing parameters, thereby\nenhancing task-level knowledge transfer. Extensive experiments on diverse\nmanipulation task suites demonstrate that SPECI consistently outperforms\nstate-of-the-art CIL methods across all evaluated metrics, revealing\nexceptional bidirectional knowledge transfer and superior overall performance.\n", "date": "2025-04-22", "categories": []}
{"id": "2410.03955", "title": "A Retention-Centric Framework for Continual Learning with Guaranteed\n  Model Developmental Safety", "abstract": "  In real-world applications, learning-enabled systems often undergo iterative\nmodel development to address challenging or emerging tasks, which involve\ncollecting new data, training a new model and validating the model. This\ncontinual model development process raises a significant issue that acquiring\nnew or improving existing capabilities may inadvertently lose good capabilities\nof the old model, also known as catastrophic forgetting. While existing\ncontinual learning aims to mitigate catastrophic forgetting by trading off\nperformance on previous tasks and new tasks to ensure good average performance,\nit often falls short in cost-sensitive applications, where failing to preserve\nessential established capabilities introduces unforeseen costs and risks and\nsubstantial expenses for re-improving these capabilities. To address this\nissue, we impose a requirement on learning systems to ensure that a new model\nstrictly retains important capabilities of the old model while improving\ntarget-task performance, which we term model developmental safety. To ensure\nmodel developmental safety, we propose a retention-centric framework with\ndata-dependent constraints, and study how to continually develop a pretrained\nCLIP model for acquiring new or improving existing capabilities of image\nclassification. We propose an efficient constrained optimization algorithm with\ntheoretical guarantees and use its insights to finetune the CLIP model with\ntask-dependent heads for promoting the model developmental safety. Experiments\non autonomous driving and scene recognition datasets validate the efficacy of\nour method.\n", "date": "2024-10-04", "categories": []}
{"id": "2501.14278", "title": "Active Learning for Continual Learning: Keeping the Past Alive in the\n  Present", "abstract": "  Continual learning (CL) enables deep neural networks to adapt to\never-changing data distributions. In practice, there may be scenarios where\nannotation is costly, leading to active continual learning (ACL), which\nperforms active learning (AL) for the CL scenarios when reducing the labeling\ncost by selecting the most informative subset is preferable. However,\nconventional AL strategies are not suitable for ACL, as they focus solely on\nlearning the new knowledge, leading to catastrophic forgetting of previously\nlearned tasks. Therefore, ACL requires a new AL strategy that can balance the\nprevention of catastrophic forgetting and the ability to quickly learn new\ntasks. In this paper, we propose AccuACL, Accumulated informativeness-based\nActive Continual Learning, by the novel use of the Fisher information matrix as\na criterion for sample selection, derived from a theoretical analysis of the\nFisher-optimality preservation properties within the framework of ACL, while\nalso addressing the scalability issue of Fisher information-based AL. Extensive\nexperiments demonstrate that AccuACL significantly outperforms AL baselines\nacross various CL algorithms, increasing the average accuracy and forgetting by\n23.8% and 17.0%, respectively, on average.\n", "date": "2025-01-24", "categories": []}
{"id": "2504.14677", "title": "Evaluating Temporal Plasticity in Foundation Time Series Models for\n  Incremental Fine-tuning", "abstract": "  Time series foundation models excel at diverse time series forecasting tasks,\nbut their capacity for continuous improvement through incremental learning\nremains unexplored. We present the first comprehensive study investigating\nthese models' temporal plasticity - their ability to progressively enhance\nperformance through continual learning while maintaining existing capabilities.\nThrough experiments on real-world datasets exhibiting distribution shifts, we\nevaluate both conventional deep learning models and foundation models using a\nnovel continual learning framework. Our findings reveal that while traditional\nmodels struggle with performance deterioration during incremental fine-tuning,\nfoundation models like Time-MoE and Chronos demonstrate sustained improvement\nin predictive accuracy. This suggests that optimizing foundation model\nfine-tuning strategies may be more valuable than developing domain-specific\nsmall models. Our research introduces new evaluation methodologies and insights\nfor developing foundation time series models with robust continuous learning\ncapabilities.\n", "date": "2025-04-20", "categories": []}
{"id": "2504.14727", "title": "Semi-parametric Memory Consolidation: Towards Brain-like Deep Continual\n  Learning", "abstract": "  Humans and most animals inherently possess a distinctive capacity to\ncontinually acquire novel experiences and accumulate worldly knowledge over\ntime. This ability, termed continual learning, is also critical for deep neural\nnetworks (DNNs) to adapt to the dynamically evolving world in open\nenvironments. However, DNNs notoriously suffer from catastrophic forgetting of\npreviously learned knowledge when trained on sequential tasks. In this work,\ninspired by the interactive human memory and learning system, we propose a\nnovel biomimetic continual learning framework that integrates semi-parametric\nmemory and the wake-sleep consolidation mechanism. For the first time, our\nmethod enables deep neural networks to retain high performance on novel tasks\nwhile maintaining prior knowledge in real-world challenging continual learning\nscenarios, e.g., class-incremental learning on ImageNet. This study\ndemonstrates that emulating biological intelligence provides a promising path\nto enable deep neural networks with continual learning capabilities.\n", "date": "2025-04-20", "categories": []}
{"id": "2504.15171", "title": "Audio-Visual Class-Incremental Learning for Fish Feeding intensity\n  Assessment in Aquaculture", "abstract": "  Fish Feeding Intensity Assessment (FFIA) is crucial in industrial aquaculture\nmanagement. Recent multi-modal approaches have shown promise in improving FFIA\nrobustness and efficiency. However, these methods face significant challenges\nwhen adapting to new fish species or environments due to catastrophic\nforgetting and the lack of suitable datasets. To address these limitations, we\nfirst introduce AV-CIL-FFIA, a new dataset comprising 81,932 labelled\naudio-visual clips capturing feeding intensities across six different fish\nspecies in real aquaculture environments. Then, we pioneer audio-visual class\nincremental learning (CIL) for FFIA and demonstrate through benchmarking on\nAV-CIL-FFIA that it significantly outperforms single-modality methods. Existing\nCIL methods rely heavily on historical data. Exemplar-based approaches store\nraw samples, creating storage challenges, while exemplar-free methods avoid\ndata storage but struggle to distinguish subtle feeding intensity variations\nacross different fish species. To overcome these limitations, we introduce\nHAIL-FFIA, a novel audio-visual class-incremental learning framework that\nbridges this gap with a prototype-based approach that achieves exemplar-free\nefficiency while preserving essential knowledge through compact feature\nrepresentations. Specifically, HAIL-FFIA employs hierarchical representation\nlearning with a dual-path knowledge preservation mechanism that separates\ngeneral intensity knowledge from fish-specific characteristics. Additionally,\nit features a dynamic modality balancing system that adaptively adjusts the\nimportance of audio versus visual information based on feeding behaviour\nstages. Experimental results show that HAIL-FFIA is superior to SOTA methods on\nAV-CIL-FFIA, achieving higher accuracy with lower storage needs while\neffectively mitigating catastrophic forgetting in incremental fish species\nlearning.\n", "date": "2025-04-21", "categories": []}
{"id": "2504.13218", "title": "Harmony: A Unified Framework for Modality Incremental Learning", "abstract": "  Incremental learning aims to enable models to continuously acquire knowledge\nfrom evolving data streams while preserving previously learned capabilities.\nWhile current research predominantly focuses on unimodal incremental learning\nand multimodal incremental learning where the modalities are consistent,\nreal-world scenarios often present data from entirely new modalities, posing\nadditional challenges. This paper investigates the feasibility of developing a\nunified model capable of incremental learning across continuously evolving\nmodal sequences. To this end, we introduce a novel paradigm called Modality\nIncremental Learning (MIL), where each learning stage involves data from\ndistinct modalities. To address this task, we propose a novel framework named\nHarmony, designed to achieve modal alignment and knowledge retention, enabling\nthe model to reduce the modal discrepancy and learn from a sequence of distinct\nmodalities, ultimately completing tasks across multiple modalities within a\nunified framework. Our approach introduces the adaptive compatible feature\nmodulation and cumulative modal bridging. Through constructing historical modal\nfeatures and performing modal knowledge accumulation and alignment, the\nproposed components collaboratively bridge modal differences and maintain\nknowledge retention, even with solely unimodal data available at each learning\nphase.These components work in concert to establish effective modality\nconnections and maintain knowledge retention, even when only unimodal data is\navailable at each learning stage. Extensive experiments on the MIL task\ndemonstrate that our proposed method significantly outperforms existing\nincremental learning methods, validating its effectiveness in MIL scenarios.\n", "date": "2025-04-17", "categories": []}
{"id": "2504.13407", "title": "LoRA-Based Continual Learning with Constraints on Critical Parameter\n  Changes", "abstract": "  LoRA-based continual learning represents a promising avenue for leveraging\npre-trained models in downstream continual learning tasks. Recent studies have\nshown that orthogonal LoRA tuning effectively mitigates forgetting. However,\nthis work unveils that under orthogonal LoRA tuning, the critical parameters\nfor pre-tasks still change notably after learning post-tasks. To address this\nproblem, we directly propose freezing the most critical parameter matrices in\nthe Vision Transformer (ViT) for pre-tasks before learning post-tasks. In\naddition, building on orthogonal LoRA tuning, we propose orthogonal LoRA\ncomposition (LoRAC) based on QR decomposition, which may further enhance the\nplasticity of our method. Elaborate ablation studies and extensive comparisons\ndemonstrate the effectiveness of our proposed method. Our results indicate that\nour method achieves state-of-the-art (SOTA) performance on several well-known\ncontinual learning benchmarks. For instance, on the Split CIFAR-100 dataset,\nour method shows a 6.35\\% improvement in accuracy and a 3.24\\% reduction in\nforgetting compared to previous methods. Our code is available at\nhttps://github.com/learninginvision/LoRAC-IPC.\n", "date": "2025-04-18", "categories": []}
{"id": "2504.13569", "title": "Bayesian continual learning and forgetting in neural networks", "abstract": "  Biological synapses effortlessly balance memory retention and flexibility,\nyet artificial neural networks still struggle with the extremes of catastrophic\nforgetting and catastrophic remembering. Here, we introduce Metaplasticity from\nSynaptic Uncertainty (MESU), a Bayesian framework that updates network\nparameters according their uncertainty. This approach allows a principled\ncombination of learning and forgetting that ensures that critical knowledge is\npreserved while unused or outdated information is gradually released. Unlike\nstandard Bayesian approaches -- which risk becoming overly constrained, and\npopular continual-learning methods that rely on explicit task boundaries, MESU\nseamlessly adapts to streaming data. It further provides reliable epistemic\nuncertainty estimates, allowing out-of-distribution detection, the only\ncomputational cost being to sample the weights multiple times to provide proper\noutput statistics. Experiments on image-classification benchmarks demonstrate\nthat MESU mitigates catastrophic forgetting, while maintaining plasticity for\nnew tasks. When training 200 sequential permuted MNIST tasks, MESU outperforms\nestablished continual learning techniques in terms of accuracy, capability to\nlearn additional tasks, and out-of-distribution data detection. Additionally,\ndue to its non-reliance on task boundaries, MESU outperforms conventional\nlearning techniques on the incremental training of CIFAR-100 tasks consistently\nin a wide range of scenarios. Our results unify ideas from metaplasticity,\nBayesian inference, and Hessian-based regularization, offering a\nbiologically-inspired pathway to robust, perpetual learning.\n", "date": "2025-04-18", "categories": []}
{"id": "2407.05229", "title": "HiDe-PET: Continual Learning via Hierarchical Decomposition of\n  Parameter-Efficient Tuning", "abstract": "  The deployment of pre-trained models (PTMs) has greatly advanced the field of\ncontinual learning (CL), enabling positive knowledge transfer and resilience to\ncatastrophic forgetting. To sustain these advantages for sequentially arriving\ntasks, a promising direction involves keeping the pre-trained backbone frozen\nwhile employing parameter-efficient tuning (PET) techniques to instruct\nrepresentation learning. Despite the popularity of Prompt-based PET for CL, its\nempirical design often leads to sub-optimal performance in our evaluation of\ndifferent PTMs and target tasks. To this end, we propose a unified framework\nfor CL with PTMs and PET that provides both theoretical and empirical\nadvancements. We first perform an in-depth theoretical analysis of the CL\nobjective in a pre-training context, decomposing it into hierarchical\ncomponents namely within-task prediction, task-identity inference and\ntask-adaptive prediction. We then present Hierarchical Decomposition PET\n(HiDe-PET), an innovative approach that explicitly optimizes the decomposed\nobjective through incorporating task-specific and task-shared knowledge via\nmainstream PET techniques along with efficient recovery of pre-trained\nrepresentations. Leveraging this framework, we delve into the distinct impacts\nof implementation strategy, PET technique and PET architecture, as well as\nadaptive knowledge accumulation amidst pronounced distribution changes.\nFinally, across various CL scenarios, our approach demonstrates remarkably\nsuperior performance over a broad spectrum of recent strong baselines.\n", "date": "2024-07-07", "categories": []}
{"id": "2504.10976", "title": "Adaptive Decision Boundary for Few-Shot Class-Incremental Learning", "abstract": "  Few-Shot Class-Incremental Learning (FSCIL) aims to continuously learn new\nclasses from a limited set of training samples without forgetting knowledge of\npreviously learned classes. Conventional FSCIL methods typically build a robust\nfeature extractor during the base training session with abundant training\nsamples and subsequently freeze this extractor, only fine-tuning the classifier\nin subsequent incremental phases. However, current strategies primarily focus\non preventing catastrophic forgetting, considering only the relationship\nbetween novel and base classes, without paying attention to the specific\ndecision spaces of each class. To address this challenge, we propose a\nplug-and-play Adaptive Decision Boundary Strategy (ADBS), which is compatible\nwith most FSCIL methods. Specifically, we assign a specific decision boundary\nto each class and adaptively adjust these boundaries during training to\noptimally refine the decision spaces for the classes in each session.\nFurthermore, to amplify the distinctiveness between classes, we employ a novel\ninter-class constraint loss that optimizes the decision boundaries and\nprototypes for each class. Extensive experiments on three benchmarks, namely\nCIFAR100, miniImageNet, and CUB200, demonstrate that incorporating our ADBS\nmethod with existing FSCIL techniques significantly improves performance,\nachieving overall state-of-the-art results.\n", "date": "2025-04-15", "categories": []}
{"id": "2504.12503", "title": "Continual Learning Strategies for 3D Engineering Regression Problems: A\n  Benchmarking Study", "abstract": "  Engineering problems that apply machine learning often involve\ncomputationally intensive methods but rely on limited datasets. As engineering\ndata evolves with new designs and constraints, models must incorporate new\nknowledge over time. However, high computational costs make retraining models\nfrom scratch infeasible. Continual learning (CL) offers a promising solution by\nenabling models to learn from sequential data while mitigating catastrophic\nforgetting, where a model forgets previously learned mappings. This work\nintroduces CL to engineering design by benchmarking several CL methods on\nrepresentative regression tasks. We apply these strategies to five engineering\ndatasets and construct nine new engineering CL benchmarks to evaluate their\nability to address forgetting and improve generalization. Preliminary results\nshow that applying existing CL methods to these tasks improves performance over\nnaive baselines. In particular, the Replay strategy achieved performance\ncomparable to retraining in several benchmarks while reducing training time by\nnearly half, demonstrating its potential for real-world engineering workflows.\nThe code and datasets used in this work will be available at:\nhttps://github.com/kmsamuel/cl-for-engineering-release.\n", "date": "2025-04-16", "categories": []}
{"id": "2410.15954", "title": "TS-ACL: Closed-Form Solution for Time Series-oriented Continual Learning", "abstract": "  Time series classification underpins critical applications such as healthcare\ndiagnostics and gesture-driven interactive systems in multimedia scenarios.\nHowever, time series class-incremental learning (TSCIL) faces two major\nchallenges: catastrophic forgetting and intra-class variations. Catastrophic\nforgetting occurs because gradient-based parameter update strategies inevitably\nerase past knowledge. And unlike images, time series data exhibits\nsubject-specific patterns, also known as intra-class variations, which refer to\ndifferences in patterns observed within the same class. While exemplar-based\nmethods fail to cover diverse variation with limited samples, existing\nexemplar-free methods lack explicit mechanisms to handle intra-class\nvariations. To address these two challenges, we propose TS-ACL, which leverages\na gradient-free closed-form solution to avoid the catastrophic forgetting\nproblem inherent in gradient-based optimization methods while simultaneously\nlearning global distributions to resolve intra-class variations. Additionally,\nit provides privacy protection and efficiency. Extensive experiments on five\nbenchmark datasets covering various sensor modalities and tasks demonstrate\nthat TS-ACL achieves performance close to joint training on four datasets,\noutperforming existing methods and establishing a new state-of-the-art (SOTA)\nfor TSCIL.\n", "date": "2024-10-21", "categories": []}
{"id": "2501.08878", "title": "Incrementally Learning Multiple Diverse Data Domains via Multi-Source\n  Dynamic Expansion Model", "abstract": "  Continual Learning seeks to develop a model capable of incrementally\nassimilating new information while retaining prior knowledge. However, current\nresearch predominantly addresses a straightforward learning context, wherein\nall data samples originate from a singular data domain. This paper shifts focus\nto a more complex and realistic learning environment, characterized by data\nsamples sourced from multiple distinct domains. We tackle this intricate\nlearning challenge by introducing a novel methodology, termed the Multi-Source\nDynamic Expansion Model (MSDEM), which leverages various pre-trained models as\nbackbones and progressively establishes new experts based on them to adapt to\nemerging tasks. Additionally, we propose an innovative dynamic expandable\nattention mechanism designed to selectively harness knowledge from multiple\nbackbones, thereby accelerating the new task learning. Moreover, we introduce a\ndynamic graph weight router that strategically reuses all previously acquired\nparameters and representations for new task learning, maximizing the positive\nknowledge transfer effect, which further improves generalization performance.\nWe conduct a comprehensive series of experiments, and the empirical findings\nindicate that our proposed approach achieves state-of-the-art performance.\n", "date": "2025-01-15", "categories": []}
{"id": "2502.11019", "title": "Unlocking the Power of Function Vectors for Characterizing and\n  Mitigating Catastrophic Forgetting in Continual Instruction Tuning", "abstract": "  Catastrophic forgetting (CF) poses a significant challenge in machine\nlearning, where a model forgets previously learned information upon learning\nnew tasks. Despite the advanced capabilities of Large Language Models (LLMs),\nthey continue to face challenges with CF during continual learning. The\nmajority of existing research focuses on analyzing forgetting patterns through\na singular training sequence, thereby overlooking the intricate effects that\ndiverse tasks have on model behavior. Our study explores CF across various\nsettings, discovering that model forgetting is influenced by both the specific\ntraining tasks and the models themselves. To this end, we interpret forgetting\nby examining the function vector (FV), a compact representation of functions in\nLLMs, offering a model-dependent indicator for the occurrence of CF. Through\ntheoretical and empirical analyses, we demonstrated that CF in LLMs primarily\nstems from biases in function activation rather than the overwriting of task\nprocessing functions. Leveraging these insights, we propose a novel function\nvector guided training methodology, incorporating a regularization technique to\nstabilize the FV and mitigate forgetting. Empirical tests on four benchmarks\nconfirm the effectiveness of our proposed training method, substantiating our\ntheoretical framework concerning CF and model function dynamics. We plan to\nmake our code publicly accessible in the near future.\n", "date": "2025-02-16", "categories": []}
{"id": "2504.10561", "title": "Self-Controlled Dynamic Expansion Model for Continual Learning", "abstract": "  Continual Learning (CL) epitomizes an advanced training paradigm wherein\nprior data samples remain inaccessible during the acquisition of new tasks.\nNumerous investigations have delved into leveraging a pre-trained Vision\nTransformer (ViT) to enhance model efficacy in continual learning. Nonetheless,\nthese approaches typically utilize a singular, static backbone, which\ninadequately adapts to novel tasks, particularly when engaging with diverse\ndata domains, due to a substantial number of inactive parameters. This paper\naddresses this limitation by introducing an innovative Self-Controlled Dynamic\nExpansion Model (SCDEM), which orchestrates multiple distinct trainable\npre-trained ViT backbones to furnish diverse and semantically enriched\nrepresentations. Specifically, by employing the multi-backbone architecture as\na shared module, the proposed SCDEM dynamically generates a new expert with\nminimal parameters to accommodate a new task. A novel Collaborative\nOptimization Mechanism (COM) is introduced to synergistically optimize multiple\nbackbones by harnessing prediction signals from historical experts, thereby\nfacilitating new task learning without erasing previously acquired knowledge.\nAdditionally, a novel Feature Distribution Consistency (FDC) approach is\nproposed to align semantic similarity between previously and currently learned\nrepresentations through an optimal transport distance-based mechanism,\neffectively mitigating negative knowledge transfer effects. Furthermore, to\nalleviate over-regularization challenges, this paper presents a novel Dynamic\nLayer-Wise Feature Attention Mechanism (DLWFAM) to autonomously determine the\npenalization intensity on each trainable representation layer. An extensive\nseries of experiments have been conducted to evaluate the proposed\nmethodology's efficacy, with empirical results corroborating that the approach\nattains state-of-the-art performance.\n", "date": "2025-04-14", "categories": []}
{"id": "2504.08626", "title": "Task-conditioned Ensemble of Expert Models for Continuous Learning", "abstract": "  One of the major challenges in machine learning is maintaining the accuracy\nof the deployed model (e.g., a classifier) in a non-stationary environment. The\nnon-stationary environment results in distribution shifts and, consequently, a\ndegradation in accuracy. Continuous learning of the deployed model with new\ndata could be one remedy. However, the question arises as to how we should\nupdate the model with new training data so that it retains its accuracy on the\nold data while adapting to the new data. In this work, we propose a\ntask-conditioned ensemble of models to maintain the performance of the existing\nmodel. The method involves an ensemble of expert models based on task\nmembership information. The in-domain models-based on the local outlier concept\n(different from the expert models) provide task membership information\ndynamically at run-time to each probe sample. To evaluate the proposed method,\nwe experiment with three setups: the first represents distribution shift\nbetween tasks (LivDet-Iris-2017), the second represents distribution shift both\nbetween and within tasks (LivDet-Iris-2020), and the third represents disjoint\ndistribution between tasks (Split MNIST). The experiments highlight the\nbenefits of the proposed method. The source code is available at\nhttps://github.com/iPRoBe-lab/Continuous_Learning_FE_DM.\n", "date": "2025-04-11", "categories": []}
{"id": "2504.10409", "title": "GPS: Distilling Compact Memories via Grid-based Patch Sampling for\n  Efficient Online Class-Incremental Learning", "abstract": "  Online class-incremental learning aims to enable models to continuously adapt\nto new classes with limited access to past data, while mitigating catastrophic\nforgetting. Replay-based methods address this by maintaining a small memory\nbuffer of previous samples, achieving competitive performance. For effective\nreplay under constrained storage, recent approaches leverage distilled data to\nenhance the informativeness of memory. However, such approaches often involve\nsignificant computational overhead due to the use of bi-level optimization.\nMotivated by these limitations, we introduce Grid-based Patch Sampling (GPS), a\nlightweight and effective strategy for distilling informative memory samples\nwithout relying on a trainable model. GPS generates informative samples by\nsampling a subset of pixels from the original image, yielding compact\nlow-resolution representations that preserve both semantic content and\nstructural information. During replay, these representations are reassembled to\nsupport training and evaluation. Experiments on extensive benchmarks\ndemonstrate that GRS can be seamlessly integrated into existing replay\nframeworks, leading to 3%-4% improvements in average end accuracy under\nmemory-constrained settings, with limited computational overhead.\n", "date": "2025-04-14", "categories": []}
{"id": "2502.12736", "title": "Cross-Domain Continual Learning for Edge Intelligence in Wireless ISAC\n  Networks", "abstract": "  In wireless networks with integrated sensing and communications (ISAC), edge\nintelligence (EI) is expected to be developed at edge devices (ED) for sensing\nuser activities based on channel state information (CSI). However, due to the\nCSI being highly specific to users' characteristics, the CSI-activity\nrelationship is notoriously domain dependent, essentially demanding EI to learn\nsufficient datasets from various domains in order to gain cross-domain sensing\ncapability. This poses a crucial challenge owing to the EDs' limited resources,\nfor which storing datasets across all domains will be a significant burden. In\nthis paper, we propose the EdgeCL framework, enabling the EI to continually\nlearn-then-discard each incoming dataset, while remaining resilient to\ncatastrophic forgetting. We design a transformer-based discriminator for\nhandling sequences of noisy and nonequispaced CSI samples. Besides, we propose\na distilled core-set based knowledge retention method with robustness-enhanced\noptimization to train the discriminator, preserving its performance for\nprevious domains while preventing future forgetting. Experimental evaluations\nshow that EdgeCL achieves 89% of performance compared to cumulative training\nwhile consuming only 3% of its memory, mitigating forgetting by 79%.\n", "date": "2025-02-18", "categories": []}
{"id": "2503.14301", "title": "FeNeC: Enhancing Continual Learning via Feature Clustering with\n  Neighbor- or Logit-Based Classification", "abstract": "  The ability of deep learning models to learn continuously is essential for\nadapting to new data categories and evolving data distributions. In recent\nyears, approaches leveraging frozen feature extractors after an initial\nlearning phase have been extensively studied. Many of these methods estimate\nper-class covariance matrices and prototypes based on backbone-derived feature\nrepresentations. Within this paradigm, we introduce FeNeC (Feature Neighborhood\nClassifier) and FeNeC-Log, its variant based on the log-likelihood function.\nOur approach generalizes the existing concept by incorporating data clustering\nto capture greater intra-class variability. Utilizing the Mahalanobis distance,\nour models classify samples either through a nearest neighbor approach or\ntrainable logit values assigned to consecutive classes. Our proposition may be\nreduced to the existing approaches in a special case while extending them with\nthe ability of more flexible adaptation to data. We demonstrate that two FeNeC\nvariants achieve competitive performance in scenarios where task identities are\nunknown and establish state-of-the-art results on several benchmarks.\n", "date": "2025-03-18", "categories": []}
{"id": "2504.01648", "title": "ProtoGuard-guided PROPEL: Class-Aware Prototype Enhancement and\n  Progressive Labeling for Incremental 3D Point Cloud Segmentation", "abstract": "  3D point cloud semantic segmentation technology has been widely used.\nHowever, in real-world scenarios, the environment is evolving. Thus,\noffline-trained segmentation models may lead to catastrophic forgetting of\npreviously seen classes. Class-incremental learning (CIL) is designed to\naddress the problem of catastrophic forgetting. While point clouds are common,\nwe observe high similarity and unclear boundaries between different classes.\nMeanwhile, they are known to be imbalanced in class distribution. These lead to\nissues including misclassification between similar classes and the long-tail\nproblem, which have not been adequately addressed in previous CIL methods. We\nthus propose ProtoGuard and PROPEL (Progressive Refinement Of PsEudo-Labels).\nIn the base-class training phase, ProtoGuard maintains geometric and semantic\nprototypes for each class, which are combined into prototype features using an\nattention mechanism. In the novel-class training phase, PROPEL inherits the\nbase feature extractor and classifier, guiding pseudo-label propagation and\nupdates based on density distribution and semantic similarity. Extensive\nexperiments show that our approach achieves remarkable results on both the\nS3DIS and ScanNet datasets, improving the mIoU of 3D point cloud segmentation\nby a maximum of 20.39% under the 5-step CIL scenario on S3DIS.\n", "date": "2025-04-02", "categories": []}
{"id": "2504.06521", "title": "DUKAE: DUal-level Knowledge Accumulation and Ensemble for Pre-Trained\n  Model-Based Continual Learning", "abstract": "  Pre-trained model-based continual learning (PTMCL) has garnered growing\nattention, as it enables more rapid acquisition of new knowledge by leveraging\nthe extensive foundational understanding inherent in pre-trained model (PTM).\nMost existing PTMCL methods use Parameter-Efficient Fine-Tuning (PEFT) to learn\nnew knowledge while consolidating existing memory. However, they often face\nsome challenges. A major challenge lies in the misalignment of classification\nheads, as the classification head of each task is trained within a distinct\nfeature space, leading to inconsistent decision boundaries across tasks and,\nconsequently, increased forgetting. Another critical limitation stems from the\nrestricted feature-level knowledge accumulation, with feature learning\ntypically restricted to the initial task only, which constrains the model's\nrepresentation capabilities. To address these issues, we propose a method named\nDUal-level Knowledge Accumulation and Ensemble (DUKAE) that leverages both\nfeature-level and decision-level knowledge accumulation by aligning\nclassification heads into a unified feature space through Gaussian distribution\nsampling and introducing an adaptive expertise ensemble to fuse knowledge\nacross feature subspaces. Extensive experiments on CIFAR-100, ImageNet-R,\nCUB-200, and Cars-196 datasets demonstrate the superior performance of our\napproach.\n", "date": "2025-04-09", "categories": []}
{"id": "2504.08823", "title": "FM-LoRA: Factorized Low-Rank Meta-Prompting for Continual Learning", "abstract": "  How to adapt a pre-trained model continuously for sequential tasks with\ndifferent prediction class labels and domains and finally learn a generalizable\nmodel across diverse tasks is a long-lasting challenge. Continual learning (CL)\nhas emerged as a promising approach to leverage pre-trained models (e.g.,\nTransformers) for sequential tasks. While many existing CL methods\nincrementally store additional learned structures, such as Low-Rank Adaptation\n(LoRA) adapters or prompts and sometimes even preserve features from previous\nsamples to maintain performance. This leads to unsustainable parameter growth\nand escalating storage costs as the number of tasks increases. Moreover,\ncurrent approaches often lack task similarity awareness, which further hinders\nthe models ability to effectively adapt to new tasks without interfering with\npreviously acquired knowledge. To address these challenges, we propose FM-LoRA,\na novel and efficient low-rank adaptation method that integrates both a dynamic\nrank selector (DRS) and dynamic meta-prompting (DMP). This framework allocates\nmodel capacity more effectively across tasks by leveraging a shared low-rank\nsubspace critical for preserving knowledge, thereby avoiding continual\nparameter expansion. Extensive experiments on various CL benchmarks, including\nImageNet-R, CIFAR100, and CUB200 for class-incremental learning (CIL), and\nDomainNet for domain-incremental learning (DIL), with Transformers backbone\ndemonstrate that FM-LoRA effectively mitigates catastrophic forgetting while\ndelivering robust performance across a diverse range of tasks and domains.\n", "date": "2025-04-09", "categories": []}
{"id": "2504.10151", "title": "Continual learning for rotating machinery fault diagnosis with\n  cross-domain environmental and operational variations", "abstract": "  Although numerous machine learning models exist to detect issues like rolling\nbearing strain and deformation, typically caused by improper mounting,\noverloading, or poor lubrication, these models often struggle to isolate faults\nfrom the noise of real-world operational and environmental variability.\nConditions such as variable loads, high temperatures, stress, and rotational\nspeeds can mask early signs of failure, making reliable detection challenging.\nTo address these limitations, this work proposes a continual deep learning\napproach capable of learning across domains that share underlying structure\nover time. This approach goes beyond traditional accuracy metrics by addressing\nfour second-order challenges: catastrophic forgetting (where new learning\noverwrites past knowledge), lack of plasticity (where models fail to adapt to\nnew data), forward transfer (using past knowledge to improve future learning),\nand backward transfer (refining past knowledge with insights from new domains).\nThe method comprises a feature generator and domain-specific classifiers,\nallowing capacity to grow as new domains emerge with minimal interference,\nwhile an experience replay mechanism selectively revisits prior domains to\nmitigate forgetting. Moreover, nonlinear dependencies across domains are\nexploited by prioritizing replay from those with the highest prior errors,\nrefining models based on most informative past experiences. Experiments show\nhigh average domain accuracy (up to 88.96%), with forgetting measures as low as\n.0027 across non-stationary class-incremental environments.\n", "date": "2025-04-14", "categories": []}
{"id": "2310.08948", "title": "Federated Class-Incremental Learning with Prompting", "abstract": "  As Web technology continues to develop, it has become increasingly common to\nuse data stored on different clients. At the same time, federated learning has\nreceived widespread attention due to its ability to protect data privacy when\nlet models learn from data which is distributed across various clients.\nHowever, most existing works assume that the client's data are fixed. In\nreal-world scenarios, such an assumption is most likely not true as data may be\ncontinuously generated and new classes may also appear. To this end, we focus\non the practical and challenging federated class-incremental learning (FCIL)\nproblem. For FCIL, the local and global models may suffer from catastrophic\nforgetting on old classes caused by the arrival of new classes and the data\ndistributions of clients are non-independent and identically distributed\n(non-iid).\n  In this paper, we propose a novel method called Federated Class-Incremental\nLearning with PrompTing (FCILPT). Given the privacy and limited memory, FCILPT\ndoes not use a rehearsal-based buffer to keep exemplars of old data. We choose\nto use prompts to ease the catastrophic forgetting of the old classes.\nSpecifically, we encode the task-relevant and task-irrelevant knowledge into\nprompts, preserving the old and new knowledge of the local clients and solving\nthe problem of catastrophic forgetting. We first sort the task information in\nthe prompt pool in the local clients to align the task information on different\nclients before global aggregation. It ensures that the same task's knowledge\nare fully integrated, solving the problem of non-iid caused by the lack of\nclasses among different clients in the same incremental task. Experiments on\nCIFAR-100, Mini-ImageNet, and Tiny-ImageNet demonstrate that FCILPT achieves\nsignificant accuracy improvements over the state-of-the-art methods.\n", "date": "2023-10-13", "categories": []}
{"id": "2504.08149", "title": "LoRAX: LoRA eXpandable Networks for Continual Synthetic Image\n  Attribution", "abstract": "  As generative AI image technologies become more widespread and advanced,\nthere is a growing need for strong attribution models. These models are crucial\nfor verifying the authenticity of images and identifying the architecture of\ntheir originating generative models-key to maintaining media integrity.\nHowever, attribution models struggle to generalize to unseen models, and\ntraditional fine-tuning methods for updating these models have shown to be\nimpractical in real-world settings. To address these challenges, we propose\nLoRA eXpandable Networks (LoRAX), a parameter-efficient class incremental\nalgorithm that adapts to novel generative image models without the need for\nfull retraining. Our approach trains an extremely parameter-efficient feature\nextractor per continual learning task via Low Rank Adaptation. Each\ntask-specific feature extractor learns distinct features while only requiring a\nsmall fraction of the parameters present in the underlying feature extractor's\nbackbone model. Our extensive experimentation shows LoRAX outperforms or\nremains competitive with state-of-the-art class incremental learning algorithms\non the Continual Deepfake Detection benchmark across all training scenarios and\nmemory settings, while requiring less than 3% of the number of trainable\nparameters per feature extractor compared to the full-rank implementation.\nLoRAX code is available at: https://github.com/mit-ll/lorax_cil.\n", "date": "2025-04-10", "categories": []}
{"id": "2504.08422", "title": "CMIP-CIL: A Cross-Modal Benchmark for Image-Point Class Incremental\n  Learning", "abstract": "  Image-point class incremental learning helps the 3D-points-vision robots\ncontinually learn category knowledge from 2D images, improving their perceptual\ncapability in dynamic environments. However, some incremental learning methods\naddress unimodal forgetting but fail in cross-modal cases, while others handle\nmodal differences within training/testing datasets but assume no modal gaps\nbetween them. We first explore this cross-modal task, proposing a benchmark\nCMIP-CIL and relieving the cross-modal catastrophic forgetting problem. It\nemploys masked point clouds and rendered multi-view images within a contrastive\nlearning framework in pre-training, empowering the vision model with the\ngeneralizations of image-point correspondence. In the incremental stage, by\nfreezing the backbone and promoting object representations close to their\nrespective prototypes, the model effectively retains and generalizes knowledge\nacross previously seen categories while continuing to learn new ones. We\nconduct comprehensive experiments on the benchmark datasets. Experiments prove\nthat our method achieves state-of-the-art results, outperforming the baseline\nmethods by a large margin.\n", "date": "2025-04-11", "categories": []}
{"id": "2504.08550", "title": "Proxy-Anchor and EVT-Driven Continual Learning Method for Generalized\n  Category Discovery", "abstract": "  Continual generalized category discovery has been introduced and studied in\nthe literature as a method that aims to continuously discover and learn novel\ncategories in incoming data batches while avoiding catastrophic forgetting of\npreviously learned categories. A key component in addressing this challenge is\nthe model's ability to separate novel samples, where Extreme Value Theory (EVT)\nhas been effectively employed. In this work, we propose a novel method that\nintegrates EVT with proxy anchors to define boundaries around proxies using a\nprobability of inclusion function, enabling the rejection of unknown samples.\nAdditionally, we introduce a novel EVT-based loss function to enhance the\nlearned representation, achieving superior performance compared to other\ndeep-metric learning methods in similar settings. Using the derived probability\nfunctions, novel samples are effectively separated from previously known\ncategories. However, category discovery within these novel samples can\nsometimes overestimate the number of new categories. To mitigate this issue, we\npropose a novel EVT-based approach to reduce the model size and discard\nredundant proxies. We also incorporate experience replay and knowledge\ndistillation mechanisms during the continual learning stage to prevent\ncatastrophic forgetting. Experimental results demonstrate that our proposed\napproach outperforms state-of-the-art methods in continual generalized category\ndiscovery scenarios.\n", "date": "2025-04-11", "categories": []}
{"id": "2504.08554", "title": "Boosting-inspired online learning with transfer for railway maintenance", "abstract": "  The integration of advanced sensor technologies with deep learning algorithms\nhas revolutionized fault diagnosis in railway systems, particularly at the\nwheel-track interface. Although numerous models have been proposed to detect\nirregularities such as wheel out-of-roundness, they often fall short in\nreal-world applications due to the dynamic and nonstationary nature of railway\noperations. This paper introduces BOLT-RM (Boosting-inspired Online Learning\nwith Transfer for Railway Maintenance), a model designed to address these\nchallenges using continual learning for predictive maintenance. By allowing the\nmodel to continuously learn and adapt as new data become available, BOLT-RM\novercomes the issue of catastrophic forgetting that often plagues traditional\nmodels. It retains past knowledge while improving predictive accuracy with each\nnew learning episode, using a boosting-like knowledge sharing mechanism to\nadapt to evolving operational conditions such as changes in speed, load, and\ntrack irregularities. The methodology is validated through comprehensive\nmulti-domain simulations of train-track dynamic interactions, which capture\nrealistic railway operating conditions. The proposed BOLT-RM model demonstrates\nsignificant improvements in identifying wheel anomalies, establishing a\nreliable sequence for maintenance interventions.\n", "date": "2025-04-11", "categories": []}
{"id": "2504.08613", "title": "Enhancing knowledge retention for continual learning with\n  domain-specific adapters and features gating", "abstract": "  Continual learning empowers models to learn from a continuous stream of data\nwhile preserving previously acquired knowledge, effectively addressing the\nchallenge of catastrophic forgetting. In this study, we propose a new approach\nthat integrates adapters within the self-attention mechanisms of Vision\nTransformers to enhance knowledge retention when sequentially adding datasets\nfrom different domains. Unlike previous methods that continue learning with\nonly one dataset, our approach introduces domain-specific output heads and\nfeature gating, allowing the model to maintain high accuracy on previously\nlearned tasks while incorporating only the essential information from multiple\ndomains. The proposed method is compared to prominent parameter-efficient\nfine-tuning methods in the current state of the art. The results provide\nevidence that our method effectively alleviates the limitations of previous\nworks. Furthermore, we conduct a comparative analysis using three datasets,\nCIFAR-100, Flowers102, and DTD, each representing a distinct domain, to\ninvestigate the impact of task order on model performance. Our findings\nunderscore the critical role of dataset sequencing in shaping learning\noutcomes, demonstrating that strategic ordering can significantly improve the\nmodel's ability to adapt to evolving data distributions over time while\npreserving the integrity of previously learned knowledge.\n", "date": "2025-04-11", "categories": []}
{"id": "2504.07240", "title": "Prototype-Based Continual Learning with Label-free Replay Buffer and\n  Cluster Preservation Loss", "abstract": "  Continual learning techniques employ simple replay sample selection processes\nand use them during subsequent tasks. Typically, they rely on labeled data. In\nthis paper, we depart from this by automatically selecting prototypes stored\nwithout labels, preserving cluster structures in the latent space across tasks.\nBy eliminating label dependence in the replay buffer and introducing cluster\npreservation loss, it is demonstrated that the proposed method can maintain\nessential information from previously encountered tasks while ensuring\nadaptation to new tasks. \"Push-away\" and \"pull-toward\" mechanisms over\npreviously learned prototypes are also introduced for class-incremental and\ndomain-incremental scenarios. These mechanisms ensure the retention of\npreviously learned information as well as adaptation to new classes or domain\nshifts. The proposed method is evaluated on several benchmarks, including\nSplitCIFAR100, SplitImageNet32, SplitTinyImageNet, and SplitCaltech256 for\nclass-incremental, as well as R-MNIST and CORe50 for domain-incremental setting\nusing pre-extracted DINOv2 features. Experimental results indicate that the\nlabel-free replay-based technique outperforms state-of-the-art continual\nlearning methods and, in some cases, even surpasses offline learning. An\nunsupervised variant of the proposed technique for the class-incremental\nsetting, avoiding labels use even on incoming data, also demonstrated\ncompetitive performance, outperforming particular supervised baselines in some\ncases. These findings underscore the effectiveness of the proposed framework in\nretaining prior information and facilitating continual adaptation.\n", "date": "2025-04-09", "categories": []}
{"id": "2411.07413", "title": "ODEStream: A Buffer-Free Online Learning Framework with ODE-based\n  Adaptor for Streaming Time Series Forecasting", "abstract": "  Addressing the challenges of irregularity and concept drift in streaming time\nseries is crucial for real-world predictive modelling. Previous studies in time\nseries continual learning often propose models that require buffering long\nsequences, potentially restricting the responsiveness of the inference system.\nMoreover, these models are typically designed for regularly sampled data, an\nunrealistic assumption in real-world scenarios. This paper introduces\nODEStream, a novel buffer-free continual learning framework that incorporates a\ntemporal isolation layer to capture temporal dependencies within the data.\nSimultaneously, it leverages the capability of neural ordinary differential\nequations to process irregular sequences and generate a continuous data\nrepresentation, enabling seamless adaptation to changing dynamics in a data\nstreaming scenario. Our approach focuses on learning how the dynamics and\ndistribution of historical data change over time, facilitating direct\nprocessing of streaming sequences. Evaluations on benchmark real-world datasets\ndemonstrate that ODEStream outperforms the state-of-the-art online learning and\nstreaming analysis baseline models, providing accurate predictions over\nextended periods while minimising performance degradation over time by learning\nhow the sequence dynamics change. The implementation of ODEStream is available\nat: https://github.com/FtoonAbushaqra/ODEStream.git.\n", "date": "2024-11-11", "categories": []}
{"id": "2504.06664", "title": "SEE: Continual Fine-tuning with Sequential Ensemble of Experts", "abstract": "  Continual fine-tuning of large language models (LLMs) suffers from\ncatastrophic forgetting. Rehearsal-based methods mitigate this problem by\nretaining a small set of old data. Nevertheless, they still suffer inevitable\nperformance loss. Although training separate experts for each task can help\nprevent forgetting, effectively assembling them remains a challenge. Some\napproaches use routers to assign tasks to experts, but in continual learning,\nthey often require retraining for optimal performance. To address these\nchallenges, we introduce the Sequential Ensemble of Experts (SEE) framework.\nSEE removes the need for an additional router, allowing each expert to\nindependently decide whether a query should be handled. The framework employs\ndistributed routing, and during continual fine-tuning, SEE only requires the\ntraining of new experts for incoming tasks rather than retraining the entire\nsystem. Experiments reveal that the SEE outperforms prior approaches, including\nmulti-task learning, in continual fine-tuning. It also demonstrates remarkable\ngeneralization ability, as the expert can effectively identify\nout-of-distribution queries, which can then be directed to a more generalized\nmodel for resolution. This work highlights the promising potential of\nintegrating routing and response mechanisms within each expert, paving the way\nfor the future of distributed model ensembling.\n", "date": "2025-04-09", "categories": []}
{"id": "2504.07097", "title": "Sculpting Subspaces: Constrained Full Fine-Tuning in LLMs for Continual\n  Learning", "abstract": "  Continual learning in large language models (LLMs) is prone to catastrophic\nforgetting, where adapting to new tasks significantly degrades performance on\npreviously learned ones. Existing methods typically rely on low-rank,\nparameter-efficient updates that limit the model's expressivity and introduce\nadditional parameters per task, leading to scalability issues. To address these\nlimitations, we propose a novel continual full fine-tuning approach leveraging\nadaptive singular value decomposition (SVD). Our method dynamically identifies\ntask-specific low-rank parameter subspaces and constrains updates to be\northogonal to critical directions associated with prior tasks, thus effectively\nminimizing interference without additional parameter overhead or storing\nprevious task gradients. We evaluate our approach extensively on standard\ncontinual learning benchmarks using both encoder-decoder (T5-Large) and\ndecoder-only (LLaMA-2 7B) models, spanning diverse tasks including\nclassification, generation, and reasoning. Empirically, our method achieves\nstate-of-the-art results, up to 7% higher average accuracy than recent\nbaselines like O-LoRA, and notably maintains the model's general linguistic\ncapabilities, instruction-following accuracy, and safety throughout the\ncontinual learning process by reducing forgetting to near-negligible levels.\nOur adaptive SVD framework effectively balances model plasticity and knowledge\nretention, providing a practical, theoretically grounded, and computationally\nscalable solution for continual learning scenarios in large language models.\n", "date": "2025-04-09", "categories": []}
{"id": "2504.05571", "title": "Knowledge-Instruct: Effective Continual Pre-training from Limited Data\n  using Instructions", "abstract": "  While Large Language Models (LLMs) acquire vast knowledge during\npre-training, they often lack domain-specific, new, or niche information.\nContinual pre-training (CPT) attempts to address this gap but suffers from\ncatastrophic forgetting and inefficiencies in low-data regimes. We introduce\nKnowledge-Instruct, a novel approach to efficiently inject knowledge from\nlimited corpora through pure instruction-tuning. By generating\ninformation-dense synthetic instruction data, it effectively integrates new\nknowledge while preserving general reasoning and instruction-following\nabilities. Knowledge-Instruct demonstrates superior factual memorization,\nminimizes catastrophic forgetting, and remains scalable by leveraging synthetic\ndata from relatively small language models. Additionally, it enhances\ncontextual understanding, including complex multi-hop reasoning, facilitating\nintegration with retrieval systems. We validate its effectiveness across\ndiverse benchmarks, including Companies, a new dataset that we release to\nmeasure knowledge injection capabilities.\n", "date": "2025-04-08", "categories": []}
{"id": "2504.05621", "title": "Continual Learning of Multiple Cognitive Functions with Brain-inspired\n  Temporal Development Mechanism", "abstract": "  Cognitive functions in current artificial intelligence networks are tied to\nthe exponential increase in network scale, whereas the human brain can\ncontinuously learn hundreds of cognitive functions with remarkably low energy\nconsumption. This advantage is in part due to the brain cross-regional temporal\ndevelopment mechanisms, where the progressive formation, reorganization, and\npruning of connections from basic to advanced regions, facilitate knowledge\ntransfer and prevent network redundancy. Inspired by these, we propose the\nContinual Learning of Multiple Cognitive Functions with Brain-inspired Temporal\nDevelopment Mechanism(TD-MCL), enabling cognitive enhancement from simple to\ncomplex in Perception-Motor-Interaction(PMI) multiple cognitive task scenarios.\nThe TD-MCL model proposes the sequential evolution of long-range connections\nbetween different cognitive modules to promote positive knowledge transfer,\nwhile using feedback-guided local connection inhibition and pruning to\neffectively eliminate redundancies in previous tasks, reducing energy\nconsumption while preserving acquired knowledge. Experiments show that the\nproposed method can achieve continual learning capabilities while reducing\nnetwork scale, without introducing regularization, replay, or freezing\nstrategies, and achieving superior accuracy on new tasks compared to direct\nlearning. The proposed method shows that the brain's developmental mechanisms\noffer a valuable reference for exploring biologically plausible, low-energy\nenhancements of general cognitive abilities.\n", "date": "2025-04-08", "categories": []}
{"id": "2503.01669", "title": "An Efficient Continual Learning Framework for Multivariate Time Series\n  Prediction Tasks with Application to Vehicle State Estimation", "abstract": "  In continual time series analysis using neural networks, catastrophic\nforgetting (CF) of previously learned models when training on new data domains\nhas always been a significant challenge. This problem is especially challenging\nin vehicle estimation and control, where new information is sequentially\nintroduced to the model. Unfortunately, existing work on continual learning has\nnot sufficiently addressed the adverse effects of catastrophic forgetting in\ntime series analysis, particularly in multivariate output environments. In this\npaper, we present EM-ReSeleCT (Efficient Multivariate Representative Selection\nfor Continual Learning in Time Series Tasks), an enhanced approach designed to\nhandle continual learning in multivariate environments. Our approach\nstrategically selects representative subsets from old and historical data and\nincorporates memory-based continual learning techniques with an improved\noptimization algorithm to adapt the pre-trained model on new information while\npreserving previously acquired information. Additionally, we develop a\nsequence-to-sequence transformer model (autoregressive model) specifically\ndesigned for vehicle state estimation. Moreover, we propose an uncertainty\nquantification framework using conformal prediction to assess the sensitivity\nof the memory size and to showcase the robustness of the proposed method.\nExperimental results from tests on an electric Equinox vehicle highlight the\nsuperiority of our method in continually learning new information while\nretaining prior knowledge, outperforming state-of-the-art continual learning\nmethods. Furthermore, EM-ReSeleCT significantly reduces training time, a\ncritical advantage in continual learning applications.\n", "date": "2025-03-03", "categories": []}
{"id": "2504.03793", "title": "Outlook Towards Deployable Continual Learning for Particle Accelerators", "abstract": "  Particle Accelerators are high power complex machines. To ensure\nuninterrupted operation of these machines, thousands of pieces of equipment\nneed to be synchronized, which requires addressing many challenges including\ndesign, optimization and control, anomaly detection and machine protection.\nWith recent advancements, Machine Learning (ML) holds promise to assist in more\nadvance prognostics, optimization, and control. While ML based solutions have\nbeen developed for several applications in particle accelerators, only few have\nreached deployment and even fewer to long term usage, due to particle\naccelerator data distribution drifts caused by changes in both measurable and\nnon-measurable parameters. In this paper, we identify some of the key areas\nwithin particle accelerators where continual learning can allow maintenance of\nML model performance with distribution drifts. Particularly, we first discuss\nexisting applications of ML in particle accelerators, and their limitations due\nto distribution drift. Next, we review existing continual learning techniques\nand investigate their potential applications to address data distribution\ndrifts in accelerators. By identifying the opportunities and challenges in\napplying continual learning, this paper seeks to open up the new field and\ninspire more research efforts towards deployable continual learning for\nparticle accelerators.\n", "date": "2025-04-04", "categories": []}
{"id": "2504.04039", "title": "Memory-Statistics Tradeoff in Continual Learning with Structural\n  Regularization", "abstract": "  We study the statistical performance of a continual learning problem with two\nlinear regression tasks in a well-specified random design setting. We consider\na structural regularization algorithm that incorporates a generalized\n$\\ell_2$-regularization tailored to the Hessian of the previous task for\nmitigating catastrophic forgetting. We establish upper and lower bounds on the\njoint excess risk for this algorithm. Our analysis reveals a fundamental\ntrade-off between memory complexity and statistical efficiency, where memory\ncomplexity is measured by the number of vectors needed to define the structural\nregularization. Specifically, increasing the number of vectors in structural\nregularization leads to a worse memory complexity but an improved excess risk,\nand vice versa. Furthermore, our theory suggests that naive continual learning\nwithout regularization suffers from catastrophic forgetting, while structural\nregularization mitigates this issue. Notably, structural regularization\nachieves comparable performance to joint training with access to both tasks\nsimultaneously. These results highlight the critical role of curvature-aware\nregularization for continual learning.\n", "date": "2025-04-05", "categories": []}
{"id": "2504.04156", "title": "CoMBO: Conflict Mitigation via Branched Optimization for Class\n  Incremental Segmentation", "abstract": "  Effective Class Incremental Segmentation (CIS) requires simultaneously\nmitigating catastrophic forgetting and ensuring sufficient plasticity to\nintegrate new classes. The inherent conflict above often leads to a\nback-and-forth, which turns the objective into finding the balance between the\nperformance of previous~(old) and incremental~(new) classes. To address this\nconflict, we introduce a novel approach, Conflict Mitigation via Branched\nOptimization~(CoMBO). Within this approach, we present the Query Conflict\nReduction module, designed to explicitly refine queries for new classes through\nlightweight, class-specific adapters. This module provides an additional branch\nfor the acquisition of new classes while preserving the original queries for\ndistillation. Moreover, we develop two strategies to further mitigate the\nconflict following the branched structure, \\textit{i.e.}, the Half-Learning\nHalf-Distillation~(HDHL) over classification probabilities, and the\nImportance-Based Knowledge Distillation~(IKD) over query features. HDHL\nselectively engages in learning for classification probabilities of queries\nthat match the ground truth of new classes, while aligning unmatched ones to\nthe corresponding old probabilities, thus ensuring retention of old knowledge\nwhile absorbing new classes via learning negative samples. Meanwhile, IKD\nassesses the importance of queries based on their matching degree to old\nclasses, prioritizing the distillation of important features and allowing less\ncritical features to evolve. Extensive experiments in Class Incremental\nPanoptic and Semantic Segmentation settings have demonstrated the superior\nperformance of CoMBO. Project page: https://guangyu-ryan.github.io/CoMBO.\n", "date": "2025-04-05", "categories": []}
{"id": "2504.02489", "title": "The Self-Learning Agent with a Progressive Neural Network Integrated\n  Transformer", "abstract": "  This paper introduces a self-learning agent that integrates LLaMA 3.2 with a\nProgressive Neural Network (PNN) for continual learning in conversational AI\nand code generation. The framework dynamically collects data, fine-tunes tasks\nwith minimal samples, and leverages Meta-Learning for rapid adaptation. LoRA\noptimizes fine-tuning, while Elastic Weight Consolidation (EWC) enhances\nknowledge retention. Experimental results demonstrate improved adaptability and\nmemory stability, positioning this approach as a scalable step toward\nArtificial General Intelligence (AGI).\n", "date": "2025-04-03", "categories": []}
{"id": "2405.16625", "title": "Consistency-Guided Asynchronous Contrastive Tuning for Few-Shot\n  Class-Incremental Tuning of Foundation Models", "abstract": "  We propose Consistency-guided Asynchronous Contrastive Tuning (CoACT), a\nnovel method for continuously tuning foundation models to learn new classes in\nfew-shot settings. CoACT consists of three key components:(i) asynchronous\ncontrastive tuning, which learns new classes by including LoRA modules in the\npre-trained encoder while enforcing consistency between two asynchronous\nencoders; (ii) controlled fine-tuning, which facilitates effective tuning of a\nsubset of the foundation model; and (iii) consistency-guided incremental\ntuning, which enforces additional regularization during later sessions to\nreduce forgetting of the learned classes. We evaluate our proposed solution on\nFew-Shot Class-Incremental Learning (FSCIL) as well as a new and more\nchallenging setup called Few-Shot Class-Incremental Tuning (FSCIT), which\nfacilitates the continual tuning of vision foundation models to learn new\nclasses with only a few samples per class. Unlike traditional FSCIL, FSCIT does\nnot require a large in-distribution base session for initial fully supervised\ntraining prior to the incremental few-shot sessions. We conduct extensive\nevaluations across 16 diverse datasets, demonstrating the effectiveness of\nCoACT in both FSCIL and FSCIT setups. CoACT outperforms existing methods by up\nto 5.02% in FSCIL and up to 12.51% in FSCIT for individual datasets, with an\naverage improvement of 2.47%. Furthermore, CoACT exhibits reduced forgetting\nand enhanced robustness in low-shot experiments. Detailed ablation and\nsensitivity studies highlight the contribution of each component of CoACT. We\nmake our code publicly available at https://github.com/ShuvenduRoy/CoACT-FSCIL.\n", "date": "2024-05-26", "categories": []}
{"id": "2504.01219", "title": "Gradient-free Continual Learning", "abstract": "  Continual learning (CL) presents a fundamental challenge in training neural\nnetworks on sequential tasks without experiencing catastrophic forgetting.\nTraditionally, the dominant approach in CL has been gradient-based\noptimization, where updates to the network parameters are performed using\nstochastic gradient descent (SGD) or its variants. However, a major limitation\narises when previous data is no longer accessible, as is often assumed in CL\nsettings. In such cases, there is no gradient information available for past\ndata, leading to uncontrolled parameter changes and consequently severe\nforgetting of previously learned tasks. By shifting focus from data\navailability to gradient availability, this work opens up new avenues for\naddressing forgetting in CL. We explore the hypothesis that gradient-free\noptimization methods can provide a robust alternative to conventional\ngradient-based continual learning approaches. We discuss the theoretical\nunderpinnings of such method, analyze their potential advantages and\nlimitations, and present empirical evidence supporting their effectiveness. By\nreconsidering the fundamental cause of forgetting, this work aims to contribute\na fresh perspective to the field of continual learning and inspire novel\nresearch directions.\n", "date": "2025-04-01", "categories": []}
{"id": "2504.01241", "title": "Catastrophic Forgetting in LLMs: A Comparative Analysis Across Language\n  Tasks", "abstract": "  Large Language Models (LLMs) have significantly advanced Natural Language\nProcessing (NLP), particularly in Natural Language Understanding (NLU) tasks.\nAs we progress toward an agentic world where LLM-based agents autonomously\nhandle specialized tasks, it becomes crucial for these models to adapt to new\ntasks without forgetting previously learned information - a challenge known as\ncatastrophic forgetting. This study evaluates the continual fine-tuning of\nvarious open-source LLMs with different parameter sizes (specifically models\nunder 10 billion parameters) on key NLU tasks from the GLUE benchmark,\nincluding SST-2, MRPC, CoLA, and MNLI. By employing prompt engineering and\ntask-specific adjustments, we assess and compare the models' abilities to\nretain prior knowledge while learning new tasks. Our results indicate that\nmodels such as Phi-3.5-mini exhibit minimal forgetting while maintaining strong\nlearning capabilities, making them well-suited for continual learning\nenvironments. Additionally, models like Orca-2-7b and Qwen2.5-7B demonstrate\nimpressive learning abilities and overall performance after fine-tuning. This\nwork contributes to understanding catastrophic forgetting in LLMs and\nhighlights prompting engineering to optimize model performance for continual\nlearning scenarios.\n", "date": "2025-04-01", "categories": []}
{"id": "2405.13900", "title": "Rehearsal-free Federated Domain-incremental Learning", "abstract": "  We introduce a rehearsal-free federated domain incremental learning\nframework, RefFiL, based on a global prompt-sharing paradigm to alleviate\ncatastrophic forgetting challenges in federated domain-incremental learning,\nwhere unseen domains are continually learned. Typical methods for mitigating\nforgetting, such as the use of additional datasets and the retention of private\ndata from earlier tasks, are not viable in federated learning (FL) due to\ndevices' limited resources. Our method, RefFiL, addresses this by learning\ndomain-invariant knowledge and incorporating various domain-specific prompts\nfrom the domains represented by different FL participants. A key feature of\nRefFiL is the generation of local fine-grained prompts by our domain adaptive\nprompt generator, which effectively learns from local domain knowledge while\nmaintaining distinctive boundaries on a global scale. We also introduce a\ndomain-specific prompt contrastive learning loss that differentiates between\nlocally generated prompts and those from other domains, enhancing RefFiL's\nprecision and effectiveness. Compared to existing methods, RefFiL significantly\nalleviates catastrophic forgetting without requiring extra memory space, making\nit ideal for privacy-sensitive and resource-constrained devices.\n", "date": "2024-05-22", "categories": []}
{"id": "2504.00561", "title": "Continual Cross-Modal Generalization", "abstract": "  Cross-modal generalization aims to learn a shared discrete representation\nspace from multimodal pairs, enabling knowledge transfer across unannotated\nmodalities. However, achieving a unified representation for all modality pairs\nrequires extensive paired data, which is often impractical. Inspired by the\navailability of abundant bimodal data (e.g., in ImageBind), we explore a\ncontinual learning approach that incrementally maps new modalities into a\nshared discrete codebook via a mediator modality. We propose the Continual\nMixture of Experts Adapter (CMoE-Adapter) to project diverse modalities into a\nunified space while preserving prior knowledge. To align semantics across\nstages, we introduce a Pseudo-Modality Replay (PMR) mechanism with a\ndynamically expanding codebook, enabling the model to adaptively incorporate\nnew modalities using learned ones as guidance. Extensive experiments on\nimage-text, audio-text, video-text, and speech-text show that our method\nachieves strong performance on various cross-modal generalization tasks. Code\nis provided in the supplementary material.\n", "date": "2025-04-01", "categories": []}
{"id": "2405.11067", "title": "Bayesian Learning-driven Prototypical Contrastive Loss for\n  Class-Incremental Learning", "abstract": "  The primary objective of methods in continual learning is to learn tasks in a\nsequential manner over time (sometimes from a stream of data), while mitigating\nthe detrimental phenomenon of catastrophic forgetting. This paper proposes a\nmethod to learn an effective representation between previous and newly\nencountered class prototypes. We propose a prototypical network with a Bayesian\nlearning-driven contrastive loss (BLCL), tailored specifically for\nclass-incremental learning scenarios. We introduce a contrastive loss that\nincorporates novel classes into the latent representation by reducing\nintra-class and increasing inter-class distance. Our approach dynamically\nadapts the balance between the cross-entropy and contrastive loss functions\nwith a Bayesian learning technique. Experimental results conducted on the\nCIFAR-10, CIFAR-100, and ImageNet100 datasets for image classification and\nimages of a GNSS-based dataset for interference classification validate the\nefficacy of our method, showcasing its superiority over existing\nstate-of-the-art approaches. Git:\nhttps://gitlab.cc-asp.fraunhofer.de/darcy_gnss/gnss_class_incremental_learning\n", "date": "2024-05-17", "categories": []}
{"id": "2503.18985", "title": "LoRA Subtraction for Drift-Resistant Space in Exemplar-Free Continual\n  Learning", "abstract": "  In continual learning (CL), catastrophic forgetting often arises due to\nfeature drift. This challenge is particularly prominent in the exemplar-free\ncontinual learning (EFCL) setting, where samples from previous tasks cannot be\nretained, making it difficult to preserve prior knowledge. To address this\nissue, some EFCL methods aim to identify feature spaces that minimize the\nimpact on previous tasks while accommodating new ones. However, they rely on\nstatic features or outdated statistics stored from old tasks, which prevents\nthem from capturing the dynamic evolution of the feature space in CL, leading\nto performance degradation over time. In this paper, we introduce the\nDrift-Resistant Space (DRS), which effectively handles feature drifts without\nrequiring explicit feature modeling or the storage of previous tasks. A novel\nparameter-efficient fine-tuning approach called Low-Rank Adaptation Subtraction\n(LoRA-) is proposed to develop the DRS. This method subtracts the LoRA weights\nof old tasks from the initial pre-trained weight before processing new task\ndata to establish the DRS for model training. Therefore, LoRA- enhances\nstability, improves efficiency, and simplifies implementation. Furthermore,\nstabilizing feature drifts allows for better plasticity by learning with a\ntriplet loss. Our method consistently achieves state-of-the-art results,\nespecially for long task sequences, across multiple datasets.\n", "date": "2025-03-23", "categories": []}
{"id": "2503.22729", "title": "Ancestral Mamba: Enhancing Selective Discriminant Space Model with\n  Online Visual Prototype Learning for Efficient and Robust Discriminant\n  Approach", "abstract": "  In the realm of computer graphics, the ability to learn continuously from\nnon-stationary data streams while adapting to new visual patterns and\nmitigating catastrophic forgetting is of paramount importance. Existing\napproaches often struggle to capture and represent the essential\ncharacteristics of evolving visual concepts, hindering their applicability to\ndynamic graphics tasks. In this paper, we propose Ancestral Mamba, a novel\napproach that integrates online prototype learning into a selective\ndiscriminant space model for efficient and robust online continual learning.\nThe key components of our approach include Ancestral Prototype Adaptation\n(APA), which continuously refines and builds upon learned visual prototypes,\nand Mamba Feedback (MF), which provides targeted feedback to adapt to\nchallenging visual patterns. APA enables the model to continuously adapt its\nprototypes, building upon ancestral knowledge to tackle new challenges, while\nMF acts as a targeted feedback mechanism, focusing on challenging classes and\nrefining their representations. Extensive experiments on graphics-oriented\ndatasets, such as CIFAR-10 and CIFAR-100, demonstrate the superior performance\nof Ancestral Mamba compared to state-of-the-art baselines, achieving\nsignificant improvements in accuracy and forgetting mitigation.\n", "date": "2025-03-26", "categories": []}
{"id": "2503.23046", "title": "VLM-C4L: Continual Core Dataset Learning with Corner Case Optimization\n  via Vision-Language Models for Autonomous Driving", "abstract": "  With the widespread adoption and deployment of autonomous driving, handling\ncomplex environments has become an unavoidable challenge. Due to the scarcity\nand diversity of extreme scenario datasets, current autonomous driving models\nstruggle to effectively manage corner cases. This limitation poses a\nsignificant safety risk, according to the National Highway Traffic Safety\nAdministration (NHTSA), autonomous vehicle systems have been involved in\nhundreds of reported crashes annually in the United States, occurred in corner\ncases like sun glare and fog, which caused a few fatal accident. Furthermore,\nin order to consistently maintain a robust and reliable autonomous driving\nsystem, it is essential for models not only to perform well on routine\nscenarios but also to adapt to newly emerging scenarios, especially those\ncorner cases that deviate from the norm. This requires a learning mechanism\nthat incrementally integrates new knowledge without degrading previously\nacquired capabilities. However, to the best of our knowledge, no existing\ncontinual learning methods have been proposed to ensure consistent and scalable\ncorner case learning in autonomous driving. To address these limitations, we\npropose VLM-C4L, a continual learning framework that introduces Vision-Language\nModels (VLMs) to dynamically optimize and enhance corner case datasets, and\nVLM-C4L combines VLM-guided high-quality data extraction with a core data\nreplay strategy, enabling the model to incrementally learn from diverse corner\ncases while preserving performance on previously routine scenarios, thus\nensuring long-term stability and adaptability in real-world autonomous driving.\nWe evaluate VLM-C4L on large-scale real-world autonomous driving datasets,\nincluding Waymo and the corner case dataset CODA.\n", "date": "2025-03-29", "categories": []}
{"id": "2503.23283", "title": "Language Guided Concept Bottleneck Models for Interpretable Continual\n  Learning", "abstract": "  Continual learning (CL) aims to enable learning systems to acquire new\nknowledge constantly without forgetting previously learned information. CL\nfaces the challenge of mitigating catastrophic forgetting while maintaining\ninterpretability across tasks. Most existing CL methods focus primarily on\npreserving learned knowledge to improve model performance. However, as new\ninformation is introduced, the interpretability of the learning process becomes\ncrucial for understanding the evolving decision-making process, yet it is\nrarely explored. In this paper, we introduce a novel framework that integrates\nlanguage-guided Concept Bottleneck Models (CBMs) to address both challenges.\nOur approach leverages the Concept Bottleneck Layer, aligning semantic\nconsistency with CLIP models to learn human-understandable concepts that can\ngeneralize across tasks. By focusing on interpretable concepts, our method not\nonly enhances the models ability to retain knowledge over time but also\nprovides transparent decision-making insights. We demonstrate the effectiveness\nof our approach by achieving superior performance on several datasets,\noutperforming state-of-the-art methods with an improvement of up to 3.06% in\nfinal average accuracy on ImageNet-subset. Additionally, we offer concept\nvisualizations for model predictions, further advancing the understanding of\ninterpretable continual learning.\n", "date": "2025-03-30", "categories": []}
{"id": "2503.23390", "title": "Pareto Continual Learning: Preference-Conditioned Learning and Adaption\n  for Dynamic Stability-Plasticity Trade-off", "abstract": "  Continual learning aims to learn multiple tasks sequentially. A key challenge\nin continual learning is balancing between two objectives: retaining knowledge\nfrom old tasks (stability) and adapting to new tasks (plasticity). Experience\nreplay methods, which store and replay past data alongside new data, have\nbecome a widely adopted approach to mitigate catastrophic forgetting. However,\nthese methods neglect the dynamic nature of the stability-plasticity trade-off\nand aim to find a fixed and unchanging balance, resulting in suboptimal\nadaptation during training and inference. In this paper, we propose Pareto\nContinual Learning (ParetoCL), a novel framework that reformulates the\nstability-plasticity trade-off in continual learning as a multi-objective\noptimization (MOO) problem. ParetoCL introduces a preference-conditioned model\nto efficiently learn a set of Pareto optimal solutions representing different\ntrade-offs and enables dynamic adaptation during inference. From a\ngeneralization perspective, ParetoCL can be seen as an objective augmentation\napproach that learns from different objective combinations of stability and\nplasticity. Extensive experiments across multiple datasets and settings\ndemonstrate that ParetoCL outperforms state-of-the-art methods and adapts to\ndiverse continual learning scenarios.\n", "date": "2025-03-30", "categories": []}
{"id": "2411.11396", "title": "Stacking Brick by Brick: Aligned Feature Isolation for Incremental Face\n  Forgery Detection", "abstract": "  The rapid advancement of face forgery techniques has introduced a growing\nvariety of forgeries. Incremental Face Forgery Detection (IFFD), involving\ngradually adding new forgery data to fine-tune the previously trained model,\nhas been introduced as a promising strategy to deal with evolving forgery\nmethods. However, a naively trained IFFD model is prone to catastrophic\nforgetting when new forgeries are integrated, as treating all forgeries as a\nsingle ''Fake\" class in the Real/Fake classification can cause different\nforgery types overriding one another, thereby resulting in the forgetting of\nunique characteristics from earlier tasks and limiting the model's\neffectiveness in learning forgery specificity and generality. In this paper, we\npropose to stack the latent feature distributions of previous and new tasks\nbrick by brick, $\\textit{i.e.}$, achieving $\\textbf{aligned feature\nisolation}$. In this manner, we aim to preserve learned forgery information and\naccumulate new knowledge by minimizing distribution overriding, thereby\nmitigating catastrophic forgetting. To achieve this, we first introduce Sparse\nUniform Replay (SUR) to obtain the representative subsets that could be treated\nas the uniformly sparse versions of the previous global distributions. We then\npropose a Latent-space Incremental Detector (LID) that leverages SUR data to\nisolate and align distributions. For evaluation, we construct a more advanced\nand comprehensive benchmark tailored for IFFD. The leading experimental results\nvalidate the superiority of our method.\n", "date": "2024-11-18", "categories": []}
{"id": "2503.19503", "title": "Adaptive Weighted Parameter Fusion with CLIP for Class-Incremental\n  Learning", "abstract": "  Class-incremental Learning (CIL) enables the model to incrementally absorb\nknowledge from new classes and build a generic classifier across all previously\nencountered classes. When the model optimizes with new classes, the knowledge\nof previous classes is inevitably erased, leading to catastrophic forgetting.\nAddressing this challenge requires making a trade-off between retaining old\nknowledge and accommodating new information. However, this balancing process\noften requires sacrificing some information, which can lead to a partial loss\nin the model's ability to discriminate between classes. To tackle this issue,\nwe design the adaptive weighted parameter fusion with Contrastive\nLanguage-Image Pre-training (CLIP), which not only takes into account the\nvariability of the data distribution of different tasks, but also retains all\nthe effective information of the parameter matrix to the greatest extent. In\naddition, we introduce a balance factor that can balance the data distribution\nalignment and distinguishability of adjacent tasks. Experimental results on\nseveral traditional benchmarks validate the superiority of the proposed method.\n", "date": "2025-03-25", "categories": []}
{"id": "2503.22068", "title": "A Proposal for Networks Capable of Continual Learning", "abstract": "  We analyze the ability of computational units to retain past responses after\nparameter updates, a key property for system-wide continual learning. Neural\nnetworks trained with gradient descent lack this capability, prompting us to\npropose Modelleyen, an alternative approach with inherent response\npreservation. We demonstrate through experiments on modeling the dynamics of a\nsimple environment and on MNIST that, despite increased computational\ncomplexity and some representational limitations at its current stage,\nModelleyen achieves continual learning without relying on sample replay or\npredefined task boundaries.\n", "date": "2025-03-28", "categories": []}
{"id": "2503.22136", "title": "Beyond Background Shift: Rethinking Instance Replay in Continual\n  Semantic Segmentation", "abstract": "  In this work, we focus on continual semantic segmentation (CSS), where\nsegmentation networks are required to continuously learn new classes without\nerasing knowledge of previously learned ones. Although storing images of old\nclasses and directly incorporating them into the training of new models has\nproven effective in mitigating catastrophic forgetting in classification tasks,\nthis strategy presents notable limitations in CSS. Specifically, the stored and\nnew images with partial category annotations leads to confusion between\nunannotated categories and the background, complicating model fitting. To\ntackle this issue, this paper proposes a novel Enhanced Instance Replay (EIR)\nmethod, which not only preserves knowledge of old classes while simultaneously\neliminating background confusion by instance storage of old classes, but also\nmitigates background shifts in the new images by integrating stored instances\nwith new images. By effectively resolving background shifts in both stored and\nnew images, EIR alleviates catastrophic forgetting in the CSS task, thereby\nenhancing the model's capacity for CSS. Experimental results validate the\nefficacy of our approach, which significantly outperforms state-of-the-art CSS\nmethods.\n", "date": "2025-03-28", "categories": []}
{"id": "2503.22163", "title": "T-CIL: Temperature Scaling using Adversarial Perturbation for\n  Calibration in Class-Incremental Learning", "abstract": "  We study model confidence calibration in class-incremental learning, where\nmodels learn from sequential tasks with different class sets. While existing\nworks primarily focus on accuracy, maintaining calibrated confidence has been\nlargely overlooked. Unfortunately, most post-hoc calibration techniques are not\ndesigned to work with the limited memories of old-task data typical in\nclass-incremental learning, as retaining a sufficient validation set would be\nimpractical. Thus, we propose T-CIL, a novel temperature scaling approach for\nclass-incremental learning without a validation set for old tasks, that\nleverages adversarially perturbed exemplars from memory. Directly using\nexemplars is inadequate for temperature optimization, since they are already\nused for training. The key idea of T-CIL is to perturb exemplars more strongly\nfor old tasks than for the new task by adjusting the perturbation direction\nbased on feature distance, with the single magnitude determined using the\nnew-task validation set. This strategy makes the perturbation magnitude\ncomputed from the new task also applicable to old tasks, leveraging the\ntendency that the accuracy of old tasks is lower than that of the new task. We\nempirically show that T-CIL significantly outperforms various baselines in\nterms of calibration on real datasets and can be integrated with existing\nclass-incremental learning techniques with minimal impact on accuracy.\n", "date": "2025-03-28", "categories": []}
{"id": "2503.22175", "title": "Efficient Continual Learning through Frequency Decomposition and\n  Integration", "abstract": "  Continual learning (CL) aims to learn new tasks while retaining past\nknowledge, addressing the challenge of forgetting during task adaptation.\nRehearsal-based methods, which replay previous samples, effectively mitigate\nforgetting. However, research on enhancing the efficiency of these methods,\nespecially in resource-constrained environments, remains limited, hindering\ntheir application in real-world systems with dynamic data streams. The human\nperceptual system processes visual scenes through complementary frequency\nchannels: low-frequency signals capture holistic cues, while high-frequency\ncomponents convey structural details vital for fine-grained discrimination.\nInspired by this, we propose the Frequency Decomposition and Integration\nNetwork (FDINet), a novel framework that decomposes and integrates information\nacross frequencies. FDINet designs two lightweight networks to independently\nprocess low- and high-frequency components of images. When integrated with\nrehearsal-based methods, this frequency-aware design effectively enhances\ncross-task generalization through low-frequency information, preserves\nclass-specific details using high-frequency information, and facilitates\nefficient training due to its lightweight architecture. Experiments demonstrate\nthat FDINet reduces backbone parameters by 78%, improves accuracy by up to\n7.49% over state-of-the-art (SOTA) methods, and decreases peak memory usage by\nup to 80%. Additionally, on edge devices, FDINet accelerates training by up to\n5$\\times$.\n", "date": "2025-03-28", "categories": []}
{"id": "2403.18886", "title": "Self-Expansion of Pre-trained Models with Mixture of Adapters for\n  Continual Learning", "abstract": "  Continual learning (CL) aims to continually accumulate knowledge from a\nnon-stationary data stream without catastrophic forgetting of learned\nknowledge, requiring a balance between stability and adaptability. Relying on\nthe generalizable representation in pre-trained models (PTMs), PTM-based CL\nmethods perform effective continual adaptation on downstream tasks by adding\nlearnable adapters or prompts upon the frozen PTMs. However, many existing\nPTM-based CL methods use restricted adaptation on a fixed set of these modules\nto avoid forgetting, suffering from limited CL ability. Periodically adding\ntask-specific modules results in linear model growth rate and impaired\nknowledge reuse. We propose Self-Expansion of pre-trained models with\nModularized Adaptation (SEMA), a novel approach to enhance the control of\nstability-plasticity balance in PTM-based CL. SEMA automatically decides to\nreuse or add adapter modules on demand in CL, depending on whether significant\ndistribution shift that cannot be handled is detected at different\nrepresentation levels. We design modular adapter consisting of a functional\nadapter and a representation descriptor. The representation descriptors are\ntrained as a distribution shift indicator and used to trigger self-expansion\nsignals. For better composing the adapters, an expandable weighting router is\nlearned jointly for mixture of adapter outputs. SEMA enables better knowledge\nreuse and sub-linear expansion rate. Extensive experiments demonstrate the\neffectiveness of the proposed self-expansion method, achieving state-of-the-art\nperformance compared to PTM-based CL methods without memory rehearsal. Code is\navailable at https://github.com/huiyiwang01/SEMA-CL.\n", "date": "2024-03-27", "categories": []}
{"id": "2411.01739", "title": "Not Just Object, But State: Compositional Incremental Learning without\n  Forgetting", "abstract": "  Most incremental learners excessively prioritize coarse classes of objects\nwhile neglecting various kinds of states (e.g. color and material) attached to\nthe objects. As a result, they are limited in the ability to reason\nfine-grained compositionality of state-object pairs. To remedy this limitation,\nwe propose a novel task called Compositional Incremental Learning\n(composition-IL), enabling the model to recognize state-object compositions as\na whole in an incremental learning fashion. Since the lack of suitable\nbenchmarks, we re-organize two existing datasets and make them tailored for\ncomposition-IL. Then, we propose a prompt-based Composition Incremental Learner\n(CompILer), to overcome the ambiguous composition boundary problem which\nchallenges composition-IL largely. Specifically, we exploit multi-pool prompt\nlearning, which is regularized by inter-pool prompt discrepancy and intra-pool\nprompt diversity. Besides, we devise object-injected state prompting by using\nobject prompts to guide the selection of state prompts. Furthermore, we fuse\nthe selected prompts by a generalized-mean strategy, to eliminate irrelevant\ninformation learned in the prompts. Extensive experiments on two datasets\nexhibit state-of-the-art performance achieved by CompILer.\n", "date": "2024-11-04", "categories": []}
{"id": "2503.20808", "title": "Dynamic Allocation Hypernetwork with Adaptive Model Recalibration for\n  Federated Continual Learning", "abstract": "  Federated continual learning (FCL) offers an emerging pattern to facilitate\nthe applicability of federated learning (FL) in real-world scenarios, where\ntasks evolve dynamically and asynchronously across clients, especially in\nmedical scenario. Existing server-side FCL methods in nature domain construct a\ncontinually learnable server model by client aggregation on all-involved tasks.\nHowever, they are challenged by: (1) Catastrophic forgetting for previously\nlearned tasks, leading to error accumulation in server model, making it\ndifficult to sustain comprehensive knowledge across all tasks. (2) Biased\noptimization due to asynchronous tasks handled across different clients,\nleading to the collision of optimization targets of different clients at the\nsame time steps. In this work, we take the first step to propose a novel\nserver-side FCL pattern in medical domain, Dynamic Allocation Hypernetwork with\nadaptive model recalibration (FedDAH). It is to facilitate collaborative\nlearning under the distinct and dynamic task streams across clients. To\nalleviate the catastrophic forgetting, we propose a dynamic allocation\nhypernetwork (DAHyper) where a continually updated hypernetwork is designed to\nmanage the mapping between task identities and their associated model\nparameters, enabling the dynamic allocation of the model across clients. For\nthe biased optimization, we introduce a novel adaptive model recalibration\n(AMR) to incorporate the candidate changes of historical models into current\nserver updates, and assign weights to identical tasks across different time\nsteps based on the similarity for continual optimization. Extensive experiments\non the AMOS dataset demonstrate the superiority of our FedDAH to other FCL\nmethods on sites with different task streams. The code is\navailable:https://github.com/jinlab-imvr/FedDAH.\n", "date": "2025-03-25", "categories": []}
{"id": "2503.21076", "title": "KAC: Kolmogorov-Arnold Classifier for Continual Learning", "abstract": "  Continual learning requires models to train continuously across consecutive\ntasks without forgetting. Most existing methods utilize linear classifiers,\nwhich struggle to maintain a stable classification space while learning new\ntasks. Inspired by the success of Kolmogorov-Arnold Networks (KAN) in\npreserving learning stability during simple continual regression tasks, we set\nout to explore their potential in more complex continual learning scenarios. In\nthis paper, we introduce the Kolmogorov-Arnold Classifier (KAC), a novel\nclassifier developed for continual learning based on the KAN structure. We\ndelve into the impact of KAN's spline functions and introduce Radial Basis\nFunctions (RBF) for improved compatibility with continual learning. We replace\nlinear classifiers with KAC in several recent approaches and conduct\nexperiments across various continual learning benchmarks, all of which\ndemonstrate performance improvements, highlighting the effectiveness and\nrobustness of KAC in continual learning. The code is available at\nhttps://github.com/Ethanhuhuhu/KAC.\n", "date": "2025-03-27", "categories": []}
{"id": "2503.21436", "title": "Stochastic Engrams for Efficient Continual Learning with Binarized\n  Neural Networks", "abstract": "  The ability to learn continuously in artificial neural networks (ANNs) is\noften limited by catastrophic forgetting, a phenomenon in which new knowledge\nbecomes dominant. By taking mechanisms of memory encoding in neuroscience (aka.\nengrams) as inspiration, we propose a novel approach that integrates\nstochastically-activated engrams as a gating mechanism for metaplastic\nbinarized neural networks (mBNNs). This method leverages the computational\nefficiency of mBNNs combined with the robustness of probabilistic memory traces\nto mitigate forgetting and maintain the model's reliability. Previously\nvalidated metaplastic optimization techniques have been incorporated to enhance\nsynaptic stability further. Compared to baseline binarized models and benchmark\nfully connected continual learning approaches, our method is the only strategy\ncapable of reaching average accuracies over 20% in class-incremental scenarios\nand achieving comparable domain-incremental results to full precision\nstate-of-the-art methods. Furthermore, we achieve a significant reduction in\npeak GPU and RAM usage, under 5% and 20%, respectively. Our findings\ndemonstrate (A) an improved stability vs. plasticity trade-off, (B) a reduced\nmemory intensiveness, and (C) an enhanced performance in binarized\narchitectures. By uniting principles of neuroscience and efficient computing,\nwe offer new insights into the design of scalable and robust deep learning\nsystems.\n", "date": "2025-03-27", "categories": []}
{"id": "2503.19939", "title": "Continual Learning With Quasi-Newton Methods", "abstract": "  Catastrophic forgetting remains a major challenge when neural networks learn\ntasks sequentially. Elastic Weight Consolidation (EWC) attempts to address this\nproblem by introducing a Bayesian-inspired regularization loss to preserve\nknowledge of previously learned tasks. However, EWC relies on a Laplace\napproximation where the Hessian is simplified to the diagonal of the Fisher\ninformation matrix, assuming uncorrelated model parameters. This overly\nsimplistic assumption often leads to poor Hessian estimates, limiting its\neffectiveness. To overcome this limitation, we introduce Continual Learning\nwith Sampled Quasi-Newton (CSQN), which leverages Quasi-Newton methods to\ncompute more accurate Hessian approximations. CSQN captures parameter\ninteractions beyond the diagonal without requiring architecture-specific\nmodifications, making it applicable across diverse tasks and architectures.\nExperimental results across four benchmarks demonstrate that CSQN consistently\noutperforms EWC and other state-of-the-art baselines, including rehearsal-based\nmethods. CSQN reduces EWC's forgetting by 50 percent and improves its\nperformance by 8 percent on average. Notably, CSQN achieves superior results on\nthree out of four benchmarks, including the most challenging scenarios,\nhighlighting its potential as a robust solution for continual learning.\n", "date": "2025-03-25", "categories": []}
{"id": "2503.20018", "title": "Experience Replay Addresses Loss of Plasticity in Continual Learning", "abstract": "  Loss of plasticity is one of the main challenges in continual learning with\ndeep neural networks, where neural networks trained via backpropagation\ngradually lose their ability to adapt to new tasks and perform significantly\nworse than their freshly initialized counterparts. The main contribution of\nthis paper is to propose a new hypothesis that experience replay addresses the\nloss of plasticity in continual learning. Here, experience replay is a form of\nmemory. We provide supporting evidence for this hypothesis. In particular, we\ndemonstrate in multiple different tasks, including regression, classification,\nand policy evaluation, that by simply adding an experience replay and\nprocessing the data in the experience replay with Transformers, the loss of\nplasticity disappears. Notably, we do not alter any standard components of deep\nlearning. For example, we do not change backpropagation. We do not modify the\nactivation functions. And we do not use any regularization. We conjecture that\nexperience replay and Transformers can address the loss of plasticity because\nof the in-context learning phenomenon.\n", "date": "2025-03-25", "categories": []}
{"id": "2503.20725", "title": "Continual learning via probabilistic exchangeable sequence modelling", "abstract": "  Continual learning (CL) refers to the ability to continuously learn and\naccumulate new knowledge while retaining useful information from past\nexperiences. Although numerous CL methods have been proposed in recent years,\nit is not straightforward to deploy them directly to real-world decision-making\nproblems due to their computational cost and lack of uncertainty\nquantification. To address these issues, we propose CL-BRUNO, a probabilistic,\nNeural Process-based CL model that performs scalable and tractable Bayesian\nupdate and prediction. Our proposed approach uses deep-generative models to\ncreate a unified probabilistic framework capable of handling different types of\nCL problems such as task- and class-incremental learning, allowing users to\nintegrate information across different CL scenarios using a single model. Our\napproach is able to prevent catastrophic forgetting through distributional and\nfunctional regularisation without the need of retaining any previously seen\nsamples, making it appealing to applications where data privacy or storage\ncapacity is of concern. Experiments show that CL-BRUNO outperforms existing\nmethods on both natural image and biomedical data sets, confirming its\neffectiveness in real-world applications.\n", "date": "2025-03-26", "categories": []}
{"id": "2503.06056", "title": "Pathological Prior-Guided Multiple Instance Learning For Mitigating\n  Catastrophic Forgetting in Breast Cancer Whole Slide Image Classification", "abstract": "  In histopathology, intelligent diagnosis of Whole Slide Images (WSIs) is\nessential for automating and objectifying diagnoses, reducing the workload of\npathologists. However, diagnostic models often face the challenge of forgetting\npreviously learned data during incremental training on datasets from different\nsources. To address this issue, we propose a new framework PaGMIL to mitigate\ncatastrophic forgetting in breast cancer WSI classification. Our framework\nintroduces two key components into the common MIL model architecture. First, it\nleverages microscopic pathological prior to select more accurate and diverse\nrepresentative patches for MIL. Secondly, it trains separate classification\nheads for each task and uses macroscopic pathological prior knowledge, treating\nthe thumbnail as a prompt guide (PG) to select the appropriate classification\nhead. We evaluate the continual learning performance of PaGMIL across several\npublic breast cancer datasets. PaGMIL achieves a better balance between the\nperformance of the current task and the retention of previous tasks,\noutperforming other continual learning methods. Our code will be open-sourced\nupon acceptance.\n", "date": "2025-03-08", "categories": []}
{"id": "2503.19212", "title": "Continual Reinforcement Learning for HVAC Systems Control: Integrating\n  Hypernetworks and Transfer Learning", "abstract": "  Buildings with Heating, Ventilation, and Air Conditioning (HVAC) systems play\na crucial role in ensuring indoor comfort and efficiency. While traditionally\ngoverned by physics-based models, the emergence of big data has enabled\ndata-driven methods like Deep Reinforcement Learning (DRL). However,\nReinforcement Learning (RL)-based techniques often suffer from sample\ninefficiency and limited generalization, especially across varying HVAC\nsystems. We introduce a model-based reinforcement learning framework that uses\na Hypernetwork to continuously learn environment dynamics across tasks with\ndifferent action spaces. This enables efficient synthetic rollout generation\nand improved sample usage. Our approach demonstrates strong backward transfer\nin a continual learning setting after training on a second task, minimal\nfine-tuning on the first task allows rapid convergence within just 5 episodes\nand thus outperforming Model Free Reinforcement Learning (MFRL) and effectively\nmitigating catastrophic forgetting. These findings have significant\nimplications for reducing energy consumption and operational costs in building\nmanagement, thus supporting global sustainability goals.\n  Keywords: Deep Reinforcement Learning, HVAC Systems Control, Hypernetworks,\nTransfer and Continual Learning, Catastrophic Forgetting\n", "date": "2025-03-24", "categories": []}
{"id": "2211.11174", "title": "Unveiling the Tapestry: the Interplay of Generalization and Forgetting\n  in Continual Learning", "abstract": "  In AI, generalization refers to a model's ability to perform well on\nout-of-distribution data related to the given task, beyond the data it was\ntrained on. For an AI agent to excel, it must also possess the continual\nlearning capability, whereby an agent incrementally learns to perform a\nsequence of tasks without forgetting the previously acquired knowledge to solve\nthe old tasks. Intuitively, generalization within a task allows the model to\nlearn underlying features that can readily be applied to novel tasks,\nfacilitating quicker learning and enhanced performance in subsequent tasks\nwithin a continual learning framework. Conversely, continual learning methods\noften include mechanisms to mitigate catastrophic forgetting, ensuring that\nknowledge from earlier tasks is retained. This preservation of knowledge over\ntasks plays a role in enhancing generalization for the ongoing task at hand.\nDespite the intuitive appeal of the interplay of both abilities, existing\nliterature on continual learning and generalization has proceeded separately.\nIn the preliminary effort to promote studies that bridge both fields, we first\npresent empirical evidence showing that each of these fields has a mutually\npositive effect on the other. Next, building upon this finding, we introduce a\nsimple and effective technique known as Shape-Texture Consistency\nRegularization (STCR), which caters to continual learning. STCR learns both\nshape and texture representations for each task, consequently enhancing\ngeneralization and thereby mitigating forgetting. Remarkably, extensive\nexperiments validate that our STCR, can be seamlessly integrated with existing\ncontinual learning methods, including replay-free approaches. Its performance\nsurpasses these continual learning methods in isolation or when combined with\nestablished generalization techniques by a large margin.\n", "date": "2022-11-21", "categories": []}
{"id": "2412.18219", "title": "Adapter Merging with Centroid Prototype Mapping for Scalable\n  Class-Incremental Learning", "abstract": "  We propose Adapter Merging with Centroid Prototype Mapping (ACMap), an\nexemplar-free framework for class-incremental learning (CIL) that addresses\nboth catastrophic forgetting and scalability. While existing methods involve a\ntrade-off between inference time and accuracy, ACMap consolidates task-specific\nadapters into a single adapter, thus achieving constant inference time across\ntasks without sacrificing accuracy. The framework employs adapter merging to\nbuild a shared subspace that aligns task representations and mitigates\nforgetting, while centroid prototype mapping maintains high accuracy by\nconsistently adapting representations within the shared subspace. To further\nimprove scalability, an early stopping strategy limits adapter merging as tasks\nincrease. Extensive experiments on five benchmark datasets demonstrate that\nACMap matches state-of-the-art accuracy while maintaining inference time\ncomparable to the fastest existing methods. The code is available at\nhttps://github.com/tf63/ACMap.\n", "date": "2024-12-24", "categories": []}
{"id": "2503.17677", "title": "Reducing Class-wise Confusion for Incremental Learning with Disentangled\n  Manifolds", "abstract": "  Class incremental learning (CIL) aims to enable models to continuously learn\nnew classes without catastrophically forgetting old ones. A promising direction\nis to learn and use prototypes of classes during incremental updates. Despite\nsimplicity and intuition, we find that such methods suffer from inadequate\nrepresentation capability and unsatisfied feature overlap. These two factors\ncause class-wise confusion and limited performance. In this paper, we develop a\nConfusion-REduced AuTo-Encoder classifier (CREATE) for CIL. Specifically, our\nmethod employs a lightweight auto-encoder module to learn compact manifold for\neach class in the latent subspace, constraining samples to be well\nreconstructed only on the semantically correct auto-encoder. Thus, the\nrepresentation stability and capability of class distributions are enhanced,\nalleviating the potential class-wise confusion problem. To further distinguish\nthe overlapped features, we propose a confusion-aware latent space separation\nloss that ensures samples are closely distributed in their corresponding\nlow-dimensional manifold while keeping away from the distributions of features\nfrom other classes. Our method demonstrates stronger representational capacity\nand discrimination ability by learning disentangled manifolds and reduces class\nconfusion. Extensive experiments on multiple datasets and settings show that\nCREATE outperforms other state-of-the-art methods up to 5.41%.\n", "date": "2025-03-22", "categories": []}
{"id": "2503.18042", "title": "DualCP: Rehearsal-Free Domain-Incremental Learning via Dual-Level\n  Concept Prototype", "abstract": "  Domain-Incremental Learning (DIL) enables vision models to adapt to changing\nconditions in real-world environments while maintaining the knowledge acquired\nfrom previous domains. Given privacy concerns and training time, Rehearsal-Free\nDIL (RFDIL) is more practical. Inspired by the incremental cognitive process of\nthe human brain, we design Dual-level Concept Prototypes (DualCP) for each\nclass to address the conflict between learning new knowledge and retaining old\nknowledge in RFDIL. To construct DualCP, we propose a Concept Prototype\nGenerator (CPG) that generates both coarse-grained and fine-grained prototypes\nfor each class. Additionally, we introduce a Coarse-to-Fine calibrator (C2F) to\nalign image features with DualCP. Finally, we propose a Dual Dot-Regression\n(DDR) loss function to optimize our C2F module. Extensive experiments on the\nDomainNet, CDDB, and CORe50 datasets demonstrate the effectiveness of our\nmethod.\n", "date": "2025-03-23", "categories": []}
{"id": "2503.18064", "title": "Dynamic Allocation Hypernetwork with Adaptive Model Recalibration for\n  FCL", "abstract": "  Federated continual learning (FCL) offers an emerging pattern to facilitate\nthe applicability of federated learning (FL) in real-world scenarios, where\ntasks evolve dynamically and asynchronously across clients, especially in\nmedical scenario. Existing server-side FCL methods in nature domain construct a\ncontinually learnable server model by client aggregation on all-involved tasks.\nHowever, they are challenged by: (1) Catastrophic forgetting for previously\nlearned tasks, leading to error accumulation in server model, making it\ndifficult to sustain comprehensive knowledge across all tasks. (2) Biased\noptimization due to asynchronous tasks handled across different clients,\nleading to the collision of optimization targets of different clients at the\nsame time steps. In this work, we take the first step to propose a novel\nserver-side FCL pattern in medical domain, Dynamic Allocation Hypernetwork with\nadaptive model recalibration (\\textbf{FedDAH}). It is to facilitate\ncollaborative learning under the distinct and dynamic task streams across\nclients. To alleviate the catastrophic forgetting, we propose a dynamic\nallocation hypernetwork (DAHyper) where a continually updated hypernetwork is\ndesigned to manage the mapping between task identities and their associated\nmodel parameters, enabling the dynamic allocation of the model across clients.\nFor the biased optimization, we introduce a novel adaptive model recalibration\n(AMR) to incorporate the candidate changes of historical models into current\nserver updates, and assign weights to identical tasks across different time\nsteps based on the similarity for continual optimization. Extensive experiments\non the AMOS dataset demonstrate the superiority of our FedDAH to other FCL\nmethods on sites with different task streams. The code is\navailable:https://github.com/jinlab-imvr/FedDAH.\n", "date": "2025-03-23", "categories": []}
{"id": "2503.18371", "title": "Do Your Best and Get Enough Rest for Continual Learning", "abstract": "  According to the forgetting curve theory, we can enhance memory retention by\nlearning extensive data and taking adequate rest. This means that in order to\neffectively retain new knowledge, it is essential to learn it thoroughly and\nensure sufficient rest so that our brain can memorize without forgetting. The\nmain takeaway from this theory is that learning extensive data at once\nnecessitates sufficient rest before learning the same data again. This aspect\nof human long-term memory retention can be effectively utilized to address the\ncontinual learning of neural networks. Retaining new knowledge for a long\nperiod of time without catastrophic forgetting is the critical problem of\ncontinual learning. Therefore, based on Ebbinghaus' theory, we introduce the\nview-batch model that adjusts the learning schedules to optimize the recall\ninterval between retraining the same samples. The proposed view-batch model\nallows the network to get enough rest to learn extensive knowledge from the\nsame samples with a recall interval of sufficient length. To this end, we\nspecifically present two approaches: 1) a replay method that guarantees the\noptimal recall interval, and 2) a self-supervised learning that acquires\nextensive knowledge from a single training sample at a time. We empirically\nshow that these approaches of our method are aligned with the forgetting curve\ntheory, which can enhance long-term memory. In our experiments, we also\ndemonstrate that our method significantly improves many state-of-the-art\ncontinual learning methods in various protocols and scenarios. We open-source\nthis project at https://github.com/hankyul2/ViewBatchModel.\n", "date": "2025-03-24", "categories": []}
{"id": "2503.18403", "title": "Knowledge Graph Enhanced Generative Multi-modal Models for\n  Class-Incremental Learning", "abstract": "  Continual learning in computer vision faces the critical challenge of\ncatastrophic forgetting, where models struggle to retain prior knowledge while\nadapting to new tasks. Although recent studies have attempted to leverage the\ngeneralization capabilities of pre-trained models to mitigate overfitting on\ncurrent tasks, models still tend to forget details of previously learned\ncategories as tasks progress, leading to misclassification. To address these\nlimitations, we introduce a novel Knowledge Graph Enhanced Generative\nMulti-modal model (KG-GMM) that builds an evolving knowledge graph throughout\nthe learning process. Our approach utilizes relationships within the knowledge\ngraph to augment the class labels and assigns different relations to similar\ncategories to enhance model differentiation. During testing, we propose a\nKnowledge Graph Augmented Inference method that locates specific categories by\nanalyzing relationships within the generated text, thereby reducing the loss of\ndetailed information about old classes when learning new knowledge and\nalleviating forgetting. Experiments demonstrate that our method effectively\nleverages relational information to help the model correct mispredictions,\nachieving state-of-the-art results in both conventional CIL and few-shot CIL\nsettings, confirming the efficacy of knowledge graphs at preserving knowledge\nin the continual learning scenarios.\n", "date": "2025-03-24", "categories": []}
{"id": "2503.18511", "title": "Global Convergence of Continual Learning on Non-IID Data", "abstract": "  Continual learning, which aims to learn multiple tasks sequentially, has\ngained extensive attention. However, most existing work focuses on empirical\nstudies, and the theoretical aspect remains under-explored. Recently, a few\ninvestigations have considered the theory of continual learning only for linear\nregressions, establishes the results based on the strict independent and\nidentically distributed (i.i.d.) assumption and the persistent excitation on\nthe feature data that may be difficult to verify or guarantee in practice. To\novercome this fundamental limitation, in this paper, we provide a general and\ncomprehensive theoretical analysis for continual learning of regression models.\nBy utilizing the stochastic Lyapunov function and martingale estimation\ntechniques, we establish the almost sure convergence results of continual\nlearning under a general data condition for the first time. Additionally,\nwithout any excitation condition imposed on the data, the convergence rates for\nthe forgetting and regret metrics are provided.\n", "date": "2025-03-24", "categories": []}
{"id": "2311.15965", "title": "FALCON: Fairness Learning via Contrastive Attention Approach to\n  Continual Semantic Scene Understanding", "abstract": "  Continual Learning in semantic scene segmentation aims to continually learn\nnew unseen classes in dynamic environments while maintaining previously learned\nknowledge. Prior studies focused on modeling the catastrophic forgetting and\nbackground shift challenges in continual learning. However, fairness, another\nmajor challenge that causes unfair predictions leading to low performance among\nmajor and minor classes, still needs to be well addressed. In addition, prior\nmethods have yet to model the unknown classes well, thus resulting in producing\nnon-discriminative features among unknown classes. This work presents a novel\nFairness Learning via Contrastive Attention Approach to continual learning in\nsemantic scene understanding. In particular, we first introduce a new Fairness\nContrastive Clustering loss to address the problems of catastrophic forgetting\nand fairness. Then, we propose an attention-based visual grammar approach to\neffectively model the background shift problem and unknown classes, producing\nbetter feature representations for different unknown classes. Through our\nexperiments, our proposed approach achieves State-of-the-Art (SoTA) performance\non different continual learning benchmarks, i.e., ADE20K, Cityscapes, and\nPascal VOC. It promotes the fairness of the continual semantic segmentation\nmodel.\n", "date": "2023-11-27", "categories": []}
{"id": "2503.16793", "title": "Restoring Forgotten Knowledge in Non-Exemplar Class Incremental Learning\n  through Test-Time Semantic Evolution", "abstract": "  Continual learning aims to accumulate knowledge over a data stream while\nmitigating catastrophic forgetting. In Non-exemplar Class Incremental Learning\n(NECIL), forgetting arises during incremental optimization because old classes\nare inaccessible, hindering the retention of prior knowledge. To solve this,\nprevious methods struggle in achieving the stability-plasticity balance in the\ntraining stages. However, we note that the testing stage is rarely considered\namong them, but is promising to be a solution to forgetting. Therefore, we\npropose RoSE, which is a simple yet effective method that\n\\textbf{R}est\\textbf{o}res forgotten knowledge through test-time\n\\textbf{S}emantic \\textbf{E}volution. Specifically designed for minimizing\nforgetting, RoSE is a test-time semantic drift compensation framework that\nenables more accurate drift estimation in a self-supervised manner. Moreover,\nto avoid incomplete optimization during online testing, we derive an analytical\nsolution as an alternative to gradient descent. We evaluate RoSE on CIFAR-100,\nTinyImageNet, and ImageNet100 datasets, under both cold-start and warm-start\nsettings. Our method consistently outperforms most state-of-the-art (SOTA)\nmethods across various scenarios, validating the potential and feasibility of\ntest-time evolution in NECIL.\n", "date": "2025-03-21", "categories": []}
{"id": "2503.17017", "title": "Specifying What You Know or Not for Multi-Label Class-Incremental\n  Learning", "abstract": "  Existing class incremental learning is mainly designed for single-label\nclassification task, which is ill-equipped for multi-label scenarios due to the\ninherent contradiction of learning objectives for samples with incomplete\nlabels. We argue that the main challenge to overcome this contradiction in\nmulti-label class-incremental learning (MLCIL) lies in the model's inability to\nclearly distinguish between known and unknown knowledge. This ambiguity hinders\nthe model's ability to retain historical knowledge, master current classes, and\nprepare for future learning simultaneously. In this paper, we target at\nspecifying what is known or not to accommodate Historical, Current, and\nProspective knowledge for MLCIL and propose a novel framework termed as HCP.\nSpecifically, (i) we clarify the known classes by dynamic feature purification\nand recall enhancement with distribution prior, enhancing the precision and\nretention of known information. (ii) We design prospective knowledge mining to\nprobe the unknown, preparing the model for future learning. Extensive\nexperiments validate that our method effectively alleviates catastrophic\nforgetting in MLCIL, surpassing the previous state-of-the-art by 3.3% on\naverage accuracy for MS-COCO B0-C10 setting without replay buffers.\n", "date": "2025-03-21", "categories": []}
{"id": "2408.12629", "title": "Continual Gesture Learning without Data via Synthetic Feature Sampling", "abstract": "  Data-Free Class Incremental Learning (DFCIL) aims to enable models to\ncontinuously learn new classes while retraining knowledge of old classes, even\nwhen the training data for old classes is unavailable. Although explored\nprimarily with image datasets by researchers, this study focuses on\ninvestigating DFCIL for skeleton-based gesture classification due to its\nsignificant real-world implications, particularly considering the growing\nprevalence of VR/AR headsets where gestures serve as the primary means of\ncontrol and interaction. In this work, we made an intriguing observation:\nskeleton models trained with base classes(even very limited) demonstrate strong\ngeneralization capabilities to unseen classes without requiring additional\ntraining. Building on this insight, we developed Synthetic Feature Replay (SFR)\nthat can sample synthetic features from class prototypes to replay for old\nclasses and augment for new classes (under a few-shot setting). Our proposed\nmethod showcases significant advancements over the state-of-the-art, achieving\nup to 15% enhancements in mean accuracy across all steps and largely mitigating\nthe accuracy imbalance between base classes and new classes.\n", "date": "2024-08-21", "categories": []}
{"id": "2408.16939", "title": "Theoretical Insights into Overparameterized Models in Multi-Task and\n  Replay-Based Continual Learning", "abstract": "  Multi-task learning (MTL) is a machine learning paradigm that aims to improve\nthe generalization performance of a model on multiple related tasks by training\nit simultaneously on those tasks. Unlike MTL, where the model has instant\naccess to the training data of all tasks, continual learning (CL) involves\nadapting to new sequentially arriving tasks over time without forgetting the\npreviously acquired knowledge. Despite the wide practical adoption of CL and\nMTL and extensive literature on both areas, there remains a gap in the\ntheoretical understanding of these methods when used with overparameterized\nmodels such as deep neural networks. This paper studies the overparameterized\nlinear models as a proxy for more complex models. We develop theoretical\nresults describing the effect of various system parameters on the model's\nperformance in an MTL setup. Specifically, we study the impact of model size,\ndataset size, and task similarity on the generalization error and knowledge\ntransfer. Additionally, we present theoretical results to characterize the\nperformance of replay-based CL models. Our results reveal the impact of buffer\nsize and model capacity on the forgetting rate in a CL setup and help shed\nlight on some of the state-of-the-art CL methods. Finally, through extensive\nempirical evaluations, we demonstrate that our theoretical findings are also\napplicable to deep neural networks, offering valuable guidance for designing\nMTL and CL models in practice.\n", "date": "2024-08-29", "categories": []}
{"id": "2410.09449", "title": "Continuous Risk Prediction", "abstract": "  Lifelong learning (LL) capabilities are essential for QA models to excel in\nreal-world applications, and architecture-based LL approaches have proven to be\na promising direction for achieving this goal. However, adapting existing\nmethods to QA tasks is far from straightforward. Many prior approaches either\nrely on access to task identities during testing or fail to adequately model\nsamples from unseen tasks, which limits their practical applicability. To\novercome these limitations, we introduce Diana , a novel\n\\underline{d}ynam\\underline{i}c \\underline{a}rchitecture-based\nlifelo\\underline{n}g Q\\underline{A} framework designed to learn a sequence of\nQA tasks using a prompt-enhanced language model.Diana leverages four\nhierarchically structured types of prompts to capture QA knowledge at multiple\nlevels of granularity. Task-level prompts are specifically designed to encode\ntask-specific knowledge, ensuring strong lifelong learning performance.\nMeanwhile, instance-level prompts are utilized to capture shared knowledge\nacross diverse input samples, enhancing the model's generalization\ncapabilities. Additionally, Diana incorporates dedicated prompts to explicitly\nhandle unseen tasks and introduces a set of prompt key vectors that facilitate\nefficient knowledge transfer and sharing between tasks. Through extensive\nexperimentation, we demonstrate that Diana achieves state-of-the-art\nperformance among lifelong QA models, with particularly notable improvements in\nits ability to handle previously unseen tasks. This makes Diana a significant\nadvancement in the field of lifelong learning for question-answering systems.\n", "date": "2024-10-12", "categories": []}
{"id": "2503.15697", "title": "Technical Report for the 5th CLVision Challenge at CVPR: Addressing the\n  Class-Incremental with Repetition using Unlabeled Data -- 4th Place Solution", "abstract": "  This paper outlines our approach to the 5th CLVision challenge at CVPR, which\naddresses the Class-Incremental with Repetition (CIR) scenario. In contrast to\ntraditional class incremental learning, this novel setting introduces unique\nchallenges and research opportunities, particularly through the integration of\nunlabeled data into the training process. In the CIR scenario, encountered\nclasses may reappear in later learning experiences, and each experience may\ninvolve only a subset of the overall class distribution. Additionally, the\nunlabeled data provided during training may include instances of unseen\nclasses, or irrelevant classes which should be ignored. Our approach focuses on\nretaining previously learned knowledge by utilizing knowledge distillation and\npseudo-labeling techniques. The key characteristic of our method is the\nexploitation of unlabeled data during training, in order to maintain optimal\nperformance on instances of previously encountered categories and reduce the\ndetrimental effects of catastrophic forgetting. Our method achieves an average\naccuracy of 16.68\\% during the pre-selection phase and 21.19% during the final\nevaluation phase, outperforming the baseline accuracy of 9.39%. We provide the\nimplementation code at\nhttps://github.com/panagiotamoraiti/continual-learning-challenge-2024 .\n", "date": "2025-03-19", "categories": []}
{"id": "2503.15924", "title": "Towards Automatic Continual Learning: A Self-Adaptive Framework for\n  Continual Instruction Tuning", "abstract": "  Continual instruction tuning enables large language models (LLMs) to learn\nincrementally while retaining past knowledge, whereas existing methods\nprimarily focus on how to retain old knowledge rather than on selecting which\nnew knowledge to learn. In domain-specific contexts, maintaining data quality\nand managing system constraints remain key challenges. To address these issues,\nwe propose an automated continual instruction tuning framework that dynamically\nfilters incoming data, which identify and reduce redundant data across\nsuccessive updates. Our approach utilizes a small proxy model for efficient\nperplexity-based filtering, and updates the proxy to ensure that the filtering\ncriteria remain aligned with the evolving state of the deployed model. Compared\nto existing static data selection methods, our framework can effectively handle\nincrementally acquired data and shifting distributions. Additionally, it\naddresses practical deployment challenges by enabling seamless model updates,\nsupporting version rollback and incorporating automatic checkpoint evaluation.\nWe evaluated the system in real-world medical scenarios. It reduced\ncomputational costs by 66.7% and improved model performance, and achieved\nautonomous updates, thus demonstrating its effectiveness for automatic\ncontinual instruction tuning.\n", "date": "2025-03-20", "categories": []}
{"id": "2503.14832", "title": "H2ST: Hierarchical Two-Sample Tests for Continual Out-of-Distribution\n  Detection", "abstract": "  Task Incremental Learning (TIL) is a specialized form of Continual Learning\n(CL) in which a model incrementally learns from non-stationary data streams.\nExisting TIL methodologies operate under the closed-world assumption, presuming\nthat incoming data remains in-distribution (ID). However, in an open-world\nsetting, incoming samples may originate from out-of-distribution (OOD) sources,\nwith their task identities inherently unknown. Continually detecting OOD\nsamples presents several challenges for current OOD detection methods: reliance\non model outputs leads to excessive dependence on model performance, selecting\nsuitable thresholds is difficult, hindering real-world deployment, and binary\nID/OOD classification fails to provide task-level identification. To address\nthese issues, we propose a novel continual OOD detection method called the\nHierarchical Two-sample Tests (H2ST). H2ST eliminates the need for threshold\nselection through hypothesis testing and utilizes feature maps to better\nexploit model capabilities without excessive dependence on model performance.\nThe proposed hierarchical architecture enables task-level detection with\nsuperior performance and lower overhead compared to non-hierarchical classifier\ntwo-sample tests. Extensive experiments and analysis validate the effectiveness\nof H2ST in open-world TIL scenarios and its superiority to the existing\nmethods. Code is available at\n\\href{https://github.com/YuhangLiuu/H2ST}{https://github.com/YuhangLiuu/H2ST}.\n", "date": "2025-03-19", "categories": []}
{"id": "2411.02210", "title": "One VLM to Keep it Learning: Generation and Balancing for Data-free\n  Continual Visual Question Answering", "abstract": "  Vision-Language Models (VLMs) have shown significant promise in Visual\nQuestion Answering (VQA) tasks by leveraging web-scale multimodal datasets.\nHowever, these models often struggle with continual learning due to\ncatastrophic forgetting when adapting to new tasks. As an effective remedy to\nmitigate catastrophic forgetting, rehearsal strategy uses the data of past\ntasks upon learning new task. However, such strategy incurs the need of storing\npast data, which might not be feasible due to hardware constraints or privacy\nconcerns. In this work, we propose the first data-free method that leverages\nthe language generation capability of a VLM, instead of relying on external\nmodels, to produce pseudo-rehearsal data for addressing continual VQA. Our\nproposal, named as GaB, generates pseudo-rehearsal data by posing previous task\nquestions on new task data. Yet, despite being effective, the distribution of\ngenerated questions skews towards the most frequently posed questions due to\nthe limited and task-specific training data. To mitigate this issue, we\nintroduce a pseudo-rehearsal balancing module that aligns the generated data\ntowards the ground-truth data distribution using either the question\nmeta-statistics or an unsupervised clustering method. We evaluate our proposed\nmethod on two recent benchmarks, \\ie VQACL-VQAv2 and CLOVE-function benchmarks.\nGaB outperforms all the data-free baselines with substantial improvement in\nmaintaining VQA performance across evolving tasks, while being on-par with\nmethods with access to the past data.\n", "date": "2024-11-04", "categories": []}
{"id": "2502.20032", "title": "Order-Robust Class Incremental Learning: Graph-Driven Dynamic Similarity\n  Grouping", "abstract": "  Class Incremental Learning (CIL) aims to enable models to learn new classes\nsequentially while retaining knowledge of previous ones. Although current\nmethods have alleviated catastrophic forgetting (CF), recent studies highlight\nthat the performance of CIL models is highly sensitive to the order of class\narrival, particularly when sequentially introduced classes exhibit high\ninter-class similarity. To address this critical yet understudied challenge of\nclass order sensitivity, we first extend existing CIL frameworks through\ntheoretical analysis, proving that grouping classes with lower pairwise\nsimilarity during incremental phases significantly improves model robustness to\norder variations. Building on this insight, we propose Graph-Driven Dynamic\nSimilarity Grouping (GDDSG), a novel method that employs graph coloring\nalgorithms to dynamically partition classes into similarity-constrained groups.\nEach group trains an isolated CIL sub-model and constructs meta-features for\nclass group identification. Experimental results demonstrate that our method\neffectively addresses the issue of class order sensitivity while achieving\noptimal performance in both model accuracy and anti-forgetting capability. Our\ncode is available at https://github.com/AIGNLAI/GDDSG.\n", "date": "2025-02-27", "categories": []}
{"id": "2503.13869", "title": "Robust3D-CIL: Robust Class-Incremental Learning for 3D Perception", "abstract": "  3D perception plays a crucial role in real-world applications such as\nautonomous driving, robotics, and AR/VR. In practical scenarios, 3D perception\nmodels must continuously adapt to new data and emerging object categories, but\nretraining from scratch incurs prohibitive costs. Therefore, adopting\nclass-incremental learning (CIL) becomes particularly essential. However,\nreal-world 3D point cloud data often include corrupted samples, which poses\nsignificant challenges for existing CIL methods and leads to more severe\nforgetting on corrupted data. To address these challenges, we consider the\nscenario in which a CIL model can be updated using point clouds with unknown\ncorruption to better simulate real-world conditions. Inspired by Farthest Point\nSampling, we propose a novel exemplar selection strategy that effectively\npreserves intra-class diversity when selecting replay exemplars, mitigating\nforgetting induced by data corruption. Furthermore, we introduce a point cloud\ndownsampling-based replay method to utilize the limited replay buffer memory\nmore efficiently, thereby further enhancing the model's continual learning\nability. Extensive experiments demonstrate that our method improves the\nperformance of replay-based CIL baselines by 2% to 11%, proving its\neffectiveness and promising potential for real-world 3D applications.\n", "date": "2025-03-18", "categories": []}
{"id": "2410.15143", "title": "Budgeted Online Continual Learning by Adaptive Layer Freezing and\n  Frequency-based Sampling", "abstract": "  The majority of online continual learning (CL) advocates single-epoch\ntraining and imposes restrictions on the size of replay memory. However,\nsingle-epoch training would incur a different amount of computations per CL\nalgorithm, and the additional storage cost to store logit or model in addition\nto replay memory is largely ignored in calculating the storage budget. Arguing\ndifferent computational and storage budgets hinder fair comparison among CL\nalgorithms in practice, we propose to use floating point operations (FLOPs) and\ntotal memory size in Byte as a metric for computational and memory budgets,\nrespectively, to compare and develop CL algorithms in the same 'total resource\nbudget.' To improve a CL method in a limited total budget, we propose adaptive\nlayer freezing that does not update the layers for less informative batches to\nreduce computational costs with a negligible loss of accuracy. In addition, we\npropose a memory retrieval method that allows the model to learn the same\namount of knowledge as using random retrieval in fewer iterations. Empirical\nvalidations on the CIFAR-10/100, CLEAR-10/100, and ImageNet-1K datasets\ndemonstrate that the proposed approach outperforms the state-of-the-art methods\nwithin the same total budget\n", "date": "2024-10-19", "categories": []}
{"id": "2503.12053", "title": "Ferret: An Efficient Online Continual Learning Framework under Varying\n  Memory Constraints", "abstract": "  In the realm of high-frequency data streams, achieving real-time learning\nwithin varying memory constraints is paramount. This paper presents Ferret, a\ncomprehensive framework designed to enhance online accuracy of Online Continual\nLearning (OCL) algorithms while dynamically adapting to varying memory budgets.\nFerret employs a fine-grained pipeline parallelism strategy combined with an\niterative gradient compensation algorithm, ensuring seamless handling of\nhigh-frequency data with minimal latency, and effectively counteracting the\nchallenge of stale gradients in parallel training. To adapt to varying memory\nbudgets, its automated model partitioning and pipeline planning optimizes\nperformance regardless of memory limitations. Extensive experiments across 20\nbenchmarks and 5 integrated OCL algorithms show Ferret's remarkable efficiency,\nachieving up to 3.7$\\times$ lower memory overhead to reach the same online\naccuracy compared to competing methods. Furthermore, Ferret consistently\noutperforms these methods across diverse memory budgets, underscoring its\nsuperior adaptability. These findings position Ferret as a premier solution for\nefficient and adaptive OCL framework in real-time environments.\n", "date": "2025-03-15", "categories": []}
{"id": "2503.12193", "title": "S2IL: Structurally Stable Incremental Learning", "abstract": "  Feature Distillation (FD) strategies are proven to be effective in mitigating\nCatastrophic Forgetting (CF) seen in Class Incremental Learning (CIL). However,\ncurrent FD approaches enforce strict alignment of feature magnitudes and\ndirections across incremental steps, limiting the model's ability to adapt to\nnew knowledge. In this paper we propose Structurally Stable Incremental\nLearning(S22IL), a FD method for CIL that mitigates CF by focusing on\npreserving the overall spatial patterns of features which promote flexible\n(plasticity) yet stable representations that preserve old knowledge\n(stability). We also demonstrate that our proposed method S2IL achieves strong\nincremental accuracy and outperforms other FD methods on SOTA benchmark\ndatasets CIFAR-100, ImageNet-100 and ImageNet-1K. Notably, S2IL outperforms\nother methods by a significant margin in scenarios that have a large number of\nincremental tasks.\n", "date": "2025-03-15", "categories": []}
{"id": "2503.12635", "title": "Hybrid Learners Do Not Forget: A Brain-Inspired Neuro-Symbolic Approach\n  to Continual Learning", "abstract": "  Continual learning is crucial for creating AI agents that can learn and\nimprove themselves autonomously. A primary challenge in continual learning is\nto learn new tasks without losing previously learned knowledge. Current\ncontinual learning methods primarily focus on enabling a neural network with\nmechanisms that mitigate forgetting effects. Inspired by the two distinct\nsystems in the human brain, System 1 and System 2, we propose a Neuro-Symbolic\nBrain-Inspired Continual Learning (NeSyBiCL) framework that incorporates two\nsubsystems to solve continual learning: A neural network model responsible for\nquickly adapting to the most recent task, together with a symbolic reasoner\nresponsible for retaining previously acquired knowledge from previous tasks.\nMoreover, we design an integration mechanism between these components to\nfacilitate knowledge transfer from the symbolic reasoner to the neural network.\nWe also introduce two compositional continual learning benchmarks and\ndemonstrate that NeSyBiCL is effective and leads to superior performance\ncompared to continual learning methods that merely rely on neural architectures\nto address forgetting.\n", "date": "2025-03-16", "categories": []}
{"id": "2503.12745", "title": "ProtoDepth: Unsupervised Continual Depth Completion with Prototypes", "abstract": "  We present ProtoDepth, a novel prototype-based approach for continual\nlearning of unsupervised depth completion, the multimodal 3D reconstruction\ntask of predicting dense depth maps from RGB images and sparse point clouds.\nThe unsupervised learning paradigm is well-suited for continual learning, as\nground truth is not needed. However, when training on new non-stationary\ndistributions, depth completion models will catastrophically forget previously\nlearned information. We address forgetting by learning prototype sets that\nadapt the latent features of a frozen pretrained model to new domains. Since\nthe original weights are not modified, ProtoDepth does not forget when\ntest-time domain identity is known. To extend ProtoDepth to the challenging\nsetting where the test-time domain identity is withheld, we propose to learn\ndomain descriptors that enable the model to select the appropriate prototype\nset for inference. We evaluate ProtoDepth on benchmark dataset sequences, where\nwe reduce forgetting compared to baselines by 52.2% for indoor and 53.2% for\noutdoor to achieve the state of the art.\n", "date": "2025-03-17", "categories": []}
{"id": "2503.12923", "title": "Lifelong Reinforcement Learning with Similarity-Driven Weighting by\n  Large Models", "abstract": "  Lifelong Reinforcement Learning (LRL) holds significant potential for\naddressing sequential tasks, but it still faces considerable challenges. A key\ndifficulty lies in effectively preventing catastrophic forgetting and\nfacilitating knowledge transfer while maintaining reliable decision-making\nperformance across subsequent tasks in dynamic environments. To tackle this, we\npropose a novel framework, SDW (Similarity-Driven Weighting Framework), which\nleverages large-language-model-generated dynamic functions to precisely control\nthe training process. The core of SDW lies in two functions pre-generated by\nlarge models: the task similarity function and the weight computation function.\nThe task similarity function extracts multidimensional features from task\ndescriptions to quantify the similarities and differences between tasks in\nterms of states, actions, and rewards. The weight computation function\ndynamically generates critical training parameters based on the similarity\ninformation, including the proportion of old task data stored in the Replay\nBuffer and the strategy consistency weight in the loss function, enabling an\nadaptive balance between learning new tasks and transferring knowledge from\nprevious tasks. By generating function code offline prior to training, rather\nthan relying on large-model inference during the training process, the SDW\nframework reduces computational overhead while maintaining efficiency in\nsequential task scenarios. Experimental results on Atari and MiniHack\nsequential tasks demonstrate that SDW significantly outperforms existing\nlifelong reinforcement learning methods.\n", "date": "2025-03-17", "categories": []}
{"id": "2404.06466", "title": "Hyperparameter Selection in Continual Learning", "abstract": "  In continual learning (CL) -- where a learner trains on a stream of data --\nstandard hyperparameter optimisation (HPO) cannot be applied, as a learner does\nnot have access to all of the data at the same time. This has prompted the\ndevelopment of CL-specific HPO frameworks. The most popular way to tune\nhyperparameters in CL is to repeatedly train over the whole data stream with\ndifferent hyperparameter settings. However, this end-of-training HPO is\nunusable in practice since a learner can only see the stream once. Hence, there\nis an open question: what HPO framework should a practitioner use for a CL\nproblem in reality? This paper looks at this question by comparing several\nrealistic HPO frameworks. We find that none of the HPO frameworks considered,\nincluding end-of-training HPO, perform consistently better than the rest on\npopular CL benchmarks. We therefore arrive at a twofold conclusion: a) to be\nable to discriminate between HPO frameworks there is a need to move beyond the\ncurrent most commonly used CL benchmarks, and b) on the popular CL benchmarks\nexamined, a CL practitioner should use a realistic HPO framework and can select\nit based on factors separate from performance, for example compute efficiency.\n", "date": "2024-04-09", "categories": []}
{"id": "2411.16870", "title": "RECAST: Reparameterized, Compact weight Adaptation for Sequential Tasks", "abstract": "  Incremental learning aims to adapt to new sets of categories over time with\nminimal computational overhead. Prior work often addresses this task by\ntraining efficient task-specific adaptors that modify frozen layer weights or\nfeatures to capture relevant information without affecting predictions on\npreviously learned categories. While these adaptors are generally more\nefficient than finetuning the entire network, they still require tens to\nhundreds of thousands of task-specific trainable parameters even for relatively\nsmall networks, making it challenging to operate on resource-constrained\nenvironments with high communication costs like edge devices or mobile phones.\nThus, we propose Reparameterized, Compact weight Adaptation for Sequential\nTasks (RECAST), a novel method that dramatically reduces task-specific\ntrainable parameters to fewer than 50 - several orders of magnitude less than\ncompeting methods like LoRA. RECAST accomplishes this efficiency by learning to\ndecompose layer weights into a soft parameter-sharing framework consisting of\nshared weight templates and very few module-specific scaling factors or\ncoefficients. This soft parameter-sharing framework allows for effective\ntask-wise reparameterization by tuning only these coefficients while keeping\ntemplates frozen.A key innovation of RECAST is the novel weight reconstruction\npipeline called Neural Mimicry, which eliminates the need for pretraining from\nscratch. This allows for high-fidelity emulation of existing pretrained weights\nwithin our framework and provides quick adaptability to any model scale and\narchitecture. Extensive experiments across six datasets demonstrate RECAST\noutperforms the state-of-the-art by up to 3% across various scales,\narchitectures, and parameter spaces Moreover, we show that RECAST's\narchitecture-agnostic nature allows for seamless integration with existing\nmethods, further boosting performance.\n", "date": "2024-11-25", "categories": []}
{"id": "2501.15356", "title": "Federated Class-Incremental Learning: A Hybrid Approach Using Latent\n  Exemplars and Data-Free Techniques to Address Local and Global Forgetting", "abstract": "  Federated Class-Incremental Learning (FCIL) refers to a scenario where a\ndynamically changing number of clients collaboratively learn an ever-increasing\nnumber of incoming tasks. FCIL is known to suffer from local forgetting due to\nclass imbalance at each client and global forgetting due to class imbalance\nacross clients. We develop a mathematical framework for FCIL that formulates\nlocal and global forgetting. Then, we propose an approach called Hybrid\nRehearsal (HR), which utilizes latent exemplars and data-free techniques to\naddress local and global forgetting, respectively. HR employs a customized\nautoencoder designed for both data classification and the generation of\nsynthetic data. To determine the embeddings of new tasks for all clients in the\nlatent space of the encoder, the server uses the Lennard-Jones Potential\nformulations. Meanwhile, at the clients, the decoder decodes the stored\nlow-dimensional latent space exemplars back to the high-dimensional input\nspace, used to address local forgetting. To overcome global forgetting, the\ndecoder generates synthetic data. Furthermore, our mathematical framework\nproves that our proposed approach HR can, in principle, tackle the two local\nand global forgetting challenges. In practice, extensive experiments\ndemonstrate that while preserving privacy, our proposed approach outperforms\nthe state-of-the-art baselines on multiple FCIL benchmarks with low compute and\nmemory footprints.\n", "date": "2025-01-26", "categories": []}
{"id": "2503.01066", "title": "Alchemist: Towards the Design of Efficient Online Continual Learning\n  System", "abstract": "  Continual learning has become a promising solution to refine large language\nmodels incrementally by leveraging user feedback. In particular, online\ncontinual learning - iteratively training the model with small batches of user\nfeedback - has demonstrated notable performance improvements. However, the\nexisting practice of separating training and serving processes forces the\nonline trainer to recompute the intermediate results already done during\nserving. Such redundant computations can account for 30%-42% of total training\ntime.\n  In this paper, we propose Alchemist, to the best of our knowledge, the first\nonline continual learning system that efficiently reuses serving activations to\nincrease training throughput. Alchemist introduces two key techniques: (1)\nrecording and storing activations and KV cache only during the prefill phase to\nminimize latency and memory overhead; and (2) smart activation offloading and\nhedging. Evaluations with inputs of varied token length sampled from ShareGPT\ndataset show that compared with a separate training cluster, Alchemist\nsignificantly increases training throughput by up to 1.72x, reduces up to 47%\nmemory usage during training, and supports up to 2x more training tokens - all\nwhile maintaining negligible impact on serving latency.\n", "date": "2025-03-03", "categories": []}
{"id": "2503.11349", "title": "An experimental approach on Few Shot Class Incremental Learning", "abstract": "  Few-Shot Class-Incremental Learning (FSCIL) represents a cutting-edge\nparadigm within the broader scope of machine learning, designed to empower\nmodels with the ability to assimilate new classes of data with limited examples\nwhile safeguarding existing knowledge. The paper will present different\nsolutions which contain extensive experiments across large-scale datasets,\ndomain shifts, and network architectures to evaluate and compare the selected\nmethods. We highlight their advantages and then present an experimental\napproach with the purpose of improving the most promising one by replacing the\nvisual-language (V-L) model (CLIP) with another V-L model (CLOOB) that seem to\noutperform it on zero-shot learning tasks. The aim of this report is to present\nan experimental method for FSCIL that would improve its performance. We also\nplan to offer an overview followed by an analysis of the recent advancements in\nFSCIL domain, focusing on various strategies to mitigate catastrophic\nforgetting and improve the adaptability of models to evolving tasks and\ndatasets.\n", "date": "2025-03-14", "categories": []}
{"id": "2410.04891", "title": "Low-Rank Continual Personalization of Diffusion Models", "abstract": "  Recent personalization methods for diffusion models, such as Dreambooth and\nLoRA, allow fine-tuning pre-trained models to generate new concepts. However,\napplying these techniques across consecutive tasks in order to include, e.g.,\nnew objects or styles, leads to a forgetting of previous knowledge due to\nmutual interference between their adapters. In this work, we tackle the problem\nof continual customization under a rigorous regime with no access to past\ntasks' adapters. In such a scenario, we investigate how different adapters'\ninitialization and merging methods can improve the quality of the final model.\nTo that end, we evaluate the naive continual fine-tuning of customized models\nand compare this approach with three methods for consecutive adapters'\ntraining: sequentially merging new adapters, merging orthogonally initialized\nadapters, and updating only relevant task-specific weights. In our experiments,\nwe show that the proposed techniques mitigate forgetting when compared to the\nnaive approach. In our studies, we show different traits of selected techniques\nand their effect on the plasticity and stability of the continually adapted\nmodel. Repository with the code is available at\nhttps://github.com/luk-st/continual-lora.\n", "date": "2024-10-07", "categories": []}
{"id": "2503.10214", "title": "Singular Value Fine-tuning for Few-Shot Class-Incremental Learning", "abstract": "  Class-Incremental Learning (CIL) aims to prevent catastrophic forgetting of\npreviously learned classes while sequentially incorporating new ones. The more\nchallenging Few-shot CIL (FSCIL) setting further complicates this by providing\nonly a limited number of samples for each new class, increasing the risk of\noverfitting in addition to standard CIL challenges. While catastrophic\nforgetting has been extensively studied, overfitting in FSCIL, especially with\nlarge foundation models, has received less attention. To fill this gap, we\npropose the Singular Value Fine-tuning for FSCIL (SVFCL) and compared it with\nexisting approaches for adapting foundation models to FSCIL, which primarily\nbuild on Parameter Efficient Fine-Tuning (PEFT) methods like prompt tuning and\nLow-Rank Adaptation (LoRA). Specifically, SVFCL applies singular value\ndecomposition to the foundation model weights, keeping the singular vectors\nfixed while fine-tuning the singular values for each task, and then merging\nthem. This simple yet effective approach not only alleviates the forgetting\nproblem but also mitigates overfitting more effectively while significantly\nreducing trainable parameters. Extensive experiments on four benchmark\ndatasets, along with visualizations and ablation studies, validate the\neffectiveness of SVFCL. The code will be made available.\n", "date": "2025-03-13", "categories": []}
{"id": "2503.10333", "title": "Generative Binary Memory: Pseudo-Replay Class-Incremental Learning on\n  Binarized Embeddings", "abstract": "  In dynamic environments where new concepts continuously emerge, Deep Neural\nNetworks (DNNs) must adapt by learning new classes while retaining previously\nacquired ones. This challenge is addressed by Class-Incremental Learning (CIL).\nThis paper introduces Generative Binary Memory (GBM), a novel CIL pseudo-replay\napproach which generates synthetic binary pseudo-exemplars. Relying on\nBernoulli Mixture Models (BMMs), GBM effectively models the multi-modal\ncharacteristics of class distributions, in a latent, binary space. With a\nspecifically-designed feature binarizer, our approach applies to any\nconventional DNN. GBM also natively supports Binary Neural Networks (BNNs) for\nhighly-constrained model sizes in embedded systems. The experimental results\ndemonstrate that GBM achieves higher than state-of-the-art average accuracy on\nCIFAR100 (+2.9%) and TinyImageNet (+1.5%) for a ResNet-18 equipped with our\nbinarizer. GBM also outperforms emerging CIL methods for BNNs, with +3.1% in\nfinal accuracy and x4.7 memory reduction, on CORE50.\n", "date": "2025-03-13", "categories": []}
{"id": "2503.09106", "title": "Freeze and Cluster: A Simple Baseline for Rehearsal-Free Continual\n  Category Discovery", "abstract": "  This paper addresses the problem of Rehearsal-Free Continual Category\nDiscovery (RF-CCD), which focuses on continuously identifying novel class by\nleveraging knowledge from labeled data. Existing methods typically train from\nscratch, overlooking the potential of base models, and often resort to data\nstorage to prevent forgetting. Moreover, because RF-CCD encompasses both\ncontinual learning and novel class discovery, previous approaches have\nstruggled to effectively integrate advanced techniques from these fields,\nresulting in less convincing comparisons and failing to reveal the unique\nchallenges posed by RF-CCD. To address these challenges, we lead the way in\nintegrating advancements from both domains and conducting extensive experiments\nand analyses. Our findings demonstrate that this integration can achieve\nstate-of-the-art results, leading to the conclusion that in the presence of\npre-trained models, the representation does not improve and may even degrade\nwith the introduction of unlabeled data. To mitigate representation\ndegradation, we propose a straightforward yet highly effective baseline method.\nThis method first utilizes prior knowledge of known categories to estimate the\nnumber of novel classes. It then acquires representations using a model\nspecifically trained on the base classes, generates high-quality pseudo-labels\nthrough k-means clustering, and trains only the classifier layer. We validate\nour conclusions and methods by conducting extensive experiments across multiple\nbenchmarks, including the Stanford Cars, CUB, iNat, and Tiny-ImageNet datasets.\nThe results clearly illustrate our findings, demonstrate the effectiveness of\nour baseline, and pave the way for future advancements in RF-CCD.\n", "date": "2025-03-12", "categories": []}
{"id": "2309.07117", "title": "PILOT: A Pre-Trained Model-Based Continual Learning Toolbox", "abstract": "  While traditional machine learning can effectively tackle a wide range of\nproblems, it primarily operates within a closed-world setting, which presents\nlimitations when dealing with streaming data. As a solution, incremental\nlearning emerges to address real-world scenarios involving new data's arrival.\nRecently, pre-training has made significant advancements and garnered the\nattention of numerous researchers. The strong performance of these pre-trained\nmodels (PTMs) presents a promising avenue for developing continual learning\nalgorithms that can effectively adapt to real-world scenarios. Consequently,\nexploring the utilization of PTMs in incremental learning has become essential.\nThis paper introduces a pre-trained model-based continual learning toolbox\nknown as PILOT. On the one hand, PILOT implements some state-of-the-art\nclass-incremental learning algorithms based on pre-trained models, such as L2P,\nDualPrompt, and CODA-Prompt. On the other hand, PILOT also fits typical\nclass-incremental learning algorithms (e.g., DER, FOSTER, and MEMO) within the\ncontext of pre-trained models to evaluate their effectiveness.\n", "date": "2023-09-13", "categories": []}
{"id": "2401.06548", "title": "Enhancing Consistency and Mitigating Bias: A Data Replay Approach for\n  Incremental Learning", "abstract": "  Deep learning systems are prone to catastrophic forgetting when learning from\na sequence of tasks, as old data from previous tasks is unavailable when\nlearning a new task. To address this, some methods propose replaying data from\nprevious tasks during new task learning, typically using extra memory to store\nreplay data. However, it is not expected in practice due to memory constraints\nand data privacy issues. Instead, data-free replay methods invert samples from\nthe classification model. While effective, these methods face inconsistencies\nbetween inverted and real training data, overlooked in recent works. To that\neffect, we propose to measure the data consistency quantitatively by some\nsimplification and assumptions. Using this measurement, we gain insight to\ndevelop a novel loss function that reduces inconsistency. Specifically, the\nloss minimizes the KL divergence between distributions of inverted and real\ndata under a tied multivariate Gaussian assumption, which is simple to\nimplement in continual learning. Additionally, we observe that old class weight\nnorms decrease continually as learning progresses. We analyze the reasons and\npropose a regularization term to balance class weights, making old class\nsamples more distinguishable. To conclude, we introduce Consistency-enhanced\ndata replay with a Debiased classifier for class incremental learning (CwD).\nExtensive experiments on CIFAR-100, Tiny-ImageNet, and ImageNet100 show\nconsistently improved performance of CwD compared to previous approaches.\n", "date": "2024-01-12", "categories": []}
{"id": "2405.16498", "title": "On Sequential Maximum a Posteriori Inference for Continual Learning", "abstract": "  We formulate sequential maximum a posteriori inference as a recursion of loss\nfunctions and reduce the problem of continual learning to approximating the\nprevious loss function. We then propose two coreset-free methods: autodiff\nquadratic consolidation, which uses an accurate and full quadratic\napproximation, and neural consolidation, which uses a neural network\napproximation. These methods are not scalable with respect to the neural\nnetwork size, and we study them for classification tasks in combination with a\nfixed pre-trained feature extractor. We also introduce simple but challenging\nclassical task sequences based on Iris and Wine datasets. We find that neural\nconsolidation performs well in the classical task sequences, where the input\ndimension is small, while autodiff quadratic consolidation performs\nconsistently well in image task sequences with a fixed pre-trained feature\nextractor, achieving comparable performance to joint maximum a posteriori\ntraining in many cases.\n", "date": "2024-05-26", "categories": []}
{"id": "2407.10545", "title": "LightCL: Compact Continual Learning with Low Memory Footprint For Edge\n  Device", "abstract": "  Continual learning (CL) is a technique that enables neural networks to\nconstantly adapt to their dynamic surroundings. Despite being overlooked for a\nlong time, this technology can considerably address the customized needs of\nusers in edge devices. Actually, most CL methods require huge resource\nconsumption by the training behavior to acquire generalizability among all\ntasks for delaying forgetting regardless of edge scenarios. Therefore, this\npaper proposes a compact algorithm called LightCL, which evaluates and\ncompresses the redundancy of already generalized components in structures of\nthe neural network. Specifically, we consider two factors of generalizability,\nlearning plasticity and memory stability, and design metrics of both to\nquantitatively assess generalizability of neural networks during CL. This\nevaluation shows that generalizability of different layers in a neural network\nexhibits a significant variation. Thus, we $\\textit{Maintain Generalizability}$\nby freezing generalized parts without the resource-intensive training process\nand $\\textit{Memorize Feature Patterns}$ by stabilizing feature extracting of\nprevious tasks to enhance generalizability for less-generalized parts with a\nlittle extra memory, which is far less than the reduction by freezing.\nExperiments illustrate that LightCL outperforms other state-of-the-art methods\nand reduces at most $\\textbf{6.16$\\times$}$ memory footprint. We also verify\nthe effectiveness of LightCL on the edge device.\n", "date": "2024-07-15", "categories": []}
{"id": "2410.04327", "title": "Leveraging Hierarchical Taxonomies in Prompt-based Continual Learning", "abstract": "  Humans perceive the world as a series of sequential events, which can be\nhierarchically organized with different levels of abstraction based on\nconceptual knowledge. Drawing inspiration from human learning behaviors, this\nwork proposes a novel approach to mitigate catastrophic forgetting in\nPrompt-based Continual Learning models by exploiting the relationships between\ncontinuously emerging class data. We find that applying human habits of\norganizing and connecting information can serve as an efficient strategy when\ntraining deep learning models. Specifically, by building a hierarchical tree\nstructure based on the expanding set of labels, we gain fresh insights into the\ndata, identifying groups of similar classes could easily cause confusion.\nAdditionally, we delve deeper into the hidden connections between classes by\nexploring the original pretrained model's behavior through an optimal\ntransport-based approach. From these insights, we propose a novel\nregularization loss function that encourages models to focus more on\nchallenging knowledge areas, thereby enhancing overall performance.\nExperimentally, our method demonstrated significant superiority over the most\nrobust state-of-the-art models on various benchmarks.\n", "date": "2024-10-06", "categories": []}
{"id": "2410.17961", "title": "Closed-form merging of parameter-efficient modules for Federated\n  Continual Learning", "abstract": "  Model merging has emerged as a crucial technique in Deep Learning, enabling\nthe integration of multiple models into a unified system while preserving\nperfor-mance and scalability. In this respect, the compositional properties of\nlow-rank adaptation techniques (e.g., LoRA) have proven beneficial, as simple\naveraging LoRA modules yields a single model that mostly integrates the\ncapabilities of all individual modules. Building on LoRA, we take a step\nfurther by imposing that the merged model matches the responses of all learned\nmodules. Solving this objective in closed form yields an indeterminate system\nwith A and B as unknown variables, indicating the existence of infinitely many\nclosed-form solutions. To address this challenge, we introduce LoRM, an\nalternating optimization strategy that trains one LoRA matrix at a time. This\nallows solving for each unknown variable individually, thus finding a unique\nsolution. We apply our proposed methodology to Federated Class-Incremental\nLearning (FCIL), ensuring alignment of model responses both between clients and\nacross tasks. Our method demonstrates state-of-the-art performance across a\nrange of FCIL scenarios. The code to reproduce our experiments is available at\ngithub.com/aimagelab/fed-mammoth.\n", "date": "2024-10-23", "categories": []}
{"id": "2503.05246", "title": "Mastering Continual Reinforcement Learning through Fine-Grained Sparse\n  Network Allocation and Dormant Neuron Exploration", "abstract": "  Continual Reinforcement Learning (CRL) is essential for developing agents\nthat can learn, adapt, and accumulate knowledge over time. However, a\nfundamental challenge persists as agents must strike a delicate balance between\nplasticity, which enables rapid skill acquisition, and stability, which ensures\nlong-term knowledge retention while preventing catastrophic forgetting. In this\npaper, we introduce SSDE, a novel structure-based approach that enhances\nplasticity through a fine-grained allocation strategy with Structured Sparsity\nand Dormant-guided Exploration. SSDE decomposes the parameter space into\nforward-transfer (frozen) parameters and task-specific (trainable) parameters.\nCrucially, these parameters are allocated by an efficient co-allocation scheme\nunder sparse coding, ensuring sufficient trainable capacity for new tasks while\npromoting efficient forward transfer through frozen parameters. However,\nstructure-based methods often suffer from rigidity due to the accumulation of\nnon-trainable parameters, limiting exploration and adaptability. To address\nthis, we further introduce a sensitivity-guided neuron reactivation mechanism\nthat systematically identifies and resets dormant neurons, which exhibit\nminimal influence in the sparse policy network during inference. This approach\neffectively enhance exploration while preserving structural efficiency.\nExtensive experiments on the CW10-v1 Continual World benchmark demonstrate that\nSSDE achieves state-of-the-art performance, reaching a success rate of 95%,\nsurpassing prior methods significantly in both plasticity and stability\ntrade-offs (code is available at: https://github.com/chengqiArchy/SSDE).\n", "date": "2025-03-07", "categories": []}
{"id": "2503.06175", "title": "Minion Gated Recurrent Unit for Continual Learning", "abstract": "  The increasing demand for continual learning in sequential data processing\nhas led to progressively complex training methodologies and larger recurrent\nnetwork architectures. Consequently, this has widened the knowledge gap between\ncontinual learning with recurrent neural networks (RNNs) and their ability to\noperate on devices with limited memory and compute. To address this challenge,\nwe investigate the effectiveness of simplifying RNN architectures, particularly\ngated recurrent unit (GRU), and its impact on both single-task and multitask\nsequential learning. We propose a new variant of GRU, namely the minion\nrecurrent unit (MiRU). MiRU replaces conventional gating mechanisms with\nscaling coefficients to regulate dynamic updates of hidden states and\nhistorical context, reducing computational costs and memory requirements.\nDespite its simplified architecture, MiRU maintains performance comparable to\nthe standard GRU while achieving 2.90x faster training and reducing parameter\nusage by 2.88x, as demonstrated through evaluations on sequential image\nclassification and natural language processing benchmarks. The impact of model\nsimplification on its learning capacity is also investigated by performing\ncontinual learning tasks with a rehearsal-based strategy and global inhibition.\nWe find that MiRU demonstrates stable performance in multitask learning even\nwhen using only rehearsal, unlike the standard GRU and its variants. These\nfeatures position MiRU as a promising candidate for edge-device applications.\n", "date": "2025-03-08", "categories": []}
{"id": "2503.06213", "title": "Lifelong Learning with Task-Specific Adaptation: Addressing the\n  Stability-Plasticity Dilemma", "abstract": "  Lifelong learning (LL) aims to continuously acquire new knowledge while\nretaining previously learned knowledge. A central challenge in LL is the\nstability-plasticity dilemma, which requires models to balance the preservation\nof previous knowledge (stability) with the ability to learn new tasks\n(plasticity). While parameter-efficient fine-tuning (PEFT) has been widely\nadopted in large language models, its application to lifelong learning remains\nunderexplored. To bridge this gap, this paper proposes AdaLL, an adapter-based\nframework designed to address the dilemma through a simple, universal, and\neffective strategy. AdaLL co-trains the backbone network and adapters under\nregularization constraints, enabling the backbone to capture task-invariant\nfeatures while allowing the adapters to specialize in task-specific\ninformation. Unlike methods that freeze the backbone network, AdaLL\nincrementally enhances the backbone's capabilities across tasks while\nminimizing interference through backbone regularization. This architectural\ndesign significantly improves both stability and plasticity, effectively\neliminating the stability-plasticity dilemma. Extensive experiments demonstrate\nthat AdaLL consistently outperforms existing methods across various\nconfigurations, including dataset choices, task sequences, and task scales.\n", "date": "2025-03-08", "categories": []}
{"id": "2503.07107", "title": "Towards Experience Replay for Class-Incremental Learning in Fully-Binary\n  Networks", "abstract": "  Binary Neural Networks (BNNs) are a promising approach to enable Artificial\nNeural Network (ANN) implementation on ultra-low power edge devices. Such\ndevices may compute data in highly dynamic environments, in which the classes\ntargeted for inference can evolve or even novel classes may arise, requiring\ncontinual learning. Class Incremental Learning (CIL) is a common type of\ncontinual learning for classification problems, that has been scarcely\naddressed in the context of BNNs. Furthermore, most of existing BNNs models are\nnot fully binary, as they require several real-valued network layers, at the\ninput, the output, and for batch normalization. This paper goes a step further,\nenabling class incremental learning in Fully-Binarized NNs (FBNNs) through four\nmain contributions. We firstly revisit the FBNN design and its training\nprocedure that is suitable to CIL. Secondly, we explore loss balancing, a\nmethod to trade-off the performance of past and current classes. Thirdly, we\npropose a semi-supervised method to pre-train the feature extractor of the FBNN\nfor transferable representations. Fourthly, two conventional CIL methods, \\ie,\nLatent and Native replay, are thoroughly compared. These contributions are\nexemplified first on the CIFAR100 dataset, before being scaled up to address\nthe CORE50 continual learning benchmark. The final results based on our 3Mb\nFBNN on CORE50 exhibit at par and better performance than conventional\nreal-valued larger NN models.\n", "date": "2025-03-10", "categories": []}
{"id": "2503.07153", "title": "PTMs-TSCIL Pre-Trained Models Based Class-Incremental Learning", "abstract": "  Class-incremental learning (CIL) for time series data faces critical\nchallenges in balancing stability against catastrophic forgetting and\nplasticity for new knowledge acquisition, particularly under real-world\nconstraints where historical data access is restricted. While pre-trained\nmodels (PTMs) have shown promise in CIL for vision and NLP domains, their\npotential in time series class-incremental learning (TSCIL) remains\nunderexplored due to the scarcity of large-scale time series pre-trained\nmodels. Prompted by the recent emergence of large-scale pre-trained models\n(PTMs) for time series data, we present the first exploration of PTM-based Time\nSeries Class-Incremental Learning (TSCIL). Our approach leverages frozen PTM\nbackbones coupled with incrementally tuning the shared adapter, preserving\ngeneralization capabilities while mitigating feature drift through knowledge\ndistillation. Furthermore, we introduce a Feature Drift Compensation Network\n(DCN), designed with a novel two-stage training strategy to precisely model\nfeature space transformations across incremental tasks. This allows for\naccurate projection of old class prototypes into the new feature space. By\nemploying DCN-corrected prototypes, we effectively enhance the unified\nclassifier retraining, mitigating model feature drift and alleviating\ncatastrophic forgetting. Extensive experiments on five real-world datasets\ndemonstrate state-of-the-art performance, with our method yielding final\naccuracy gains of 1.4%-6.1% across all datasets compared to existing PTM-based\napproaches. Our work establishes a new paradigm for TSCIL, providing insights\ninto stability-plasticity optimization for continual learning systems.\n", "date": "2025-03-10", "categories": []}
{"id": "2410.00064", "title": "M2Distill: Multi-Modal Distillation for Lifelong Imitation Learning", "abstract": "  Lifelong imitation learning for manipulation tasks poses significant\nchallenges due to distribution shifts that occur in incremental learning steps.\nExisting methods often focus on unsupervised skill discovery to construct an\never-growing skill library or distillation from multiple policies, which can\nlead to scalability issues as diverse manipulation tasks are continually\nintroduced and may fail to ensure a consistent latent space throughout the\nlearning process, leading to catastrophic forgetting of previously learned\nskills. In this paper, we introduce M2Distill, a multi-modal distillation-based\nmethod for lifelong imitation learning focusing on preserving consistent latent\nspace across vision, language, and action distributions throughout the learning\nprocess. By regulating the shifts in latent representations across different\nmodalities from previous to current steps, and reducing discrepancies in\nGaussian Mixture Model (GMM) policies between consecutive learning steps, we\nensure that the learned policy retains its ability to perform previously\nlearned tasks while seamlessly integrating new skills. Extensive evaluations on\nthe LIBERO lifelong imitation learning benchmark suites, including\nLIBERO-OBJECT, LIBERO-GOAL, and LIBERO-SPATIAL, demonstrate that our method\nconsistently outperforms prior state-of-the-art methods across all evaluated\nmetrics.\n", "date": "2024-09-30", "categories": []}
{"id": "2501.13198", "title": "SD-LoRA: Scalable Decoupled Low-Rank Adaptation for Class Incremental\n  Learning", "abstract": "  Continual Learning (CL) with foundation models has recently emerged as a\npromising paradigm to exploit abundant knowledge acquired during pre-training\nfor tackling sequential tasks. However, existing prompt-based and Low-Rank\nAdaptation-based (LoRA-based) methods often require expanding a prompt/LoRA\npool or retaining samples of previous tasks, which poses significant\nscalability challenges as the number of tasks grows. To address these\nlimitations, we propose Scalable Decoupled LoRA (SD-LoRA) for class incremental\nlearning, which continually separates the learning of the magnitude and\ndirection of LoRA components without rehearsal. Our empirical and theoretical\nanalysis reveals that SD-LoRA tends to follow a low-loss trajectory and\nconverges to an overlapping low-loss region for all learned tasks, resulting in\nan excellent stability-plasticity trade-off. Building upon these insights, we\nintroduce two variants of SD-LoRA with further improved parameter efficiency.\nAll parameters of SD-LoRAs can be end-to-end optimized for CL objectives.\nMeanwhile, they support efficient inference by allowing direct evaluation with\nthe finally trained model, obviating the need for component selection.\nExtensive experiments across multiple CL benchmarks and foundation models\nconsistently validate the effectiveness of SD-LoRA. The code is available at\nhttps://github.com/WuYichen-97/SD-Lora-CL.\n", "date": "2025-01-22", "categories": []}
{"id": "2409.08388", "title": "Continual Learning in 3D Point Clouds: Employing Spectral Techniques for\n  Exemplar Selection", "abstract": "  We introduce a novel framework for Continual Learning in 3D object\nclassification. Our approach, CL3D, is based on the selection of prototypes\nfrom each class using spectral clustering. For non-Euclidean data such as point\nclouds, spectral clustering can be employed as long as one can define a\ndistance measure between pairs of samples. Choosing the appropriate distance\nmeasure enables us to leverage 3D geometric characteristics to identify\nrepresentative prototypes for each class. We explore the effectiveness of\nclustering in the input space (3D points), local feature space\n(1024-dimensional points), and global feature space. We conduct experiments on\nthe ModelNet40, ShapeNet, and ScanNet datasets, achieving state-of-the-art\naccuracy exclusively through the use of input space features. By leveraging the\ncombined input, local, and global features, we have improved the\nstate-of-the-art on ModelNet and ShapeNet, utilizing nearly half the memory\nused by competing approaches. For the challenging ScanNet dataset, our method\nenhances accuracy by 4.1% while consuming just 28% of the memory used by our\ncompetitors, demonstrating the scalability of our approach.\n", "date": "2024-09-12", "categories": []}
{"id": "2503.04229", "title": "Synthetic Data is an Elegant GIFT for Continual Vision-Language Models", "abstract": "  Pre-trained Vision-Language Models (VLMs) require Continual Learning (CL) to\nefficiently update their knowledge and adapt to various downstream tasks\nwithout retraining from scratch. However, for VLMs, in addition to the loss of\nknowledge previously learned from downstream tasks, pre-training knowledge is\nalso corrupted during continual fine-tuning. This issue is exacerbated by the\nunavailability of original pre-training data, leaving VLM's generalization\nability degrading. In this paper, we propose GIFT, a novel continual\nfine-tuning approach that utilizes synthetic data to overcome catastrophic\nforgetting in VLMs. Taking advantage of recent advances in text-to-image\nsynthesis, we employ a pre-trained diffusion model to recreate both\npre-training and learned downstream task data. In this way, the VLM can revisit\nprevious knowledge through distillation on matching diffusion-generated images\nand corresponding text prompts. Leveraging the broad distribution and high\nalignment between synthetic image-text pairs in VLM's feature space, we propose\na contrastive distillation loss along with an image-text alignment constraint.\nTo further combat in-distribution overfitting and enhance distillation\nperformance with limited amount of generated data, we incorporate adaptive\nweight consolidation, utilizing Fisher information from these synthetic\nimage-text pairs and achieving a better stability-plasticity balance. Extensive\nexperiments demonstrate that our method consistently outperforms previous\nstate-of-the-art approaches across various settings.\n", "date": "2025-03-06", "categories": []}
{"id": "2503.04688", "title": "Teach YOLO to Remember: A Self-Distillation Approach for Continual\n  Object Detection", "abstract": "  Real-time object detectors like YOLO achieve exceptional performance when\ntrained on large datasets for multiple epochs. However, in real-world scenarios\nwhere data arrives incrementally, neural networks suffer from catastrophic\nforgetting, leading to a loss of previously learned knowledge. To address this,\nprior research has explored strategies for Class Incremental Learning (CIL) in\nContinual Learning for Object Detection (CLOD), with most approaches focusing\non two-stage object detectors. However, existing work suggests that Learning\nwithout Forgetting (LwF) may be ineffective for one-stage anchor-free detectors\nlike YOLO due to noisy regression outputs, which risk transferring corrupted\nknowledge. In this work, we introduce YOLO LwF, a self-distillation approach\ntailored for YOLO-based continual object detection. We demonstrate that when\ncoupled with a replay memory, YOLO LwF significantly mitigates forgetting.\nCompared to previous approaches, it achieves state-of-the-art performance,\nimproving mAP by +2.1% and +2.9% on the VOC and COCO benchmarks, respectively.\n", "date": "2025-03-06", "categories": []}
{"id": "2408.09838", "title": "Mitigating the Stability-Plasticity Dilemma in Adaptive Train Scheduling\n  with Curriculum-Driven Continual DQN Expansion", "abstract": "  A continual learning agent builds on previous experiences to develop\nincreasingly complex behaviors by adapting to non-stationary and dynamic\nenvironments while preserving previously acquired knowledge. However, scaling\nthese systems presents significant challenges, particularly in balancing the\npreservation of previous policies with the adaptation of new ones to current\nenvironments. This balance, known as the stability-plasticity dilemma, is\nespecially pronounced in complex multi-agent domains such as the train\nscheduling problem, where environmental and agent behaviors are constantly\nchanging, and the search space is vast. In this work, we propose addressing\nthese challenges in the train scheduling problem using curriculum learning. We\ndesign a curriculum with adjacent skills that build on each other to improve\ngeneralization performance. Introducing a curriculum with distinct tasks\nintroduces non-stationarity, which we address by proposing a new algorithm:\nContinual Deep Q-Network (DQN) Expansion (CDE). Our approach dynamically\ngenerates and adjusts Q-function subspaces to handle environmental changes and\ntask requirements. CDE mitigates catastrophic forgetting through EWC while\nensuring high plasticity using adaptive rational activation functions.\nExperimental results demonstrate significant improvements in learning\nefficiency and adaptability compared to RL baselines and other adapted methods\nfor continual learning, highlighting the potential of our method in managing\nthe stability-plasticity dilemma in the adaptive train scheduling setting.\n", "date": "2024-08-19", "categories": []}
{"id": "2409.15729", "title": "Sequential Learning in the Dense Associative Memory", "abstract": "  Sequential learning involves learning tasks in a sequence, and proves\nchallenging for most neural networks. Biological neural networks regularly\nconquer the sequential learning challenge and are even capable of transferring\nknowledge both forward and backwards between tasks. Artificial neural networks\noften totally fail to transfer performance between tasks, and regularly suffer\nfrom degraded performance or catastrophic forgetting on previous tasks. Models\nof associative memory have been used to investigate the discrepancy between\nbiological and artificial neural networks due to their biological ties and\ninspirations, of which the Hopfield network is the most studied model. The\nDense Associative Memory (DAM), or modern Hopfield network, generalizes the\nHopfield network, allowing for greater capacities and prototype learning\nbehaviors, while still retaining the associative memory structure. We give a\nsubstantial review of the sequential learning space with particular respect to\nthe Hopfield network and associative memories. We perform foundational\nbenchmarks of sequential learning in the DAM using various sequential learning\ntechniques, and analyze the results of the sequential learning to demonstrate\npreviously unseen transitions in the behavior of the DAM. This paper also\ndiscusses the departure from biological plausibility that may affect the\nutility of the DAM as a tool for studying biological neural networks. We\npresent our findings, including the effectiveness of a range of\nstate-of-the-art sequential learning methods when applied to the DAM, and use\nthese methods to further the understanding of DAM properties and behaviors.\n", "date": "2024-09-24", "categories": []}
{"id": "2412.18355", "title": "Handling Spatial-Temporal Data Heterogeneity for Federated Continual\n  Learning via Tail Anchor", "abstract": "  Federated continual learning (FCL) allows each client to continually update\nits knowledge from task streams, enhancing the applicability of federated\nlearning in real-world scenarios. However, FCL needs to address not only\nspatial data heterogeneity between clients but also temporal data heterogeneity\nbetween tasks. In this paper, empirical experiments demonstrate that such\ninput-level heterogeneity significantly affects the model's internal parameters\nand outputs, leading to severe spatial-temporal catastrophic forgetting of\nlocal and previous knowledge. To this end, we propose Federated Tail Anchor\n(FedTA) to mix trainable Tail Anchor with the frozen output features to adjust\ntheir position in the feature space, thereby overcoming parameter-forgetting\nand output-forgetting. Three novel components are also included: Input\nEnhancement for improving the performance of pre-trained models on downstream\ntasks; Selective Input Knowledge Fusion for fusion of heterogeneous local\nknowledge on the server; and Best Global Prototype Selection for finding the\nbest anchor point for each class in the feature space. Extensive experiments\ndemonstrate that FedTA not only outperforms existing FCL methods but also\neffectively preserves the relative positions of features.\n", "date": "2024-12-24", "categories": []}
{"id": "2311.14028", "title": "Continual Learning of Diffusion Models with Generative Distillation", "abstract": "  Diffusion models are powerful generative models that achieve state-of-the-art\nperformance in image synthesis. However, training them demands substantial\namounts of data and computational resources. Continual learning would allow for\nincrementally learning new tasks and accumulating knowledge, thus enabling the\nreuse of trained models for further learning. One potentially suitable\ncontinual learning approach is generative replay, where a copy of a generative\nmodel trained on previous tasks produces synthetic data that are interleaved\nwith data from the current task. However, standard generative replay applied to\ndiffusion models results in a catastrophic loss in denoising capabilities. In\nthis paper, we propose generative distillation, an approach that distils the\nentire reverse process of a diffusion model. We demonstrate that our approach\nsubstantially improves the continual learning performance of generative replay\nwith only a modest increase in the computational costs.\n", "date": "2023-11-23", "categories": []}
{"id": "2312.16731", "title": "Infinite dSprites for Disentangled Continual Learning: Separating Memory\n  Edits from Generalization", "abstract": "  The ability of machine learning systems to learn continually is hindered by\ncatastrophic forgetting, the tendency of neural networks to overwrite\npreviously acquired knowledge when learning a new task. Existing methods\nmitigate this problem through regularization, parameter isolation, or\nrehearsal, but they are typically evaluated on benchmarks comprising only a\nhandful of tasks. In contrast, humans are able to learn over long time horizons\nin dynamic, open-world environments, effortlessly memorizing unfamiliar objects\nand reliably recognizing them under various transformations. To make progress\ntowards closing this gap, we introduce Infinite dSprites, a parsimonious tool\nfor creating continual classification and disentanglement benchmarks of\narbitrary length and with full control over generative factors. We show that\nover a sufficiently long time horizon, the performance of all major types of\ncontinual learning methods deteriorates on this simple benchmark. This result\nhighlights an important and previously overlooked aspect of continual learning:\ngiven a finite modelling capacity and an arbitrarily long learning horizon,\nefficient learning requires memorizing class-specific information and\naccumulating knowledge about general mechanisms. In a simple setting with\ndirect supervision on the generative factors, we show how learning\nclass-agnostic transformations offers a way to circumvent catastrophic\nforgetting and improve classification accuracy over time. Our approach sets the\nstage for continual learning over hundreds of tasks with explicit control over\nmemorization and forgetting, emphasizing open-set classification and one-shot\ngeneralization.\n", "date": "2023-12-27", "categories": []}
{"id": "2409.05650", "title": "Replay Consolidation with Label Propagation for Continual Object\n  Detection", "abstract": "  Continual Learning (CL) aims to learn new data while remembering previously\nacquired knowledge. In contrast to CL for image classification, CL for Object\nDetection faces additional challenges such as the missing annotations problem.\nIn this scenario, images from previous tasks may contain instances of unknown\nclasses that could reappear as labeled in future tasks, leading to task\ninterference in replay-based approaches. Consequently, most approaches in the\nliterature have focused on distillation-based techniques, which are effective\nwhen there is a significant class overlap between tasks. In our work, we\npropose an alternative to distillation-based approaches with a novel approach\ncalled Replay Consolidation with Label Propagation for Object Detection\n(RCLPOD). RCLPOD enhances the replay memory by improving the quality of the\nstored samples through a technique that promotes class balance while also\nimproving the quality of the ground truth associated with these samples through\na technique called label propagation. RCLPOD outperforms existing techniques on\nwell-established benchmarks such as VOC and COC. Moreover, our approach is\ndeveloped to work with modern architectures like YOLOv8, making it suitable for\ndynamic, real-world applications such as autonomous driving and robotics, where\ncontinuous learning and resource efficiency are essential.\n", "date": "2024-09-09", "categories": []}
{"id": "2410.00911", "title": "Dual Consolidation for Pre-Trained Model-Based Domain-Incremental\n  Learning", "abstract": "  Domain-Incremental Learning (DIL) involves the progressive adaptation of a\nmodel to new concepts across different domains. While recent advances in\npre-trained models provide a solid foundation for DIL, learning new concepts\noften results in the catastrophic forgetting of pre-trained knowledge.\nSpecifically, sequential model updates can overwrite both the representation\nand the classifier with knowledge from the latest domain. Thus, it is crucial\nto develop a representation and corresponding classifier that accommodate all\nseen domains throughout the learning process. To this end, we propose DUal\nConsolidaTion (Duct) to unify and consolidate historical knowledge at both the\nrepresentation and classifier levels. By merging the backbone of different\nstages, we create a representation space suitable for multiple domains\nincrementally. The merged representation serves as a balanced intermediary that\ncaptures task-specific features from all seen domains. Additionally, to address\nthe mismatch between consolidated embeddings and the classifier, we introduce\nan extra classifier consolidation process. Leveraging class-wise semantic\ninformation, we estimate the classifier weights of old domains within the\nlatest embedding space. By merging historical and estimated classifiers, we\nalign them with the consolidated embedding space, facilitating incremental\nclassification. Extensive experimental results on four benchmark datasets\ndemonstrate Duct's state-of-the-art performance. Code is available at\nhttps://github.com/Estrella-fugaz/CVPR25-Duct\n", "date": "2024-10-01", "categories": []}
{"id": "2503.01887", "title": "When Continue Learning Meets Multimodal Large Language Model: A Survey", "abstract": "  Recent advancements in Artificial Intelligence have led to the development of\nMultimodal Large Language Models (MLLMs). However, adapting these pre-trained\nmodels to dynamic data distributions and various tasks efficiently remains a\nchallenge. Fine-tuning MLLMs for specific tasks often causes performance\ndegradation in the model's prior knowledge domain, a problem known as\n'Catastrophic Forgetting'. While this issue has been well-studied in the\nContinual Learning (CL) community, it presents new challenges for MLLMs. This\nreview paper, the first of its kind in MLLM continual learning, presents an\noverview and analysis of 440 research papers in this area.The review is\nstructured into four sections. First, it discusses the latest research on\nMLLMs, covering model innovations, benchmarks, and applications in various\nfields. Second, it categorizes and overviews the latest studies on continual\nlearning, divided into three parts: non-large language models unimodal\ncontinual learning (Non-LLM Unimodal CL), non-large language models multimodal\ncontinual learning (Non-LLM Multimodal CL), and continual learning in large\nlanguage models (CL in LLM). The third section provides a detailed analysis of\nthe current state of MLLM continual learning research, including benchmark\nevaluations, architectural innovations, and a summary of theoretical and\nempirical studies.Finally, the paper discusses the challenges and future\ndirections of continual learning in MLLMs, aiming to inspire future research\nand development in the field. This review connects the foundational concepts,\ntheoretical insights, method innovations, and practical applications of\ncontinual learning for multimodal large models, providing a comprehensive\nunderstanding of the research progress and challenges in this field, aiming to\ninspire researchers in the field and promote the advancement of related\ntechnologies.\n", "date": "2025-02-27", "categories": []}
{"id": "2503.02117", "title": "Parabolic Continual Learning", "abstract": "  Regularizing continual learning techniques is important for anticipating\nalgorithmic behavior under new realizations of data. We introduce a new\napproach to continual learning by imposing the properties of a parabolic\npartial differential equation (PDE) to regularize the expected behavior of the\nloss over time. This class of parabolic PDEs has a number of favorable\nproperties that allow us to analyze the error incurred through forgetting and\nthe error induced through generalization. Specifically, we do this through\nimposing boundary conditions where the boundary is given by a memory buffer. By\nusing the memory buffer as a boundary, we can enforce long term dependencies by\nbounding the expected error by the boundary loss. Finally, we illustrate the\nempirical performance of the method on a series of continual learning tasks.\n", "date": "2025-03-03", "categories": []}
{"id": "2503.02526", "title": "A Theory of Initialisation's Impact on Specialisation", "abstract": "  Prior work has demonstrated a consistent tendency in neural networks engaged\nin continual learning tasks, wherein intermediate task similarity results in\nthe highest levels of catastrophic interference. This phenomenon is attributed\nto the network's tendency to reuse learned features across tasks. However, this\nexplanation heavily relies on the premise that neuron specialisation occurs,\ni.e. the emergence of localised representations. Our investigation challenges\nthe validity of this assumption. Using theoretical frameworks for the analysis\nof neural networks, we show a strong dependence of specialisation on the\ninitial condition. More precisely, we show that weight imbalance and high\nweight entropy can favour specialised solutions. We then apply these insights\nin the context of continual learning, first showing the emergence of a\nmonotonic relation between task-similarity and forgetting in non-specialised\nnetworks. {Finally, we show that specialization by weight imbalance is\nbeneficial on the commonly employed elastic weight consolidation regularisation\ntechnique.\n", "date": "2025-03-04", "categories": []}
{"id": "2405.16350", "title": "A Second-Order Perspective on Model Compositionality and Incremental\n  Learning", "abstract": "  The fine-tuning of deep pre-trained models has revealed compositional\nproperties, with multiple specialized modules that can be arbitrarily composed\ninto a single, multi-task model. However, identifying the conditions that\npromote compositionality remains an open issue, with recent efforts\nconcentrating mainly on linearized networks. We conduct a theoretical study\nthat attempts to demystify compositionality in standard non-linear networks\nthrough the second-order Taylor approximation of the loss function. The\nproposed formulation highlights the importance of staying within the\npre-training basin to achieve composable modules. Moreover, it provides the\nbasis for two dual incremental training algorithms: the one from the\nperspective of multiple models trained individually, while the other aims to\noptimize the composed model as a whole. We probe their application in\nincremental classification tasks and highlight some valuable skills. In fact,\nthe pool of incrementally learned modules not only supports the creation of an\neffective multi-task model but also enables unlearning and specialization in\ncertain tasks. Code available at https://github.com/aimagelab/mammoth.\n", "date": "2024-05-25", "categories": []}
{"id": "2406.07413", "title": "Towards Continuous Reuse of Graph Models via Holistic Memory\n  Diversification", "abstract": "  This paper addresses the challenge of incremental learning in growing graphs\nwith increasingly complex tasks. The goal is to continuously train a graph\nmodel to handle new tasks while retaining proficiency in previous tasks via\nmemory replay. Existing methods usually overlook the importance of memory\ndiversity, limiting in selecting high-quality memory from previous tasks and\nremembering broad previous knowledge within the scarce memory on graphs. To\naddress that, we introduce a novel holistic Diversified Memory Selection and\nGeneration (DMSG) framework for incremental learning in graphs, which first\nintroduces a buffer selection strategy that considers both intra-class and\ninter-class diversities, employing an efficient greedy algorithm for sampling\nrepresentative training nodes from graphs into memory buffers after learning\neach new task. Then, to adequately rememorize the knowledge preserved in the\nmemory buffer when learning new tasks, a diversified memory generation replay\nmethod is introduced. This method utilizes a variational layer to generate the\ndistribution of buffer node embeddings and sample synthesized ones for\nreplaying. Furthermore, an adversarial variational embedding learning method\nand a reconstruction-based decoder are proposed to maintain the integrity and\nconsolidate the generalization of the synthesized node embeddings,\nrespectively. Extensive experimental results on publicly accessible datasets\ndemonstrate the superiority of \\method{} over state-of-the-art methods.\n", "date": "2024-06-11", "categories": []}
{"id": "2409.12326", "title": "ReFu: Recursive Fusion for Exemplar-Free 3D Class-Incremental Learning", "abstract": "  We introduce a novel Recursive Fusion model, dubbed ReFu, designed to\nintegrate point clouds and meshes for exemplar-free 3D Class-Incremental\nLearning, where the model learns new 3D classes while retaining knowledge of\npreviously learned ones. Unlike existing methods that either rely on storing\nhistorical data to mitigate forgetting or focus on single data modalities, ReFu\neliminates the need for exemplar storage while utilizing the complementary\nstrengths of both point clouds and meshes. To achieve this, we introduce a\nrecursive method which continuously accumulates knowledge by updating the\nregularized auto-correlation matrix. Furthermore, we propose a fusion module,\nfeaturing a Pointcloud-guided Mesh Attention Layer that learns correlations\nbetween the two modalities. This mechanism effectively integrates point cloud\nand mesh features, leading to more robust and stable continual learning.\nExperiments across various datasets demonstrate that our proposed framework\noutperforms existing methods in 3D class-incremental learning.\n", "date": "2024-09-18", "categories": []}
{"id": "2410.23751", "title": "EXACFS -- A CIL Method to mitigate Catastrophic Forgetting", "abstract": "  Deep neural networks (DNNS) excel at learning from static datasets but\nstruggle with continual learning, where data arrives sequentially. Catastrophic\nforgetting, the phenomenon of forgetting previously learned knowledge, is a\nprimary challenge. This paper introduces EXponentially Averaged Class-wise\nFeature Significance (EXACFS) to mitigate this issue in the class incremental\nlearning (CIL) setting. By estimating the significance of model features for\neach learned class using loss gradients, gradually aging the significance\nthrough the incremental tasks and preserving the significant features through a\ndistillation loss, EXACFS effectively balances remembering old knowledge\n(stability) and learning new knowledge (plasticity). Extensive experiments on\nCIFAR-100 and ImageNet-100 demonstrate EXACFS's superior performance in\npreserving stability while acquiring plasticity.\n", "date": "2024-10-31", "categories": []}
{"id": "2411.06916", "title": "Slowing Down Forgetting in Continual Learning", "abstract": "  A common challenge in continual learning (CL) is catastrophic forgetting,\nwhere the performance on old tasks drops after new, additional tasks are\nlearned. In this paper, we propose a novel framework called ReCL to slow down\nforgetting in CL. Our framework exploits an implicit bias of gradient-based\nneural networks due to which these converge to margin maximization points. Such\nconvergence points allow us to reconstruct old data from previous tasks, which\nwe then combine with the current training data. Our framework is flexible and\ncan be applied on top of existing, state-of-the-art CL methods. We further\ndemonstrate the performance gain from our framework across a large series of\nexperiments, including two challenging CL scenarios (class incremental and\ndomain incremental learning), different datasets (MNIST, CIFAR10,\nTinyImagenet), and different network architectures. Across all experiments, we\nfind large performance gains through ReCL. To the best of our knowledge, our\nframework is the first to address catastrophic forgetting by leveraging models\nin CL as their own memory buffers.\n", "date": "2024-11-11", "categories": []}
{"id": "2502.12949", "title": "Efficient Learning Under Density Shift in Incremental Settings Using\n  Cram\\'er-Rao-Based Regularization", "abstract": "  The continuous surge in data volume and velocity is often dealt with using\ndata orchestration and distributed processing approaches, abstracting away the\nmachine learning challenges that exist at the algorithmic level. With growing\ninterest in automating the learning loop, training with data that arrive in a\nsequence rather than in the classical in-memory training data form will face a\nmachine learning challenge because of evolving feature distributions across\nbatches of training data biasing the cross-validation step\n(\\cite{sugiyama2012machine}). This work takes a distributed density estimation\nangle to the problem where data are temporally distributed. It processes data\nin batches and allows a neural network to treat a batch as training data. The\nmethod accumulates knowledge about the data density via posterior probability\nabsorption using the Fisher Information Matrix, which contains information\nabout the local optimization gradients for the batch. This is then used as a\nregularizer for the loss in the following batch, and therefore the density\nestimate for the entire dataset constructively gets more robust to the non-iid\ndistribution shift. This needs the presence of a pair of batches in memory at a\ntime, so the space cost is not a function of the size of the complete,\ndistributed dataset. We proposed a novel regularization-based approach\nCovariate Shift Correction $C^{2}A$ that leverages Fisher information and\nKullback-Leibler divergence to adapt to both natural and sequential covariate\nshift caused by dataset fragmentation. $C^{2}A$ achieves $19\\%$ accuracy at\nmaximum against state-of-the-art methods.\n", "date": "2025-02-18", "categories": []}
{"id": "2503.00413", "title": "CL-MoE: Enhancing Multimodal Large Language Model with Dual Momentum\n  Mixture-of-Experts for Continual Visual Question Answering", "abstract": "  Multimodal large language models (MLLMs) have garnered widespread attention\nfrom researchers due to their remarkable understanding and generation\ncapabilities in visual language tasks (e.g., visual question answering).\nHowever, the rapid pace of knowledge updates in the real world makes offline\ntraining of MLLMs costly, and when faced with non-stationary data streams,\nMLLMs suffer from catastrophic forgetting during learning. In this paper, we\npropose an MLLMs-based dual momentum Mixture-of-Experts (CL-MoE) framework for\ncontinual visual question answering (VQA). We integrate MLLMs with continual\nlearning to utilize the rich commonsense knowledge in LLMs. We introduce a\nDual-Router MoE (RMoE) strategy to select the global and local experts using\ntask-level and instance-level routers, to robustly assign weights to the\nexperts most appropriate for the task. Then, we design a dynamic Momentum MoE\n(MMoE) to update the parameters of experts dynamically based on the\nrelationships between the experts and tasks/instances, so that the model can\nabsorb new knowledge while maintaining existing knowledge. The extensive\nexperimental results indicate that our method achieves state-of-the-art\nperformance on 10 VQA tasks, proving the effectiveness of our approach.\n", "date": "2025-03-01", "categories": []}
{"id": "2503.00528", "title": "Efficient Prompting for Continual Adaptation to Missing Modalities", "abstract": "  Missing modality issues are common in real-world applications, arising from\nfactors such as equipment failures and privacy concerns. When fine-tuning\npre-trained models on downstream datasets with missing modalities, performance\ncan degrade significantly. Current methods often aggregate various missing\ncases to train recovery modules or align multimodal features, resulting in\nsuboptimal performance, high computational costs, and the risk of catastrophic\nforgetting in continual environments where data arrives sequentially. In this\npaper, we formulate the dynamic missing modality problem as a continual\nlearning task and introduce the continual multimodal missing modality task. To\naddress this challenge efficiently, we introduce three types of prompts:\nmodality-specific, task-aware, and task-specific prompts. These prompts enable\nthe model to learn intra-modality, inter-modality, intra-task, and inter-task\nfeatures. Furthermore, we propose a contrastive task interaction strategy to\nexplicitly learn prompts correlating different modalities. We conduct extensive\nexperiments on three public datasets, where our method consistently outperforms\nstate-of-the-art approaches.\n", "date": "2025-03-01", "categories": []}
{"id": "2503.00677", "title": "Advancing Prompt-Based Methods for Replay-Independent General Continual\n  Learning", "abstract": "  General continual learning (GCL) is a broad concept to describe real-world\ncontinual learning (CL) problems, which are often characterized by online data\nstreams without distinct transitions between tasks, i.e., blurry task\nboundaries. Such requirements result in poor initial performance, limited\ngeneralizability, and severe catastrophic forgetting, heavily impacting the\neffectiveness of mainstream GCL models trained from scratch. While the use of a\nfrozen pretrained backbone with appropriate prompt tuning can partially address\nthese challenges, such prompt-based methods remain suboptimal for CL of\nremaining tunable parameters on the fly. In this regard, we propose an\ninnovative approach named MISA (Mask and Initial Session Adaption) to advance\nprompt-based methods in GCL. It includes a forgetting-aware initial session\nadaption that employs pretraining data to initialize prompt parameters and\nimprove generalizability, as well as a non-parametric logit mask of the output\nlayers to mitigate catastrophic forgetting. Empirical results demonstrate\nsubstantial performance gains of our approach compared to recent competitors,\nespecially without a replay buffer (e.g., up to 18.39%, 22.06%, and 11.96%\nperformance lead on CIFAR-100, Tiny-ImageNet, and ImageNet-R, respectively).\nMoreover, our approach features the plug-in nature for prompt-based methods,\nindependence of replay, ease of implementation, and avoidance of CL-relevant\nhyperparameters, serving as a strong baseline for GCL research. Our source code\nis publicly available at https://github.com/kangzhiq/MISA\n", "date": "2025-03-02", "categories": []}
{"id": "2503.00823", "title": "Task-Agnostic Guided Feature Expansion for Class-Incremental Learning", "abstract": "  The ability to learn new concepts while preserve the learned knowledge is\ndesirable for learning systems in Class-Incremental Learning (CIL). Recently,\nfeature expansion of the model become a prevalent solution for CIL, where the\nold features are fixed during the training of the new task while new features\nare expanded for the new tasks. However, such task-specific features learned\nfrom the new task may collide with the old features, leading to\nmisclassification between tasks. Therefore, the expanded model is often\nencouraged to capture diverse features from the new task, aiming to avoid such\ncollision. However, the existing solution is largely restricted to the samples\nfrom the current task, because of the poor accessibility to previous samples.\nTo promote the learning and transferring of diverse features across tasks, we\npropose a framework called Task-Agnostic Guided Feature Expansion (TagFex).\nFirstly, it captures task-agnostic features continually with a separate model,\nproviding extra task-agnostic features for subsequent tasks. Secondly, to\nobtain useful features from the task-agnostic model for the current task, it\naggregates the task-agnostic features with the task-specific feature using a\nmerge attention. Then the aggregated feature is transferred back into the\ntask-specific feature for inference, helping the task-specific model capture\ndiverse features. Extensive experiments show the effectiveness and superiority\nof TagFex on various CIL settings. Code is available at\nhttps://github.com/bwnzheng/TagFex_CVPR2025.\n", "date": "2025-03-02", "categories": []}
{"id": "2503.01580", "title": "A Selective Learning Method for Temporal Graph Continual Learning", "abstract": "  Node classification is a key task in temporal graph learning (TGL). Real-life\ntemporal graphs often introduce new node classes over time, but existing TGL\nmethods assume a fixed set of classes. This assumption brings limitations, as\nupdating models with full data is costly, while focusing only on new classes\nresults in forgetting old ones. Graph continual learning (GCL) methods mitigate\nforgetting using old-class subsets but fail to account for their evolution. We\ndefine this novel problem as temporal graph continual learning (TGCL), which\nfocuses on efficiently maintaining up-to-date knowledge of old classes. To\ntackle TGCL, we propose a selective learning framework that substitutes the\nold-class data with its subsets, Learning Towards the Future (LTF). We derive\nan upper bound on the error caused by such replacement and transform it into\nobjectives for selecting and learning subsets that minimize classification\nerror while preserving the distribution of the full old-class data. Experiments\non three real-world datasets validate the effectiveness of LTF on TGCL.\n", "date": "2025-03-03", "categories": []}
{"id": "2503.01595", "title": "STAR: Stability-Inducing Weight Perturbation for Continual Learning", "abstract": "  Humans can naturally learn new and varying tasks in a sequential manner.\nContinual learning is a class of learning algorithms that updates its learned\nmodel as it sees new data (on potentially new tasks) in a sequence. A key\nchallenge in continual learning is that as the model is updated to learn new\ntasks, it becomes susceptible to catastrophic forgetting, where knowledge of\npreviously learned tasks is lost. A popular approach to mitigate forgetting\nduring continual learning is to maintain a small buffer of previously-seen\nsamples and to replay them during training. However, this approach is limited\nby the small buffer size, and while forgetting is reduced, it is still present.\nIn this paper, we propose a novel loss function, STAR, that exploits the\nworst-case parameter perturbation that reduces the KL-divergence of model\npredictions with that of its local parameter neighborhood to promote stability\nand alleviate forgetting. STAR can be combined with almost any existing\nrehearsal-based method as a plug-and-play component. We empirically show that\nSTAR consistently improves the performance of existing methods by up to 15%\nacross varying baselines and achieves superior or competitive accuracy to that\nof state-of-the-art methods aimed at improving rehearsal-based continual\nlearning.\n", "date": "2025-03-03", "categories": []}
{"id": "2502.20596", "title": "Few-Shot, No Problem: Descriptive Continual Relation Extraction", "abstract": "  Few-shot Continual Relation Extraction is a crucial challenge for enabling AI\nsystems to identify and adapt to evolving relationships in dynamic real-world\ndomains. Traditional memory-based approaches often overfit to limited samples,\nfailing to reinforce old knowledge, with the scarcity of data in few-shot\nscenarios further exacerbating these issues by hindering effective data\naugmentation in the latent space. In this paper, we propose a novel\nretrieval-based solution, starting with a large language model to generate\ndescriptions for each relation. From these descriptions, we introduce a\nbi-encoder retrieval training paradigm to enrich both sample and class\nrepresentation learning. Leveraging these enhanced representations, we design a\nretrieval-based prediction method where each sample \"retrieves\" the best\nfitting relation via a reciprocal rank fusion score that integrates both\nrelation description vectors and class prototypes. Extensive experiments on\nmultiple datasets demonstrate that our method significantly advances the\nstate-of-the-art by maintaining robust performance across sequential tasks,\neffectively addressing catastrophic forgetting.\n", "date": "2025-02-27", "categories": []}
{"id": "2502.20677", "title": "FoCTTA: Low-Memory Continual Test-Time Adaptation with Focus", "abstract": "  Continual adaptation to domain shifts at test time (CTTA) is crucial for\nenhancing the intelligence of deep learning enabled IoT applications. However,\nprevailing TTA methods, which typically update all batch normalization (BN)\nlayers, exhibit two memory inefficiencies. First, the reliance on BN layers for\nadaptation necessitates large batch sizes, leading to high memory usage.\nSecond, updating all BN layers requires storing the activations of all BN\nlayers for backpropagation, exacerbating the memory demand. Both factors lead\nto substantial memory costs, making existing solutions impractical for IoT\ndevices. In this paper, we present FoCTTA, a low-memory CTTA strategy. The key\nis to automatically identify and adapt a few drift-sensitive representation\nlayers, rather than blindly update all BN layers. The shift from BN to\nrepresentation layers eliminates the need for large batch sizes. Also, by\nupdating adaptation-critical layers only, FoCTTA avoids storing excessive\nactivations. This focused adaptation approach ensures that FoCTTA is not only\nmemory-efficient but also maintains effective adaptation. Evaluations show that\nFoCTTA improves the adaptation accuracy over the state-of-the-arts by 4.5%,\n4.9%, and 14.8% on CIFAR10-C, CIFAR100-C, and ImageNet-C under the same memory\nconstraints. Across various batch sizes, FoCTTA reduces the memory usage by\n3-fold on average, while improving the accuracy by 8.1%, 3.6%, and 0.2%,\nrespectively, on the three datasets.\n", "date": "2025-02-28", "categories": []}
{"id": "2502.21147", "title": "Same accuracy, twice as fast: continuous training surpasses retraining\n  from scratch", "abstract": "  Continual learning aims to enable models to adapt to new datasets without\nlosing performance on previously learned data, often assuming that prior data\nis no longer available. However, in many practical scenarios, both old and new\ndata are accessible. In such cases, good performance on both datasets is\ntypically achieved by abandoning the model trained on the previous data and\nre-training a new model from scratch on both datasets. This training from\nscratch is computationally expensive. In contrast, methods that leverage the\npreviously trained model and old data are worthy of investigation, as they\ncould significantly reduce computational costs. Our evaluation framework\nquantifies the computational savings of such methods while maintaining or\nexceeding the performance of training from scratch. We identify key\noptimization aspects -- initialization, regularization, data selection, and\nhyper-parameters -- that can each contribute to reducing computational costs.\nFor each aspect, we propose effective first-step methods that already yield\nsubstantial computational savings. By combining these methods, we achieve up to\n2.7x reductions in computation time across various computer vision tasks,\nhighlighting the potential for further advancements in this area.\n", "date": "2025-02-28", "categories": []}
{"id": "2404.03200", "title": "Future-Proofing Class-Incremental Learning", "abstract": "  Exemplar-Free Class Incremental Learning is a highly challenging setting\nwhere replay memory is unavailable. Methods relying on frozen feature\nextractors have drawn attention recently in this setting due to their\nimpressive performances and lower computational costs. However, those methods\nare highly dependent on the data used to train the feature extractor and may\nstruggle when an insufficient amount of classes are available during the first\nincremental step. To overcome this limitation, we propose to use a pre-trained\ntext-to-image diffusion model in order to generate synthetic images of future\nclasses and use them to train the feature extractor. Experiments on the\nstandard benchmarks CIFAR100 and ImageNet-Subset demonstrate that our proposed\nmethod can be used to improve state-of-the-art methods for exemplar-free class\nincremental learning, especially in the most difficult settings where the first\nincremental step only contains few classes. Moreover, we show that using\nsynthetic samples of future classes achieves higher performance than using real\ndata from different classes, paving the way for better and less costly\npre-training methods for incremental learning.\n", "date": "2024-04-04", "categories": []}
{"id": "2408.17394", "title": "Continual learning with the neural tangent ensemble", "abstract": "  A natural strategy for continual learning is to weigh a Bayesian ensemble of\nfixed functions. This suggests that if a (single) neural network could be\ninterpreted as an ensemble, one could design effective algorithms that learn\nwithout forgetting. To realize this possibility, we observe that a neural\nnetwork classifier with N parameters can be interpreted as a weighted ensemble\nof N classifiers, and that in the lazy regime limit these classifiers are fixed\nthroughout learning. We call these classifiers the neural tangent experts and\nshow they output valid probability distributions over the labels. We then\nderive the likelihood and posterior probability of each expert given past data.\nSurprisingly, the posterior updates for these experts are equivalent to a\nscaled and projected form of stochastic gradient descent (SGD) over the network\nweights. Away from the lazy regime, networks can be seen as ensembles of\nadaptive experts which improve over time. These results offer a new\ninterpretation of neural networks as Bayesian ensembles of experts, providing a\nprincipled framework for understanding and mitigating catastrophic forgetting\nin continual learning settings.\n", "date": "2024-08-30", "categories": []}
{"id": "2411.07175", "title": "Continual Memorization of Factoids in Language Models", "abstract": "  As new knowledge rapidly accumulates, language models (LMs) with pretrained\nknowledge quickly become obsolete. A common approach to updating LMs is\nfine-tuning them directly on new knowledge. However, recent studies have shown\nthat fine-tuning for memorization may be ineffective in storing knowledge or\nmay exacerbate hallucinations. In this work, we introduce a setting we call\ncontinual memorization, where a model must memorize and retain a set of\nfactoids through multiple stages of fine-tuning on subsequent datasets. We\ncharacterized the forgetting patterns through extensive experiments and show\nthat LMs widely suffer from forgetting, especially when needing to memorize\nfactoids in the second stage. We posit that forgetting can be alleviated by\nmodifying training dynamics: (1) protecting the memorization process when\nlearning factoids or (2) reducing interference from subsequent training stages.\nIntriguingly, we find that mixing randomly generated word sequences or generic\ndata sampled from pretraining corpora at different training stages effectively\nmitigates forgetting REMIX: Random and Generic Data Mixing). REMIX can recover\nperformance from severe forgetting, outperforming replay methods and other\ncontinual learning baselines. We analyze how REMIX influences the learning\nprocess and find that robust memorization follows a distinct pattern: the model\nstores factoids in earlier layers than usual and diversifies the layers that\nretain them, which results in easier recall and manipulate of the learned\nfactoids.\n", "date": "2024-11-11", "categories": []}
{"id": "2502.19922", "title": "Incremental Learning with Repetition via Pseudo-Feature Projection", "abstract": "  Incremental Learning scenarios do not always represent real-world inference\nuse-cases, which tend to have less strict task boundaries, and exhibit\nrepetition of common classes and concepts in their continual data stream. To\nbetter represent these use-cases, new scenarios with partial repetition and\nmixing of tasks are proposed, where the repetition patterns are innate to the\nscenario and unknown to the strategy. We investigate how exemplar-free\nincremental learning strategies are affected by data repetition, and we adapt a\nseries of state-of-the-art approaches to analyse and fairly compare them under\nboth settings. Further, we also propose a novel method (Horde), able to\ndynamically adjust an ensemble of self-reliant feature extractors, and align\nthem by exploiting class repetition. Our proposed exemplar-free method achieves\ncompetitive results in the classic scenario without repetition, and\nstate-of-the-art performance in the one with repetition.\n", "date": "2025-02-27", "categories": []}
{"id": "2502.20124", "title": "Exploring Open-world Continual Learning with Knowns-Unknowns Knowledge\n  Transfer", "abstract": "  Open-World Continual Learning (OWCL) is a challenging paradigm where models\nmust incrementally learn new knowledge without forgetting while operating under\nan open-world assumption. This requires handling incomplete training data and\nrecognizing unknown samples during inference. However, existing OWCL methods\noften treat open detection and continual learning as separate tasks, limiting\ntheir ability to integrate open-set detection and incremental classification in\nOWCL. Moreover, current approaches primarily focus on transferring knowledge\nfrom known samples, neglecting the insights derived from unknown/open samples.\nTo address these limitations, we formalize four distinct OWCL scenarios and\nconduct comprehensive empirical experiments to explore potential challenges in\nOWCL. Our findings reveal a significant interplay between the open detection of\nunknowns and incremental classification of knowns, challenging a widely held\nassumption that unknown detection and known classification are orthogonal\nprocesses. Building on our insights, we propose \\textbf{HoliTrans} (Holistic\nKnowns-Unknowns Knowledge Transfer), a novel OWCL framework that integrates\nnonlinear random projection (NRP) to create a more linearly separable embedding\nspace and distribution-aware prototypes (DAPs) to construct an adaptive\nknowledge space. Particularly, our HoliTrans effectively supports knowledge\ntransfer for both known and unknown samples while dynamically updating\nrepresentations of open samples during OWCL. Extensive experiments across\nvarious OWCL scenarios demonstrate that HoliTrans outperforms 22 competitive\nbaselines, bridging the gap between OWCL theory and practice and providing a\nrobust, scalable framework for advancing open-world learning paradigms.\n", "date": "2025-02-27", "categories": []}
{"id": "2502.12388", "title": "Achieving Upper Bound Accuracy of Joint Training in Continual Learning", "abstract": "  Continual learning has been an active research area in machine learning,\nfocusing on incrementally learning a sequence of tasks. A key challenge is\ncatastrophic forgetting (CF), and most research efforts have been directed\ntoward mitigating this issue. However, a significant gap remains between the\naccuracy achieved by state-of-the-art continual learning algorithms and the\nideal or upper-bound accuracy achieved by training all tasks together jointly.\nThis gap has hindered or even prevented the adoption of continual learning in\napplications, as accuracy is often of paramount importance. Recently, another\nchallenge, termed inter-task class separation (ICS), was also identified, which\nspurred a theoretical study into principled approaches for solving continual\nlearning. Further research has shown that by leveraging the theory and the\npower of large foundation models, it is now possible to achieve upper-bound\naccuracy, which has been empirically validated using both text and image\nclassification datasets. Continual learning is now ready for real-life\napplications. This paper surveys the main research leading to this achievement,\njustifies the approach both intuitively and from neuroscience research, and\ndiscusses insights gained.\n", "date": "2025-02-17", "categories": []}
{"id": "2502.18923", "title": "Brain-inspired analogical mixture prototypes for few-shot\n  class-incremental learning", "abstract": "  Few-shot class-incremental learning (FSCIL) poses significant challenges for\nartificial neural networks due to the need to efficiently learn from limited\ndata while retaining knowledge of previously learned tasks. Inspired by the\nbrain's mechanisms for categorization and analogical learning, we propose a\nnovel approach called Brain-inspired Analogical Mixture Prototypes (BAMP). BAMP\nhas three components: mixed prototypical feature learning, statistical analogy,\nand soft voting. Starting from a pre-trained Vision Transformer (ViT), mixed\nprototypical feature learning represents each class using a mixture of\nprototypes and fine-tunes these representations during the base session. The\nstatistical analogy calibrates the mean and covariance matrix of prototypes for\nnew classes according to similarity to the base classes, and computes\nclassification score with Mahalanobis distance. Soft voting combines both\nmerits of statistical analogy and an off-shelf FSCIL method. Our experiments on\nbenchmark datasets demonstrate that BAMP outperforms state-of-the-art on both\ntraditional big start FSCIL setting and challenging small start FSCIL setting.\nThe study suggests that brain-inspired analogical mixture prototypes can\nalleviate catastrophic forgetting and over-fitting problems in FSCIL.\n", "date": "2025-02-26", "categories": []}
{"id": "2407.00717", "title": "Learning System Dynamics without Forgetting", "abstract": "  Observation-based trajectory prediction for systems with unknown dynamics is\nessential in fields such as physics and biology. Most existing approaches are\nlimited to learning within a single system with fixed dynamics patterns.\nHowever, many real-world applications require learning across systems with\nevolving dynamics patterns, a challenge that has been largely overlooked. To\naddress this, we systematically investigate the problem of Continual Dynamics\nLearning (CDL), examining task configurations and evaluating the applicability\nof existing techniques, while identifying key challenges. In response, we\npropose the Mode-switching Graph ODE (MS-GODE) model, which integrates the\nstrengths LG-ODE and sub-network learning with a mode-switching module,\nenabling efficient learning over varying dynamics. Moreover, we construct a\nnovel benchmark of biological dynamic systems for CDL, Bio-CDL, featuring\ndiverse systems with disparate dynamics and significantly enriching the\nresearch field of machine learning for dynamic systems. Our code available at\nhttps://github.com/QueuQ/MS-GODE.\n", "date": "2024-06-30", "categories": []}
{"id": "2502.16648", "title": "Few-shot Continual Relation Extraction via Open Information Extraction", "abstract": "  Typically, Few-shot Continual Relation Extraction (FCRE) models must balance\nretaining prior knowledge while adapting to new tasks with extremely limited\ndata. However, real-world scenarios may also involve unseen or undetermined\nrelations that existing methods still struggle to handle. To address these\nchallenges, we propose a novel approach that leverages the Open Information\nExtraction concept of Knowledge Graph Construction (KGC). Our method not only\nexposes models to all possible pairs of relations, including determined and\nundetermined labels not available in the training set, but also enriches model\nknowledge with diverse relation descriptions, thereby enhancing knowledge\nretention and adaptability in this challenging scenario. In the perspective of\nKGC, this is the first work explored in the setting of Continual Learning,\nallowing efficient expansion of the graph as the data evolves. Experimental\nresults demonstrate our superior performance compared to other state-of-the-art\nFCRE baselines, as well as the efficiency in handling dynamic graph\nconstruction in this setting.\n", "date": "2025-02-23", "categories": []}
{"id": "2502.17788", "title": "On-device edge learning for IoT data streams: a survey", "abstract": "  This literature review explores continual learning methods for on-device\ntraining in the context of neural networks (NNs) and decision trees (DTs) for\nclassification tasks on smart environments. We highlight key constraints, such\nas data architecture (batch vs. stream) and network capacity (cloud vs. edge),\nwhich impact TinyML algorithm design, due to the uncontrolled natural arrival\nof data streams. The survey details the challenges of deploying deep learners\non resource-constrained edge devices, including catastrophic forgetting, data\ninefficiency, and the difficulty of handling IoT tabular data in open-world\nsettings. While decision trees are more memory-efficient for on-device\ntraining, they are limited in expressiveness, requiring dynamic adaptations,\nlike pruning and meta-learning, to handle complex patterns and concept drifts.\nWe emphasize the importance of multi-criteria performance evaluation tailored\nto edge applications, which assess both output-based and internal\nrepresentation metrics. The key challenge lies in integrating these building\nblocks into autonomous online systems, taking into account stability-plasticity\ntrade-offs, forward-backward transfer, and model convergence.\n", "date": "2025-02-25", "categories": []}
{"id": "2502.17794", "title": "PVBF: A Framework for Mitigating Parameter Variation Imbalance in Online\n  Continual Learning", "abstract": "  Online continual learning (OCL), which enables AI systems to adaptively learn\nfrom non-stationary data streams, is commonly achieved using experience replay\n(ER)-based methods that retain knowledge by replaying stored past during\ntraining. However, these methods face challenges of prediction bias, stemming\nfrom deviations in parameter update directions during task transitions. This\npaper identifies parameter variation imbalance as a critical factor\ncontributing to prediction bias in ER-based OCL. Specifically, using the\nproposed parameter variation evaluation method, we highlight two types of\nimbalance: correlation-induced imbalance, where certain parameters are\ndisproportionately updated across tasks, and layer-wise imbalance, where output\nlayer parameters update faster than those in preceding layers. To mitigate the\nabove imbalances, we propose the Parameter Variation Balancing Framework\n(PVBF), which incorporates: 1) a novel method to compute parameter correlations\nwith previous tasks based on parameter variations, 2) an\nencourage-and-consolidate (E&C) method utilizing parameter correlations to\nperform gradient adjustments across all parameters during training, 3) a\ndual-layer copy weights with reinit (D-CWR) strategy to slowly update output\nlayer parameters for frequently occuring sample categories. Experiments on\nshort and long task sequences demonstrate that PVBF significantly reduces\nprediction bias and improves OCL performance, achieving up to 47\\% higher\naccuracy compared to existing ER-based methods.\n", "date": "2025-02-25", "categories": []}
{"id": "2502.17920", "title": "C-LoRA: Continual Low-Rank Adaptation for Pre-trained Models", "abstract": "  Low-Rank Adaptation (LoRA) is an efficient fine-tuning method that has been\nextensively applied in areas such as natural language processing and computer\nvision. Existing LoRA fine-tuning approaches excel in static environments but\nstruggle in dynamic learning due to reliance on multiple adapter modules,\nincreasing overhead and complicating inference. We propose Continual Low-Rank\nAdaptation (C-LoRA), a novel extension of LoRA for continual learning. C-LoRA\nuses a learnable routing matrix to dynamically manage parameter updates across\ntasks, ensuring efficient reuse of learned subspaces while enforcing\northogonality to minimize interference and forgetting. Unlike existing\napproaches that require separate adapters for each task, C-LoRA enables a\nintegrated approach for task adaptation, achieving both scalability and\nparameter efficiency in sequential learning scenarios. C-LoRA achieves\nstate-of-the-art accuracy and parameter efficiency on benchmarks while\nproviding theoretical insights into its routing matrix's role in retaining and\ntransferring knowledge, establishing a scalable framework for continual\nlearning.\n", "date": "2025-02-25", "categories": []}
{"id": "2403.16707", "title": "One-Shot Domain Incremental Learning", "abstract": "  Domain incremental learning (DIL) has been discussed in previous studies on\ndeep neural network models for classification. In DIL, we assume that samples\non new domains are observed over time. The models must classify inputs on all\ndomains. In practice, however, we may encounter a situation where we need to\nperform DIL under the constraint that the samples on the new domain are\nobserved only infrequently. Therefore, in this study, we consider the extreme\ncase where we have only one sample from the new domain, which we call one-shot\nDIL. We first empirically show that existing DIL methods do not work well in\none-shot DIL. We have analyzed the reason for this failure through various\ninvestigations. According to our analysis, we clarify that the difficulty of\none-shot DIL is caused by the statistics in the batch normalization layers.\nTherefore, we propose a technique regarding these statistics and demonstrate\nthe effectiveness of our technique through experiments on open datasets. The\ncode is available at https://github.com/ToyotaCRDL/OneShotDIL.\n", "date": "2024-03-25", "categories": []}
{"id": "2502.15922", "title": "On the Design of Safe Continual RL Methods for Control of Nonlinear\n  Systems", "abstract": "  Reinforcement learning (RL) algorithms have been successfully applied to\ncontrol tasks associated with unmanned aerial vehicles and robotics. In recent\nyears, safe RL has been proposed to allow the safe execution of RL algorithms\nin industrial and mission-critical systems that operate in closed loops.\nHowever, if the system operating conditions change, such as when an unknown\nfault occurs in the system, typical safe RL algorithms are unable to adapt\nwhile retaining past knowledge. Continual reinforcement learning algorithms\nhave been proposed to address this issue. However, the impact of continual\nadaptation on the system's safety is an understudied problem. In this paper, we\nstudy the intersection of safe and continual RL. First, we empirically\ndemonstrate that a popular continual RL algorithm, online elastic weight\nconsolidation, is unable to satisfy safety constraints in non-linear systems\nsubject to varying operating conditions. Specifically, we study the MuJoCo\nHalfCheetah and Ant environments with velocity constraints and sudden joint\nloss non-stationarity. Then, we show that an agent trained using constrained\npolicy optimization, a safe RL algorithm, experiences catastrophic forgetting\nin continual learning settings. With this in mind, we explore a simple\nreward-shaping method to ensure that elastic weight consolidation prioritizes\nremembering both safety and task performance for safety-constrained,\nnon-linear, and non-stationary dynamical systems.\n", "date": "2025-02-21", "categories": []}
{"id": "2405.15773", "title": "Feature Aggregation with Latent Generative Replay for Federated\n  Continual Learning of Socially Appropriate Robot Behaviours", "abstract": "  It is critical for robots to explore Federated Learning (FL) settings where\nseveral robots, deployed in parallel, can learn independently while also\nsharing their learning with each other. This collaborative learning in\nreal-world environments requires social robots to adapt dynamically to changing\nand unpredictable situations and varying task settings. Our work contributes to\naddressing these challenges by exploring a simulated living room environment\nwhere robots need to learn the social appropriateness of their actions. First,\nwe propose Federated Root (FedRoot) averaging, a novel weight aggregation\nstrategy which disentangles feature learning across clients from individual\ntask-based learning. Second, to adapt to challenging environments, we extend\nFedRoot to Federated Latent Generative Replay (FedLGR), a novel Federated\nContinual Learning (FCL) strategy that uses FedRoot-based weight aggregation\nand embeds each client with a generator model for pseudo-rehearsal of learnt\nfeature embeddings to mitigate forgetting in a resource-efficient manner. Our\nresults show that FedRoot-based methods offer competitive performance while\nalso resulting in a sizeable reduction in resource consumption (up to 86% for\nCPU usage and up to 72% for GPU usage). Additionally, our results demonstrate\nthat FedRoot-based FCL methods outperform other methods while also offering an\nefficient solution (up to 84% CPU and 92% GPU usage reduction), with FedLGR\nproviding the best results across evaluations.\n", "date": "2024-03-16", "categories": []}
{"id": "2502.10789", "title": "ReReLRP -- Remembering and Recognizing Tasks with LRP", "abstract": "  Deep neural networks have revolutionized numerous research fields and\napplications. Despite their widespread success, a fundamental limitation known\nas catastrophic forgetting remains, where models fail to retain their ability\nto perform previously learned tasks after being trained on new ones. This\nlimitation is particularly acute in certain continual learning scenarios, where\nmodels must integrate the knowledge from new domains with their existing\ncapabilities. Traditional approaches to mitigate this problem typically rely on\nmemory replay mechanisms, storing either original data samples, prototypes, or\nactivation patterns. Although effective, these methods often introduce\nsignificant computational overhead, raise privacy concerns, and require the use\nof dedicated architectures. In this work we present ReReLRP (Remembering and\nRecognizing with LRP), a novel solution that leverages Layerwise Relevance\nPropagation (LRP) to preserve information across tasks. Our contribution\nprovides increased privacy of existing replay-free methods while additionally\noffering built-in explainability, flexibility of model architecture and\ndeployment, and a new mechanism to increase memory storage efficiency. We\nvalidate our approach on a wide variety of datasets, demonstrating results\ncomparable with a well-known replay-based method in selected scenarios.\n", "date": "2025-02-15", "categories": []}
{"id": "2502.14205", "title": "Accurate Forgetting for Heterogeneous Federated Continual Learning", "abstract": "  Recent years have witnessed a burgeoning interest in federated learning (FL).\nHowever, the contexts in which clients engage in sequential learning remain\nunder-explored. Bridging FL and continual learning (CL) gives rise to a\nchallenging practical problem: federated continual learning (FCL). Existing\nresearch in FCL primarily focuses on mitigating the catastrophic forgetting\nissue of continual learning while collaborating with other clients. We argue\nthat the forgetting phenomena are not invariably detrimental. In this paper, we\nconsider a more practical and challenging FCL setting characterized by\npotentially unrelated or even antagonistic data/tasks across different clients.\nIn the FL scenario, statistical heterogeneity and data noise among clients may\nexhibit spurious correlations which result in biased feature learning. While\nexisting CL strategies focus on a complete utilization of previous knowledge,\nwe found that forgetting biased information is beneficial in our study.\nTherefore, we propose a new concept accurate forgetting (AF) and develop a\nnovel generative-replay method~\\method~which selectively utilizes previous\nknowledge in federated networks. We employ a probabilistic framework based on a\nnormalizing flow model to quantify the credibility of previous knowledge.\nComprehensive experiments affirm the superiority of our method over baselines.\n", "date": "2025-02-20", "categories": []}
{"id": "2406.16437", "title": "Theory on Mixture-of-Experts in Continual Learning", "abstract": "  Continual learning (CL) has garnered significant attention because of its\nability to adapt to new tasks that arrive over time. Catastrophic forgetting\n(of old tasks) has been identified as a major issue in CL, as the model adapts\nto new tasks. The Mixture-of-Experts (MoE) model has recently been shown to\neffectively mitigate catastrophic forgetting in CL, by employing a gating\nnetwork to sparsify and distribute diverse tasks among multiple experts.\nHowever, there is a lack of theoretical analysis of MoE and its impact on the\nlearning performance in CL. This paper provides the first theoretical results\nto characterize the impact of MoE in CL via the lens of overparameterized\nlinear regression tasks. We establish the benefit of MoE over a single expert\nby proving that the MoE model can diversify its experts to specialize in\ndifferent tasks, while its router learns to select the right expert for each\ntask and balance the loads across all experts. Our study further suggests an\nintriguing fact that the MoE in CL needs to terminate the update of the gating\nnetwork after sufficient training rounds to attain system convergence, which is\nnot needed in the existing MoE studies that do not consider the continual task\narrival. Furthermore, we provide explicit expressions for the expected\nforgetting and overall generalization error to characterize the benefit of MoE\nin the learning performance in CL. Interestingly, adding more experts requires\nadditional rounds before convergence, which may not enhance the learning\nperformance. Finally, we conduct experiments on both synthetic and real\ndatasets to extend these insights from linear models to deep neural networks\n(DNNs), which also shed light on the practical algorithm design for MoE in CL.\n", "date": "2024-06-24", "categories": []}
{"id": "2304.01239", "title": "Online Distillation with Continual Learning for Cyclic Domain Shifts", "abstract": "  In recent years, online distillation has emerged as a powerful technique for\nadapting real-time deep neural networks on the fly using a slow, but accurate\nteacher model. However, a major challenge in online distillation is\ncatastrophic forgetting when the domain shifts, which occurs when the student\nmodel is updated with data from the new domain and forgets previously learned\nknowledge. In this paper, we propose a solution to this issue by leveraging the\npower of continual learning methods to reduce the impact of domain shifts.\nSpecifically, we integrate several state-of-the-art continual learning methods\nin the context of online distillation and demonstrate their effectiveness in\nreducing catastrophic forgetting. Furthermore, we provide a detailed analysis\nof our proposed solution in the case of cyclic domain shifts. Our experimental\nresults demonstrate the efficacy of our approach in improving the robustness\nand accuracy of online distillation, with potential applications in domains\nsuch as video surveillance or autonomous driving. Overall, our work represents\nan important step forward in the field of online distillation and continual\nlearning, with the potential to significantly impact real-world applications.\n", "date": "2023-04-03", "categories": []}
{"id": "2312.00276", "title": "Metalearning Continual Learning Algorithms", "abstract": "  General-purpose learning systems should improve themselves in open-ended\nfashion in ever-changing environments. Conventional learning algorithms for\nneural networks, however, suffer from catastrophic forgetting (CF), i.e.,\npreviously acquired skills are forgotten when a new task is learned. Instead of\nhand-crafting new algorithms for avoiding CF, we propose Automated Continual\nLearning (ACL) to train self-referential neural networks to metalearn their own\nin-context continual (meta)learning algorithms. ACL encodes continual learning\n(CL) desiderata -- good performance on both old and new tasks -- into its\nmetalearning objectives. Our experiments demonstrate that ACL effectively\nresolves \"in-context catastrophic forgetting,\" a problem that naive in-context\nlearning algorithms suffer from; ACL-learned algorithms outperform both\nhand-crafted learning algorithms and popular meta-continual learning methods on\nthe Split-MNIST benchmark in the replay-free setting, and enables continual\nlearning of diverse tasks consisting of multiple standard image classification\ndatasets. We also discuss the current limitations of in-context CL by comparing\nACL with state-of-the-art CL methods that leverage pre-trained models. Overall,\nwe bring several novel perspectives into the long-standing problem of CL.\n", "date": "2023-12-01", "categories": []}
{"id": "2407.00673", "title": "TEAL: New Selection Strategy for Small Buffers in Experience Replay\n  Class Incremental Learning", "abstract": "  Continual Learning is an unresolved challenge, whose relevance increases when\nconsidering modern applications. Unlike the human brain, trained deep neural\nnetworks suffer from a phenomenon called catastrophic forgetting, wherein they\nprogressively lose previously acquired knowledge upon learning new tasks. To\nmitigate this problem, numerous methods have been developed, many relying on\nthe replay of past exemplars during new task training. However, as the memory\nallocated for replay decreases, the effectiveness of these approaches\ndiminishes. On the other hand, maintaining a large memory for the purpose of\nreplay is inefficient and often impractical. Here we introduce TEAL, a novel\napproach to populate the memory with exemplars, that can be integrated with\nvarious experience-replay methods and significantly enhance their performance\nwith small memory buffers. We show that TEAL enhances the average accuracy of\nexisting class-incremental methods and outperforms other selection strategies,\nachieving state-of-the-art performance even with small memory buffers of 1-3\nexemplars per class in the final task. This confirms our initial hypothesis\nthat when memory is scarce, it is best to prioritize the most typical data.\nCode is available at this https URL: https://github.com/shahariel/TEAL.\n", "date": "2024-06-30", "categories": []}
{"id": "2502.07560", "title": "Navigating Semantic Drift in Task-Agnostic Class-Incremental Learning", "abstract": "  Class-incremental learning (CIL) seeks to enable a model to sequentially\nlearn new classes while retaining knowledge of previously learned ones.\nBalancing flexibility and stability remains a significant challenge,\nparticularly when the task ID is unknown. To address this, our study reveals\nthat the gap in feature distribution between novel and existing tasks is\nprimarily driven by differences in mean and covariance moments. Building on\nthis insight, we propose a novel semantic drift calibration method that\nincorporates mean shift compensation and covariance calibration. Specifically,\nwe calculate each class's mean by averaging its sample embeddings and estimate\ntask shifts using weighted embedding changes based on their proximity to the\nprevious mean, effectively capturing mean shifts for all learned classes with\neach new task. We also apply Mahalanobis distance constraint for covariance\ncalibration, aligning class-specific embedding covariances between old and\ncurrent networks to mitigate the covariance shift. Additionally, we integrate a\nfeature-level self-distillation approach to enhance generalization.\nComprehensive experiments on commonly used datasets demonstrate the\neffectiveness of our approach. The source code is available at\n\\href{https://github.com/fwu11/MACIL.git}{https://github.com/fwu11/MACIL.git}.\n", "date": "2025-02-11", "categories": []}
{"id": "2502.10442", "title": "Analysis of Overparameterization in Continual Learning under a Linear\n  Model", "abstract": "  Autonomous machine learning systems that learn many tasks in sequence are\nprone to the catastrophic forgetting problem. Mathematical theory is needed in\norder to understand the extent of forgetting during continual learning. As a\nfoundational step towards this goal, we study continual learning and\ncatastrophic forgetting from a theoretical perspective in the simple setting of\ngradient descent with no explicit algorithmic mechanism to prevent forgetting.\nIn this setting, we analytically demonstrate that overparameterization alone\ncan mitigate forgetting in the context of a linear regression model. We\nconsider a two-task setting motivated by permutation tasks, and show that as\nthe overparameterization ratio becomes sufficiently high, a model trained on\nboth tasks in sequence results in a low-risk estimator for the first task. As\npart of this work, we establish a non-asymptotic bound of the risk of a single\nlinear regression task, which may be of independent interest to the field of\ndouble descent theory.\n", "date": "2025-02-11", "categories": []}
{"id": "2502.10966", "title": "Neural Networks Remember More: The Power of Parameter Isolation and\n  Combination", "abstract": "  Catastrophic forgetting is a pervasive issue for pre-trained language models\n(PLMs) during continual learning, where models lose previously acquired\nknowledge when sequentially trained on a series of tasks. The model's ability\nto retain old tasks is referred to as stability, while its adaptability to new\ntasks is called plasticity. Therefore, the key to solving this problem is to\nfind a trade-off between the plasticity and stability of the model. To address\nthis issue, in this paper, we propose a novel method to achieve a balance\nbetween model stability and plasticity, thereby mitigating catastrophic\nforgetting. More specifically, our proposed approach leverages parameter\nisolation and a subsequent combination strategy. Initially, in the training\nstage, the model adapts to each downstream task via a parameter isolation\nmethod to prevent potential interference among different tasks. We then combine\nall trained parameters, which contain acquired knowledge, using the task\narithmetic method and finally apply them to the backbone model. Empirical\nevaluations on continual language learning benchmarks substantiate the\neffectiveness of our approach, revealing a marked enhancement over existing\nstate-of-the-art approaches.\n", "date": "2025-02-16", "categories": []}
{"id": "2502.11305", "title": "Non-Uniform Memory Sampling in Experience Replay", "abstract": "  Continual learning is the process of training machine learning models on a\nsequence of tasks where data distributions change over time. A well-known\nobstacle in this setting is catastrophic forgetting, a phenomenon in which a\nmodel drastically loses performance on previously learned tasks when learning\nnew ones. A popular strategy to alleviate this problem is experience replay, in\nwhich a subset of old samples is stored in a memory buffer and replayed with\nnew data. Despite continual learning advances focusing on which examples to\nstore and how to incorporate them into the training loss, most approaches\nassume that sampling from this buffer is uniform by default.\n  We challenge the assumption that uniform sampling is necessarily optimal. We\nconduct an experiment in which the memory buffer updates the same way in every\ntrial, but the replay probability of each stored sample changes between trials\nbased on different random weight distributions. Specifically, we generate 50\ndifferent non-uniform sampling probability weights for each trial and compare\ntheir final accuracy to the uniform sampling baseline. We find that there is\nalways at least one distribution that significantly outperforms the baseline\nacross multiple buffer sizes, models, and datasets. These results suggest that\nmore principled adaptive replay policies could yield further gains. We discuss\nhow exploiting this insight could inspire new research on non-uniform memory\nsampling in continual learning to better mitigate catastrophic forgetting.\n  The code supporting this study is available at\n$\\href{https://github.com/DentonJC/memory-sampling}{https://github.com/DentonJC/memory-sampling}$.\n", "date": "2025-02-16", "categories": []}
{"id": "2502.11482", "title": "DATA: Decomposed Attention-based Task Adaptation for Rehearsal-Free\n  Continual Learning", "abstract": "  Continual learning (CL) is essential for Large Language Models (LLMs) to\nadapt to evolving real-world demands, yet they are susceptible to catastrophic\nforgetting (CF). While traditional CF solutions rely on expensive data\nrehearsal, recent rehearsal-free methods employ model-based and\nregularization-based strategies to address this issue. However, these\napproaches often neglect the model's plasticity, which is crucial to achieving\noptimal performance on newly learned tasks. Consequently, a key challenge in CL\nis striking a balance between preserving plasticity and mitigating CF. To\ntackle this challenge, we propose the $\\textbf{D}$ecomposed\n$\\textbf{A}$ttention-based $\\textbf{T}$ask $\\textbf{A}$daptation (DATA), which\nexplicitly decouples and learns both task-specific and task-shared knowledge\nusing high-rank and low-rank task adapters (e.g., LoRAs). For new tasks, DATA\ndynamically adjusts the weights of adapters of different ranks based on their\nrelevance and distinction from previous tasks, allowing the model to acquire\nnew task-specific skills while effectively retaining previously learned\nknowledge. Specifically, we implement a decomposed component weighting strategy\ncomprising learnable components that collectively generate attention-based\nweights, allowing the model to integrate and utilize diverse knowledge from\neach DATA. Extensive experiments on three widely used benchmarks demonstrate\nthat our proposed method achieves state-of-the-art performance. Notably, our\napproach significantly enhances model plasticity and mitigates CF by extending\nlearnable components and employing stochastic restoration during training\niterations.\n", "date": "2025-02-17", "categories": []}
{"id": "2502.11756", "title": "On the Computation of the Fisher Information in Continual Learning", "abstract": "  One of the most popular methods for continual learning with deep neural\nnetworks is Elastic Weight Consolidation (EWC), which involves computing the\nFisher Information. The exact way in which the Fisher Information is computed\nis however rarely described, and multiple different implementations for it can\nbe found online. This blog post discusses and empirically compares several\noften-used implementations, which highlights that many currently reported\nresults for EWC could likely be improved by changing the way the Fisher\nInformation is computed.\n", "date": "2025-02-17", "categories": []}
{"id": "2502.11927", "title": "Continual Learning Should Move Beyond Incremental Classification", "abstract": "  Continual learning (CL) is the sub-field of machine learning concerned with\naccumulating knowledge in dynamic environments. So far, CL research has mainly\nfocused on incremental classification tasks, where models learn to classify new\ncategories while retaining knowledge of previously learned ones. Here, we argue\nthat maintaining such a focus limits both theoretical development and practical\napplicability of CL methods. Through a detailed analysis of concrete examples -\nincluding multi-target classification, robotics with constrained output spaces,\nlearning in continuous task domains, and higher-level concept memorization - we\ndemonstrate how current CL approaches often fail when applied beyond standard\nclassification. We identify three fundamental challenges: (C1) the nature of\ncontinuity in learning problems, (C2) the choice of appropriate spaces and\nmetrics for measuring similarity, and (C3) the role of learning objectives\nbeyond classification. For each challenge, we provide specific recommendations\nto help move the field forward, including formalizing temporal dynamics through\ndistribution processes, developing principled approaches for continuous task\nspaces, and incorporating density estimation and generative objectives. In so\ndoing, this position paper aims to broaden the scope of CL research while\nstrengthening its theoretical foundations, making it more applicable to\nreal-world problems.\n", "date": "2025-02-17", "categories": []}
{"id": "2502.09500", "title": "Eidetic Learning: an Efficient and Provable Solution to Catastrophic\n  Forgetting", "abstract": "  Catastrophic forgetting -- the phenomenon of a neural network learning a task\nt1 and losing the ability to perform it after being trained on some other task\nt2 -- is a long-standing problem for neural networks [McCloskey and Cohen,\n1989]. We present a method, Eidetic Learning, that provably solves catastrophic\nforgetting. A network trained with Eidetic Learning -- here, an EideticNet --\nrequires no rehearsal or replay. We consider successive discrete tasks and show\nhow at inference time an EideticNet automatically routes new instances without\nauxiliary task information. An EideticNet bears a family resemblance to the\nsparsely-gated Mixture-of-Experts layer Shazeer et al. [2016] in that network\ncapacity is partitioned across tasks and the network itself performs\ndata-conditional routing. An EideticNet is easy to implement and train, is\nefficient, and has time and space complexity linear in the number of\nparameters. The guarantee of our method holds for normalization layers of\nmodern neural networks during both pre-training and fine-tuning. We show with a\nvariety of network architectures and sets of tasks that EideticNets are immune\nto forgetting. While the practical benefits of EideticNets are substantial, we\nbelieve they can be benefit practitioners and theorists alike. The code for\ntraining EideticNets is available at\nhttps://github.com/amazon-science/eideticnet-training.\n", "date": "2025-02-13", "categories": []}
{"id": "2502.09140", "title": "Replay-free Online Continual Learning with Self-Supervised MultiPatches", "abstract": "  Online Continual Learning (OCL) methods train a model on a non-stationary\ndata stream where only a few examples are available at a time, often leveraging\nreplay strategies. However, usage of replay is sometimes forbidden, especially\nin applications with strict privacy regulations. Therefore, we propose\nContinual MultiPatches (CMP), an effective plug-in for existing OCL\nself-supervised learning strategies that avoids the use of replay samples. CMP\ngenerates multiple patches from a single example and projects them into a\nshared feature space, where patches coming from the same example are pushed\ntogether without collapsing into a single point. CMP surpasses replay and other\nSSL-based strategies on OCL streams, challenging the role of replay as a go-to\nsolution for self-supervised OCL.\n", "date": "2025-02-13", "categories": []}
{"id": "2502.09143", "title": "Feature-based Graph Attention Networks Improve Online Continual Learning", "abstract": "  Online continual learning for image classification is crucial for models to\nadapt to new data while retaining knowledge of previously learned tasks. This\ncapability is essential to address real-world challenges involving dynamic\nenvironments and evolving data distributions. Traditional approaches\npredominantly employ Convolutional Neural Networks, which are limited to\nprocessing images as grids and primarily capture local patterns rather than\nrelational information. Although the emergence of transformer architectures has\nimproved the ability to capture relationships, these models often require\nsignificantly larger resources. In this paper, we present a novel online\ncontinual learning framework based on Graph Attention Networks (GATs), which\neffectively capture contextual relationships and dynamically update the\ntask-specific representation via learned attention weights. Our approach\nutilizes a pre-trained feature extractor to convert images into graphs using\nhierarchical feature maps, representing information at varying levels of\ngranularity. These graphs are then processed by a GAT and incorporate an\nenhanced global pooling strategy to improve classification performance for\ncontinual learning. In addition, we propose the rehearsal memory duplication\ntechnique that improves the representation of the previous tasks while\nmaintaining the memory budget. Comprehensive evaluations on benchmark datasets,\nincluding SVHN, CIFAR10, CIFAR100, and MiniImageNet, demonstrate the\nsuperiority of our method compared to the state-of-the-art methods.\n", "date": "2025-02-13", "categories": []}
{"id": "2305.16332", "title": "Continual Learning through Human-Robot Interaction: Human Perceptions of\n  a Continual Learning Robot in Repeated Interactions", "abstract": "  For long-term deployment in dynamic real-world environments, assistive robots\nmust continue to learn and adapt to their environments. Researchers have\ndeveloped various computational models for continual learning (CL) that can\nallow robots to continually learn from limited training data, and avoid\nforgetting previous knowledge. While these CL models can mitigate forgetting on\nstatic, systematically collected datasets, it is unclear how human users might\nperceive a robot that continually learns over multiple interactions with them.\nIn this paper, we developed a system that integrates CL models for object\nrecognition with a Fetch mobile manipulator robot and allows human participants\nto directly teach and test the robot over multiple sessions. We conducted an\nin-person study with 60 participants that interacted with our system in 300\nsessions (5 sessions per participant). We conducted a between-subject study\nwith three different CL models to understand human perceptions of continual\nlearning robots over multiple sessions. Our results suggest that participants'\nperceptions of trust, competence, and usability of a continual learning robot\nsignificantly decrease over multiple sessions if the robot forgets previously\nlearned objects. However, the perceived task load on participants for teaching\nand testing the robot remains the same over multiple sessions even if the robot\nforgets previously learned objects. Our results also indicate that\nstate-of-the-art CL models might perform unreliably when applied on robots\ninteracting with human participants. Further, continual learning robots are not\nperceived as very trustworthy or competent by human participants, regardless of\nthe underlying continual learning model or the session number.\n", "date": "2023-05-22", "categories": []}
{"id": "2305.19270", "title": "Learning without Forgetting for Vision-Language Models", "abstract": "  Class-Incremental Learning (CIL) or continual learning is a desired\ncapability in the real world, which requires a learning system to adapt to new\ntasks without forgetting former ones. While traditional CIL methods focus on\nvisual information to grasp core features, recent advances in Vision-Language\nModels (VLM) have shown promising capabilities in learning generalizable\nrepresentations with the aid of textual information. However, when continually\ntrained with new classes, VLMs often suffer from catastrophic forgetting of\nformer knowledge. Applying VLMs to CIL poses two major challenges: 1) how to\nadapt the model without forgetting; and 2) how to make full use of the\nmulti-modal information. To this end, we propose PROjectiOn Fusion (PROOF) that\nenables VLMs to learn without forgetting. To handle the first challenge, we\npropose training task-specific projections based on the frozen image/text\nencoders. When facing new tasks, new projections are expanded and former\nprojections are fixed, alleviating the forgetting of old concepts. For the\nsecond challenge, we propose the fusion module to better utilize the\ncross-modality information. By jointly adjusting visual and textual features,\nthe model can capture semantic information with stronger representation\nability. Extensive experiments on nine benchmark datasets validate PROOF\nachieves state-of-the-art performance. Code is available at\nhttps://github.com/zhoudw-zdw/PROOF\n", "date": "2023-05-30", "categories": []}
{"id": "2402.17400", "title": "Investigating Continual Pretraining in Large Language Models: Insights\n  and Implications", "abstract": "  Continual learning (CL) in large language models (LLMs) is an evolving domain\nthat focuses on developing efficient and sustainable training strategies to\nadapt models to emerging knowledge and achieve robustness in dynamic\nenvironments. Our primary emphasis is on continual domain-adaptive pretraining,\na process designed to equip LLMs with the ability to integrate new information\nfrom various domains while retaining previously learned knowledge. Since\nexisting works concentrate mostly on continual fine-tuning for a limited\nselection of downstream tasks or training domains, we introduce a new benchmark\ndesigned to measure the adaptability of LLMs to changing pretraining data\nlandscapes. We further examine the impact of model size on learning efficacy\nand forgetting, as well as how the progression and similarity of emerging\ndomains affect the knowledge transfer within these models.\n  Our findings uncover several key insights: (i) continual pretraining\nconsistently improves <1.5B models studied in this work and is also superior to\ndomain adaptation, (ii) larger models always achieve better perplexity than\nsmaller ones when continually pretrained on the same corpus, (iii) smaller\nmodels are particularly sensitive to continual pretraining, showing the most\nsignificant rates of both learning and forgetting, (iv) continual pretraining\nboosts downstream task performance of GPT-2 family, (v) continual pretraining\nenables LLMs to specialize better when the sequence of domains shows semantic\nsimilarity while randomizing training domains leads to better transfer and\nfinal performance otherwise. We posit that our research establishes a new\nbenchmark for CL in LLMs, providing a more realistic evaluation of knowledge\nretention and transfer across diverse domains.\n", "date": "2024-02-27", "categories": []}
{"id": "2502.08181", "title": "Latest Advancements Towards Catastrophic Forgetting under Data Scarcity:\n  A Comprehensive Survey on Few-Shot Class Incremental Learning", "abstract": "  Data scarcity significantly complicates the continual learning problem, i.e.,\nhow a deep neural network learns in dynamic environments with very few samples.\nHowever, the latest progress of few-shot class incremental learning (FSCIL)\nmethods and related studies show insightful knowledge on how to tackle the\nproblem. This paper presents a comprehensive survey on FSCIL that highlights\nseveral important aspects i.e. comprehensive and formal objectives of FSCIL\napproaches, the importance of prototype rectifications, the new learning\nparadigms based on pre-trained model and language-guided mechanism, the deeper\nanalysis of FSCIL performance metrics and evaluation, and the practical\ncontexts of FSCIL in various areas. Our extensive discussion presents the open\nchallenges, potential solutions, and future directions of FSCIL.\n", "date": "2025-02-12", "categories": []}
{"id": "2402.08096", "title": "An Efficient Rehearsal Scheme for Catastrophic Forgetting Mitigation\n  during Multi-stage Fine-tuning", "abstract": "  Incrementally fine-tuning foundational models on new tasks or domains is now\nthe de facto approach in NLP. A known pitfall of this approach is the\n\\emph{catastrophic forgetting} of prior knowledge that happens during\nfine-tuning. A common approach to alleviate such forgetting is to rehearse\nsamples from prior tasks during fine-tuning. Several existing works assume a\nfixed memory buffer to store prior task examples, while relying on inferences\n(forward passes) with the model at hand for choosing examples for rehearsal\nfrom the buffer. However, given the increasing computational cost of model\ninference, and decreasing cost of data storage, we focus on the setting to\nrehearse samples with a fixed computational budget instead of a fixed memory\nbudget. We propose a sampling scheme, \\texttt{\\bf mix-cd}, that prioritizes\nrehearsal of ``collateral damage'' samples, which are samples predicted\ncorrectly by the prior model but forgotten by the incrementally tuned one. The\ncrux of our scheme is a procedure to efficiently estimate the density of\ncollateral damage samples without incurring additional model inferences. Our\napproach is computationally efficient, easy to implement, and outperforms\nseveral leading continual learning methods in compute-constrained settings. All\nthe code will be publicly available at\nhttps://github.com/jybai/mix-cd-rehearsal.\n", "date": "2024-02-12", "categories": []}
{"id": "2410.00700", "title": "Mining Your Own Secrets: Diffusion Classifier Scores for Continual\n  Personalization of Text-to-Image Diffusion Models", "abstract": "  Personalized text-to-image diffusion models have grown popular for their\nability to efficiently acquire a new concept from user-defined text\ndescriptions and a few images. However, in the real world, a user may wish to\npersonalize a model on multiple concepts but one at a time, with no access to\nthe data from previous concepts due to storage/privacy concerns. When faced\nwith this continual learning (CL) setup, most personalization methods fail to\nfind a balance between acquiring new concepts and retaining previous ones -- a\nchallenge that continual personalization (CP) aims to solve. Inspired by the\nsuccessful CL methods that rely on class-specific information for\nregularization, we resort to the inherent class-conditioned density estimates,\nalso known as diffusion classifier (DC) scores, for continual personalization\nof text-to-image diffusion models. Namely, we propose using DC scores for\nregularizing the parameter-space and function-space of text-to-image diffusion\nmodels, to achieve continual personalization. Using several diverse evaluation\nsetups, datasets, and metrics, we show that our proposed regularization-based\nCP methods outperform the state-of-the-art C-LoRA, and other baselines.\nFinally, by operating in the replay-free CL setup and on low-rank adapters, our\nmethod incurs zero storage and parameter overhead, respectively, over the\nstate-of-the-art. Our project page:\nhttps://srvcodes.github.io/continual_personalization/\n", "date": "2024-10-01", "categories": []}
{"id": "2502.06327", "title": "Prompt-Driven Continual Graph Learning", "abstract": "  Continual Graph Learning (CGL), which aims to accommodate new tasks over\nevolving graph data without forgetting prior knowledge, is garnering\nsignificant research interest. Mainstream solutions adopt the memory\nreplay-based idea, ie, caching representative data from earlier tasks for\nretraining the graph model. However, this strategy struggles with scalability\nissues for constantly evolving graphs and raises concerns regarding data\nprivacy. Inspired by recent advancements in the prompt-based learning paradigm,\nthis paper introduces a novel prompt-driven continual graph learning\n(PROMPTCGL) framework, which learns a separate prompt for each incoming task\nand maintains the underlying graph neural network model fixed. In this way,\nPROMPTCGL naturally avoids catastrophic forgetting of knowledge from previous\ntasks. More specifically, we propose hierarchical prompting to instruct the\nmodel from both feature- and topology-level to fully address the variability of\ntask graphs in dynamic continual learning. Additionally, we develop a\npersonalized prompt generator to generate tailored prompts for each graph node\nwhile minimizing the number of prompts needed, leading to constant memory\nconsumption regardless of the graph scale. Extensive experiments on four\nbenchmarks show that PROMPTCGL achieves superior performance against existing\nCGL approaches while significantly reducing memory consumption. Our code is\navailable at https://github.com/QiWang98/PromptCGL.\n", "date": "2025-02-10", "categories": []}
{"id": "2502.06544", "title": "Sequence Transferability and Task Order Selection in Continual Learning", "abstract": "  In continual learning, understanding the properties of task sequences and\ntheir relationships to model performance is important for developing advanced\nalgorithms with better accuracy. However, efforts in this direction remain\nunderdeveloped despite encouraging progress in methodology development. In this\nwork, we investigate the impacts of sequence transferability on continual\nlearning and propose two novel measures that capture the total transferability\nof a task sequence, either in the forward or backward direction. Based on the\nempirical properties of these measures, we then develop a new method for the\ntask order selection problem in continual learning. Our method can be shown to\noffer a better performance than the conventional strategy of random task\nselection.\n", "date": "2025-02-10", "categories": []}
{"id": "2502.04870", "title": "IPSeg: Image Posterior Mitigates Semantic Drift in Class-Incremental\n  Segmentation", "abstract": "  Class incremental learning aims to enable models to learn from sequential,\nnon-stationary data streams across different tasks without catastrophic\nforgetting. In class incremental semantic segmentation (CISS), the semantic\ncontent of image pixels evolves over incremental phases, known as semantic\ndrift. In this work, we identify two critical challenges in CISS that\ncontribute to semantic drift and degrade performance. First, we highlight the\nissue of separate optimization, where different parts of the model are\noptimized in distinct incremental stages, leading to misaligned probability\nscales. Second, we identify noisy semantics arising from inappropriate\npseudo-labeling, which results in sub-optimal results. To address these\nchallenges, we propose a novel and effective approach, Image Posterior and\nSemantics Decoupling for Segmentation (IPSeg). IPSeg introduces two key\nmechanisms: (1) leveraging image posterior probabilities to align optimization\nacross stages and mitigate the effects of separate optimization, and (2)\nemploying semantics decoupling to handle noisy semantics and tailor learning\nstrategies for different semantics. Extensive experiments on the Pascal VOC\n2012 and ADE20K datasets demonstrate that IPSeg achieves superior performance\ncompared to state-of-the-art methods, particularly in challenging long-term\nincremental scenarios.\n", "date": "2025-02-07", "categories": []}
{"id": "2407.16920", "title": "Train-Attention: Meta-Learning Where to Focus in Continual Knowledge\n  Learning", "abstract": "  Previous studies on continual knowledge learning (CKL) in large language\nmodels (LLMs) have predominantly focused on approaches such as regularization,\narchitectural modifications, and rehearsal techniques to mitigate catastrophic\nforgetting. However, these methods naively inherit the inefficiencies of\nstandard training procedures, indiscriminately applying uniform weight across\nall tokens, which can lead to unnecessary parameter updates and increased\nforgetting. To address these shortcomings, we propose a novel CKL approach\ntermed Train-Attention-Augmented Language Model (TAALM), which enhances\nlearning efficiency by dynamically predicting and applying weights to tokens\nbased on their usefulness. This method employs a meta-learning framework that\noptimizes token importance predictions, facilitating targeted knowledge updates\nand minimizing forgetting. Also, we observe that existing benchmarks do not\nclearly exhibit the trade-off between learning and retaining, therefore we\npropose a new benchmark, \\textsc{LAMA-ckl}, to address this issue. Through\nexperiments conducted on both newly introduced and established CKL benchmarks,\nTAALM proves the state-of-the-art performance upon the baselines, and also\nshows synergistic compatibility when integrated with previous CKL approaches.\n", "date": "2024-07-24", "categories": []}
{"id": "2502.02909", "title": "SPARC: Subspace-Aware Prompt Adaptation for Robust Continual Learning in\n  LLMs", "abstract": "  We propose SPARC, a lightweight continual learning framework for large\nlanguage models (LLMs) that enables efficient task adaptation through prompt\ntuning in a lower-dimensional space. By leveraging principal component analysis\n(PCA), we identify a compact subspace of the training data. Optimizing prompts\nin this lower-dimensional space enhances training efficiency, as it focuses\nupdates on the most relevant features while reducing computational overhead.\nFurthermore, since the model's internal structure remains unaltered, the\nextensive knowledge gained from pretraining is fully preserved, ensuring that\npreviously learned information is not compromised during adaptation. Our method\nachieves high knowledge retention in both task-incremental and\ndomain-incremental continual learning setups while fine-tuning only 0.04% of\nthe model's parameters. Additionally, by integrating LoRA, we enhance\nadaptability to computational constraints, allowing for a tradeoff between\naccuracy and training cost. Experiments on the SuperGLUE benchmark demonstrate\nthat our PCA-based prompt tuning combined with LoRA maintains full knowledge\nretention while improving accuracy, utilizing only 1% of the model's\nparameters. These results establish our approach as a scalable and\nresource-efficient solution for continual learning in LLMs.\n", "date": "2025-02-05", "categories": []}
{"id": "2406.02596", "title": "Slow and Steady Wins the Race: Maintaining Plasticity with Hare and\n  Tortoise Networks", "abstract": "  This study investigates the loss of generalization ability in neural\nnetworks, revisiting warm-starting experiments from Ash & Adams. Our empirical\nanalysis reveals that common methods designed to enhance plasticity by\nmaintaining trainability provide limited benefits to generalization. While\nreinitializing the network can be effective, it also risks losing valuable\nprior knowledge. To this end, we introduce the Hare & Tortoise, inspired by the\nbrain's complementary learning system. Hare & Tortoise consists of two\ncomponents: the Hare network, which rapidly adapts to new information\nanalogously to the hippocampus, and the Tortoise network, which gradually\nintegrates knowledge akin to the neocortex. By periodically reinitializing the\nHare network to the Tortoise's weights, our method preserves plasticity while\nretaining general knowledge. Hare & Tortoise can effectively maintain the\nnetwork's ability to generalize, which improves advanced reinforcement learning\nalgorithms on the Atari-100k benchmark. The code is available at\nhttps://github.com/dojeon-ai/hare-tortoise.\n", "date": "2024-06-01", "categories": []}
{"id": "2502.01866", "title": "Online Curvature-Aware Replay: Leveraging $\\mathbf{2^{nd}}$ Order\n  Information for Online Continual Learning", "abstract": "  Online Continual Learning (OCL) models continuously adapt to nonstationary\ndata streams, usually without task information. These settings are complex and\nmany traditional CL methods fail, while online methods (mainly replay-based)\nsuffer from instabilities after the task shift. To address this issue, we\nformalize replay-based OCL as a second-order online joint optimization with\nexplicit KL-divergence constraints on replay data. We propose Online\nCurvature-Aware Replay (OCAR) to solve the problem: a method that leverages\nsecond-order information of the loss using a K-FAC approximation of the Fisher\nInformation Matrix (FIM) to precondition the gradient. The FIM acts as a\nstabilizer to prevent forgetting while also accelerating the optimization in\nnon-interfering directions. We show how to adapt the estimation of the FIM to a\ncontinual setting stabilizing second-order optimization for non-iid data,\nuncovering the role of the Tikhonov regularization in the stability-plasticity\ntradeoff. Empirical results show that OCAR outperforms state-of-the-art methods\nin continual metrics achieving higher average accuracy throughout the training\nprocess in three different benchmarks.\n", "date": "2025-02-03", "categories": []}
{"id": "2502.01427", "title": "Structural features of the fly olfactory circuit mitigate the\n  stability-plasticity dilemma in continual learning", "abstract": "  Artificial neural networks face the stability-plasticity dilemma in continual\nlearning, while the brain can maintain memories and remain adaptable. However,\nthe biological strategies for continual learning and their potential to inspire\nlearning algorithms in neural networks are poorly understood. This study\npresents a minimal model of the fly olfactory circuit to investigate the\nbiological strategies that support continual odor learning. We introduce the\nfly olfactory circuit as a plug-and-play component, termed the Fly Model, which\ncan integrate with modern machine learning methods to address this dilemma. Our\nfindings demonstrate that the Fly Model enhances both memory stability and\nlearning plasticity, overcoming the limitations of current continual learning\nstrategies. We validated its effectiveness across various challenging continual\nlearning scenarios using commonly used datasets. The fly olfactory system\nserves as an elegant biological circuit for lifelong learning, offering a\nmodule that enhances continual learning with minimal additional computational\ncost for machine learning.\n", "date": "2025-02-03", "categories": []}
{"id": "2501.16922", "title": "Agential AI for Integrated Continual Learning, Deliberative Behavior,\n  and Comprehensible Models", "abstract": "  Contemporary machine learning paradigm excels in statistical data analysis,\nsolving problems that classical AI couldn't. However, it faces key limitations,\nsuch as a lack of integration with planning, incomprehensible internal\nstructure, and inability to learn continually. We present the initial design\nfor an AI system, Agential AI (AAI), in principle operating independently or on\ntop of statistical methods, designed to overcome these issues. AAI's core is a\nlearning method that models temporal dynamics with guarantees of completeness,\nminimality, and continual learning, using component-level variation and\nselection to learn the structure of the environment. It integrates this with a\nbehavior algorithm that plans on a learned model and encapsulates high-level\nbehavior patterns. Preliminary experiments on a simple environment show AAI's\neffectiveness and potential.\n", "date": "2025-01-28", "categories": []}
{"id": "2407.10315", "title": "Order parameters and phase transitions of continual learning in deep\n  neural networks", "abstract": "  Continual learning (CL) enables animals to learn new tasks without erasing\nprior knowledge. CL in artificial neural networks (NNs) is challenging due to\ncatastrophic forgetting, where new learning degrades performance on older\ntasks. While various techniques exist to mitigate forgetting, theoretical\ninsights into when and why CL fails in NNs are lacking. Here, we present a\nstatistical-mechanics theory of CL in deep, wide NNs, which characterizes the\nnetwork's input-output mapping as it learns a sequence of tasks. It gives rise\nto order parameters (OPs) that capture how task relations and network\narchitecture influence forgetting and anterograde interference, as verified by\nnumerical evaluations. For networks with a shared readout for all tasks\n(single-head CL), the relevant-feature and rule similarity between tasks,\nrespectively measured by two OPs, are sufficient to predict a wide range of CL\nbehaviors. In addition, the theory predicts that increasing the network depth\ncan effectively reduce interference between tasks, thereby lowering forgetting.\nFor networks with task-specific readouts (multi-head CL), the theory identifies\na phase transition where CL performance shifts dramatically as tasks become\nless similar, as measured by another task-similarity OP. While forgetting is\nrelatively mild compared to single-head CL across all tasks, sufficiently low\nsimilarity leads to catastrophic anterograde interference, where the network\nretains old tasks perfectly but completely fails to generalize new learning.\nOur results delineate important factors affecting CL performance and suggest\nstrategies for mitigating forgetting.\n", "date": "2024-07-14", "categories": []}
{"id": "2501.15454", "title": "On the Discrimination and Consistency for Exemplar-Free Class\n  Incremental Learning", "abstract": "  Exemplar-free class incremental learning (EF-CIL) is a nontrivial task that\nrequires continuously enriching model capability with new classes while\nmaintaining previously learned knowledge without storing and replaying any old\nclass exemplars. An emerging theory-guided framework for CIL trains\ntask-specific models for a shared network, shifting the pressure of forgetting\nto task-id prediction. In EF-CIL, task-id prediction is more challenging due to\nthe lack of inter-task interaction (e.g., replays of exemplars). To address\nthis issue, we conduct a theoretical analysis of the importance and feasibility\nof preserving a discriminative and consistent feature space, upon which we\npropose a novel method termed DCNet. Concretely, it progressively maps class\nrepresentations into a hyperspherical space, in which different classes are\northogonally distributed to achieve ample inter-class separation. Meanwhile, it\nalso introduces compensatory training to adaptively adjust supervision\nintensity, thereby aligning the degree of intra-class aggregation. Extensive\nexperiments and theoretical analysis verified the superiority of the proposed\nDCNet.\n", "date": "2025-01-26", "categories": []}
{"id": "2501.15998", "title": "Controllable Forgetting Mechanism for Few-Shot Class-Incremental\n  Learning", "abstract": "  Class-incremental learning in the context of limited personal labeled samples\n(few-shot) is critical for numerous real-world applications, such as smart home\ndevices. A key challenge in these scenarios is balancing the trade-off between\nadapting to new, personalized classes and maintaining the performance of the\nmodel on the original, base classes. Fine-tuning the model on novel classes\noften leads to the phenomenon of catastrophic forgetting, where the accuracy of\nbase classes declines unpredictably and significantly. In this paper, we\npropose a simple yet effective mechanism to address this challenge by\ncontrolling the trade-off between novel and base class accuracy. We\nspecifically target the ultra-low-shot scenario, where only a single example is\navailable per novel class. Our approach introduces a Novel Class Detection\n(NCD) rule, which adjusts the degree of forgetting a priori while\nsimultaneously enhancing performance on novel classes. We demonstrate the\nversatility of our solution by applying it to state-of-the-art Few-Shot\nClass-Incremental Learning (FSCIL) methods, showing consistent improvements\nacross different settings. To better quantify the trade-off between novel and\nbase class performance, we introduce new metrics: NCR@2FOR and NCR@5FOR. Our\napproach achieves up to a 30% improvement in novel class accuracy on the\nCIFAR100 dataset (1-shot, 1 novel class) while maintaining a controlled base\nclass forgetting rate of 2%.\n", "date": "2025-01-27", "categories": []}
{"id": "2405.14318", "title": "Adaptive Retention & Correction: Test-Time Training for Continual\n  Learning", "abstract": "  Continual learning, also known as lifelong learning or incremental learning,\nrefers to the process by which a model learns from a stream of incoming data\nover time. A common problem in continual learning is the classification layer's\nbias towards the most recent task. Traditionally, methods have relied on\nincorporating data from past tasks during training to mitigate this issue.\nHowever, the recent shift in continual learning to memory-free environments has\nrendered these approaches infeasible. In this study, we propose a solution\nfocused on the testing phase. We first introduce a simple Out-of-Task Detection\nmethod, OTD, designed to accurately identify samples from past tasks during\ntesting. Leveraging OTD, we then propose: (1) an Adaptive Retention mechanism\nfor dynamically tuning the classifier layer on past task data; (2) an Adaptive\nCorrection mechanism for revising predictions when the model classifies data\nfrom previous tasks into classes from the current task. We name our approach\nAdaptive Retention & Correction (ARC). While designed for memory-free\nenvironments, ARC also proves effective in memory-based settings. Extensive\nexperiments show that our proposed method can be plugged in to virtually any\nexisting continual learning approach without requiring any modifications to its\ntraining procedure. Specifically, when integrated with state-of-the-art\napproaches, ARC achieves an average performance increase of 2.7% and 2.6% on\nthe CIFAR-100 and Imagenet-R datasets, respectively.\n", "date": "2024-05-23", "categories": []}
{"id": "2501.14369", "title": "Low-rank Prompt Interaction for Continual Vision-Language Retrieval", "abstract": "  Research on continual learning in multi-modal tasks has been receiving\nincreasing attention. However, most existing work overlooks the explicit\ncross-modal and cross-task interactions. In this paper, we innovatively propose\nthe Low-rank Prompt Interaction (LPI) to address this general problem of\nmulti-modal understanding, which considers both cross-modal and cross-task\ninteractions. Specifically, as for the former, we employ multi-modal\ncorrelation modules for corresponding Transformer layers. Considering that the\ntraining parameters scale to the number of layers and tasks, we propose\nlow-rank interaction-augmented decomposition to avoid memory explosion while\nenhancing the cross-modal association through sharing and separating\ncommon-specific low-rank factors. In addition, due to the multi-modal semantic\ndifferences carried by the low-rank initialization, we adopt hierarchical\nlow-rank contrastive learning to ensure training robustness. As for the latter,\nwe initially employ a visual analysis and identify that different tasks have\nclear distinctions in proximity. Therefore, we introduce explicit task\ncontrastive constraints in the prompt learning process based on task semantic\ndistances. Experiments on two retrieval tasks show performance improvements\nwith the introduction of a minimal number of parameters, demonstrating the\neffectiveness of our method. Code is available at\nhttps://github.com/Kelvin-ywc/LPI.\n", "date": "2025-01-24", "categories": []}
{"id": "2405.07295", "title": "Bridging Neuroscience and AI: Environmental Enrichment as a Model for\n  Forward Knowledge Transfer", "abstract": "  Continual learning (CL) refers to an agent's capability to learn from a\ncontinuous stream of data and transfer knowledge without forgetting old\ninformation. One crucial aspect of CL is forward transfer, i.e., improved and\nfaster learning on a new task by leveraging information from prior knowledge.\nWhile this ability comes naturally to biological brains, it poses a significant\nchallenge for artificial intelligence (AI). Here, we suggest that environmental\nenrichment (EE) can be used as a biological model for studying forward\ntransfer, inspiring human-like AI development. EE refers to animal studies that\nenhance cognitive, social, motor, and sensory stimulation and is a model for\nwhat, in humans, is referred to as 'cognitive reserve'. Enriched animals show\nsignificant improvement in learning speed and performance on new tasks,\ntypically exhibiting forward transfer. We explore anatomical, molecular, and\nneuronal changes post-EE and discuss how artificial neural networks (ANNs) can\nbe used to predict neural computation changes after enriched experiences.\nFinally, we provide a synergistic way of combining neuroscience and AI research\nthat paves the path toward developing AI capable of rapid and efficient new\ntask learning.\n", "date": "2024-05-12", "categories": []}
{"id": "2408.05200", "title": "KIF: Knowledge Identification and Fusion for Language Model Continual\n  Learning", "abstract": "  Language model continual learning (CL) has recently attracted significant\ninterest for its ability to adapt large language models (LLMs) to dynamic\nreal-world scenarios without retraining. A major challenge in this domain is\ncatastrophic forgetting, where models lose previously acquired knowledge upon\nlearning new tasks. Existing approaches commonly utilize multiple\nparameter-efficient fine-tuning (PEFT) blocks to acquire task-specific\nknowledge, yet these methods are inefficient and fail to leverage potential\nknowledge transfer across tasks. In this paper, we introduce a novel CL\nframework for language models, named Knowledge Identification and Fusion (KIF),\nwhich boosts knowledge transfer without depending on memory replay. KIF\ninitially segregates the model into 'skill units' based on parameter\ndependencies, allowing for more precise control. Subsequently, it employs a\nnovel group-wise knowledge identification technique to ascertain the importance\ndistribution of skill units for a new task. By comparing this importance\ndistribution with those from previous tasks, we implement a fine-grained\nknowledge fusion strategy that retains task-specific knowledge, thereby\npreventing forgetting, and updates task-shared knowledge, which facilitates\nbi-directional knowledge transfer. As a result, KIF achieves an optimal balance\nbetween retaining prior knowledge and excelling in new tasks. KIF also\ndemonstrates strong generalizability, making it suitable for various base\nmodels and adaptable to PEFT methods like LoRA. Furthermore, it offers notable\nextensibility, supporting enhancements through integration with memory replay\ntechniques. Comprehensive experiments conducted on two CL benchmarks, involving\nmodels ranging from 220M to 7B parameters, affirm the effectiveness of KIF and\nits variants across different settings.\n", "date": "2024-08-09", "categories": []}
{"id": "2501.13453", "title": "Spurious Forgetting in Continual Learning of Language Models", "abstract": "  Recent advancements in large language models (LLMs) reveal a perplexing\nphenomenon in continual learning: despite extensive training, models experience\nsignificant performance declines, raising questions about task alignment and\nunderlying knowledge retention. This study first explores the concept of\n\"spurious forgetting\", proposing that such performance drops often reflect a\ndecline in task alignment rather than true knowledge loss. Through controlled\nexperiments with a synthesized dataset, we investigate the dynamics of model\nperformance during the initial training phases of new tasks, discovering that\nearly optimization steps can disrupt previously established task alignments.\nOur theoretical analysis connects these shifts to orthogonal updates in model\nweights, providing a robust framework for understanding this behavior.\nUltimately, we introduce a Freezing strategy that fix the bottom layers of the\nmodel, leading to substantial improvements in four continual learning\nscenarios. Our findings underscore the critical distinction between task\nalignment and knowledge retention, paving the way for more effective strategies\nin continual learning.\n", "date": "2025-01-23", "categories": []}
{"id": "2409.13265", "title": "Towards LifeSpan Cognitive Systems", "abstract": "  Building a human-like system that continuously interacts with complex\nenvironments -- whether simulated digital worlds or human society -- presents\nseveral key challenges. Central to this is enabling continuous, high-frequency\ninteractions, where the interactions are termed experiences. We refer to this\nenvisioned system as the LifeSpan Cognitive System (LSCS). A critical feature\nof LSCS is its ability to engage in incremental and rapid updates while\nretaining and accurately recalling past experiences. In this paper we focus on\nthe domain of Large Language Models (LLMs), where we identify two major\nchallenges: (1) Abstraction and Experience Merging, and (2) Long-term Retention\nwith Accurate Recall. These properties are essential for storing new\nexperiences, organizing past experiences, and responding to the environment in\nways that leverage relevant historical data. Unlike language models with\ncontinual learning, which typically rely on large corpora for fine-tuning and\nfocus on improving performance within specific domains or tasks, LSCS must\nrapidly and incrementally update with new information from its environment at a\nhigh frequency. Existing technologies with the potential of solving the above\ntwo major challenges can be classified into four classes based on a conceptual\nmetric called Storage Complexity, which measures the relative space required to\nstore past experiences. Each of these four classes of technologies has its own\nstrengths and limitations while we argue none of them alone can achieve LSCS\nalone. To this end, we propose a potential instantiation for LSCS that can\nintegrate all four classes of technologies. The new instantiation, serving as a\nconjecture, operates through two core processes: Absorbing Experiences and\nGenerating Responses.\n", "date": "2024-09-20", "categories": []}
{"id": "2410.22658", "title": "Incremental Learning of Retrievable Skills For Efficient Continual Task\n  Adaptation", "abstract": "  Continual Imitation Learning (CiL) involves extracting and accumulating task\nknowledge from demonstrations across multiple stages and tasks to achieve a\nmulti-task policy. With recent advancements in foundation models, there has\nbeen a growing interest in adapter-based CiL approaches, where adapters are\nestablished parameter-efficiently for tasks newly demonstrated. While these\napproaches isolate parameters for specific tasks and tend to mitigate\ncatastrophic forgetting, they limit knowledge sharing among different\ndemonstrations. We introduce IsCiL, an adapter-based CiL framework that\naddresses this limitation of knowledge sharing by incrementally learning\nshareable skills from different demonstrations, thus enabling sample-efficient\ntask adaptation using the skills particularly in non-stationary CiL\nenvironments. In IsCiL, demonstrations are mapped into the state embedding\nspace, where proper skills can be retrieved upon input states through\nprototype-based memory. These retrievable skills are incrementally learned on\ntheir corresponding adapters. Our CiL experiments with complex tasks in\nFranka-Kitchen and Meta-World demonstrate robust performance of IsCiL in both\ntask adaptation and sample-efficiency. We also show a simple extension of IsCiL\nfor task unlearning scenarios.\n", "date": "2024-10-30", "categories": []}
{"id": "2412.08285", "title": "Adaptive Prompting for Continual Relation Extraction: A Within-Task\n  Variance Perspective", "abstract": "  To address catastrophic forgetting in Continual Relation Extraction (CRE),\nmany current approaches rely on memory buffers to rehearse previously learned\nknowledge while acquiring new tasks. Recently, prompt-based methods have\nemerged as potent alternatives to rehearsal-based strategies, demonstrating\nstrong empirical performance. However, upon analyzing existing prompt-based\napproaches for CRE, we identified several critical limitations, such as\ninaccurate prompt selection, inadequate mechanisms for mitigating forgetting in\nshared parameters, and suboptimal handling of cross-task and within-task\nvariances. To overcome these challenges, we draw inspiration from the\nrelationship between prefix-tuning and mixture of experts, proposing a novel\napproach that employs a prompt pool for each task, capturing variations within\neach task while enhancing cross-task variances. Furthermore, we incorporate a\ngenerative model to consolidate prior knowledge within shared parameters,\neliminating the need for explicit data storage. Extensive experiments validate\nthe efficacy of our approach, demonstrating superior performance over\nstate-of-the-art prompt-based and rehearsal-free methods in continual relation\nextraction.\n", "date": "2024-12-11", "categories": []}
{"id": "2501.09525", "title": "Class Incremental Fault Diagnosis under Limited Fault Data via\n  Supervised Contrastive Knowledge Distillation", "abstract": "  Class-incremental fault diagnosis requires a model to adapt to new fault\nclasses while retaining previous knowledge. However, limited research exists\nfor imbalanced and long-tailed data. Extracting discriminative features from\nfew-shot fault data is challenging, and adding new fault classes often demands\ncostly model retraining. Moreover, incremental training of existing methods\nrisks catastrophic forgetting, and severe class imbalance can bias the model's\ndecisions toward normal classes. To tackle these issues, we introduce a\nSupervised Contrastive knowledge distiLlation for class Incremental Fault\nDiagnosis (SCLIFD) framework proposing supervised contrastive knowledge\ndistillation for improved representation learning capability and less\nforgetting, a novel prioritized exemplar selection method for sample replay to\nalleviate catastrophic forgetting, and the Random Forest Classifier to address\nthe class imbalance. Extensive experimentation on simulated and real-world\nindustrial datasets across various imbalance ratios demonstrates the\nsuperiority of SCLIFD over existing approaches. Our code can be found at\nhttps://github.com/Zhang-Henry/SCLIFD_TII.\n", "date": "2025-01-16", "categories": []}
{"id": "2501.10395", "title": "Towards General Purpose Robots at Scale: Lifelong Learning and Learning\n  to Use Memory", "abstract": "  The widespread success of artificial intelligence in fields like natural\nlanguage processing and computer vision has not yet fully transferred to\nrobotics, where progress is hindered by the lack of large-scale training data\nand the complexity of real-world tasks. To address this, many robot learning\nresearchers are pushing to get robots deployed at scale in everyday\nunstructured environments like our homes to initiate a data flywheel. While\ncurrent robot learning systems are effective for certain short-horizon tasks,\nthey are not designed to autonomously operate over long time horizons in\nunstructured environments. This thesis focuses on addressing two key challenges\nfor robots operating over long time horizons: memory and lifelong learning.\n  We propose two novel methods to advance these capabilities. First, we\nintroduce t-DGR, a trajectory-based deep generative replay method that achieves\nstate-of-the-art performance on Continual World benchmarks, advancing lifelong\nlearning. Second, we develop a framework that leverages human demonstrations to\nteach agents effective memory utilization, improving learning efficiency and\nsuccess rates on Memory Gym tasks. Finally, we discuss future directions for\nachieving the lifelong learning and memory capabilities necessary for robots to\nfunction at scale in real-world settings.\n", "date": "2024-12-28", "categories": []}
{"id": "2501.10861", "title": "Dynamic Continual Learning: Harnessing Parameter Uncertainty for\n  Improved Network Adaptation", "abstract": "  When fine-tuning Deep Neural Networks (DNNs) to new data, DNNs are prone to\noverwriting network parameters required for task-specific functionality on\npreviously learned tasks, resulting in a loss of performance on those tasks. We\npropose using parameter-based uncertainty to determine which parameters are\nrelevant to a network's learned function and regularize training to prevent\nchange in these important parameters. We approach this regularization in two\nways: (1), we constrain critical parameters from significant changes by\nassociating more critical parameters with lower learning rates, thereby\nlimiting alterations in those parameters; (2), important parameters are\nrestricted from change by imposing a higher regularization weighting, causing\nparameters to revert to their states prior to the learning of subsequent tasks.\nWe leverage a Bayesian Moment Propagation framework which learns network\nparameters concurrently with their associated uncertainties while allowing each\nparameter to contribute uncertainty to the network's predictive distribution,\navoiding the pitfalls of existing sampling-based methods. The proposed approach\nis evaluated for common sequential benchmark datasets and compared to existing\npublished approaches from the Continual Learning community. Ultimately, we show\nimproved Continual Learning performance for Average Test Accuracy and Backward\nTransfer metrics compared to sampling-based methods and other\nnon-uncertainty-based approaches.\n", "date": "2025-01-18", "categories": []}
{"id": "2501.11140", "title": "CLOFAI: A Dataset of Real And Fake Image Classification Tasks for\n  Continual Learning", "abstract": "  The rapid advancement of generative AI models capable of creating realistic\nmedia has led to a need for classifiers that can accurately distinguish between\ngenuine and artificially-generated images. A significant challenge for these\nclassifiers emerges when they encounter images from generative models that are\nnot represented in their training data, usually resulting in diminished\nperformance. A typical approach is to periodically update the classifier's\ntraining data with images from the new generative models then retrain the\nclassifier on the updated dataset. However, in some real-life scenarios,\nstorage, computational, or privacy constraints render this approach\nimpractical. Additionally, models used in security applications may be required\nto rapidly adapt. In these circumstances, continual learning provides a\npromising alternative, as the classifier can be updated without retraining on\nthe entire dataset. In this paper, we introduce a new dataset called CLOFAI\n(Continual Learning On Fake and Authentic Images), which takes the form of a\ndomain-incremental image classification problem. Moreover, we showcase the\napplicability of this dataset as a benchmark for evaluating continual learning\nmethodologies. In doing this, we set a baseline on our novel dataset using\nthree foundational continual learning methods -- EWC, GEM, and Experience\nReplay -- and find that EWC performs poorly, while GEM and Experience Replay\nshow promise, performing significantly better than a Naive baseline. The\ndataset and code to run the experiments can be accessed from the following\nGitHub repository: https://github.com/Will-Doherty/CLOFAI.\n", "date": "2025-01-19", "categories": []}
{"id": "2410.05820", "title": "IncSAR: A Dual Fusion Incremental Learning Framework for SAR Target\n  Recognition", "abstract": "  Deep learning techniques have achieved significant success in Synthetic\nAperture Radar (SAR) target recognition using predefined datasets in static\nscenarios. However, real-world applications demand that models incrementally\nlearn new information without forgetting previously acquired knowledge. The\nchallenge of catastrophic forgetting, where models lose past knowledge when\nadapting to new tasks, remains a critical issue. In this paper, we introduce\nIncSAR, an incremental learning framework designed to tackle catastrophic\nforgetting in SAR target recognition. IncSAR combines the power of a Vision\nTransformer (ViT) and a custom-designed Convolutional Neural Network (CNN) in a\ndual-branch architecture, integrated via a late-fusion strategy. Additionally,\nwe explore the use of TinyViT to reduce computational complexity and propose an\nattention mechanism to dynamically enhance feature representation. To mitigate\nthe speckle noise inherent in SAR images, we employ a denoising module based on\na neural network approximation of Robust Principal Component Analysis (RPCA),\nleveraging a simple neural network for efficient noise reduction in SAR\nimagery. Moreover, a random projection layer improves the linear separability\nof features, and a variant of Linear Discriminant Analysis (LDA) decorrelates\nextracted class prototypes for better generalization. Extensive experiments on\nthe MSTAR, SAR-AIRcraft-1.0, and OpenSARShip benchmark datasets demonstrate\nthat IncSAR significantly outperforms state-of-the-art approaches, achieving a\n99.63\\% average accuracy and a 0.33\\% performance drop, representing an 89\\%\nimprovement in retention compared to existing techniques. The source code is\navailable at https://github.com/geokarant/IncSAR.\n", "date": "2024-10-08", "categories": []}
{"id": "2501.09352", "title": "PAL: Prompting Analytic Learning with Missing Modality for Multi-Modal\n  Class-Incremental Learning", "abstract": "  Multi-modal class-incremental learning (MMCIL) seeks to leverage multi-modal\ndata, such as audio-visual and image-text pairs, thereby enabling models to\nlearn continuously across a sequence of tasks while mitigating forgetting.\nWhile existing studies primarily focus on the integration and utilization of\nmulti-modal information for MMCIL, a critical challenge remains: the issue of\nmissing modalities during incremental learning phases. This oversight can\nexacerbate severe forgetting and significantly impair model performance. To\nbridge this gap, we propose PAL, a novel exemplar-free framework tailored to\nMMCIL under missing-modality scenarios. Concretely, we devise modality-specific\nprompts to compensate for missing information, facilitating the model to\nmaintain a holistic representation of the data. On this foundation, we\nreformulate the MMCIL problem into a Recursive Least-Squares task, delivering\nan analytical linear solution. Building upon these, PAL not only alleviates the\ninherent under-fitting limitation in analytic learning but also preserves the\nholistic representation of missing-modality data, achieving superior\nperformance with less forgetting across various multi-modal incremental\nscenarios. Extensive experiments demonstrate that PAL significantly outperforms\ncompetitive methods across various datasets, including UPMC-Food101 and\nN24News, showcasing its robustness towards modality absence and its\nanti-forgetting ability to maintain high incremental accuracy.\n", "date": "2025-01-16", "categories": []}
{"id": "2501.09361", "title": "Strategic Base Representation Learning via Feature Augmentations for\n  Few-Shot Class Incremental Learning", "abstract": "  Few-shot class incremental learning implies the model to learn new classes\nwhile retaining knowledge of previously learned classes with a small number of\ntraining instances. Existing frameworks typically freeze the parameters of the\npreviously learned classes during the incorporation of new classes. However,\nthis approach often results in suboptimal class separation of previously\nlearned classes, leading to overlap between old and new classes. Consequently,\nthe performance of old classes degrades on new classes. To address these\nchallenges, we propose a novel feature augmentation driven contrastive learning\nframework designed to enhance the separation of previously learned classes to\naccommodate new classes. Our approach involves augmenting feature vectors and\nassigning proxy labels to these vectors. This strategy expands the feature\nspace, ensuring seamless integration of new classes within the expanded space.\nAdditionally, we employ a self-supervised contrastive loss to improve the\nseparation between previous classes. We validate our framework through\nexperiments on three FSCIL benchmark datasets: CIFAR100, miniImageNet, and\nCUB200. The results demonstrate that our Feature Augmentation driven\nContrastive Learning framework significantly outperforms other approaches,\nachieving state-of-the-art performance.\n", "date": "2025-01-16", "categories": []}
{"id": "2409.02512", "title": "Continual Diffuser (CoD): Mastering Continual Offline Reinforcement\n  Learning with Experience Rehearsal", "abstract": "  Artificial neural networks, especially recent diffusion-based models, have\nshown remarkable superiority in gaming, control, and QA systems, where the\ntraining tasks' datasets are usually static. However, in real-world\napplications, such as robotic control of reinforcement learning (RL), the tasks\nare changing, and new tasks arise in a sequential order. This situation poses\nthe new challenge of plasticity-stability trade-off for training an agent who\ncan adapt to task changes and retain acquired knowledge. In view of this, we\npropose a rehearsal-based continual diffusion model, called Continual Diffuser\n(CoD), to endow the diffuser with the capabilities of quick adaptation\n(plasticity) and lasting retention (stability). Specifically, we first\nconstruct an offline benchmark that contains 90 tasks from multiple domains.\nThen, we train the CoD on each task with sequential modeling and conditional\ngeneration for making decisions. Next, we preserve a small portion of previous\ndatasets as the rehearsal buffer and replay it to retain the acquired\nknowledge. Extensive experiments on a series of tasks show CoD can achieve a\npromising plasticity-stability trade-off and outperform existing\ndiffusion-based methods and other representative baselines on most tasks.\n", "date": "2024-09-04", "categories": []}
{"id": "2312.16409", "title": "Dynamic Sub-graph Distillation for Robust Semi-supervised Continual\n  Learning", "abstract": "  Continual learning (CL) has shown promising results and comparable\nperformance to learning at once in a fully supervised manner. However, CL\nstrategies typically require a large number of labeled samples, making their\nreal-life deployment challenging. In this work, we focus on semi-supervised\ncontinual learning (SSCL), where the model progressively learns from partially\nlabeled data with unknown categories. We provide a comprehensive analysis of\nSSCL and demonstrate that unreliable distributions of unlabeled data lead to\nunstable training and refinement of the progressing stages. This problem\nseverely impacts the performance of SSCL. To address the limitations, we\npropose a novel approach called Dynamic Sub-Graph Distillation (DSGD) for\nsemi-supervised continual learning, which leverages both semantic and\nstructural information to achieve more stable knowledge distillation on\nunlabeled data and exhibit robustness against distribution bias. Firstly, we\nformalize a general model of structural distillation and design a dynamic graph\nconstruction for the continual learning progress. Next, we define a structure\ndistillation vector and design a dynamic sub-graph distillation algorithm,\nwhich enables end-to-end training and adaptability to scale up tasks. The\nentire proposed method is adaptable to various CL methods and supervision\nsettings. Finally, experiments conducted on three datasets CIFAR10, CIFAR100,\nand ImageNet-100, with varying supervision ratios, demonstrate the\neffectiveness of our proposed approach in mitigating the catastrophic\nforgetting problem in semi-supervised continual learning scenarios.\n", "date": "2023-12-27", "categories": []}
{"id": "2501.08245", "title": "Continual Deep Active Learning for Medical Imaging: Replay-Base\n  Architecture for Context Adaptation", "abstract": "  Deep Learning for medical imaging faces challenges in adapting and\ngeneralizing to new contexts. Additionally, it often lacks sufficient labeled\ndata for specific tasks requiring significant annotation effort. Continual\nLearning (CL) tackles adaptability and generalizability by enabling lifelong\nlearning from a data stream while mitigating forgetting of previously learned\nknowledge. Active Learning (AL) reduces the number of required annotations for\neffective training. This work explores both approaches (CAL) to develop a novel\nframework for robust medical image analysis. Based on the automatic recognition\nof shifts in image characteristics, Replay-Base Architecture for Context\nAdaptation (RBACA) employs a CL rehearsal method to continually learn from\ndiverse contexts, and an AL component to select the most informative instances\nfor annotation. A novel approach to evaluate CAL methods is established using a\ndefined metric denominated IL-Score, which allows for the simultaneous\nassessment of transfer learning, forgetting, and final model performance. We\nshow that RBACA works in domain and class-incremental learning scenarios, by\nassessing its IL-Score on the segmentation and diagnosis of cardiac images. The\nresults show that RBACA outperforms a baseline framework without CAL, and a\nstate-of-the-art CAL method across various memory sizes and annotation budgets.\nOur code is available in https://github.com/RuiDaniel/RBACA .\n", "date": "2025-01-14", "categories": []}
{"id": "2403.15517", "title": "Improving Forward Compatibility in Class Incremental Learning by\n  Increasing Representation Rank and Feature Richness", "abstract": "  Class Incremental Learning (CIL) constitutes a pivotal subfield within\ncontinual learning, aimed at enabling models to progressively learn new\nclassification tasks while retaining knowledge obtained from prior tasks.\nAlthough previous studies have predominantly focused on backward compatible\napproaches to mitigate catastrophic forgetting, recent investigations have\nintroduced forward compatible methods to enhance performance on novel tasks and\ncomplement existing backward compatible methods. In this study, we introduce an\neffective-Rank based Feature Richness enhancement (RFR) method, designed for\nimproving forward compatibility. Specifically, this method increases the\neffective rank of representations during the base session, thereby facilitating\nthe incorporation of more informative features pertinent to unseen novel tasks.\nConsequently, RFR achieves dual objectives in backward and forward\ncompatibility: minimizing feature extractor modifications and enhancing novel\ntask performance, respectively. To validate the efficacy of our approach, we\nestablish a theoretical connection between effective rank and the Shannon\nentropy of representations. Subsequently, we conduct comprehensive experiments\nby integrating RFR into eleven well-known CIL methods. Our results demonstrate\nthe effectiveness of our approach in enhancing novel-task performance while\nmitigating catastrophic forgetting. Furthermore, our method notably improves\nthe average incremental accuracy across all eleven cases examined.\n", "date": "2024-03-22", "categories": []}
{"id": "2501.02198", "title": "Fresh-CL: Feature Realignment through Experts on Hypersphere in\n  Continual Learning", "abstract": "  Continual Learning enables models to learn and adapt to new tasks while\nretaining prior knowledge. Introducing new tasks, however, can naturally lead\nto feature entanglement across tasks, limiting the model's capability to\ndistinguish between new domain data. In this work, we propose a method called\nFeature Realignment through Experts on hyperSpHere in Continual Learning\n(Fresh-CL). By leveraging predefined and fixed simplex equiangular tight frame\n(ETF) classifiers on a hypersphere, our model improves feature separation both\nintra and inter tasks. However, the projection to a simplex ETF shifts with new\ntasks, disrupting structured feature representation of previous tasks and\ndegrading performance. Therefore, we propose a dynamic extension of ETF through\nmixture of experts, enabling adaptive projections onto diverse subspaces to\nenhance feature representation. Experiments on 11 datasets demonstrate a 2%\nimprovement in accuracy compared to the strongest baseline, particularly in\nfine-grained datasets, confirming the efficacy of combining ETF and MoE to\nimprove feature distinction in continual learning scenarios.\n", "date": "2025-01-04", "categories": []}
{"id": "2501.07236", "title": "CSTA: Spatial-Temporal Causal Adaptive Learning for Exemplar-Free Video\n  Class-Incremental Learning", "abstract": "  Continual learning aims to acquire new knowledge while retaining past\ninformation. Class-incremental learning (CIL) presents a challenging scenario\nwhere classes are introduced sequentially. For video data, the task becomes\nmore complex than image data because it requires learning and preserving both\nspatial appearance and temporal action involvement. To address this challenge,\nwe propose a novel exemplar-free framework that equips separate spatiotemporal\nadapters to learn new class patterns, accommodating the incremental information\nrepresentation requirements unique to each class. While separate adapters are\nproven to mitigate forgetting and fit unique requirements, naively applying\nthem hinders the intrinsic connection between spatial and temporal information\nincrements, affecting the efficiency of representing newly learned class\ninformation. Motivated by this, we introduce two key innovations from a causal\nperspective. First, a causal distillation module is devised to maintain the\nrelation between spatial-temporal knowledge for a more efficient\nrepresentation. Second, a causal compensation mechanism is proposed to reduce\nthe conflicts during increment and memorization between different types of\ninformation. Extensive experiments conducted on benchmark datasets demonstrate\nthat our framework can achieve new state-of-the-art results, surpassing current\nexample-based methods by 4.2% in accuracy on average.\n", "date": "2025-01-13", "categories": []}
{"id": "2501.07382", "title": "Information-Theoretic Dual Memory System for Continual Learning", "abstract": "  Continuously acquiring new knowledge from a dynamic environment is a\nfundamental capability for animals, facilitating their survival and ability to\naddress various challenges. This capability is referred to as continual\nlearning, which focuses on the ability to learn a sequence of tasks without the\ndetriment of previous knowledge. A prevalent strategy to tackle continual\nlearning involves selecting and storing numerous essential data samples from\nprior tasks within a fixed-size memory buffer. However, the majority of current\nmemory-based techniques typically utilize a single memory buffer, which poses\nchallenges in concurrently managing newly acquired and previously learned\nsamples. Drawing inspiration from the Complementary Learning Systems (CLS)\ntheory, which defines rapid and gradual learning mechanisms for processing\ninformation, we propose an innovative dual memory system called the\nInformation-Theoretic Dual Memory System (ITDMS). This system comprises a fast\nmemory buffer designed to retain temporary and novel samples, alongside a slow\nmemory buffer dedicated to preserving critical and informative samples. The\nfast memory buffer is optimized employing an efficient reservoir sampling\nprocess. Furthermore, we introduce a novel information-theoretic memory\noptimization strategy that selectively identifies and retains diverse and\ninformative data samples for the slow memory buffer. Additionally, we propose a\nnovel balanced sample selection procedure that automatically identifies and\neliminates redundant memorized samples, thus freeing up memory capacity for new\ndata acquisitions, which can deal with a growing array of tasks. Our\nmethodology is rigorously assessed through a series of continual learning\nexperiments, with empirical results underscoring the effectiveness of the\nproposed system.\n", "date": "2025-01-13", "categories": []}
{"id": "2501.05495", "title": "LSEBMCL: A Latent Space Energy-Based Model for Continual Learning", "abstract": "  Continual learning has become essential in many practical applications such\nas online news summaries and product classification. The primary challenge is\nknown as catastrophic forgetting, a phenomenon where a model inadvertently\ndiscards previously learned knowledge when it is trained on new tasks. Existing\nsolutions involve storing exemplars from previous classes, regularizing\nparameters during the fine-tuning process, or assigning different model\nparameters to each task. The proposed solution LSEBMCL (Latent Space\nEnergy-Based Model for Continual Learning) in this work is to use energy-based\nmodels (EBMs) to prevent catastrophic forgetting by sampling data points from\nprevious tasks when training on new ones. The EBM is a machine learning model\nthat associates an energy value with each input data point. The proposed method\nuses an EBM layer as an outer-generator in the continual learning framework for\nNLP tasks. The study demonstrates the efficacy of EBM in NLP tasks, achieving\nstate-of-the-art results in all experiments.\n", "date": "2025-01-09", "categories": []}
{"id": "2501.05559", "title": "Soup to go: mitigating forgetting during continual learning with model\n  averaging", "abstract": "  In continual learning, where task data arrives in a sequence, fine-tuning on\nlater tasks will often lead to performance degradation on earlier tasks. This\nis especially pronounced when these tasks come from diverse domains. In this\nsetting, how can we mitigate catastrophic forgetting of earlier tasks and\nretain what the model has learned with minimal computational expenses? Inspired\nby other merging methods, and L2-regression, we propose Sequential Fine-tuning\nwith Averaging (SFA), a method that merges currently training models with\nearlier checkpoints during the course of training. SOTA approaches typically\nmaintain a data buffer of past tasks or impose a penalty at each gradient step.\nIn contrast, our method achieves comparable results without the need to store\npast data, or multiple copies of parameters for each gradient step.\nFurthermore, our method outperforms common merging techniques such as Task\nArithmetic, TIES Merging, and WiSE-FT, as well as other penalty methods like L2\nand Elastic Weight Consolidation. In turn, our method offers insight into the\nbenefits of merging partially-trained models during training across both image\nand language domains.\n", "date": "2025-01-09", "categories": []}
{"id": "2501.05862", "title": "Language-Inspired Relation Transfer for Few-shot Class-Incremental\n  Learning", "abstract": "  Depicting novel classes with language descriptions by observing few-shot\nsamples is inherent in human-learning systems. This lifelong learning\ncapability helps to distinguish new knowledge from old ones through the\nincrease of open-world learning, namely Few-Shot Class-Incremental Learning\n(FSCIL). Existing works to solve this problem mainly rely on the careful tuning\nof visual encoders, which shows an evident trade-off between the base knowledge\nand incremental ones. Motivated by human learning systems, we propose a new\nLanguage-inspired Relation Transfer (LRT) paradigm to understand objects by\njoint visual clues and text depictions, composed of two major steps. We first\ntransfer the pretrained text knowledge to the visual domains by proposing a\ngraph relation transformation module and then fuse the visual and language\nembedding by a text-vision prototypical fusion module. Second, to mitigate the\ndomain gap caused by visual finetuning, we propose context prompt learning for\nfast domain alignment and imagined contrastive learning to alleviate the\ninsufficient text data during alignment. With collaborative learning of domain\nalignments and text-image transfer, our proposed LRT outperforms the\nstate-of-the-art models by over $13\\%$ and $7\\%$ on the final session of\nmini-ImageNet and CIFAR-100 FSCIL benchmarks.\n", "date": "2025-01-10", "categories": []}
{"id": "2308.06764", "title": "Few-shot Class-incremental Learning for Classification and Object\n  Detection: A Survey", "abstract": "  Few-shot Class-Incremental Learning (FSCIL) presents a unique challenge in\nMachine Learning (ML), as it necessitates the Incremental Learning (IL) of new\nclasses from sparsely labeled training samples without forgetting previous\nknowledge. While this field has seen recent progress, it remains an active\nexploration area. This paper aims to provide a comprehensive and systematic\nreview of FSCIL. In our in-depth examination, we delve into various facets of\nFSCIL, encompassing the problem definition, the discussion of the primary\nchallenges of unreliable empirical risk minimization and the\nstability-plasticity dilemma, general schemes, and relevant problems of IL and\nFew-shot Learning (FSL). Besides, we offer an overview of benchmark datasets\nand evaluation metrics. Furthermore, we introduce the Few-shot\nClass-incremental Classification (FSCIC) methods from data-based,\nstructure-based, and optimization-based approaches and the Few-shot\nClass-incremental Object Detection (FSCIOD) methods from anchor-free and\nanchor-based approaches. Beyond these, we present several promising research\ndirections within FSCIL that merit further investigation.\n", "date": "2023-08-13", "categories": []}
{"id": "2409.07387", "title": "A Contrastive Symmetric Forward-Forward Algorithm (SFFA) for Continual\n  Learning Tasks", "abstract": "  The so-called Forward-Forward Algorithm (FFA) has recently gained momentum as\nan alternative to the conventional back-propagation algorithm for neural\nnetwork learning, yielding competitive performance across various modeling\ntasks. By replacing the backward pass of gradient back-propagation with two\ncontrastive forward passes, the FFA avoids several shortcomings undergone by\nits predecessor (e.g., vanishing/exploding gradient) by enabling layer-wise\ntraining heuristics. In classification tasks, this contrastive method has been\nproven to effectively create a latent sparse representation of the input data,\nultimately favoring discriminability. However, FFA exhibits an inherent\nasymmetric gradient behavior due to an imbalanced loss function between\npositive and negative data, adversely impacting on the model's generalization\ncapabilities and leading to an accuracy degradation. To address this issue,\nthis work proposes the Symmetric Forward-Forward Algorithm (SFFA), a novel\nmodification of the original FFA which partitions each layer into positive and\nnegative neurons. This allows the local fitness function to be defined as the\nratio between the activation of positive neurons and the overall layer\nactivity, resulting in a symmetric loss landscape during the training phase. To\nevaluate the enhanced convergence of our method, we conduct several experiments\nusing multiple image classification benchmarks, comparing the accuracy of\nmodels trained with SFFA to those trained with its FFA counterpart. As a\nbyproduct of this reformulation, we explore the advantages of using a\nlayer-wise training algorithm for Continual Learning (CL) tasks. The\nspecialization of neurons and the sparsity of their activations induced by\nlayer-wise training algorithms enable efficient CL strategies that incorporate\nnew knowledge (classes) into the neural network, while preventing catastrophic\nforgetting of previously...\n", "date": "2024-09-11", "categories": []}
{"id": "2501.04897", "title": "Online Continual Learning: A Systematic Literature Review of Approaches,\n  Challenges, and Benchmarks", "abstract": "  Online Continual Learning (OCL) is a critical area in machine learning,\nfocusing on enabling models to adapt to evolving data streams in real-time\nwhile addressing challenges such as catastrophic forgetting and the\nstability-plasticity trade-off. This study conducts the first comprehensive\nSystematic Literature Review (SLR) on OCL, analyzing 81 approaches, extracting\nover 1,000 features (specific tasks addressed by these approaches), and\nidentifying more than 500 components (sub-models within approaches, including\nalgorithms and tools). We also review 83 datasets spanning applications like\nimage classification, object detection, and multimodal vision-language tasks.\nOur findings highlight key challenges, including reducing computational\noverhead, developing domain-agnostic solutions, and improving scalability in\nresource-constrained environments. Furthermore, we identify promising\ndirections for future research, such as leveraging self-supervised learning for\nmultimodal and sequential data, designing adaptive memory mechanisms that\nintegrate sparse retrieval and generative replay, and creating efficient\nframeworks for real-world applications with noisy or evolving task boundaries.\nBy providing a rigorous and structured synthesis of the current state of OCL,\nthis review offers a valuable resource for advancing this field and addressing\nits critical challenges and opportunities. The complete SLR methodology steps\nand extracted data are publicly available through the provided link:\nhttps://github.com/kiyan-rezaee/\nSystematic-Literature-Review-on-Online-Continual-Learning\n", "date": "2025-01-09", "categories": []}
{"id": "2501.05089", "title": "Supervised Learning with Evolving Tasks and Performance Guarantees", "abstract": "  Multiple supervised learning scenarios are composed by a sequence of\nclassification tasks. For instance, multi-task learning and continual learning\naim to learn a sequence of tasks that is either fixed or grows over time.\nExisting techniques for learning tasks that are in a sequence are tailored to\nspecific scenarios, lacking adaptability to others. In addition, most of\nexisting techniques consider situations in which the order of the tasks in the\nsequence is not relevant. However, it is common that tasks in a sequence are\nevolving in the sense that consecutive tasks often have a higher similarity.\nThis paper presents a learning methodology that is applicable to multiple\nsupervised learning scenarios and adapts to evolving tasks. Differently from\nexisting techniques, we provide computable tight performance guarantees and\nanalytically characterize the increase in the effective sample size.\nExperiments on benchmark datasets show the performance improvement of the\nproposed methodology in multiple scenarios and the reliability of the presented\nperformance guarantees.\n", "date": "2025-01-09", "categories": []}
{"id": "2501.05246", "title": "Domain-Incremental Semantic Segmentation for Autonomous Driving under\n  Adverse Driving Conditions", "abstract": "  Semantic segmentation for autonomous driving is an even more challenging task\nwhen faced with adverse driving conditions. Standard models trained on data\nrecorded under ideal conditions show a deteriorated performance in unfavorable\nweather or illumination conditions. Fine-tuning on the new task or condition\nwould lead to overwriting the previously learned information resulting in\ncatastrophic forgetting. Adapting to the new conditions through traditional\ndomain adaption methods improves the performance on the target domain at the\nexpense of the source domain. Addressing these issues, we propose an\narchitecture-based domain-incremental learning approach called Progressive\nSemantic Segmentation (PSS). PSS is a task-agnostic, dynamically growing\ncollection of domain-specific segmentation models. The task of inferring the\ndomain and subsequently selecting the appropriate module for segmentation is\ncarried out using a collection of convolutional autoencoders. We extensively\nevaluate our proposed approach using several datasets at varying levels of\ngranularity in the categorization of adverse driving conditions. Furthermore,\nwe demonstrate the generalization of the proposed approach to similar and\nunseen domains.\n", "date": "2025-01-09", "categories": []}
{"id": "2301.01828", "title": "On Sequential Bayesian Inference for Continual Learning", "abstract": "  Sequential Bayesian inference can be used for continual learning to prevent\ncatastrophic forgetting of past tasks and provide an informative prior when\nlearning new tasks. We revisit sequential Bayesian inference and test whether\nhaving access to the true posterior is guaranteed to prevent catastrophic\nforgetting in Bayesian neural networks. To do this we perform sequential\nBayesian inference using Hamiltonian Monte Carlo. We propagate the posterior as\na prior for new tasks by fitting a density estimator on Hamiltonian Monte Carlo\nsamples. We find that this approach fails to prevent catastrophic forgetting\ndemonstrating the difficulty in performing sequential Bayesian inference in\nneural networks. From there we study simple analytical examples of sequential\nBayesian inference and CL and highlight the issue of model misspecification\nwhich can lead to sub-optimal continual learning performance despite exact\ninference. Furthermore, we discuss how task data imbalances can cause\nforgetting. From these limitations, we argue that we need probabilistic models\nof the continual learning generative process rather than relying on sequential\nBayesian inference over Bayesian neural network weights. In this vein, we also\npropose a simple baseline called Prototypical Bayesian Continual Learning,\nwhich is competitive with state-of-the-art Bayesian continual learning methods\non class incremental continual learning vision benchmarks.\n", "date": "2023-01-04", "categories": []}
{"id": "2312.17051", "title": "FILP-3D: Enhancing 3D Few-shot Class-incremental Learning with\n  Pre-trained Vision-Language Models", "abstract": "  Few-shot class-incremental learning (FSCIL) aims to mitigate the catastrophic\nforgetting issue when a model is incrementally trained on limited data.\nHowever, many of these works lack effective exploration of prior knowledge,\nrendering them unable to effectively address the domain gap issue in the\ncontext of 3D FSCIL, thereby leading to catastrophic forgetting. The\nContrastive Vision-Language Pre-Training (CLIP) model serves as a highly\nsuitable backbone for addressing the challenges of 3D FSCIL due to its abundant\nshape-related prior knowledge. Unfortunately, its direct application to 3D\nFSCIL still faces the incompatibility between 3D data representation and the 2D\nfeatures, primarily manifested as feature space misalignment and significant\nnoise. To address the above challenges, we introduce the FILP-3D framework with\ntwo novel components: the Redundant Feature Eliminator (RFE) for feature space\nmisalignment and the Spatial Noise Compensator (SNC) for significant noise. RFE\naligns the feature spaces of input point clouds and their embeddings by\nperforming a unique dimensionality reduction on the feature space of\npre-trained models (PTMs), effectively eliminating redundant information\nwithout compromising semantic integrity. On the other hand, SNC is a\ngraph-based 3D model designed to capture robust geometric information within\npoint clouds, thereby augmenting the knowledge lost due to projection,\nparticularly when processing real-world scanned data. Moreover, traditional\naccuracy metrics are proven to be biased due to the imbalance in existing 3D\ndatasets. Therefore we propose 3D FSCIL benchmark FSCIL3D-XL and novel\nevaluation metrics that offer a more nuanced assessment of a 3D FSCIL model.\nExperimental results on both established and our proposed benchmarks\ndemonstrate that our approach significantly outperforms existing\nstate-of-the-art methods.\n", "date": "2023-12-28", "categories": []}
{"id": "2501.04588", "title": "Federated-Continual Dynamic Segmentation of Histopathology guided by\n  Barlow Continuity", "abstract": "  Federated- and Continual Learning have been established as approaches to\nenable privacy-aware learning on continuously changing data, as required for\ndeploying AI systems in histopathology images. However, data shifts can occur\nin a dynamic world, spatially between institutions and temporally, due to\nchanging data over time. This leads to two issues: Client Drift, where the\ncentral model degrades from aggregating data from clients trained on shifted\ndata, and Catastrophic Forgetting, from temporal shifts such as changes in\npatient populations. Both tend to degrade the model's performance of previously\nseen data or spatially distributed training. Despite both problems arising from\nthe same underlying problem of data shifts, existing research addresses them\nonly individually. In this work, we introduce a method that can jointly\nalleviate Client Drift and Catastrophic Forgetting by using our proposed\nDynamic Barlow Continuity that evaluates client updates on a public reference\ndataset and uses this to guide the training process to a spatially and\ntemporally shift-invariant model. We evaluate our approach on the\nhistopathology datasets BCSS and Semicol and prove our method to be highly\neffective by jointly improving the dice score as much as from 15.8% to 71.6% in\nClient Drift and from 42.5% to 62.8% in Catastrophic Forgetting. This enables\nDynamic Learning by establishing spatio-temporal shift-invariance.\n", "date": "2025-01-08", "categories": []}
{"id": "2411.16532", "title": "Continual Deep Reinforcement Learning with Task-Agnostic Policy\n  Distillation", "abstract": "  Central to the development of universal learning systems is the ability to\nsolve multiple tasks without retraining from scratch when new data arrives.\nThis is crucial because each task requires significant training time.\nAddressing the problem of continual learning necessitates various methods due\nto the complexity of the problem space. This problem space includes: (1)\naddressing catastrophic forgetting to retain previously learned tasks, (2)\ndemonstrating positive forward transfer for faster learning, (3) ensuring\nscalability across numerous tasks, and (4) facilitating learning without\nrequiring task labels, even in the absence of clear task boundaries. In this\npaper, the Task-Agnostic Policy Distillation (TAPD) framework is introduced.\nThis framework alleviates problems (1)-(4) by incorporating a task-agnostic\nphase, where an agent explores its environment without any external goal and\nmaximizes only its intrinsic motivation. The knowledge gained during this phase\nis later distilled for further exploration. Therefore, the agent acts in a\nself-supervised manner by systematically seeking novel states. By utilizing\ntask-agnostic distilled knowledge, the agent can solve downstream tasks more\nefficiently, leading to improved sample efficiency. Our code is available at\nthe repository: https://github.com/wabbajack1/TAPD.\n", "date": "2024-11-25", "categories": []}
{"id": "2308.08747", "title": "An Empirical Study of Catastrophic Forgetting in Large Language Models\n  During Continual Fine-tuning", "abstract": "  Catastrophic forgetting (CF) is a phenomenon that occurs in machine learning\nwhen a model forgets previously learned information while acquiring new\nknowledge for achieving a satisfactory performance in downstream tasks. As\nlarge language models (LLMs) have demonstrated remarkable performance, it is\nintriguing to investigate whether CF exists during the continual instruction\ntuning of LLMs. This study empirically evaluates the forgetting phenomenon in\nLLMs' knowledge during continual instruction tuning from the perspectives of\ndomain knowledge, reasoning, and reading comprehension. The experiments reveal\nthat catastrophic forgetting is generally observed in LLMs ranging from 1b to\n7b parameters. Surprisingly, as the model scale increases, the severity of\nforgetting intensifies in such a model sale range which may result from the\nmuch significant initial performance in the larger LLM. Comparing the\ndecoder-only model BLOOMZ with the encoder-decoder model mT0, BLOOMZ exhibits\nless forgetting and retains more knowledge. Interestingly, we also observe that\nLLMs can mitigate language biases, such as gender bias, during continual\nfine-tuning. Furthermore, our findings indicate that general instruction tuning\ncan help alleviate the forgetting phenomenon in LLMs during subsequent\nfine-tuning.\n", "date": "2023-08-17", "categories": []}
{"id": "2405.14124", "title": "Mixture of Experts Meets Prompt-Based Continual Learning", "abstract": "  Exploiting the power of pre-trained models, prompt-based approaches stand out\ncompared to other continual learning solutions in effectively preventing\ncatastrophic forgetting, even with very few learnable parameters and without\nthe need for a memory buffer. While existing prompt-based continual learning\nmethods excel in leveraging prompts for state-of-the-art performance, they\noften lack a theoretical explanation for the effectiveness of prompting. This\npaper conducts a theoretical analysis to unravel how prompts bestow such\nadvantages in continual learning, thus offering a new perspective on prompt\ndesign. We first show that the attention block of pre-trained models like\nVision Transformers inherently encodes a special mixture of experts\narchitecture, characterized by linear experts and quadratic gating score\nfunctions. This realization drives us to provide a novel view on prefix tuning,\nreframing it as the addition of new task-specific experts, thereby inspiring\nthe design of a novel gating mechanism termed Non-linear Residual Gates\n(NoRGa). Through the incorporation of non-linear activation and residual\nconnection, NoRGa enhances continual learning performance while preserving\nparameter efficiency. The effectiveness of NoRGa is substantiated both\ntheoretically and empirically across diverse benchmarks and pretraining\nparadigms. Our code is publicly available at\nhttps://github.com/Minhchuyentoancbn/MoE_PromptCL\n", "date": "2024-05-23", "categories": []}
{"id": "2406.07590", "title": "StreamFP: Learnable Fingerprint-guided Data Selection for Efficient\n  Stream Learning", "abstract": "  Stream Learning (SL) requires models that can quickly adapt to continuously\nevolving data, posing significant challenges in both computational efficiency\nand learning accuracy. Effective data selection is critical in SL to ensure a\nbalance between information retention and training efficiency. Traditional\nrule-based data selection methods struggle to accommodate the dynamic nature of\nstreaming data, highlighting the necessity for innovative solutions that\neffectively address these challenges. Recent approaches to handling changing\ndata distributions face challenges that limit their effectiveness in fast-paced\nenvironments. In response, we propose StreamFP, a novel approach that uniquely\nemploys dynamic, learnable parameters called fingerprints to enhance data\nselection efficiency and adaptability in stream learning. StreamFP optimizes\ncoreset selection through its unique fingerprint-guided mechanism for efficient\ntraining while ensuring robust buffer updates that adaptively respond to data\ndynamics, setting it apart from existing methods in stream learning.\nExperimental results demonstrate that StreamFP outperforms state-of-the-art\nmethods by achieving accuracy improvements of 15.99%, 29.65%, and 51.24%\ncompared to baseline models across varying data arrival rates, alongside a\ntraining throughput increase of 4.6x.\n", "date": "2024-06-11", "categories": []}
{"id": "2309.15038", "title": "HPCR: Holistic Proxy-based Contrastive Replay for Online Continual\n  Learning", "abstract": "  Online continual learning, aimed at developing a neural network that\ncontinuously learns new data from a single pass over an online data stream,\ngenerally suffers from catastrophic forgetting. Existing replay-based methods\nalleviate forgetting by replaying partial old data in a proxy-based or\ncontrastive-based replay manner, each with its own shortcomings. Our previous\nwork proposes a novel replay-based method called proxy-based contrastive replay\n(PCR), which handles the shortcomings by achieving complementary advantages of\nboth replay manners. In this work, we further conduct gradient and limitation\nanalysis of PCR. The analysis results show that PCR still can be further\nimproved in feature extraction, generalization, and anti-forgetting\ncapabilities of the model. Hence, we develop a more advanced method named\nholistic proxy-based contrastive replay (HPCR). HPCR consists of three\ncomponents, each tackling one of the limitations of PCR. The contrastive\ncomponent conditionally incorporates anchor-to-sample pairs to PCR, improving\nthe feature extraction ability. The second is a temperature component that\ndecouples the temperature coefficient into two parts based on their gradient\nimpacts and sets different values for them to enhance the generalization\nability. The third is a distillation component that constrains the learning\nprocess with additional loss terms to improve the anti-forgetting ability.\nExperiments on four datasets consistently demonstrate the superiority of HPCR\nover various state-of-the-art methods.\n", "date": "2023-09-26", "categories": []}
{"id": "2408.01356", "title": "Balanced Residual Distillation Learning for 3D Point Cloud\n  Class-Incremental Semantic Segmentation", "abstract": "  Class-incremental learning (CIL) enables continuous learning of new classes\nwhile mitigating catastrophic forgetting of old ones. For the performance\nbreakthrough of CIL, it is essential yet challenging to effectively refine past\nknowledge from the base model and balance it with new learning. However, such a\nchallenge has not been considered in current research. This work proposes a\nbalanced residual distillation learning framework (BRDL) to address this gap\nand advance CIL performance. BRDL introduces a residual distillation strategy\nto dynamically refine past knowledge by expanding the network structure and a\nbalanced pseudo-label learning strategy to mitigate class bias and balance\nlearning between old and new classes. We apply the proposed BRDL to a\nchallenging 3D point cloud semantic segmentation task where the data is\nunordered and unstructured. Extensive experimental results demonstrate that\nBRDL sets a new benchmark with an outstanding balance capability in\nclass-biased scenarios.\n", "date": "2024-08-02", "categories": []}
{"id": "2501.00340", "title": "Dynamic Prompt Adjustment for Multi-Label Class-Incremental Learning", "abstract": "  Significant advancements have been made in single label incremental learning\n(SLCIL),yet the more practical and challenging multi label class incremental\nlearning (MLCIL) remains understudied. Recently,visual language models such as\nCLIP have achieved good results in classification tasks. However,directly using\nCLIP to solve MLCIL issue can lead to catastrophic forgetting. To tackle this\nissue, we integrate an improved data replay mechanism and prompt loss to curb\nknowledge forgetting. Specifically,our model enhances the prompt information to\nbetter adapt to multi-label classification tasks and employs confidence-based\nreplay strategy to select representative samples. Moreover, the prompt loss\nsignificantly reduces the model's forgetting of previous knowledge.\nExperimental results demonstrate that our method has substantially improved the\nperformance of MLCIL tasks across multiple benchmark datasets,validating its\neffectiveness.\n", "date": "2024-12-31", "categories": []}
{"id": "2501.00237", "title": "Make Domain Shift a Catastrophic Forgetting Alleviator in\n  Class-Incremental Learning", "abstract": "  In the realm of class-incremental learning (CIL), alleviating the\ncatastrophic forgetting problem is a pivotal challenge. This paper discovers a\ncounter-intuitive observation: by incorporating domain shift into CIL tasks,\nthe forgetting rate is significantly reduced. Our comprehensive studies\ndemonstrate that incorporating domain shift leads to a clearer separation in\nthe feature distribution across tasks and helps reduce parameter interference\nduring the learning process. Inspired by this observation, we propose a simple\nyet effective method named DisCo to deal with CIL tasks. DisCo introduces a\nlightweight prototype pool that utilizes contrastive learning to promote\ndistinct feature distributions for the current task relative to previous ones,\neffectively mitigating interference across tasks. DisCo can be easily\nintegrated into existing state-of-the-art class-incremental learning methods.\nExperimental results show that incorporating our method into various CIL\nmethods achieves substantial performance improvements, validating the benefits\nof our approach in enhancing class-incremental learning by separating feature\nrepresentation and reducing interference. These findings illustrate that DisCo\ncan serve as a robust fashion for future research in class-incremental\nlearning.\n", "date": "2024-12-31", "categories": []}
{"id": "2412.20034", "title": "Maintain Plasticity in Long-timescale Continual Test-time Adaptation", "abstract": "  Continual test-time domain adaptation (CTTA) aims to adjust pre-trained\nsource models to perform well over time across non-stationary target\nenvironments. While previous methods have made considerable efforts to optimize\nthe adaptation process, a crucial question remains: can the model adapt to\ncontinually-changing environments with preserved plasticity over a long time?\nThe plasticity refers to the model's capability to adjust predictions in\nresponse to non-stationary environments continually. In this work, we explore\nplasticity, this essential but often overlooked aspect of continual adaptation\nto facilitate more sustained adaptation in the long run. First, we observe that\nmost CTTA methods experience a steady and consistent decline in plasticity\nduring the long-timescale continual adaptation phase. Moreover, we find that\nthe loss of plasticity is strongly associated with the change in label flip.\nBased on this correlation, we propose a simple yet effective policy, Adaptive\nShrink-Restore (ASR), towards preserving the model's plasticity. In particular,\nASR does the weight re-initialization by the adaptive intervals. The adaptive\ninterval is determined based on the change in label flipping. Our method is\nvalidated on extensive CTTA benchmarks, achieving excellent performance.\n", "date": "2024-12-28", "categories": []}
{"id": "2408.11374", "title": "A Unified Framework for Continual Learning and Unlearning", "abstract": "  Continual learning and machine unlearning are crucial challenges in machine\nlearning, typically addressed separately. Continual learning focuses on\nadapting to new knowledge while preserving past information, whereas unlearning\ninvolves selectively forgetting specific subsets of data. In this paper, we\nintroduce a new framework that jointly tackles both tasks by leveraging\ncontrolled knowledge distillation. Our approach enables efficient learning with\nminimal forgetting and effective targeted unlearning. By incorporating a fixed\nmemory buffer, the system supports learning new concepts while retaining prior\nknowledge. The distillation process is carefully managed to ensure a balance\nbetween acquiring new information and forgetting specific data as needed.\nExperimental results on benchmark datasets show that our method matches or\nexceeds the performance of existing approaches in both continual learning and\nmachine unlearning. This unified framework is the first to address both\nchallenges simultaneously, paving the way for adaptable models capable of\ndynamic learning and forgetting while maintaining strong overall performance.\nSource code: \\textcolor{blue}{https://respailab.github.io/CLMUL}\n", "date": "2024-08-21", "categories": []}
{"id": "2412.10900", "title": "PEARL: Input-Agnostic Prompt Enhancement with Negative Feedback\n  Regulation for Class-Incremental Learning", "abstract": "  Class-incremental learning (CIL) aims to continuously introduce novel\ncategories into a classification system without forgetting previously learned\nones, thus adapting to evolving data distributions. Researchers are currently\nfocusing on leveraging the rich semantic information of pre-trained models\n(PTMs) in CIL tasks. Prompt learning has been adopted in CIL for its ability to\nadjust data distribution to better align with pre-trained knowledge. This paper\ncritically examines the limitations of existing methods from the perspective of\nprompt learning, which heavily rely on input information. To address this\nissue, we propose a novel PTM-based CIL method called Input-Agnostic Prompt\nEnhancement with Negative Feedback Regulation (PEARL). In PEARL, we implement\nan input-agnostic global prompt coupled with an adaptive momentum update\nstrategy to reduce the model's dependency on data distribution, thereby\neffectively mitigating catastrophic forgetting. Guided by negative feedback\nregulation, this adaptive momentum update addresses the parameter sensitivity\ninherent in fixed-weight momentum updates. Furthermore, it fosters the\ncontinuous enhancement of the prompt for new tasks by harnessing correlations\nbetween different tasks in CIL. Experiments on six benchmarks demonstrate that\nour method achieves state-of-the-art performance. The code is available at:\nhttps://github.com/qinyongchun/PEARL.\n", "date": "2024-12-14", "categories": []}
{"id": "2410.10179", "title": "Is Parameter Collision Hindering Continual Learning in LLMs?", "abstract": "  Large Language Models (LLMs) often suffer from catastrophic forgetting when\nlearning multiple tasks sequentially, making continual learning (CL) essential\nfor their dynamic deployment. Existing state-of-the-art (SOTA) methods, such as\nO-LoRA, typically focus on constructing orthogonality tasks to decouple\nparameter interdependence from various domains.In this paper, we reveal that\nbuilding non-collision parameters is a more critical factor in addressing CL\nchallenges. Our theoretical and experimental analyses demonstrate that\nnon-collision parameters can provide better task orthogonality, which is a\nsufficient but unnecessary condition. Furthermore, knowledge from multiple\ndomains will be preserved in non-collision parameter subspaces, making it more\ndifficult to forget previously seen data. Leveraging this insight, we propose\nNon-collision Low-Rank Adaptation (N-LoRA), a simple yet effective approach\nleveraging low collision rates to enhance CL in LLMs. Experimental results on\nmultiple CL benchmarks indicate that N-LoRA achieves superior performance\n(+2.9), higher task orthogonality (*4.1 times), and lower parameter collision\n(*58.1 times) than SOTA methods.\n", "date": "2024-10-14", "categories": []}
{"id": "2411.01904", "title": "FPPL: An Efficient and Non-IID Robust Federated Continual Learning\n  Framework", "abstract": "  Federated continual learning (FCL) aims to learn from sequential data stream\nin the decentralized federated learning setting, while simultaneously\nmitigating the catastrophic forgetting issue in classical continual learning.\nExisting FCL methods usually employ typical rehearsal mechanisms, which could\nresult in privacy violations or additional onerous storage and computational\nburdens. In this work, an efficient and non-IID robust federated continual\nlearning framework, called Federated Prototype-Augmented Prompt Learning\n(FPPL), is proposed. The FPPL can collaboratively learn lightweight prompts\naugmented by prototypes without rehearsal. On the client side, a fusion\nfunction is employed to fully leverage the knowledge contained in task-specific\nprompts for alleviating catastrophic forgetting. Additionally, global\nprototypes aggregated from the server are used to obtain unified representation\nthrough contrastive learning, mitigating the impact of non-IID-derived data\nheterogeneity. On the server side, locally uploaded prototypes are utilized to\nperform debiasing on the classifier, further alleviating the performance\ndegradation caused by both non-IID and catastrophic forgetting. Empirical\nevaluations demonstrate the effectiveness of FPPL, achieving notable\nperformance with an efficient design while remaining robust to diverse non-IID\ndegrees. Code is available at: https://github.com/ycheoo/FPPL.\n", "date": "2024-11-04", "categories": []}
{"id": "2412.16830", "title": "Algorithm Design for Continual Learning in IoT Networks", "abstract": "  Continual learning (CL) is a new online learning technique over sequentially\ngenerated streaming data from different tasks, aiming to maintain a small\nforgetting loss on previously-learned tasks. Existing work focuses on reducing\nthe forgetting loss under a given task sequence. However, if similar tasks\ncontinuously appear to the end time, the forgetting loss is still huge on prior\ndistinct tasks. In practical IoT networks, an autonomous vehicle to sample data\nand learn different tasks can route and alter the order of task pattern at\nincreased travelling cost. To our best knowledge, we are the first to study how\nto opportunistically route the testing object and alter the task sequence in\nCL. We formulate a new optimization problem and prove it NP-hard. We propose a\npolynomial-time algorithm to achieve approximation ratios of $\\frac{3}{2}$ for\nunderparameterized case and $\\frac{3}{2} + r^{1-T}$ for overparameterized case,\nrespectively, where $r:=1-\\frac{n}{m}$ is a parameter of feature number $m$ and\nsample number $n$ and $T$ is the task number. Simulation results verify our\nalgorithm's close-to-optimum performance.\n", "date": "2024-12-22", "categories": []}
{"id": "2412.18177", "title": "Enhancing Online Continual Learning with Plug-and-Play State Space Model\n  and Class-Conditional Mixture of Discretization", "abstract": "  Online continual learning (OCL) seeks to learn new tasks from data streams\nthat appear only once, while retaining knowledge of previously learned tasks.\nMost existing methods rely on replay, focusing on enhancing memory retention\nthrough regularization or distillation. However, they often overlook the\nadaptability of the model, limiting the ability to learn generalizable and\ndiscriminative features incrementally from online training data. To address\nthis, we introduce a plug-and-play module, S6MOD, which can be integrated into\nmost existing methods and directly improve adaptability. Specifically, S6MOD\nintroduces an extra branch after the backbone, where a mixture of\ndiscretization selectively adjusts parameters in a selective state space model,\nenriching selective scan patterns such that the model can adaptively select the\nmost sensitive discretization method for current dynamics. We further design a\nclass-conditional routing algorithm for dynamic, uncertainty-based adjustment\nand implement a contrastive discretization loss to optimize it. Extensive\nexperiments combining our module with various models demonstrate that S6MOD\nsignificantly enhances model adaptability, leading to substantial performance\ngains and achieving the state-of-the-art results.\n", "date": "2024-12-24", "categories": []}
{"id": "2412.18231", "title": "Towards Macro-AUC oriented Imbalanced Multi-Label Continual Learning", "abstract": "  In Continual Learning (CL), while existing work primarily focuses on the\nmulti-class classification task, there has been limited research on Multi-Label\nLearning (MLL). In practice, MLL datasets are often class-imbalanced, making it\ninherently challenging, a problem that is even more acute in CL. Due to its\nsensitivity to imbalance, Macro-AUC is an appropriate and widely used measure\nin MLL. However, there is no research to optimize Macro-AUC in MLCL\nspecifically. To fill this gap, in this paper, we propose a new memory\nreplay-based method to tackle the imbalance issue for Macro-AUC-oriented MLCL.\nSpecifically, inspired by recent theory work, we propose a new Reweighted\nLabel-Distribution-Aware Margin (RLDAM) loss. Furthermore, to be compatible\nwith the RLDAM loss, a new memory-updating strategy named Weight Retain\nUpdating (WRU) is proposed to maintain the numbers of positive and negative\ninstances of the original dataset in memory. Theoretically, we provide superior\ngeneralization analyses of the RLDAM-based algorithm in terms of Macro-AUC,\nseparately in batch MLL and MLCL settings. This is the first work to offer\ntheoretical generalization analyses in MLCL to our knowledge. Finally, a series\nof experimental results illustrate the effectiveness of our method over several\nbaselines. Our codes are available at\nhttps://github.com/ML-Group-SDU/Macro-AUC-CL.\n", "date": "2024-12-24", "categories": []}
{"id": "2402.14228", "title": "COPR: Continual Human Preference Learning via Optimal Policy\n  Regularization", "abstract": "  Reinforcement Learning from Human Feedback (RLHF) is commonly utilized to\nimprove the alignment of Large Language Models (LLMs) with human preferences.\nGiven the evolving nature of human preferences, continual alignment becomes\nmore crucial and practical in comparison to traditional static alignment.\nNevertheless, making RLHF compatible with Continual Learning (CL) is\nchallenging due to its complex process. Meanwhile, directly learning new human\npreferences may lead to Catastrophic Forgetting (CF) of historical preferences,\nresulting in helpless or harmful outputs. To overcome these challenges, we\npropose the Continual Optimal Policy Regularization (COPR) method, which draws\ninspiration from the optimal policy theory. COPR utilizes a sampling\ndistribution as a demonstration and regularization constraints for CL. It\nadopts the Lagrangian Duality (LD) method to dynamically regularize the current\npolicy based on the historically optimal policy, which prevents CF and avoids\nover-emphasizing unbalanced objectives. We also provide formal proof for the\nlearnability of COPR. The experimental results show that COPR outperforms\nstrong CL baselines on our proposed benchmark, in terms of reward-based, GPT-4\nevaluations and human assessment. Furthermore, we validate the robustness of\nCOPR under various CL settings, including different backbones, replay memory\nsizes, and learning orders.\n", "date": "2024-02-22", "categories": []}
{"id": "2405.17466", "title": "Distributed Continual Learning", "abstract": "  This work studies the intersection of continual and federated learning, in\nwhich independent agents face unique tasks in their environments and\nincrementally develop and share knowledge. We introduce a mathematical\nframework capturing the essential aspects of distributed continual learning,\nincluding agent model and statistical heterogeneity, continual distribution\nshift, network topology, and communication constraints. Operating on the thesis\nthat distributed continual learning enhances individual agent performance over\nsingle-agent learning, we identify three modes of information exchange: data\ninstances, full model parameters, and modular (partial) model parameters. We\ndevelop algorithms for each sharing mode and conduct extensive empirical\ninvestigations across various datasets, topology structures, and communication\nlimits. Our findings reveal three key insights: sharing parameters is more\nefficient than sharing data as tasks become more complex; modular parameter\nsharing yields the best performance while minimizing communication costs; and\ncombining sharing modes can cumulatively improve performance.\n", "date": "2024-05-23", "categories": []}
{"id": "2412.15979", "title": "MR-GDINO: Efficient Open-World Continual Object Detection", "abstract": "  Open-world (OW) recognition and detection models show strong zero- and\nfew-shot adaptation abilities, inspiring their use as initializations in\ncontinual learning methods to improve performance. Despite promising results on\nseen classes, such OW abilities on unseen classes are largely degenerated due\nto catastrophic forgetting. To tackle this challenge, we propose an open-world\ncontinual object detection task, requiring detectors to generalize to old, new,\nand unseen categories in continual learning scenarios. Based on this task, we\npresent a challenging yet practical OW-COD benchmark to assess detection\nabilities. The goal is to motivate OW detectors to simultaneously preserve\nlearned classes, adapt to new classes, and maintain open-world capabilities\nunder few-shot adaptations. To mitigate forgetting in unseen categories, we\npropose MR-GDINO, a strong, efficient and scalable baseline via memory and\nretrieval mechanisms within a highly scalable memory pool. Experimental results\nshow that existing continual detectors suffer from severe forgetting for both\nseen and unseen categories. In contrast, MR-GDINO largely mitigates forgetting\nwith only 0.1% activated extra parameters, achieving state-of-the-art\nperformance for old, new, and unseen categories.\n", "date": "2024-12-20", "categories": []}
{"id": "2412.16409", "title": "Uncertainty Quantification in Continual Open-World Learning", "abstract": "  AI deployed in the real-world should be capable of autonomously adapting to\nnovelties encountered after deployment. Yet, in the field of continual\nlearning, the reliance on novelty and labeling oracles is commonplace albeit\nunrealistic. This paper addresses a challenging and under-explored problem: a\ndeployed AI agent that continuously encounters unlabeled data - which may\ninclude both unseen samples of known classes and samples from novel (unknown)\nclasses - and must adapt to it continuously. To tackle this challenge, we\npropose our method COUQ \"Continual Open-world Uncertainty Quantification\", an\niterative uncertainty estimation algorithm tailored for learning in generalized\ncontinual open-world multi-class settings. We rigorously apply and evaluate\nCOUQ on key sub-tasks in the Continual Open-World: continual novelty detection,\nuncertainty guided active learning, and uncertainty guided pseudo-labeling for\nsemi-supervised CL. We demonstrate the effectiveness of our method across\nmultiple datasets, ablations, backbones and performance superior to\nstate-of-the-art.\n", "date": "2024-12-21", "categories": []}
{"id": "2412.17188", "title": "Hierarchically Gated Experts for Efficient Online Continual Learning", "abstract": "  Continual Learning models aim to learn a set of tasks under the constraint\nthat the tasks arrive sequentially with no way to access data from previous\ntasks. The Online Continual Learning framework poses a further challenge where\nthe tasks are unknown and instead the data arrives as a single stream. Building\non existing work, we propose a method for identifying these underlying tasks:\nthe Gated Experts (GE) algorithm, where a dynamically growing set of experts\nallows for new knowledge to be acquired without catastrophic forgetting.\nFurthermore, we extend GE to Hierarchically Gated Experts (HGE), a method which\nis able to efficiently select the best expert for each data sample by\norganising the experts into a hierarchical structure. On standard Continual\nLearning benchmarks, GE and HGE are able to achieve results comparable with\ncurrent methods, with HGE doing so more efficiently.\n", "date": "2024-12-22", "categories": []}
{"id": "2407.03111", "title": "Compressed Latent Replays for Lightweight Continual Learning on Spiking\n  Neural Networks", "abstract": "  Rehearsal-based Continual Learning (CL) has been intensely investigated in\nDeep Neural Networks (DNNs). However, its application in Spiking Neural\nNetworks (SNNs) has not been explored in depth. In this paper we introduce the\nfirst memory-efficient implementation of Latent Replay (LR)-based CL for SNNs,\ndesigned to seamlessly integrate with resource-constrained devices. LRs combine\nnew samples with latent representations of previously learned data, to mitigate\nforgetting. Experiments on the Heidelberg SHD dataset with Sample and\nClass-Incremental tasks reach a Top-1 accuracy of 92.5% and 92%, respectively,\nwithout forgetting the previously learned information. Furthermore, we minimize\nthe LRs' requirements by applying a time-domain compression, reducing by two\norders of magnitude their memory requirement, with respect to a naive rehearsal\nsetup, with a maximum accuracy drop of 4%. On a Multi-Class-Incremental task,\nour SNN learns 10 new classes from an initial set of 10, reaching a Top-1\naccuracy of 78.4% on the full test set.\n", "date": "2024-05-08", "categories": []}
{"id": "2407.18042", "title": "Lifelong Graph Learning for Graph Summarization", "abstract": "  Summarizing web graphs is challenging due to the heterogeneity of the modeled\ninformation and its changes over time. We investigate the use of neural\nnetworks for lifelong graph summarization. Assuming we observe the web graph at\na certain time, we train the networks to summarize graph vertices. We apply\nthis trained network to summarize the vertices of the changed graph at the next\npoint in time. Subsequently, we continue training and evaluating the network to\nperform lifelong graph summarization. We use the GNNs Graph-MLP and GraphSAINT,\nas well as an MLP baseline, to summarize the temporal graphs. We compare\n$1$-hop and $2$-hop summaries. We investigate the impact of reusing parameters\nfrom a previous snapshot by measuring the backward and forward transfer and the\nforgetting rate of the neural networks. Our extensive experiments on ten weekly\nsnapshots of a web graph with over $100$M edges, sampled in 2012 and 2022, show\nthat all networks predominantly use $1$-hop information to determine the\nsummary, even when performing $2$-hop summarization. Due to the heterogeneity\nof web graphs, in some snapshots, the $2$-hop summary produces over ten times\nmore vertex summaries than the $1$-hop summary. When using the network trained\non the last snapshot from 2012 and applying it to the first snapshot of 2022,\nwe observe a strong drop in accuracy. We attribute this drop over the ten-year\ntime warp to the strongly increased heterogeneity of the web graph in 2022.\n", "date": "2024-07-25", "categories": []}
{"id": "2412.15479", "title": "Continual Learning Using Only Large Language Model Prompting", "abstract": "  We introduce CLOB, a novel continual learning (CL) paradigm wherein a large\nlanguage model (LLM) is regarded as a black box. Learning is done incrementally\nvia only verbal prompting. CLOB does not fine-tune any part of the LLM or add\nany trainable parameters to it. It is particularly suitable for LLMs that are\naccessible via APIs. We also propose a new CL technique, called CIS, based on\nincremental summarization that also overcomes the LLM's input length limit.\nExperiments show CIS outperforms baselines by a very large margin.\n", "date": "2024-12-20", "categories": []}
{"id": "2412.15563", "title": "In-context Continual Learning Assisted by an External Continual Learner", "abstract": "  Existing continual learning (CL) methods mainly rely on fine-tuning or\nadapting large language models (LLMs). They still suffer from catastrophic\nforgetting (CF). Little work has been done to exploit in-context learning (ICL)\nto leverage the extensive knowledge within LLMs for CL without updating any\nparameters. However, incrementally learning each new task in ICL necessitates\nadding training examples from each class of the task to the prompt, which\nhampers scalability as the prompt length increases. This issue not only leads\nto excessively long prompts that exceed the input token limit of the underlying\nLLM but also degrades the model's performance due to the overextended context.\nTo address this, we introduce InCA, a novel approach that integrates an\nexternal continual learner (ECL) with ICL to enable scalable CL without CF. The\nECL is built incrementally to pre-select a small subset of likely classes for\neach test instance. By restricting the ICL prompt to only these selected\nclasses, InCA prevents prompt lengths from becoming excessively long, while\nmaintaining high performance. Experimental results demonstrate that InCA\nsignificantly outperforms existing CL baselines, achieving substantial\nperformance gains.\n", "date": "2024-12-20", "categories": []}
{"id": "2412.15571", "title": "Continual Learning Using a Kernel-Based Method Over Foundation Models", "abstract": "  Continual learning (CL) learns a sequence of tasks incrementally. This paper\nstudies the challenging CL setting of class-incremental learning (CIL). CIL has\ntwo key challenges: catastrophic forgetting (CF) and inter-task class\nseparation (ICS). Despite numerous proposed methods, these issues remain\npersistent obstacles. This paper proposes a novel CIL method, called Kernel\nLinear Discriminant Analysis (KLDA), that can effectively avoid CF and ICS\nproblems. It leverages only the powerful features learned in a foundation model\n(FM). However, directly using these features proves suboptimal. To address\nthis, KLDA incorporates the Radial Basis Function (RBF) kernel and its Random\nFourier Features (RFF) to enhance the feature representations from the FM,\nleading to improved performance. When a new task arrives, KLDA computes only\nthe mean for each class in the task and updates a shared covariance matrix for\nall learned classes based on the kernelized features. Classification is\nperformed using Linear Discriminant Analysis. Our empirical evaluation using\ntext and image classification datasets demonstrates that KLDA significantly\noutperforms baselines. Remarkably, without relying on replay data, KLDA\nachieves accuracy comparable to joint training of all classes, which is\nconsidered the upper bound for CIL performance. The KLDA code is available at\nhttps://github.com/salehmomeni/klda.\n", "date": "2024-12-20", "categories": []}
{"id": "2312.11973", "title": "Continual Learning: Forget-free Winning Subnetworks for Video\n  Representations", "abstract": "  Inspired by the Lottery Ticket Hypothesis (LTH), which highlights the\nexistence of efficient subnetworks within larger, dense networks, a\nhigh-performing Winning Subnetwork (WSN) in terms of task performance under\nappropriate sparsity conditions is considered for various continual learning\ntasks. It leverages pre-existing weights from dense networks to achieve\nefficient learning in Task Incremental Learning (TIL) and Task-agnostic\nIncremental Learning (TaIL) scenarios. In Few-Shot Class Incremental Learning\n(FSCIL), a variation of WSN referred to as the Soft subnetwork (SoftNet) is\ndesigned to prevent overfitting when the data samples are scarce. Furthermore,\nthe sparse reuse of WSN weights is considered for Video Incremental Learning\n(VIL). The use of Fourier Subneural Operator (FSO) within WSN is considered. It\nenables compact encoding of videos and identifies reusable subnetworks across\nvarying bandwidths. We have integrated FSO into different architectural\nframeworks for continual learning, including VIL, TIL, and FSCIL. Our\ncomprehensive experiments demonstrate FSO's effectiveness, significantly\nimproving task performance at various convolutional representational levels.\nSpecifically, FSO enhances higher-layer performance in TIL and FSCIL and\nlower-layer performance in VIL.\n", "date": "2023-12-19", "categories": []}
{"id": "2412.14430", "title": "Balanced Gradient Sample Retrieval for Enhanced Knowledge Retention in\n  Proxy-based Continual Learning", "abstract": "  Continual learning in deep neural networks often suffers from catastrophic\nforgetting, where representations for previous tasks are overwritten during\nsubsequent training. We propose a novel sample retrieval strategy from the\nmemory buffer that leverages both gradient-conflicting and gradient-aligned\nsamples to effectively retain knowledge about past tasks within a supervised\ncontrastive learning framework. Gradient-conflicting samples are selected for\ntheir potential to reduce interference by re-aligning gradients, thereby\npreserving past task knowledge. Meanwhile, gradient-aligned samples are\nincorporated to reinforce stable, shared representations across tasks. By\nbalancing gradient correction from conflicting samples with alignment\nreinforcement from aligned ones, our approach increases the diversity among\nretrieved instances and achieves superior alignment in parameter space,\nsignificantly enhancing knowledge retention and mitigating proxy drift.\nEmpirical results demonstrate that using both sample types outperforms methods\nrelying solely on one sample type or random retrieval. Experiments on popular\ncontinual learning benchmarks in computer vision validate our method's\nstate-of-the-art performance in mitigating forgetting while maintaining\ncompetitive accuracy on new tasks.\n", "date": "2024-12-19", "categories": []}
{"id": "2406.18868", "title": "Advancing Cross-domain Discriminability in Continual Learning of\n  Vision-Language Models", "abstract": "  Continual learning (CL) with Vision-Language Models (VLMs) has overcome the\nconstraints of traditional CL, which only focuses on previously encountered\nclasses. During the CL of VLMs, we need not only to prevent the catastrophic\nforgetting on incrementally learned knowledge but also to preserve the\nzero-shot ability of VLMs. However, existing methods require additional\nreference datasets to maintain such zero-shot ability and rely on\ndomain-identity hints to classify images across different domains. In this\nstudy, we propose Regression-based Analytic Incremental Learning (RAIL), which\nutilizes a recursive ridge regression-based adapter to learn from a sequence of\ndomains in a non-forgetting manner and decouple the cross-domain correlations\nby projecting features to a higher-dimensional space. Cooperating with a\ntraining-free fusion module, RAIL absolutely preserves the VLM's zero-shot\nability on unseen domains without any reference data. Additionally, we\nintroduce Cross-domain Task-Agnostic Incremental Learning (X-TAIL) setting. In\nthis setting, a CL learner is required to incrementally learn from multiple\ndomains and classify test images from both seen and unseen domains without any\ndomain-identity hint. We theoretically prove RAIL's absolute memorization on\nincrementally learned domains. Experiment results affirm RAIL's\nstate-of-the-art performance in both X-TAIL and existing Multi-domain\nTask-Incremental Learning settings. The code is released at\nhttps://github.com/linghan1997/Regression-based-Analytic-Incremental-Learning.\n", "date": "2024-06-27", "categories": []}
{"id": "2407.11780", "title": "SwitchCIT: Switching for Continual Instruction Tuning", "abstract": "  Large language models (LLMs) and multimodal models (MMs) have exhibited\nimpressive capabilities in various domains, particularly in general language\nunderstanding and visual reasoning. However, these models, trained on massive\ndata, may not be finely optimized for specific tasks triggered by instructions.\nContinual instruction tuning is crucial to adapt a large model to evolving\ntasks and domains, ensuring their effectiveness and relevance across a wide\nrange of applications. In the context of continual instruction tuning, where\nmodels are sequentially trained on different tasks, catastrophic forgetting can\noccur, leading to performance degradation on previously learned tasks. This\nwork addresses the catastrophic forgetting in continual instruction learning\nthrough a switching mechanism for routing computations to parameter-efficient\ntuned models. We demonstrate the effectiveness of our method through\nexperiments on continual instruction tuning of different natural language\ngeneration tasks and vision-language tasks. We also showcase the advantages of\nour proposed method in terms of efficiency, scalability, portability, and\nprivacy preservation.\n", "date": "2024-07-16", "categories": []}
{"id": "2412.13714", "title": "AnchorInv: Few-Shot Class-Incremental Learning of Physiological Signals\n  via Representation Space Guided Inversion", "abstract": "  Deep learning models have demonstrated exceptional performance in a variety\nof real-world applications. These successes are often attributed to strong base\nmodels that can generalize to novel tasks with limited supporting data while\nkeeping prior knowledge intact. However, these impressive results are based on\nthe availability of a large amount of high-quality data, which is often lacking\nin specialized biomedical applications. In such fields, models are usually\ndeveloped with limited data that arrive incrementally with novel categories.\nThis requires the model to adapt to new information while preserving existing\nknowledge. Few-Shot Class-Incremental Learning (FSCIL) methods offer a\npromising approach to addressing these challenges, but they also depend on\nstrong base models that face the same aforementioned limitations. To overcome\nthese constraints, we propose AnchorInv following the straightforward and\nefficient buffer-replay strategy. Instead of selecting and storing raw data,\nAnchorInv generates synthetic samples guided by anchor points in the feature\nspace. This approach protects privacy and regularizes the model for adaptation.\nWhen evaluated on three public physiological time series datasets, AnchorInv\nexhibits efficient knowledge forgetting prevention and improved adaptation to\nnovel classes, surpassing state-of-the-art baselines.\n", "date": "2024-12-18", "categories": []}
{"id": "2412.13860", "title": "Domain-adaptative Continual Learning for Low-resource Tasks: Evaluation\n  on Nepali", "abstract": "  Continual learning has emerged as an important research direction due to the\ninfeasibility of retraining large language models (LLMs) from scratch in the\nevent of new data availability. Of great interest is the domain-adaptive\npre-training (DAPT) paradigm, which focuses on continually training a\npre-trained language model to adapt it to a domain it was not originally\ntrained on. In this work, we evaluate the feasibility of DAPT in a low-resource\nsetting, namely the Nepali language. We use synthetic data to continue training\nLlama 3 8B to adapt it to the Nepali language in a 4-bit QLoRA setting. We\nevaluate the adapted model on its performance, forgetting, and knowledge\nacquisition. We compare the base model and the final model on their Nepali\ngeneration abilities, their performance on popular benchmarks, and run\ncase-studies to probe their linguistic knowledge in Nepali. We see some\nunsurprising forgetting in the final model, but also surprisingly find that\nincreasing the number of shots during evaluation yields better percent\nincreases in the final model (as high as 19.29% increase) compared to the base\nmodel (4.98%), suggesting latent retention. We also explore layer-head\nself-attention heatmaps to establish dependency resolution abilities of the\nfinal model in Nepali.\n", "date": "2024-12-18", "categories": []}
{"id": "2406.19753", "title": "Attack On Prompt: Backdoor Attack in Prompt-Based Continual Learning", "abstract": "  Prompt-based approaches offer a cutting-edge solution to data privacy issues\nin continual learning, particularly in scenarios involving multiple data\nsuppliers where long-term storage of private user data is prohibited. Despite\ndelivering state-of-the-art performance, its impressive remembering capability\ncan become a double-edged sword, raising security concerns as it might\ninadvertently retain poisoned knowledge injected during learning from private\nuser data. Following this insight, in this paper, we expose continual learning\nto a potential threat: backdoor attack, which drives the model to follow a\ndesired adversarial target whenever a specific trigger is present while still\nperforming normally on clean samples. We highlight three critical challenges in\nexecuting backdoor attacks on incremental learners and propose corresponding\nsolutions: (1) \\emph{Transferability}: We employ a surrogate dataset and\nmanipulate prompt selection to transfer backdoor knowledge to data from other\nsuppliers; (2) \\emph{Resiliency}: We simulate static and dynamic states of the\nvictim to ensure the backdoor trigger remains robust during intense incremental\nlearning processes; and (3) \\emph{Authenticity}: We apply binary cross-entropy\nloss as an anti-cheating factor to prevent the backdoor trigger from devolving\ninto adversarial noise. Extensive experiments across various benchmark datasets\nand continual learners validate our continual backdoor framework, achieving up\nto $100\\%$ attack success rate, with further ablation studies confirming our\ncontributions' effectiveness.\n", "date": "2024-06-28", "categories": []}
{"id": "2412.06418", "title": "Continual Learning for Segment Anything Model Adaptation", "abstract": "  Although the current different types of SAM adaptation methods have achieved\npromising performance for various downstream tasks, such as prompt-based ones\nand adapter-based ones, most of them belong to the one-step adaptation\nparadigm. In real-world scenarios, we are generally confronted with the dynamic\nscenario where the data comes in a streaming manner. Driven by the practical\nneed, in this paper, we first propose a novel Continual SAM adaptation (CoSAM)\nbenchmark with 8 different task domains and carefully analyze the limitations\nof the existing SAM one-step adaptation methods in the continual segmentation\nscenario. Then we propose a novel simple-yet-effective Mixture of Domain\nAdapters (MoDA) algorithm which utilizes the Global Feature Tokens (GFT) and\nGlobal Assistant Tokens (GAT) modules to help the SAM encoder extract\nwell-separated features for different task domains, and then provide the\naccurate task-specific information for continual learning. Extensive\nexperiments demonstrate that our proposed MoDA obviously surpasses the existing\nclassic continual learning methods, as well as prompt-based and adapter-based\napproaches for continual segmentation. Moreover, after sequential learning on\nthe CoSAM benchmark with diverse data distributions, our MoDA maintains highly\ncompetitive results in the natural image domain, approaching the zero-shot\nperformance of the original SAM, demonstrating its superior capability in\nknowledge preservation. Notably, the proposed MoDA can be seamlessly integrated\ninto various one-step adaptation methods of SAM, which can consistently bring\nobvious performance gains. Code is available at\n\\url{https://github.com/yangjl1215/CoSAM}\n", "date": "2024-12-09", "categories": []}
{"id": "2412.09224", "title": "DASK: Distribution Rehearsing via Adaptive Style Kernel Learning for\n  Exemplar-Free Lifelong Person Re-Identification", "abstract": "  Lifelong person re-identification (LReID) is an important but challenging\ntask that suffers from catastrophic forgetting due to significant domain gaps\nbetween training steps. Existing LReID approaches typically rely on data replay\nand knowledge distillation to mitigate this issue. However, data replay methods\ncompromise data privacy by storing historical exemplars, while knowledge\ndistillation methods suffer from limited performance due to the cumulative\nforgetting of undistilled knowledge. To overcome these challenges, we propose a\nnovel paradigm that models and rehearses the distribution of the old domains to\nenhance knowledge consolidation during the new data learning, possessing a\nstrong anti-forgetting capacity without storing any exemplars. Specifically, we\nintroduce an exemplar-free LReID method called Distribution Rehearsing via\nAdaptive Style Kernel Learning (DASK). DASK includes a Distribution Rehearser\nLearning (DRL) mechanism that learns to transform arbitrary distribution data\ninto the current data style at each learning step. To enhance the style\ntransfer capacity of DRL, an Adaptive Kernel Prediction Network (AKPNet) is\nexplored to achieve an instance-specific distribution adjustment. Additionally,\nwe design a Distribution Rehearsing-driven LReID Training (DRRT) module, which\nrehearses old distribution based on the new data via the old AKPNet model,\nachieving effective new-old knowledge accumulation under a joint knowledge\nconsolidation scheme. Experimental results show our DASK outperforms the\nexisting methods by 3.6%-6.8% and 4.5%-6.5% on anti-forgetting and\ngeneralization capacity, respectively. Our code is available at\nhttps://github.com/zhoujiahuan1991/AAAI2025-LReID-DASK\n", "date": "2024-12-12", "categories": []}
{"id": "2412.11017", "title": "On Distilling the Displacement Knowledge for Few-Shot Class-Incremental\n  Learning", "abstract": "  Few-shot Class-Incremental Learning (FSCIL) addresses the challenges of\nevolving data distributions and the difficulty of data acquisition in\nreal-world scenarios. To counteract the catastrophic forgetting typically\nencountered in FSCIL, knowledge distillation is employed as a way to maintain\nthe knowledge from learned data distribution. Recognizing the limitations of\ngenerating discriminative feature representations in a few-shot context, our\napproach incorporates structural information between samples into knowledge\ndistillation. This structural information serves as a remedy for the low\nquality of features. Diverging from traditional structured distillation methods\nthat compute sample similarity, we introduce the Displacement Knowledge\nDistillation (DKD) method. DKD utilizes displacement rather than similarity\nbetween samples, incorporating both distance and angular information to\nsignificantly enhance the information density retained through knowledge\ndistillation. Observing performance disparities in feature distribution between\nbase and novel classes, we propose the Dual Distillation Network (DDNet). This\nnetwork applies traditional knowledge distillation to base classes and DKD to\nnovel classes, challenging the conventional integration of novel classes with\nbase classes. Additionally, we implement an instance-aware sample selector\nduring inference to dynamically adjust dual branch weights, thereby leveraging\nthe complementary strengths of each approach. Extensive testing on three\nbenchmarks demonstrates that DDNet achieves state-of-the-art results. Moreover,\nthrough rigorous experimentation and comparison, we establish the robustness\nand general applicability of our proposed DKD method.\n", "date": "2024-12-15", "categories": []}
{"id": "2412.12654", "title": "CALA: A Class-Aware Logit Adapter for Few-Shot Class-Incremental\n  Learning", "abstract": "  Few-Shot Class-Incremental Learning (FSCIL) defines a practical but\nchallenging task where models are required to continuously learn novel concepts\nwith only a few training samples. Due to data scarcity, existing FSCIL methods\nresort to training a backbone with abundant base data and then keeping it\nfrozen afterward. However, the above operation often causes the backbone to\noverfit to base classes while overlooking the novel ones, leading to severe\nconfusion between them. To address this issue, we propose Class-Aware Logit\nAdapter (CALA). Our method involves a lightweight adapter that learns to\nrectify biased predictions through a pseudo-incremental learning paradigm. In\nthe real FSCIL process, we use the learned adapter to dynamically generate\nrobust balancing factors. These factors can adjust confused novel instances\nback to their true label space based on their similarity to base classes.\nSpecifically, when confusion is more likely to occur in novel instances that\nclosely resemble base classes, greater rectification is required. Notably, CALA\noperates on the classifier level, preserving the original feature space, thus\nit can be flexibly plugged into most of the existing FSCIL works for improved\nperformance. Experiments on three benchmark datasets consistently validate the\neffectiveness and flexibility of CALA. Codes will be available upon acceptance.\n", "date": "2024-12-17", "categories": []}
{"id": "2412.12669", "title": "Adaptive Prototype Replay for Class Incremental Semantic Segmentation", "abstract": "  Class incremental semantic segmentation (CISS) aims to segment new classes\nduring continual steps while preventing the forgetting of old knowledge.\nExisting methods alleviate catastrophic forgetting by replaying distributions\nof previously learned classes using stored prototypes or features. However,\nthey overlook a critical issue: in CISS, the representation of class knowledge\nis updated continuously through incremental learning, whereas prototype replay\nmethods maintain fixed prototypes. This mismatch between updated representation\nand fixed prototypes limits the effectiveness of the prototype replay strategy.\nTo address this issue, we propose the Adaptive prototype replay (Adapter) for\nCISS in this paper. Adapter comprises an adaptive deviation compen sation (ADC)\nstrategy and an uncertainty-aware constraint (UAC) loss. Specifically, the ADC\nstrategy dynamically updates the stored prototypes based on the estimated\nrepresentation shift distance to match the updated representation of old class.\nThe UAC loss reduces prediction uncertainty, aggregating discriminative\nfeatures to aid in generating compact prototypes. Additionally, we introduce a\ncompensation-based prototype similarity discriminative (CPD) loss to ensure\nadequate differentiation between similar prototypes, thereby enhancing the\nefficiency of the adaptive prototype replay strategy. Extensive experiments on\nPascal VOC and ADE20K datasets demonstrate that Adapter achieves\nstate-of-the-art results and proves effective across various CISS tasks,\nparticularly in challenging multi-step scenarios. The code and model is\navailable at https://github.com/zhu-gl-ux/Adapter.\n", "date": "2024-12-17", "categories": []}
{"id": "2412.12801", "title": "Multi-View Incremental Learning with Structured Hebbian Plasticity for\n  Enhanced Fusion Efficiency", "abstract": "  The rapid evolution of multimedia technology has revolutionized human\nperception, paving the way for multi-view learning. However, traditional\nmulti-view learning approaches are tailored for scenarios with fixed data\nviews, falling short of emulating the intricate cognitive procedures of the\nhuman brain processing signals sequentially. Our cerebral architecture\nseamlessly integrates sequential data through intricate feed-forward and\nfeedback mechanisms. In stark contrast, traditional methods struggle to\ngeneralize effectively when confronted with data spanning diverse domains,\nhighlighting the need for innovative strategies that can mimic the brain's\nadaptability and dynamic integration capabilities. In this paper, we propose a\nbio-neurologically inspired multi-view incremental framework named MVIL aimed\nat emulating the brain's fine-grained fusion of sequentially arriving views.\nMVIL lies two fundamental modules: structured Hebbian plasticity and synaptic\npartition learning. The structured Hebbian plasticity reshapes the structure of\nweights to express the high correlation between view representations,\nfacilitating a fine-grained fusion of view representations. Moreover, synaptic\npartition learning is efficient in alleviating drastic changes in weights and\nalso retaining old knowledge by inhibiting partial synapses. These modules\nbionically play a central role in reinforcing crucial associations between\nnewly acquired information and existing knowledge repositories, thereby\nenhancing the network's capacity for generalization. Experimental results on\nsix benchmark datasets show MVIL's effectiveness over state-of-the-art methods.\n", "date": "2024-12-17", "categories": []}
{"id": "2404.08707", "title": "CEM: A Data-Efficient Method for Large Language Models to Continue\n  Evolving From Mistakes", "abstract": "  As world knowledge advances and new task schemas emerge, Continual Learning\n(CL) becomes essential for keeping Large Language Models (LLMs) current and\naddressing their shortcomings. This process typically involves continual\ninstruction tuning (CIT) and continual pre-training (CPT) to enable these\nmodels to adapt to novel tasks and acquire critical knowledge. However,\ncollecting sufficient CPT data and efficiently bridging knowledge gaps remain\nsignificant challenges. Inspired by the 'summarizing mistakes' strategy, we\npropose the Continue Evolving from Mistakes (CEM) method, a data-efficient\napproach aiming to collect CPT data and continually improve LLMs' performance\nthrough iterative evaluation and supplementation with mistake-relevant\nknowledge. To further optimize data usage and mitigate forgetting, we introduce\na novel training paradigm that combines CIT and CPT. Experiments show that CEM\nsubstantially enhances multiple models' performance on both in-domain and\nout-of-domain QA tasks, achieving gains of up to 29.63%. Code and datasets are\navailable on https://anonymous.4open.science/r/cem-BB25.\n", "date": "2024-04-11", "categories": []}
{"id": "2406.13123", "title": "ViLCo-Bench: VIdeo Language COntinual learning Benchmark", "abstract": "  Video language continual learning involves continuously adapting to\ninformation from video and text inputs, enhancing a model's ability to handle\nnew tasks while retaining prior knowledge. This field is a relatively\nunder-explored area, and establishing appropriate datasets is crucial for\nfacilitating communication and research in this field. In this study, we\npresent the first dedicated benchmark, ViLCo-Bench, designed to evaluate\ncontinual learning models across a range of video-text tasks. The dataset\ncomprises ten-minute-long videos and corresponding language queries collected\nfrom publicly available datasets. Additionally, we introduce a novel\nmemory-efficient framework that incorporates self-supervised learning and\nmimics long-term and short-term memory effects. This framework addresses\nchallenges including memory complexity from long video clips, natural language\ncomplexity from open queries, and text-video misalignment. We posit that\nViLCo-Bench, with greater complexity compared to existing continual learning\nbenchmarks, would serve as a critical tool for exploring the video-language\ndomain, extending beyond conventional class-incremental tasks, and addressing\ncomplex and limited annotation issues. The curated data, evaluations, and our\nnovel method are available at https://github.com/cruiseresearchgroup/ViLCo.\n", "date": "2024-06-19", "categories": []}
{"id": "2412.08929", "title": "CAPrompt: Cyclic Prompt Aggregation for Pre-Trained Model Based Class\n  Incremental Learning", "abstract": "  Recently, prompt tuning methods for pre-trained models have demonstrated\npromising performance in Class Incremental Learning (CIL). These methods\ntypically involve learning task-specific prompts and predicting the task ID to\nselect the appropriate prompts for inference. However, inaccurate task ID\npredictions can cause severe inconsistencies between the prompts used during\ntraining and inference, leading to knowledge forgetting and performance\ndegradation. Additionally, existing prompt tuning methods rely solely on the\npre-trained model to predict task IDs, without fully leveraging the knowledge\nembedded in the learned prompt parameters, resulting in inferior prediction\nperformance. To address these issues, we propose a novel Cyclic Prompt\nAggregation (CAPrompt) method that eliminates the dependency on task ID\nprediction by cyclically aggregating the knowledge from different prompts.\nSpecifically, rather than predicting task IDs, we introduce an innovative\nprompt aggregation strategy during both training and inference to overcome\nprompt inconsistency by utilizing a weighted sum of different prompts. Thorough\ntheoretical analysis demonstrates that under concave conditions, the aggregated\nprompt achieves lower error compared to selecting a single task-specific\nprompt. Consequently, we incorporate a concave constraint and a linear\nconstraint to guide prompt learning, ensuring compliance with the concave\ncondition requirement. Furthermore, to fully exploit the prompts and achieve\nmore accurate prompt weights, we develop a cyclic weight prediction strategy.\nThis strategy begins with equal weights for each task and automatically adjusts\nthem to more appropriate values in a cyclical manner. Experiments on various\ndatasets demonstrate that our proposed CAPrompt outperforms state-of-the-art\nmethods by 2%-3%. Our code is available at\nhttps://github.com/zhoujiahuan1991/AAAI2025-CAPrompt.\n", "date": "2024-12-12", "categories": []}
{"id": "2412.10473", "title": "CONCLAD: COntinuous Novel CLAss Detector", "abstract": "  In the field of continual learning, relying on so-called oracles for novelty\ndetection is commonplace albeit unrealistic. This paper introduces CONCLAD\n(\"COntinuous Novel CLAss Detector\"), a comprehensive solution to the\nunder-explored problem of continual novel class detection in post-deployment\ndata. At each new task, our approach employs an iterative uncertainty\nestimation algorithm to differentiate between known and novel class(es)\nsamples, and to further discriminate between the different novel classes\nthemselves. Samples predicted to be from a novel class with high-confidence are\nautomatically pseudo-labeled and used to update our model. Simultaneously, a\ntiny supervision budget is used to iteratively query ambiguous novel class\npredictions, which are also used during update. Evaluation across multiple\ndatasets, ablations and experimental settings demonstrate our method's\neffectiveness at separating novel and old class samples continuously. We will\nrelease our code upon acceptance.\n", "date": "2024-12-13", "categories": []}
{"id": "2412.10687", "title": "Linked Adapters: Linking Past and Future to Present for Effective\n  Continual Learning", "abstract": "  Continual learning allows the system to learn and adapt to new tasks while\nretaining the knowledge acquired from previous tasks. However, deep learning\nmodels suffer from catastrophic forgetting of knowledge learned from earlier\ntasks while learning a new task. Moreover, retraining large models like\ntransformers from scratch for every new task is costly. An effective approach\nto address continual learning is to use a large pre-trained model with\ntask-specific adapters to adapt to the new tasks. Though this approach can\nmitigate catastrophic forgetting, they fail to transfer knowledge across tasks\nas each task is learning adapters separately. To address this, we propose a\nnovel approach Linked Adapters that allows knowledge transfer through a\nweighted attention mechanism to other task-specific adapters. Linked adapters\nuse a multi-layer perceptron (MLP) to model the attention weights, which\novercomes the challenge of backward knowledge transfer in continual learning in\naddition to modeling the forward knowledge transfer. During inference, our\nproposed approach effectively leverages knowledge transfer through MLP-based\nattention weights across all the lateral task adapters. Through numerous\nexperiments conducted on diverse image classification datasets, we effectively\ndemonstrated the improvement in performance on the continual learning tasks\nusing Linked Adapters.\n", "date": "2024-12-14", "categories": []}
{"id": "2412.10780", "title": "Continual Learning for Behavior-based Driver Identification", "abstract": "  Behavior-based Driver Identification is an emerging technology that\nrecognizes drivers based on their unique driving behaviors, offering important\napplications such as vehicle theft prevention and personalized driving\nexperiences. However, most studies fail to account for the real-world\nchallenges of deploying Deep Learning models within vehicles. These challenges\ninclude operating under limited computational resources, adapting to new\ndrivers, and changes in driving behavior over time. The objective of this study\nis to evaluate if Continual Learning (CL) is well-suited to address these\nchallenges, as it enables models to retain previously learned knowledge while\ncontinually adapting with minimal computational overhead and resource\nrequirements. We tested several CL techniques across three scenarios of\nincreasing complexity based on the well-known OCSLab dataset. This work\nprovides an important step forward in scalable driver identification solutions,\ndemonstrating that CL approaches, such as DER, can obtain strong performance,\nwith only an 11% reduction in accuracy compared to the static scenario.\nFurthermore, to enhance the performance, we propose two new methods, SmooER and\nSmooDER, that leverage the temporal continuity of driver identity over time to\nenhance classification accuracy. Our novel method, SmooDER, achieves optimal\nresults with only a 2% reduction compared to the 11\\% of the DER approach. In\nconclusion, this study proves the feasibility of CL approaches to address the\nchallenges of Driver Identification in dynamic environments, making them\nsuitable for deployment on cloud infrastructure or directly within vehicles.\n", "date": "2024-12-14", "categories": []}
{"id": "2412.11551", "title": "Region-Based Optimization in Continual Learning for Audio Deepfake\n  Detection", "abstract": "  Rapid advancements in speech synthesis and voice conversion bring convenience\nbut also new security risks, creating an urgent need for effective audio\ndeepfake detection. Although current models perform well, their effectiveness\ndiminishes when confronted with the diverse and evolving nature of real-world\ndeepfakes. To address this issue, we propose a continual learning method named\nRegion-Based Optimization (RegO) for audio deepfake detection. Specifically, we\nuse the Fisher information matrix to measure important neuron regions for real\nand fake audio detection, dividing them into four regions. First, we directly\nfine-tune the less important regions to quickly adapt to new tasks. Next, we\napply gradient optimization in parallel for regions important only to real\naudio detection, and in orthogonal directions for regions important only to\nfake audio detection. For regions that are important to both, we use sample\nproportion-based adaptive gradient optimization. This region-adaptive\noptimization ensures an appropriate trade-off between memory stability and\nlearning plasticity. Additionally, to address the increase of redundant neurons\nfrom old tasks, we further introduce the Ebbinghaus forgetting mechanism to\nrelease them, thereby promoting the capability of the model to learn more\ngeneralized discriminative features. Experimental results show our method\nachieves a 21.3% improvement in EER over the state-of-the-art continual\nlearning approach RWM for audio deepfake detection. Moreover, the effectiveness\nof RegO extends beyond the audio deepfake detection domain, showing potential\nsignificance in other tasks, such as image recognition. The code is available\nat https://github.com/cyjie429/RegO\n", "date": "2024-12-16", "categories": []}
{"id": "2412.09701", "title": "CUAL: Continual Uncertainty-aware Active Learner", "abstract": "  AI deployed in many real-world use cases should be capable of adapting to\nnovelties encountered after deployment. Here, we consider a challenging,\nunder-explored and realistic continual adaptation problem: a deployed AI agent\nis continuously provided with unlabeled data that may contain not only unseen\nsamples of known classes but also samples from novel (unknown) classes. In such\na challenging setting, it has only a tiny labeling budget to query the most\ninformative samples to help it continuously learn. We present a comprehensive\nsolution to this complex problem with our model \"CUAL\" (Continual\nUncertainty-aware Active Learner). CUAL leverages an uncertainty estimation\nalgorithm to prioritize active labeling of ambiguous (uncertain) predicted\nnovel class samples while also simultaneously pseudo-labeling the most certain\npredictions of each class. Evaluations across multiple datasets, ablations,\nsettings and backbones (e.g. ViT foundation model) demonstrate our method's\neffectiveness. We will release our code upon acceptance.\n", "date": "2024-12-12", "categories": []}
{"id": "2412.10244", "title": "Efficient Continual Pre-training of LLMs for Low-resource Languages", "abstract": "  Open-source Large Language models (OsLLMs) propel the democratization of\nnatural language research by giving the flexibility to augment or update model\nparameters for performance improvement. Nevertheless, like proprietary LLMs,\nOs-LLMs offer poorer performance on low-resource languages (LRLs) than\nhigh-resource languages (HRLs), owing to smaller amounts of training data and\nunderrepresented vocabulary. On the other hand, continual pre-training (CPT)\nwith large amounts of language-specific data is a costly proposition in terms\nof data acquisition and computational resources. Our goal is to drastically\nreduce CPT cost. To that end, we first develop a new algorithm to select a\nsubset of texts from a larger corpus. We show the effectiveness of our\ntechnique using very little CPT data. In search of further improvement, we\ndesign a new algorithm to select tokens to include in the LLM vocabulary. We\nexperiment with the recent Llama-3 model and nine Indian languages with diverse\nscripts and extent of resource availability. For evaluation, we use\nIndicGenBench, a generation task benchmark dataset for Indic languages. We\nexperiment with various CPT corpora and augmented vocabulary size and offer\ninsights across language families.\n", "date": "2024-12-13", "categories": []}
{"id": "2412.09549", "title": "Exemplar Masking for Multimodal Incremental Learning", "abstract": "  Multimodal incremental learning needs to digest the information from multiple\nmodalities while concurrently learning new knowledge without forgetting the\npreviously learned information. There are numerous challenges for this task,\nmainly including the larger storage size of multimodal data in exemplar-based\nmethods and the computational requirement of finetuning on huge multimodal\nmodels. In this paper, we leverage the parameter-efficient tuning scheme to\nreduce the burden of fine-tuning and propose the exemplar masking framework to\nefficiently replay old knowledge. Specifically, the non-important tokens are\nmasked based on the attention weights and the correlation across different\nmodalities, significantly reducing the storage size of an exemplar and\nconsequently saving more exemplars under the same memory buffer. Moreover, we\ndesign a multimodal data augmentation technique to diversify exemplars for\nreplaying prior knowledge. In experiments, we not only evaluate our method in\nexisting multimodal datasets but also extend the ImageNet-R dataset to a\nmultimodal dataset as a real-world application, where captions are generated by\nquerying multimodal large language models (e.g., InstructBLIP). Extensive\nexperiments show that our exemplar masking framework is more efficient and\nrobust to catastrophic forgetting under the same limited memory buffer. Code is\navailable at https://github.com/YiLunLee/Exemplar_Masking_MCIL.\n", "date": "2024-12-12", "categories": []}
{"id": "2312.08977", "title": "Weighted Ensemble Models Are Strong Continual Learners", "abstract": "  In this work, we study the problem of continual learning (CL) where the goal\nis to learn a model on a sequence of tasks, such that the data from the\nprevious tasks becomes unavailable while learning on the current task data. CL\nis essentially a balancing act between being able to learn on the new task\n(i.e., plasticity) and maintaining the performance on the previously learned\nconcepts (i.e., stability). Intending to address the stability-plasticity\ntrade-off, we propose to perform weight-ensembling of the model parameters of\nthe previous and current tasks. This weighted-ensembled model, which we call\nContinual Model Averaging (or CoMA), attains high accuracy on the current task\nby leveraging plasticity, while not deviating too far from the previous weight\nconfiguration, ensuring stability. We also propose an improved variant of CoMA,\nnamed Continual Fisher-weighted Model Averaging (or CoFiMA), that selectively\nweighs each parameter in the weights ensemble by leveraging the Fisher\ninformation of the weights of the model. Both variants are conceptually simple,\neasy to implement, and effective in attaining state-of-the-art performance on\nseveral standard CL benchmarks. Code is available at:\nhttps://github.com/IemProg/CoFiMA.\n", "date": "2023-12-14", "categories": []}
{"id": "2402.01865", "title": "What Will My Model Forget? Forecasting Forgotten Examples in Language\n  Model Refinement", "abstract": "  Language models deployed in the wild make errors. However, simply updating\nthe model with the corrected error instances causes catastrophic forgetting --\nthe updated model makes errors on instances learned during the instruction\ntuning or upstream training phase. Randomly replaying upstream data yields\nunsatisfactory performance and often comes with high variance and poor\ncontrollability. To this end, we try to forecast upstream examples that will be\nforgotten due to a model update for improved controllability of the replay\nprocess and interpretability. We train forecasting models given a collection of\nonline learned examples and corresponding forgotten upstream pre-training\nexamples. We propose a partially interpretable forecasting model based on the\nobservation that changes in pre-softmax logit scores of pretraining examples\nresemble that of online learned examples, which performs decently on BART but\nfails on T5 models. We further show a black-box classifier based on inner\nproducts of example representations achieves better forecasting performance\nover a series of setups. Finally, we show that we reduce forgetting of upstream\npretraining examples by replaying examples that are forecasted to be forgotten,\ndemonstrating the practical utility of forecasting example forgetting.\n", "date": "2024-02-02", "categories": []}
{"id": "2412.07224", "title": "Parseval Regularization for Continual Reinforcement Learning", "abstract": "  Loss of plasticity, trainability loss, and primacy bias have been identified\nas issues arising when training deep neural networks on sequences of tasks --\nall referring to the increased difficulty in training on new tasks. We propose\nto use Parseval regularization, which maintains orthogonality of weight\nmatrices, to preserve useful optimization properties and improve training in a\ncontinual reinforcement learning setting. We show that it provides significant\nbenefits to RL agents on a suite of gridworld, CARL and MetaWorld tasks. We\nconduct comprehensive ablations to identify the source of its benefits and\ninvestigate the effect of certain metrics associated to network trainability\nincluding weight matrix rank, weight norms and policy entropy.\n", "date": "2024-12-10", "categories": []}
{"id": "2412.07246", "title": "Filling Memory Gaps: Enhancing Continual Semantic Parsing via SQL Syntax\n  Variance-Guided LLMs without Real Data Replay", "abstract": "  Continual Semantic Parsing (CSP) aims to train parsers to convert natural\nlanguage questions into SQL across tasks with limited annotated examples,\nadapting to the real-world scenario of dynamically updated databases. Previous\nstudies mitigate this challenge by replaying historical data or employing\nparameter-efficient tuning (PET), but they often violate data privacy or rely\non ideal continual learning settings. To address these problems, we propose a\nnew Large Language Model (LLM)-Enhanced Continuous Semantic Parsing method,\nnamed LECSP, which alleviates forgetting while encouraging generalization,\nwithout requiring real data replay or ideal settings. Specifically, it first\nanalyzes the commonalities and differences between tasks from the SQL syntax\nperspective to guide LLMs in reconstructing key memories and improving memory\naccuracy through a calibration strategy. Then, it uses a task-aware\ndual-teacher distillation framework to promote the accumulation and transfer of\nknowledge during sequential training. Experimental results on two CSP\nbenchmarks show that our method significantly outperforms existing methods,\neven those utilizing data replay or ideal settings. Additionally, we achieve\ngeneralization performance beyond the upper limits, better adapting to unseen\ntasks.\n", "date": "2024-12-10", "categories": []}
{"id": "2412.07393", "title": "CMT: A Memory Compression Method for Continual Knowledge Learning of\n  Large Language Models", "abstract": "  Large Language Models (LLMs) need to adapt to the continuous changes in data,\ntasks, and user preferences. Due to their massive size and the high costs\nassociated with training, LLMs are not suitable for frequent retraining.\nHowever, updates are necessary to keep them in sync with rapidly evolving human\nknowledge. To address these challenges, this paper proposes the Compression\nMemory Training (CMT) method, an efficient and effective online adaptation\nframework for LLMs that features robust knowledge retention capabilities.\nInspired by human memory mechanisms, CMT compresses and extracts information\nfrom new documents to be stored in a memory bank. When answering to queries\nrelated to these new documents, the model aggregates these document memories\nfrom the memory bank to better answer user questions. The parameters of the LLM\nitself do not change during training and inference, reducing the risk of\ncatastrophic forgetting. To enhance the encoding, retrieval, and aggregation of\nmemory, we further propose three new general and flexible techniques, including\nmemory-aware objective, self-matching and top-aggregation. Extensive\nexperiments conducted on three continual learning datasets (i.e., StreamingQA,\nSQuAD and ArchivalQA) demonstrate that the proposed method improves model\nadaptability and robustness across multiple base LLMs (e.g., +4.07 EM & +4.19\nF1 in StreamingQA with Llama-2-7b).\n", "date": "2024-12-10", "categories": []}
{"id": "2412.05573", "title": "Neighborhood Commonality-aware Evolution Network for Continuous\n  Generalized Category Discovery", "abstract": "  Continuous Generalized Category Discovery (C-GCD) aims to continually\ndiscover novel classes from unlabelled image sets while maintaining performance\non old classes. In this paper, we propose a novel learning framework, dubbed\nNeighborhood Commonality-aware Evolution Network (NCENet) that conquers this\ntask from the perspective of representation learning. Concretely, to learn\ndiscriminative representations for novel classes, a Neighborhood\nCommonality-aware Representation Learning (NCRL) is designed, which exploits\nlocal commonalities derived neighborhoods to guide the learning of\nrepresentational differences between instances of different classes. To\nmaintain the representation ability for old classes, a Bi-level Contrastive\nKnowledge Distillation (BCKD) module is designed, which leverages contrastive\nlearning to perceive the learning and learned knowledge and conducts knowledge\ndistillation. Extensive experiments conducted on CIFAR10, CIFAR100, and\nTiny-ImageNet demonstrate the superior performance of NCENet compared to the\nprevious state-of-the-art method. Particularly, in the last incremental\nlearning session on CIFAR100, the clustering accuracy of NCENet outperforms the\nsecond-best method by a margin of 3.09\\% on old classes and by a margin of\n6.32\\% on new classes. Our code will be publicly available at\n\\href{https://github.com/xjtuYW/NCENet.git}{https://github.com/xjtuYW/NCENet.git}.\n\\end{abstract}\n", "date": "2024-12-07", "categories": []}
{"id": "2412.05840", "title": "LVP-CLIP:Revisiting CLIP for Continual Learning with Label Vector Pool", "abstract": "  Continual learning aims to update a model so that it can sequentially learn\nnew tasks without forgetting previously acquired knowledge. Recent continual\nlearning approaches often leverage the vision-language model CLIP for its\nhigh-dimensional feature space and cross-modality feature matching. Traditional\nCLIP-based classification methods identify the most similar text label for a\ntest image by comparing their embeddings. However, these methods are sensitive\nto the quality of text phrases and less effective for classes lacking\nmeaningful text labels. In this work, we rethink CLIP-based continual learning\nand introduce the concept of Label Vector Pool (LVP). LVP replaces text labels\nwith training images as similarity references, eliminating the need for ideal\ntext descriptions. We present three variations of LVP and evaluate their\nperformance on class and domain incremental learning tasks. Leveraging CLIP's\nhigh dimensional feature space, LVP learning algorithms are task-order\ninvariant. The new knowledge does not modify the old knowledge, hence, there is\nminimum forgetting. Different tasks can be learned independently and in\nparallel with low computational and memory demands. Experimental results show\nthat proposed LVP-based methods outperform the current state-of-the-art\nbaseline by a significant margin of 40.7%.\n", "date": "2024-12-08", "categories": []}
{"id": "2412.06642", "title": "Class Balance Matters to Active Class-Incremental Learning", "abstract": "  Few-Shot Class-Incremental Learning has shown remarkable efficacy in\nefficient learning new concepts with limited annotations. Nevertheless, the\nheuristic few-shot annotations may not always cover the most informative\nsamples, which largely restricts the capability of incremental learner. We aim\nto start from a pool of large-scale unlabeled data and then annotate the most\ninformative samples for incremental learning. Based on this premise, this paper\nintroduces the Active Class-Incremental Learning (ACIL). The objective of ACIL\nis to select the most informative samples from the unlabeled pool to\neffectively train an incremental learner, aiming to maximize the performance of\nthe resulting model. Note that vanilla active learning algorithms suffer from\nclass-imbalanced distribution among annotated samples, which restricts the\nability of incremental learning. To achieve both class balance and\ninformativeness in chosen samples, we propose Class-Balanced Selection (CBS)\nstrategy. Specifically, we first cluster the features of all unlabeled images\ninto multiple groups. Then for each cluster, we employ greedy selection\nstrategy to ensure that the Gaussian distribution of the sampled features\nclosely matches the Gaussian distribution of all unlabeled features within the\ncluster. Our CBS can be plugged and played into those CIL methods which are\nbased on pretrained models with prompts tunning technique. Extensive\nexperiments under ACIL protocol across five diverse datasets demonstrate that\nCBS outperforms both random selection and other SOTA active learning\napproaches. Code is publicly available at https://github.com/1170300714/CBS.\n", "date": "2024-12-09", "categories": []}
{"id": "2408.14471", "title": "A Practitioner's Guide to Continual Multimodal Pretraining", "abstract": "  Multimodal foundation models serve numerous applications at the intersection\nof vision and language. Still, despite being pretrained on extensive data, they\nbecome outdated over time. To keep models updated, research into continual\npretraining mainly explores scenarios with either (1) infrequent,\nindiscriminate updates on large-scale new data, or (2) frequent, sample-level\nupdates. However, practical model deployment often operates in the gap between\nthese two limit cases, as real-world applications often demand adaptation to\nspecific subdomains, tasks or concepts -- spread over the entire, varying life\ncycle of a model. In this work, we complement current perspectives on continual\npretraining through a research test bed as well as provide comprehensive\nguidance for effective continual model updates in such scenarios. We first\nintroduce FoMo-in-Flux, a continual multimodal pretraining benchmark with\nrealistic compute constraints and practical deployment requirements,\nconstructed over 63 datasets with diverse visual and semantic coverage. Using\nFoMo-in-Flux, we explore the complex landscape of practical continual\npretraining through multiple perspectives: (1) A data-centric investigation of\ndata mixtures and stream orderings that emulate real-world deployment\nsituations, (2) a method-centric investigation ranging from simple fine-tuning\nand traditional continual learning strategies to parameter-efficient updates\nand model merging, (3) meta learning rate schedules and mechanistic design\nchoices, and (4) the influence of model and compute scaling. Together, our\ninsights provide a practitioner's guide to continual multimodal pretraining for\nreal-world deployment. Our benchmark and code is here:\nhttps://github.com/ExplainableML/fomo_in_flux.\n", "date": "2024-08-26", "categories": []}
{"id": "2410.18156", "title": "Dreaming Learning", "abstract": "  Incorporating novelties into deep learning systems remains a challenging\nproblem. Introducing new information to a machine learning system can interfere\nwith previously stored data and potentially alter the global model paradigm,\nespecially when dealing with non-stationary sources. In such cases, traditional\napproaches based on validation error minimization offer limited advantages. To\naddress this, we propose a training algorithm inspired by Stuart Kauffman's\nnotion of the Adjacent Possible. This novel training methodology explores new\ndata spaces during the learning phase. It predisposes the neural network to\nsmoothly accept and integrate data sequences with different statistical\ncharacteristics than expected. The maximum distance compatible with such\ninclusion depends on a specific parameter: the sampling temperature used in the\nexplorative phase of the present method. This algorithm, called Dreaming\nLearning, anticipates potential regime shifts over time, enhancing the neural\nnetwork's responsiveness to non-stationary events that alter statistical\nproperties. To assess the advantages of this approach, we apply this\nmethodology to unexpected statistical changes in Markov chains and\nnon-stationary dynamics in textual sequences. We demonstrated its ability to\nimprove the auto-correlation of generated textual sequences by $\\sim 29\\%$ and\nenhance the velocity of loss convergence by $\\sim 100\\%$ in the case of a\nparadigm shift in Markov chains.\n", "date": "2024-10-23", "categories": []}
{"id": "2412.02865", "title": "Memory-efficient Continual Learning with Neural Collapse Contrastive", "abstract": "  Contrastive learning has significantly improved representation quality,\nenhancing knowledge transfer across tasks in continual learning (CL). However,\ncatastrophic forgetting remains a key challenge, as contrastive based methods\nprimarily focus on \"soft relationships\" or \"softness\" between samples, which\nshift with changing data distributions and lead to representation overlap\nacross tasks. Recently, the newly identified Neural Collapse phenomenon has\nshown promise in CL by focusing on \"hard relationships\" or \"hardness\" between\nsamples and fixed prototypes. However, this approach overlooks \"softness\",\ncrucial for capturing intra-class variability, and this rigid focus can also\npull old class representations toward current ones, increasing forgetting.\nBuilding on these insights, we propose Focal Neural Collapse Contrastive\n(FNC^2), a novel representation learning loss that effectively balances both\nsoft and hard relationships. Additionally, we introduce the Hardness-Softness\nDistillation (HSD) loss to progressively preserve the knowledge gained from\nthese relationships across tasks. Our method outperforms state-of-the-art\napproaches, particularly in minimizing memory reliance. Remarkably, even\nwithout the use of memory, our approach rivals rehearsal-based methods,\noffering a compelling solution for data privacy concerns.\n", "date": "2024-12-03", "categories": []}
{"id": "2403.12486", "title": "NTK-Guided Few-Shot Class Incremental Learning", "abstract": "  The proliferation of Few-Shot Class Incremental Learning (FSCIL)\nmethodologies has highlighted the critical challenge of maintaining robust\nanti-amnesia capabilities in FSCIL learners. In this paper, we present a novel\nconceptualization of anti-amnesia in terms of mathematical generalization,\nleveraging the Neural Tangent Kernel (NTK) perspective. Our method focuses on\ntwo key aspects: ensuring optimal NTK convergence and minimizing NTK-related\ngeneralization loss, which serve as the theoretical foundation for cross-task\ngeneralization. To achieve global NTK convergence, we introduce a principled\nmeta-learning mechanism that guides optimization within an expanded network\narchitecture. Concurrently, to reduce the NTK-related generalization loss, we\nsystematically optimize its constituent factors. Specifically, we initiate\nself-supervised pre-training on the base session to enhance NTK-related\ngeneralization potential. These self-supervised weights are then carefully\nrefined through curricular alignment, followed by the application of dual NTK\nregularization tailored specifically for both convolutional and linear layers.\nThrough the combined effects of these measures, our network acquires robust NTK\nproperties, ensuring optimal convergence and stability of the NTK matrix and\nminimizing the NTK-related generalization loss, significantly enhancing its\ntheoretical generalization. On popular FSCIL benchmark datasets, our NTK-FSCIL\nsurpasses contemporary state-of-the-art approaches, elevating end-session\naccuracy by 2.9\\% to 9.3\\%.\n", "date": "2024-03-19", "categories": []}
{"id": "2412.04190", "title": "Directed Structural Adaptation to Overcome Statistical Conflicts and\n  Enable Continual Learning", "abstract": "  Adaptive networks today rely on overparameterized fixed topologies that\ncannot break through the statistical conflicts they encounter in the data they\nare exposed to, and are prone to \"catastrophic forgetting\" as the network\nattempts to reuse the existing structures to learn new task. We propose a\nstructural adaptation method, DIRAD, that can complexify as needed and in a\ndirected manner without being limited by statistical conflicts within a\ndataset. We then extend this method and present the PREVAL framework, designed\nto prevent \"catastrophic forgetting\" in continual learning by detection of new\ndata and assigning encountered data to suitable models adapted to process them,\nwithout needing task labels anywhere in the workflow. We show the reliability\nof the DIRAD in growing a network with high performance and orders-of-magnitude\nsimpler than fixed topology networks; and demonstrate the proof-of-concept\noperation of PREVAL, in which continual adaptation to new tasks is observed\nwhile being able to detect and discern previously-encountered tasks.\n", "date": "2024-12-05", "categories": []}
{"id": "2401.01598", "title": "Learning Prompt with Distribution-Based Feature Replay for Few-Shot\n  Class-Incremental Learning", "abstract": "  Few-shot Class-Incremental Learning (FSCIL) aims to continuously learn new\nclasses based on very limited training data without forgetting the old ones\nencountered. Existing studies solely relied on pure visual networks, while in\nthis paper we solved FSCIL by leveraging the Vision-Language model (e.g., CLIP)\nand propose a simple yet effective framework, named Learning Prompt with\nDistribution-based Feature Replay (LP-DiF). We observe that simply using CLIP\nfor zero-shot evaluation can substantially outperform the most influential\nmethods. Then, prompt tuning technique is involved to further improve its\nadaptation ability, allowing the model to continually capture specific\nknowledge from each session. To prevent the learnable prompt from forgetting\nold knowledge in the new session, we propose a pseudo-feature replay approach.\nSpecifically, we preserve the old knowledge of each class by maintaining a\nfeature-level Gaussian distribution with a diagonal covariance matrix, which is\nestimated by the image features of training images and synthesized features\ngenerated from a VAE. When progressing to a new session, pseudo-features are\nsampled from old-class distributions combined with training images of the\ncurrent session to optimize the prompt, thus enabling the model to learn new\nknowledge while retaining old knowledge. Experiments on three prevalent\nbenchmarks, i.e., CIFAR100, mini-ImageNet, CUB-200, and two more challenging\nbenchmarks, i.e., SUN-397 and CUB-200$^*$ proposed in this paper showcase the\nsuperiority of LP-DiF, achieving new state-of-the-art (SOTA) in FSCIL. Code is\npublicly available at https://github.com/1170300714/LP-DiF.\n", "date": "2024-01-03", "categories": []}
{"id": "2406.08830", "title": "Center-Sensitive Kernel Optimization for Efficient On-Device Incremental\n  Learning", "abstract": "  To facilitate the evolution of edge intelligence in ever-changing\nenvironments, we study on-device incremental learning constrained in limited\ncomputation resource in this paper. Current on-device training methods just\nfocus on efficient training without considering the catastrophic forgetting,\npreventing the model getting stronger when continually exploring the world. To\nsolve this problem, a direct solution is to involve the existing incremental\nlearning mechanisms into the on-device training framework. Unfortunately, such\na manner cannot work well as those mechanisms usually introduce large\nadditional computational cost to the network optimization process, which would\ninevitably exceed the memory capacity of the edge devices. To address this\nissue, this paper makes an early effort to propose a simple but effective\nedge-friendly incremental learning framework. Based on an empirical study on\nthe knowledge intensity of the kernel elements of the neural network, we find\nthat the center kernel is the key for maximizing the knowledge intensity for\nlearning new data, while freezing the other kernel elements would get a good\nbalance on the model's capacity for overcoming catastrophic forgetting. Upon\nthis finding, we further design a center-sensitive kernel optimization\nframework to largely alleviate the cost of the gradient computation and\nback-propagation. Besides, a dynamic channel element selection strategy is also\nproposed to facilitate a sparse orthogonal gradient projection for further\nreducing the optimization complexity, upon the knowledge explored from the new\ntask data. Extensive experiments validate our method is efficient and\neffective, e.g., our method achieves average accuracy boost of 38.08% with even\nless memory and approximate computation compared to existing on-device training\nmethods, indicating its significant potential for on-device incremental\nlearning.\n", "date": "2024-06-13", "categories": []}
{"id": "2412.02627", "title": "Continual Learning of Personalized Generative Face Models with\n  Experience Replay", "abstract": "  We introduce a novel continual learning problem: how to sequentially update\nthe weights of a personalized 2D and 3D generative face model as new batches of\nphotos in different appearances, styles, poses, and lighting are captured\nregularly. We observe that naive sequential fine-tuning of the model leads to\ncatastrophic forgetting of past representations of the individual's face. We\nthen demonstrate that a simple random sampling-based experience replay method\nis effective at mitigating catastrophic forgetting when a relatively large\nnumber of images can be stored and replayed. However, for long-term deployment\nof these models with relatively smaller storage, this simple random\nsampling-based replay technique also forgets past representations. Thus, we\nintroduce a novel experience replay algorithm that combines random sampling\nwith StyleGAN's latent space to represent the buffer as an optimal convex hull.\nWe observe that our proposed convex hull-based experience replay is more\neffective in preventing forgetting than a random sampling baseline and the\nlower bound.\n", "date": "2024-12-03", "categories": []}
{"id": "2405.04101", "title": "Continual Learning in the Presence of Repetition", "abstract": "  Continual learning (CL) provides a framework for training models in\never-evolving environments. Although re-occurrence of previously seen objects\nor tasks is common in real-world problems, the concept of repetition in the\ndata stream is not often considered in standard benchmarks for CL. Unlike with\nthe rehearsal mechanism in buffer-based strategies, where sample repetition is\ncontrolled by the strategy, repetition in the data stream naturally stems from\nthe environment. This report provides a summary of the CLVision challenge at\nCVPR 2023, which focused on the topic of repetition in class-incremental\nlearning. The report initially outlines the challenge objective and then\ndescribes three solutions proposed by finalist teams that aim to effectively\nexploit the repetition in the stream to learn continually. The experimental\nresults from the challenge highlight the effectiveness of ensemble-based\nsolutions that employ multiple versions of similar modules, each trained on\ndifferent but overlapping subsets of classes. This report underscores the\ntransformative potential of taking a different perspective in CL by employing\nrepetition in the data stream to foster innovative strategy design.\n", "date": "2024-05-07", "categories": []}
{"id": "2405.18653", "title": "Recent Advances of Foundation Language Models-based Continual Learning:\n  A Survey", "abstract": "  Recently, foundation language models (LMs) have marked significant\nachievements in the domains of natural language processing (NLP) and computer\nvision (CV). Unlike traditional neural network models, foundation LMs obtain a\ngreat ability for transfer learning by acquiring rich commonsense knowledge\nthrough pre-training on extensive unsupervised datasets with a vast number of\nparameters. However, they still can not emulate human-like continuous learning\ndue to catastrophic forgetting. Consequently, various continual learning\n(CL)-based methodologies have been developed to refine LMs, enabling them to\nadapt to new tasks without forgetting previous knowledge. However, a systematic\ntaxonomy of existing approaches and a comparison of their performance are still\nlacking, which is the gap that our survey aims to fill. We delve into a\ncomprehensive review, summarization, and classification of the existing\nliterature on CL-based approaches applied to foundation language models, such\nas pre-trained language models (PLMs), large language models (LLMs) and\nvision-language models (VLMs). We divide these studies into offline CL and\nonline CL, which consist of traditional methods, parameter-efficient-based\nmethods, instruction tuning-based methods and continual pre-training methods.\nOffline CL encompasses domain-incremental learning, task-incremental learning,\nand class-incremental learning, while online CL is subdivided into hard task\nboundary and blurry task boundary settings. Additionally, we outline the\ntypical datasets and metrics employed in CL research and provide a detailed\nanalysis of the challenges and future work for LMs-based continual learning.\n", "date": "2024-05-28", "categories": []}
{"id": "2411.18788", "title": "Investigating Plausibility of Biologically Inspired Bayesian Learning in\n  ANNs", "abstract": "  Catastrophic forgetting has been the leading issue in the domain of lifelong\nlearning in artificial systems. Current artificial systems are reasonably good\nat learning domains they have seen before; however, as soon as they encounter\nsomething new, they either go through a significant performance deterioration\nor if you try to teach them the new distribution of data, they forget what they\nhave learned before. Additionally, they are also prone to being overly\nconfident when performing inference on seen as well as unseen data, causing\nsignificant reliability issues when lives are at stake. Therefore, it is\nextremely important to dig into this problem and formulate an approach that\nwill be continually adaptable as well as reliable. If we move away from the\nengineering domain of such systems and look into biological systems, we can\nrealize that these very systems are very efficient at computing the reliance as\nwell as the uncertainty of accurate predictions that further help them refine\nthe inference in a life-long setting. These systems are not perfect; however,\nthey do give us a solid understanding of the reasoning under uncertainty which\ntakes us to the domain of Bayesian reasoning. We incorporate this Bayesian\ninference with thresholding mechanism as to mimic more biologically inspired\nmodels, but only at spatial level. Further, we reproduce a recent study on\nBayesian Inference with Spiking Neural Networks for Continual Learning to\ncompare against it as a suitable biologically inspired Bayesian framework.\nOverall, we investigate the plausibility of biologically inspired Bayesian\nLearning in artificial systems on a vision dataset, MNIST, and show relative\nperformance improvement under the conditions when the model is forced to\npredict VS when the model is not.\n", "date": "2024-11-27", "categories": []}
{"id": "2411.19154", "title": "DESIRE: Dynamic Knowledge Consolidation for Rehearsal-Free Continual\n  Learning", "abstract": "  Continual learning aims to equip models with the ability to retain previously\nlearned knowledge like a human. Recent work incorporating Parameter-Efficient\nFine-Tuning has revitalized the field by introducing lightweight extension\nmodules. However, existing methods usually overlook the issue of information\nleakage caused by the fact that the experiment data have been used in\npre-trained models. Once these duplicate data are removed in the pre-training\nphase, their performance can be severely affected. In this paper, we propose a\nnew LoRA-based rehearsal-free method named DESIRE. Our method avoids imposing\nadditional constraints during training to mitigate catastrophic forgetting,\nthereby maximizing the learning of new classes. To integrate knowledge from old\nand new tasks, we propose two efficient post-processing modules. On the one\nhand, we retain only two sets of LoRA parameters for merging and propose\ndynamic representation consolidation to calibrate the merged feature\nrepresentation. On the other hand, we propose decision boundary refinement to\naddress classifier bias when training solely on new class data. Extensive\nexperiments demonstrate that our method achieves state-of-the-art performance\non multiple datasets and strikes an effective balance between stability and\nplasticity. Our code will be publicly available.\n", "date": "2024-11-28", "categories": []}
{"id": "2411.04723", "title": "Exploring the Stability Gap in Continual Learning: The Role of the\n  Classification Head", "abstract": "  Continual learning (CL) has emerged as a critical area in machine learning,\nenabling neural networks to learn from evolving data distributions while\nmitigating catastrophic forgetting. However, recent research has identified the\nstability gap -- a phenomenon where models initially lose performance on\npreviously learned tasks before partially recovering during training. Such\nlearning dynamics are contradictory to the intuitive understanding of stability\nin continual learning where one would expect the performance to degrade\ngradually instead of rapidly decreasing and then partially recovering later. To\nbetter understand and alleviate the stability gap, we investigate it at\ndifferent levels of the neural network architecture, particularly focusing on\nthe role of the classification head. We introduce the nearest-mean classifier\n(NMC) as a tool to attribute the influence of the backbone and the\nclassification head on the stability gap. Our experiments demonstrate that NMC\nnot only improves final performance, but also significantly enhances training\nstability across various continual learning benchmarks, including CIFAR100,\nImageNet100, CUB-200, and FGVC Aircrafts. Moreover, we find that NMC also\nreduces task-recency bias. Our analysis provides new insights into the\nstability gap and suggests that the primary contributor to this phenomenon is\nthe linear head, rather than the insufficient representation learning.\n", "date": "2024-11-06", "categories": []}
{"id": "2411.17610", "title": "Modality-Incremental Learning with Disjoint Relevance Mapping Networks\n  for Image-based Semantic Segmentation", "abstract": "  In autonomous driving, environment perception has significantly advanced with\nthe utilization of deep learning techniques for diverse sensors such as\ncameras, depth sensors, or infrared sensors. The diversity in the sensor stack\nincreases the safety and contributes to robustness against adverse weather and\nlighting conditions. However, the variance in data acquired from different\nsensors poses challenges. In the context of continual learning (CL),\nincremental learning is especially challenging for considerably large domain\nshifts, e.g. different sensor modalities. This amplifies the problem of\ncatastrophic forgetting. To address this issue, we formulate the concept of\nmodality-incremental learning and examine its necessity, by contrasting it with\nexisting incremental learning paradigms. We propose the use of a modified\nRelevance Mapping Network (RMN) to incrementally learn new modalities while\npreserving performance on previously learned modalities, in which relevance\nmaps are disjoint. Experimental results demonstrate that the prevention of\nshared connections in this approach helps alleviate the problem of forgetting\nwithin the constraints of a strict continual learning framework.\n", "date": "2024-11-26", "categories": []}
{"id": "2311.02058", "title": "LOTUS: Continual Imitation Learning for Robot Manipulation Through\n  Unsupervised Skill Discovery", "abstract": "  We introduce LOTUS, a continual imitation learning algorithm that empowers a\nphysical robot to continuously and efficiently learn to solve new manipulation\ntasks throughout its lifespan. The core idea behind LOTUS is constructing an\never-growing skill library from a sequence of new tasks with a small number of\nhuman demonstrations. LOTUS starts with a continual skill discovery process\nusing an open-vocabulary vision model, which extracts skills as recurring\npatterns presented in unsegmented demonstrations. Continual skill discovery\nupdates existing skills to avoid catastrophic forgetting of previous tasks and\nadds new skills to solve novel tasks. LOTUS trains a meta-controller that\nflexibly composes various skills to tackle vision-based manipulation tasks in\nthe lifelong learning process. Our comprehensive experiments show that LOTUS\noutperforms state-of-the-art baselines by over 11% in success rate, showing its\nsuperior knowledge transfer ability compared to prior methods. More results and\nvideos can be found on the project website:\nhttps://ut-austin-rpl.github.io/Lotus/.\n", "date": "2023-11-03", "categories": []}
{"id": "2403.09613", "title": "Reawakening knowledge: Anticipatory recovery from catastrophic\n  interference via structured training", "abstract": "  We explore the training dynamics of neural networks in a structured non-IID\nsetting where documents are presented cyclically in a fixed, repeated sequence.\nTypically, networks suffer from catastrophic interference when training on a\nsequence of documents; however, we discover a curious and remarkable property\nof LLMs finetuned sequentially in this setting: they exhibit anticipatory\nbehavior, recovering from the forgetting on documents before encountering them\nagain. This behavior occurs even though the documents are never presented in\ncontext together. The behavior emerges and becomes more robust as the\narchitecture scales up its number of parameters. Through comprehensive\nexperiments and visualizations, we demonstrate a new mechanism by which\nover-parametrized neural networks can recover from catastrophic interference\nand uncover new insights into training over-parameterized networks in\ncyclically structured environments.\n", "date": "2024-03-14", "categories": []}
{"id": "2404.16789", "title": "Continual Learning of Large Language Models: A Comprehensive Survey", "abstract": "  The recent success of large language models (LLMs) trained on static,\npre-collected, general datasets has sparked numerous research directions and\napplications. One such direction addresses the non-trivial challenge of\nintegrating pre-trained LLMs into dynamic data distributions, task structures,\nand user preferences. Pre-trained LLMs, when tailored for specific needs, often\nexperience significant performance degradation in previous knowledge domains --\na phenomenon known as \"catastrophic forgetting\". While extensively studied in\nthe continual learning (CL) community, it presents new manifestations in the\nrealm of LLMs. In this survey, we provide a comprehensive overview of the\ncurrent research progress on LLMs within the context of CL. This survey is\nstructured into four main sections: we first describe an overview of\ncontinually learning LLMs, consisting of two directions of continuity: vertical\ncontinuity (or vertical continual learning), i.e., continual adaptation from\ngeneral to specific capabilities, and horizontal continuity (or horizontal\ncontinual learning), i.e., continual adaptation across time and domains\n(Section 3). We then summarize three stages of learning LLMs in the context of\nmodern CL: Continual Pre-Training (CPT), Domain-Adaptive Pre-training (DAP),\nand Continual Fine-Tuning (CFT) (Section 4). Then we provide an overview of\nevaluation protocols for continual learning with LLMs, along with the current\navailable data sources (Section 5). Finally, we discuss intriguing questions\npertaining to continual learning for LLMs (Section 6). The full list of papers\nexamined in this survey is available at\nhttps://github.com/Wang-ML-Lab/llm-continual-learning-survey.\n", "date": "2024-04-25", "categories": []}
{"id": "2411.01916", "title": "Masked Autoencoders are Parameter-Efficient Federated Continual Learners", "abstract": "  Federated learning is a specific distributed learning paradigm in which a\ncentral server aggregates updates from multiple clients' local models, thereby\nenabling the server to learn without requiring clients to upload their private\ndata, maintaining data privacy. While existing federated learning methods are\nprimarily designed for static data, real-world applications often require\nclients to learn new categories over time. This challenge necessitates the\nintegration of continual learning techniques, leading to federated continual\nlearning (FCL). To address both catastrophic forgetting and non-IID issues, we\npropose to use masked autoencoders (MAEs) as parameter-efficient federated\ncontinual learners, called pMAE. pMAE learns reconstructive prompt on the\nclient side through image reconstruction using MAE. On the server side, it\nreconstructs the uploaded restore information to capture the data distribution\nacross previous tasks and different clients, using these reconstructed images\nto fine-tune discriminative prompt and classifier parameters tailored for\nclassification, thereby alleviating catastrophic forgetting and non-IID issues\non a global scale. Experimental results demonstrate that pMAE achieves\nperformance comparable to existing prompt-based methods and can enhance their\neffectiveness, particularly when using self-supervised pre-trained transformers\nas the backbone. Code is available at: https://github.com/ycheoo/pMAE.\n", "date": "2024-11-04", "categories": []}
{"id": "2308.08812", "title": "A Fusion of Variational Distribution Priors and Saliency Map Replay for\n  Continual 3D Reconstruction", "abstract": "  Single-image 3D reconstruction is a research challenge focused on predicting\n3D object shapes from single-view images. This task requires significant data\nacquisition to predict both visible and occluded portions of the shape.\nFurthermore, learning-based methods face the difficulty of creating a\ncomprehensive training dataset for all possible classes. To this end, we\npropose a continual learning-based 3D reconstruction method where our goal is\nto design a model using Variational Priors that can still reconstruct the\npreviously seen classes reasonably even after training on new classes.\nVariational Priors represent abstract shapes and combat forgetting, whereas\nsaliency maps preserve object attributes with less memory usage. This is vital\ndue to resource constraints in storing extensive training data. Additionally,\nwe introduce saliency map-based experience replay to capture global and\ndistinct object features. Thorough experiments show competitive results\ncompared to established methods, both quantitatively and qualitatively.\n", "date": "2023-08-17", "categories": []}
{"id": "2311.15414", "title": "KOPPA: Improving Prompt-based Continual Learning with Key-Query\n  Orthogonal Projection and Prototype-based One-Versus-All", "abstract": "  Drawing inspiration from prompt tuning techniques applied to Large Language\nModels, recent methods based on pre-trained ViT networks have achieved\nremarkable results in the field of Continual Learning. Specifically, these\napproaches propose to maintain a set of prompts and allocate a subset of them\nto learn each task using a key-query matching strategy. However, they may\nencounter limitations when lacking control over the correlations between old\ntask queries and keys of future tasks, the shift of features in the latent\nspace, and the relative separation of latent vectors learned in independent\ntasks. In this work, we introduce a novel key-query learning strategy based on\northogonal projection, inspired by model-agnostic meta-learning, to enhance\nprompt matching efficiency and address the challenge of shifting features.\nFurthermore, we introduce a One-Versus-All (OVA) prototype-based component that\nenhances the classification head distinction. Experimental results on benchmark\ndatasets demonstrate that our method empowers the model to achieve results\nsurpassing those of current state-of-the-art approaches by a large margin of up\nto 20%.\n", "date": "2023-11-26", "categories": []}
{"id": "2407.14242", "title": "Continual Panoptic Perception: Towards Multi-modal Incremental\n  Interpretation of Remote Sensing Images", "abstract": "  Continual learning (CL) breaks off the one-way training manner and enables a\nmodel to adapt to new data, semantics and tasks continuously. However, current\nCL methods mainly focus on single tasks. Besides, CL models are plagued by\ncatastrophic forgetting and semantic drift since the lack of old data, which\noften occurs in remote-sensing interpretation due to the intricate fine-grained\nsemantics. In this paper, we propose Continual Panoptic Perception (CPP), a\nunified continual learning model that leverages multi-task joint learning\ncovering pixel-level classification, instance-level segmentation and\nimage-level perception for universal interpretation in remote sensing images.\nConcretely, we propose a collaborative cross-modal encoder (CCE) to extract the\ninput image features, which supports pixel classification and caption\ngeneration synchronously. To inherit the knowledge from the old model without\nexemplar memory, we propose a task-interactive knowledge distillation (TKD)\nmethod, which leverages cross-modal optimization and task-asymmetric\npseudo-labeling (TPL) to alleviate catastrophic forgetting. Furthermore, we\nalso propose a joint optimization mechanism to achieve end-to-end multi-modal\npanoptic perception. Experimental results on the fine-grained panoptic\nperception dataset validate the effectiveness of the proposed model, and also\nprove that joint optimization can boost sub-task CL efficiency with over 13\\%\nrelative improvement on panoptic quality.\n", "date": "2024-07-19", "categories": []}
{"id": "2411.13740", "title": "Federated Continual Learning for Edge-AI: A Comprehensive Survey", "abstract": "  Edge-AI, the convergence of edge computing and artificial intelligence (AI),\nhas become a promising paradigm that enables the deployment of advanced AI\nmodels at the network edge, close to users. In Edge-AI, federated continual\nlearning (FCL) has emerged as an imperative framework, which fuses knowledge\nfrom different clients while preserving data privacy and retaining knowledge\nfrom previous tasks as it learns new ones. By so doing, FCL aims to ensure\nstable and reliable performance of learning models in dynamic and distributed\nenvironments. In this survey, we thoroughly review the state-of-the-art\nresearch and present the first comprehensive survey of FCL for Edge-AI. We\ncategorize FCL methods based on three task characteristics: federated class\ncontinual learning, federated domain continual learning, and federated task\ncontinual learning. For each category, an in-depth investigation and review of\nthe representative methods are provided, covering background, challenges,\nproblem formalisation, solutions, and limitations. Besides, existing real-world\napplications empowered by FCL are reviewed, indicating the current progress and\npotential of FCL in diverse application domains. Furthermore, we discuss and\nhighlight several prospective research directions of FCL such as\nalgorithm-hardware co-design for FCL and FCL with foundation models, which\ncould provide insights into the future development and practical deployment of\nFCL in the era of Edge-AI.\n", "date": "2024-11-20", "categories": []}
{"id": "2411.13852", "title": "Dealing with Synthetic Data Contamination in Online Continual Learning", "abstract": "  Image generation has shown remarkable results in generating high-fidelity\nrealistic images, in particular with the advancement of diffusion-based models.\nHowever, the prevalence of AI-generated images may have side effects for the\nmachine learning community that are not clearly identified. Meanwhile, the\nsuccess of deep learning in computer vision is driven by the massive dataset\ncollected on the Internet. The extensive quantity of synthetic data being added\nto the Internet would become an obstacle for future researchers to collect\n\"clean\" datasets without AI-generated content. Prior research has shown that\nusing datasets contaminated by synthetic images may result in performance\ndegradation when used for training. In this paper, we investigate the potential\nimpact of contaminated datasets on Online Continual Learning (CL) research. We\nexperimentally show that contaminated datasets might hinder the training of\nexisting online CL methods. Also, we propose Entropy Selection with\nReal-synthetic similarity Maximization (ESRM), a method to alleviate the\nperformance deterioration caused by synthetic images when training online CL\nmodels. Experiments show that our method can significantly alleviate\nperformance deterioration, especially when the contamination is severe. For\nreproducibility, the source code of our work is available at\nhttps://github.com/maorong-wang/ESRM.\n", "date": "2024-11-21", "categories": []}
{"id": "2411.14202", "title": "Revised Regularization for Efficient Continual Learning through\n  Correlation-Based Parameter Update in Bayesian Neural Networks", "abstract": "  We propose a Bayesian neural network-based continual learning algorithm using\nVariational Inference, aiming to overcome several drawbacks of existing\nmethods. Specifically, in continual learning scenarios, storing network\nparameters at each step to retain knowledge poses challenges. This is\ncompounded by the crucial need to mitigate catastrophic forgetting,\nparticularly given the limited access to past datasets, which complicates\nmaintaining correspondence between network parameters and datasets across all\nsessions. Current methods using Variational Inference with KL divergence risk\ncatastrophic forgetting during uncertain node updates and coupled disruptions\nin certain nodes. To address these challenges, we propose the following\nstrategies. To reduce the storage of the dense layer parameters, we propose a\nparameter distribution learning method that significantly reduces the storage\nrequirements. In the continual learning framework employing variational\ninference, our study introduces a regularization term that specifically targets\nthe dynamics and population of the mean and variance of the parameters. This\nterm aims to retain the benefits of KL divergence while addressing related\nchallenges. To ensure proper correspondence between network parameters and the\ndata, our method introduces an importance-weighted Evidence Lower Bound term to\ncapture data and parameter correlations. This enables storage of common and\ndistinctive parameter hyperspace bases. The proposed method partitions the\nparameter space into common and distinctive subspaces, with conditions for\neffective backward and forward knowledge transfer, elucidating the\nnetwork-parameter dataset correspondence. The experimental results demonstrate\nthe effectiveness of our method across diverse datasets and various\ncombinations of sequential datasets, yielding superior performance compared to\nexisting approaches.\n", "date": "2024-11-21", "categories": []}
{"id": "2402.08823", "title": "Random Representations Outperform Online Continually Learned\n  Representations", "abstract": "  Continual learning has primarily focused on the issue of catastrophic\nforgetting and the associated stability-plasticity tradeoffs. However, little\nattention has been paid to the efficacy of continually learned representations,\nas representations are learned alongside classifiers throughout the learning\nprocess. Our primary contribution is empirically demonstrating that existing\nonline continually trained deep networks produce inferior representations\ncompared to a simple pre-defined random transforms. Our approach projects raw\npixels using a fixed random transform, approximating an RBF-Kernel initialized\nbefore any data is seen. We then train a simple linear classifier on top\nwithout storing any exemplars, processing one sample at a time in an online\ncontinual learning setting. This method, called RanDumb, significantly\noutperforms state-of-the-art continually learned representations across all\nstandard online continual learning benchmarks. Our study reveals the\nsignificant limitations of representation learning, particularly in\nlow-exemplar and online continual learning scenarios. Extending our\ninvestigation to popular exemplar-free scenarios with pretrained models, we\nfind that training only a linear classifier on top of pretrained\nrepresentations surpasses most continual fine-tuning and prompt-tuning\nstrategies. Overall, our investigation challenges the prevailing assumptions\nabout effective representation learning in online continual learning. Our code\nis available at://github.com/drimpossible/RanDumb.\n", "date": "2024-02-13", "categories": []}
{"id": "2312.02491", "title": "Pseudo Replay-based Class Continual Learning for Online New Category\n  Anomaly Detection in Advanced Manufacturing", "abstract": "  The incorporation of advanced sensors and machine learning techniques has\nenabled modern manufacturing enterprises to perform data-driven\nclassification-based anomaly detection based on the sensor data collected in\nmanufacturing processes. However, one critical challenge is that newly\npresented defect category may manifest as the manufacturing process continues,\nresulting in monitoring performance deterioration of previously trained machine\nlearning models. Hence, there is an increasing need for empowering machine\nlearning models to learn continually. Among all continual learning methods,\nmemory-based continual learning has the best performance but faces the\nconstraints of data storage capacity. To address this issue, this paper\ndevelops a novel pseudo replay-based continual learning framework by\nintegrating class incremental learning and oversampling-based data generation.\nWithout storing all the data, the developed framework could generate\nhigh-quality data representing previous classes to train machine learning model\nincrementally when new category anomaly occurs. In addition, it could even\nenhance the monitoring performance since it also effectively improves the data\nquality. The effectiveness of the proposed framework is validated in three\ncases studies, which leverages supervised classification problem for anomaly\ndetection. The experimental results show that the developed method is very\npromising in detecting novel anomaly while maintaining a good performance on\nthe previous task and brings up more flexibility in model architecture.\n", "date": "2023-12-05", "categories": []}
{"id": "2411.10809", "title": "Stable Continual Reinforcement Learning via Diffusion-based Trajectory\n  Replay", "abstract": "  Given the inherent non-stationarity prevalent in real-world applications,\ncontinual Reinforcement Learning (RL) aims to equip the agent with the\ncapability to address a series of sequentially presented decision-making tasks.\nWithin this problem setting, a pivotal challenge revolves around\n\\textit{catastrophic forgetting} issue, wherein the agent is prone to\neffortlessly erode the decisional knowledge associated with past encountered\ntasks when learning the new one. In recent progresses, the \\textit{generative\nreplay} methods have showcased substantial potential by employing generative\nmodels to replay data distribution of past tasks. Compared to storing the data\nfrom past tasks directly, this category of methods circumvents the growing\nstorage overhead and possible data privacy concerns. However, constrained by\nthe expressive capacity of generative models, existing \\textit{generative\nreplay} methods face challenges in faithfully reconstructing the data\ndistribution of past tasks, particularly in scenarios with a myriad of tasks or\nhigh-dimensional data. Inspired by the success of diffusion models in various\ngenerative tasks, this paper introduces a novel continual RL algorithm DISTR\n(Diffusion-based Trajectory Replay) that employs a diffusion model to memorize\nthe high-return trajectory distribution of each encountered task and wakeups\nthese distributions during the policy learning on new tasks. Besides,\nconsidering the impracticality of replaying all past data each time, a\nprioritization mechanism is proposed to prioritize the trajectory replay of\npivotal tasks in our method. Empirical experiments on the popular continual RL\nbenchmark \\texttt{Continual World} demonstrate that our proposed method obtains\na favorable balance between \\textit{stability} and \\textit{plasticity},\nsurpassing various existing continual RL baselines in average success rate.\n", "date": "2024-11-16", "categories": []}
{"id": "2411.10367", "title": "Continual Adversarial Reinforcement Learning (CARL) of False Data\n  Injection detection: forgetting and explainability", "abstract": "  False data injection attacks (FDIAs) on smart inverters are a growing concern\nlinked to increased renewable energy production. While data-based FDIA\ndetection methods are also actively developed, we show that they remain\nvulnerable to impactful and stealthy adversarial examples that can be crafted\nusing Reinforcement Learning (RL). We propose to include such adversarial\nexamples in data-based detection training procedure via a continual adversarial\nRL (CARL) approach. This way, one can pinpoint the deficiencies of data-based\ndetection, thereby offering explainability during their incremental\nimprovement. We show that a continual learning implementation is subject to\ncatastrophic forgetting, and additionally show that forgetting can be addressed\nby employing a joint training strategy on all generated FDIA scenarios.\n", "date": "2024-11-15", "categories": []}
{"id": "2410.06645", "title": "Continual Learning in the Frequency Domain", "abstract": "  Continual learning (CL) is designed to learn new tasks while preserving\nexisting knowledge. Replaying samples from earlier tasks has proven to be an\neffective method to mitigate the forgetting of previously acquired knowledge.\nHowever, the current research on the training efficiency of rehearsal-based\nmethods is insufficient, which limits the practical application of CL systems\nin resource-limited scenarios. The human visual system (HVS) exhibits varying\nsensitivities to different frequency components, enabling the efficient\nelimination of visually redundant information. Inspired by HVS, we propose a\nnovel framework called Continual Learning in the Frequency Domain (CLFD). To\nour knowledge, this is the first study to utilize frequency domain features to\nenhance the performance and efficiency of CL training on edge devices. For the\ninput features of the feature extractor, CLFD employs wavelet transform to map\nthe original input image into the frequency domain, thereby effectively\nreducing the size of input feature maps. Regarding the output features of the\nfeature extractor, CLFD selectively utilizes output features for distinct\nclasses for classification, thereby balancing the reusability and interference\nof output features based on the frequency domain similarity of the classes\nacross various tasks. Optimizing only the input and output features of the\nfeature extractor allows for seamless integration of CLFD with various\nrehearsal-based methods. Extensive experiments conducted in both cloud and edge\nenvironments demonstrate that CLFD consistently improves the performance of\nstate-of-the-art (SOTA) methods in both precision and training efficiency.\nSpecifically, CLFD can increase the accuracy of the SOTA CL method by up to\n6.83% and reduce the training time by 2.6$\\times$.\n", "date": "2024-10-09", "categories": []}
{"id": "2408.09984", "title": "Boosting Open-Domain Continual Learning via Leveraging Intra-domain\n  Category-aware Prototype", "abstract": "  Despite recent progress in enhancing the efficacy of Open-Domain Continual\nLearning (ODCL) in Vision-Language Models (VLM), failing to (1) correctly\nidentify the Task-ID of a test image and (2) use only the category set\ncorresponding to the Task-ID, while preserving the knowledge related to each\ndomain, cannot address the two primary challenges of ODCL: forgetting old\nknowledge and maintaining zero-shot capabilities, as well as the confusions\ncaused by category-relatedness between domains. In this paper, we propose a\nsimple yet effective solution: leveraging intra-domain category-aware\nprototypes for ODCL in CLIP (DPeCLIP), where the prototype is the key to\nbridging the above two processes. Concretely, we propose a training-free\nTask-ID discriminator method, by utilizing prototypes as classifiers for\nidentifying Task-IDs. Furthermore, to maintain the knowledge corresponding to\neach domain, we incorporate intra-domain category-aware prototypes as domain\nprior prompts into the training process. Extensive experiments conducted on 11\ndifferent datasets demonstrate the effectiveness of our approach, achieving\n2.37% and 1.14% average improvement in class-incremental and task-incremental\nsettings, respectively.\n", "date": "2024-08-19", "categories": []}
{"id": "2410.20178", "title": "LLMs Can Evolve Continually on Modality for X-Modal Reasoning", "abstract": "  Multimodal Large Language Models (MLLMs) have gained significant attention\ndue to their impressive capabilities in multimodal understanding. However,\nexisting methods rely heavily on extensive modal-specific pretraining and\njoint-modal tuning, leading to significant computational burdens when expanding\nto new modalities. In this paper, we propose PathWeave, a flexible and scalable\nframework with modal-Path sWitching and ExpAnsion abilities that enables MLLMs\nto continually EVolve on modalities for $\\mathbb{X}$-modal reasoning. We\nleverage the concept of Continual Learning and develop an incremental training\nstrategy atop pre-trained MLLMs, enabling their expansion to new modalities\nusing uni-modal data, without executing joint-modal pretraining. In detail, a\nnovel Adapter-in-Adapter (AnA) framework is introduced, in which uni-modal and\ncross-modal adapters are seamlessly integrated to facilitate efficient modality\nalignment and collaboration. Additionally, an MoE-based gating module is\napplied between two types of adapters to further enhance the multimodal\ninteraction. To investigate the proposed method, we establish a challenging\nbenchmark called Continual Learning of Modality (MCL), which consists of\nhigh-quality QA data from five distinct modalities: image, video, audio, depth\nand point cloud. Extensive experiments demonstrate the effectiveness of the\nproposed AnA framework on learning plasticity and memory stability during\ncontinual learning. Furthermore, PathWeave performs comparably to\nstate-of-the-art MLLMs while concurrently reducing parameter training burdens\nby 98.73%. Our code locates at https://github.com/JiazuoYu/PathWeave\n", "date": "2024-10-26", "categories": []}
{"id": "2401.14807", "title": "PL-FSCIL: Harnessing the Power of Prompts for Few-Shot Class-Incremental\n  Learning", "abstract": "  Few-Shot Class-Incremental Learning (FSCIL) aims to enable deep neural\nnetworks to learn new tasks incrementally from a small number of labeled\nsamples without forgetting previously learned tasks, closely mimicking human\nlearning patterns. In this paper, we propose a novel approach called Prompt\nLearning for FSCIL (PL-FSCIL), which harnesses the power of prompts in\nconjunction with a pre-trained Vision Transformer (ViT) model to address the\nchallenges of FSCIL effectively. Our work pioneers the use of visual prompts in\nFSCIL, which is characterized by its notable simplicity. PL-FSCIL consists of\ntwo distinct prompts: the Domain Prompt and the FSCIL Prompt. Both are vectors\nthat augment the model by embedding themselves into the attention layer of the\nViT model. Specifically, the Domain Prompt assists the ViT model in adapting to\nnew data domains. The task-specific FSCIL Prompt, coupled with a prototype\nclassifier, amplifies the model's ability to effectively handle FSCIL tasks. We\nvalidate the efficacy of PL-FSCIL on widely used benchmark datasets such as\nCIFAR-100 and CUB-200. The results showcase competitive performance,\nunderscoring its promising potential for real-world applications where\nhigh-quality data is often scarce. The source code is available at:\nhttps://github.com/TianSongS/PL-FSCIL.\n", "date": "2024-01-26", "categories": []}
{"id": "2411.02564", "title": "Continual LLaVA: Continual Instruction Tuning in Large Vision-Language\n  Models", "abstract": "  Instruction tuning constitutes a prevalent technique for tailoring Large\nVision Language Models (LVLMs) to meet individual task requirements. To date,\nmost of the existing approaches are confined to single-task adaptation, whereas\nthe requirements in real-world scenarios are inherently varied and continually\nevolving. Thus an ideal LVLM should sustain continual instruction tuning in the\nface of stream-task distributions (i.e., different domains, emerging\ncapabilities, and new datasets) while minimizing the forgetting of previously\nacquired knowledge. To achieve this, we propose a new benchmark for COntinuAl\ninStruction Tuning on LVLMs (COAST), which encompasses the aforementioned\ndomain-incremental, capability-incremental, and dataset-incremental\nconfigurations. In terms of methodology, we propose Continual LLaVA, a\nrehearsal-free method tailored for continual instruction tuning in LVLMs. To\ncircumvent the additional overhead associated with experience replay, we freeze\nLVLMs and construct the dual increment embeddings for each input instruction to\nfacilitate parameter-efficient tuning. Specifically, the increment embeddings\ncan be decomposed into two principal components: 1) intrinsic increment\nembeddings to encode task-specific characteristics. To achieve this, we set up\na low-rank pool containing candidate embeddings, from which we select the\nrelevant ones based on their similarity with the user instructions; 2)\ncontextual increment embeddings to investigate the inter-dependencies across\ntasks. In this regard, the low-rank embeddings chosen in the previous tasks are\naggregated via learnable weighted sum to provide complementary hints. Extensive\nexperiments indicate that the proposed Continual LLaVA outperforms previous\nmethods by significantly reducing the forgetting during the continual\ninstruction tuning process.\n", "date": "2024-11-04", "categories": []}
{"id": "2411.05846", "title": "Reducing catastrophic forgetting of incremental learning in the absence\n  of rehearsal memory with task-specific token", "abstract": "  Deep learning models generally display catastrophic forgetting when learning\nnew data continuously. Many incremental learning approaches address this\nproblem by reusing data from previous tasks while learning new tasks. However,\nthe direct access to past data generates privacy and security concerns. To\naddress these issues, we present a novel method that preserves previous\nknowledge without storing previous data. This method is inspired by the\narchitecture of a vision transformer and employs a unique token capable of\nencapsulating the compressed knowledge of each task. This approach generates\ntask-specific embeddings by directing attention differently based on the task\nassociated with the data, thereby effectively mimicking the impact of having\nmultiple models through tokens. Our method incorporates a distillation process\nthat ensures efficient interactions even after multiple additional learning\nsteps, thereby optimizing the model against forgetting. We measured the\nperformance of our model in terms of accuracy and backward transfer using a\nbenchmark dataset for different task-incremental learning scenarios. Our\nresults demonstrate the superiority of our approach, which achieved the highest\naccuracy and lowest backward transfer among the compared methods. In addition\nto presenting a new model, our approach lays the foundation for various\nextensions within the spectrum of vision-transformer architectures.\n", "date": "2024-11-06", "categories": []}
{"id": "2411.06171", "title": "SEEKR: Selective Attention-Guided Knowledge Retention for Continual\n  Learning of Large Language Models", "abstract": "  Continual learning (CL) is crucial for language models to dynamically adapt\nto the evolving real-world demands. To mitigate the catastrophic forgetting\nproblem in CL, data replay has been proven a simple and effective strategy, and\nthe subsequent data-replay-based distillation can further enhance the\nperformance. However, existing methods fail to fully exploit the knowledge\nembedded in models from previous tasks, resulting in the need for a relatively\nlarge number of replay samples to achieve good results. In this work, we first\nexplore and emphasize the importance of attention weights in knowledge\nretention, and then propose a SElective attEntion-guided Knowledge Retention\nmethod (SEEKR) for data-efficient replay-based continual learning of large\nlanguage models (LLMs). Specifically, SEEKR performs attention distillation on\nthe selected attention heads for finer-grained knowledge retention, where the\nproposed forgettability-based and task-sensitivity-based measures are used to\nidentify the most valuable attention heads. Experimental results on two\ncontinual learning benchmarks for LLMs demonstrate the superiority of SEEKR\nover the existing methods on both performance and efficiency. Explicitly, SEEKR\nachieves comparable or even better performance with only 1/10 of the replayed\ndata used by other methods, and reduces the proportion of replayed data to 1%.\n", "date": "2024-11-09", "categories": []}
{"id": "2411.06618", "title": "Using Diffusion Models as Generative Replay in Continual Federated\n  Learning -- What will Happen?", "abstract": "  Federated learning (FL) has become a cornerstone in decentralized learning,\nwhere, in many scenarios, the incoming data distribution will change\ndynamically over time, introducing continuous learning (CL) problems. This\ncontinual federated learning (CFL) task presents unique challenges,\nparticularly regarding catastrophic forgetting and non-IID input data. Existing\nsolutions include using a replay buffer to store historical data or leveraging\ngenerative adversarial networks. Nevertheless, motivated by recent advancements\nin the diffusion model for generative tasks, this paper introduces DCFL, a\nnovel framework tailored to address the challenges of CFL in dynamic\ndistributed learning environments. Our approach harnesses the power of the\nconditional diffusion model to generate synthetic historical data at each local\ndevice during communication, effectively mitigating latent shifts in dynamic\ndata distribution inputs. We provide the convergence bound for the proposed CFL\nframework and demonstrate its promising performance across multiple datasets,\nshowcasing its effectiveness in tackling the complexities of CFL tasks.\n", "date": "2024-11-10", "categories": []}
{"id": "2411.06634", "title": "Inductive Graph Few-shot Class Incremental Learning", "abstract": "  Node classification with Graph Neural Networks (GNN) under a fixed set of\nlabels is well known in contrast to Graph Few-Shot Class Incremental Learning\n(GFSCIL), which involves learning a GNN classifier as graph nodes and classes\ngrowing over time sporadically. We introduce inductive GFSCIL that continually\nlearns novel classes with newly emerging nodes while maintaining performance on\nold classes without accessing previous data. This addresses the practical\nconcern of transductive GFSCIL, which requires storing the entire graph with\nhistorical data. Compared to the transductive GFSCIL, the inductive setting\nexacerbates catastrophic forgetting due to inaccessible previous data during\nincremental training, in addition to overfitting issue caused by label\nsparsity. Thus, we propose a novel method, called Topology-based class\nAugmentation and Prototype calibration (TAP). To be specific, it first creates\na triple-branch multi-topology class augmentation method to enhance model\ngeneralization ability. As each incremental session receives a disjoint\nsubgraph with nodes of novel classes, the multi-topology class augmentation\nmethod helps replicate such a setting in the base session to boost backbone\nversatility. In incremental learning, given the limited number of novel class\nsamples, we propose an iterative prototype calibration to improve the\nseparation of class prototypes. Furthermore, as backbone fine-tuning poses the\nfeature distribution drift, prototypes of old classes start failing over time,\nwe propose the prototype shift method for old classes to compensate for the\ndrift. We showcase the proposed method on four datasets.\n", "date": "2024-11-11", "categories": []}
{"id": "2411.06659", "title": "An Efficient Memory Module for Graph Few-Shot Class-Incremental Learning", "abstract": "  Incremental graph learning has gained significant attention for its ability\nto address the catastrophic forgetting problem in graph representation\nlearning. However, traditional methods often rely on a large number of labels\nfor node classification, which is impractical in real-world applications. This\nmakes few-shot incremental learning on graphs a pressing need. Current methods\ntypically require extensive training samples from meta-learning to build memory\nand perform intensive fine-tuning of GNN parameters, leading to high memory\nconsumption and potential loss of previously learned knowledge. To tackle these\nchallenges, we introduce Mecoin, an efficient method for building and\nmaintaining memory. Mecoin employs Structured Memory Units to cache prototypes\nof learned categories, as well as Memory Construction Modules to update these\nprototypes for new categories through interactions between the nodes and the\ncached prototypes. Additionally, we have designed a Memory Representation\nAdaptation Module to store probabilities associated with each class prototype,\nreducing the need for parameter fine-tuning and lowering the forgetting rate.\nWhen a sample matches its corresponding class prototype, the relevant\nprobabilities are retrieved from the MRaM. Knowledge is then distilled back\ninto the GNN through a Graph Knowledge Distillation Module, preserving the\nmodel's memory. We analyze the effectiveness of Mecoin in terms of\ngeneralization error and explore the impact of different distillation\nstrategies on model performance through experiments and VC-dimension analysis.\nCompared to other related works, Mecoin shows superior performance in accuracy\nand forgetting rate. Our code is publicly available on the\nhttps://github.com/Arvin0313/Mecoin-GFSCIL.git .\n", "date": "2024-11-11", "categories": []}
{"id": "2411.06764", "title": "Multi-Stage Knowledge Integration of Vision-Language Models for\n  Continual Learning", "abstract": "  Vision Language Models (VLMs), pre-trained on large-scale image-text\ndatasets, enable zero-shot predictions for unseen data but may underperform on\nspecific unseen tasks. Continual learning (CL) can help VLMs effectively adapt\nto new data distributions without joint training, but faces challenges of\ncatastrophic forgetting and generalization forgetting. Although significant\nprogress has been achieved by distillation-based methods, they exhibit two\nsevere limitations. One is the popularly adopted single-teacher paradigm fails\nto impart comprehensive knowledge, The other is the existing methods\ninadequately leverage the multimodal information in the original training\ndataset, instead they rely on additional data for distillation, which increases\ncomputational and storage overhead. To mitigate both limitations, by drawing on\nKnowledge Integration Theory (KIT), we propose a Multi-Stage Knowledge\nIntegration network (MulKI) to emulate the human learning process in\ndistillation methods. MulKI achieves this through four stages, including\nEliciting Ideas, Adding New Ideas, Distinguishing Ideas, and Making\nConnections. During the four stages, we first leverage prototypes to align\nacross modalities, eliciting cross-modal knowledge, then adding new knowledge\nby constructing fine-grained intra- and inter-modality relationships with\nprototypes. After that, knowledge from two teacher models is adaptively\ndistinguished and re-weighted. Finally, we connect between models from intra-\nand inter-task, integrating preceding and new knowledge. Our method\ndemonstrates significant improvements in maintaining zero-shot capabilities\nwhile supporting continual learning across diverse downstream tasks, showcasing\nits potential in adapting VLMs to evolving data distributions.\n", "date": "2024-11-11", "categories": []}
{"id": "2306.08448", "title": "Kalman Filter for Online Classification of Non-Stationary Data", "abstract": "  In Online Continual Learning (OCL) a learning system receives a stream of\ndata and sequentially performs prediction and training steps. Important\nchallenges in OCL are concerned with automatic adaptation to the particular\nnon-stationary structure of the data, and with quantification of predictive\nuncertainty. Motivated by these challenges we introduce a probabilistic\nBayesian online learning model by using a (possibly pretrained) neural\nrepresentation and a state space model over the linear predictor weights.\nNon-stationarity over the linear predictor weights is modelled using a\nparameter drift transition density, parametrized by a coefficient that\nquantifies forgetting. Inference in the model is implemented with efficient\nKalman filter recursions which track the posterior distribution over the linear\nweights, while online SGD updates over the transition dynamics coefficient\nallows to adapt to the non-stationarity seen in data. While the framework is\ndeveloped assuming a linear Gaussian model, we also extend it to deal with\nclassification problems and for fine-tuning the deep learning representation.\nIn a set of experiments in multi-class classification using data sets such as\nCIFAR-100 and CLOC we demonstrate the predictive ability of the model and its\nflexibility to capture non-stationarity.\n", "date": "2023-06-14", "categories": []}
{"id": "2311.05241", "title": "When Meta-Learning Meets Online and Continual Learning: A Survey", "abstract": "  Over the past decade, deep neural networks have demonstrated significant\nsuccess using the training scheme that involves mini-batch stochastic gradient\ndescent on extensive datasets. Expanding upon this accomplishment, there has\nbeen a surge in research exploring the application of neural networks in other\nlearning scenarios. One notable framework that has garnered significant\nattention is meta-learning. Often described as \"learning to learn,\"\nmeta-learning is a data-driven approach to optimize the learning algorithm.\nOther branches of interest are continual learning and online learning, both of\nwhich involve incrementally updating a model with streaming data. While these\nframeworks were initially developed independently, recent works have started\ninvestigating their combinations, proposing novel problem settings and learning\nalgorithms. However, due to the elevated complexity and lack of unified\nterminology, discerning differences between the learning frameworks can be\nchallenging even for experienced researchers. To facilitate a clear\nunderstanding, this paper provides a comprehensive survey that organizes\nvarious problem settings using consistent terminology and formal descriptions.\nBy offering an overview of these learning paradigms, our work aims to foster\nfurther advancements in this promising area of research.\n", "date": "2023-11-09", "categories": []}
{"id": "2411.04034", "title": "Non-Stationary Learning of Neural Networks with Automatic Soft Parameter\n  Reset", "abstract": "  Neural networks are traditionally trained under the assumption that data come\nfrom a stationary distribution. However, settings which violate this assumption\nare becoming more popular; examples include supervised learning under\ndistributional shifts, reinforcement learning, continual learning and\nnon-stationary contextual bandits. In this work we introduce a novel learning\napproach that automatically models and adapts to non-stationarity, via an\nOrnstein-Uhlenbeck process with an adaptive drift parameter. The adaptive drift\ntends to draw the parameters towards the initialisation distribution, so the\napproach can be understood as a form of soft parameter reset. We show\nempirically that our approach performs well in non-stationary supervised and\noff-policy reinforcement learning settings.\n", "date": "2024-11-06", "categories": []}
{"id": "2411.05663", "title": "Online-LoRA: Task-free Online Continual Learning via Low Rank Adaptation", "abstract": "  Catastrophic forgetting is a significant challenge in online continual\nlearning (OCL), especially for non-stationary data streams that do not have\nwell-defined task boundaries. This challenge is exacerbated by the memory\nconstraints and privacy concerns inherent in rehearsal buffers. To tackle\ncatastrophic forgetting, in this paper, we introduce Online-LoRA, a novel\nframework for task-free OCL. Online-LoRA allows to finetune pre-trained Vision\nTransformer (ViT) models in real-time to address the limitations of rehearsal\nbuffers and leverage pre-trained models' performance benefits. As the main\ncontribution, our approach features a novel online weight regularization\nstrategy to identify and consolidate important model parameters. Moreover,\nOnline-LoRA leverages the training dynamics of loss values to enable the\nautomatic recognition of the data distribution shifts. Extensive experiments\nacross many task-free OCL scenarios and benchmark datasets (including\nCIFAR-100, ImageNet-R, ImageNet-S, CUB-200 and CORe50) demonstrate that\nOnline-LoRA can be robustly adapted to various ViT architectures, while\nachieving better performance compared to SOTA methods. Our code will be\npublicly available at:\nhttps://github.com/Christina200/Online-LoRA-official.git.\n", "date": "2024-11-08", "categories": []}
{"id": "2402.05797", "title": "TaE: Task-aware Expandable Representation for Long Tail Class\n  Incremental Learning", "abstract": "  Class-incremental learning is dedicated to the development of deep learning\nmodels that are capable of acquiring new knowledge while retaining previously\nlearned information. Most methods focus on balanced data distribution for each\ntask, overlooking real-world long-tailed distributions. Therefore, Long-Tailed\nClass-Incremental Learning has been introduced, which trains on data where head\nclasses have more samples than tail classes. Existing methods mainly focus on\npreserving representative samples from previous classes to combat catastrophic\nforgetting. Recently, dynamic network algorithms freeze old network structures\nand expand new ones, achieving significant performance. However, with the\nintroduction of the long-tail problem, merely extending Determined blocks can\nlead to miscalibrated predictions, while expanding the entire backbone results\nin an explosion of memory size. To address these issues, we introduce a novel\nTask-aware Expandable (TaE) framework, dynamically allocating and updating\ntask-specific trainable parameters to learn diverse representations from each\nincremental task while resisting forgetting through the majority of frozen\nmodel parameters. To further encourage the class-specific feature\nrepresentation, we develop a Centroid-Enhanced (CEd) method to guide the update\nof these task-aware parameters. This approach is designed to adaptively\nallocate feature space for every class by adjusting the distance between intra-\nand inter-class features, which can extend to all \"training from sketch\"\nalgorithms. Extensive experiments demonstrate that TaE achieves\nstate-of-the-art performance.\n", "date": "2024-02-08", "categories": []}
{"id": "2307.05707", "title": "MoP-CLIP: A Mixture of Prompt-Tuned CLIP Models for Domain Incremental\n  Learning", "abstract": "  Despite the recent progress in incremental learning, addressing catastrophic\nforgetting under distributional drift is still an open and important problem.\nIndeed, while state-of-the-art domain incremental learning (DIL) methods\nperform satisfactorily within known domains, their performance largely degrades\nin the presence of novel domains. This limitation hampers their\ngeneralizability, and restricts their scalability to more realistic settings\nwhere train and test data are drawn from different distributions. To address\nthese limitations, we present a novel DIL approach based on a mixture of\nprompt-tuned CLIP models (MoP-CLIP), which generalizes the paradigm of\nS-Prompting to handle both in-distribution and out-of-distribution data at\ninference. In particular, at the training stage we model the features\ndistribution of every class in each domain, learning individual text and visual\nprompts to adapt to a given domain. At inference, the learned distributions\nallow us to identify whether a given test sample belongs to a known domain,\nselecting the correct prompt for the classification task, or from an unseen\ndomain, leveraging a mixture of the prompt-tuned CLIP models. Our empirical\nevaluation reveals the poor performance of existing DIL methods under domain\nshift, and suggests that the proposed MoP-CLIP performs competitively in the\nstandard DIL settings while outperforming state-of-the-art methods in OOD\nscenarios. These results demonstrate the superiority of MoP-CLIP, offering a\nrobust and general solution to the problem of domain incremental learning.\n", "date": "2023-07-11", "categories": []}
{"id": "2405.17779", "title": "Online Analytic Exemplar-Free Continual Learning with Large Models for\n  Imbalanced Autonomous Driving Task", "abstract": "  In autonomous driving, even a meticulously trained model can encounter\nfailures when facing unfamiliar scenarios. One of these scenarios can be\nformulated as an online continual learning (OCL) problem. That is, data come in\nan online fashion, and models are updated according to these streaming data.\nTwo major OCL challenges are catastrophic forgetting and data imbalance. To\naddress these challenges, in this paper, we propose an Analytic Exemplar-Free\nOnline Continual Learning algorithm (AEF-OCL). The AEF-OCL leverages analytic\ncontinual learning principles and employs ridge regression as a classifier for\nfeatures extracted by a large backbone network. It solves the OCL problem by\nrecursively calculating the analytical solution, ensuring an equalization\nbetween the continual learning and its joint-learning counterpart, and works\nwithout the need to save any used samples (i.e., exemplar-free). Additionally,\nwe introduce a Pseudo-Features Generator (PFG) module that recursively\nestimates the mean and the variance of real features for each class. It\nover-samples offset pseudo-features from the same normal distribution as the\nreal features, thereby addressing the data imbalance issue. Experimental\nresults demonstrate that despite being an exemplar-free strategy, our method\noutperforms various methods on the autonomous driving SODA10M dataset. Source\ncode is available at https://github.com/ZHUANGHP/Analytic-continual-learning.\n", "date": "2024-05-28", "categories": []}
{"id": "2411.02860", "title": "Continual Audio-Visual Sound Separation", "abstract": "  In this paper, we introduce a novel continual audio-visual sound separation\ntask, aiming to continuously separate sound sources for new classes while\npreserving performance on previously learned classes, with the aid of visual\nguidance. This problem is crucial for practical visually guided auditory\nperception as it can significantly enhance the adaptability and robustness of\naudio-visual sound separation models, making them more applicable for\nreal-world scenarios where encountering new sound sources is commonplace. The\ntask is inherently challenging as our models must not only effectively utilize\ninformation from both modalities in current tasks but also preserve their\ncross-modal association in old tasks to mitigate catastrophic forgetting during\naudio-visual continual learning. To address these challenges, we propose a\nnovel approach named ContAV-Sep (\\textbf{Cont}inual\n\\textbf{A}udio-\\textbf{V}isual Sound \\textbf{Sep}aration). ContAV-Sep presents\na novel Cross-modal Similarity Distillation Constraint (CrossSDC) to uphold the\ncross-modal semantic similarity through incremental tasks and retain previously\nacquired knowledge of semantic similarity in old models, mitigating the risk of\ncatastrophic forgetting. The CrossSDC can seamlessly integrate into the\ntraining process of different audio-visual sound separation frameworks.\nExperiments demonstrate that ContAV-Sep can effectively mitigate catastrophic\nforgetting and achieve significantly better performance compared to other\ncontinual learning baselines for audio-visual sound separation. Code is\navailable at: \\url{https://github.com/weiguoPian/ContAV-Sep_NeurIPS2024}.\n", "date": "2024-11-05", "categories": []}
{"id": "2411.03008", "title": "Hierarchical Orchestra of Policies", "abstract": "  Continual reinforcement learning poses a major challenge due to the tendency\nof agents to experience catastrophic forgetting when learning sequential tasks.\nIn this paper, we introduce a modularity-based approach, called Hierarchical\nOrchestra of Policies (HOP), designed to mitigate catastrophic forgetting in\nlifelong reinforcement learning. HOP dynamically forms a hierarchy of policies\nbased on a similarity metric between the current observations and previously\nencountered observations in successful tasks. Unlike other state-of-the-art\nmethods, HOP does not require task labelling, allowing for robust adaptation in\nenvironments where boundaries between tasks are ambiguous. Our experiments,\nconducted across multiple tasks in a procedurally generated suite of\nenvironments, demonstrate that HOP significantly outperforms baseline methods\nin retaining knowledge across tasks and performs comparably to state-of-the-art\ntransfer methods that require task labelling. Moreover, HOP achieves this\nwithout compromising performance when tasks remain constant, highlighting its\nversatility.\n", "date": "2024-11-05", "categories": []}
{"id": "2403.04398", "title": "MAGR: Manifold-Aligned Graph Regularization for Continual Action Quality\n  Assessment", "abstract": "  Action Quality Assessment (AQA) evaluates diverse skills but models struggle\nwith non-stationary data. We propose Continual AQA (CAQA) to refine models\nusing sparse new data. Feature replay preserves memory without storing raw\ninputs. However, the misalignment between static old features and the\ndynamically changing feature manifold causes severe catastrophic forgetting. To\naddress this novel problem, we propose Manifold-Aligned Graph Regularization\n(MAGR), which first aligns deviated old features to the current feature\nmanifold, ensuring representation consistency. It then constructs a graph\njointly arranging old and new features aligned with quality scores. Experiments\nshow MAGR outperforms recent strong baselines with up to 6.56%, 5.66%, 15.64%,\nand 9.05% correlation gains on the MTL-AQA, FineDiving, UNLV-Dive, and JDM-MSA\nsplit datasets, respectively. This validates MAGR for continual assessment\nchallenges arising from non-stationary skill variations. Code is available at\nhttps://github.com/ZhouKanglei/MAGR_CAQA}{https://github.com/ZhouKanglei/MAGR_CAQA.\n", "date": "2024-03-07", "categories": []}
{"id": "2403.15706", "title": "GACL: Exemplar-Free Generalized Analytic Continual Learning", "abstract": "  Class incremental learning (CIL) trains a network on sequential tasks with\nseparated categories in each task but suffers from catastrophic forgetting,\nwhere models quickly lose previously learned knowledge when acquiring new\ntasks. The generalized CIL (GCIL) aims to address the CIL problem in a more\nreal-world scenario, where incoming data have mixed data categories and unknown\nsample size distribution. Existing attempts for the GCIL either have poor\nperformance or invade data privacy by saving exemplars. In this paper, we\npropose a new exemplar-free GCIL technique named generalized analytic continual\nlearning (GACL). The GACL adopts analytic learning (a gradient-free training\ntechnique) and delivers an analytical (i.e., closed-form) solution to the GCIL\nscenario. This solution is derived via decomposing the incoming data into\nexposed and unexposed classes, thereby attaining a weight-invariant property, a\nrare yet valuable property supporting an equivalence between incremental\nlearning and its joint training. Such an equivalence is crucial in GCIL\nsettings as data distributions among different tasks no longer pose challenges\nto adopting our GACL. Theoretically, this equivalence property is validated\nthrough matrix analysis tools. Empirically, we conduct extensive experiments\nwhere, compared with existing GCIL methods, our GACL exhibits a consistently\nleading performance across various datasets and GCIL settings. Source code is\navailable at https://github.com/CHEN-YIZHU/GACL.\n", "date": "2024-03-23", "categories": []}
{"id": "2403.15751", "title": "F-OAL: Forward-only Online Analytic Learning with Fast Training and Low\n  Memory Footprint in Class Incremental Learning", "abstract": "  Online Class Incremental Learning (OCIL) aims to train models incrementally,\nwhere data arrive in mini-batches, and previous data are not accessible. A\nmajor challenge in OCIL is Catastrophic Forgetting, i.e., the loss of\npreviously learned knowledge. Among existing baselines, replay-based methods\nshow competitive results but requires extra memory for storing exemplars, while\nexemplar-free (i.e., data need not be stored for replay in production) methods\nare resource-friendly but often lack accuracy. In this paper, we propose an\nexemplar-free approach--Forward-only Online Analytic Learning (F-OAL). Unlike\ntraditional methods, F-OAL does not rely on back-propagation and is\nforward-only, significantly reducing memory usage and computational time.\nCooperating with a pre-trained frozen encoder with Feature Fusion, F-OAL only\nneeds to update a linear classifier by recursive least square. This approach\nsimultaneously achieves high accuracy and low resource consumption. Extensive\nexperiments on benchmark datasets demonstrate F-OAL's robust performance in\nOCIL scenarios. Code is available at https://github.com/liuyuchen-cz/F-OAL.\n", "date": "2024-03-23", "categories": []}
{"id": "2407.18145", "title": "Taxonomy-Aware Continual Semantic Segmentation in Hyperbolic Spaces for\n  Open-World Perception", "abstract": "  Semantic segmentation models are typically trained on a fixed set of classes,\nlimiting their applicability in open-world scenarios. Class-incremental\nsemantic segmentation aims to update models with emerging new classes while\npreventing catastrophic forgetting of previously learned ones. However,\nexisting methods impose strict rigidity on old classes, reducing their\neffectiveness in learning new incremental classes. In this work, we propose\nTaxonomy-Oriented Poincar\\'e-regularized Incremental-Class Segmentation\n(TOPICS) that learns feature embeddings in hyperbolic space following explicit\ntaxonomy-tree structures. This supervision provides plasticity for old classes,\nupdating ancestors based on new classes while integrating new classes at\nfitting positions. Additionally, we maintain implicit class relational\nconstraints on the geometric basis of the Poincar\\'e ball. This ensures that\nthe latent space can continuously adapt to new constraints while maintaining a\nrobust structure to combat catastrophic forgetting. We also establish eight\nrealistic incremental learning protocols for autonomous driving scenarios,\nwhere novel classes can originate from known classes or the background.\nExtensive evaluations of TOPICS on the Cityscapes and Mapillary Vistas 2.0\nbenchmarks demonstrate that it achieves state-of-the-art performance. We make\nthe code and trained models publicly available at\nhttp://topics.cs.uni-freiburg.de.\n", "date": "2024-07-25", "categories": []}
{"id": "2411.01172", "title": "Covariance-based Space Regularization for Few-shot Class Incremental\n  Learning", "abstract": "  Few-shot Class Incremental Learning (FSCIL) presents a challenging yet\nrealistic scenario, which requires the model to continually learn new classes\nwith limited labeled data (i.e., incremental sessions) while retaining\nknowledge of previously learned base classes (i.e., base sessions). Due to the\nlimited data in incremental sessions, models are prone to overfitting new\nclasses and suffering catastrophic forgetting of base classes. To tackle these\nissues, recent advancements resort to prototype-based approaches to constrain\nthe base class distribution and learn discriminative representations of new\nclasses. Despite the progress, the limited data issue still induces ill-divided\nfeature space, leading the model to confuse the new class with old classes or\nfail to facilitate good separation among new classes. In this paper, we aim to\nmitigate these issues by directly constraining the span of each class\ndistribution from a covariance perspective. In detail, we propose a simple yet\neffective covariance constraint loss to force the model to learn each class\ndistribution with the same covariance matrix. In addition, we propose a\nperturbation approach to perturb the few-shot training samples in the feature\nspace, which encourages the samples to be away from the weighted distribution\nof other classes. Regarding perturbed samples as new class data, the classifier\nis forced to establish explicit boundaries between each new class and the\nexisting ones. Our approach is easy to integrate into existing FSCIL approaches\nto boost performance. Experiments on three benchmarks validate the\neffectiveness of our approach, achieving a new state-of-the-art performance of\nFSCIL.\n", "date": "2024-11-02", "categories": []}
{"id": "2411.01465", "title": "Efficient Non-Exemplar Class-Incremental Learning with Retrospective\n  Feature Synthesis", "abstract": "  Despite the outstanding performance in many individual tasks, deep neural\nnetworks suffer from catastrophic forgetting when learning from continuous data\nstreams in real-world scenarios. Current Non-Exemplar Class-Incremental\nLearning (NECIL) methods mitigate forgetting by storing a single prototype per\nclass, which serves to inject previous information when sequentially learning\nnew classes. However, these stored prototypes or their augmented variants often\nfail to simultaneously capture spatial distribution diversity and precision\nneeded for representing old classes. Moreover, as the model acquires new\nknowledge, these prototypes gradually become outdated, making them less\neffective. To overcome these limitations, we propose a more efficient NECIL\nmethod that replaces prototypes with synthesized retrospective features for old\nclasses. Specifically, we model each old class's feature space using a\nmultivariate Gaussian distribution and generate deep representations by\nsampling from high-likelihood regions. Additionally, we introduce a\nsimilarity-based feature compensation mechanism that integrates generated old\nclass features with similar new class features to synthesize robust\nretrospective representations. These retrospective features are then\nincorporated into our incremental learning framework to preserve the decision\nboundaries of previous classes while learning new ones. Extensive experiments\non CIFAR-100, TinyImageNet, and ImageNet-Subset demonstrate that our method\nsignificantly improves the efficiency of non-exemplar class-incremental\nlearning and achieves state-of-the-art performance.\n", "date": "2024-11-03", "categories": []}
{"id": "2411.02175", "title": "SAFE: Slow and Fast Parameter-Efficient Tuning for Continual Learning\n  with Pre-Trained Models", "abstract": "  Continual learning aims to incrementally acquire new concepts in data streams\nwhile resisting forgetting previous knowledge. With the rise of powerful\npre-trained models (PTMs), there is a growing interest in training incremental\nlearning systems using these foundation models, rather than learning from\nscratch. Existing works often view PTMs as a strong initial point and directly\napply parameter-efficient tuning (PET) in the first session for adapting to\ndownstream tasks. In the following sessions, most methods freeze model\nparameters for tackling forgetting issues. However, applying PET directly to\ndownstream data cannot fully explore the inherent knowledge in PTMs.\nAdditionally, freezing the parameters in incremental sessions hinders models'\nplasticity to novel concepts not covered in the first session. To solve the\nabove issues, we propose a Slow And Fast parameter-Efficient tuning (SAFE)\nframework. In particular, to inherit general knowledge from foundation models,\nwe include a transfer loss function by measuring the correlation between the\nPTM and the PET-applied model. After calibrating in the first session, the slow\nefficient tuning parameters can capture more informative features, improving\ngeneralization to incoming classes. Moreover, to further incorporate novel\nconcepts, we strike a balance between stability and plasticity by fixing slow\nefficient tuning parameters and continuously updating the fast ones.\nSpecifically, a cross-classification loss with feature alignment is proposed to\ncircumvent catastrophic forgetting. During inference, we introduce an\nentropy-based aggregation strategy to dynamically utilize the complementarity\nin the slow and fast learners. Extensive experiments on seven benchmark\ndatasets verify the effectiveness of our method by significantly surpassing the\nstate-of-the-art.\n", "date": "2024-11-04", "categories": []}
{"id": "2306.06190", "title": "$FastDoc$: Domain-Specific Fast Continual Pre-training Technique using\n  Document-Level Metadata and Taxonomy", "abstract": "  In this paper, we propose $FastDoc$ (Fast Continual Pre-training Technique\nusing Document Level Metadata and Taxonomy), a novel, compute-efficient\nframework that utilizes Document metadata and Domain-Specific Taxonomy as\nsupervision signals to continually pre-train transformer encoder on a\ndomain-specific corpus. The main innovation is that during domain-specific\npretraining, an open-domain encoder is continually pre-trained using\nsentence-level embeddings as inputs (to accommodate long documents), however,\nfine-tuning is done with token-level embeddings as inputs to this encoder. We\nperform such domain-specific pre-training on three different domains namely\ncustomer support, scientific, and legal domains, and compare performance on 6\ndifferent downstream tasks and 9 different datasets. The novel use of\ndocument-level supervision along with sentence-level embedding input for\npre-training reduces pre-training compute by around $1,000$, $4,500$, and $500$\ntimes compared to MLM and/or NSP in Customer Support, Scientific, and Legal\nDomains, respectively. The reduced training time does not lead to a\ndeterioration in performance. In fact we show that $FastDoc$ either outperforms\nor performs on par with several competitive transformer-based baselines in\nterms of character-level F1 scores and other automated metrics in the Customer\nSupport, Scientific, and Legal Domains. Moreover, reduced training aids in\nmitigating the risk of catastrophic forgetting. Thus, unlike baselines,\n$FastDoc$ shows a negligible drop in performance on open domain.\n", "date": "2023-06-09", "categories": []}
{"id": "2404.00986", "title": "Make Continual Learning Stronger via C-Flat", "abstract": "  Model generalization ability upon incrementally acquiring dynamically\nupdating knowledge from sequentially arriving tasks is crucial to tackle the\nsensitivity-stability dilemma in Continual Learning (CL). Weight loss landscape\nsharpness minimization seeking for flat minima lying in neighborhoods with\nuniform low loss or smooth gradient is proven to be a strong training regime\nimproving model generalization compared with loss minimization based optimizer\nlike SGD. Yet only a few works have discussed this training regime for CL,\nproving that dedicated designed zeroth-order sharpness optimizer can improve CL\nperformance. In this work, we propose a Continual Flatness (C-Flat) method\nfeaturing a flatter loss landscape tailored for CL. C-Flat could be easily\ncalled with only one line of code and is plug-and-play to any CL methods. A\ngeneral framework of C-Flat applied to all CL categories and a thorough\ncomparison with loss minima optimizer and flat minima based CL approaches is\npresented in this paper, showing that our method can boost CL performance in\nalmost all cases. Code is available at https://github.com/WanNaa/C-Flat.\n", "date": "2024-04-01", "categories": []}
{"id": "2410.23495", "title": "DASH: Warm-Starting Neural Network Training in Stationary Settings\n  without Loss of Plasticity", "abstract": "  Warm-starting neural network training by initializing networks with\npreviously learned weights is appealing, as practical neural networks are often\ndeployed under a continuous influx of new data. However, it often leads to loss\nof plasticity, where the network loses its ability to learn new information,\nresulting in worse generalization than training from scratch. This occurs even\nunder stationary data distributions, and its underlying mechanism is poorly\nunderstood. We develop a framework emulating real-world neural network training\nand identify noise memorization as the primary cause of plasticity loss when\nwarm-starting on stationary data. Motivated by this, we propose Direction-Aware\nSHrinking (DASH), a method aiming to mitigate plasticity loss by selectively\nforgetting memorized noise while preserving learned features. We validate our\napproach on vision tasks, demonstrating improvements in test accuracy and\ntraining efficiency.\n", "date": "2024-10-30", "categories": []}
{"id": "2311.04898", "title": "Two Complementary Perspectives to Continual Learning: Ask Not Only What\n  to Optimize, But Also How", "abstract": "  Recent years have seen considerable progress in the continual training of\ndeep neural networks, predominantly thanks to approaches that add replay or\nregularization terms to the loss function to approximate the joint loss over\nall tasks so far. However, we show that even with a perfect approximation to\nthe joint loss, these approaches still suffer from temporary but substantial\nforgetting when starting to train on a new task. Motivated by this 'stability\ngap', we propose that continual learning strategies should focus not only on\nthe optimization objective, but also on the way this objective is optimized.\nWhile there is some continual learning work that alters the optimization\ntrajectory (e.g., using gradient projection techniques), this line of research\nis positioned as alternative to improving the optimization objective, while we\nargue it should be complementary. In search of empirical support for our\nproposition, we perform a series of pre-registered experiments combining\nreplay-approximated joint objectives with gradient projection-based\noptimization routines. However, this first experimental attempt fails to show\nclear and consistent benefits. Nevertheless, our conceptual arguments, as well\nas some of our empirical results, demonstrate the distinctive importance of the\noptimization trajectory in continual learning, thereby opening up a new\ndirection for continual learning research.\n", "date": "2023-11-08", "categories": []}
{"id": "2403.19137", "title": "CLAP4CLIP: Continual Learning with Probabilistic Finetuning for\n  Vision-Language Models", "abstract": "  Continual learning (CL) aims to help deep neural networks learn new knowledge\nwhile retaining what has been learned. Owing to their powerful\ngeneralizability, pre-trained vision-language models such as Contrastive\nLanguage-Image Pre-training (CLIP) have lately gained traction as practical CL\ncandidates. However, the domain mismatch between the pre-training and the\ndownstream CL tasks often calls for finetuning of the CLIP on the latter. Most\nexisting finetuning methods exhibit deterministic nature. This makes them\noverlook the many possible interactions across the input modalities and deems\nthem unsafe for high-risk tasks requiring reliable uncertainty estimation. To\naddress these, our work proposes Continual LeArning with Probabilistic\nfinetuning (CLAP) - a probabilistic modeling framework over visual-guided text\nfeatures per task, thus providing more calibrated CL finetuning. Unlike recent\ndata-hungry anti-forgetting CL techniques, CLAP alleviates forgetting by\nexploiting the rich pre-trained knowledge of CLIP for weight initialization and\ndistribution regularization of task-specific parameters. Cooperating with the\ndiverse range of existing prompting methods, CLAP can surpass the predominant\ndeterministic finetuning approaches for CL with CLIP. We conclude with\nout-of-the-box applications of superior uncertainty estimation abilities of\nCLAP including novel data detection and exemplar selection within the existing\nCL setups. Our code is available at\n\\url{https://github.com/srvCodes/clap4clip}.\n", "date": "2024-03-28", "categories": []}
{"id": "2410.23368", "title": "NCAdapt: Dynamic adaptation with domain-specific Neural Cellular\n  Automata for continual hippocampus segmentation", "abstract": "  Continual learning (CL) in medical imaging presents a unique challenge,\nrequiring models to adapt to new domains while retaining previously acquired\nknowledge. We introduce NCAdapt, a Neural Cellular Automata (NCA) based method\ndesigned to address this challenge. NCAdapt features a domain-specific\nmulti-head structure, integrating adaptable convolutional layers into the NCA\nbackbone for each new domain encountered. After initial training, the NCA\nbackbone is frozen, and only the newly added adaptable convolutional layers,\nconsisting of 384 parameters, are trained along with domain-specific NCA\nconvolutions. We evaluate NCAdapt on hippocampus segmentation tasks,\nbenchmarking its performance against Lifelong nnU-Net and U-Net models with\nstate-of-the-art (SOTA) CL methods. Our lightweight approach achieves SOTA\nperformance, underscoring its effectiveness in addressing CL challenges in\nmedical imaging. Upon acceptance, we will make our code base publicly\naccessible to support reproducibility and foster further advancements in\nmedical CL.\n", "date": "2024-10-30", "categories": []}
{"id": "2211.14568", "title": "BeGin: Extensive Benchmark Scenarios and An Easy-to-use Framework for\n  Graph Continual Learning", "abstract": "  Continual Learning (CL) is the process of learning ceaselessly a sequence of\ntasks. Most existing CL methods deal with independent data (e.g., images and\ntext) for which many benchmark frameworks and results under standard\nexperimental settings are available. Compared to them, however, CL methods for\ngraph data (graph CL) are relatively underexplored because of (a) the lack of\nstandard experimental settings, especially regarding how to deal with the\ndependency between instances, (b) the lack of benchmark datasets and scenarios,\nand (c) high complexity in implementation and evaluation due to the dependency.\nIn this paper, regarding (a) we define four standard incremental settings\n(task-, class-, domain-, and time-incremental) for node-, link-, and\ngraph-level problems, extending the previously explored scope. Regarding (b),\nwe provide 35 benchmark scenarios based on 24 real-world graphs. Regarding (c),\nwe develop BeGin, an easy and fool-proof framework for graph CL. BeGin is\neasily extended since it is modularized with reusable modules for data\nprocessing, algorithm design, and evaluation. Especially, the evaluation module\nis completely separated from user code to eliminate potential mistakes.\nRegarding benchmark results, we cover 3x more combinations of incremental\nsettings and levels of problems than the latest benchmark. All assets for the\nbenchmark framework are publicly available at\nhttps://github.com/ShinhwanKang/BeGin.\n", "date": "2022-11-26", "categories": []}
{"id": "2405.16642", "title": "Fast TRAC: A Parameter-Free Optimizer for Lifelong Reinforcement\n  Learning", "abstract": "  A key challenge in lifelong reinforcement learning (RL) is the loss of\nplasticity, where previous learning progress hinders an agent's adaptation to\nnew tasks. While regularization and resetting can help, they require precise\nhyperparameter selection at the outset and environment-dependent adjustments.\nBuilding on the principled theory of online convex optimization, we present a\nparameter-free optimizer for lifelong RL, called TRAC, which requires no tuning\nor prior knowledge about the distribution shifts. Extensive experiments on\nProcgen, Atari, and Gym Control environments show that TRAC works surprisingly\nwell-mitigating loss of plasticity and rapidly adapting to challenging\ndistribution shifts-despite the underlying optimization problem being nonconvex\nand nonstationary.\n", "date": "2024-05-26", "categories": []}
{"id": "2408.09053", "title": "Learning to Route for Dynamic Adapter Composition in Continual Learning\n  with Language Models", "abstract": "  Parameter-efficient fine-tuning (PEFT) methods are increasingly used with\npre-trained language models (PLMs) for continual learning (CL). These methods\ntypically involve training a PEFT module for each new task and employing\nsimilarity-based selection to route modules during inference. However, they\nface two major limitations: 1) interference during module training with already\nlearned modules and 2) suboptimal routing when composing modules. In this\npaper, we present L2R, a method that isolates the training of new PEFT modules\nto ensure their task specialization. L2R then learns to compose the learned\nmodules by training a network of routers that leverages a small memory\ncontaining examples of previously seen tasks. We evaluate our method in two CL\nsetups using various benchmarks. Our results demonstrate that L2R provides an\neffective composition of PEFT modules, leading to improved generalization and\nperformance compared to other methods.\n", "date": "2024-08-16", "categories": []}
{"id": "2410.22695", "title": "Permutation Invariant Learning with High-Dimensional Particle Filters", "abstract": "  Sequential learning in deep models often suffers from challenges such as\ncatastrophic forgetting and loss of plasticity, largely due to the permutation\ndependence of gradient-based algorithms, where the order of training data\nimpacts the learning outcome. In this work, we introduce a novel\npermutation-invariant learning framework based on high-dimensional particle\nfilters. We theoretically demonstrate that particle filters are invariant to\nthe sequential ordering of training minibatches or tasks, offering a principled\nsolution to mitigate catastrophic forgetting and loss-of-plasticity. We develop\nan efficient particle filter for optimizing high-dimensional models, combining\nthe strengths of Bayesian methods with gradient-based optimization. Through\nextensive experiments on continual supervised and reinforcement learning\nbenchmarks, including SplitMNIST, SplitCIFAR100, and ProcGen, we empirically\nshow that our method consistently improves performance, while reducing variance\ncompared to standard baselines.\n", "date": "2024-10-30", "categories": []}
{"id": "2410.14753", "title": "Collaboratively adding new knowledge to an LLM", "abstract": "  We address the question of how to successively add new knowledge to an LLM\nwhilst retaining previously-added knowledge. We consider two settings,\nsemi-cooperative and fully-cooperative. Overall, LoRA performs better in most\ncases than full-fine tuning of all parameters when both new knowledge\nacquisition and retention of old, including recent, knowledge are taken into\naccount. In the semi-cooperative setting, where datasets are not available\nafter training, MOE mixing, model merging, and LoRA-based orthogonal subspace\nsequential learning, using a small weight on the orthogonality term, perform\nwell. In the fully-cooperative setting where datasets remain available, joint\ntraining and sequential training with replay are both effective approaches with\nLoRA training generally preferable to full fine-tuning. The codes needed to\nreproduce the results are provided in an open source repository.\n", "date": "2024-10-18", "categories": []}
{"id": "2406.05658", "title": "Visual Prompt Tuning in Null Space for Continual Learning", "abstract": "  Existing prompt-tuning methods have demonstrated impressive performances in\ncontinual learning (CL), by selecting and updating relevant prompts in the\nvision-transformer models. On the contrary, this paper aims to learn each task\nby tuning the prompts in the direction orthogonal to the subspace spanned by\nprevious tasks' features, so as to ensure no interference on tasks that have\nbeen learned to overcome catastrophic forgetting in CL. However, different from\nthe orthogonal projection in the traditional CNN architecture, the prompt\ngradient orthogonal projection in the ViT architecture shows completely\ndifferent and greater challenges, i.e., 1) the high-order and non-linear\nself-attention operation; 2) the drift of prompt distribution brought by the\nLayerNorm in the transformer block. Theoretically, we have finally deduced two\nconsistency conditions to achieve the prompt gradient orthogonal projection,\nwhich provide a theoretical guarantee of eliminating interference on previously\nlearned knowledge via the self-attention mechanism in visual prompt tuning. In\npractice, an effective null-space-based approximation solution has been\nproposed to implement the prompt gradient orthogonal projection. Extensive\nexperimental results demonstrate the effectiveness of anti-forgetting on four\nclass-incremental benchmarks with diverse pre-trained baseline models, and our\napproach achieves superior performances to state-of-the-art methods. Our code\nis available at https://github.com/zugexiaodui/VPTinNSforCL.\n", "date": "2024-06-09", "categories": []}
{"id": "2406.06811", "title": "Learning Continually by Spectral Regularization", "abstract": "  Loss of plasticity is a phenomenon where neural networks can become more\ndifficult to train over the course of learning. Continual learning algorithms\nseek to mitigate this effect by sustaining good performance while maintaining\nnetwork trainability. We develop a new technique for improving continual\nlearning inspired by the observation that the singular values of the neural\nnetwork parameters at initialization are an important factor for trainability\nduring early phases of learning. From this perspective, we derive a new\nspectral regularizer for continual learning that better sustains these\nbeneficial initialization properties throughout training. In particular, the\nregularizer keeps the maximum singular value of each layer close to one.\nSpectral regularization directly ensures that gradient diversity is maintained\nthroughout training, which promotes continual trainability, while minimally\ninterfering with performance in a single task. We present an experimental\nanalysis that shows how the proposed spectral regularizer can sustain\ntrainability and performance across a range of model architectures in continual\nsupervised and reinforcement learning settings. Spectral regularization is less\nsensitive to hyperparameters while demonstrating better training in individual\ntasks, sustaining trainability as new tasks arrive, and achieving better\ngeneralization performance.\n", "date": "2024-06-10", "categories": []}
{"id": "2407.15793", "title": "CLIP with Generative Latent Replay: a Strong Baseline for Incremental\n  Learning", "abstract": "  With the emergence of Transformers and Vision-Language Models (VLMs) such as\nCLIP, fine-tuning large pre-trained models has recently become a prevalent\nstrategy in Continual Learning. This has led to the development of numerous\nprompting strategies to adapt transformer-based models without incurring\ncatastrophic forgetting. However, these strategies often compromise the\noriginal zero-shot capabilities of the pre-trained CLIP model and struggle to\nadapt to domains that significantly deviate from the pre-training data. In this\nwork, we propose Continual Generative training for Incremental prompt-Learning,\na simple and novel approach to mitigate forgetting while adapting CLIP.\nBriefly, we employ Variational Autoencoders (VAEs) to learn class-conditioned\ndistributions within the embedding space of the visual encoder. We then exploit\nthese distributions to sample new synthetic visual embeddings and train the\ncorresponding class-specific textual prompts during subsequent tasks. Through\nextensive experiments on different domains, we show that such a generative\nreplay approach can adapt to new tasks while improving zero-shot capabilities,\nevaluated using a novel metric tailored for CL scenarios. Notably, further\nanalysis reveals that our approach can bridge the gap with joint prompt tuning.\nThe codebase is available at https://github.com/aimagelab/mammoth.\n", "date": "2024-07-22", "categories": []}
{"id": "2409.18265", "title": "Task-recency bias strikes back: Adapting covariances in Exemplar-Free\n  Class Incremental Learning", "abstract": "  Exemplar-Free Class Incremental Learning (EFCIL) tackles the problem of\ntraining a model on a sequence of tasks without access to past data. Existing\nstate-of-the-art methods represent classes as Gaussian distributions in the\nfeature extractor's latent space, enabling Bayes classification or training the\nclassifier by replaying pseudo features. However, we identify two critical\nissues that compromise their efficacy when the feature extractor is updated on\nincremental tasks. First, they do not consider that classes' covariance\nmatrices change and must be adapted after each task. Second, they are\nsusceptible to a task-recency bias caused by dimensionality collapse occurring\nduring training. In this work, we propose AdaGauss -- a novel method that\nadapts covariance matrices from task to task and mitigates the task-recency\nbias owing to the additional anti-collapse loss function. AdaGauss yields\nstate-of-the-art results on popular EFCIL benchmarks and datasets when training\nfrom scratch or starting from a pre-trained backbone. The code is available at:\nhttps://github.com/grypesc/AdaGauss.\n", "date": "2024-09-26", "categories": []}
{"id": "2410.10341", "title": "Replay-and-Forget-Free Graph Class-Incremental Learning: A Task\n  Profiling and Prompting Approach", "abstract": "  Class-incremental learning (CIL) aims to continually learn a sequence of\ntasks, with each task consisting of a set of unique classes. Graph CIL (GCIL)\nfollows the same setting but needs to deal with graph tasks (e.g., node\nclassification in a graph). The key characteristic of CIL lies in the absence\nof task identifiers (IDs) during inference, which causes a significant\nchallenge in separating classes from different tasks (i.e., inter-task class\nseparation). Being able to accurately predict the task IDs can help address\nthis issue, but it is a challenging problem. In this paper, we show\ntheoretically that accurate task ID prediction on graph data can be achieved by\na Laplacian smoothing-based graph task profiling approach, in which each graph\ntask is modeled by a task prototype based on Laplacian smoothing over the\ngraph. It guarantees that the task prototypes of the same graph task are nearly\nthe same with a large smoothing step, while those of different tasks are\ndistinct due to differences in graph structure and node attributes. Further, to\navoid the catastrophic forgetting of the knowledge learned in previous graph\ntasks, we propose a novel graph prompting approach for GCIL which learns a\nsmall discriminative graph prompt for each task, essentially resulting in a\nseparate classification model for each task. The prompt learning requires the\ntraining of a single graph neural network (GNN) only once on the first task,\nand no data replay is required thereafter, thereby obtaining a GCIL model being\nboth replay-free and forget-free. Extensive experiments on four GCIL benchmarks\nshow that i) our task prototype-based method can achieve 100% task ID\nprediction accuracy on all four datasets, ii) our GCIL model significantly\noutperforms state-of-the-art competing methods by at least 18% in average CIL\naccuracy, and iii) our model is fully free of forgetting on the four datasets.\n", "date": "2024-10-14", "categories": []}
{"id": "2410.20634", "title": "Plastic Learning with Deep Fourier Features", "abstract": "  Deep neural networks can struggle to learn continually in the face of\nnon-stationarity. This phenomenon is known as loss of plasticity. In this\npaper, we identify underlying principles that lead to plastic algorithms. In\nparticular, we provide theoretical results showing that linear function\napproximation, as well as a special case of deep linear networks, do not suffer\nfrom loss of plasticity. We then propose deep Fourier features, which are the\nconcatenation of a sine and cosine in every layer, and we show that this\ncombination provides a dynamic balance between the trainability obtained\nthrough linearity and the effectiveness obtained through the nonlinearity of\nneural networks. Deep networks composed entirely of deep Fourier features are\nhighly trainable and sustain their trainability over the course of learning.\nOur empirical results show that continual learning performance can be\ndrastically improved by replacing ReLU activations with deep Fourier features.\nThese results hold for different continual learning scenarios (e.g., label\nnoise, class incremental learning, pixel permutations) on all major supervised\nlearning datasets used for continual learning research, such as CIFAR10,\nCIFAR100, and tiny-ImageNet.\n", "date": "2024-10-27", "categories": []}
{"id": "2410.20768", "title": "Task Confusion and Catastrophic Forgetting in Class-Incremental\n  Learning: A Mathematical Framework for Discriminative and Generative\n  Modelings", "abstract": "  In class-incremental learning (class-IL), models must classify all previously\nseen classes at test time without task-IDs, leading to task confusion. Despite\nbeing a key challenge, task confusion lacks a theoretical understanding. We\npresent a novel mathematical framework for class-IL and prove the Infeasibility\nTheorem, showing optimal class-IL is impossible with discriminative modeling\ndue to task confusion. However, we establish the Feasibility Theorem,\ndemonstrating that generative modeling can achieve optimal class-IL by\novercoming task confusion. We then assess popular class-IL strategies,\nincluding regularization, bias-correction, replay, and generative classifier,\nusing our framework. Our analysis suggests that adopting generative modeling,\neither for generative replay or direct classification (generative classifier),\nis essential for optimal class-IL.\n", "date": "2024-10-28", "categories": []}
{"id": "2305.05738", "title": "DOCTOR: A Multi-Disease Detection Continual Learning Framework Based on\n  Wearable Medical Sensors", "abstract": "  Modern advances in machine learning (ML) and wearable medical sensors (WMSs)\nin edge devices have enabled ML-driven disease detection for smart healthcare.\nConventional ML-driven methods for disease detection rely on customizing\nindividual models for each disease and its corresponding WMS data. However,\nsuch methods lack adaptability to distribution shifts and new task\nclassification classes. In addition, they need to be rearchitected and\nretrained from scratch for each new disease. Moreover, installing multiple ML\nmodels in an edge device consumes excessive memory, drains the battery faster,\nand complicates the detection process. To address these challenges, we propose\nDOCTOR, a multi-disease detection continual learning (CL) framework based on\nWMSs. It employs a multi-headed deep neural network (DNN) and a replay-style CL\nalgorithm. The CL algorithm enables the framework to continually learn new\nmissions where different data distributions, classification classes, and\ndisease detection tasks are introduced sequentially. It counteracts\ncatastrophic forgetting with a data preservation method and a synthetic data\ngeneration (SDG) module. The data preservation method preserves the most\ninformative subset of real training data from previous missions for exemplar\nreplay. The SDG module models the probability distribution of the real training\ndata and generates synthetic data for generative replay while retaining data\nprivacy. The multi-headed DNN enables DOCTOR to detect multiple diseases\nsimultaneously based on user WMS data. We demonstrate DOCTOR's efficacy in\nmaintaining high disease classification accuracy with a single DNN model in\nvarious CL experiments. In complex scenarios, DOCTOR achieves 1.43 times better\naverage test accuracy, 1.25 times better F1-score, and 0.41 higher backward\ntransfer than the naive fine-tuning framework with a small model size of less\nthan 350KB.\n", "date": "2023-05-09", "categories": []}
{"id": "2308.11958", "title": "Maintaining Plasticity in Continual Learning via Regenerative\n  Regularization", "abstract": "  In continual learning, plasticity refers to the ability of an agent to\nquickly adapt to new information. Neural networks are known to lose plasticity\nwhen processing non-stationary data streams. In this paper, we propose L2 Init,\na simple approach for maintaining plasticity by incorporating in the loss\nfunction L2 regularization toward initial parameters. This is very similar to\nstandard L2 regularization (L2), the only difference being that L2 regularizes\ntoward the origin. L2 Init is simple to implement and requires selecting only a\nsingle hyper-parameter. The motivation for this method is the same as that of\nmethods that reset neurons or parameter values. Intuitively, when recent losses\nare insensitive to particular parameters, these parameters should drift toward\ntheir initial values. This prepares parameters to adapt quickly to new tasks.\nOn problems representative of different types of nonstationarity in continual\nsupervised learning, we demonstrate that L2 Init most consistently mitigates\nplasticity loss compared to previously proposed approaches.\n", "date": "2023-08-23", "categories": []}
{"id": "2309.14763", "title": "ConPET: Continual Parameter-Efficient Tuning for Large Language Models", "abstract": "  Continual learning necessitates the continual adaptation of models to newly\nemerging tasks while minimizing the catastrophic forgetting of old ones. This\nis extremely challenging for large language models (LLMs) with vanilla\nfull-parameter tuning due to high computation costs, memory consumption, and\nforgetting issue. Inspired by the success of parameter-efficient tuning (PET),\nwe propose Continual Parameter-Efficient Tuning (ConPET), a generalizable\nparadigm for continual task adaptation of LLMs with task-number-independent\ntraining complexity. ConPET includes two versions with different application\nscenarios. First, Static ConPET can adapt former continual learning methods\noriginally designed for relatively smaller models to LLMs through PET and a\ndynamic replay strategy, which largely reduces the tuning costs and alleviates\nthe over-fitting and forgetting issue. Furthermore, to maintain scalability,\nDynamic ConPET adopts separate PET modules for different tasks and a PET module\nselector for dynamic optimal selection. In our extensive experiments, the\nadaptation of Static ConPET helps multiple former methods reduce the scale of\ntunable parameters by over 3,000 times and surpass the PET-only baseline by at\nleast 5 points on five smaller benchmarks, while Dynamic ConPET gains its\nadvantage on the largest dataset. The codes and datasets are available at\nhttps://github.com/Raincleared-Song/ConPET.\n", "date": "2023-09-26", "categories": []}
{"id": "2410.19239", "title": "Prompting Continual Person Search", "abstract": "  The development of person search techniques has been greatly promoted in\nrecent years for its superior practicality and challenging goals. Despite their\nsignificant progress, existing person search models still lack the ability to\ncontinually learn from increaseing real-world data and adaptively process input\nfrom different domains. To this end, this work introduces the continual person\nsearch task that sequentially learns on multiple domains and then performs\nperson search on all seen domains. This requires balancing the stability and\nplasticity of the model to continually learn new knowledge without catastrophic\nforgetting. For this, we propose a Prompt-based Continual Person Search (PoPS)\nmodel in this paper. First, we design a compositional person search transformer\nto construct an effective pre-trained transformer without exhaustive\npre-training from scratch on large-scale person search data. This serves as the\nfundamental for prompt-based continual learning. On top of that, we design a\ndomain incremental prompt pool with a diverse attribute matching module. For\neach domain, we independently learn a set of prompts to encode the\ndomain-oriented knowledge. Meanwhile, we jointly learn a group of diverse\nattribute projections and prototype embeddings to capture discriminative domain\nattributes. By matching an input image with the learned attributes across\ndomains, the learned prompts can be properly selected for model inference.\nExtensive experiments are conducted to validate the proposed method for\ncontinual person search. The source code is available at\nhttps://github.com/PatrickZad/PoPS.\n", "date": "2024-10-25", "categories": []}
{"id": "2410.18378", "title": "Delta: A Cloud-assisted Data Enrichment Framework for On-Device\n  Continual Learning", "abstract": "  In modern mobile applications, users frequently encounter various new\ncontexts, necessitating on-device continual learning (CL) to ensure consistent\nmodel performance. While existing research predominantly focused on developing\nlightweight CL frameworks, we identify that data scarcity is a critical\nbottleneck for on-device CL. In this work, we explore the potential of\nleveraging abundant cloud-side data to enrich scarce on-device data, and\npropose a private, efficient and effective data enrichment framework Delta.\nSpecifically, Delta first introduces a directory dataset to decompose the data\nenrichment problem into device-side and cloud-side sub-problems without sharing\nsensitive data. Next, Delta proposes a soft data matching strategy to\neffectively solve the device-side sub-problem with sparse user data, and an\noptimal data sampling scheme for cloud server to retrieve the most suitable\ndataset for enrichment with low computational complexity. Further, Delta\nrefines the data sampling scheme by jointly considering the impact of enriched\ndata on both new and past contexts, mitigating the catastrophic forgetting\nissue from a new aspect. Comprehensive experiments across four typical mobile\ncomputing tasks with varied data modalities demonstrate that Delta could\nenhance the overall model accuracy by an average of 15.1%, 12.4%, 1.1% and 5.6%\nfor visual, IMU, audio and textual tasks compared with few-shot CL, and\nconsistently reduce the communication costs by over 90% compared to federated\nCL.\n", "date": "2024-10-24", "categories": []}
{"id": "2403.08350", "title": "CoIN: A Benchmark of Continual Instruction tuNing for Multimodel Large\n  Language Model", "abstract": "  Instruction tuning represents a prevalent strategy employed by Multimodal\nLarge Language Models (MLLMs) to align with human instructions and adapt to new\ntasks. Nevertheless, MLLMs encounter the challenge of adapting to users'\nevolving knowledge and demands. Therefore, how to retain existing skills while\nacquiring new knowledge needs to be investigated. In this paper, we present a\ncomprehensive benchmark, namely Continual Instruction tuNing (CoIN), to assess\nexisting MLLMs in the sequential instruction tuning paradigm. CoIN comprises 10\ncommonly used datasets spanning 8 task categories, ensuring a diverse range of\ninstructions and tasks. Besides, the trained model is evaluated from two\naspects: Instruction Following and General Knowledge, which assess the\nalignment with human intention and knowledge preserved for reasoning,\nrespectively. Experiments on CoIN demonstrate that current powerful MLLMs still\nsuffer catastrophic forgetting, and the failure in intention alignment assumes\nthe main responsibility, instead of the knowledge forgetting. To this end, we\nintroduce MoELoRA to MLLMs which is effective to retain the previous\ninstruction alignment. Experimental results consistently illustrate the\nforgetting decreased from this method on CoIN.\n", "date": "2024-03-13", "categories": []}
{"id": "2410.17594", "title": "How to Continually Adapt Text-to-Image Diffusion Models for Flexible\n  Customization?", "abstract": "  Custom diffusion models (CDMs) have attracted widespread attention due to\ntheir astonishing generative ability for personalized concepts. However, most\nexisting CDMs unreasonably assume that personalized concepts are fixed and\ncannot change over time. Moreover, they heavily suffer from catastrophic\nforgetting and concept neglect on old personalized concepts when continually\nlearning a series of new concepts. To address these challenges, we propose a\nnovel Concept-Incremental text-to-image Diffusion Model (CIDM), which can\nresolve catastrophic forgetting and concept neglect to learn new customization\ntasks in a concept-incremental manner. Specifically, to surmount the\ncatastrophic forgetting of old concepts, we develop a concept consolidation\nloss and an elastic weight aggregation module. They can explore task-specific\nand task-shared knowledge during training, and aggregate all low-rank weights\nof old concepts based on their contributions during inference. Moreover, in\norder to address concept neglect, we devise a context-controllable synthesis\nstrategy that leverages expressive region features and noise estimation to\ncontrol the contexts of generated images according to user conditions.\nExperiments validate that our CIDM surpasses existing custom diffusion models.\nThe source codes are available at https://github.com/JiahuaDong/CIFC.\n", "date": "2024-10-23", "categories": []}
{"id": "2410.17715", "title": "Continual Learning on a Data Diet", "abstract": "  Continual Learning (CL) methods usually learn from all available data.\nHowever, this is not the case in human cognition which efficiently focuses on\nkey experiences while disregarding the redundant information. Similarly, not\nall data points in a dataset have equal potential; some can be more informative\nthan others. This disparity may significantly impact the performance, as both\nthe quality and quantity of samples directly influence the model's\ngeneralizability and efficiency. Drawing inspiration from this, we explore the\npotential of learning from important samples and present an empirical study for\nevaluating coreset selection techniques in the context of CL to stimulate\nresearch in this unexplored area. We train different continual learners on\nincreasing amounts of selected samples and investigate the learning-forgetting\ndynamics by shedding light on the underlying mechanisms driving their improved\nstability-plasticity balance. We present several significant observations:\nlearning from selectively chosen samples (i) enhances incremental accuracy,\n(ii) improves knowledge retention of previous tasks, and (iii) refines learned\nrepresentations. This analysis contributes to a deeper understanding of\nselective learning strategies in CL scenarios.\n", "date": "2024-10-23", "categories": []}
{"id": "2404.04846", "title": "F-MALLOC: Feed-forward Memory Allocation for Continual Learning in\n  Neural Machine Translation", "abstract": "  In the evolving landscape of Neural Machine Translation (NMT), the\npretrain-then-finetune paradigm has yielded impressive results. However, the\npersistent challenge of Catastrophic Forgetting (CF) remains a hurdle. While\nprevious work has introduced Continual Learning (CL) methods to address CF,\nthese approaches grapple with the delicate balance between avoiding forgetting\nand maintaining system extensibility. To address this, we propose a CL method,\nnamed $\\textbf{F-MALLOC}$ ($\\textbf{F}$eed-forward $\\textbf{M}$emory\n$\\textbf{ALLOC}ation)$. F-MALLOC is inspired by recent insights highlighting\nthat feed-forward layers emulate neural memories and encapsulate crucial\ntranslation knowledge. It decomposes feed-forward layers into discrete memory\ncells and allocates these memories to different tasks. By learning to allocate\nand safeguard these memories, our method effectively alleviates CF while\nensuring robust extendability. Besides, we propose a comprehensive assessment\nprotocol for multi-stage CL of NMT systems. Experiments conducted following\nthis new protocol showcase the superior performance of F-MALLOC, evidenced by\nhigher BLEU scores and almost zero forgetting.\n", "date": "2024-04-07", "categories": []}
{"id": "2410.04183", "title": "Unsupervised Assessment of Landscape Shifts Based on Persistent Entropy\n  and Topological Preservation", "abstract": "  In Continual Learning (CL) contexts, concept drift typically refers to the\nanalysis of changes in data distribution. A drift in the input data can have\nnegative consequences on a learning predictor and the system's stability. The\nmajority of concept drift methods emphasize the analysis of statistical changes\nin non-stationary data over time. In this context, we consider another\nperspective, where the concept drift also integrates substantial changes in the\ntopological characteristics of the data stream. In this article, we introduce a\nnovel framework for monitoring changes in multi-dimensional data streams. We\nexplore variations in the topological structures of the data, presenting\nanother angle on the standard concept drift. Our developed approach is based on\npersistent entropy and topology-preserving projections in a continual learning\nscenario. The framework operates in both unsupervised and supervised\nenvironments. To show the utility of the proposed framework, we analyze the\nmodel across three scenarios using data streams generated with MNIST samples.\nThe obtained results reveal the potential of applying topological data analysis\nfor shift detection and encourage further research in this area.\n", "date": "2024-10-05", "categories": []}
{"id": "2304.10038", "title": "Open-World Continual Learning: Unifying Novelty Detection and Continual\n  Learning", "abstract": "  As AI agents are increasingly used in the real open world with unknowns or\nnovelties, they need the ability to (1) recognize objects that (a) they have\nlearned before and (b) detect items that they have never seen or learned, and\n(2) learn the new items incrementally to become more and more knowledgeable and\npowerful. (1) is called novelty detection or out-of-distribution (OOD)\ndetection and (2) is called class incremental learning (CIL), which is a\nsetting of continual learning (CL). In existing research, OOD detection and CIL\nare regarded as two completely different problems. This paper first provides a\ntheoretical proof that good OOD detection for each task within the set of\nlearned tasks (called closed-world OOD detection) is necessary for successful\nCIL. We show this by decomposing CIL into two sub-problems: within-task\nprediction (WP) and task-id prediction (TP), and proving that TP is correlated\nwith closed-world OOD detection. The key theoretical result is that regardless\nof whether WP and OOD detection (or TP) are defined explicitly or implicitly by\na CIL algorithm, good WP and good closed-world OOD detection are necessary and\nsufficient conditions for good CIL, which unifies novelty or OOD detection and\ncontinual learning (CIL, in particular). We call this traditional CIL the\nclosed-world CIL as it does not detect future OOD data in the open world. The\npaper then proves that the theory can be generalized or extended to open-world\nCIL, which is the proposed open-world continual learning, that can perform CIL\nin the open world and detect future or open-world OOD data. Based on the\ntheoretical results, new CIL methods are also designed, which outperform strong\nbaselines in CIL accuracy and in continual OOD detection by a large margin.\n", "date": "2023-04-20", "categories": []}
{"id": "2310.07996", "title": "Reset It and Forget It: Relearning Last-Layer Weights Improves Continual\n  and Transfer Learning", "abstract": "  This work identifies a simple pre-training mechanism that leads to\nrepresentations exhibiting better continual and transfer learning. This\nmechanism -- the repeated resetting of weights in the last layer, which we\nnickname \"zapping\" -- was originally designed for a meta-continual-learning\nprocedure, yet we show it is surprisingly applicable in many settings beyond\nboth meta-learning and continual learning. In our experiments, we wish to\ntransfer a pre-trained image classifier to a new set of classes, in a few\nshots. We show that our zapping procedure results in improved transfer accuracy\nand/or more rapid adaptation in both standard fine-tuning and continual\nlearning settings, while being simple to implement and computationally\nefficient. In many cases, we achieve performance on par with state of the art\nmeta-learning without needing the expensive higher-order gradients, by using a\ncombination of zapping and sequential learning. An intuitive explanation for\nthe effectiveness of this zapping procedure is that representations trained\nwith repeated zapping learn features that are capable of rapidly adapting to\nnewly initialized classifiers. Such an approach may be considered a\ncomputationally cheaper type of, or alternative to, meta-learning rapidly\nadaptable features with higher-order gradients. This adds to recent work on the\nusefulness of resetting neural network parameters during training, and invites\nfurther investigation of this mechanism.\n", "date": "2023-10-12", "categories": []}
{"id": "2410.15318", "title": "SNAP: Stopping Catastrophic Forgetting in Hebbian Learning with\n  Sigmoidal Neuronal Adaptive Plasticity", "abstract": "  Artificial Neural Networks (ANNs) suffer from catastrophic forgetting, where\nthe learning of new tasks causes the catastrophic forgetting of old tasks.\nExisting Machine Learning (ML) algorithms, including those using Stochastic\nGradient Descent (SGD) and Hebbian Learning typically update their weights\nlinearly with experience i.e., independently of their current strength. This\ncontrasts with biological neurons, which at intermediate strengths are very\nplastic, but consolidate with Long-Term Potentiation (LTP) once they reach a\ncertain strength. We hypothesize this mechanism might help mitigate\ncatastrophic forgetting. We introduce Sigmoidal Neuronal Adaptive Plasticity\n(SNAP) an artificial approximation to Long-Term Potentiation for ANNs by having\nthe weights follow a sigmoidal growth behaviour allowing the weights to\nconsolidate and stabilize when they reach sufficiently large or small values.\nWe then compare SNAP to linear weight growth and exponential weight growth and\nsee that SNAP completely prevents the forgetting of previous tasks for Hebbian\nLearning but not for SGD-base learning.\n", "date": "2024-10-20", "categories": []}
{"id": "2410.15372", "title": "Hybrid Memory Replay: Blending Real and Distilled Data for Class\n  Incremental Learning", "abstract": "  Incremental learning (IL) aims to acquire new knowledge from current tasks\nwhile retaining knowledge learned from previous tasks. Replay-based IL methods\nstore a set of exemplars from previous tasks in a buffer and replay them when\nlearning new tasks. However, there is usually a size-limited buffer that cannot\nstore adequate real exemplars to retain the knowledge of previous tasks. In\ncontrast, data distillation (DD) can reduce the exemplar buffer's size, by\ncondensing a large real dataset into a much smaller set of more\ninformation-compact synthetic exemplars. Nevertheless, DD's performance gain on\nIL quickly vanishes as the number of synthetic exemplars grows. To overcome the\nweaknesses of real-data and synthetic-data buffers, we instead optimize a\nhybrid memory including both types of data. Specifically, we propose an\ninnovative modification to DD that distills synthetic data from a sliding\nwindow of checkpoints in history (rather than checkpoints on multiple training\ntrajectories). Conditioned on the synthetic data, we then optimize the\nselection of real exemplars to provide complementary improvement to the DD\nobjective. The optimized hybrid memory combines the strengths of synthetic and\nreal exemplars, effectively mitigating catastrophic forgetting in Class IL\n(CIL) when the buffer size for exemplars is limited. Notably, our method can be\nseamlessly integrated into most existing replay-based CIL models. Extensive\nexperiments across multiple benchmarks demonstrate that our method\nsignificantly outperforms existing replay-based baselines.\n", "date": "2024-10-20", "categories": []}
{"id": "2410.15698", "title": "Solving Continual Offline RL through Selective Weights Activation on\n  Aligned Spaces", "abstract": "  Continual offline reinforcement learning (CORL) has shown impressive ability\nin diffusion-based lifelong learning systems by modeling the joint\ndistributions of trajectories. However, most research only focuses on limited\ncontinual task settings where the tasks have the same observation and action\nspace, which deviates from the realistic demands of training agents in various\nenvironments. In view of this, we propose Vector-Quantized Continual Diffuser,\nnamed VQ-CD, to break the barrier of different spaces between various tasks.\nSpecifically, our method contains two complementary sections, where the\nquantization spaces alignment provides a unified basis for the selective\nweights activation. In the quantized spaces alignment, we leverage vector\nquantization to align the different state and action spaces of various tasks,\nfacilitating continual training in the same space. Then, we propose to leverage\na unified diffusion model attached by the inverse dynamic model to master all\ntasks by selectively activating different weights according to the task-related\nsparse masks. Finally, we conduct extensive experiments on 15 continual\nlearning (CL) tasks, including conventional CL task settings (identical state\nand action spaces) and general CL task settings (various state and action\nspaces). Compared with 16 baselines, our method reaches the SOTA performance.\n", "date": "2024-10-21", "categories": []}
{"id": "2410.16154", "title": "Unsupervised Replay Strategies for Continual Learning with Limited Data", "abstract": "  Artificial neural networks (ANNs) show limited performance with scarce or\nimbalanced training data and face challenges with continuous learning, such as\nforgetting previously learned data after new tasks training. In contrast, the\nhuman brain can learn continuously and from just a few examples. This research\nexplores the impact of 'sleep', an unsupervised phase incorporating stochastic\nactivation with local Hebbian learning rules, on ANNs trained incrementally\nwith limited and imbalanced datasets, specifically MNIST and Fashion MNIST. We\ndiscovered that introducing a sleep phase significantly enhanced accuracy in\nmodels trained with limited data. When a few tasks were trained sequentially,\nsleep replay not only rescued previously learned information that had been\ncatastrophically forgetting following new task training but often enhanced\nperformance in prior tasks, especially those trained with limited data. This\nstudy highlights the multifaceted role of sleep replay in augmenting learning\nefficiency and facilitating continual learning in ANNs.\n", "date": "2024-10-21", "categories": []}
{"id": "2405.16922", "title": "Theories of synaptic memory consolidation and intelligent plasticity for\n  continual learning", "abstract": "  Humans and animals learn throughout life. Such continual learning is crucial\nfor intelligence. In this chapter, we examine the pivotal role plasticity\nmechanisms with complex internal synaptic dynamics could play in enabling this\nability in neural networks. By surveying theoretical research, we highlight two\nfundamental enablers for continual learning. First, synaptic plasticity\nmechanisms must maintain and evolve an internal state over several behaviorally\nrelevant timescales. Second, plasticity algorithms must leverage the internal\nstate to intelligently regulate plasticity at individual synapses to facilitate\nthe seamless integration of new memories while avoiding detrimental\ninterference with existing ones. Our chapter covers successful applications of\nthese principles to deep neural networks and underscores the significance of\nsynaptic metaplasticity in sustaining continual learning capabilities. Finally,\nwe outline avenues for further research to understand the brain's superb\ncontinual learning abilities and harness similar mechanisms for artificial\nintelligence systems.\n", "date": "2024-05-27", "categories": []}
{"id": "2301.10931", "title": "Towards Continual Egocentric Activity Recognition: A Multi-modal\n  Egocentric Activity Dataset for Continual Learning", "abstract": "  With the rapid development of wearable cameras, a massive collection of\negocentric video for first-person visual perception becomes available. Using\negocentric videos to predict first-person activity faces many challenges,\nincluding limited field of view, occlusions, and unstable motions. Observing\nthat sensor data from wearable devices facilitates human activity recognition,\nmulti-modal activity recognition is attracting increasing attention. However,\nthe deficiency of related dataset hinders the development of multi-modal deep\nlearning for egocentric activity recognition. Nowadays, deep learning in real\nworld has led to a focus on continual learning that often suffers from\ncatastrophic forgetting. But the catastrophic forgetting problem for egocentric\nactivity recognition, especially in the context of multiple modalities, remains\nunexplored due to unavailability of dataset. In order to assist this research,\nwe present a multi-modal egocentric activity dataset for continual learning\nnamed UESTC-MMEA-CL, which is collected by self-developed glasses integrating a\nfirst-person camera and wearable sensors. It contains synchronized data of\nvideos, accelerometers, and gyroscopes, for 32 types of daily activities,\nperformed by 10 participants. Its class types and scale are compared with other\npublicly available datasets. The statistical analysis of the sensor data is\ngiven to show the auxiliary effects for different behaviors. And results of\negocentric activity recognition are reported when using separately, and\njointly, three modalities: RGB, acceleration, and gyroscope, on a base network\narchitecture. To explore the catastrophic forgetting in continual learning\ntasks, four baseline methods are extensively evaluated with different\nmulti-modal combinations. We hope the UESTC-MMEA-CL can promote future studies\non continual learning for first-person activity recognition in wearable\napplications.\n", "date": "2023-01-26", "categories": []}
{"id": "2410.01024", "title": "GPTreeO: An R package for continual regression with dividing local\n  Gaussian processes", "abstract": "  We introduce GPTreeO, a flexible R package for scalable Gaussian process (GP)\nregression, particularly tailored to continual learning problems. GPTreeO\nbuilds upon the Dividing Local Gaussian Processes (DLGP) algorithm, in which a\nbinary tree of local GP regressors is dynamically constructed using a continual\nstream of input data. In GPTreeO we extend the original DLGP algorithm by\nallowing continual optimisation of the GP hyperparameters, incorporating\nuncertainty calibration, and introducing new strategies for how the local\npartitions are created. Moreover, the modular code structure allows users to\ninterface their favourite GP library to perform the local GP regression in\nGPTreeO. The flexibility of GPTreeO gives the user fine-grained control of the\nbalance between computational speed, accuracy, stability and smoothness. We\nconduct a sensitivity analysis to show how GPTreeO's configurable features\nimpact the regression performance in a continual learning setting.\n", "date": "2024-10-01", "categories": []}
{"id": "2410.13094", "title": "Task Consistent Prototype Learning for Incremental Few-shot Semantic\n  Segmentation", "abstract": "  Incremental Few-Shot Semantic Segmentation (iFSS) tackles a task that\nrequires a model to continually expand its segmentation capability on novel\nclasses using only a few annotated examples. Typical incremental approaches\nencounter a challenge that the objective of the base training phase (fitting\nbase classes with sufficient instances) does not align with the incremental\nlearning phase (rapidly adapting to new classes with less forgetting). This\ndisconnect can result in suboptimal performance in the incremental setting.\nThis study introduces a meta-learning-based prototype approach that encourages\nthe model to learn how to adapt quickly while preserving previous knowledge.\nConcretely, we mimic the incremental evaluation protocol during the base\ntraining session by sampling a sequence of pseudo-incremental tasks. Each task\nin the simulated sequence is trained using a meta-objective to enable rapid\nadaptation without forgetting. To enhance discrimination among class\nprototypes, we introduce prototype space redistribution learning, which\ndynamically updates class prototypes to establish optimal inter-prototype\nboundaries within the prototype space. Extensive experiments on iFSS datasets\nbuilt upon PASCAL and COCO benchmarks show the advanced performance of the\nproposed approach, offering valuable insights for addressing iFSS challenges.\n", "date": "2024-10-16", "categories": []}
{"id": "2408.13452", "title": "Data Augmentation for Continual RL via Adversarial Gradient Episodic\n  Memory", "abstract": "  Data efficiency of learning, which plays a key role in the Reinforcement\nLearning (RL) training process, becomes even more important in continual RL\nwith sequential environments. In continual RL, the learner interacts with\nnon-stationary, sequential tasks and is required to learn new tasks without\nforgetting previous knowledge. However, there is little work on implementing\ndata augmentation for continual RL. In this paper, we investigate the efficacy\nof data augmentation for continual RL. Specifically, we provide benchmarking\ndata augmentations for continual RL, by (1) summarising existing data\naugmentation methods and (2) including a new augmentation method for continual\nRL: Adversarial Augmentation with Gradient Episodic Memory (Adv-GEM). Extensive\nexperiments show that data augmentations, such as random amplitude scaling,\nstate-switch, mixup, adversarial augmentation, and Adv-GEM, can improve\nexisting continual RL algorithms in terms of their average performance,\ncatastrophic forgetting, and forward transfer, on robot control tasks. All data\naugmentation methods are implemented as plug-in modules for trivial integration\ninto continual RL methods.\n", "date": "2024-08-24", "categories": []}
{"id": "2404.10984", "title": "Graph Continual Learning with Debiased Lossless Memory Replay", "abstract": "  Real-life graph data often expands continually, rendering the learning of\ngraph neural networks (GNNs) on static graph data impractical. Graph continual\nlearning (GCL) tackles this problem by continually adapting GNNs to the\nexpanded graph of the current task while maintaining the performance over the\ngraph of previous tasks. Memory replay-based methods, which aim to replay data\nof previous tasks when learning new tasks, have been explored as one principled\napproach to mitigate the forgetting of the knowledge learned from the previous\ntasks. In this paper we extend this methodology with a novel framework, called\nDebiased Lossless Memory replay (DeLoMe). Unlike existing methods that sample\nnodes/edges of previous graphs to construct the memory, DeLoMe learns small\nlossless synthetic node representations as the memory. The learned memory can\nnot only preserve the graph data privacy but also capture the holistic graph\ninformation, for which the sampling-based methods are not viable. Further,\nprior methods suffer from bias toward the current task due to the data\nimbalance between the classes in the memory data and the current data. A\ndebiased GCL loss function is devised in DeLoMe to effectively alleviate this\nbias. Extensive experiments on four graph datasets show the effectiveness of\nDeLoMe under both class- and task-incremental learning settings.\n", "date": "2024-04-17", "categories": []}
{"id": "2410.10923", "title": "ATLAS: Adapter-Based Multi-Modal Continual Learning with a Two-Stage\n  Learning Strategy", "abstract": "  While vision-and-language models significantly advance in many fields, the\nchallenge of continual learning is unsolved. Parameter-efficient modules like\nadapters and prompts present a promising way to alleviate catastrophic\nforgetting. However, existing works usually learn individual adapters for each\ntask, which may result in redundant knowledge among adapters. Moreover, they\ncontinue to use the original pre-trained model to initialize the downstream\nmodel, leading to negligible changes in the model's generalization compared to\nthe original model. In addition, there is still a lack of research\ninvestigating the consequences of integrating a multi-modal model into the\nupdating procedure for both uni-modal and multi-modal tasks and the subsequent\nimpacts it has on downstream tasks. In this paper, we propose an adapter-based\ntwo-stage learning paradigm, a multi-modal continual learning scheme that\nconsists of experience-based learning and novel knowledge expansion, which\nhelps the model fully use experience knowledge and compensate for novel\nknowledge. Extensive experiments demonstrate that our method is proficient for\ncontinual learning. It expands the distribution of representation upstream\nwhile also minimizing the negative impact of forgetting previous tasks.\nAdditionally, it enhances the generalization capability for downstream tasks.\nFurthermore, we incorporate both multi-modal and uni-modal tasks into upstream\ncontinual learning. We observe that learning from upstream tasks can help with\ndownstream tasks. Our code will be available at:\nhttps://github.com/lihong2303/ATLAS.\n", "date": "2024-10-14", "categories": []}
{"id": "2410.09237", "title": "Foundation Model-Powered 3D Few-Shot Class Incremental Learning via\n  Training-free Adaptor", "abstract": "  Recent advances in deep learning for processing point clouds hold increased\ninterest in Few-Shot Class Incremental Learning (FSCIL) for 3D computer vision.\nThis paper introduces a new method to tackle the Few-Shot Continual Incremental\nLearning (FSCIL) problem in 3D point cloud environments. We leverage a\nfoundational 3D model trained extensively on point cloud data. Drawing from\nrecent improvements in foundation models, known for their ability to work well\nacross different tasks, we propose a novel strategy that does not require\nadditional training to adapt to new tasks. Our approach uses a dual cache\nsystem: first, it uses previous test samples based on how confident the model\nwas in its predictions to prevent forgetting, and second, it includes a small\nnumber of new task samples to prevent overfitting. This dynamic adaptation\nensures strong performance across different learning tasks without needing lots\nof fine-tuning. We tested our approach on datasets like ModelNet, ShapeNet,\nScanObjectNN, and CO3D, showing that it outperforms other FSCIL methods and\ndemonstrating its effectiveness and versatility. The code is available at\n\\url{https://github.com/ahmadisahar/ACCV_FCIL3D}.\n", "date": "2024-10-11", "categories": []}
{"id": "2410.10521", "title": "Continual Deep Reinforcement Learning to Prevent Catastrophic Forgetting\n  in Jamming Mitigation", "abstract": "  Deep Reinforcement Learning (DRL) has been highly effective in learning from\nand adapting to RF environments and thus detecting and mitigating jamming\neffects to facilitate reliable wireless communications. However, traditional\nDRL methods are susceptible to catastrophic forgetting (namely forgetting old\ntasks when learning new ones), especially in dynamic wireless environments\nwhere jammer patterns change over time. This paper considers an anti-jamming\nsystem and addresses the challenge of catastrophic forgetting in DRL applied to\njammer detection and mitigation. First, we demonstrate the impact of\ncatastrophic forgetting in DRL when applied to jammer detection and mitigation\ntasks, where the network forgets previously learned jammer patterns while\nadapting to new ones. This catastrophic interference undermines the\neffectiveness of the system, particularly in scenarios where the environment is\nnon-stationary. We present a method that enables the network to retain\nknowledge of old jammer patterns while learning to handle new ones. Our\napproach substantially reduces catastrophic forgetting, allowing the\nanti-jamming system to learn new tasks without compromising its ability to\nperform previously learned tasks effectively. Furthermore, we introduce a\nsystematic methodology for sequentially learning tasks in the anti-jamming\nframework. By leveraging continual DRL techniques based on PackNet, we achieve\nsuperior anti-jamming performance compared to standard DRL methods. Our\nproposed approach not only addresses catastrophic forgetting but also enhances\nthe adaptability and robustness of the system in dynamic jamming environments.\nWe demonstrate the efficacy of our method in preserving knowledge of past\njammer patterns, learning new tasks efficiently, and achieving superior\nanti-jamming performance compared to traditional DRL approaches.\n", "date": "2024-10-14", "categories": []}
{"id": "2409.19245", "title": "Forgetting, Ignorance or Myopia: Revisiting Key Challenges in Online\n  Continual Learning", "abstract": "  Online continual learning requires the models to learn from constant, endless\nstreams of data. While significant efforts have been made in this field, most\nwere focused on mitigating the catastrophic forgetting issue to achieve better\nclassification ability, at the cost of a much heavier training workload. They\noverlooked that in real-world scenarios, e.g., in high-speed data stream\nenvironments, data do not pause to accommodate slow models. In this paper, we\nemphasize that model throughput -- defined as the maximum number of training\nsamples that a model can process within a unit of time -- is equally important.\nIt directly limits how much data a model can utilize and presents a challenging\ndilemma for current methods. With this understanding, we revisit key challenges\nin OCL from both empirical and theoretical perspectives, highlighting two\ncritical issues beyond the well-documented catastrophic forgetting: Model's\nignorance: the single-pass nature of OCL challenges models to learn effective\nfeatures within constrained training time and storage capacity, leading to a\ntrade-off between effective learning and model throughput; Model's myopia: the\nlocal learning nature of OCL on the current task leads the model to adopt\noverly simplified, task-specific features and excessively sparse classifier,\nresulting in the gap between the optimal solution for the current task and the\nglobal objective. To tackle these issues, we propose the Non-sparse Classifier\nEvolution framework (NsCE) to facilitate effective global discriminative\nfeature learning with minimal time cost. NsCE integrates non-sparse maximum\nseparation regularization and targeted experience replay techniques with the\nhelp of pre-trained models, enabling rapid acquisition of new globally\ndiscriminative features.\n", "date": "2024-09-28", "categories": []}
{"id": "2410.08905", "title": "Lifelong Event Detection via Optimal Transport", "abstract": "  Continual Event Detection (CED) poses a formidable challenge due to the\ncatastrophic forgetting phenomenon, where learning new tasks (with new coming\nevent types) hampers performance on previous ones. In this paper, we introduce\na novel approach, Lifelong Event Detection via Optimal Transport (LEDOT), that\nleverages optimal transport principles to align the optimization of our\nclassification module with the intrinsic nature of each class, as defined by\ntheir pre-trained language modeling. Our method integrates replay sets,\nprototype latent representations, and an innovative Optimal Transport\ncomponent. Extensive experiments on MAVEN and ACE datasets demonstrate LEDOT's\nsuperior performance, consistently outperforming state-of-the-art baselines.\nThe results underscore LEDOT as a pioneering solution in continual event\ndetection, offering a more effective and nuanced approach to addressing\ncatastrophic forgetting in evolving environments.\n", "date": "2024-10-11", "categories": []}
{"id": "2312.17004", "title": "Continual Learning in Medical Image Analysis: A Comprehensive Review of\n  Recent Advancements and Future Prospects", "abstract": "  Medical imaging analysis has witnessed remarkable advancements even\nsurpassing human-level performance in recent years, driven by the rapid\ndevelopment of advanced deep-learning algorithms. However, when the inference\ndataset slightly differs from what the model has seen during one-time training,\nthe model performance is greatly compromised. The situation requires restarting\nthe training process using both the old and the new data which is\ncomputationally costly, does not align with the human learning process, and\nimposes storage constraints and privacy concerns. Alternatively, continual\nlearning has emerged as a crucial approach for developing unified and\nsustainable deep models to deal with new classes, tasks, and the drifting\nnature of data in non-stationary environments for various application areas.\nContinual learning techniques enable models to adapt and accumulate knowledge\nover time, which is essential for maintaining performance on evolving datasets\nand novel tasks. This systematic review paper provides a comprehensive overview\nof the state-of-the-art in continual learning techniques applied to medical\nimaging analysis. We present an extensive survey of existing research, covering\ntopics including catastrophic forgetting, data drifts, stability, and\nplasticity requirements. Further, an in-depth discussion of key components of a\ncontinual learning framework such as continual learning scenarios, techniques,\nevaluation schemes, and metrics is provided. Continual learning techniques\nencompass various categories, including rehearsal, regularization,\narchitectural, and hybrid strategies. We assess the popularity and\napplicability of continual learning categories in various medical sub-fields\nlike radiology and histopathology...\n", "date": "2023-12-28", "categories": []}
{"id": "2410.06535", "title": "Happy: A Debiased Learning Framework for Continual Generalized Category\n  Discovery", "abstract": "  Constantly discovering novel concepts is crucial in evolving environments.\nThis paper explores the underexplored task of Continual Generalized Category\nDiscovery (C-GCD), which aims to incrementally discover new classes from\nunlabeled data while maintaining the ability to recognize previously learned\nclasses. Although several settings are proposed to study the C-GCD task, they\nhave limitations that do not reflect real-world scenarios. We thus study a more\npractical C-GCD setting, which includes more new classes to be discovered over\na longer period, without storing samples of past classes. In C-GCD, the model\nis initially trained on labeled data of known classes, followed by multiple\nincremental stages where the model is fed with unlabeled data containing both\nold and new classes. The core challenge involves two conflicting objectives:\ndiscover new classes and prevent forgetting old ones. We delve into the\nconflicts and identify that models are susceptible to prediction bias and\nhardness bias. To address these issues, we introduce a debiased learning\nframework, namely Happy, characterized by Hardness-aware prototype sampling and\nsoft entropy regularization. For the prediction bias, we first introduce\nclustering-guided initialization to provide robust features. In addition, we\npropose soft entropy regularization to assign appropriate probabilities to new\nclasses, which can significantly enhance the clustering performance of new\nclasses. For the harness bias, we present the hardness-aware prototype\nsampling, which can effectively reduce the forgetting issue for previously seen\nclasses, especially for difficult classes. Experimental results demonstrate our\nmethod proficiently manages the conflicts of C-GCD and achieves remarkable\nperformance across various datasets, e.g., 7.5% overall gains on ImageNet-100.\nOur code is publicly available at https://github.com/mashijie1028/Happy-CGCD.\n", "date": "2024-10-09", "categories": []}
{"id": "2410.05627", "title": "CLOSER: Towards Better Representation Learning for Few-Shot\n  Class-Incremental Learning", "abstract": "  Aiming to incrementally learn new classes with only few samples while\npreserving the knowledge of base (old) classes, few-shot class-incremental\nlearning (FSCIL) faces several challenges, such as overfitting and catastrophic\nforgetting. Such a challenging problem is often tackled by fixing a feature\nextractor trained on base classes to reduce the adverse effects of overfitting\nand forgetting. Under such formulation, our primary focus is representation\nlearning on base classes to tackle the unique challenge of FSCIL:\nsimultaneously achieving the transferability and the discriminability of the\nlearned representation. Building upon the recent efforts for enhancing\ntransferability, such as promoting the spread of features, we find that trying\nto secure the spread of features within a more confined feature space enables\nthe learned representation to strike a better balance between transferability\nand discriminability. Thus, in stark contrast to prior beliefs that the\ninter-class distance should be maximized, we claim that the closer different\nclasses are, the better for FSCIL. The empirical results and analysis from the\nperspective of information bottleneck theory justify our simple yet seemingly\ncounter-intuitive representation learning method, raising research questions\nand suggesting alternative research directions. The code is available at\nhttps://github.com/JungHunOh/CLOSER_ECCV2024.\n", "date": "2024-10-08", "categories": []}
{"id": "2410.05648", "title": "Does RoBERTa Perform Better than BERT in Continual Learning: An\n  Attention Sink Perspective", "abstract": "  Continual learning (CL) aims to train models that can sequentially learn new\ntasks without forgetting previous tasks' knowledge. Although previous works\nobserved that pre-training can benefit CL, it remains unclear whether a\npre-trained model with higher downstream capacity also performs better in CL.\nIn this paper, we observe that pre-trained models may allocate high attention\nscores to some 'sink' tokens, such as [SEP] tokens, which are ubiquitous across\nvarious tasks. Such attention sinks may lead to models' over-smoothing in\nsingle-task learning and interference in sequential tasks' learning, which may\ncompromise the models' CL performance despite their high pre-trained\ncapabilities. To reduce these effects, we propose a pre-scaling mechanism that\nencourages attention diversity across all tokens. Specifically, it first scales\nthe task's attention to the non-sink tokens in a probing stage, and then\nfine-tunes the model with scaling. Experiments show that pre-scaling yields\nsubstantial improvements in CL without experience replay, or progressively\nstoring parameters from previous tasks.\n", "date": "2024-10-08", "categories": []}
{"id": "2410.05800", "title": "Core Tokensets for Data-efficient Sequential Training of Transformers", "abstract": "  Deep networks are frequently tuned to novel tasks and continue learning from\nongoing data streams. Such sequential training requires consolidation of new\nand past information, a challenge predominantly addressed by retaining the most\nimportant data points - formally known as coresets. Traditionally, these\ncoresets consist of entire samples, such as images or sentences. However,\nrecent transformer architectures operate on tokens, leading to the famous\nassertion that an image is worth 16x16 words. Intuitively, not all of these\ntokens are equally informative or memorable. Going beyond coresets, we thus\npropose to construct a deeper-level data summary on the level of tokens. Our\nrespectively named core tokensets both select the most informative data points\nand leverage feature attribution to store only their most relevant features. We\ndemonstrate that core tokensets yield significant performance retention in\nincremental image classification, open-ended visual question answering, and\ncontinual image captioning with significantly reduced memory. In fact, we\nempirically find that a core tokenset of 1\\% of the data performs comparably to\nat least a twice as large and up to 10 times larger coreset.\n", "date": "2024-10-08", "categories": []}
{"id": "2410.05899", "title": "Brain-inspired continual pre-trained learner via silent synaptic\n  consolidation", "abstract": "  Pre-trained models have demonstrated impressive generalization capabilities,\nyet they remain vulnerable to catastrophic forgetting when incrementally\ntrained on new tasks. Existing architecture-based strategies encounter two\nprimary challenges: 1) Integrating a pre-trained network with a trainable\nsub-network complicates the delicate balance between learning plasticity and\nmemory stability across evolving tasks during learning. 2) The absence of\nrobust interconnections between pre-trained networks and various sub-networks\nlimits the effective retrieval of pertinent information during inference. In\nthis study, we introduce the Artsy, inspired by the activation mechanisms of\nsilent synapses via spike-timing-dependent plasticity observed in mature\nbrains, to enhance the continual learning capabilities of pre-trained models.\nThe Artsy integrates two key components: During training, the Artsy mimics\nmature brain dynamics by maintaining memory stability for previously learned\nknowledge within the pre-trained network while simultaneously promoting\nlearning plasticity in task-specific sub-networks. During inference, artificial\nsilent and functional synapses are utilized to establish precise connections\nbetween the pre-synaptic neurons in the pre-trained network and the\npost-synaptic neurons in the sub-networks, facilitated through synaptic\nconsolidation, thereby enabling effective extraction of relevant information\nfrom test samples. Comprehensive experimental evaluations reveal that our model\nsignificantly outperforms conventional methods on class-incremental learning\ntasks, while also providing enhanced biological interpretability for\narchitecture-based approaches. Moreover, we propose that the Artsy offers a\npromising avenue for simulating biological synaptic mechanisms, potentially\nadvancing our understanding of neural plasticity in both artificial and\nbiological systems.\n", "date": "2024-10-08", "categories": []}
{"id": "2410.07110", "title": "Continual Learning: Less Forgetting, More OOD Generalization via\n  Adaptive Contrastive Replay", "abstract": "  Machine learning models often suffer from catastrophic forgetting of\npreviously learned knowledge when learning new classes. Various methods have\nbeen proposed to mitigate this issue. However, rehearsal-based learning, which\nretains samples from previous classes, typically achieves good performance but\ntends to memorize specific instances, struggling with Out-of-Distribution (OOD)\ngeneralization. This often leads to high forgetting rates and poor\ngeneralization. Surprisingly, the OOD generalization capabilities of these\nmethods have been largely unexplored. In this paper, we highlight this issue\nand propose a simple yet effective strategy inspired by contrastive learning\nand data-centric principles to address it. We introduce Adaptive Contrastive\nReplay (ACR), a method that employs dual optimization to simultaneously train\nboth the encoder and the classifier. ACR adaptively populates the replay buffer\nwith misclassified samples while ensuring a balanced representation of classes\nand tasks. By refining the decision boundary in this way, ACR achieves a\nbalance between stability and plasticity. Our method significantly outperforms\nprevious approaches in terms of OOD generalization, achieving an improvement of\n13.41\\% on Split CIFAR-100, 9.91\\% on Split Mini-ImageNet, and 5.98\\% on Split\nTiny-ImageNet.\n", "date": "2024-10-09", "categories": []}
{"id": "2307.01430", "title": "Continual Learning in Open-vocabulary Classification with Complementary\n  Memory Systems", "abstract": "  We introduce a method for flexible and efficient continual learning in\nopen-vocabulary image classification, drawing inspiration from the\ncomplementary learning systems observed in human cognition. Specifically, we\npropose to combine predictions from a CLIP zero-shot model and the\nexemplar-based model, using the zero-shot estimated probability that a sample's\nclass is within the exemplar classes. We also propose a \"tree probe\" method, an\nadaption of lazy learning principles, which enables fast learning from new\nexamples with competitive accuracy to batch-trained linear models. We test in\ndata incremental, class incremental, and task incremental settings, as well as\nability to perform flexible inference on varying subsets of zero-shot and\nlearned categories. Our proposed method achieves a good balance of learning\nspeed, target task effectiveness, and zero-shot effectiveness. Code will be\navailable at https://github.com/jessemelpolio/TreeProbe.\n", "date": "2023-07-04", "categories": []}
{"id": "2312.00246", "title": "Directions of Curvature as an Explanation for Loss of Plasticity", "abstract": "  Loss of plasticity is a phenomenon in which neural networks lose their\nability to learn from new experience. Despite being empirically observed in\nseveral problem settings, little is understood about the mechanisms that lead\nto loss of plasticity. In this paper, we offer a consistent explanation for\nloss of plasticity: Neural networks lose directions of curvature during\ntraining and that loss of plasticity can be attributed to this reduction in\ncurvature. To support such a claim, we provide a systematic investigation of\nloss of plasticity across continual learning tasks using MNIST, CIFAR-10 and\nImageNet. Our findings illustrate that loss of curvature directions coincides\nwith loss of plasticity, while also showing that previous explanations are\ninsufficient to explain loss of plasticity in all settings. Lastly, we show\nthat regularizers which mitigate loss of plasticity also preserve curvature,\nmotivating a simple distributional regularizer that proves to be effective\nacross the problem settings we considered.\n", "date": "2023-11-30", "categories": []}
{"id": "2406.17245", "title": "Unlocking Continual Learning Abilities in Language Models", "abstract": "  Language models (LMs) exhibit impressive performance and generalization\ncapabilities. However, LMs struggle with the persistent challenge of\ncatastrophic forgetting, which undermines their long-term sustainability in\ncontinual learning (CL). Existing approaches usually address the issue by\nincorporating old task data or task-wise inductive bias into LMs. However, old\ndata and accurate task information are often unavailable or costly to collect,\nhindering the availability of current CL approaches for LMs. To address this\nlimitation, we introduce $\\textbf{MIGU}$ ($\\textbf{M}$agn$\\textbf{I}$tude-based\n$\\textbf{G}$radient $\\textbf{U}$pdating for continual learning), a\nrehearsal-free and task-label-free method that only updates the model\nparameters with large magnitudes of output in LMs' linear layers. MIGU is based\non our observation that the L1-normalized magnitude distribution of the output\nin LMs' linear layers is different when the LM models deal with different task\ndata. By imposing this simple constraint on the gradient update process, we can\nleverage the inherent behaviors of LMs, thereby unlocking their innate CL\nabilities. Our experiments demonstrate that MIGU is universally applicable to\nall three LM architectures (T5, RoBERTa, and Llama2), delivering\nstate-of-the-art or on-par performance across continual finetuning and\ncontinual pre-training settings on four CL benchmarks. For example, MIGU brings\na 15.2% average accuracy improvement over conventional parameter-efficient\nfinetuning baselines in a 15-task CL benchmark. MIGU can also seamlessly\nintegrate with all three existing CL types to further enhance performance. Code\nis available at https://github.com/wenyudu/MIGU.\n", "date": "2024-06-25", "categories": []}
{"id": "2407.17467", "title": "CMR Scaling Law: Predicting Critical Mixture Ratios for Continual\n  Pre-training of Language Models", "abstract": "  Large Language Models (LLMs) excel in diverse tasks but often underperform in\nspecialized fields due to limited domain-specific or proprietary corpus.\nContinual pre-training (CPT) enhances LLM capabilities by imbuing new\ndomain-specific or proprietary knowledge while replaying general corpus to\nprevent catastrophic forgetting. The data mixture ratio of general corpus and\ndomain-specific corpus, however, has been chosen heuristically, leading to\nsub-optimal training efficiency in practice. In this context, we attempt to\nre-visit the scaling behavior of LLMs under the hood of CPT, and discover a\npower-law relationship between loss, mixture ratio, and training tokens scale.\nWe formalize the trade-off between general and domain-specific capabilities,\nleading to a well-defined Critical Mixture Ratio (CMR) of general and domain\ndata. By striking the balance, CMR maintains the model's general ability and\nachieves the desired domain transfer, ensuring the highest utilization of\navailable resources. Considering the balance between efficiency and\neffectiveness, CMR can be regarded as the optimal mixture ratio. Through\nextensive experiments, we ascertain the predictability of CMR, propose CMR\nscaling law and have substantiated its generalization. These findings offer\npractical guidelines for optimizing LLM training in specialized domains,\nensuring both general and domain-specific performance while efficiently\nmanaging training resources.\n", "date": "2024-07-24", "categories": []}
{"id": "2409.05898", "title": "Simplex-enabled Safe Continual Learning Machine", "abstract": "  This paper proposes the SeC-Learning Machine: Simplex-enabled safe continual\nlearning for safety-critical autonomous systems. The SeC-learning machine is\nbuilt on Simplex logic (that is, ``using simplicity to control complexity'')\nand physics-regulated deep reinforcement learning (Phy-DRL). The SeC-learning\nmachine thus constitutes HP (high performance)-Student, HA (high\nassurance)-Teacher, and Coordinator. Specifically, the HP-Student is a\npre-trained high-performance but not fully verified Phy-DRL, continuing to\nlearn in a real plant to tune the action policy to be safe. In contrast, the\nHA-Teacher is a mission-reduced, physics-model-based, and verified design. As a\ncomplementary, HA-Teacher has two missions: backing up safety and correcting\nunsafe learning. The Coordinator triggers the interaction and the switch\nbetween HP-Student and HA-Teacher. Powered by the three interactive components,\nthe SeC-learning machine can i) assure lifetime safety (i.e., safety guarantee\nin any continual-learning stage, regardless of HP-Student's success or\nconvergence), ii) address the Sim2Real gap, and iii) learn to tolerate unknown\nunknowns in real plants. The experiments on a cart-pole system and a real\nquadruped robot demonstrate the distinguished features of the SeC-learning\nmachine, compared with continual learning built on state-of-the-art safe DRL\nframeworks with approaches to addressing the Sim2Real gap.\n", "date": "2024-09-05", "categories": []}
{"id": "2309.03487", "title": "Privacy-preserving Continual Federated Clustering via Adaptive Resonance\n  Theory", "abstract": "  With the increasing importance of data privacy protection, various\nprivacy-preserving machine learning methods have been proposed. In the\nclustering domain, various algorithms with a federated learning framework\n(i.e., federated clustering) have been actively studied and showed high\nclustering performance while preserving data privacy. However, most of the base\nclusterers (i.e., clustering algorithms) used in existing federated clustering\nalgorithms need to specify the number of clusters in advance. These algorithms,\ntherefore, are unable to deal with data whose distributions are unknown or\ncontinually changing. To tackle this problem, this paper proposes a\nprivacy-preserving continual federated clustering algorithm. In the proposed\nalgorithm, an adaptive resonance theory-based clustering algorithm capable of\ncontinual learning is used as a base clusterer. Therefore, the proposed\nalgorithm inherits the ability of continual learning. Experimental results with\nsynthetic and real-world datasets show that the proposed algorithm has superior\nclustering performance to state-of-the-art federated clustering algorithms\nwhile realizing data privacy protection and continual learning ability. The\nsource code is available at \\url{https://github.com/Masuyama-lab/FCAC}.\n", "date": "2023-09-07", "categories": []}
{"id": "2410.00074", "title": "Collaborative Knowledge Distillation via a Learning-by-Education Node\n  Community", "abstract": "  A novel Learning-by-Education Node Community framework (LENC) for\nCollaborative Knowledge Distillation (CKD) is presented, which facilitates\ncontinual collective learning through effective knowledge exchanges among\ndiverse deployed Deep Neural Network (DNN) peer nodes. These DNNs dynamically\nand autonomously adopt either the role of a student, seeking knowledge, or that\nof a teacher, imparting knowledge, fostering a collaborative learning\nenvironment. The proposed framework enables efficient knowledge transfer among\nparticipating DNN nodes as needed, while enhancing their learning capabilities\nand promoting their collaboration. LENC addresses the challenges of handling\ndiverse training data distributions and the limitations of individual DNN node\nlearning abilities. It ensures the exploitation of the best available teacher\nknowledge upon learning a new task and protects the DNN nodes from catastrophic\nforgetting. Additionally, it innovates by enabling collaborative multitask\nknowledge distillation, while addressing the problem of task-agnostic continual\nlearning, as DNN nodes have no information on task boundaries. Experimental\nevaluation on a proof-of-concept implementation demonstrates the LENC\nframework's functionalities and benefits across multiple DNN learning and\ninference scenarios. The conducted experiments showcase its ability to\ngradually maximize the average test accuracy of the community of interacting\nDNN nodes in image classification problems, by appropriately leveraging the\ncollective knowledge of all node peers. The LENC framework achieves\nstate-of-the-art performance in on-line unlabelled CKD.\n", "date": "2024-09-30", "categories": []}
{"id": "2410.00334", "title": "Preserving Generalization of Language models in Few-shot Continual\n  Relation Extraction", "abstract": "  Few-shot Continual Relations Extraction (FCRE) is an emerging and dynamic\narea of study where models can sequentially integrate knowledge from new\nrelations with limited labeled data while circumventing catastrophic forgetting\nand preserving prior knowledge from pre-trained backbones. In this work, we\nintroduce a novel method that leverages often-discarded language model heads.\nBy employing these components via a mutual information maximization strategy,\nour approach helps maintain prior knowledge from the pre-trained backbone and\nstrategically aligns the primary classification head, thereby enhancing model\nperformance. Furthermore, we explore the potential of Large Language Models\n(LLMs), renowned for their wealth of knowledge, in addressing FCRE challenges.\nOur comprehensive experimental results underscore the efficacy of the proposed\nmethod and offer valuable insights for future work.\n", "date": "2024-10-01", "categories": []}
{"id": "2410.00868", "title": "Fine-Grained Gradient Restriction: A Simple Approach for Mitigating\n  Catastrophic Forgetting", "abstract": "  A fundamental challenge in continual learning is to balance the trade-off\nbetween learning new tasks and remembering the previously acquired knowledge.\nGradient Episodic Memory (GEM) achieves this balance by utilizing a subset of\npast training samples to restrict the update direction of the model parameters.\nIn this work, we start by analyzing an often overlooked hyper-parameter in GEM,\nthe memory strength, which boosts the empirical performance by further\nconstraining the update direction. We show that memory strength is effective\nmainly because it improves GEM's generalization ability and therefore leads to\na more favorable trade-off. By this finding, we propose two approaches that\nmore flexibly constrain the update direction. Our methods are able to achieve\nuniformly better Pareto Frontiers of remembering old and learning new knowledge\nthan using memory strength. We further propose a computationally efficient\nmethod to approximately solve the optimization problem with more constraints.\n", "date": "2024-10-01", "categories": []}
{"id": "2311.06460", "title": "Online Continual Learning via Logit Adjusted Softmax", "abstract": "  Online continual learning is a challenging problem where models must learn\nfrom a non-stationary data stream while avoiding catastrophic forgetting.\nInter-class imbalance during training has been identified as a major cause of\nforgetting, leading to model prediction bias towards recently learned classes.\nIn this paper, we theoretically analyze that inter-class imbalance is entirely\nattributed to imbalanced class-priors, and the function learned from\nintra-class intrinsic distributions is the Bayes-optimal classifier. To that\nend, we present that a simple adjustment of model logits during training can\neffectively resist prior class bias and pursue the corresponding Bayes-optimum.\nOur proposed method, Logit Adjusted Softmax, can mitigate the impact of\ninter-class imbalance not only in class-incremental but also in realistic\ngeneral setups, with little additional computational cost. We evaluate our\napproach on various benchmarks and demonstrate significant performance\nimprovements compared to prior arts. For example, our approach improves the\nbest baseline by 4.6% on CIFAR10.\n", "date": "2023-11-11", "categories": []}
{"id": "2405.03537", "title": "Exploring the Efficacy of Federated-Continual Learning Nodes with\n  Attention-Based Classifier for Robust Web Phishing Detection: An Empirical\n  Investigation", "abstract": "  Web phishing poses a dynamic threat, requiring detection systems to quickly\nadapt to the latest tactics. Traditional approaches of accumulating data and\nperiodically retraining models are outpaced. We propose a novel paradigm\ncombining federated learning and continual learning, enabling distributed nodes\nto continually update models on streams of new phishing data, without\naccumulating data. These locally adapted models are then aggregated at a\ncentral server via federated learning. To enhance detection, we introduce a\ncustom attention-based classifier model with residual connections, tailored for\nweb phishing, leveraging attention mechanisms to capture intricate phishing\npatterns. We evaluate our hybrid learning paradigm across continual learning\nstrategies (cumulative, replay, MIR, LwF) and model architectures through an\nempirical investigation. Our main contributions are: (1) a new hybrid\nfederated-continual learning paradigm for robust web phishing detection, and\n(2) a novel attention + residual connections based model explicitly designed\nfor this task, attaining 0.93 accuracy, 0.90 precision, 0.96 recall and 0.93\nf1-score with the LwF strategy, outperforming traditional approaches in\ndetecting emerging phishing threats while retaining past knowledge.\n", "date": "2024-05-06", "categories": []}
{"id": "2409.05312", "title": "Open-World Dynamic Prompt and Continual Visual Representation Learning", "abstract": "  The open world is inherently dynamic, characterized by ever-evolving concepts\nand distributions. Continual learning (CL) in this dynamic open-world\nenvironment presents a significant challenge in effectively generalizing to\nunseen test-time classes. To address this challenge, we introduce a new\npractical CL setting tailored for open-world visual representation learning. In\nthis setting, subsequent data streams systematically introduce novel classes\nthat are disjoint from those seen in previous training phases, while also\nremaining distinct from the unseen test classes. In response, we present\nDynamic Prompt and Representation Learner (DPaRL), a simple yet effective\nPrompt-based CL (PCL) method. Our DPaRL learns to generate dynamic prompts for\ninference, as opposed to relying on a static prompt pool in previous PCL\nmethods. In addition, DPaRL jointly learns dynamic prompt generation and\ndiscriminative representation at each training stage whereas prior PCL methods\nonly refine the prompt learning throughout the process. Our experimental\nresults demonstrate the superiority of our approach, surpassing\nstate-of-the-art methods on well-established open-world image retrieval\nbenchmarks by an average of 4.7% improvement in Recall@1 performance.\n", "date": "2024-09-09", "categories": []}
{"id": "2409.19611", "title": "Learning Attentional Mixture of LoRAs for Language Model Continual\n  Learning", "abstract": "  Fine-tuning large language models (LLMs) with Low-Rank adaption (LoRA) is\nwidely acknowledged as an effective approach for continual learning for new\ntasks. However, it often suffers from catastrophic forgetting when dealing with\nmultiple tasks sequentially. To this end, we propose Attentional Mixture of\nLoRAs (AM-LoRA), a continual learning approach tailored for LLMs. Specifically,\nAM-LoRA learns a sequence of LoRAs for a series of tasks to continually learn\nknowledge from different tasks. The key of our approach is that we devise an\nattention mechanism as a knowledge mixture module to adaptively integrate\ninformation from each LoRA. With the attention mechanism, AM-LoRA can\nefficiently leverage the distinctive contributions of each LoRA, while\nmitigating the risk of mutually negative interactions among them that may lead\nto catastrophic forgetting. Moreover, we further introduce $L1$ norm in the\nlearning process to make the attention vector more sparse. The sparse\nconstraints can enable the model to lean towards selecting a few highly\nrelevant LoRAs, rather than aggregating and weighting all LoRAs collectively,\nwhich can further reduce the impact stemming from mutual interference.\nExperimental results on continual learning benchmarks indicate the superiority\nof our proposed method.\n", "date": "2024-09-29", "categories": []}
{"id": "2409.13550", "title": "A preliminary study on continual learning in computer vision using\n  Kolmogorov-Arnold Networks", "abstract": "  Deep learning has long been dominated by multi-layer perceptrons (MLPs),\nwhich have demonstrated superiority over other optimizable models in various\ndomains. Recently, a new alternative to MLPs has emerged - Kolmogorov-Arnold\nNetworks (KAN)- which are based on a fundamentally different mathematical\nframework. According to their authors, KANs address several major issues in\nMLPs, such as catastrophic forgetting in continual learning scenarios. However,\nthis claim has only been supported by results from a regression task on a toy\n1D dataset. In this paper, we extend the investigation by evaluating the\nperformance of KANs in continual learning tasks within computer vision,\nspecifically using the MNIST datasets. To this end, we conduct a structured\nanalysis of the behavior of MLPs and two KAN-based models in a\nclass-incremental learning scenario, ensuring that the architectures involved\nhave the same number of trainable parameters. Our results demonstrate that an\nefficient version of KAN outperforms both traditional MLPs and the original KAN\nimplementation. We further analyze the influence of hyperparameters in MLPs and\nKANs, as well as the impact of certain trainable parameters in KANs, such as\nbias and scale weights. Additionally, we provide a preliminary investigation of\nrecent KAN-based convolutional networks and compare their performance with that\nof traditional convolutional neural networks. Our codes can be found at\nhttps://github.com/MrPio/KAN-Continual_Learning_tests.\n", "date": "2024-09-20", "categories": []}
{"id": "2409.18664", "title": "How green is continual learning, really? Analyzing the energy\n  consumption in continual training of vision foundation models", "abstract": "  With the ever-growing adoption of AI, its impact on the environment is no\nlonger negligible. Despite the potential that continual learning could have\ntowards Green AI, its environmental sustainability remains relatively\nuncharted. In this work we aim to gain a systematic understanding of the energy\nefficiency of continual learning algorithms. To that end, we conducted an\nextensive set of empirical experiments comparing the energy consumption of\nrecent representation-, prompt-, and exemplar-based continual learning\nalgorithms and two standard baseline (fine tuning and joint training) when used\nto continually adapt a pre-trained ViT-B/16 foundation model. We performed our\nexperiments on three standard datasets: CIFAR-100, ImageNet-R, and DomainNet.\nAdditionally, we propose a novel metric, the Energy NetScore, which we use\nmeasure the algorithm efficiency in terms of energy-accuracy trade-off. Through\nnumerous evaluations varying the number and size of the incremental learning\nsteps, our experiments demonstrate that different types of continual learning\nalgorithms have very different impacts on energy consumption during both\ntraining and inference. Although often overlooked in the continual learning\nliterature, we found that the energy consumed during the inference phase is\ncrucial for evaluating the environmental sustainability of continual learning\nmodels.\n", "date": "2024-09-27", "categories": []}
{"id": "2301.05180", "title": "Effective Decision Boundary Learning for Class Incremental Learning", "abstract": "  Rehearsal approaches in class incremental learning (CIL) suffer from decision\nboundary overfitting to new classes, which is mainly caused by two factors:\ninsufficiency of old classes data for knowledge distillation and imbalanced\ndata learning between the learned and new classes because of the limited\nstorage memory. In this work, we present a simple but effective approach to\ntackle these two factors. First, we employ a re-sampling strategy and Mixup\nK}nowledge D}istillation (Re-MKD) to improve the performances of KD, which\nwould greatly alleviate the overfitting problem. Specifically, we combine mixup\nand re-sampling strategies to synthesize adequate data used in KD training that\nare more consistent with the latent distribution between the learned and new\nclasses. Second, we propose a novel incremental influence balance (IIB) method\nfor CIL to tackle the classification of imbalanced data by extending the\ninfluence balance method into the CIL setting, which re-weights samples by\ntheir influences to create a proper decision boundary. With these two\nimprovements, we present the effective decision boundary learning algorithm\n(EDBL) which improves the performance of KD and deals with the imbalanced data\nlearning simultaneously. Experiments show that the proposed EDBL achieves\nstate-of-the-art performances on several CIL benchmarks.\n", "date": "2023-01-12", "categories": []}
{"id": "2409.17300", "title": "Neural Network Plasticity and Loss Sharpness", "abstract": "  In recent years, continual learning, a prediction setting in which the\nproblem environment may evolve over time, has become an increasingly popular\nresearch field due to the framework's gearing towards complex, non-stationary\nobjectives. Learning such objectives requires plasticity, or the ability of a\nneural network to adapt its predictions to a different task. Recent findings\nindicate that plasticity loss on new tasks is highly related to loss landscape\nsharpness in non-stationary RL frameworks. We explore the usage of sharpness\nregularization techniques, which seek out smooth minima and have been touted\nfor their generalization capabilities in vanilla prediction settings, in\nefforts to combat plasticity loss. Our findings indicate that such techniques\nhave no significant effect on reducing plasticity loss.\n", "date": "2024-09-25", "categories": []}
{"id": "2409.16391", "title": "Patch-Based Contrastive Learning and Memory Consolidation for Online\n  Unsupervised Continual Learning", "abstract": "  We focus on a relatively unexplored learning paradigm known as {\\em Online\nUnsupervised Continual Learning} (O-UCL), where an agent receives a\nnon-stationary, unlabeled data stream and progressively learns to identify an\nincreasing number of classes. This paradigm is designed to model real-world\napplications where encountering novelty is the norm, such as exploring a\nterrain with several unknown and time-varying entities. Unlike prior work in\nunsupervised, continual, or online learning, O-UCL combines all three areas\ninto a single challenging and realistic learning paradigm. In this setting,\nagents are frequently evaluated and must aim to maintain the best possible\nrepresentation at any point of the data stream, rather than at the end of\npre-specified offline tasks. The proposed approach, called \\textbf{P}atch-based\n\\textbf{C}ontrastive learning and \\textbf{M}emory \\textbf{C}onsolidation\n(PCMC), builds a compositional understanding of data by identifying and\nclustering patch-level features. Embeddings for these patch-level features are\nextracted with an encoder trained via patch-based contrastive learning. PCMC\nincorporates new data into its distribution while avoiding catastrophic\nforgetting, and it consolidates memory examples during ``sleep\" periods. We\nevaluate PCMC's performance on streams created from the ImageNet and Places365\ndatasets. Additionally, we explore various versions of the PCMC algorithm and\ncompare its performance against several existing methods and simple baselines.\n", "date": "2024-09-24", "categories": []}
{"id": "2406.17935", "title": "Sequential Editing for Lifelong Training of Speech Recognition Models", "abstract": "  Automatic Speech Recognition (ASR) traditionally assumes known domains, but\nadding data from a new domain raises concerns about computational\ninefficiencies linked to retraining models on both existing and new domains.\nFine-tuning solely on new domain risks Catastrophic Forgetting (CF). To address\nthis, Lifelong Learning (LLL) algorithms have been proposed for ASR. Prior\nresearch has explored techniques such as Elastic Weight Consolidation,\nKnowledge Distillation, and Replay, all of which necessitate either additional\nparameters or access to prior domain data. We propose Sequential Model Editing\nas a novel method to continually learn new domains in ASR systems. Different\nthan previous methods, our approach does not necessitate access to prior\ndatasets or the introduction of extra parameters. Our study demonstrates up to\n15% Word Error Rate Reduction (WERR) over fine-tuning baseline, and superior\nefficiency over other LLL techniques on CommonVoice English multi-accent\ndataset.\n", "date": "2024-06-25", "categories": []}
{"id": "2409.15651", "title": "SurgIRL: Towards Life-Long Learning for Surgical Automation by\n  Incremental Reinforcement Learning", "abstract": "  Surgical automation holds immense potential to improve the outcome and\naccessibility of surgery. Recent studies use reinforcement learning to learn\npolicies that automate different surgical tasks. However, these policies are\ndeveloped independently and are limited in their reusability when the task\nchanges, making it more time-consuming when robots learn to solve multiple\ntasks. Inspired by how human surgeons build their expertise, we train surgical\nautomation policies through Surgical Incremental Reinforcement Learning\n(SurgIRL). SurgIRL aims to (1) acquire new skills by referring to external\npolicies (knowledge) and (2) accumulate and reuse these skills to solve\nmultiple unseen tasks incrementally (incremental learning). Our SurgIRL\nframework includes three major components. We first define an expandable\nknowledge set containing heterogeneous policies that can be helpful for\nsurgical tasks. Then, we propose Knowledge Inclusive Attention Network with\nmAximum Coverage Exploration (KIAN-ACE), which improves learning efficiency by\nmaximizing the coverage of the knowledge set during the exploration process.\nFinally, we develop incremental learning pipelines based on KIAN-ACE to\naccumulate and reuse learned knowledge and solve multiple surgical tasks\nsequentially. Our simulation experiments show that KIAN-ACE efficiently learns\nto automate ten surgical tasks separately or incrementally. We also evaluate\nour learned policies on the da Vinci Research Kit (dVRK) and demonstrate\nsuccessful sim-to-real transfers.\n", "date": "2024-09-24", "categories": []}
{"id": "2306.09970", "title": "Continual Adaptation of Vision Transformers for Federated Learning", "abstract": "  In this paper, we focus on the important yet understudied problem of\nContinual Federated Learning (CFL), where a server communicates with a set of\nclients to incrementally learn new concepts over time without sharing or\nstoring any data. The complexity of this problem is compounded by challenges\nfrom both the Continual and Federated Learning perspectives. Specifically,\nmodels trained in a CFL setup suffer from catastrophic forgetting which is\nexacerbated by data heterogeneity across clients. Existing attempts at this\nproblem tend to impose large overheads on clients and communication channels or\nrequire access to stored data which renders them unsuitable for real-world use\ndue to privacy. In this paper, we attempt to tackle forgetting and\nheterogeneity while minimizing overhead costs and without requiring access to\nany stored data. We study this problem in the context of Vision Transformers\nand explore parameter-efficient approaches to adapt to dynamic distributions\nwhile minimizing forgetting. We achieve this by leveraging a prompting based\napproach (such that only prompts and classifier heads have to be communicated)\nand proposing a novel and lightweight generation and distillation scheme to\nconsolidate client models at the server. We formulate this problem for image\nclassification and establish strong baselines for comparison, conduct\nexperiments on CIFAR-100 as well as challenging, large-scale datasets like\nImageNet-R and DomainNet. Our approach outperforms both existing methods and\nour own baselines by as much as 7% while significantly reducing communication\nand client-level computation costs. Code available at\nhttps://github.com/shaunak27/hepco-fed.\n", "date": "2023-06-16", "categories": []}
{"id": "2311.15161", "title": "Hessian Aware Low-Rank Perturbation for Order-Robust Continual Learning", "abstract": "  Continual learning aims to learn a series of tasks sequentially without\nforgetting the knowledge acquired from the previous ones. In this work, we\npropose the Hessian Aware Low-Rank Perturbation algorithm for continual\nlearning. By modeling the parameter transitions along the sequential tasks with\nthe weight matrix transformation, we propose to apply the low-rank\napproximation on the task-adaptive parameters in each layer of the neural\nnetworks. Specifically, we theoretically demonstrate the quantitative\nrelationship between the Hessian and the proposed low-rank approximation. The\napproximation ranks are then globally determined according to the marginal\nincrement of the empirical loss estimated by the layer-specific gradient and\nlow-rank approximation error. Furthermore, we control the model capacity by\npruning less important parameters to diminish the parameter growth. We conduct\nextensive experiments on various benchmarks, including a dataset with\nlarge-scale tasks, and compare our method against some recent state-of-the-art\nmethods to demonstrate the effectiveness and scalability of our proposed\nmethod. Empirical results show that our method performs better on different\nbenchmarks, especially in achieving task order robustness and handling the\nforgetting issue. The source code is at https://github.com/lijiaqi/HALRP.\n", "date": "2023-11-26", "categories": []}
{"id": "2312.08917", "title": "An Incremental Unified Framework for Small Defect Inspection", "abstract": "  Artificial Intelligence (AI)-driven defect inspection is pivotal in\nindustrial manufacturing. Yet, many methods, tailored to specific pipelines,\ngrapple with diverse product portfolios and evolving processes. Addressing\nthis, we present the Incremental Unified Framework (IUF), which can reduce the\nfeature conflict problem when continuously integrating new objects in the\npipeline, making it advantageous in object-incremental learning scenarios.\nEmploying a state-of-the-art transformer, we introduce Object-Aware\nSelf-Attention (OASA) to delineate distinct semantic boundaries. Semantic\nCompression Loss (SCL) is integrated to optimize non-primary semantic space,\nenhancing network adaptability for novel objects. Additionally, we prioritize\nretaining the features of established objects during weight updates.\nDemonstrating prowess in both image and pixel-level defect inspection, our\napproach achieves state-of-the-art performance, proving indispensable for\ndynamic and scalable industrial inspections. Our code will be released at\nhttps://github.com/jqtangust/IUF.\n", "date": "2023-12-14", "categories": []}
{"id": "2409.13997", "title": "Drift to Remember", "abstract": "  Lifelong learning in artificial intelligence (AI) aims to mimic the\nbiological brain's ability to continuously learn and retain knowledge, yet it\nfaces challenges such as catastrophic forgetting. Recent neuroscience research\nsuggests that neural activity in biological systems undergoes representational\ndrift, where neural responses evolve over time, even with consistent inputs and\ntasks. We hypothesize that representational drift can alleviate catastrophic\nforgetting in AI during new task acquisition. To test this, we introduce\nDriftNet, a network designed to constantly explore various local minima in the\nloss landscape while dynamically retrieving relevant tasks. This approach\nensures efficient integration of new information and preserves existing\nknowledge. Experimental studies in image classification and natural language\nprocessing demonstrate that DriftNet outperforms existing models in lifelong\nlearning. Importantly, DriftNet is scalable in handling a sequence of tasks\nsuch as sentiment analysis and question answering using large language models\n(LLMs) with billions of parameters on a single Nvidia A100 GPU. DriftNet\nefficiently updates LLMs using only new data, avoiding the need for full\ndataset retraining. Tested on GPT-2 and RoBERTa, DriftNet is a robust,\ncost-effective solution for lifelong learning in LLMs. This study not only\nadvances AI systems to emulate biological learning, but also provides insights\ninto the adaptive mechanisms of biological neural systems, deepening our\nunderstanding of lifelong learning in nature.\n", "date": "2024-09-21", "categories": []}
{"id": "2409.13275", "title": "Adaptive Margin Global Classifier for Exemplar-Free Class-Incremental\n  Learning", "abstract": "  Exemplar-free class-incremental learning (EFCIL) presents a significant\nchallenge as the old class samples are absent for new task learning. Due to the\nsevere imbalance between old and new class samples, the learned classifiers can\nbe easily biased toward the new ones. Moreover, continually updating the\nfeature extractor under EFCIL can compromise the discriminative power of old\nclass features, e.g., leading to less compact and more overlapping\ndistributions across classes. Existing methods mainly focus on handling biased\nclassifier learning. In this work, both cases are considered using the proposed\nmethod. Specifically, we first introduce a Distribution-Based Global Classifier\n(DBGC) to avoid bias factors in existing methods, such as data imbalance and\nsampling. More importantly, the compromised distributions of old classes are\nsimulated via a simple operation, variance enlarging (VE). Incorporating VE\nbased on DBGC results in a novel classification loss for EFCIL. This loss is\nproven equivalent to an Adaptive Margin Softmax Cross Entropy (AMarX). The\nproposed method is thus called Adaptive Margin Global Classifier (AMGC). AMGC\nis simple yet effective. Extensive experiments show that AMGC achieves superior\nimage classification results on its own under a challenging EFCIL setting.\nDetailed analysis is also provided for further demonstration.\n", "date": "2024-09-20", "categories": []}
{"id": "2310.05566", "title": "Aggregated f-average Neural Network applied to Few-Shot Class\n  Incremental Learning", "abstract": "  Ensemble learning leverages multiple models (i.e., weak learners) on a common\nmachine learning task to enhance prediction performance. Basic ensembling\napproaches average the weak learners outputs, while more sophisticated ones\nstack a machine learning model in between the weak learners outputs and the\nfinal prediction. This work fuses both aforementioned frameworks. We introduce\nan aggregated f-average (AFA) shallow neural network which models and combines\ndifferent types of averages to perform an optimal aggregation of the weak\nlearners predictions. We emphasise its interpretable architecture and simple\ntraining strategy, and illustrate its good performance on the problem of\nfew-shot class incremental learning.\n", "date": "2023-10-09", "categories": []}
{"id": "2402.16933", "title": "Incremental Concept Formation over Visual Images Without Catastrophic\n  Forgetting", "abstract": "  Deep neural networks have excelled in machine learning, particularly in\nvision tasks, however, they often suffer from catastrophic forgetting when\nlearning new tasks sequentially. In this work, we introduce Cobweb4V, an\nalternative to traditional neural network approaches. Cobweb4V is a novel\nvisual classification method that builds on Cobweb, a human like learning\nsystem that is inspired by the way humans incrementally learn new concepts over\ntime. In this research, we conduct a comprehensive evaluation, showcasing\nCobweb4Vs proficiency in learning visual concepts, requiring less data to\nachieve effective learning outcomes compared to traditional methods,\nmaintaining stable performance over time, and achieving commendable asymptotic\nbehavior, without catastrophic forgetting effects. These characteristics align\nwith learning strategies in human cognition, positioning Cobweb4V as a\npromising alternative to neural network approaches.\n", "date": "2024-02-26", "categories": []}
{"id": "2409.11657", "title": "Few-Shot Class-Incremental Learning with Non-IID Decentralized Data", "abstract": "  Few-shot class-incremental learning is crucial for developing scalable and\nadaptive intelligent systems, as it enables models to acquire new classes with\nminimal annotated data while safeguarding the previously accumulated knowledge.\nNonetheless, existing methods deal with continuous data streams in a\ncentralized manner, limiting their applicability in scenarios that prioritize\ndata privacy and security. To this end, this paper introduces federated\nfew-shot class-incremental learning, a decentralized machine learning paradigm\ntailored to progressively learn new classes from scarce data distributed across\nmultiple clients. In this learning paradigm, clients locally update their\nmodels with new classes while preserving data privacy, and then transmit the\nmodel updates to a central server where they are aggregated globally. However,\nthis paradigm faces several issues, such as difficulties in few-shot learning,\ncatastrophic forgetting, and data heterogeneity. To address these challenges,\nwe present a synthetic data-driven framework that leverages replay buffer data\nto maintain existing knowledge and facilitate the acquisition of new knowledge.\nWithin this framework, a noise-aware generative replay module is developed to\nfine-tune local models with a balance of new and replay data, while generating\nsynthetic data of new classes to further expand the replay buffer for future\ntasks. Furthermore, a class-specific weighted aggregation strategy is designed\nto tackle data heterogeneity by adaptively aggregating class-specific\nparameters based on local models performance on synthetic data. This enables\neffective global model optimization without direct access to client data.\nComprehensive experiments across three widely-used datasets underscore the\neffectiveness and preeminence of the introduced framework.\n", "date": "2024-09-18", "categories": []}
{"id": "2409.11770", "title": "Knowledge Adaptation Network for Few-Shot Class-Incremental Learning", "abstract": "  Few-shot class-incremental learning (FSCIL) aims to incrementally recognize\nnew classes using a few samples while maintaining the performance on previously\nlearned classes. One of the effective methods to solve this challenge is to\nconstruct prototypical evolution classifiers. Despite the advancement achieved\nby most existing methods, the classifier weights are simply initialized using\nmean features. Because representations for new classes are weak and biased, we\nargue such a strategy is suboptimal. In this paper, we tackle this issue from\ntwo aspects. Firstly, thanks to the development of foundation models, we employ\na foundation model, the CLIP, as the network pedestal to provide a general\nrepresentation for each class. Secondly, to generate a more reliable and\ncomprehensive instance representation, we propose a Knowledge Adapter (KA)\nmodule that summarizes the data-specific knowledge from training data and fuses\nit into the general representation. Additionally, to tune the knowledge learned\nfrom the base classes to the upcoming classes, we propose a mechanism of\nIncremental Pseudo Episode Learning (IPEL) by simulating the actual FSCIL.\nTaken together, our proposed method, dubbed as Knowledge Adaptation Network\n(KANet), achieves competitive performance on a wide range of datasets,\nincluding CIFAR100, CUB200, and ImageNet-R.\n", "date": "2024-09-18", "categories": []}
{"id": "2306.01904", "title": "Overcoming the Stability Gap in Continual Learning", "abstract": "  Pre-trained deep neural networks (DNNs) are being widely deployed by industry\nfor making business decisions and to serve users; however, a major problem is\nmodel decay, where the DNN's predictions become more erroneous over time,\nresulting in revenue loss or unhappy users. To mitigate model decay, DNNs are\nretrained from scratch using old and new data. This is computationally\nexpensive, so retraining happens only once performance significantly decreases.\nHere, we study how continual learning (CL) could potentially overcome model\ndecay in large pre-trained DNNs and greatly reduce computational costs for\nkeeping DNNs up-to-date. We identify the \"stability gap\" as a major obstacle in\nour setting. The stability gap refers to a phenomenon where learning new data\ncauses large drops in performance for past tasks before CL mitigation methods\neventually compensate for this drop. We test two hypotheses to investigate the\nfactors influencing the stability gap and identify a method that vastly reduces\nthis gap. In large-scale experiments for both easy and hard CL distributions\n(e.g., class incremental learning), we demonstrate that our method reduces the\nstability gap and greatly increases computational efficiency. Our work aligns\nCL with the goals of the production setting, where CL is needed for many\napplications.\n", "date": "2023-06-02", "categories": []}
{"id": "2409.10956", "title": "Versatile Incremental Learning: Towards Class and Domain-Agnostic\n  Incremental Learning", "abstract": "  Incremental Learning (IL) aims to accumulate knowledge from sequential input\ntasks while overcoming catastrophic forgetting. Existing IL methods typically\nassume that an incoming task has only increments of classes or domains,\nreferred to as Class IL (CIL) or Domain IL (DIL), respectively. In this work,\nwe consider a more challenging and realistic but under-explored IL scenario,\nnamed Versatile Incremental Learning (VIL), in which a model has no prior of\nwhich of the classes or domains will increase in the next task. In the proposed\nVIL scenario, the model faces intra-class domain confusion and inter-domain\nclass confusion, which makes the model fail to accumulate new knowledge without\ninterference with learned knowledge. To address these issues, we propose a\nsimple yet effective IL framework, named Incremental Classifier with Adaptation\nShift cONtrol (ICON). Based on shifts of learnable modules, we design a novel\nregularization method called Cluster-based Adaptation Shift conTrol (CAST) to\ncontrol the model to avoid confusion with the previously learned knowledge and\nthereby accumulate the new knowledge more effectively. Moreover, we introduce\nan Incremental Classifier (IC) which expands its output nodes to address the\noverwriting issue from different domains corresponding to a single class while\nmaintaining the previous knowledge. We conducted extensive experiments on three\nbenchmarks, showcasing the effectiveness of our method across all the\nscenarios, particularly in cases where the next task can be randomly altered.\nOur implementation code is available at https://github.com/KHU-AGI/VIL.\n", "date": "2024-09-17", "categories": []}
{"id": "2409.11329", "title": "Reducing Catastrophic Forgetting in Online Class Incremental Learning\n  Using Self-Distillation", "abstract": "  In continual learning, there is a serious problem of catastrophic forgetting,\nin which previous knowledge is forgotten when a model learns new tasks. Various\nmethods have been proposed to solve this problem. Replay methods which replay\ndata from previous tasks in later training, have shown good accuracy. However,\nreplay methods have a generalizability problem from a limited memory buffer. In\nthis paper, we tried to solve this problem by acquiring transferable knowledge\nthrough self-distillation using highly generalizable output in shallow layer as\na teacher. Furthermore, when we deal with a large number of classes or\nchallenging data, there is a risk of learning not converging and not\nexperiencing overfitting. Therefore, we attempted to achieve more efficient and\nthorough learning by prioritizing the storage of easily misclassified samples\nthrough a new method of memory update. We confirmed that our proposed method\noutperformed conventional methods by experiments on CIFAR10, CIFAR100, and\nMiniimageNet datasets.\n", "date": "2024-09-17", "categories": []}
{"id": "2409.08518", "title": "Anytime Continual Learning for Open Vocabulary Classification", "abstract": "  We propose an approach for anytime continual learning (AnytimeCL) for open\nvocabulary image classification. The AnytimeCL problem aims to break away from\nbatch training and rigid models by requiring that a system can predict any set\nof labels at any time and efficiently update and improve when receiving one or\nmore training samples at any time. Despite the challenging goal, we achieve\nsubstantial improvements over recent methods. We propose a dynamic weighting\nbetween predictions of a partially fine-tuned model and a fixed open vocabulary\nmodel that enables continual improvement when training samples are available\nfor a subset of a task's labels. We also propose an attention-weighted PCA\ncompression of training features that reduces storage and computation with\nlittle impact to model accuracy. Our methods are validated with experiments\nthat test flexibility of learning and inference. Code is available at\nhttps://github.com/jessemelpolio/AnytimeCL.\n", "date": "2024-09-13", "categories": []}
{"id": "2309.00023", "title": "Continual Learning From a Stream of APIs", "abstract": "  Continual learning (CL) aims to learn new tasks without forgetting previous\ntasks. However, existing CL methods require a large amount of raw data, which\nis often unavailable due to copyright considerations and privacy risks.\nInstead, stakeholders usually release pre-trained machine learning models as a\nservice (MLaaS), which users can access via APIs. This paper considers two\npractical-yet-novel CL settings: data-efficient CL (DECL-APIs) and data-free CL\n(DFCL-APIs), which achieve CL from a stream of APIs with partial or no raw\ndata. Performing CL under these two new settings faces several challenges:\nunavailable full raw data, unknown model parameters, heterogeneous models of\narbitrary architecture and scale, and catastrophic forgetting of previous APIs.\nTo overcome these issues, we propose a novel data-free cooperative continual\ndistillation learning framework that distills knowledge from a stream of APIs\ninto a CL model by generating pseudo data, just by querying APIs. Specifically,\nour framework includes two cooperative generators and one CL model, forming\ntheir training as an adversarial game. We first use the CL model and the\ncurrent API as fixed discriminators to train generators via a derivative-free\nmethod. Generators adversarially generate hard and diverse synthetic data to\nmaximize the response gap between the CL model and the API. Next, we train the\nCL model by minimizing the gap between the responses of the CL model and the\nblack-box API on synthetic data, to transfer the API's knowledge to the CL\nmodel. Furthermore, we propose a new regularization term based on network\nsimilarity to prevent catastrophic forgetting of previous APIs.Our method\nperforms comparably to classic CL with full raw data on the MNIST and SVHN in\nthe DFCL-APIs setting. In the DECL-APIs setting, our method achieves 0.97x,\n0.75x and 0.69x performance of classic CL on CIFAR10, CIFAR100, and\nMiniImageNet.\n", "date": "2023-08-31", "categories": []}
{"id": "2409.07757", "title": "From Uncertainty to Clarity: Uncertainty-Guided Class-Incremental\n  Learning for Limited Biomedical Samples via Semantic Expansion", "abstract": "  In real-world clinical settings, data distributions evolve over time, with a\ncontinuous influx of new, limited disease cases. Therefore, class incremental\nlearning is of great significance, i.e., deep learning models are required to\nlearn new class knowledge while maintaining accurate recognition of previous\ndiseases. However, traditional deep neural networks often suffer from severe\nforgetting of prior knowledge when adapting to new data unless trained from\nscratch, which undesirably costs much time and computational burden.\nAdditionally, the sample sizes for different diseases can be highly imbalanced,\nwith newly emerging diseases typically having much fewer instances,\nconsequently causing the classification bias. To tackle these challenges, we\nare the first to propose a class-incremental learning method under limited\nsamples in the biomedical field. First, we propose a novel cumulative entropy\nprediction module to measure the uncertainty of the samples, of which the most\nuncertain samples are stored in a memory bank as exemplars for the model's\nlater review. Furthermore, we theoretically demonstrate its effectiveness in\nmeasuring uncertainty. Second, we developed a fine-grained semantic expansion\nmodule through various augmentations, leading to more compact distributions\nwithin the feature space and creating sufficient room for generalization to new\nclasses. Besides, a cosine classifier is utilized to mitigate classification\nbias caused by imbalanced datasets. Across four imbalanced data distributions\nover two datasets, our method achieves optimal performance, surpassing\nstate-of-the-art methods by as much as 53.54% in accuracy.\n", "date": "2024-09-12", "categories": []}
{"id": "2303.10147", "title": "CoDEPS: Online Continual Learning for Depth Estimation and Panoptic\n  Segmentation", "abstract": "  Operating a robot in the open world requires a high level of robustness with\nrespect to previously unseen environments. Optimally, the robot is able to\nadapt by itself to new conditions without human supervision, e.g.,\nautomatically adjusting its perception system to changing lighting conditions.\nIn this work, we address the task of continual learning for deep learning-based\nmonocular depth estimation and panoptic segmentation in new environments in an\nonline manner. We introduce CoDEPS to perform continual learning involving\nmultiple real-world domains while mitigating catastrophic forgetting by\nleveraging experience replay. In particular, we propose a novel domain-mixing\nstrategy to generate pseudo-labels to adapt panoptic segmentation. Furthermore,\nwe explicitly address the limited storage capacity of robotic systems by\nleveraging sampling strategies for constructing a fixed-size replay buffer\nbased on rare semantic class sampling and image diversity. We perform extensive\nevaluations of CoDEPS on various real-world datasets demonstrating that it\nsuccessfully adapts to unseen environments without sacrificing performance on\nprevious domains while achieving state-of-the-art results. The code of our work\nis publicly available at http://codeps.cs.uni-freiburg.de.\n", "date": "2023-03-17", "categories": []}
{"id": "2409.07446", "title": "Adaptive Adapter Routing for Long-Tailed Class-Incremental Learning", "abstract": "  In our ever-evolving world, new data exhibits a long-tailed distribution,\nsuch as e-commerce platform reviews. This necessitates continuous model\nlearning imbalanced data without forgetting, addressing the challenge of\nlong-tailed class-incremental learning (LTCIL). Existing methods often rely on\nretraining linear classifiers with former data, which is impractical in\nreal-world settings. In this paper, we harness the potent representation\ncapabilities of pre-trained models and introduce AdaPtive Adapter RouTing\n(APART) as an exemplar-free solution for LTCIL. To counteract forgetting, we\ntrain inserted adapters with frozen pre-trained weights for deeper adaptation\nand maintain a pool of adapters for selection during sequential model updates.\nAdditionally, we present an auxiliary adapter pool designed for effective\ngeneralization, especially on minority classes. Adaptive instance routing\nacross these pools captures crucial correlations, facilitating a comprehensive\nrepresentation of all classes. Consequently, APART tackles the imbalance\nproblem as well as catastrophic forgetting in a unified framework. Extensive\nbenchmark experiments validate the effectiveness of APART. Code is available\nat: https://github.com/vita-qzh/APART\n", "date": "2024-09-11", "categories": []}
{"id": "2409.06455", "title": "Continual Domain Incremental Learning for Privacy-aware Digital\n  Pathology", "abstract": "  In recent years, there has been remarkable progress in the field of digital\npathology, driven by the ability to model complex tissue patterns using\nadvanced deep-learning algorithms. However, the robustness of these models is\noften severely compromised in the presence of data shifts (e.g., different\nstains, organs, centers, etc.). Alternatively, continual learning (CL)\ntechniques aim to reduce the forgetting of past data when learning new data\nwith distributional shift conditions. Specifically, rehearsal-based CL\ntechniques, which store some past data in a buffer and then replay it with new\ndata, have proven effective in medical image analysis tasks. However, privacy\nconcerns arise as these approaches store past data, prompting the development\nof our novel Generative Latent Replay-based CL (GLRCL) approach. GLRCL captures\nthe previous distribution through Gaussian Mixture Models instead of storing\npast samples, which are then utilized to generate features and perform latent\nreplay with new data. We systematically evaluate our proposed framework under\ndifferent shift conditions in histopathology data, including stain and organ\nshift. Our approach significantly outperforms popular buffer-free CL approaches\nand performs similarly to rehearsal-based CL approaches that require large\nbuffers causing serious privacy violations.\n", "date": "2024-09-10", "categories": []}
{"id": "2403.14392", "title": "A Bag of Tricks for Few-Shot Class-Incremental Learning", "abstract": "  We present a bag of tricks framework for few-shot class-incremental learning\n(FSCIL), which is a challenging form of continual learning that involves\ncontinuous adaptation to new tasks with limited samples. FSCIL requires both\nstability and adaptability, i.e., preserving proficiency in previously learned\ntasks while learning new ones. Our proposed bag of tricks brings together six\nkey and highly influential techniques that improve stability, adaptability, and\noverall performance under a unified framework for FSCIL. We organize these\ntricks into three categories: stability tricks, adaptability tricks, and\ntraining tricks. Stability tricks aim to mitigate the forgetting of previously\nlearned classes by enhancing the separation between the embeddings of learned\nclasses and minimizing interference when learning new ones. On the other hand,\nadaptability tricks focus on the effective learning of new classes. Finally,\ntraining tricks improve the overall performance without compromising stability\nor adaptability. We perform extensive experiments on three benchmark datasets,\nCIFAR-100, CUB-200, and miniIMageNet, to evaluate the impact of our proposed\nframework. Our detailed analysis shows that our approach substantially improves\nboth stability and adaptability, establishing a new state-of-the-art by\noutperforming prior works in the area. We believe our method provides a go-to\nsolution and establishes a robust baseline for future research in this area.\n", "date": "2024-03-21", "categories": []}
{"id": "2409.05620", "title": "Joint Input and Output Coordination for Class-Incremental Learning", "abstract": "  Incremental learning is nontrivial due to severe catastrophic forgetting.\nAlthough storing a small amount of data on old tasks during incremental\nlearning is a feasible solution, current strategies still do not 1) adequately\naddress the class bias problem, and 2) alleviate the mutual interference\nbetween new and old tasks, and 3) consider the problem of class bias within\ntasks. This motivates us to propose a joint input and output coordination\n(JIOC) mechanism to address these issues. This mechanism assigns different\nweights to different categories of data according to the gradient of the output\nscore, and uses knowledge distillation (KD) to reduce the mutual interference\nbetween the outputs of old and new tasks. The proposed mechanism is general and\nflexible, and can be incorporated into different incremental learning\napproaches that use memory storage. Extensive experiments show that our\nmechanism can significantly improve their performance.\n", "date": "2024-09-09", "categories": []}
{"id": "2212.11192", "title": "Continual Learning Approaches for Anomaly Detection", "abstract": "  Anomaly Detection is a relevant problem that arises in numerous real-world\napplications, especially when dealing with images. However, there has been\nlittle research for this task in the Continual Learning setting. In this work,\nwe introduce a novel approach called SCALE (SCALing is Enough) to perform\nCompressed Replay in a framework for Anomaly Detection in Continual Learning\nsetting. The proposed technique scales and compresses the original images using\na Super Resolution model which, to the best of our knowledge, is studied for\nthe first time in the Continual Learning setting. SCALE can achieve a high\nlevel of compression while maintaining a high level of image reconstruction\nquality. In conjunction with other Anomaly Detection approaches, it can achieve\noptimal results. To validate the proposed approach, we use a real-world dataset\nof images with pixel-based anomalies, with the scope to provide a reliable\nbenchmark for Anomaly Detection in the context of Continual Learning, serving\nas a foundation for further advancements in the field.\n", "date": "2022-12-21", "categories": []}
{"id": "2409.03354", "title": "Few-Shot Continual Learning for Activity Recognition in Classroom\n  Surveillance Images", "abstract": "  The application of activity recognition in the \"AI + Education\" field is\ngaining increasing attention. However, current work mainly focuses on the\nrecognition of activities in manually captured videos and a limited number of\nactivity types, with little attention given to recognizing activities in\nsurveillance images from real classrooms. In real classroom settings, normal\nteaching activities such as reading, account for a large proportion of samples,\nwhile rare non-teaching activities such as eating, continue to appear. This\nrequires a model that can learn non-teaching activities from few samples\nwithout forgetting the normal teaching activities, which necessitates fewshot\ncontinual learning (FSCL) capability. To address this gap, we constructed a\ncontinual learning dataset focused on classroom surveillance image activity\nrecognition called ARIC (Activity Recognition in Classroom). The dataset has\nadvantages such as multiple perspectives, a wide variety of activities, and\nreal-world scenarios, but it also presents challenges like similar activities\nand imbalanced sample distribution. To overcome these challenges, we designed a\nfew-shot continual learning method that combines supervised contrastive\nlearning (SCL) and an adaptive covariance classifier (ACC). During the base\nphase, we proposed a SCL approach based on feature augmentation to enhance the\nmodel's generalization ability. In the incremental phase, we employed an ACC to\nmore accurately describe the distribution of new classes. Experimental results\ndemonstrate that our method outperforms other existing methods on the ARIC\ndataset.\n", "date": "2024-09-05", "categories": []}
{"id": "2311.11293", "title": "From Categories to Classifiers: Name-Only Continual Learning by\n  Exploring the Web", "abstract": "  Continual Learning (CL) often relies on the availability of extensive\nannotated datasets, an assumption that is unrealistically time-consuming and\ncostly in practice. We explore a novel paradigm termed name-only continual\nlearning where time and cost constraints prohibit manual annotation. In this\nscenario, learners adapt to new category shifts using only category names\nwithout the luxury of annotated training data. Our proposed solution leverages\nthe expansive and ever-evolving internet to query and download uncurated\nwebly-supervised data for image classification. We investigate the reliability\nof our web data and find them comparable, and in some cases superior, to\nmanually annotated datasets. Additionally, we show that by harnessing the web,\nwe can create support sets that surpass state-of-the-art name-only\nclassification that create support sets using generative models or image\nretrieval from LAION-5B, achieving up to 25% boost in accuracy. When applied\nacross varied continual learning contexts, our method consistently exhibits a\nsmall performance gap in comparison to models trained on manually annotated\ndatasets. We present EvoTrends, a class-incremental dataset made from the web\nto capture real-world trends, created in just minutes. Overall, this paper\nunderscores the potential of using uncurated webly-supervised data to mitigate\nthe challenges associated with manual data labeling in continual learning.\n", "date": "2023-11-19", "categories": []}
{"id": "2403.08763", "title": "Simple and Scalable Strategies to Continually Pre-train Large Language\n  Models", "abstract": "  Large language models (LLMs) are routinely pre-trained on billions of tokens,\nonly to start the process over again once new data becomes available. A much\nmore efficient solution is to continually pre-train these models, saving\nsignificant compute compared to re-training. However, the distribution shift\ninduced by new data typically results in degraded performance on previous data\nor poor adaptation to the new data. In this work, we show that a simple and\nscalable combination of learning rate (LR) re-warming, LR re-decaying, and\nreplay of previous data is sufficient to match the performance of fully\nre-training from scratch on all available data, as measured by the final loss\nand the average score on several language model (LM) evaluation benchmarks.\nSpecifically, we show this for a weak but realistic distribution shift between\ntwo commonly used LLM pre-training datasets (English$\\rightarrow$English) and a\nstronger distribution shift (English$\\rightarrow$German) at the $405$M\nparameter model scale with large dataset sizes (hundreds of billions of\ntokens). Selecting the weak but realistic shift for larger-scale experiments,\nwe also find that our continual learning strategies match the re-training\nbaseline for a 10B parameter LLM. Our results demonstrate that LLMs can be\nsuccessfully updated via simple and scalable continual learning strategies,\nmatching the re-training baseline using only a fraction of the compute.\nFinally, inspired by previous work, we propose alternatives to the cosine\nlearning rate schedule that help circumvent forgetting induced by LR re-warming\nand that are not bound to a fixed token budget.\n", "date": "2024-03-13", "categories": []}
{"id": "2409.01128", "title": "Diffusion-Driven Data Replay: A Novel Approach to Combat Forgetting in\n  Federated Class Continual Learning", "abstract": "  Federated Class Continual Learning (FCCL) merges the challenges of\ndistributed client learning with the need for seamless adaptation to new\nclasses without forgetting old ones. The key challenge in FCCL is catastrophic\nforgetting, an issue that has been explored to some extent in Continual\nLearning (CL). However, due to privacy preservation requirements, some\nconventional methods, such as experience replay, are not directly applicable to\nFCCL. Existing FCCL methods mitigate forgetting by generating historical data\nthrough federated training of GANs or data-free knowledge distillation.\nHowever, these approaches often suffer from unstable training of generators or\nlow-quality generated data, limiting their guidance for the model. To address\nthis challenge, we propose a novel method of data replay based on diffusion\nmodels. Instead of training a diffusion model, we employ a pre-trained\nconditional diffusion model to reverse-engineer each class, searching the\ncorresponding input conditions for each class within the model's input space,\nsignificantly reducing computational resources and time consumption while\nensuring effective generation. Furthermore, we enhance the classifier's domain\ngeneralization ability on generated and real data through contrastive learning,\nindirectly improving the representational capability of generated data for real\ndata. Comprehensive experiments demonstrate that our method significantly\noutperforms existing baselines. Code is available at\nhttps://github.com/jinglin-liang/DDDR.\n", "date": "2024-09-02", "categories": []}
{"id": "2409.01585", "title": "Buffer-based Gradient Projection for Continual Federated Learning", "abstract": "  Continual Federated Learning (CFL) is essential for enabling real-world\napplications where multiple decentralized clients adaptively learn from\ncontinuous data streams. A significant challenge in CFL is mitigating\ncatastrophic forgetting, where models lose previously acquired knowledge when\nlearning new information. Existing approaches often face difficulties due to\nthe constraints of device storage capacities and the heterogeneous nature of\ndata distributions among clients. While some CFL algorithms have addressed\nthese challenges, they frequently rely on unrealistic assumptions about the\navailability of task boundaries (i.e., knowing when new tasks begin). To\naddress these limitations, we introduce Fed-A-GEM, a federated adaptation of\nthe A-GEM method (Chaudhry et al., 2019), which employs a buffer-based gradient\nprojection approach. Fed-A-GEM alleviates catastrophic forgetting by leveraging\nlocal buffer samples and aggregated buffer gradients, thus preserving knowledge\nacross multiple clients. Our method is combined with existing CFL techniques,\nenhancing their performance in the CFL context. Our experiments on standard\nbenchmarks show consistent performance improvements across diverse scenarios.\nFor example, in a task-incremental learning scenario using the CIFAR-100\ndataset, our method can increase the accuracy by up to 27%. Our code is\navailable at https://github.com/shenghongdai/Fed-A-GEM.\n", "date": "2024-09-03", "categories": []}
{"id": "2206.03934", "title": "A Study of Continual Learning Methods for Q-Learning", "abstract": "  We present an empirical study on the use of continual learning (CL) methods\nin a reinforcement learning (RL) scenario, which, to the best of our knowledge,\nhas not been described before. CL is a very active recent research topic\nconcerned with machine learning under non-stationary data distributions.\nAlthough this naturally applies to RL, the use of dedicated CL methods is still\nuncommon. This may be due to the fact that CL methods often assume a\ndecomposition of CL problems into disjoint sub-tasks of stationary\ndistribution, that the onset of these sub-tasks is known, and that sub-tasks\nare non-contradictory. In this study, we perform an empirical comparison of\nselected CL methods in a RL problem where a physically simulated robot must\nfollow a racetrack by vision. In order to make CL methods applicable, we\nrestrict the RL setting and introduce non-conflicting subtasks of known onset,\nwhich are however not disjoint and whose distribution, from the learner's point\nof view, is still non-stationary. Our results show that dedicated CL methods\ncan significantly improve learning when compared to the baseline technique of\n\"experience replay\".\n", "date": "2022-06-08", "categories": []}
{"id": "2409.01872", "title": "Latent Distillation for Continual Object Detection at the Edge", "abstract": "  While numerous methods achieving remarkable performance exist in the Object\nDetection literature, addressing data distribution shifts remains challenging.\nContinual Learning (CL) offers solutions to this issue, enabling models to\nadapt to new data while maintaining performance on previous data. This is\nparticularly pertinent for edge devices, common in dynamic environments like\nautomotive and robotics. In this work, we address the memory and computation\nconstraints of edge devices in the Continual Learning for Object Detection\n(CLOD) scenario. Specifically, (i) we investigate the suitability of an\nopen-source, lightweight, and fast detector, namely NanoDet, for CLOD on edge\ndevices, improving upon larger architectures used in the literature. Moreover,\n(ii) we propose a novel CL method, called Latent Distillation~(LD), that\nreduces the number of operations and the memory required by state-of-the-art CL\napproaches without significantly compromising detection performance. Our\napproach is validated using the well-known VOC and COCO benchmarks, reducing\nthe distillation parameter overhead by 74\\% and the Floating Points\nOperations~(FLOPs) by 56\\% per model update compared to other distillation\nmethods.\n", "date": "2024-09-03", "categories": []}
{"id": "2408.16517", "title": "Adaptive Variational Continual Learning via Task-Heuristic Modelling", "abstract": "  Variational continual learning (VCL) is a turn-key learning algorithm that\nhas state-of-the-art performance among the best continual learning models. In\nour work, we explore an extension of the generalized variational continual\nlearning (GVCL) model, named AutoVCL, which combines task heuristics for\ninformed learning and model optimization. We demonstrate that our model\noutperforms the standard GVCL with fixed hyperparameters, benefiting from the\nautomatic adjustment of the hyperparameter based on the difficulty and\nsimilarity of the incoming task compared to the previous tasks.\n", "date": "2024-08-29", "categories": []}
{"id": "2401.10393", "title": "Natural Mitigation of Catastrophic Interference: Continual Learning in\n  Power-Law Learning Environments", "abstract": "  Neural networks often suffer from catastrophic interference (CI): performance\non previously learned tasks drops off significantly when learning a new task.\nThis contrasts strongly with humans, who can continually learn new tasks\nwithout appreciably forgetting previous tasks. Prior work has explored various\ntechniques for mitigating CI and promoting continual learning such as\nregularization, rehearsal, generative replay, and context-specific components.\nThis paper takes a different approach, one guided by cognitive science research\nshowing that in naturalistic environments, the probability of encountering a\ntask decreases as a power-law of the time since it was last performed. We argue\nthat techniques for mitigating CI should be compared against the intrinsic\nmitigation in simulated naturalistic learning environments. Thus, we evaluate\nthe extent of the natural mitigation of CI when training models in power-law\nenvironments, similar to those humans face. Our results show that natural\nrehearsal environments are better at mitigating CI than existing methods,\ncalling for the need for better evaluation processes. The benefits of this\nenvironment include simplicity, rehearsal that is agnostic to both tasks and\nmodels, and the lack of a need for extra neural circuitry. In addition, we\nexplore popular mitigation techniques in power-law environments to create new\nbaselines for continual learning research.\n", "date": "2024-01-18", "categories": []}
{"id": "2407.21687", "title": "Dynamic Object Queries for Transformer-based Incremental Object\n  Detection", "abstract": "  Incremental object detection (IOD) aims to sequentially learn new classes,\nwhile maintaining the capability to locate and identify old ones. As the\ntraining data arrives with annotations only with new classes, IOD suffers from\ncatastrophic forgetting. Prior methodologies mainly tackle the forgetting issue\nthrough knowledge distillation and exemplar replay, ignoring the conflict\nbetween limited model capacity and increasing knowledge. In this paper, we\nexplore \\textit{dynamic object queries} for incremental object detection built\non Transformer architecture. We propose the \\textbf{Dy}namic object\n\\textbf{Q}uery-based \\textbf{DE}tection \\textbf{TR}ansformer (DyQ-DETR), which\nincrementally expands the model representation ability to achieve\nstability-plasticity tradeoff. First, a new set of learnable object queries are\nfed into the decoder to represent new classes. These new object queries are\naggregated with those from previous phases to adapt both old and new knowledge\nwell. Second, we propose the isolated bipartite matching for object queries in\ndifferent phases, based on disentangled self-attention. The interaction among\nthe object queries at different phases is eliminated to reduce inter-class\nconfusion. Thanks to the separate supervision and computation over object\nqueries, we further present the risk-balanced partial calibration for effective\nexemplar replay. Extensive experiments demonstrate that DyQ-DETR significantly\nsurpasses the state-of-the-art methods, with limited parameter overhead. Code\nwill be made publicly available.\n", "date": "2024-07-31", "categories": []}
{"id": "2408.14572", "title": "CURLoRA: Stable LLM Continual Fine-Tuning and Catastrophic Forgetting\n  Mitigation", "abstract": "  This paper introduces CURLoRA, a novel approach to fine-tuning large language\nmodels (LLMs) that leverages CUR matrix decomposition in the context of\nLow-Rank Adaptation (LoRA). Our method addresses two critical challenges in LLM\nfine-tuning: mitigating catastrophic forgetting during continual learning and\nreducing the number of trainable parameters. We propose a unique modification\nto the CUR decomposition process, utilizing inverted probabilities for column\nand row selection which acts as an implicit regularization, and initializing\nthe $U$ matrix as a zero matrix, and only fine-tuning it. We demonstrate\nthrough experiments on multiple datasets that CURLoRA outperforms standard LoRA\nin mitigating catastrophic forgetting. It maintains model stability and\nperformance across tasks while significantly reducing the number of trainable\nparameters. Our results show that CURLoRA achieves very good and stable task\naccuracy while maintaining base model's perplexity scores fixed compared to\nLoRA upon continual fine-tuning, particularly in scenarios with limited data.\n", "date": "2024-08-26", "categories": []}
{"id": "2408.14976", "title": "Prior-free Balanced Replay: Uncertainty-guided Reservoir Sampling for\n  Long-Tailed Continual Learning", "abstract": "  Even in the era of large models, one of the well-known issues in continual\nlearning (CL) is catastrophic forgetting, which is significantly challenging\nwhen the continual data stream exhibits a long-tailed distribution, termed as\nLong-Tailed Continual Learning (LTCL). Existing LTCL solutions generally\nrequire the label distribution of the data stream to achieve re-balance\ntraining. However, obtaining such prior information is often infeasible in real\nscenarios since the model should learn without pre-identifying the majority and\nminority classes. To this end, we propose a novel Prior-free Balanced Replay\n(PBR) framework to learn from long-tailed data stream with less forgetting.\nConcretely, motivated by our experimental finding that the minority classes are\nmore likely to be forgotten due to the higher uncertainty, we newly design an\nuncertainty-guided reservoir sampling strategy to prioritize rehearsing\nminority data without using any prior information, which is based on the mutual\ndependence between the model and samples. Additionally, we incorporate two\nprior-free components to further reduce the forgetting issue: (1) Boundary\nconstraint is to preserve uncertain boundary supporting samples for continually\nre-estimating task boundaries. (2) Prototype constraint is to maintain the\nconsistency of learned class prototypes along with training. Our approach is\nevaluated on three standard long-tailed benchmarks, demonstrating superior\nperformance to existing CL methods and previous SOTA LTCL approach in both\ntask- and class-incremental learning settings, as well as ordered- and\nshuffled-LTCL settings.\n", "date": "2024-08-27", "categories": []}
{"id": "2408.13492", "title": "Online Continuous Generalized Category Discovery", "abstract": "  With the advancement of deep neural networks in computer vision, artificial\nintelligence (AI) is widely employed in real-world applications. However, AI\nstill faces limitations in mimicking high-level human capabilities, such as\nnovel category discovery, for practical use. While some methods utilizing\noffline continual learning have been proposed for novel category discovery,\nthey neglect the continuity of data streams in real-world settings. In this\nwork, we introduce Online Continuous Generalized Category Discovery (OCGCD),\nwhich considers the dynamic nature of data streams where data can be created\nand deleted in real time. Additionally, we propose a novel method, DEAN,\nDiscovery via Energy guidance and feature AugmentatioN, which can discover\nnovel categories in an online manner through energy-guided discovery and\nfacilitate discriminative learning via energy-based contrastive loss.\nFurthermore, DEAN effectively pseudo-labels unlabeled data through\nvariance-based feature augmentation. Experimental results demonstrate that our\nproposed DEAN achieves outstanding performance in proposed OCGCD scenario.\n", "date": "2024-08-24", "categories": []}
{"id": "2408.13991", "title": "Dual-CBA: Improving Online Continual Learning via Dual Continual Bias\n  Adaptors from a Bi-level Optimization Perspective", "abstract": "  In online continual learning (CL), models trained on changing distributions\neasily forget previously learned knowledge and bias toward newly received\ntasks. To address this issue, we present Continual Bias Adaptor (CBA), a\nbi-level framework that augments the classification network to adapt to\ncatastrophic distribution shifts during training, enabling the network to\nachieve a stable consolidation of all seen tasks. However, the CBA module\nadjusts distribution shifts in a class-specific manner, exacerbating the\nstability gap issue and, to some extent, fails to meet the need for continual\ntesting in online CL. To mitigate this challenge, we further propose a novel\nclass-agnostic CBA module that separately aggregates the posterior\nprobabilities of classes from new and old tasks, and applies a stable\nadjustment to the resulting posterior probabilities. We combine the two kinds\nof CBA modules into a unified Dual-CBA module, which thus is capable of\nadapting to catastrophic distribution shifts and simultaneously meets the\nreal-time testing requirements of online CL. Besides, we propose Incremental\nBatch Normalization (IBN), a tailored BN module to re-estimate its population\nstatistics for alleviating the feature bias arising from the inner loop\noptimization problem of our bi-level framework. To validate the effectiveness\nof the proposed method, we theoretically provide some insights into how it\nmitigates catastrophic distribution shifts, and empirically demonstrate its\nsuperiority through extensive experiments based on four rehearsal-based\nbaselines and three public continual learning benchmarks.\n", "date": "2024-08-26", "categories": []}
{"id": "2408.14284", "title": "May the Forgetting Be with You: Alternate Replay for Learning with Noisy\n  Labels", "abstract": "  Forgetting presents a significant challenge during incremental training,\nmaking it particularly demanding for contemporary AI systems to assimilate new\nknowledge in streaming data environments. To address this issue, most\napproaches in Continual Learning (CL) rely on the replay of a restricted buffer\nof past data. However, the presence of noise in real-world scenarios, where\nhuman annotation is constrained by time limitations or where data is\nautomatically gathered from the web, frequently renders these strategies\nvulnerable. In this study, we address the problem of CL under Noisy Labels\n(CLN) by introducing Alternate Experience Replay (AER), which takes advantage\nof forgetting to maintain a clear distinction between clean, complex, and noisy\nsamples in the memory buffer. The idea is that complex or mislabeled examples,\nwhich hardly fit the previously learned data distribution, are most likely to\nbe forgotten. To grasp the benefits of such a separation, we equip AER with\nAsymmetric Balanced Sampling (ABS): a new sample selection strategy that\nprioritizes purity on the current task while retaining relevant samples from\nthe past. Through extensive computational comparisons, we demonstrate the\neffectiveness of our approach in terms of both accuracy and purity of the\nobtained buffer, resulting in a remarkable average gain of 4.71% points in\naccuracy with respect to existing loss-based purification strategies. Code is\navailable at https://github.com/aimagelab/mammoth.\n", "date": "2024-08-26", "categories": []}
{"id": "2303.15975", "title": "Large-scale Pre-trained Models are Surprisingly Strong in Incremental\n  Novel Class Discovery", "abstract": "  Discovering novel concepts in unlabelled datasets and in a continuous manner\nis an important desideratum of lifelong learners. In the literature such\nproblems have been partially addressed under very restricted settings, where\nnovel classes are learned by jointly accessing a related labelled set (e.g.,\nNCD) or by leveraging only a supervisedly pre-trained model (e.g., class-iNCD).\nIn this work we challenge the status quo in class-iNCD and propose a learning\nparadigm where class discovery occurs continuously and truly unsupervisedly,\nwithout needing any related labelled set. In detail, we propose to exploit the\nricher priors from strong self-supervised pre-trained models (PTM). To this\nend, we propose simple baselines, composed of a frozen PTM backbone and a\nlearnable linear classifier, that are not only simple to implement but also\nresilient under longer learning scenarios. We conduct extensive empirical\nevaluation on a multitude of benchmarks and show the effectiveness of our\nproposed baselines when compared with sophisticated state-of-the-art methods.\nThe code is open source.\n", "date": "2023-03-28", "categories": []}
{"id": "2310.14277", "title": "A Survey on Continual Semantic Segmentation: Theory, Challenge, Method\n  and Application", "abstract": "  Continual learning, also known as incremental learning or life-long learning,\nstands at the forefront of deep learning and AI systems. It breaks through the\nobstacle of one-way training on close sets and enables continuous adaptive\nlearning on open-set conditions. In the recent decade, continual learning has\nbeen explored and applied in multiple fields especially in computer vision\ncovering classification, detection and segmentation tasks. Continual semantic\nsegmentation (CSS), of which the dense prediction peculiarity makes it a\nchallenging, intricate and burgeoning task. In this paper, we present a review\nof CSS, committing to building a comprehensive survey on problem formulations,\nprimary challenges, universal datasets, neoteric theories and multifarious\napplications. Concretely, we begin by elucidating the problem definitions and\nprimary challenges. Based on an in-depth investigation of relevant approaches,\nwe sort out and categorize current CSS models into two main branches including\ndata-replay and data-free sets. In each branch, the corresponding approaches\nare similarity-based clustered and thoroughly analyzed, following qualitative\ncomparison and quantitative reproductions on relevant datasets. Besides, we\nalso introduce four CSS specialities with diverse application scenarios and\ndevelopment tendencies. Furthermore, we develop a benchmark for CSS\nencompassing representative references, evaluation results and reproductions,\nwhich is available at~\\url{https://github.com/YBIO/SurveyCSS}. We hope this\nsurvey can serve as a reference-worthy and stimulating contribution to the\nadvancement of the life-long learning field, while also providing valuable\nperspectives for related fields.\n", "date": "2023-10-22", "categories": []}
{"id": "2408.12161", "title": "Rebalancing Multi-Label Class-Incremental Learning", "abstract": "  Multi-label class-incremental learning (MLCIL) is essential for real-world\nmulti-label applications, allowing models to learn new labels while retaining\npreviously learned knowledge continuously. However, recent MLCIL approaches can\nonly achieve suboptimal performance due to the oversight of the\npositive-negative imbalance problem, which manifests at both the label and loss\nlevels because of the task-level partial label issue. The imbalance at the\nlabel level arises from the substantial absence of negative labels, while the\nimbalance at the loss level stems from the asymmetric contributions of the\npositive and negative loss parts to the optimization. To address the issue\nabove, we propose a Rebalance framework for both the Loss and Label levels\n(RebLL), which integrates two key modules: asymmetric knowledge distillation\n(AKD) and online relabeling (OR). AKD is proposed to rebalance at the loss\nlevel by emphasizing the negative label learning in classification loss and\ndown-weighting the contribution of overconfident predictions in distillation\nloss. OR is designed for label rebalance, which restores the original class\ndistribution in memory by online relabeling the missing classes. Our\ncomprehensive experiments on the PASCAL VOC and MS-COCO datasets demonstrate\nthat this rebalancing strategy significantly improves performance, achieving\nnew state-of-the-art results even with a vanilla CNN backbone.\n", "date": "2024-08-22", "categories": []}
{"id": "2408.11742", "title": "CluMo: Cluster-based Modality Fusion Prompt for Continual Learning in\n  Visual Question Answering", "abstract": "  Large vision-language models (VLMs) have shown significant performance boost\nin various application domains. However, adopting them to deal with several\nsequentially encountered tasks has been challenging because finetuning a VLM on\na task normally leads to reducing its generalization power and the capacity of\nlearning new tasks as well as causing catastrophic forgetting on previously\nlearned tasks. Enabling using VLMs in multimodal continual learning (CL)\nsettings can help to address such scenarios. To improve generalization capacity\nand prevent catastrophic forgetting, we propose a novel prompt-based CL method\nfor VLMs, namely $\\textbf{Clu}$ster-based $\\textbf{Mo}$dality Fusion Prompt\n(\\textbf{CluMo}). We design a novel \\textbf{Key-Key-Prompt} pair, where each\nprompt is associated with a visual prompt key and a textual prompt key. We\nadopt a two-stage training strategy. During the first stage, the single-modal\nkeys are trained via $K$-means clustering algorithm to help select the best\nsemantically matched prompt. During the second stage, the prompt keys are\nfrozen, the selected prompt is attached to the input for training the VLM in\nthe CL scenario. Experiments on two benchmarks demonstrate that our method\nachieves SOTA performance.\n", "date": "2024-08-21", "categories": []}
{"id": "2408.10349", "title": "AIR: Analytic Imbalance Rectifier for Continual Learning", "abstract": "  Continual learning enables AI models to learn new data sequentially without\nretraining in real-world scenarios. Most existing methods assume the training\ndata are balanced, aiming to reduce the catastrophic forgetting problem that\nmodels tend to forget previously generated data. However, data imbalance and\nthe mixture of new and old data in real-world scenarios lead the model to\nignore categories with fewer training samples. To solve this problem, we\npropose an analytic imbalance rectifier algorithm (AIR), a novel online\nexemplar-free continual learning method with an analytic (i.e., closed-form)\nsolution for data-imbalanced class-incremental learning (CIL) and generalized\nCIL scenarios in real-world continual learning. AIR introduces an analytic\nre-weighting module (ARM) that calculates a re-weighting factor for each class\nfor the loss function to balance the contribution of each category to the\noverall loss and solve the problem of imbalanced training data. AIR uses the\nleast squares technique to give a non-discriminatory optimal classifier and its\niterative update method in continual learning. Experimental results on multiple\ndatasets show that AIR significantly outperforms existing methods in\nlong-tailed and generalized CIL scenarios. The source code is available at\nhttps://github.com/fang-d/AIR.\n", "date": "2024-08-19", "categories": []}
{"id": "2301.04584", "title": "Continual HyperTransformer: A Meta-Learner for Continual Few-Shot\n  Learning", "abstract": "  We focus on the problem of learning without forgetting from multiple tasks\narriving sequentially, where each task is defined using a few-shot episode of\nnovel or already seen classes. We approach this problem using the recently\npublished HyperTransformer (HT), a Transformer-based hypernetwork that\ngenerates specialized task-specific CNN weights directly from the support set.\nIn order to learn from a continual sequence of tasks, we propose to recursively\nre-use the generated weights as input to the HT for the next task. This way,\nthe generated CNN weights themselves act as a representation of previously\nlearned tasks, and the HT is trained to update these weights so that the new\ntask can be learned without forgetting past tasks. This approach is different\nfrom most continual learning algorithms that typically rely on using replay\nbuffers, weight regularization or task-dependent architectural changes. We\ndemonstrate that our proposed Continual HyperTransformer method equipped with a\nprototypical loss is capable of learning and retaining knowledge about past\ntasks for a variety of scenarios, including learning from mini-batches, and\ntask-incremental and class-incremental learning scenarios.\n", "date": "2023-01-11", "categories": []}
{"id": "2405.03244", "title": "Examining Changes in Internal Representations of Continual Learning\n  Models Through Tensor Decomposition", "abstract": "  Continual learning (CL) has spurred the development of several methods aimed\nat consolidating previous knowledge across sequential learning. Yet, the\nevaluations of these methods have primarily focused on the final output, such\nas changes in the accuracy of predicted classes, overlooking the issue of\nrepresentational forgetting within the model. In this paper, we propose a novel\nrepresentation-based evaluation framework for CL models. This approach involves\ngathering internal representations from throughout the continual learning\nprocess and formulating three-dimensional tensors. The tensors are formed by\nstacking representations, such as layer activations, generated from several\ninputs and model `snapshots', throughout the learning process. By conducting\ntensor component analysis (TCA), we aim to uncover meaningful patterns about\nhow the internal representations evolve, expecting to highlight the merits or\nshortcomings of examined CL strategies. We conduct our analyses across\ndifferent model architectures and importance-based continual learning\nstrategies, with a curated task selection. While the results of our approach\nmirror the difference in performance of various CL strategies, we found that\nour methodology did not directly highlight specialized clusters of neurons, nor\nprovide an immediate understanding the evolution of filters. We believe a\nscaled down version of our approach will provide insight into the benefits and\npitfalls of using TCA to study continual learning dynamics.\n", "date": "2024-05-06", "categories": []}
{"id": "2407.09271", "title": "iNeMo: Incremental Neural Mesh Models for Robust Class-Incremental\n  Learning", "abstract": "  Different from human nature, it is still common practice today for vision\ntasks to train deep learning models only initially and on fixed datasets. A\nvariety of approaches have recently addressed handling continual data streams.\nHowever, extending these methods to manage out-of-distribution (OOD) scenarios\nhas not effectively been investigated. On the other hand, it has recently been\nshown that non-continual neural mesh models exhibit strong performance in\ngeneralizing to such OOD scenarios. To leverage this decisive property in a\ncontinual learning setting, we propose incremental neural mesh models that can\nbe extended with new meshes over time. In addition, we present a latent space\ninitialization strategy that enables us to allocate feature space for future\nunseen classes in advance and a positional regularization term that forces the\nfeatures of the different classes to consistently stay in respective latent\nspace regions. We demonstrate the effectiveness of our method through extensive\nexperiments on the Pascal3D and ObjectNet3D datasets and show that our approach\noutperforms the baselines for classification by $2-6\\%$ in the in-domain and by\n$6-50\\%$ in the OOD setting. Our work also presents the first incremental\nlearning approach for pose estimation. Our code and model can be found at\nhttps://github.com/Fischer-Tom/iNeMo.\n", "date": "2024-07-12", "categories": []}
{"id": "2408.09350", "title": "E-CGL: An Efficient Continual Graph Learner", "abstract": "  Continual learning has emerged as a crucial paradigm for learning from\nsequential data while preserving previous knowledge. In the realm of continual\ngraph learning, where graphs continuously evolve based on streaming graph data,\ncontinual graph learning presents unique challenges that require adaptive and\nefficient graph learning methods in addition to the problem of catastrophic\nforgetting. The first challenge arises from the interdependencies between\ndifferent graph data, where previous graphs can influence new data\ndistributions. The second challenge lies in the efficiency concern when dealing\nwith large graphs. To addresses these two problems, we produce an Efficient\nContinual Graph Learner (E-CGL) in this paper. We tackle the interdependencies\nissue by demonstrating the effectiveness of replay strategies and introducing a\ncombined sampling strategy that considers both node importance and diversity.\nTo overcome the limitation of efficiency, E-CGL leverages a simple yet\neffective MLP model that shares weights with a GCN during training, achieving\nacceleration by circumventing the computationally expensive message passing\nprocess. Our method comprehensively surpasses nine baselines on four graph\ncontinual learning datasets under two settings, meanwhile E-CGL largely reduces\nthe catastrophic forgetting problem down to an average of -1.1%. Additionally,\nE-CGL achieves an average of 15.83x training time acceleration and 4.89x\ninference time acceleration across the four datasets. These results indicate\nthat E-CGL not only effectively manages the correlation between different graph\ndata during continual training but also enhances the efficiency of continual\nlearning on large graphs. The code is publicly available at\nhttps://github.com/aubreygjh/E-CGL.\n", "date": "2024-08-18", "categories": []}
{"id": "2408.10046", "title": "Exploiting Fine-Grained Prototype Distribution for Boosting Unsupervised\n  Class Incremental Learning", "abstract": "  The dynamic nature of open-world scenarios has attracted more attention to\nclass incremental learning (CIL). However, existing CIL methods typically\npresume the availability of complete ground-truth labels throughout the\ntraining process, an assumption rarely met in practical applications.\nConsequently, this paper explores a more challenging problem of unsupervised\nclass incremental learning (UCIL). The essence of addressing this problem lies\nin effectively capturing comprehensive feature representations and discovering\nunknown novel classes. To achieve this, we first model the knowledge of class\ndistribution by exploiting fine-grained prototypes. Subsequently, a granularity\nalignment technique is introduced to enhance the unsupervised class discovery.\nAdditionally, we proposed a strategy to minimize overlap between novel and\nexisting classes, thereby preserving historical knowledge and mitigating the\nphenomenon of catastrophic forgetting. Extensive experiments on the five\ndatasets demonstrate that our approach significantly outperforms current\nstate-of-the-art methods, indicating the effectiveness of the proposed method.\n", "date": "2024-08-19", "categories": []}
{"id": "2405.01040", "title": "Few Shot Class Incremental Learning using Vision-Language models", "abstract": "  Recent advancements in deep learning have demonstrated remarkable performance\ncomparable to human capabilities across various supervised computer vision\ntasks. However, the prevalent assumption of having an extensive pool of\ntraining data encompassing all classes prior to model training often diverges\nfrom real-world scenarios, where limited data availability for novel classes is\nthe norm. The challenge emerges in seamlessly integrating new classes with few\nsamples into the training data, demanding the model to adeptly accommodate\nthese additions without compromising its performance on base classes. To\naddress this exigency, the research community has introduced several solutions\nunder the realm of few-shot class incremental learning (FSCIL).\n  In this study, we introduce an innovative FSCIL framework that utilizes\nlanguage regularizer and subspace regularizer. During base training, the\nlanguage regularizer helps incorporate semantic information extracted from a\nVision-Language model. The subspace regularizer helps in facilitating the\nmodel's acquisition of nuanced connections between image and text semantics\ninherent to base classes during incremental training. Our proposed framework\nnot only empowers the model to embrace novel classes with limited data, but\nalso ensures the preservation of performance on base classes. To substantiate\nthe efficacy of our approach, we conduct comprehensive experiments on three\ndistinct FSCIL benchmarks, where our framework attains state-of-the-art\nperformance.\n", "date": "2024-05-02", "categories": []}
{"id": "2408.08084", "title": "An Efficient Replay for Class-Incremental Learning with Pre-trained\n  Models", "abstract": "  In general class-incremental learning, researchers typically use sample sets\nas a tool to avoid catastrophic forgetting during continuous learning. At the\nsame time, researchers have also noted the differences between\nclass-incremental learning and Oracle training and have attempted to make\ncorrections. In recent years, researchers have begun to develop\nclass-incremental learning algorithms utilizing pre-trained models, achieving\nsignificant results. This paper observes that in class-incremental learning,\nthe steady state among the weight guided by each class center is disrupted,\nwhich is significantly correlated with catastrophic forgetting. Based on this,\nwe propose a new method to overcoming forgetting . In some cases, by retaining\nonly a single sample unit of each class in memory for replay and applying\nsimple gradient constraints, very good results can be achieved. Experimental\nresults indicate that under the condition of pre-trained models, our method can\nachieve competitive performance with very low computational cost and by simply\nusing the cross-entropy loss.\n", "date": "2024-08-15", "categories": []}
{"id": "2408.08295", "title": "SLCA++: Unleash the Power of Sequential Fine-tuning for Continual\n  Learning with Pre-training", "abstract": "  In recent years, continual learning with pre-training (CLPT) has received\nwidespread interest, instead of its traditional focus of training from scratch.\nThe use of strong pre-trained models (PTMs) can greatly facilitate knowledge\ntransfer and alleviate catastrophic forgetting, but also suffers from\nprogressive overfitting of pre-trained knowledge into specific downstream\ntasks. A majority of current efforts often keep the PTMs frozen and incorporate\ntask-specific prompts to instruct representation learning, coupled with a\nprompt selection process for inference. However, due to the limited capacity of\nprompt parameters, this strategy demonstrates only sub-optimal performance in\ncontinual learning. In comparison, tuning all parameters of PTMs often provides\nthe greatest potential for representation learning, making sequential\nfine-tuning (Seq FT) a fundamental baseline that has been overlooked in CLPT.\nTo this end, we present an in-depth analysis of the progressive overfitting\nproblem from the lens of Seq FT. Considering that the overly fast\nrepresentation learning and the biased classification layer constitute this\nparticular problem, we introduce the advanced Slow Learner with Classifier\nAlignment (SLCA++) framework to unleash the power of Seq FT, serving as a\nstrong baseline approach for CLPT. Our approach involves a Slow Learner to\nselectively reduce the learning rate of backbone parameters, and a Classifier\nAlignment to align the disjoint classification layers in a post-hoc fashion. We\nfurther enhance the efficacy of SL with a symmetric cross-entropy loss, as well\nas employ a parameter-efficient strategy to implement Seq FT with SLCA++.\nAcross a variety of continual learning scenarios on image classification\nbenchmarks, our approach provides substantial improvements and outperforms\nstate-of-the-art methods by a large margin. Code:\nhttps://github.com/GengDavid/SLCA.\n", "date": "2024-08-15", "categories": []}
{"id": "2408.03219", "title": "Learning to Learn without Forgetting using Attention", "abstract": "  Continual learning (CL) refers to the ability to continually learn over time\nby accommodating new knowledge while retaining previously learned experience.\nWhile this concept is inherent in human learning, current machine learning\nmethods are highly prone to overwrite previously learned patterns and thus\nforget past experience. Instead, model parameters should be updated selectively\nand carefully, avoiding unnecessary forgetting while optimally leveraging\npreviously learned patterns to accelerate future learning. Since hand-crafting\neffective update mechanisms is difficult, we propose meta-learning a\ntransformer-based optimizer to enhance CL. This meta-learned optimizer uses\nattention to learn the complex relationships between model parameters across a\nstream of tasks, and is designed to generate effective weight updates for the\ncurrent task while preventing catastrophic forgetting on previously encountered\ntasks. Evaluations on benchmark datasets like SplitMNIST, RotatedMNIST, and\nSplitCIFAR-100 affirm the efficacy of the proposed approach in terms of both\nforward and backward transfer, even on small sets of labeled data, highlighting\nthe advantages of integrating a meta-learned optimizer within the continual\nlearning framework.\n", "date": "2024-08-06", "categories": []}
{"id": "2402.11196", "title": "Maintaining Adversarial Robustness in Continuous Learning", "abstract": "  Adversarial robustness is essential for security and reliability of machine\nlearning systems. However, adversarial robustness enhanced by defense\nalgorithms is easily erased as the neural network's weights update to learn new\ntasks. To address this vulnerability, it is essential to improve the capability\nof neural networks in terms of robust continual learning. Specially, we propose\na novel gradient projection technique that effectively stabilizes sample\ngradients from previous data by orthogonally projecting back-propagation\ngradients onto a crucial subspace before using them for weight updates. This\ntechnique can maintaining robustness by collaborating with a class of defense\nalgorithms through sample gradient smoothing. The experimental results on four\nbenchmarks including Split-CIFAR100 and Split-miniImageNet, demonstrate that\nthe superiority of the proposed approach in mitigating rapidly degradation of\nrobustness during continual learning even when facing strong adversarial\nattacks.\n", "date": "2024-02-17", "categories": []}
{"id": "2404.09447", "title": "kNN-CLIP: Retrieval Enables Training-Free Segmentation on Continually\n  Expanding Large Vocabularies", "abstract": "  Continual segmentation has not yet tackled the challenge of improving\nopen-vocabulary segmentation models with training data for accurate\nsegmentation across large, continually expanding vocabularies. We discover that\ntraditional continual training results in severe catastrophic forgetting,\nfailing to outperform a zero-shot segmentation baseline. We introduce a novel\ntraining-free strategy, kNN-CLIP, which augments the model with a database of\ninstance embeddings for semantic and panoptic segmentation that achieves zero\nforgetting. We demonstrate that kNN-CLIP can adapt to continually growing\nvocabularies without the need for retraining or large memory costs. kNN-CLIP\nenables open-vocabulary segmentation methods to expand their vocabularies on\nany domain with a single pass through the data, while only storing compact\nembeddings. This approach minimizes both compute and memory costs. kNN-CLIP\nachieves state-of-the-art performance across large-vocabulary semantic and\npanoptic segmentation datasets. We hope kNN-CLIP represents a significant step\nforward in enabling more efficient and adaptable continual segmentation, paving\nthe way for advances in real-world large-vocabulary continual segmentation\nmethods.\n", "date": "2024-04-15", "categories": []}
{"id": "2403.12559", "title": "Confidence Self-Calibration for Multi-Label Class-Incremental Learning", "abstract": "  The partial label challenge in Multi-Label Class-Incremental Learning (MLCIL)\narises when only the new classes are labeled during training, while past and\nfuture labels remain unavailable. This issue leads to a proliferation of\nfalse-positive errors due to erroneously high confidence multi-label\npredictions, exacerbating catastrophic forgetting within the disjoint label\nspace. In this paper, we aim to refine multi-label confidence calibration in\nMLCIL and propose a Confidence Self-Calibration (CSC) approach. Firstly, for\nlabel relationship calibration, we introduce a class-incremental graph\nconvolutional network that bridges the isolated label spaces by constructing\nlearnable, dynamically extended label relationship graph. Then, for confidence\ncalibration, we present a max-entropy regularization for each multi-label\nincrement, facilitating confidence self-calibration through the penalization of\nover-confident output distributions. Our approach attains new state-of-the-art\nresults in MLCIL tasks on both MS-COCO and PASCAL VOC datasets, with the\ncalibration of label confidences confirmed through our methodology.\n", "date": "2024-03-19", "categories": []}
{"id": "2404.19132", "title": "Integrating Present and Past in Unsupervised Continual Learning", "abstract": "  We formulate a unifying framework for unsupervised continual learning (UCL),\nwhich disentangles learning objectives that are specific to the present and the\npast data, encompassing stability, plasticity, and cross-task consolidation.\nThe framework reveals that many existing UCL approaches overlook cross-task\nconsolidation and try to balance plasticity and stability in a shared embedding\nspace. This results in worse performance due to a lack of within-task data\ndiversity and reduced effectiveness in learning the current task. Our method,\nOsiris, which explicitly optimizes all three objectives on separate embedding\nspaces, achieves state-of-the-art performance on all benchmarks, including two\nnovel benchmarks proposed in this paper featuring semantically structured task\nsequences. Compared to standard benchmarks, these two structured benchmarks\nmore closely resemble visual signals received by humans and animals when\nnavigating real-world environments. Finally, we show some preliminary evidence\nthat continual models can benefit from such realistic learning scenarios.\n", "date": "2024-04-29", "categories": []}
{"id": "2312.07887", "title": "Learn or Recall? Revisiting Incremental Learning with Pre-trained\n  Language Models", "abstract": "  Incremental Learning (IL) has been a long-standing problem in both vision and\nNatural Language Processing (NLP) communities. In recent years, as Pre-trained\nLanguage Models (PLMs) have achieved remarkable progress in various NLP\ndownstream tasks, utilizing PLMs as backbones has become a common practice in\nrecent research of IL in NLP. Most assume that catastrophic forgetting is the\nbiggest obstacle to achieving superior IL performance and propose various\ntechniques to overcome this issue. However, we find that this assumption is\nproblematic. Specifically, we revisit more than 20 methods on four\nclassification tasks (Text Classification, Intent Classification, Relation\nExtraction, and Named Entity Recognition) under the two most popular IL\nsettings (Class-Incremental and Task-Incremental) and reveal that most of them\nseverely underestimate the inherent anti-forgetting ability of PLMs. Based on\nthe observation, we propose a frustratingly easy method called SEQ* for IL with\nPLMs. The results show that SEQ* has competitive or superior performance\ncompared to state-of-the-art (SOTA) IL methods and requires considerably less\ntrainable parameters and training time. These findings urge us to revisit the\nIL with PLMs and encourage future studies to have a fundamental understanding\nof the catastrophic forgetting in PLMs. The data, code and scripts are publicly\navailable at\nhttps://github.com/zzz47zzz/codebase-for-incremental-learning-with-llm.\n", "date": "2023-12-13", "categories": []}
{"id": "2407.19429", "title": "FTF-ER: Feature-Topology Fusion-Based Experience Replay Method for\n  Continual Graph Learning", "abstract": "  Continual graph learning (CGL) is an important and challenging task that aims\nto extend static GNNs to dynamic task flow scenarios. As one of the mainstream\nCGL methods, the experience replay (ER) method receives widespread attention\ndue to its superior performance. However, existing ER methods focus on\nidentifying samples by feature significance or topological relevance, which\nlimits their utilization of comprehensive graph data. In addition, the\ntopology-based ER methods only consider local topological information and add\nneighboring nodes to the buffer, which ignores the global topological\ninformation and increases memory overhead. To bridge these gaps, we propose a\nnovel method called Feature-Topology Fusion-based Experience Replay (FTF-ER) to\neffectively mitigate the catastrophic forgetting issue with enhanced\nefficiency. Specifically, from an overall perspective to maximize the\nutilization of the entire graph data, we propose a highly complementary\napproach including both feature and global topological information, which can\nsignificantly improve the effectiveness of the sampled nodes. Moreover, to\nfurther utilize global topological information, we propose Hodge Potential\nScore (HPS) as a novel module to calculate the topological importance of nodes.\nHPS derives a global node ranking via Hodge decomposition on graphs, providing\nmore accurate global topological information compared to neighbor sampling. By\nexcluding neighbor sampling, HPS significantly reduces buffer storage costs for\nacquiring topological information and simultaneously decreases training time.\nCompared with state-of-the-art methods, FTF-ER achieves a significant\nimprovement of 3.6% in AA and 7.1% in AF on the OGB-Arxiv dataset,\ndemonstrating its superior performance in the class-incremental learning\nsetting.\n", "date": "2024-07-28", "categories": []}
{"id": "2408.04347", "title": "AggSS: An Aggregated Self-Supervised Approach for Class-Incremental\n  Learning", "abstract": "  This paper investigates the impact of self-supervised learning, specifically\nimage rotations, on various class-incremental learning paradigms. Here, each\nimage with a predefined rotation is considered as a new class for training. At\ninference, all image rotation predictions are aggregated for the final\nprediction, a strategy we term Aggregated Self-Supervision (AggSS). We observe\na shift in the deep neural network's attention towards intrinsic object\nfeatures as it learns through AggSS strategy. This learning approach\nsignificantly enhances class-incremental learning by promoting robust feature\nlearning. AggSS serves as a plug-and-play module that can be seamlessly\nincorporated into any class-incremental learning framework, leveraging its\npowerful feature learning capabilities to enhance performance across various\nclass-incremental learning approaches. Extensive experiments conducted on\nstandard incremental learning datasets CIFAR-100 and ImageNet-Subset\ndemonstrate the significant role of AggSS in improving performance within these\nparadigms.\n", "date": "2024-08-08", "categories": []}
{"id": "2407.19265", "title": "Towards Robust Few-shot Class Incremental Learning in Audio\n  Classification using Contrastive Representation", "abstract": "  In machine learning applications, gradual data ingress is common, especially\nin audio processing where incremental learning is vital for real-time\nanalytics. Few-shot class-incremental learning addresses challenges arising\nfrom limited incoming data. Existing methods often integrate additional\ntrainable components or rely on a fixed embedding extractor post-training on\nbase sessions to mitigate concerns related to catastrophic forgetting and the\ndangers of model overfitting. However, using cross-entropy loss alone during\nbase session training is suboptimal for audio data. To address this, we propose\nincorporating supervised contrastive learning to refine the representation\nspace, enhancing discriminative power and leading to better generalization\nsince it facilitates seamless integration of incremental classes, upon arrival.\nExperimental results on NSynth and LibriSpeech datasets with 100 classes, as\nwell as ESC dataset with 50 and 10 classes, demonstrate state-of-the-art\nperformance.\n", "date": "2024-07-27", "categories": []}
{"id": "2408.02695", "title": "Distribution-Level Memory Recall for Continual Learning: Preserving\n  Knowledge and Avoiding Confusion", "abstract": "  Continual Learning (CL) aims to enable Deep Neural Networks (DNNs) to learn\nnew data without forgetting previously learned knowledge. The key to achieving\nthis goal is to avoid confusion at the feature level, i.e., avoiding confusion\nwithin old tasks and between new and old tasks. Previous prototype-based CL\nmethods generate pseudo features for old knowledge replay by adding Gaussian\nnoise to the centroids of old classes. However, the distribution in the feature\nspace exhibits anisotropy during the incremental process, which prevents the\npseudo features from faithfully reproducing the distribution of old knowledge\nin the feature space, leading to confusion in classification boundaries within\nold tasks. To address this issue, we propose the Distribution-Level Memory\nRecall (DMR) method, which uses a Gaussian mixture model to precisely fit the\nfeature distribution of old knowledge at the distribution level and generate\npseudo features in the next stage. Furthermore, resistance to confusion at the\ndistribution level is also crucial for multimodal learning, as the problem of\nmultimodal imbalance results in significant differences in feature responses\nbetween different modalities, exacerbating confusion within old tasks in\nprototype-based CL methods. Therefore, we mitigate the multi-modal imbalance\nproblem by using the Inter-modal Guidance and Intra-modal Mining (IGIM) method\nto guide weaker modalities with prior information from dominant modalities and\nfurther explore useful information within modalities. For the second key, We\npropose the Confusion Index to quantitatively describe a model's ability to\ndistinguish between new and old tasks, and we use the Incremental Mixup Feature\nEnhancement (IMFE) method to enhance pseudo features with new sample features,\nalleviating classification confusion between new and old knowledge.\n", "date": "2024-08-04", "categories": []}
{"id": "2408.02930", "title": "The Need for a Big World Simulator: A Scientific Challenge for Continual\n  Learning", "abstract": "  The \"small agent, big world\" frame offers a conceptual view that motivates\nthe need for continual learning. The idea is that a small agent operating in a\nmuch bigger world cannot store all information that the world has to offer. To\nperform well, the agent must be carefully designed to ingest, retain, and eject\nthe right information. To enable the development of performant continual\nlearning agents, a number of synthetic environments have been proposed.\nHowever, these benchmarks suffer from limitations, including unnatural\ndistribution shifts and a lack of fidelity to the \"small agent, big world\"\nframing. This paper aims to formalize two desiderata for the design of future\nsimulated environments. These two criteria aim to reflect the objectives and\ncomplexity of continual learning in practical settings while enabling rapid\nprototyping of algorithms on a smaller scale.\n", "date": "2024-08-06", "categories": []}
{"id": "2408.02983", "title": "Diffusion Model Meets Non-Exemplar Class-Incremental Learning and Beyond", "abstract": "  Non-exemplar class-incremental learning (NECIL) is to resist catastrophic\nforgetting without saving old class samples. Prior methodologies generally\nemploy simple rules to generate features for replaying, suffering from large\ndistribution gap between replayed features and real ones. To address the\naforementioned issue, we propose a simple, yet effective\n\\textbf{Diff}usion-based \\textbf{F}eature \\textbf{R}eplay (\\textbf{DiffFR})\nmethod for NECIL. First, to alleviate the limited representational capacity\ncaused by fixing the feature extractor, we employ Siamese-based self-supervised\nlearning for initial generalizable features. Second, we devise diffusion models\nto generate class-representative features highly similar to real features,\nwhich provides an effective way for exemplar-free knowledge memorization.\nThird, we introduce prototype calibration to direct the diffusion model's focus\ntowards learning the distribution shapes of features, rather than the entire\ndistribution. Extensive experiments on public datasets demonstrate significant\nperformance gains of our DiffFR, outperforming the state-of-the-art NECIL\nmethods by 3.0\\% in average. The code will be made publicly available soon.\n", "date": "2024-08-06", "categories": []}
{"id": "2303.07338", "title": "Revisiting Class-Incremental Learning with Pre-Trained Models:\n  Generalizability and Adaptivity are All You Need", "abstract": "  Class-incremental learning (CIL) aims to adapt to emerging new classes\nwithout forgetting old ones. Traditional CIL models are trained from scratch to\ncontinually acquire knowledge as data evolves. Recently, pre-training has\nachieved substantial progress, making vast pre-trained models (PTMs) accessible\nfor CIL. Contrary to traditional methods, PTMs possess generalizable\nembeddings, which can be easily transferred for CIL. In this work, we revisit\nCIL with PTMs and argue that the core factors in CIL are adaptivity for model\nupdating and generalizability for knowledge transferring. 1) We first reveal\nthat frozen PTM can already provide generalizable embeddings for CIL.\nSurprisingly, a simple baseline (SimpleCIL) which continually sets the\nclassifiers of PTM to prototype features can beat state-of-the-art even without\ntraining on the downstream task. 2) Due to the distribution gap between\npre-trained and downstream datasets, PTM can be further cultivated with\nadaptivity via model adaptation. We propose AdaPt and mERge (APER), which\naggregates the embeddings of PTM and adapted models for classifier\nconstruction. APER is a general framework that can be orthogonally combined\nwith any parameter-efficient tuning method, which holds the advantages of PTM's\ngeneralizability and adapted model's adaptivity. 3) Additionally, considering\nprevious ImageNet-based benchmarks are unsuitable in the era of PTM due to data\noverlapping, we propose four new benchmarks for assessment, namely ImageNet-A,\nObjectNet, OmniBenchmark, and VTAB. Extensive experiments validate the\neffectiveness of APER with a unified and concise framework. Code is available\nat https://github.com/zhoudw-zdw/RevisitingCIL\n", "date": "2023-03-13", "categories": []}
{"id": "2402.12035", "title": "Class-incremental Learning for Time Series: Benchmark and Evaluation", "abstract": "  Real-world environments are inherently non-stationary, frequently introducing\nnew classes over time. This is especially common in time series classification,\nsuch as the emergence of new disease classification in healthcare or the\naddition of new activities in human activity recognition. In such cases, a\nlearning system is required to assimilate novel classes effectively while\navoiding catastrophic forgetting of the old ones, which gives rise to the\nClass-incremental Learning (CIL) problem. However, despite the encouraging\nprogress in the image and language domains, CIL for time series data remains\nrelatively understudied. Existing studies suffer from inconsistent experimental\ndesigns, necessitating a comprehensive evaluation and benchmarking of methods\nacross a wide range of datasets. To this end, we first present an overview of\nthe Time Series Class-incremental Learning (TSCIL) problem, highlight its\nunique challenges, and cover the advanced methodologies. Further, based on\nstandardized settings, we develop a unified experimental framework that\nsupports the rapid development of new algorithms, easy integration of new\ndatasets, and standardization of the evaluation process. Using this framework,\nwe conduct a comprehensive evaluation of various generic and\ntime-series-specific CIL methods in both standard and privacy-sensitive\nscenarios. Our extensive experiments not only provide a standard baseline to\nsupport future research but also shed light on the impact of various design\nfactors such as normalization layers or memory budget thresholds. Codes are\navailable at https://github.com/zqiao11/TSCIL.\n", "date": "2024-02-19", "categories": []}
{"id": "2406.11340", "title": "CM2-Net: Continual Cross-Modal Mapping Network for Driver Action\n  Recognition", "abstract": "  Driver action recognition has significantly advanced in enhancing\ndriver-vehicle interactions and ensuring driving safety by integrating multiple\nmodalities, such as infrared and depth. Nevertheless, compared to RGB modality\nonly, it is always laborious and costly to collect extensive data for all types\nof non-RGB modalities in car cabin environments. Therefore, previous works have\nsuggested independently learning each non-RGB modality by fine-tuning a model\npre-trained on RGB videos, but these methods are less effective in extracting\ninformative features when faced with newly-incoming modalities due to large\ndomain gaps. In contrast, we propose a Continual Cross-Modal Mapping Network\n(CM2-Net) to continually learn each newly-incoming modality with instructive\nprompts from the previously-learned modalities. Specifically, we have developed\nAccumulative Cross-modal Mapping Prompting (ACMP), to map the discriminative\nand informative features learned from previous modalities into the feature\nspace of newly-incoming modalities. Then, when faced with newly-incoming\nmodalities, these mapped features are able to provide effective prompts for\nwhich features should be extracted and prioritized. These prompts are\naccumulating throughout the continual learning process, thereby boosting\nfurther recognition performances. Extensive experiments conducted on the\nDrive&Act dataset demonstrate the performance superiority of CM2-Net on both\nuni- and multi-modal driver action recognition.\n", "date": "2024-06-17", "categories": []}
{"id": "2408.01970", "title": "SR-CIS: Self-Reflective Incremental System with Decoupled Memory and\n  Reasoning", "abstract": "  The ability of humans to rapidly learn new knowledge while retaining old\nmemories poses a significant challenge for current deep learning models. To\nhandle this challenge, we draw inspiration from human memory and learning\nmechanisms and propose the Self-Reflective Complementary Incremental System\n(SR-CIS). Comprising the deconstructed Complementary Inference Module (CIM) and\nComplementary Memory Module (CMM), SR-CIS features a small model for fast\ninference and a large model for slow deliberation in CIM, enabled by the\nConfidence-Aware Online Anomaly Detection (CA-OAD) mechanism for efficient\ncollaboration. CMM consists of task-specific Short-Term Memory (STM) region and\na universal Long-Term Memory (LTM) region. By setting task-specific Low-Rank\nAdaptive (LoRA) and corresponding prototype weights and biases, it instantiates\nexternal storage for parameter and representation memory, thus deconstructing\nthe memory module from the inference module. By storing textual descriptions of\nimages during training and combining them with the Scenario Replay Module (SRM)\npost-training for memory combination, along with periodic short-to-long-term\nmemory restructuring, SR-CIS achieves stable incremental memory with limited\nstorage requirements. Balancing model plasticity and memory stability under\nconstraints of limited storage and low data resources, SR-CIS surpasses\nexisting competitive baselines on multiple standard and few-shot incremental\nlearning benchmarks.\n", "date": "2024-08-04", "categories": []}
{"id": "2407.21043", "title": "CP-Prompt: Composition-Based Cross-modal Prompting for\n  Domain-Incremental Continual Learning", "abstract": "  The key challenge of cross-modal domain-incremental learning (DIL) is to\nenable the learning model to continuously learn from novel data with different\nfeature distributions under the same task without forgetting old ones. However,\nexisting top-performing methods still cause high forgetting rates, by lacking\nintra-domain knowledge extraction and inter-domain common prompting strategy.\nIn this paper, we propose a simple yet effective framework, CP-Prompt, by\ntraining limited parameters to instruct a pre-trained model to learn new\ndomains and avoid forgetting existing feature distributions. CP-Prompt captures\nintra-domain knowledge by compositionally inserting personalized prompts on\nmulti-head self-attention layers and then learns the inter-domain knowledge\nwith a common prompting strategy. CP-Prompt shows superiority compared with\nstate-of-the-art baselines among three widely evaluated DIL tasks. The source\ncode is available at https://github.com/dannis97500/CP_Prompt.\n", "date": "2024-07-22", "categories": []}
{"id": "2407.21571", "title": "PMoE: Progressive Mixture of Experts with Asymmetric Transformer for\n  Continual Learning", "abstract": "  Large Language Models (LLMs) encounter significant challenges in continual\nlearning due to catastrophic forgetting, where new information overwrites\npreviously acquired knowledge. This limitation leads to substantial\nenvironmental and economic waste. In this study, we introduce the PMoE,\nProgressive Mixture of Experts with Asymmetric Transformer, which aims to\nminimize forgetting by utilizing an asymmetric design with shallow layers\ndedicated to general knowledge and deep layers for new knowledge. PMoE\nincorporates progressively added experts in deep layers and a router that\nallocates new knowledge to the appropriate experts efficiently. The router,\npositioned adjacent to the deep layers, utilizes deep features aggregating\nconsolidated information. This enables the router to perform efficiently,\nallocating new knowledge to the appropriate experts, which progressively\nincrease in the deep layers. Extensive experiments on TRACE datasets and\ngeneral language understanding datasets demonstrate that the proposed PMoE\noutperforms previous state-of-the-art approaches.\n", "date": "2024-07-31", "categories": []}
{"id": "2407.06322", "title": "MagMax: Leveraging Model Merging for Seamless Continual Learning", "abstract": "  This paper introduces a continual learning approach named MagMax, which\nutilizes model merging to enable large pre-trained models to continuously learn\nfrom new data without forgetting previously acquired knowledge. Distinct from\ntraditional continual learning methods that aim to reduce forgetting during\ntask training, MagMax combines sequential fine-tuning with a maximum magnitude\nweight selection for effective knowledge integration across tasks. Our initial\ncontribution is an extensive examination of model merging techniques, revealing\nthat simple approaches like weight averaging and random weight selection\nsurprisingly hold up well in various continual learning contexts. More\nimportantly, we present MagMax, a novel model-merging strategy that enables\ncontinual learning of large pre-trained models for successive tasks. Our\nthorough evaluation demonstrates the superiority of MagMax in various\nscenarios, including class- and domain-incremental learning settings. The code\nis available at this URL: https://github.com/danielm1405/magmax.\n", "date": "2024-07-08", "categories": []}
{"id": "2407.20705", "title": "PIP: Prototypes-Injected Prompt for Federated Class Incremental Learning", "abstract": "  Federated Class Incremental Learning (FCIL) is a new direction in continual\nlearning (CL) for addressing catastrophic forgetting and non-IID data\ndistribution simultaneously. Existing FCIL methods call for high communication\ncosts and exemplars from previous classes. We propose a novel rehearsal-free\nmethod for FCIL named prototypes-injected prompt (PIP) that involves 3 main\nideas: a) prototype injection on prompt learning, b) prototype augmentation,\nand c) weighted Gaussian aggregation on the server side. Our experiment result\nshows that the proposed method outperforms the current state of the arts\n(SOTAs) with a significant improvement (up to 33%) in CIFAR100, MiniImageNet\nand TinyImageNet datasets. Our extensive analysis demonstrates the robustness\nof PIP in different task sizes, and the advantage of requiring smaller\nparticipating local clients, and smaller global rounds. For further study,\nsource codes of PIP, baseline, and experimental logs are shared publicly in\nhttps://github.com/anwarmaxsum/PIP.\n", "date": "2024-07-30", "categories": []}
{"id": "2407.20956", "title": "An Effective Dynamic Gradient Calibration Method for Continual Learning", "abstract": "  Continual learning (CL) is a fundamental topic in machine learning, where the\ngoal is to train a model with continuously incoming data and tasks. Due to the\nmemory limit, we cannot store all the historical data, and therefore confront\nthe ``catastrophic forgetting'' problem, i.e., the performance on the previous\ntasks can substantially decrease because of the missing information in the\nlatter period. Though a number of elegant methods have been proposed, the\ncatastrophic forgetting phenomenon still cannot be well avoided in practice. In\nthis paper, we study the problem from the gradient perspective, where our aim\nis to develop an effective algorithm to calibrate the gradient in each updating\nstep of the model; namely, our goal is to guide the model to be updated in the\nright direction under the situation that a large amount of historical data are\nunavailable. Our idea is partly inspired by the seminal stochastic variance\nreduction methods (e.g., SVRG and SAGA) for reducing the variance of gradient\nestimation in stochastic gradient descent algorithms. Another benefit is that\nour approach can be used as a general tool, which is able to be incorporated\nwith several existing popular CL methods to achieve better performance. We also\nconduct a set of experiments on several benchmark datasets to evaluate the\nperformance in practice.\n", "date": "2024-07-30", "categories": []}
{"id": "2404.14721", "title": "Dynamically Anchored Prompting for Task-Imbalanced Continual Learning", "abstract": "  Existing continual learning literature relies heavily on a strong assumption\nthat tasks arrive with a balanced data stream, which is often unrealistic in\nreal-world applications. In this work, we explore task-imbalanced continual\nlearning (TICL) scenarios where the distribution of task data is non-uniform\nacross the whole learning process. We find that imbalanced tasks significantly\nchallenge the capability of models to control the trade-off between stability\nand plasticity from the perspective of recent prompt-based continual learning\nmethods. On top of the above finding, we propose Dynamically Anchored Prompting\n(DAP), a prompt-based method that only maintains a single general prompt to\nadapt to the shifts within a task stream dynamically. This general prompt is\nregularized in the prompt space with two specifically designed prompt anchors,\ncalled boosting anchor and stabilizing anchor, to balance stability and\nplasticity in TICL. Remarkably, DAP achieves this balance by only storing a\nprompt across the data stream, therefore offering a substantial advantage in\nrehearsal-free CL. Extensive experiments demonstrate that the proposed DAP\nresults in 4.5% to 15% absolute improvements over state-of-the-art methods on\nbenchmarks under task-imbalanced settings. Our code is available at\nhttps://github.com/chenxing6666/DAP\n", "date": "2024-04-23", "categories": []}
{"id": "2403.04400", "title": "Exploring Continual Learning of Compositional Generalization in NLI", "abstract": "  Compositional Natural Language Inference has been explored to assess the true\nabilities of neural models to perform NLI. Yet, current evaluations assume\nmodels to have full access to all primitive inferences in advance, in contrast\nto humans that continuously acquire inference knowledge. In this paper, we\nintroduce the Continual Compositional Generalization in Inference (C2Gen NLI)\nchallenge, where a model continuously acquires knowledge of constituting\nprimitive inference tasks as a basis for compositional inferences. We explore\nhow continual learning affects compositional generalization in NLI, by\ndesigning a continual learning setup for compositional NLI inference tasks. Our\nexperiments demonstrate that models fail to compositionally generalize in a\ncontinual scenario. To address this problem, we first benchmark various\ncontinual learning algorithms and verify their efficacy. We then further\nanalyze C2Gen, focusing on how to order primitives and compositional inference\ntypes and examining correlations between subtasks. Our analyses show that by\nlearning subtasks continuously while observing their dependencies and\nincreasing degrees of difficulty, continual learning can enhance composition\ngeneralization ability.\n", "date": "2024-03-07", "categories": []}
{"id": "2405.07327", "title": "Liquid Ensemble Selection for Continual Learning", "abstract": "  Continual learning aims to enable machine learning models to continually\nlearn from a shifting data distribution without forgetting what has already\nbeen learned. Such shifting distributions can be broken into disjoint subsets\nof related examples; by training each member of an ensemble on a different\nsubset it is possible for the ensemble as a whole to achieve much higher\naccuracy with less forgetting than a naive model. We address the problem of\nselecting which models within an ensemble should learn on any given data, and\nwhich should predict. By drawing on work from delegative voting we develop an\nalgorithm for using delegation to dynamically select which models in an\nensemble are active. We explore a variety of delegation methods and performance\nmetrics, ultimately finding that delegation is able to provide a significant\nperformance boost over naive learning in the face of distribution shifts.\n", "date": "2024-05-12", "categories": []}
{"id": "2407.00465", "title": "Characterizing Continual Learning Scenarios and Strategies for Audio\n  Analysis", "abstract": "  Audio analysis is useful in many application scenarios. The state-of-the-art\naudio analysis approaches assume the data distribution at training and\ndeployment time will be the same. However, due to various real-life challenges,\nthe data may encounter drift in its distribution or can encounter new classes\nin the late future. Thus, a one-time trained model might not perform\nadequately. Continual learning (CL) approaches are devised to handle such\nchanges in data distribution. There have been a few attempts to use CL\napproaches for audio analysis. Yet, there is a lack of a systematic evaluation\nframework. In this paper, we create a comprehensive CL dataset and characterize\nCL approaches for audio-based monitoring tasks. We have investigated the\nfollowing CL and non-CL approaches: EWC, LwF, SI, GEM, A-GEM, GDumb, Replay,\nNaive, Cumulative, and Joint training. The study is very beneficial for\nresearchers and practitioners working in the area of audio analysis for\ndeveloping adaptive models. We observed that Replay achieved better results\nthan other methods in the DCASE challenge data. It achieved an accuracy of\n70.12% for the domain incremental scenario and an accuracy of 96.98% for the\nclass incremental scenario.\n", "date": "2024-06-29", "categories": []}
{"id": "2407.18526", "title": "Constructing Enhanced Mutual Information for Online Class-Incremental\n  Learning", "abstract": "  Online Class-Incremental continual Learning (OCIL) addresses the challenge of\ncontinuously learning from a single-channel data stream, adapting to new tasks\nwhile mitigating catastrophic forgetting. Recently, Mutual Information\n(MI)-based methods have shown promising performance in OCIL. However, existing\nMI-based methods treat various knowledge components in isolation, ignoring the\nknowledge confusion across tasks. This narrow focus on simple MI knowledge\nalignment may lead to old tasks being easily forgotten with the introduction of\nnew tasks, risking the loss of common parts between past and present\nknowledge.To address this, we analyze the MI relationships from the\nperspectives of diversity, representativeness, and separability, and propose an\nEnhanced Mutual Information (EMI) method based on knwoledge decoupling. EMI\nconsists of Diversity Mutual Information (DMI), Representativeness Mutual\nInformation (RMI) and Separability Mutual Information (SMI). DMI diversifies\nintra-class sample features by considering the similarity relationships among\ninter-class sample features to enable the network to learn more general\nknowledge. RMI summarizes representative features for each category and aligns\nsample features with these representative features, making the intra-class\nsample distribution more compact. SMI establishes MI relationships for\ninter-class representative features, enhancing the stability of representative\nfeatures while increasing the distinction between inter-class representative\nfeatures, thus creating clear boundaries between class. Extensive experimental\nresults on widely used benchmark datasets demonstrate the superior performance\nof EMI over state-of-the-art baseline methods.\n", "date": "2024-07-26", "categories": []}
{"id": "2407.18743", "title": "Towards Effective and Efficient Continual Pre-training of Large Language\n  Models", "abstract": "  Continual pre-training (CPT) has been an important approach for adapting\nlanguage models to specific domains or tasks. To make the CPT approach more\ntraceable, this paper presents a technical report for continually pre-training\nLlama-3 (8B), which significantly enhances the Chinese language ability and\nscientific reasoning ability of the backbone model. To enhance the new\nabilities while retaining the original abilities, we design specific data\nmixture and curriculum strategies by utilizing existing datasets and\nsynthesizing high-quality datasets. Specifically, we synthesize\nmultidisciplinary scientific question and answer (QA) pairs based on related\nweb pages, and subsequently incorporate these synthetic data to improve the\nscientific reasoning ability of Llama-3. We refer to the model after CPT as\nLlama-3-SynE (Synthetic data Enhanced Llama-3). We also present the tuning\nexperiments with a relatively small model -- TinyLlama, and employ the derived\nfindings to train the backbone model. Extensive experiments on a number of\nevaluation benchmarks show that our approach can largely improve the\nperformance of the backbone models, including both the general abilities (+8.81\non C-Eval and +6.31 on CMMLU) and the scientific reasoning abilities (+12.00 on\nMATH and +4.13 on SciEval), without hurting the original capacities. Our model,\ndata, and codes are available at https://github.com/RUC-GSAI/Llama-3-SynE.\n", "date": "2024-07-26", "categories": []}
{"id": "2308.12112", "title": "Category Adaptation Meets Projected Distillation in Generalized\n  Continual Category Discovery", "abstract": "  Generalized Continual Category Discovery (GCCD) tackles learning from\nsequentially arriving, partially labeled datasets while uncovering new\ncategories. Traditional methods depend on feature distillation to prevent\nforgetting the old knowledge. However, this strategy restricts the model's\nability to adapt and effectively distinguish new categories. To address this,\nwe introduce a novel technique integrating a learnable projector with feature\ndistillation, thus enhancing model adaptability without sacrificing past\nknowledge. The resulting distribution shift of the previously learned\ncategories is mitigated with the auxiliary category adaptation network. We\ndemonstrate that while each component offers modest benefits individually,\ntheir combination - dubbed CAMP (Category Adaptation Meets Projected\ndistillation) - significantly improves the balance between learning new\ninformation and retaining old. CAMP exhibits superior performance across\nseveral GCCD and Class Incremental Learning scenarios. The code is available at\nhttps://github.com/grypesc/CAMP.\n", "date": "2023-08-23", "categories": []}
{"id": "2407.17816", "title": "NC-NCD: Novel Class Discovery for Node Classification", "abstract": "  Novel Class Discovery (NCD) involves identifying new categories within\nunlabeled data by utilizing knowledge acquired from previously established\ncategories. However, existing NCD methods often struggle to maintain a balance\nbetween the performance of old and new categories. Discovering unlabeled new\ncategories in a class-incremental way is more practical but also more\nchallenging, as it is frequently hindered by either catastrophic forgetting of\nold categories or an inability to learn new ones. Furthermore, the\nimplementation of NCD on continuously scalable graph-structured data remains an\nunder-explored area. In response to these challenges, we introduce for the\nfirst time a more practical NCD scenario for node classification (i.e.,\nNC-NCD), and propose a novel self-training framework with prototype replay and\ndistillation called SWORD, adopted to our NC-NCD setting. Our approach enables\nthe model to cluster unlabeled new category nodes after learning labeled nodes\nwhile preserving performance on old categories without reliance on old category\nnodes. SWORD achieves this by employing a self-training strategy to learn new\ncategories and preventing the forgetting of old categories through the joint\nuse of feature prototypes and knowledge distillation. Extensive experiments on\nfour common benchmarks demonstrate the superiority of SWORD over other\nstate-of-the-art methods.\n", "date": "2024-07-25", "categories": []}
{"id": "2309.08546", "title": "Towards Robust Continual Learning with Bayesian Adaptive Moment\n  Regularization", "abstract": "  The pursuit of long-term autonomy mandates that machine learning models must\ncontinuously adapt to their changing environments and learn to solve new tasks.\nContinual learning seeks to overcome the challenge of catastrophic forgetting,\nwhere learning to solve new tasks causes a model to forget previously learnt\ninformation. Prior-based continual learning methods are appealing as they are\ncomputationally efficient and do not require auxiliary models or data storage.\nHowever, prior-based approaches typically fail on important benchmarks and are\nthus limited in their potential applications compared to their memory-based\ncounterparts. We introduce Bayesian adaptive moment regularization (BAdam), a\nnovel prior-based method that better constrains parameter growth, reducing\ncatastrophic forgetting. Our method boasts a range of desirable properties such\nas being lightweight and task label-free, converging quickly, and offering\ncalibrated uncertainty that is important for safe real-world deployment.\nResults show that BAdam achieves state-of-the-art performance for prior-based\nmethods on challenging single-headed class-incremental experiments such as\nSplit MNIST and Split FashionMNIST, and does so without relying on task labels\nor discrete task boundaries.\n", "date": "2023-09-15", "categories": []}
{"id": "2407.16354", "title": "Strike a Balance in Continual Panoptic Segmentation", "abstract": "  This study explores the emerging area of continual panoptic segmentation,\nhighlighting three key balances. First, we introduce past-class backtrace\ndistillation to balance the stability of existing knowledge with the\nadaptability to new information. This technique retraces the features\nassociated with past classes based on the final label assignment results,\nperforming knowledge distillation targeting these specific features from the\nprevious model while allowing other features to flexibly adapt to new\ninformation. Additionally, we introduce a class-proportional memory strategy,\nwhich aligns the class distribution in the replay sample set with that of the\nhistorical training data. This strategy maintains a balanced class\nrepresentation during replay, enhancing the utility of the limited-capacity\nreplay sample set in recalling prior classes. Moreover, recognizing that replay\nsamples are annotated only for the classes of their original step, we devise\nbalanced anti-misguidance losses, which combat the impact of incomplete\nannotations without incurring classification bias. Building upon these\ninnovations, we present a new method named Balanced Continual Panoptic\nSegmentation (BalConpas). Our evaluation on the challenging ADE20K dataset\ndemonstrates its superior performance compared to existing state-of-the-art\nmethods. The official code is available at\nhttps://github.com/jinpeng0528/BalConpas.\n", "date": "2024-07-23", "categories": []}
{"id": "2407.16611", "title": "Local vs Global continual learning", "abstract": "  Continual learning is the problem of integrating new information in a model\nwhile retaining the knowledge acquired in the past. Despite the tangible\nimprovements achieved in recent years, the problem of continual learning is\nstill an open one. A better understanding of the mechanisms behind the\nsuccesses and failures of existing continual learning algorithms can unlock the\ndevelopment of new successful strategies. In this work, we view continual\nlearning from the perspective of the multi-task loss approximation, and we\ncompare two alternative strategies, namely local and global approximations. We\nclassify existing continual learning algorithms based on the approximation\nused, and we assess the practical effects of this distinction in common\ncontinual learning settings.Additionally, we study optimal continual learning\nobjectives in the case of local polynomial approximations and we provide\nexamples of existing algorithms implementing the optimal objectives\n", "date": "2024-07-23", "categories": []}
{"id": "2311.12048", "title": "One Size Fits All for Semantic Shifts: Adaptive Prompt Tuning for\n  Continual Learning", "abstract": "  In real-world continual learning (CL) scenarios, tasks often exhibit\nintricate and unpredictable semantic shifts, posing challenges for fixed prompt\nmanagement strategies which are tailored to only handle semantic shifts of\nuniform degree (i.e., uniformly mild or uniformly abrupt). To address this\nlimitation, we propose an adaptive prompting approach that effectively\naccommodates semantic shifts of varying degree where mild and abrupt shifts are\nmixed. AdaPromptCL employs the assign-and-refine semantic grouping mechanism\nthat dynamically manages prompt groups in accordance with the semantic\nsimilarity between tasks, enhancing the quality of grouping through continuous\nrefinement. Our experiment results demonstrate that AdaPromptCL outperforms\nexisting prompting methods by up to 21.3%, especially in the benchmark datasets\nwith diverse semantic shifts between tasks.\n", "date": "2023-11-18", "categories": []}
{"id": "2403.05016", "title": "DiffClass: Diffusion-Based Class Incremental Learning", "abstract": "  Class Incremental Learning (CIL) is challenging due to catastrophic\nforgetting. On top of that, Exemplar-free Class Incremental Learning is even\nmore challenging due to forbidden access to previous task data. Recent\nexemplar-free CIL methods attempt to mitigate catastrophic forgetting by\nsynthesizing previous task data. However, they fail to overcome the\ncatastrophic forgetting due to the inability to deal with the significant\ndomain gap between real and synthetic data. To overcome these issues, we\npropose a novel exemplar-free CIL method. Our method adopts multi-distribution\nmatching (MDM) diffusion models to unify quality and bridge domain gaps among\nall domains of training data. Moreover, our approach integrates selective\nsynthetic image augmentation (SSIA) to expand the distribution of the training\ndata, thereby improving the model's plasticity and reinforcing the performance\nof our method's ultimate component, multi-domain adaptation (MDA). With the\nproposed integrations, our method then reformulates exemplar-free CIL into a\nmulti-domain adaptation problem to implicitly address the domain gap problem to\nenhance model stability during incremental training. Extensive experiments on\nbenchmark class incremental datasets and settings demonstrate that our method\nexcels previous exemplar-free CIL methods and achieves state-of-the-art\nperformance.\n", "date": "2024-03-08", "categories": []}
{"id": "2407.13768", "title": "Addressing Imbalance for Class Incremental Learning in Medical Image\n  Classification", "abstract": "  Deep convolutional neural networks have made significant breakthroughs in\nmedical image classification, under the assumption that training samples from\nall classes are simultaneously available. However, in real-world medical\nscenarios, there's a common need to continuously learn about new diseases,\nleading to the emerging field of class incremental learning (CIL) in the\nmedical domain. Typically, CIL suffers from catastrophic forgetting when\ntrained on new classes. This phenomenon is mainly caused by the imbalance\nbetween old and new classes, and it becomes even more challenging with\nimbalanced medical datasets. In this work, we introduce two simple yet\neffective plug-in methods to mitigate the adverse effects of the imbalance.\nFirst, we propose a CIL-balanced classification loss to mitigate the classifier\nbias toward majority classes via logit adjustment. Second, we propose a\ndistribution margin loss that not only alleviates the inter-class overlap in\nembedding space but also enforces the intra-class compactness. We evaluate the\neffectiveness of our method with extensive experiments on three benchmark\ndatasets (CCH5000, HAM10000, and EyePACS). The results demonstrate that our\napproach outperforms state-of-the-art methods.\n", "date": "2024-07-18", "categories": []}
{"id": "2407.14811", "title": "Decoupled Prompt-Adapter Tuning for Continual Activity Recognition", "abstract": "  Action recognition technology plays a vital role in enhancing security\nthrough surveillance systems, enabling better patient monitoring in healthcare,\nproviding in-depth performance analysis in sports, and facilitating seamless\nhuman-AI collaboration in domains such as manufacturing and assistive\ntechnologies. The dynamic nature of data in these areas underscores the need\nfor models that can continuously adapt to new video data without losing\npreviously acquired knowledge, highlighting the critical role of advanced\ncontinual action recognition. To address these challenges, we propose Decoupled\nPrompt-Adapter Tuning (DPAT), a novel framework that integrates adapters for\ncapturing spatial-temporal information and learnable prompts for mitigating\ncatastrophic forgetting through a decoupled training strategy. DPAT uniquely\nbalances the generalization benefits of prompt tuning with the plasticity\nprovided by adapters in pretrained vision models, effectively addressing the\nchallenge of maintaining model performance amidst continuous data evolution\nwithout necessitating extensive finetuning. DPAT consistently achieves\nstate-of-the-art performance across several challenging action recognition\nbenchmarks, thus demonstrating the effectiveness of our model in the domain of\ncontinual action recognition.\n", "date": "2024-07-20", "categories": []}
{"id": "2407.15429", "title": "Learning at a Glance: Towards Interpretable Data-limited Continual\n  Semantic Segmentation via Semantic-Invariance Modelling", "abstract": "  Continual semantic segmentation (CSS) based on incremental learning (IL) is a\ngreat endeavour in developing human-like segmentation models. However, current\nCSS approaches encounter challenges in the trade-off between preserving old\nknowledge and learning new ones, where they still need large-scale annotated\ndata for incremental training and lack interpretability. In this paper, we\npresent Learning at a Glance (LAG), an efficient, robust, human-like and\ninterpretable approach for CSS. Specifically, LAG is a simple and\nmodel-agnostic architecture, yet it achieves competitive CSS efficiency with\nlimited incremental data. Inspired by human-like recognition patterns, we\npropose a semantic-invariance modelling approach via semantic features\ndecoupling that simultaneously reconciles solid knowledge inheritance and\nnew-term learning. Concretely, the proposed decoupling manner includes two\nways, i.e., channel-wise decoupling and spatial-level neuron-relevant semantic\nconsistency. Our approach preserves semantic-invariant knowledge as solid\nprototypes to alleviate catastrophic forgetting, while also constraining\nsample-specific contents through an asymmetric contrastive learning method to\nenhance model robustness during IL steps. Experimental results in multiple\ndatasets validate the effectiveness of the proposed method. Furthermore, we\nintroduce a novel CSS protocol that better reflects realistic data-limited CSS\nsettings, and LAG achieves superior performance under multiple data-limited\nconditions.\n", "date": "2024-07-22", "categories": []}
{"id": "2402.09542", "title": "Layerwise Proximal Replay: A Proximal Point Method for Online Continual\n  Learning", "abstract": "  In online continual learning, a neural network incrementally learns from a\nnon-i.i.d. data stream. Nearly all online continual learning methods employ\nexperience replay to simultaneously prevent catastrophic forgetting and\nunderfitting on past data. Our work demonstrates a limitation of this approach:\nneural networks trained with experience replay tend to have unstable\noptimization trajectories, impeding their overall accuracy. Surprisingly, these\ninstabilities persist even when the replay buffer stores all previous training\nexamples, suggesting that this issue is orthogonal to catastrophic forgetting.\nWe minimize these instabilities through a simple modification of the\noptimization geometry. Our solution, Layerwise Proximal Replay (LPR), balances\nlearning from new and replay data while only allowing for gradual changes in\nthe hidden activation of past data. We demonstrate that LPR consistently\nimproves replay-based online continual learning methods across multiple problem\nsettings, regardless of the amount of available replay memory.\n", "date": "2024-02-14", "categories": []}
{"id": "2407.13974", "title": "Continual Learning for Remote Physiological Measurement: Minimize\n  Forgetting and Simplify Inference", "abstract": "  Remote photoplethysmography (rPPG) has gained significant attention in recent\nyears for its ability to extract physiological signals from facial videos.\nWhile existing rPPG measurement methods have shown satisfactory performance in\nintra-dataset and cross-dataset scenarios, they often overlook the incremental\nlearning scenario, where training data is presented sequentially, resulting in\nthe issue of catastrophic forgetting. Meanwhile, most existing class\nincremental learning approaches are unsuitable for rPPG measurement. In this\npaper, we present a novel method named ADDP to tackle continual learning for\nrPPG measurement. We first employ adapter to efficiently finetune the model on\nnew tasks. Then we design domain prototypes that are more applicable to rPPG\nsignal regression than commonly used class prototypes. Based on these\nprototypes, we propose a feature augmentation strategy to consolidate the past\nknowledge and an inference simplification strategy to convert potentially\nforgotten tasks into familiar ones for the model. To evaluate ADDP and enable\nfair comparisons, we create the first continual learning protocol for rPPG\nmeasurement. Comprehensive experiments demonstrate the effectiveness of our\nmethod for rPPG continual learning. Source code is available at\n\\url{https://github.com/MayYoY/rPPGDIL}\n", "date": "2024-07-19", "categories": []}
{"id": "2407.14029", "title": "PASS++: A Dual Bias Reduction Framework for Non-Exemplar\n  Class-Incremental Learning", "abstract": "  Class-incremental learning (CIL) aims to recognize new classes incrementally\nwhile maintaining the discriminability of old classes. Most existing CIL\nmethods are exemplar-based, i.e., storing a part of old data for retraining.\nWithout relearning old data, those methods suffer from catastrophic forgetting.\nIn this paper, we figure out two inherent problems in CIL, i.e., representation\nbias and classifier bias, that cause catastrophic forgetting of old knowledge.\nTo address these two biases, we present a simple and novel dual bias reduction\nframework that employs self-supervised transformation (SST) in input space and\nprototype augmentation (protoAug) in deep feature space. On the one hand, SST\nalleviates the representation bias by learning generic and diverse\nrepresentations that can transfer across different tasks. On the other hand,\nprotoAug overcomes the classifier bias by explicitly or implicitly augmenting\nprototypes of old classes in the deep feature space, which poses tighter\nconstraints to maintain previously learned decision boundaries. We further\npropose hardness-aware prototype augmentation and multi-view ensemble\nstrategies, leading to significant improvements. The proposed framework can be\neasily integrated with pre-trained models. Without storing any samples of old\nclasses, our method can perform comparably with state-of-the-art exemplar-based\napproaches which store plenty of old data. We hope to draw the attention of\nresearchers back to non-exemplar CIL by rethinking the necessity of storing old\nsamples in CIL.\n", "date": "2024-07-19", "categories": []}
{"id": "2407.14142", "title": "Early Preparation Pays Off: New Classifier Pre-tuning for Class\n  Incremental Semantic Segmentation", "abstract": "  Class incremental semantic segmentation aims to preserve old knowledge while\nlearning new tasks, however, it is impeded by catastrophic forgetting and\nbackground shift issues. Prior works indicate the pivotal importance of\ninitializing new classifiers and mainly focus on transferring knowledge from\nthe background classifier or preparing classifiers for future classes,\nneglecting the flexibility and variance of new classifiers. In this paper, we\npropose a new classifier pre-tuning~(NeST) method applied before the formal\ntraining process, learning a transformation from old classifiers to generate\nnew classifiers for initialization rather than directly tuning the parameters\nof new classifiers. Our method can make new classifiers align with the backbone\nand adapt to the new data, preventing drastic changes in the feature extractor\nwhen learning new classes. Besides, we design a strategy considering the\ncross-task class similarity to initialize matrices used in the transformation,\nhelping achieve the stability-plasticity trade-off. Experiments on Pascal VOC\n2012 and ADE20K datasets show that the proposed strategy can significantly\nimprove the performance of previous methods. The code is available at\n\\url{https://github.com/zhengyuan-xie/ECCV24_NeST}.\n", "date": "2024-07-19", "categories": []}
{"id": "2407.14143", "title": "Class-Incremental Learning with CLIP: Adaptive Representation Adjustment\n  and Parameter Fusion", "abstract": "  Class-incremental learning is a challenging problem, where the goal is to\ntrain a model that can classify data from an increasing number of classes over\ntime. With the advancement of vision-language pre-trained models such as CLIP,\nthey demonstrate good generalization ability that allows them to excel in\nclass-incremental learning with completely frozen parameters. However, further\nadaptation to downstream tasks by simply fine-tuning the model leads to severe\nforgetting. Most existing works with pre-trained models assume that the\nforgetting of old classes is uniform when the model acquires new knowledge. In\nthis paper, we propose a method named Adaptive Representation Adjustment and\nParameter Fusion (RAPF). During training for new data, we measure the influence\nof new classes on old ones and adjust the representations, using textual\nfeatures. After training, we employ a decomposed parameter fusion to further\nmitigate forgetting during adapter module fine-tuning. Experiments on several\nconventional benchmarks show that our method achieves state-of-the-art results.\nOur code is available at \\url{https://github.com/linlany/RAPF}.\n", "date": "2024-07-19", "categories": []}
{"id": "2407.14249", "title": "An Attention-based Representation Distillation Baseline for Multi-Label\n  Continual Learning", "abstract": "  The field of Continual Learning (CL) has inspired numerous researchers over\nthe years, leading to increasingly advanced countermeasures to the issue of\ncatastrophic forgetting. Most studies have focused on the single-class\nscenario, where each example comes with a single label. The recent literature\nhas successfully tackled such a setting, with impressive results. Differently,\nwe shift our attention to the multi-label scenario, as we feel it to be more\nrepresentative of real-world open problems. In our work, we show that existing\nstate-of-the-art CL methods fail to achieve satisfactory performance, thus\nquestioning the real advance claimed in recent years. Therefore, we assess both\nold-style and novel strategies and propose, on top of them, an approach called\nSelective Class Attention Distillation (SCAD). It relies on a knowledge\ntransfer technique that seeks to align the representations of the student\nnetwork -- which trains continuously and is subject to forgetting -- with the\nteacher ones, which is pretrained and kept frozen. Importantly, our method is\nable to selectively transfer the relevant information from the teacher to the\nstudent, thereby preventing irrelevant information from harming the student's\nperformance during online training. To demonstrate the merits of our approach,\nwe conduct experiments on two different multi-label datasets, showing that our\nmethod outperforms the current state-of-the-art Continual Learning methods. Our\nfindings highlight the importance of addressing the unique challenges posed by\nmulti-label environments in the field of Continual Learning. The code of SCAD\nis available at https://github.com/aimagelab/SCAD-LOD-2024.\n", "date": "2024-07-19", "categories": []}
{"id": "2306.03542", "title": "Masked Autoencoders are Efficient Continual Federated Learners", "abstract": "  Machine learning is typically framed from a perspective of i.i.d., and more\nimportantly, isolated data. In parts, federated learning lifts this assumption,\nas it sets out to solve the real-world challenge of collaboratively learning a\nshared model from data distributed across clients. However, motivated primarily\nby privacy and computational constraints, the fact that data may change,\ndistributions drift, or even tasks advance individually on clients, is seldom\ntaken into account. The field of continual learning addresses this separate\nchallenge and first steps have recently been taken to leverage synergies in\ndistributed supervised settings, in which several clients learn to solve\nchanging classification tasks over time without forgetting previously seen\nones. Motivated by these prior works, we posit that such federated continual\nlearning should be grounded in unsupervised learning of representations that\nare shared across clients; in the loose spirit of how humans can indirectly\nleverage others' experience without exposure to a specific task. For this\npurpose, we demonstrate that masked autoencoders for distribution estimation\nare particularly amenable to this setup. Specifically, their masking strategy\ncan be seamlessly integrated with task attention mechanisms to enable selective\nknowledge transfer between clients. We empirically corroborate the latter\nstatement through several continual federated scenarios on both image and\nbinary datasets.\n", "date": "2023-06-06", "categories": []}
{"id": "2403.06870", "title": "Semantic Residual Prompts for Continual Learning", "abstract": "  Prompt-tuning methods for Continual Learning (CL) freeze a large pre-trained\nmodel and train a few parameter vectors termed prompts. Most of these methods\norganize these vectors in a pool of key-value pairs and use the input image as\nquery to retrieve the prompts (values). However, as keys are learned while\ntasks progress, the prompting selection strategy is itself subject to\ncatastrophic forgetting, an issue often overlooked by existing approaches. For\ninstance, prompts introduced to accommodate new tasks might end up interfering\nwith previously learned prompts. To make the selection strategy more stable, we\nleverage a foundation model (CLIP) to select our prompts within a two-level\nadaptation mechanism. Specifically, the first level leverages a standard\ntextual prompt pool for the CLIP textual encoder, leading to stable class\nprototypes. The second level, instead, uses these prototypes along with the\nquery image as keys to index a second pool. The retrieved prompts serve to\nadapt a pre-trained ViT, granting plasticity. In doing so, we also propose a\nnovel residual mechanism to transfer CLIP semantics to the ViT layers. Through\nextensive analysis on established CL benchmarks, we show that our method\nsignificantly outperforms both state-of-the-art CL approaches and the zero-shot\nCLIP test. Notably, our findings hold true even for datasets with a substantial\ndomain gap w.r.t. the pre-training knowledge of the backbone model, as\nshowcased by experiments on satellite imagery and medical datasets. The\ncodebase is available at https://github.com/aimagelab/mammoth.\n", "date": "2024-03-11", "categories": []}
{"id": "2404.06859", "title": "Multi-Label Continual Learning for the Medical Domain: A Novel Benchmark", "abstract": "  Despite the critical importance of the medical domain in Deep Learning, most\nof the research in this area solely focuses on training models in static\nenvironments. It is only in recent years that research has begun to address\ndynamic environments and tackle the Catastrophic Forgetting problem through\nContinual Learning (CL) techniques. Previous studies have primarily focused on\nscenarios such as Domain Incremental Learning and Class Incremental Learning,\nwhich do not fully capture the complexity of real-world applications.\nTherefore, in this work, we propose a novel benchmark combining the challenges\nof new class arrivals and domain shifts in a single framework, by considering\nthe New Instances and New Classes (NIC) scenario. This benchmark aims to model\na realistic CL setting for the multi-label classification problem in medical\nimaging. Additionally, it encompasses a greater number of tasks compared to\npreviously tested scenarios. Specifically, our benchmark consists of two\ndatasets (NIH and CXP), nineteen classes, and seven tasks, a stream longer than\nthe previously tested ones. To solve common challenges (e.g., the task\ninference problem) found in the CIL and NIC scenarios, we propose a novel\napproach called Replay Consolidation with Label Propagation (RCLP). Our method\nsurpasses existing approaches, exhibiting superior performance with minimal\nforgetting.\n", "date": "2024-04-10", "categories": []}
{"id": "2407.05005", "title": "Personalized Federated Domain-Incremental Learning based on Adaptive\n  Knowledge Matching", "abstract": "  This paper focuses on Federated Domain-Incremental Learning (FDIL) where each\nclient continues to learn incremental tasks where their domain shifts from each\nother. We propose a novel adaptive knowledge matching-based personalized FDIL\napproach (pFedDIL) which allows each client to alternatively utilize\nappropriate incremental task learning strategy on the correlation with the\nknowledge from previous tasks. More specifically, when a new task arrives, each\nclient first calculates its local correlations with previous tasks. Then, the\nclient can choose to adopt a new initial model or a previous model with similar\nknowledge to train the new task and simultaneously migrate knowledge from\nprevious tasks based on these correlations. Furthermore, to identify the\ncorrelations between the new task and previous tasks for each client, we\nseparately employ an auxiliary classifier to each target classification model\nand propose sharing partial parameters between the target classification model\nand the auxiliary classifier to condense model parameters. We conduct extensive\nexperiments on several datasets of which results demonstrate that pFedDIL\noutperforms state-of-the-art methods by up to 14.35\\% in terms of average\naccuracy of all tasks.\n", "date": "2024-07-06", "categories": []}
{"id": "2407.09367", "title": "Reshaping the Online Data Buffering and Organizing Mechanism for\n  Continual Test-Time Adaptation", "abstract": "  Continual Test-Time Adaptation (CTTA) involves adapting a pre-trained source\nmodel to continually changing unsupervised target domains. In this paper, we\nsystematically analyze the challenges of this task: online environment,\nunsupervised nature, and the risks of error accumulation and catastrophic\nforgetting under continual domain shifts. To address these challenges, we\nreshape the online data buffering and organizing mechanism for CTTA. We propose\nan uncertainty-aware buffering approach to identify and aggregate significant\nsamples with high certainty from the unsupervised, single-pass data stream.\nBased on this, we propose a graph-based class relation preservation constraint\nto overcome catastrophic forgetting. Furthermore, a pseudo-target replay\nobjective is used to mitigate error accumulation. Extensive experiments\ndemonstrate the superiority of our method in both segmentation and\nclassification CTTA tasks. Code is available at https://github.com/z1358/OBAO.\n", "date": "2024-07-12", "categories": []}
{"id": "2311.13321", "title": "Revisiting Supervision for Continual Representation Learning", "abstract": "  In the field of continual learning, models are designed to learn tasks one\nafter the other. While most research has centered on supervised continual\nlearning, there is a growing interest in unsupervised continual learning, which\nmakes use of the vast amounts of unlabeled data. Recent studies have\nhighlighted the strengths of unsupervised methods, particularly self-supervised\nlearning, in providing robust representations. The improved transferability of\nthose representations built with self-supervised methods is often associated\nwith the role played by the multi-layer perceptron projector. In this work, we\ndepart from this observation and reexamine the role of supervision in continual\nrepresentation learning. We reckon that additional information, such as human\nannotations, should not deteriorate the quality of representations. Our\nfindings show that supervised models when enhanced with a multi-layer\nperceptron head, can outperform self-supervised models in continual\nrepresentation learning. This highlights the importance of the multi-layer\nperceptron projector in shaping feature transferability across a sequence of\ntasks in continual learning. The code is available on github:\nhttps://github.com/danielm1405/sl-vs-ssl-cl.\n", "date": "2023-11-22", "categories": []}
{"id": "2403.09296", "title": "Select and Distill: Selective Dual-Teacher Knowledge Transfer for\n  Continual Learning on Vision-Language Models", "abstract": "  Large-scale vision-language models (VLMs) have shown a strong zero-shot\ngeneralization capability on unseen-domain data. However, adapting pre-trained\nVLMs to a sequence of downstream tasks often leads to the forgetting of\npreviously learned knowledge and a reduction in zero-shot classification\nperformance. To tackle this problem, we propose a unique Selective Dual-Teacher\nKnowledge Transfer framework that leverages the most recent fine-tuned and the\noriginal pre-trained VLMs as dual teachers to preserve the previously learned\nknowledge and zero-shot capabilities, respectively. With only access to an\nunlabeled reference dataset, our proposed framework performs a selective\nknowledge distillation mechanism by measuring the feature discrepancy from the\ndual-teacher VLMs. Consequently, our selective dual-teacher knowledge\ndistillation mitigates catastrophic forgetting of previously learned knowledge\nwhile preserving the zero-shot capabilities of pre-trained VLMs. Extensive\nexperiments on benchmark datasets demonstrate that our framework is favorable\nagainst state-of-the-art continual learning approaches for preventing\ncatastrophic forgetting and zero-shot degradation. Project page:\nhttps://chuyu.org/research/snd\n", "date": "2024-03-14", "categories": []}
{"id": "2403.09857", "title": "Few-Shot Class Incremental Learning with Attention-Aware Self-Adaptive\n  Prompt", "abstract": "  Few-Shot Class-Incremental Learning (FSCIL) models aim to incrementally learn\nnew classes with scarce samples while preserving knowledge of old ones.\nExisting FSCIL methods usually fine-tune the entire backbone, leading to\noverfitting and hindering the potential to learn new classes. On the other\nhand, recent prompt-based CIL approaches alleviate forgetting by training\nprompts with sufficient data in each task. In this work, we propose a novel\nframework named Attention-aware Self-adaptive Prompt (ASP). ASP encourages\ntask-invariant prompts to capture shared knowledge by reducing specific\ninformation from the attention aspect. Additionally, self-adaptive\ntask-specific prompts in ASP provide specific information and transfer\nknowledge from old classes to new classes with an Information Bottleneck\nlearning objective. In summary, ASP prevents overfitting on base task and does\nnot require enormous data in few-shot incremental tasks. Extensive experiments\non three benchmark datasets validate that ASP consistently outperforms\nstate-of-the-art FSCIL and prompt-based CIL methods in terms of both learning\nnew classes and mitigating forgetting.\n", "date": "2024-03-14", "categories": []}
{"id": "2405.13383", "title": "Gradient Projection For Continual Parameter-Efficient Tuning", "abstract": "  Parameter-efficient tunings (PETs) have demonstrated impressive performance\nand promising perspectives in training large models, while they are still\nconfronted with a common problem: the trade-off between learning new content\nand protecting old knowledge, leading to zero-shot generalization collapse, and\ncross-modal hallucination. In this paper, we reformulate Adapter, LoRA,\nPrefix-tuning, and Prompt-tuning from the perspective of gradient projection,\nand firstly propose a unified framework called Parameter Efficient Gradient\nProjection (PEGP). We introduce orthogonal gradient projection into different\nPET paradigms and theoretically demonstrate that the orthogonal condition for\nthe gradient can effectively resist forgetting even for large-scale models. It\ntherefore modifies the gradient towards the direction that has less impact on\nthe old feature space, with less extra memory space and training time. We\nextensively evaluate our method with different backbones, including ViT and\nCLIP, on diverse datasets, and experiments comprehensively demonstrate its\nefficiency in reducing forgetting in class, online class, domain, task, and\nmulti-modality continual settings. The project page is available at\nhttps://dmcv-ecnu-pegp.github.io/.\n", "date": "2024-05-22", "categories": []}
{"id": "2407.12188", "title": "CroMo-Mixup: Augmenting Cross-Model Representations for Continual\n  Self-Supervised Learning", "abstract": "  Continual self-supervised learning (CSSL) learns a series of tasks\nsequentially on the unlabeled data. Two main challenges of continual learning\nare catastrophic forgetting and task confusion. While CSSL problem has been\nstudied to address the catastrophic forgetting challenge, little work has been\ndone to address the task confusion aspect. In this work, we show through\nextensive experiments that self-supervised learning (SSL) can make CSSL more\nsusceptible to the task confusion problem, particularly in less diverse\nsettings of class incremental learning because different classes belonging to\ndifferent tasks are not trained concurrently. Motivated by this challenge, we\npresent a novel cross-model feature Mixup (CroMo-Mixup) framework that\naddresses this issue through two key components: 1) Cross-Task data Mixup,\nwhich mixes samples across tasks to enhance negative sample diversity; and 2)\nCross-Model feature Mixup, which learns similarities between embeddings\nobtained from current and old models of the mixed sample and the original\nimages, facilitating cross-task class contrast learning and old knowledge\nretrieval. We evaluate the effectiveness of CroMo-Mixup to improve both Task-ID\nprediction and average linear accuracy across all tasks on three datasets,\nCIFAR10, CIFAR100, and tinyImageNet under different class-incremental learning\nsettings. We validate the compatibility of CroMo-Mixup on four state-of-the-art\nSSL objectives. Code is available at\n\\url{https://github.com/ErumMushtaq/CroMo-Mixup}.\n", "date": "2024-07-16", "categories": []}
{"id": "2407.12279", "title": "ER-FSL: Experience Replay with Feature Subspace Learning for Online\n  Continual Learning", "abstract": "  Online continual learning (OCL) involves deep neural networks retaining\nknowledge from old data while adapting to new data, which is accessible only\nonce. A critical challenge in OCL is catastrophic forgetting, reflected in\nreduced model performance on old data. Existing replay-based methods mitigate\nforgetting by replaying buffered samples from old data and learning current\nsamples of new data. In this work, we dissect existing methods and empirically\ndiscover that learning and replaying in the same feature space is not conducive\nto addressing the forgetting issue. Since the learned features associated with\nold data are readily changed by the features related to new data due to data\nimbalance, leading to the forgetting problem. Based on this observation, we\nintuitively explore learning and replaying in different feature spaces.\nLearning in a feature subspace is sufficient to capture novel knowledge from\nnew data while replaying in a larger feature space provides more feature space\nto maintain historical knowledge from old data. To this end, we propose a novel\nOCL approach called experience replay with feature subspace learning (ER-FSL).\nFirstly, ER-FSL divides the entire feature space into multiple subspaces, with\neach subspace used to learn current samples. Moreover, it introduces a subspace\nreuse mechanism to address situations where no blank subspaces exist. Secondly,\nER-FSL replays previous samples using an accumulated space comprising all\nlearned subspaces. Extensive experiments on three datasets demonstrate the\nsuperiority of ER-FSL over various state-of-the-art methods.\n", "date": "2024-07-17", "categories": []}
{"id": "2407.12375", "title": "FETCH: A Memory-Efficient Replay Approach for Continual Learning in\n  Image Classification", "abstract": "  Class-incremental continual learning is an important area of research, as\nstatic deep learning methods fail to adapt to changing tasks and data\ndistributions. In previous works, promising results were achieved using replay\nand compressed replay techniques. In the field of regular replay, GDumb\nachieved outstanding results but requires a large amount of memory. This\nproblem can be addressed by compressed replay techniques. The goal of this work\nis to evaluate compressed replay in the pipeline of GDumb. We propose FETCH, a\ntwo-stage compression approach. First, the samples from the continual\ndatastream are encoded by the early layers of a pre-trained neural network.\nSecond, the samples are compressed before being stored in the episodic memory.\nFollowing GDumb, the remaining classification head is trained from scratch\nusing only the decompressed samples from the reply memory. We evaluate FETCH in\ndifferent scenarios and show that this approach can increase accuracy on\nCIFAR10 and CIFAR100. In our experiments, simple compression methods (e.g.,\nquantization of tensors) outperform deep autoencoders. In the future, FETCH\ncould serve as a baseline for benchmarking compressed replay learning in\nconstrained memory scenarios.\n", "date": "2024-07-17", "categories": []}
{"id": "2407.12470", "title": "Continual Learning for Temporal-Sensitive Question Answering", "abstract": "  In this study, we explore an emerging research area of Continual Learning for\nTemporal Sensitive Question Answering (CLTSQA). Previous research has primarily\nfocused on Temporal Sensitive Question Answering (TSQA), often overlooking the\nunpredictable nature of future events. In real-world applications, it's crucial\nfor models to continually acquire knowledge over time, rather than relying on a\nstatic, complete dataset. Our paper investigates strategies that enable models\nto adapt to the ever-evolving information landscape, thereby addressing the\nchallenges inherent in CLTSQA. To support our research, we first create a novel\ndataset, divided into five subsets, designed specifically for various stages of\ncontinual learning. We then propose a training framework for CLTSQA that\nintegrates temporal memory replay and temporal contrastive learning. Our\nexperimental results highlight two significant insights: First, the CLTSQA task\nintroduces unique challenges for existing models. Second, our proposed\nframework effectively navigates these challenges, resulting in improved\nperformance.\n", "date": "2024-07-17", "categories": []}
{"id": "2407.12724", "title": "An Evaluation of Continual Learning for Advanced Node Semiconductor\n  Defect Inspection", "abstract": "  Deep learning-based semiconductor defect inspection has gained traction in\nrecent years, offering a powerful and versatile approach that provides high\naccuracy, adaptability, and efficiency in detecting and classifying nano-scale\ndefects. However, semiconductor manufacturing processes are continually\nevolving, leading to the emergence of new types of defects over time. This\npresents a significant challenge for conventional supervised defect detectors,\nas they may suffer from catastrophic forgetting when trained on new defect\ndatasets, potentially compromising performance on previously learned tasks. An\nalternative approach involves the constant storage of previously trained\ndatasets alongside pre-trained model versions, which can be utilized for\n(re-)training from scratch or fine-tuning whenever encountering a new defect\ndataset. However, adhering to such a storage template is impractical in terms\nof size, particularly when considering High-Volume Manufacturing (HVM).\nAdditionally, semiconductor defect datasets, especially those encompassing\nstochastic defects, are often limited and expensive to obtain, thus lacking\nsufficient representation of the entire universal set of defectivity. This work\nintroduces a task-agnostic, meta-learning approach aimed at addressing this\nchallenge, which enables the incremental addition of new defect classes and\nscales to create a more robust and generalized model for semiconductor defect\ninspection. We have benchmarked our approach using real resist-wafer SEM\n(Scanning Electron Microscopy) datasets for two process steps, ADI and AEI,\ndemonstrating its superior performance compared to conventional supervised\ntraining methods.\n", "date": "2024-07-17", "categories": []}
{"id": "2301.03345", "title": "Latent Spectral Regularization for Continual Learning", "abstract": "  While biological intelligence grows organically as new knowledge is gathered\nthroughout life, Artificial Neural Networks forget catastrophically whenever\nthey face a changing training data distribution. Rehearsal-based Continual\nLearning (CL) approaches have been established as a versatile and reliable\nsolution to overcome this limitation; however, sudden input disruptions and\nmemory constraints are known to alter the consistency of their predictions. We\nstudy this phenomenon by investigating the geometric characteristics of the\nlearner's latent space and find that replayed data points of different classes\nincreasingly mix up, interfering with classification. Hence, we propose a\ngeometric regularizer that enforces weak requirements on the Laplacian spectrum\nof the latent space, promoting a partitioning behavior. Our proposal, called\nContinual Spectral Regularizer for Incremental Learning (CaSpeR-IL), can be\neasily combined with any rehearsal-based CL approach and improves the\nperformance of SOTA methods on standard benchmarks.\n", "date": "2023-01-09", "categories": []}
{"id": "2401.16650", "title": "Augmenting Replay in World Models for Continual Reinforcement Learning", "abstract": "  Continual RL requires an agent to learn new tasks without forgetting previous\nones, while improving on both past and future tasks. The most common approaches\nuse model-free algorithms and replay buffers can help to mitigate catastrophic\nforgetting, but often struggle with scalability due to large memory\nrequirements. Biologically inspired replay suggests replay to a world model,\naligning with model-based RL; as opposed to the common setting of replay in\nmodel-free algorithms. Model-based RL offers benefits for continual RL by\nleveraging knowledge of the environment, independent of policy. We introduce\nWMAR (World Models with Augmented Replay), a model-based RL algorithm with a\nmemory-efficient distribution-matching replay buffer. WMAR extends the well\nknown DreamerV3 algorithm, which employs a simple FIFO buffer and was not\ntested in continual RL. We evaluated WMAR and DreamerV3, with the same-size\nreplay buffers. They were tested on two scenarios: tasks with shared structure\nusing OpenAI Procgen and tasks without shared structure using the Atari\nbenchmark. WMAR demonstrated favourable properties for continual RL considering\nmetrics for forgetting as well as skill transfer on past and future tasks.\nCompared to DreamerV3, WMAR showed slight benefits in tasks with shared\nstructure and substantially better forgetting characteristics on tasks without\nshared structure. Our results suggest that model-based RL with a\nmemory-efficient replay buffer can be an effective approach to continual RL,\njustifying further research.\n", "date": "2024-01-30", "categories": []}
{"id": "2407.11078", "title": "Overcoming Catastrophic Forgetting in Federated Class-Incremental\n  Learning via Federated Global Twin Generator", "abstract": "  Federated Class-Incremental Learning (FCIL) increasingly becomes important in\nthe decentralized setting, where it enables multiple participants to\ncollaboratively train a global model to perform well on a sequence of tasks\nwithout sharing their private data. In FCIL, conventional Federated Learning\nalgorithms such as FedAVG often suffer from catastrophic forgetting, resulting\nin significant performance declines on earlier tasks. Recent works, based on\ngenerative models, produce synthetic images to help mitigate this issue across\nall classes, but these approaches' testing accuracy on previous classes is\nstill much lower than recent classes, i.e., having better plasticity than\nstability. To overcome these issues, this paper presents Federated Global Twin\nGenerator (FedGTG), an FCIL framework that exploits privacy-preserving\ngenerative-model training on the global side without accessing client data.\nSpecifically, the server trains a data generator and a feature generator to\ncreate two types of information from all seen classes, and then it sends the\nsynthetic data to the client side. The clients then use\nfeature-direction-controlling losses to make the local models retain knowledge\nand learn new tasks well. We extensively analyze the robustness of FedGTG on\nnatural images, as well as its ability to converge to flat local minima and\nachieve better-predicting confidence (calibration). Experimental results on\nCIFAR-10, CIFAR-100, and tiny-ImageNet demonstrate the improvements in accuracy\nand forgetting measures of FedGTG compared to previous frameworks.\n", "date": "2024-07-13", "categories": []}
{"id": "2302.03648", "title": "Class-Incremental Learning: A Survey", "abstract": "  Deep models, e.g., CNNs and Vision Transformers, have achieved impressive\nachievements in many vision tasks in the closed world. However, novel classes\nemerge from time to time in our ever-changing world, requiring a learning\nsystem to acquire new knowledge continually. Class-Incremental Learning (CIL)\nenables the learner to incorporate the knowledge of new classes incrementally\nand build a universal classifier among all seen classes. Correspondingly, when\ndirectly training the model with new class instances, a fatal problem occurs --\nthe model tends to catastrophically forget the characteristics of former ones,\nand its performance drastically degrades. There have been numerous efforts to\ntackle catastrophic forgetting in the machine learning community. In this\npaper, we survey comprehensively recent advances in class-incremental learning\nand summarize these methods from several aspects. We also provide a rigorous\nand unified evaluation of 17 methods in benchmark image classification tasks to\nfind out the characteristics of different algorithms empirically. Furthermore,\nwe notice that the current comparison protocol ignores the influence of memory\nbudget in model storage, which may result in unfair comparison and biased\nresults. Hence, we advocate fair comparison by aligning the memory budget in\nevaluation, as well as several memory-agnostic performance measures. The source\ncode is available at https://github.com/zhoudw-zdw/CIL_Survey/\n", "date": "2023-02-07", "categories": []}
{"id": "2309.17105", "title": "Continual Action Assessment via Task-Consistent Score-Discriminative\n  Feature Distribution Modeling", "abstract": "  Action Quality Assessment (AQA) is a task that tries to answer how well an\naction is carried out. While remarkable progress has been achieved, existing\nworks on AQA assume that all the training data are visible for training at one\ntime, but do not enable continual learning on assessing new technical actions.\nIn this work, we address such a Continual Learning problem in AQA\n(Continual-AQA), which urges a unified model to learn AQA tasks sequentially\nwithout forgetting. Our idea for modeling Continual-AQA is to sequentially\nlearn a task-consistent score-discriminative feature distribution, in which the\nlatent features express a strong correlation with the score labels regardless\nof the task or action types.From this perspective, we aim to mitigate the\nforgetting in Continual-AQA from two aspects. Firstly, to fuse the features of\nnew and previous data into a score-discriminative distribution, a novel\nFeature-Score Correlation-Aware Rehearsal is proposed to store and reuse data\nfrom previous tasks with limited memory size. Secondly, an Action\nGeneral-Specific Graph is developed to learn and decouple the action-general\nand action-specific knowledge so that the task-consistent score-discriminative\nfeatures can be better extracted across various tasks. Extensive experiments\nare conducted to evaluate the contributions of proposed components. The\ncomparisons with the existing continual learning methods additionally verify\nthe effectiveness and versatility of our approach. Data and code are available\nat https://github.com/iSEE-Laboratory/Continual-AQA.\n", "date": "2023-09-29", "categories": []}
{"id": "2312.01203", "title": "Harnessing Discrete Representations For Continual Reinforcement Learning", "abstract": "  Reinforcement learning (RL) agents make decisions using nothing but\nobservations from the environment, and consequently, heavily rely on the\nrepresentations of those observations. Though some recent breakthroughs have\nused vector-based categorical representations of observations, often referred\nto as discrete representations, there is little work explicitly assessing the\nsignificance of such a choice. In this work, we provide a thorough empirical\ninvestigation of the advantages of representing observations as vectors of\ncategorical values within the context of reinforcement learning. We perform\nevaluations on world-model learning, model-free RL, and ultimately continual RL\nproblems, where the benefits best align with the needs of the problem setting.\nWe find that, when compared to traditional continuous representations, world\nmodels learned over discrete representations accurately model more of the world\nwith less capacity, and that agents trained with discrete representations learn\nbetter policies with less data. In the context of continual RL, these benefits\ntranslate into faster adapting agents. Additionally, our analysis suggests that\nthe observed performance improvements can be attributed to the information\ncontained within the latent vectors and potentially the encoding of the\ndiscrete representation itself.\n", "date": "2023-12-02", "categories": []}
{"id": "2401.02094", "title": "PILoRA: Prototype Guided Incremental LoRA for Federated\n  Class-Incremental Learning", "abstract": "  Existing federated learning methods have effectively dealt with decentralized\nlearning in scenarios involving data privacy and non-IID data. However, in\nreal-world situations, each client dynamically learns new classes, requiring\nthe global model to classify all seen classes. To effectively mitigate\ncatastrophic forgetting and data heterogeneity under low communication costs,\nwe propose a simple and effective method named PILoRA. On the one hand, we\nadopt prototype learning to learn better feature representations and leverage\nthe heuristic information between prototypes and class features to design a\nprototype re-weight module to solve the classifier bias caused by data\nheterogeneity without retraining the classifier. On the other hand, we view\nincremental learning as the process of learning distinct task vectors and\nencoding them within different LoRA parameters. Accordingly, we propose\nIncremental LoRA to mitigate catastrophic forgetting. Experimental results on\nstandard datasets indicate that our method outperforms the state-of-the-art\napproaches significantly. More importantly, our method exhibits strong\nrobustness and superiority in different settings and degrees of data\nheterogeneity. The code is available at\n\\url{https://github.com/Ghy0501/PILoRA}.\n", "date": "2024-01-04", "categories": []}
{"id": "2403.14797", "title": "Preventing Catastrophic Forgetting through Memory Networks in Continuous\n  Detection", "abstract": "  Modern pre-trained architectures struggle to retain previous information\nwhile undergoing continuous fine-tuning on new tasks. Despite notable progress\nin continual classification, systems designed for complex vision tasks such as\ndetection or segmentation still struggle to attain satisfactory performance. In\nthis work, we introduce a memory-based detection transformer architecture to\nadapt a pre-trained DETR-style detector to new tasks while preserving knowledge\nfrom previous tasks. We propose a novel localized query function for efficient\ninformation retrieval from memory units, aiming to minimize forgetting.\nFurthermore, we identify a fundamental challenge in continual detection\nreferred to as background relegation. This arises when object categories from\nearlier tasks reappear in future tasks, potentially without labels, leading\nthem to be implicitly treated as background. This is an inevitable issue in\ncontinual detection or segmentation. The introduced continual optimization\ntechnique effectively tackles this challenge. Finally, we assess the\nperformance of our proposed system on continual detection benchmarks and\ndemonstrate that our approach surpasses the performance of existing\nstate-of-the-art resulting in 5-7% improvements on MS-COCO and PASCAL-VOC on\nthe task of continual detection.\n", "date": "2024-03-21", "categories": []}
{"id": "2407.09642", "title": "Seq-to-Final: A Benchmark for Tuning from Sequential Distributions to a\n  Final Time Point", "abstract": "  Distribution shift over time occurs in many settings. Leveraging historical\ndata is necessary to learn a model for the last time point when limited data is\navailable in the final period, yet few methods have been developed specifically\nfor this purpose. In this work, we construct a benchmark with different\nsequences of synthetic shifts to evaluate the effectiveness of 3 classes of\nmethods that 1) learn from all data without adapting to the final period, 2)\nlearn from historical data with no regard to the sequential nature and then\nadapt to the final period, and 3) leverage the sequential nature of historical\ndata when tailoring a model to the final period. We call this benchmark\nSeq-to-Final to highlight the focus on using a sequence of time periods to\nlearn a model for the final time point. Our synthetic benchmark allows users to\nconstruct sequences with different types of shift and compare different\nmethods. We focus on image classification tasks using CIFAR-10 and CIFAR-100 as\nthe base images for the synthetic sequences. We also evaluate the same methods\non the Portraits dataset to explore the relevance to real-world shifts over\ntime. Finally, we create a visualization to contrast the initializations and\nupdates from different methods at the final time step. Our results suggest\nthat, for the sequences in our benchmark, methods that disregard the sequential\nstructure and adapt to the final time point tend to perform well. The\napproaches we evaluate that leverage the sequential nature do not offer any\nimprovement. We hope that this benchmark will inspire the development of new\nalgorithms that are better at leveraging sequential historical data or a deeper\nunderstanding of why methods that disregard the sequential nature are able to\nperform well.\n", "date": "2024-07-12", "categories": []}
{"id": "2407.09838", "title": "Background Adaptation with Residual Modeling for Exemplar-Free\n  Class-Incremental Semantic Segmentation", "abstract": "  Class Incremental Semantic Segmentation~(CISS), within Incremental Learning\nfor semantic segmentation, targets segmenting new categories while reducing the\ncatastrophic forgetting on the old categories.Besides, background shifting,\nwhere the background category changes constantly in each step, is a special\nchallenge for CISS. Current methods with a shared background classifier\nstruggle to keep up with these changes, leading to decreased stability in\nbackground predictions and reduced accuracy of segmentation. For this special\nchallenge, we designed a novel background adaptation mechanism, which\nexplicitly models the background residual rather than the background itself in\neach step, and aggregates these residuals to represent the evolving background.\nTherefore, the background adaptation mechanism ensures the stability of\nprevious background classifiers, while enabling the model to concentrate on the\neasy-learned residuals from the additional channel, which enhances background\ndiscernment for better prediction of novel categories. To precisely optimize\nthe background adaptation mechanism, we propose Pseudo Background Binary\nCross-Entropy loss and Background Adaptation losses, which amplify the\nadaptation effect. Group Knowledge Distillation and Background Feature\nDistillation strategies are designed to prevent forgetting old categories. Our\napproach, evaluated across various incremental scenarios on Pascal VOC 2012 and\nADE20K datasets, outperforms prior exemplar-free state-of-the-art methods with\nmIoU of 3.0% in VOC 10-1 and 2.0% in ADE 100-5, notably enhancing the accuracy\nof new classes while mitigating catastrophic forgetting. Code is available in\nhttps://andyzaq.github.io/barmsite/.\n", "date": "2024-07-13", "categories": []}
{"id": "2407.10281", "title": "Beyond Prompt Learning: Continual Adapter for Efficient Rehearsal-Free\n  Continual Learning", "abstract": "  The problem of Rehearsal-Free Continual Learning (RFCL) aims to continually\nlearn new knowledge while preventing forgetting of the old knowledge, without\nstoring any old samples and prototypes. The latest methods leverage large-scale\npre-trained models as the backbone and use key-query matching to generate\ntrainable prompts to learn new knowledge. However, the domain gap between the\npre-training dataset and the downstream datasets can easily lead to\ninaccuracies in key-query matching prompt selection when directly generating\nqueries using the pre-trained model, which hampers learning new knowledge.\nThus, in this paper, we propose a beyond prompt learning approach to the RFCL\ntask, called Continual Adapter (C-ADA). It mainly comprises a\nparameter-extensible continual adapter layer (CAL) and a scaling and shifting\n(S&S) module in parallel with the pre-trained model. C-ADA flexibly extends\nspecific weights in CAL to learn new knowledge for each task and freezes old\nweights to preserve prior knowledge, thereby avoiding matching errors and\noperational inefficiencies introduced by key-query matching. To reduce the gap,\nC-ADA employs an S&S module to transfer the feature space from pre-trained\ndatasets to downstream datasets. Moreover, we propose an orthogonal loss to\nmitigate the interaction between old and new knowledge. Our approach achieves\nsignificantly improved performance and training speed, outperforming the\ncurrent state-of-the-art (SOTA) method. Additionally, we conduct experiments on\ndomain-incremental learning, surpassing the SOTA, and demonstrating the\ngenerality of our approach in different settings.\n", "date": "2024-07-14", "categories": []}
{"id": "2407.10758", "title": "Continual Deep Learning on the Edge via Stochastic Local Competition\n  among Subnetworks", "abstract": "  Continual learning on edge devices poses unique challenges due to stringent\nresource constraints. This paper introduces a novel method that leverages\nstochastic competition principles to promote sparsity, significantly reducing\ndeep network memory footprint and computational demand. Specifically, we\npropose deep networks that comprise blocks of units that compete locally to win\nthe representation of each arising new task; competition takes place in a\nstochastic manner. This type of network organization results in sparse\ntask-specific representations from each network layer; the sparsity pattern is\nobtained during training and is different among tasks. Crucially, our method\nsparsifies both the weights and the weight gradients, thus facilitating\ntraining on edge devices. This is performed on the grounds of winning\nprobability for each unit in a block. During inference, the network retains\nonly the winning unit and zeroes-out all weights pertaining to non-winning\nunits for the task at hand. Thus, our approach is specifically tailored for\ndeployment on edge devices, providing an efficient and scalable solution for\ncontinual learning in resource-limited environments.\n", "date": "2024-07-15", "categories": []}
{"id": "2407.08153", "title": "Lifelong Histopathology Whole Slide Image Retrieval via Distance\n  Consistency Rehearsal", "abstract": "  Content-based histopathological image retrieval (CBHIR) has gained attention\nin recent years, offering the capability to return histopathology images that\nare content-wise similar to the query one from an established database.\nHowever, in clinical practice, the continuously expanding size of WSI databases\nlimits the practical application of the current CBHIR methods. In this paper,\nwe propose a Lifelong Whole Slide Retrieval (LWSR) framework to address the\nchallenges of catastrophic forgetting by progressive model updating on\ncontinuously growing retrieval database. Our framework aims to achieve the\nbalance between stability and plasticity during continuous learning. To\npreserve system plasticity, we utilize local memory bank with reservoir\nsampling method to save instances, which can comprehensively encompass the\nfeature spaces of both old and new tasks. Furthermore, A distance consistency\nrehearsal (DCR) module is designed to ensure the retrieval queue's consistency\nfor previous tasks, which is regarded as stability within a lifelong CBHIR\nsystem. We evaluated the proposed method on four public WSI datasets from TCGA\nprojects. The experimental results have demonstrated the proposed method is\neffective and is superior to the state-of-the-art methods.\n", "date": "2024-07-11", "categories": []}
{"id": "2407.09039", "title": "Overcoming Catastrophic Forgetting in Tabular Data Classification: A\n  Pseudorehearsal-based approach", "abstract": "  Continual learning (CL) poses the important challenge of adapting to evolving\ndata distributions without forgetting previously acquired knowledge while\nconsolidating new knowledge. In this paper, we introduce a new methodology,\ncoined as Tabular-data Rehearsal-based Incremental Lifelong Learning framework\n(TRIL3), designed to address the phenomenon of catastrophic forgetting in\ntabular data classification problems. TRIL3 uses the prototype-based\nincremental generative model XuILVQ to generate synthetic data to preserve old\nknowledge and the DNDF algorithm, which was modified to run in an incremental\nway, to learn classification tasks for tabular data, without storing old\nsamples. After different tests to obtain the adequate percentage of synthetic\ndata and to compare TRIL3 with other CL available proposals, we can conclude\nthat the performance of TRIL3 outstands other options in the literature using\nonly 50% of synthetic data.\n", "date": "2024-07-12", "categories": []}
{"id": "2407.09047", "title": "Cs2K: Class-specific and Class-shared Knowledge Guidance for Incremental\n  Semantic Segmentation", "abstract": "  Incremental semantic segmentation endeavors to segment newly encountered\nclasses while maintaining knowledge of old classes. However, existing methods\neither 1) lack guidance from class-specific knowledge (i.e., old class\nprototypes), leading to a bias towards new classes, or 2) constrain\nclass-shared knowledge (i.e., old model weights) excessively without\ndiscrimination, resulting in a preference for old classes. In this paper, to\ntrade off model performance, we propose the Class-specific and Class-shared\nKnowledge (Cs2K) guidance for incremental semantic segmentation. Specifically,\nfrom the class-specific knowledge aspect, we design a prototype-guided pseudo\nlabeling that exploits feature proximity from prototypes to correct pseudo\nlabels, thereby overcoming catastrophic forgetting. Meanwhile, we develop a\nprototype-guided class adaptation that aligns class distribution across\ndatasets via learning old augmented prototypes. Moreover, from the class-shared\nknowledge aspect, we propose a weight-guided selective consolidation to\nstrengthen old memory while maintaining new memory by integrating old and new\nmodel weights based on weight importance relative to old classes. Experiments\non public datasets demonstrate that our proposed Cs2K significantly improves\nsegmentation performance and is plug-and-play.\n", "date": "2024-07-12", "categories": []}
{"id": "2310.02206", "title": "Chunking: Continual Learning is not just about Distribution Shift", "abstract": "  Work on continual learning (CL) has thus far largely focused on the problems\narising from shifts in the data distribution. However, CL can be decomposed\ninto two sub-problems: (a) shifts in the data distribution, and (b) dealing\nwith the fact that the data is split into chunks and so only a part of the data\nis available to be trained on at any point in time. In this work, we look at\nthe latter sub-problem, the chunking of data. We show that chunking is an\nimportant part of CL, accounting for around half of the performance drop from\noffline learning in our experiments. Furthermore, our results reveal that\ncurrent CL algorithms do not address the chunking sub-problem, only performing\nas well as plain SGD training when there is no shift in the data distribution.\nTherefore, we show that chunking is both an important and currently unaddressed\nsub-problem and until it is addressed CL methods will be capped in performance.\nAdditionally, we analyse why performance drops when learning occurs on\nidentically distributed chunks of data, and find that forgetting, which is\noften seen to be a problem due to distribution shift, still arises and is a\nsignificant problem. We also show that performance on the chunking sub-problem\ncan be increased and that this performance transfers to the full CL setting,\nwhere there is distribution shift. Hence, we argue that work on chunking can\nhelp advance CL in general.\n", "date": "2023-10-03", "categories": []}
{"id": "2405.09858", "title": "Towards Realistic Incremental Scenario in Class Incremental Semantic\n  Segmentation", "abstract": "  This paper addresses the unrealistic aspect of the commonly adopted\nContinuous Incremental Semantic Segmentation (CISS) scenario, termed\noverlapped. We point out that overlapped allows the same image to reappear in\nfuture tasks with different pixel labels, which is far from practical\nincremental learning scenarios. Moreover, we identified that this flawed\nscenario may lead to biased results for two commonly used techniques in CISS,\npseudo-labeling and exemplar memory, resulting in unintended advantages or\ndisadvantages for certain techniques. To mitigate this, a practical scenario\ncalled partitioned is proposed, in which the dataset is first divided into\ndistinct subsets representing each class, and then the subsets are assigned to\neach corresponding task. This efficiently addresses the issue above while\nmeeting the requirement of CISS scenario, such as capturing the background\nshifts. Furthermore, we identify and address the code implementation issues\nrelated to retrieving data from the exemplar memory, which was ignored in\nprevious works. Lastly, we introduce a simple yet competitive memory-based\nbaseline, MiB-AugM, that handles background shifts of current tasks in the\nexemplar memory. This baseline achieves state-of-the-art results across\nmultiple tasks involving learning numerous new classes.\n", "date": "2024-05-16", "categories": []}
{"id": "2407.08041", "title": "TACLE: Task and Class-aware Exemplar-free Semi-supervised Class\n  Incremental Learning", "abstract": "  We propose a novel TACLE (TAsk and CLass-awarE) framework to address the\nrelatively unexplored and challenging problem of exemplar-free semi-supervised\nclass incremental learning. In this scenario, at each new task, the model has\nto learn new classes from both (few) labeled and unlabeled data without access\nto exemplars from previous classes. In addition to leveraging the capabilities\nof pre-trained models, TACLE proposes a novel task-adaptive threshold, thereby\nmaximizing the utilization of the available unlabeled data as incremental\nlearning progresses. Additionally, to enhance the performance of the\nunder-represented classes within each task, we propose a class-aware weighted\ncross-entropy loss. We also exploit the unlabeled data for classifier\nalignment, which further enhances the model performance. Extensive experiments\non benchmark datasets, namely CIFAR10, CIFAR100, and ImageNet-Subset100\ndemonstrate the effectiveness of the proposed TACLE framework. We further\nshowcase its effectiveness when the unlabeled data is imbalanced and also for\nthe extreme case of one labeled example per class.\n", "date": "2024-07-10", "categories": []}
{"id": "2407.08214", "title": "Towards stable training of parallel continual learning", "abstract": "  Parallel Continual Learning (PCL) tasks investigate the training methods for\ncontinual learning with multi-source input, where data from different tasks are\nlearned as they arrive. PCL offers high training efficiency and is well-suited\nfor complex multi-source data systems, such as autonomous vehicles equipped\nwith multiple sensors. However, at any time, multiple tasks need to be trained\nsimultaneously, leading to severe training instability in PCL. This instability\nmanifests during both forward and backward propagation, where features are\nentangled and gradients are conflict. This paper introduces Stable Parallel\nContinual Learning (SPCL), a novel approach that enhances the training\nstability of PCL for both forward and backward propagation. For the forward\npropagation, we apply Doubly-block Toeplit (DBT) Matrix based orthogonality\nconstraints to network parameters to ensure stable and consistent propagation.\nFor the backward propagation, we employ orthogonal decomposition for gradient\nmanagement stabilizes backpropagation and mitigates gradient conflicts across\ntasks. By optimizing gradients by ensuring orthogonality and minimizing the\ncondition number, SPCL effectively stabilizing the gradient descent in complex\noptimization tasks. Experimental results demonstrate that SPCL outperforms\nstate-of-the-art methjods and achieve better training stability.\n", "date": "2024-07-11", "categories": []}
{"id": "2407.08411", "title": "CLEO: Continual Learning of Evolving Ontologies", "abstract": "  Continual learning (CL) addresses the problem of catastrophic forgetting in\nneural networks, which occurs when a trained model tends to overwrite\npreviously learned information, when presented with a new task. CL aims to\ninstill the lifelong learning characteristic of humans in intelligent systems,\nmaking them capable of learning continuously while retaining what was already\nlearned. Current CL problems involve either learning new domains\n(domain-incremental) or new and previously unseen classes (class-incremental).\nHowever, general learning processes are not just limited to learning\ninformation, but also refinement of existing information. In this paper, we\ndefine CLEO - Continual Learning of Evolving Ontologies, as a new incremental\nlearning setting under CL to tackle evolving classes. CLEO is motivated by the\nneed for intelligent systems to adapt to real-world ontologies that change over\ntime, such as those in autonomous driving. We use Cityscapes, PASCAL VOC, and\nMapillary Vistas to define the task settings and demonstrate the applicability\nof CLEO. We highlight the shortcomings of existing CIL methods in adapting to\nCLEO and propose a baseline solution, called Modelling Ontologies (MoOn). CLEO\nis a promising new approach to CL that addresses the challenge of evolving\nontologies in real-world applications. MoOn surpasses previous CL approaches in\nthe context of CLEO.\n", "date": "2024-07-11", "categories": []}
{"id": "2407.08536", "title": "Exemplar-free Continual Representation Learning via Learnable Drift\n  Compensation", "abstract": "  Exemplar-free class-incremental learning using a backbone trained from\nscratch and starting from a small first task presents a significant challenge\nfor continual representation learning. Prototype-based approaches, when\ncontinually updated, face the critical issue of semantic drift due to which the\nold class prototypes drift to different positions in the new feature space.\nThrough an analysis of prototype-based continual learning, we show that\nforgetting is not due to diminished discriminative power of the feature\nextractor, and can potentially be corrected by drift compensation. To address\nthis, we propose Learnable Drift Compensation (LDC), which can effectively\nmitigate drift in any moving backbone, whether supervised or unsupervised. LDC\nis fast and straightforward to integrate on top of existing continual learning\napproaches. Furthermore, we showcase how LDC can be applied in combination with\nself-supervised CL methods, resulting in the first exemplar-free\nsemi-supervised continual learning approach. We achieve state-of-the-art\nperformance in both supervised and semi-supervised settings across multiple\ndatasets. Code is available at \\url{https://github.com/alviur/ldc}.\n", "date": "2024-07-11", "categories": []}
{"id": "2407.08609", "title": "BiasPruner: Debiased Continual Learning for Medical Image Classification", "abstract": "  Continual Learning (CL) is crucial for enabling networks to dynamically adapt\nas they learn new tasks sequentially, accommodating new data and classes\nwithout catastrophic forgetting. Diverging from conventional perspectives on\nCL, our paper introduces a new perspective wherein forgetting could actually\nbenefit the sequential learning paradigm. Specifically, we present BiasPruner,\na CL framework that intentionally forgets spurious correlations in the training\ndata that could lead to shortcut learning. Utilizing a new bias score that\nmeasures the contribution of each unit in the network to learning spurious\nfeatures, BiasPruner prunes those units with the highest bias scores to form a\ndebiased subnetwork preserved for a given task. As BiasPruner learns a new\ntask, it constructs a new debiased subnetwork, potentially incorporating units\nfrom previous subnetworks, which improves adaptation and performance on the new\ntask. During inference, BiasPruner employs a simple task-agnostic approach to\nselect the best debiased subnetwork for predictions. We conduct experiments on\nthree medical datasets for skin lesion classification and chest X-Ray\nclassification and demonstrate that BiasPruner consistently outperforms SOTA CL\nmethods in terms of classification performance and fairness. Our code is\navailable here.\n", "date": "2024-07-11", "categories": []}
{"id": "2303.07223", "title": "PromptFusion: Decoupling Stability and Plasticity for Continual Learning", "abstract": "  Current research on continual learning mainly focuses on relieving\ncatastrophic forgetting, and most of their success is at the cost of limiting\nthe performance of newly incoming tasks. Such a trade-off is referred to as the\nstability-plasticity dilemma and is a more general and challenging problem for\ncontinual learning. However, the inherent conflict between these two concepts\nmakes it seemingly impossible to devise a satisfactory solution to both of them\nsimultaneously. Therefore, we ask, \"is it possible to divide them into two\nseparate problems to conquer them independently?\". To this end, we propose a\nprompt-tuning-based method termed PromptFusion to enable the decoupling of\nstability and plasticity. Specifically, PromptFusion consists of a carefully\ndesigned \\stab module that deals with catastrophic forgetting and a \\boo module\nto learn new knowledge concurrently. Furthermore, to address the computational\noverhead brought by the additional architecture, we propose PromptFusion-Lite\nwhich improves PromptFusion by dynamically determining whether to activate both\nmodules for each input image. Extensive experiments show that both PromptFusion\nand PromptFusion-Lite achieve promising results on popular continual learning\ndatasets for class-incremental and domain-incremental settings. Especially on\nSplit-Imagenet-R, one of the most challenging datasets for class-incremental\nlearning, our method can exceed state-of-the-art prompt-based methods by more\nthan 5\\% in accuracy, with PromptFusion-Lite using 14.8\\% less computational\nresources than PromptFusion.\n", "date": "2023-03-13", "categories": []}
{"id": "2312.14439", "title": "PUMA: Efficient Continual Graph Learning for Node Classification with\n  Graph Condensation", "abstract": "  When handling streaming graphs, existing graph representation learning models\nencounter a catastrophic forgetting problem, where previously learned knowledge\nof these models is easily overwritten when learning with newly incoming graphs.\nIn response, Continual Graph Learning (CGL) emerges as a novel paradigm\nenabling graph representation learning from streaming graphs. Our prior work,\nCondense and Train (CaT) is a replay-based CGL framework with a balanced\ncontinual learning procedure, which designs a small yet effective memory bankn\nfor replaying. Although the CaT alleviates the catastrophic forgetting problem,\nthere exist three issues: (1) The graph condensation only focuses on labelled\nnodes while neglecting abundant information carried by unlabelled nodes; (2)\nThe continual training scheme of the CaT overemphasises on the previously\nlearned knowledge, limiting the model capacity to learn from newly added\nmemories; (3) Both the condensation process and replaying process of the CaT\nare time-consuming. In this paper, we propose a PsUdo-label guided Memory bAnk\n(PUMA) CGL framework, extending from the CaT to enhance its efficiency and\neffectiveness by overcoming the above-mentioned weaknesses and limits. To fully\nexploit the information in a graph, PUMA expands the coverage of nodes during\ngraph condensation with both labelled and unlabelled nodes. Furthermore, a\ntraining-from-scratch strategy is proposed to upgrade the previous continual\nlearning scheme for a balanced training between the historical and the new\ngraphs. Besides, PUMA uses a one-time prorogation and wide graph encoders to\naccelerate the graph condensation and the graph encoding process in the\ntraining stage to improve the efficiency of the whole framework. Extensive\nexperiments on six datasets for the node classification task demonstrate the\nstate-of-the-art performance and efficiency over existing methods.\n", "date": "2023-12-22", "categories": []}
{"id": "2407.07133", "title": "Neuromimetic metaplasticity for adaptive continual learning", "abstract": "  Conventional intelligent systems based on deep neural network (DNN) models\nencounter challenges in achieving human-like continual learning due to\ncatastrophic forgetting. Here, we propose a metaplasticity model inspired by\nhuman working memory, enabling DNNs to perform catastrophic forgetting-free\ncontinual learning without any pre- or post-processing. A key aspect of our\napproach involves implementing distinct types of synapses from stable to\nflexible, and randomly intermixing them to train synaptic connections with\ndifferent degrees of flexibility. This strategy allowed the network to\nsuccessfully learn a continuous stream of information, even under unexpected\nchanges in input length. The model achieved a balanced tradeoff between memory\ncapacity and performance without requiring additional training or structural\nmodifications, dynamically allocating memory resources to retain both old and\nnew information. Furthermore, the model demonstrated robustness against data\npoisoning attacks by selectively filtering out erroneous memories, leveraging\nthe Hebb repetition effect to reinforce the retention of significant data.\n", "date": "2024-07-09", "categories": []}
{"id": "2407.07468", "title": "Rethinking Few-shot Class-incremental Learning: Learning from Yourself", "abstract": "  Few-shot class-incremental learning (FSCIL) aims to learn sequential classes\nwith limited samples in a few-shot fashion. Inherited from the classical\nclass-incremental learning setting, the popular benchmark of FSCIL uses\naveraged accuracy (aAcc) and last-task averaged accuracy (lAcc) as the\nevaluation metrics. However, we reveal that such evaluation metrics may not\nprovide adequate emphasis on the novel class performance, and the continual\nlearning ability of FSCIL methods could be ignored under this benchmark. In\nthis work, as a complement to existing metrics, we offer a new metric called\ngeneralized average accuracy (gAcc) which is designed to provide an extra\nequitable evaluation by incorporating different perspectives of the performance\nunder the guidance of a parameter $\\alpha$. We also present an overall metric\nin the form of the area under the curve (AUC) along the $\\alpha$. Under the\nguidance of gAcc, we release the potential of intermediate features of the\nvision transformers to boost the novel-class performance. Taking information\nfrom intermediate layers which are less class-specific and more generalizable,\nwe manage to rectify the final features, leading to a more generalizable\ntransformer-based FSCIL framework. Without complex network designs or\ncumbersome training procedures, our method outperforms existing FSCIL methods\nat aAcc and gAcc on three datasets. See codes at\nhttps://github.com/iSEE-Laboratory/Revisting_FSCIL\n", "date": "2024-07-10", "categories": []}
{"id": "2209.07863", "title": "Expanding continual few-shot learning benchmarks to include recognition\n  of specific instances", "abstract": "  Continual learning and few-shot learning are important frontiers in progress\ntoward broader Machine Learning (ML) capabilities. Recently, there has been\nintense interest in combining both. One of the first examples to do so was the\nContinual few-shot Learning (CFSL) framework of Antoniou et al.\narXiv:2004.11967. In this study, we extend CFSL in two ways that capture a\nbroader range of challenges, important for intelligent agent behaviour in\nreal-world conditions. First, we increased the number of classes by an order of\nmagnitude, making the results more comparable to standard continual learning\nexperiments. Second, we introduced an 'instance test' which requires\nrecognition of specific instances of classes -- a capability of animal\ncognition that is usually neglected in ML. For an initial exploration of ML\nmodel performance under these conditions, we selected representative baseline\nmodels from the original CFSL work and added a model variant with replay. As\nexpected, learning more classes is more difficult than the original CFSL\nexperiments, and interestingly, the way in which image instances and classes\nare presented affects classification performance. Surprisingly, accuracy in the\nbaseline instance test is comparable to other classification tasks, but poor\ngiven significant occlusion and noise. The use of replay for consolidation\nsubstantially improves performance for both types of tasks, but particularly\nfor the instance test.\n", "date": "2022-08-26", "categories": []}
{"id": "2302.03534", "title": "On the Limitation and Experience Replay for GNNs in Continual Learning", "abstract": "  Continual learning seeks to empower models to progressively acquire\ninformation from a sequence of tasks. This approach is crucial for many\nreal-world systems, which are dynamic and evolve over time. Recent research has\nwitnessed a surge in the exploration of Graph Neural Networks (GNN) in\nNode-wise Graph Continual Learning (NGCL), a practical yet challenging paradigm\ninvolving the continual training of a GNN on node-related tasks. Despite recent\nadvancements in continual learning strategies for GNNs in NGCL, a thorough\ntheoretical understanding, especially regarding its learnability, is lacking.\nLearnability concerns the existence of a learning algorithm that can produce a\ngood candidate model from the hypothesis/weight space, which is crucial for\nmodel selection in NGCL development. This paper introduces the first\ntheoretical exploration of the learnability of GNN in NGCL, revealing that\nlearnability is heavily influenced by structural shifts due to the\ninterconnected nature of graph data. Specifically, GNNs may not be viable for\nNGCL under significant structural changes, emphasizing the need to manage\nstructural shifts. To mitigate the impact of structural shifts, we propose a\nnovel experience replay method termed Structure-Evolution-Aware Experience\nReplay (SEA-ER). SEA-ER features an innovative experience selection strategy\nthat capitalizes on the topological awareness of GNNs, alongside a unique\nreplay strategy that employs structural alignment to effectively counter\ncatastrophic forgetting and diminish the impact of structural shifts on GNNs in\nNGCL. Our extensive experiments validate our theoretical insights and the\neffectiveness of SEA-ER.\n", "date": "2023-02-07", "categories": []}
{"id": "2304.03980", "title": "Continual Learning for LiDAR Semantic Segmentation: Class-Incremental\n  and Coarse-to-Fine strategies on Sparse Data", "abstract": "  During the last few years, continual learning (CL) strategies for image\nclassification and segmentation have been widely investigated designing\ninnovative solutions to tackle catastrophic forgetting, like knowledge\ndistillation and self-inpainting. However, the application of continual\nlearning paradigms to point clouds is still unexplored and investigation is\nrequired, especially using architectures that capture the sparsity and uneven\ndistribution of LiDAR data. The current paper analyzes the problem of class\nincremental learning applied to point cloud semantic segmentation, comparing\napproaches and state-of-the-art architectures. To the best of our knowledge,\nthis is the first example of class-incremental continual learning for LiDAR\npoint cloud semantic segmentation. Different CL strategies were adapted to\nLiDAR point clouds and tested, tackling both classic fine-tuning scenarios and\nthe Coarse-to-Fine learning paradigm. The framework has been evaluated through\ntwo different architectures on SemanticKITTI, obtaining results in line with\nstate-of-the-art CL strategies and standard offline learning.\n", "date": "2023-04-08", "categories": []}
{"id": "2310.08815", "title": "Incremental Object Detection with CLIP", "abstract": "  In contrast to the incremental classification task, the incremental detection\ntask is characterized by the presence of data ambiguity, as an image may have\ndifferently labeled bounding boxes across multiple continuous learning stages.\nThis phenomenon often impairs the model's ability to effectively learn new\nclasses. However, existing research has paid less attention to the forward\ncompatibility of the model, which limits its suitability for incremental\nlearning. To overcome this obstacle, we propose leveraging a visual-language\nmodel such as CLIP to generate text feature embeddings for different class\nsets, which enhances the feature space globally. We then employ super-classes\nto replace the unavailable novel classes in the early learning stage to\nsimulate the incremental scenario. Finally, we utilize the CLIP image encoder\nto accurately identify potential objects. We incorporate the finely recognized\ndetection boxes as pseudo-annotations into the training process, thereby\nfurther improving the detection performance. We evaluate our approach on\nvarious incremental learning settings using the PASCAL VOC 2007 dataset, and\nour approach outperforms state-of-the-art methods, particularly for recognizing\nthe new classes.\n", "date": "2023-10-13", "categories": []}
{"id": "2310.13888", "title": "Towards a General Framework for Continual Learning with Pre-training", "abstract": "  In this work, we present a general framework for continual learning of\nsequentially arrived tasks with the use of pre-training, which has emerged as a\npromising direction for artificial intelligence systems to accommodate\nreal-world dynamics. From a theoretical perspective, we decompose its objective\ninto three hierarchical components, including within-task prediction,\ntask-identity inference, and task-adaptive prediction. Then we propose an\ninnovative approach to explicitly optimize these components with\nparameter-efficient fine-tuning (PEFT) techniques and representation\nstatistics. We empirically demonstrate the superiority and generality of our\napproach in downstream continual learning, and further explore the\napplicability of PEFT techniques in upstream continual learning. We also\ndiscuss the biological basis of the proposed framework with recent advances in\nneuroscience.\n", "date": "2023-10-21", "categories": []}
{"id": "2407.06045", "title": "OpenCIL: Benchmarking Out-of-Distribution Detection in Class-Incremental\n  Learning", "abstract": "  Class incremental learning (CIL) aims to learn a model that can not only\nincrementally accommodate new classes, but also maintain the learned knowledge\nof old classes. Out-of-distribution (OOD) detection in CIL is to retain this\nincremental learning ability, while being able to reject unknown samples that\nare drawn from different distributions of the learned classes. This capability\nis crucial to the safety of deploying CIL models in open worlds. However,\ndespite remarkable advancements in the respective CIL and OOD detection, there\nlacks a systematic and large-scale benchmark to assess the capability of\nadvanced CIL models in detecting OOD samples. To fill this gap, in this study\nwe design a comprehensive empirical study to establish such a benchmark, named\n$\\textbf{OpenCIL}$. To this end, we propose two principled frameworks for\nenabling four representative CIL models with 15 diverse OOD detection methods,\nresulting in 60 baseline models for OOD detection in CIL. The empirical\nevaluation is performed on two popular CIL datasets with six commonly-used OOD\ndatasets. One key observation we find through our comprehensive evaluation is\nthat the CIL models can be severely biased towards the OOD samples and newly\nadded classes when they are exposed to open environments. Motivated by this, we\nfurther propose a new baseline for OOD detection in CIL, namely Bi-directional\nEnergy Regularization ($\\textbf{BER}$), which is specially designed to mitigate\nthese two biases in different CIL models by having energy regularization on\nboth old and new classes. Its superior performance is justified in our\nexperiments. All codes and datasets are open-source at\nhttps://github.com/mala-lab/OpenCIL.\n", "date": "2024-07-08", "categories": []}
{"id": "2407.06697", "title": "Certified Continual Learning for Neural Network Regression", "abstract": "  On the one hand, there has been considerable progress on neural network\nverification in recent years, which makes certifying neural networks a\npossibility. On the other hand, neural networks in practice are often\nre-trained over time to cope with new data distribution or for solving\ndifferent tasks (a.k.a. continual learning). Once re-trained, the verified\ncorrectness of the neural network is likely broken, particularly in the\npresence of the phenomenon known as catastrophic forgetting. In this work, we\npropose an approach called certified continual learning which improves existing\ncontinual learning methods by preserving, as long as possible, the established\ncorrectness properties of a verified network. Our approach is evaluated with\nmultiple neural networks and on two different continual learning methods. The\nresults show that our approach is efficient and the trained models preserve\ntheir certified correctness and often maintain high utility.\n", "date": "2024-07-09", "categories": []}
{"id": "2407.04996", "title": "The Solution for the sequential task continual learning track of the 2nd\n  Greater Bay Area International Algorithm Competition", "abstract": "  This paper presents a data-free, parameter-isolation-based continual learning\nalgorithm we developed for the sequential task continual learning track of the\n2nd Greater Bay Area International Algorithm Competition. The method learns an\nindependent parameter subspace for each task within the network's convolutional\nand linear layers and freezes the batch normalization layers after the first\ntask. Specifically, for domain incremental setting where all domains share a\nclassification head, we freeze the shared classification head after first task\nis completed, effectively solving the issue of catastrophic forgetting.\nAdditionally, facing the challenge of domain incremental settings without\nproviding a task identity, we designed an inference task identity strategy,\nselecting an appropriate mask matrix for each sample. Furthermore, we\nintroduced a gradient supplementation strategy to enhance the importance of\nunselected parameters for the current task, facilitating learning for new\ntasks. We also implemented an adaptive importance scoring strategy that\ndynamically adjusts the amount of parameters to optimize single-task\nperformance while reducing parameter usage. Moreover, considering the\nlimitations of storage space and inference time, we designed a mask matrix\ncompression strategy to save storage space and improve the speed of encryption\nand decryption of the mask matrix. Our approach does not require expanding the\ncore network or using external auxiliary networks or data, and performs well\nunder both task incremental and domain incremental settings. This solution\nultimately won a second-place prize in the competition.\n", "date": "2024-07-06", "categories": []}
{"id": "2407.05342", "title": "Mind the Interference: Retaining Pre-trained Knowledge in Parameter\n  Efficient Continual Learning of Vision-Language Models", "abstract": "  This study addresses the Domain-Class Incremental Learning problem, a\nrealistic but challenging continual learning scenario where both the domain\ndistribution and target classes vary across tasks. To handle these diverse\ntasks, pre-trained Vision-Language Models (VLMs) are introduced for their\nstrong generalizability. However, this incurs a new problem: the knowledge\nencoded in the pre-trained VLMs may be disturbed when adapting to new tasks,\ncompromising their inherent zero-shot ability. Existing methods tackle it by\ntuning VLMs with knowledge distillation on extra datasets, which demands heavy\ncomputation overhead. To address this problem efficiently, we propose the\nDistribution-aware Interference-free Knowledge Integration (DIKI) framework,\nretaining pre-trained knowledge of VLMs from a perspective of avoiding\ninformation interference. Specifically, we design a fully residual mechanism to\ninfuse newly learned knowledge into a frozen backbone, while introducing\nminimal adverse impacts on pre-trained knowledge. Besides, this residual\nproperty enables our distribution-aware integration calibration scheme,\nexplicitly controlling the information implantation process for test data from\nunseen distributions. Experiments demonstrate that our DIKI surpasses the\ncurrent state-of-the-art approach using only 0.86% of the trained parameters\nand requiring substantially less training time. Code is available at:\nhttps://github.com/lloongx/DIKI .\n", "date": "2024-07-07", "categories": []}
{"id": "2210.15701", "title": "Do Pre-trained Models Benefit Equally in Continual Learning?", "abstract": "  Existing work on continual learning (CL) is primarily devoted to developing\nalgorithms for models trained from scratch. Despite their encouraging\nperformance on contrived benchmarks, these algorithms show dramatic performance\ndrops in real-world scenarios. Therefore, this paper advocates the systematic\nintroduction of pre-training to CL, which is a general recipe for transferring\nknowledge to downstream tasks but is substantially missing in the CL community.\nOur investigation reveals the multifaceted complexity of exploiting pre-trained\nmodels for CL, along three different axes, pre-trained models, CL algorithms,\nand CL scenarios. Perhaps most intriguingly, improvements in CL algorithms from\npre-training are very inconsistent an underperforming algorithm could become\ncompetitive and even state-of-the-art when all algorithms start from a\npre-trained model. This indicates that the current paradigm, where all CL\nmethods are compared in from-scratch training, is not well reflective of the\ntrue CL objective and desired progress. In addition, we make several other\nimportant observations, including that CL algorithms that exert less\nregularization benefit more from a pre-trained model; and that a stronger\npre-trained model such as CLIP does not guarantee a better improvement. Based\non these findings, we introduce a simple yet effective baseline that employs\nminimum regularization and leverages the more beneficial pre-trained model,\ncoupled with a two-stage training pipeline. We recommend including this strong\nbaseline in the future development of CL algorithms, due to its demonstrated\nstate-of-the-art performance.\n", "date": "2022-10-27", "categories": []}
{"id": "2306.00427", "title": "Out-of-distribution forgetting: vulnerability of continual learning to\n  intra-class distribution shift", "abstract": "  Continual learning (CL) is an important technique to allow artificial neural\nnetworks to work in open environments. CL enables a system to learn new tasks\nwithout severe interference to its performance on old tasks, i.e., overcome the\nproblems of catastrophic forgetting. In joint learning, it is well known that\nthe out-of-distribution (OOD) problem caused by intentional attacks or\nenvironmental perturbations will severely impair the ability of networks to\ngeneralize. In this work, we reported a special form of catastrophic forgetting\nraised by the OOD problem in continual learning settings, and we named it\nout-of-distribution forgetting (OODF). In continual image classification tasks,\nwe found that for a given category, introducing an intra-class distribution\nshift significantly impaired the recognition accuracy of CL methods for that\ncategory during subsequent learning. Interestingly, this phenomenon is special\nfor CL as the same level of distribution shift had only negligible effects in\nthe joint learning scenario. We verified that CL methods without dedicating\nsubnetworks for individual tasks are all vulnerable to OODF. Moreover, OODF\ndoes not depend on any specific way of shifting the distribution, suggesting it\nis a risk for CL in a wide range of circumstances. Taken together, our work\nidentified an under-attended risk during CL, highlighting the importance of\ndeveloping approaches that can overcome OODF. Code available:\n\\url{https://github.com/Hiroid/OODF}\n", "date": "2023-06-01", "categories": []}
{"id": "2312.08888", "title": "Read Between the Layers: Leveraging Multi-Layer Representations for\n  Rehearsal-Free Continual Learning with Pre-Trained Models", "abstract": "  We address the Continual Learning (CL) problem, wherein a model must learn a\nsequence of tasks from non-stationary distributions while preserving prior\nknowledge upon encountering new experiences. With the advancement of foundation\nmodels, CL research has pivoted from the initial learning-from-scratch paradigm\ntowards utilizing generic features from large-scale pre-training. However,\nexisting approaches to CL with pre-trained models primarily focus on separating\nclass-specific features from the final representation layer and neglect the\npotential of intermediate representations to capture low- and mid-level\nfeatures, which are more invariant to domain shifts. In this work, we propose\nLayUP, a new prototype-based approach to CL that leverages second-order feature\nstatistics from multiple intermediate layers of a pre-trained network. Our\nmethod is conceptually simple, does not require access to prior data, and works\nout of the box with any foundation model. LayUP surpasses the state of the art\nin four of the seven class-incremental learning benchmarks, all three\ndomain-incremental learning benchmarks and in six of the seven online continual\nlearning benchmarks, while significantly reducing memory and computational\nrequirements compared to existing baselines. Our results demonstrate that fully\nexhausting the representational capacities of pre-trained models in CL goes\nwell beyond their final embeddings.\n", "date": "2023-12-13", "categories": []}
{"id": "2407.03813", "title": "PECTP: Parameter-Efficient Cross-Task Prompts for Incremental Vision\n  Transformer", "abstract": "  Incremental Learning (IL) aims to learn deep models on sequential tasks\ncontinually, where each new task includes a batch of new classes and deep\nmodels have no access to task-ID information at the inference time. Recent vast\npre-trained models (PTMs) have achieved outstanding performance by prompt\ntechnique in practical IL without the old samples (rehearsal-free) and with a\nmemory constraint (memory-constrained): Prompt-extending and Prompt-fixed\nmethods. However, prompt-extending methods need a large memory buffer to\nmaintain an ever-expanding prompt pool and meet an extra challenging prompt\nselection problem. Prompt-fixed methods only learn a single set of prompts on\none of the incremental tasks and can not handle all the incremental tasks\neffectively. To achieve a good balance between the memory cost and the\nperformance on all the tasks, we propose a Parameter-Efficient Cross-Task\nPrompt (PECTP) framework with Prompt Retention Module (PRM) and classifier Head\nRetention Module (HRM). To make the final learned prompts effective on all\nincremental tasks, PRM constrains the evolution of cross-task prompts'\nparameters from Outer Prompt Granularity and Inner Prompt Granularity. Besides,\nwe employ HRM to inherit old knowledge in the previously learned classifier\nheads to facilitate the cross-task prompts' generalization ability. Extensive\nexperiments show the effectiveness of our method. The source codes will be\navailable at \\url{https://github.com/RAIAN08/PECTP}.\n", "date": "2024-07-04", "categories": []}
{"id": "2407.01704", "title": "Weight Clipping for Deep Continual and Reinforcement Learning", "abstract": "  Many failures in deep continual and reinforcement learning are associated\nwith increasing magnitudes of the weights, making them hard to change and\npotentially causing overfitting. While many methods address these learning\nfailures, they often change the optimizer or the architecture, a complexity\nthat hinders widespread adoption in various systems. In this paper, we focus on\nlearning failures that are associated with increasing weight norm and we\npropose a simple technique that can be easily added on top of existing learning\nsystems: clipping neural network weights to limit them to a specific range. We\nstudy the effectiveness of weight clipping in a series of supervised and\nreinforcement learning experiments. Our empirical results highlight the\nbenefits of weight clipping for generalization, addressing loss of plasticity\nand policy collapse, and facilitating learning with a large replay ratio.\n", "date": "2024-07-01", "categories": []}
{"id": "2407.02253", "title": "Parameter-Selective Continual Test-Time Adaptation", "abstract": "  Continual Test-Time Adaptation (CTTA) aims to adapt a pretrained model to\never-changing environments during the test time under continuous domain shifts.\nMost existing CTTA approaches are based on the Mean Teacher (MT) structure,\nwhich contains a student and a teacher model, where the student is updated\nusing the pseudo-labels from the teacher model, and the teacher is then updated\nby exponential moving average strategy. However, these methods update the MT\nmodel indiscriminately on all parameters of the model. That is, some critical\nparameters involving sharing knowledge across different domains may be erased,\nintensifying error accumulation and catastrophic forgetting. In this paper, we\nintroduce Parameter-Selective Mean Teacher (PSMT) method, which is capable of\neffectively updating the critical parameters within the MT network under domain\nshifts. First, we introduce a selective distillation mechanism in the student\nmodel, which utilizes past knowledge to regularize novel knowledge, thereby\nmitigating the impact of error accumulation. Second, to avoid catastrophic\nforgetting, in the teacher model, we create a mask through Fisher information\nto selectively update parameters via exponential moving average, with\npreservation measures applied to crucial parameters. Extensive experimental\nresults verify that PSMT outperforms state-of-the-art methods across multiple\nbenchmark datasets. Our code is available at\n\\url{https://github.com/JiaxuTian/PSMT}.\n", "date": "2024-07-02", "categories": []}
{"id": "2303.13113", "title": "AdaCL:Adaptive Continual Learning", "abstract": "  Class-Incremental Learning aims to update a deep classifier to learn new\ncategories while maintaining or improving its accuracy on previously observed\nclasses. Common methods to prevent forgetting previously learned classes\ninclude regularizing the neural network updates and storing exemplars in\nmemory, which come with hyperparameters such as the learning rate,\nregularization strength, or the number of exemplars. However, these\nhyperparameters are usually only tuned at the start and then kept fixed\nthroughout the learning sessions, ignoring the fact that newly encountered\ntasks may have varying levels of novelty or difficulty. This study investigates\nthe necessity of hyperparameter `adaptivity' in Class-Incremental Learning: the\nability to dynamically adjust hyperparameters such as the learning rate,\nregularization strength, and memory size according to the properties of the new\ntask at hand. We propose AdaCL, a Bayesian Optimization-based approach to\nautomatically and efficiently determine the optimal values for those parameters\nwith each learning task. We show that adapting hyperpararmeters on each new\ntask leads to improvement in accuracy, forgetting and memory. Code is available\nat https://github.com/ElifCerenGokYildirim/AdaCL.\n", "date": "2023-03-23", "categories": []}
{"id": "2401.13200", "title": "Topology-aware Embedding Memory for Continual Learning on Expanding\n  Networks", "abstract": "  Memory replay based techniques have shown great success for continual\nlearning with incrementally accumulated Euclidean data. Directly applying them\nto continually expanding networks, however, leads to the potential memory\nexplosion problem due to the need to buffer representative nodes and their\nassociated topological neighborhood structures. To this end, we systematically\nanalyze the key challenges in the memory explosion problem, and present a\ngeneral framework, \\textit{i.e.}, Parameter Decoupled Graph Neural Networks\n(PDGNNs) with Topology-aware Embedding Memory (TEM), to tackle this issue. The\nproposed framework not only reduces the memory space complexity from\n$\\mathcal{O}(nd^L)$ to $\\mathcal{O}(n)$~\\footnote{$n$: memory budget, $d$:\naverage node degree, $L$: the radius of the GNN receptive field}, but also\nfully utilizes the topological information for memory replay. Specifically,\nPDGNNs decouple trainable parameters from the computation ego-subnetwork via\n\\textit{Topology-aware Embeddings} (TEs), which compress ego-subnetworks into\ncompact vectors (\\textit{i.e.}, TEs) to reduce the memory consumption. Based on\nthis framework, we discover a unique \\textit{pseudo-training effect} in\ncontinual learning on expanding networks and this effect motivates us to\ndevelop a novel \\textit{coverage maximization sampling} strategy that can\nenhance the performance with a tight memory budget. Thorough empirical studies\ndemonstrate that, by tackling the memory explosion problem and incorporating\ntopological information into memory replay, PDGNNs with TEM significantly\noutperform state-of-the-art techniques, especially in the challenging\nclass-incremental setting.\n", "date": "2024-01-24", "categories": []}
{"id": "2403.10123", "title": "Regularization-Based Efficient Continual Learning in Deep State-Space\n  Models", "abstract": "  Deep state-space models (DSSMs) have gained popularity in recent years due to\ntheir potent modeling capacity for dynamic systems. However, existing DSSM\nworks are limited to single-task modeling, which requires retraining with\nhistorical task data upon revisiting a forepassed task. To address this\nlimitation, we propose continual learning DSSMs (CLDSSMs), which are capable of\nadapting to evolving tasks without catastrophic forgetting. Our proposed\nCLDSSMs integrate mainstream regularization-based continual learning (CL)\nmethods, ensuring efficient updates with constant computational and memory\ncosts for modeling multiple dynamic systems. We also conduct a comprehensive\ncost analysis of each CL method applied to the respective CLDSSMs, and\ndemonstrate the efficacy of CLDSSMs through experiments on real-world datasets.\nThe results corroborate that while various competing CL methods exhibit\ndifferent merits, the proposed CLDSSMs consistently outperform traditional\nDSSMs in terms of effectively addressing catastrophic forgetting, enabling\nswift and accurate parameter transfer to new tasks.\n", "date": "2024-03-15", "categories": []}
{"id": "2407.00113", "title": "Personalized Federated Continual Learning via Multi-granularity Prompt", "abstract": "  Personalized Federated Continual Learning (PFCL) is a new practical scenario\nthat poses greater challenges in sharing and personalizing knowledge. PFCL not\nonly relies on knowledge fusion for server aggregation at the global\nspatial-temporal perspective but also needs model improvement for each client\naccording to the local requirements. Existing methods, whether in Personalized\nFederated Learning (PFL) or Federated Continual Learning (FCL), have overlooked\nthe multi-granularity representation of knowledge, which can be utilized to\novercome Spatial-Temporal Catastrophic Forgetting (STCF) and adopt generalized\nknowledge to itself by coarse-to-fine human cognitive mechanisms. Moreover, it\nallows more effectively to personalized shared knowledge, thus serving its own\npurpose. To this end, we propose a novel concept called multi-granularity\nprompt, i.e., coarse-grained global prompt acquired through the common model\nlearning process, and fine-grained local prompt used to personalize the\ngeneralized representation. The former focuses on efficiently transferring\nshared global knowledge without spatial forgetting, and the latter emphasizes\nspecific learning of personalized local knowledge to overcome temporal\nforgetting. In addition, we design a selective prompt fusion mechanism for\naggregating knowledge of global prompts distilled from different clients. By\nthe exclusive fusion of coarse-grained knowledge, we achieve the transmission\nand refinement of common knowledge among clients, further enhancing the\nperformance of personalization. Extensive experiments demonstrate the\neffectiveness of the proposed method in addressing STCF as well as improving\npersonalized performance. Our code now is available at\nhttps://github.com/SkyOfBeginning/FedMGP.\n", "date": "2024-06-27", "categories": []}
{"id": "2407.00176", "title": "The impact of model size on catastrophic forgetting in Online Continual\n  Learning", "abstract": "  This study investigates the impact of model size on Online Continual Learning\nperformance, with a focus on catastrophic forgetting. Employing ResNet\narchitectures of varying sizes, the research examines how network depth and\nwidth affect model performance in class-incremental learning using the\nSplitCIFAR-10 dataset. Key findings reveal that larger models do not guarantee\nbetter Continual Learning performance; in fact, they often struggle more in\nadapting to new tasks, particularly in online settings. These results challenge\nthe notion that larger models inherently mitigate catastrophic forgetting,\nhighlighting the nuanced relationship between model size and Continual Learning\nefficacy. This study contributes to a deeper understanding of model scalability\nand its practical implications in Continual Learning scenarios.\n", "date": "2024-06-28", "categories": []}
{"id": "2407.00875", "title": "MoE-CT: A Novel Approach For Large Language Models Training With\n  Resistance To Catastrophic Forgetting", "abstract": "  The advent of large language models (LLMs) has predominantly catered to\nhigh-resource languages, leaving a disparity in performance for low-resource\nlanguages. Conventional Continual Training (CT) approaches to bridge this gap\noften undermine a model's original linguistic proficiency when expanding to\nmultilingual contexts. Addressing this issue, we introduce a novel MoE-CT\narchitecture, a paradigm that innovatively separates the base model's learning\nfrom the multilingual expansion process. Our design freezes the original LLM\nparameters, thus safeguarding its performance in high-resource languages, while\nan appended MoE module, trained on diverse language datasets, augments\nlow-resource language proficiency. Our approach significantly outperforms\nconventional CT methods, as evidenced by our experiments, which show marked\nimprovements in multilingual benchmarks without sacrificing the model's\noriginal language performance. Moreover, our MoE-CT framework demonstrates\nenhanced resistance to forgetting and superior transfer learning capabilities.\nBy preserving the base model's integrity and focusing on strategic parameter\nexpansion, our methodology advances multilingual language modeling and\nrepresents a significant step forward for low-resource language inclusion in\nLLMs, indicating a fruitful direction for future research in language\ntechnologies.\n", "date": "2024-06-25", "categories": []}
{"id": "2407.01397", "title": "Mask and Compress: Efficient Skeleton-based Action Recognition in\n  Continual Learning", "abstract": "  The use of skeletal data allows deep learning models to perform action\nrecognition efficiently and effectively. Herein, we believe that exploring this\nproblem within the context of Continual Learning is crucial. While numerous\nstudies focus on skeleton-based action recognition from a traditional offline\nperspective, only a handful venture into online approaches. In this respect, we\nintroduce CHARON (Continual Human Action Recognition On skeletoNs), which\nmaintains consistent performance while operating within an efficient framework.\nThrough techniques like uniform sampling, interpolation, and a memory-efficient\ntraining stage based on masking, we achieve improved recognition accuracy while\nminimizing computational overhead. Our experiments on Split NTU-60 and the\nproposed Split NTU-120 datasets demonstrate that CHARON sets a new benchmark in\nthis domain. The code is available at https://github.com/Sperimental3/CHARON.\n", "date": "2024-07-01", "categories": []}
{"id": "2304.03894", "title": "A multifidelity approach to continual learning for physical systems", "abstract": "  We introduce a novel continual learning method based on multifidelity deep\nneural networks. This method learns the correlation between the output of\npreviously trained models and the desired output of the model on the current\ntraining dataset, limiting catastrophic forgetting. On its own the\nmultifidelity continual learning method shows robust results that limit\nforgetting across several datasets. Additionally, we show that the\nmultifidelity method can be combined with existing continual learning methods,\nincluding replay and memory aware synapses, to further limit catastrophic\nforgetting. The proposed continual learning method is especially suited for\nphysical problems where the data satisfy the same physical laws on each domain,\nor for physics-informed neural networks, because in these cases we expect there\nto be a strong correlation between the output of the previous model and the\nmodel on the current training domain.\n", "date": "2023-04-08", "categories": []}
{"id": "2403.06873", "title": "Last Iterate Convergence of Incremental Methods and Applications in\n  Continual Learning", "abstract": "  Incremental gradient and incremental proximal methods are a fundamental class\nof optimization algorithms used for solving finite sum problems, broadly\nstudied in the literature. Yet, without strong convexity, their convergence\nguarantees have primarily been established for the ergodic (average) iterate.\nMotivated by applications in continual learning, we obtain the first\nconvergence guarantees for the last iterate of both incremental gradient and\nincremental proximal methods, in general convex smooth (for both) and convex\nLipschitz (for the proximal variants) settings. Our oracle complexity bounds\nfor the last iterate nearly match (i.e., match up to a square-root-log or a log\nfactor) the best known oracle complexity bounds for the average iterate, for\nboth classes of methods. We further obtain generalizations of our results to\nweighted averaging of the iterates with increasing weights and for randomly\npermuted ordering of updates. We study incremental proximal methods as a model\nof continual learning with generalization and argue that large amount of\nregularization is crucial to preventing catastrophic forgetting. Our results\ngeneralize last iterate guarantees for incremental methods compared to state of\nthe art, as such results were previously known only for overparameterized\nlinear models, which correspond to convex quadratic problems with infinitely\nmany solutions.\n", "date": "2024-03-11", "categories": []}
{"id": "2406.19796", "title": "Comprehensive Generative Replay for Task-Incremental Segmentation with\n  Concurrent Appearance and Semantic Forgetting", "abstract": "  Generalist segmentation models are increasingly favored for diverse tasks\ninvolving various objects from different image sources. Task-Incremental\nLearning (TIL) offers a privacy-preserving training paradigm using tasks\narriving sequentially, instead of gathering them due to strict data sharing\npolicies. However, the task evolution can span a wide scope that involves\nshifts in both image appearance and segmentation semantics with intricate\ncorrelation, causing concurrent appearance and semantic forgetting. To solve\nthis issue, we propose a Comprehensive Generative Replay (CGR) framework that\nrestores appearance and semantic knowledge by synthesizing image-mask pairs to\nmimic past task data, which focuses on two aspects: modeling image-mask\ncorrespondence and promoting scalability for diverse tasks. Specifically, we\nintroduce a novel Bayesian Joint Diffusion (BJD) model for high-quality\nsynthesis of image-mask pairs with their correspondence explicitly preserved by\nconditional denoising. Furthermore, we develop a Task-Oriented Adapter (TOA)\nthat recalibrates prompt embeddings to modulate the diffusion model, making the\ndata synthesis compatible with different tasks. Experiments on incremental\ntasks (cardiac, fundus and prostate segmentation) show its clear advantage for\nalleviating concurrent appearance and semantic forgetting. Code is available at\nhttps://github.com/jingyzhang/CGR.\n", "date": "2024-06-28", "categories": []}
{"id": "2311.01200", "title": "Continual Learning Under Language Shift", "abstract": "  The recent increase in data and model scale for language model pre-training\nhas led to huge training costs. In scenarios where new data become available\nover time, updating a model instead of fully retraining it would therefore\nprovide significant gains. We study the pros and cons of updating a language\nmodel when new data comes from new languages -- the case of continual learning\nunder language shift. Starting from a monolingual English language model, we\nincrementally add data from Danish, Icelandic, and Norwegian to investigate how\nforward and backward transfer effects depend on pre-training order and\ncharacteristics of languages, for three different model sizes. Our results show\nthat, while forward transfer is largely positive and independent of language\norder, backward transfer can be positive or negative depending on the order and\ncharacteristics of new languages. We explore a number of potentially\nexplanatory factors and find that a combination of language contamination and\nsyntactic similarity best fits our results.\n", "date": "2023-11-02", "categories": []}
{"id": "2401.09181", "title": "Beyond Anti-Forgetting: Multimodal Continual Instruction Tuning with\n  Positive Forward Transfer", "abstract": "  Multimodal Continual Instruction Tuning (MCIT) enables Multimodal Large\nLanguage Models (MLLMs) to meet continuously emerging requirements without\nexpensive retraining. MCIT faces two major obstacles: catastrophic forgetting\n(where old knowledge is forgotten) and negative forward transfer (where the\nperformance of future tasks is degraded). Although existing methods have\ngreatly alleviated catastrophic forgetting, they still suffer from negative\nforward transfer. We discover a large discrepancy in different input embeddings\nby performing singular value decomposition (SVD) on input embeddings. This\ndiscrepancy results in the model learning irrelevant information for old and\npre-trained tasks, leading to catastrophic forgetting and negative forward\ntransfer. To address these issues, we propose Prompt Tuning with Positive\nForward Transfer (Fwd-Prompt), a prompt-based method that projects the prompt\ngradient to the residual space to minimize interference between tasks and to\nthe pre-trained subspace for reusing pre-trained knowledge. Our experiments\ndemonstrate that Fwd-Prompt achieves state-of-the-art performance while\nupdating fewer parameters and requiring no old samples. Our research\nilluminates the potential of continuously adapting MLLMs to new tasks under the\ninstruction tuning paradigm and encourages future studies to explore MCIT.\n", "date": "2024-01-17", "categories": []}
{"id": "2406.14833", "title": "Efficient Continual Pre-training by Mitigating the Stability Gap", "abstract": "  Continual pre-training has increasingly become the predominant approach for\nadapting Large Language Models (LLMs) to new domains. This process involves\nupdating the pre-trained LLM with a corpus from a new domain, resulting in a\nshift in the training distribution. To study the behavior of LLMs during this\nshift, we measured the model's performance throughout the continual\npre-training process. we observed a temporary performance drop at the\nbeginning, followed by a recovery phase, a phenomenon known as the \"stability\ngap,\" previously noted in vision models classifying new classes. To address\nthis issue and enhance LLM performance within a fixed compute budget, we\npropose three effective strategies: (1) Continually pre-training the LLM on a\nsubset with a proper size for multiple epochs, resulting in faster performance\nrecovery than pre-training the LLM on a large corpus in a single epoch; (2)\nPre-training the LLM only on high-quality sub-corpus, which rapidly boosts\ndomain performance; and (3) Using a data mixture similar to the pre-training\ndata to reduce distribution gap. We conduct various experiments on Llama-family\nmodels to validate the effectiveness of our strategies in both medical\ncontinual pre-training and instruction tuning. For example, our strategies\nimprove the average medical task performance of the OpenLlama-3B model from\n36.2% to 40.7% with only 40% of the original training budget and enhance the\naverage general task performance without causing forgetting. Furthermore, we\napply our strategies to the Llama-3-8B model. The resulting model,\nLlama-3-Physician, achieves the best medical performance among current\nopen-source models, and performs comparably to or even better than GPT-4 on\nseveral medical benchmarks. We release our models at\n\\url{https://huggingface.co/YiDuo1999/Llama-3-Physician-8B-Instruct}.\n", "date": "2024-06-21", "categories": []}
{"id": "2406.18708", "title": "Learn it or Leave it: Module Composition and Pruning for Continual\n  Learning", "abstract": "  In real-world environments, continual learning is essential for machine\nlearning models, as they need to acquire new knowledge incrementally without\nforgetting what they have already learned. While pretrained language models\nhave shown impressive capabilities on various static tasks, applying them to\ncontinual learning poses significant challenges, including avoiding\ncatastrophic forgetting, facilitating knowledge transfer, and maintaining\nparameter efficiency. In this paper, we introduce MoCL-P, a novel lightweight\ncontinual learning method that addresses these challenges simultaneously.\nUnlike traditional approaches that continuously expand parameters for newly\narriving tasks, MoCL-P integrates task representation-guided module composition\nwith adaptive pruning, effectively balancing knowledge integration and\ncomputational overhead. Our evaluation across three continual learning\nbenchmarks with up to 176 tasks shows that MoCL-P achieves state-of-the-art\nperformance and improves parameter efficiency by up to three times,\ndemonstrating its potential for practical applications where resource\nrequirements are constrained.\n", "date": "2024-06-26", "categories": []}
{"id": "2406.19297", "title": "Enhancing Continual Learning in Visual Question Answering with\n  Modality-Aware Feature Distillation", "abstract": "  Continual learning focuses on incrementally training a model on a sequence of\ntasks with the aim of learning new tasks while minimizing performance drop on\nprevious tasks. Existing approaches at the intersection of Continual Learning\nand Visual Question Answering (VQA) do not study how the multimodal nature of\nthe input affects the learning dynamics of a model. In this paper, we\ndemonstrate that each modality evolves at different rates across a continuum of\ntasks and that this behavior occurs in established encoder-only models as well\nas modern recipes for developing Vision & Language (VL) models. Motivated by\nthis observation, we propose a modality-aware feature distillation (MAFED)\napproach which outperforms existing baselines across models of varying scale in\nthree multimodal continual learning settings. Furthermore, we provide ablations\nshowcasing that modality-aware distillation complements experience replay.\nOverall, our results emphasize the importance of addressing modality-specific\ndynamics to prevent forgetting in multimodal continual learning.\n", "date": "2024-06-27", "categories": []}
{"id": "2206.08101", "title": "Towards Diverse Evaluation of Class Incremental Learning: A\n  Representation Learning Perspective", "abstract": "  Class incremental learning (CIL) algorithms aim to continually learn new\nobject classes from incrementally arriving data while not forgetting past\nlearned classes. The common evaluation protocol for CIL algorithms is to\nmeasure the average test accuracy across all classes learned so far -- however,\nwe argue that solely focusing on maximizing the test accuracy may not\nnecessarily lead to developing a CIL algorithm that also continually learns and\nupdates the representations, which may be transferred to the downstream tasks.\nTo that end, we experimentally analyze neural network models trained by CIL\nalgorithms using various evaluation protocols in representation learning and\npropose new analysis methods. Our experiments show that most state-of-the-art\nalgorithms prioritize high stability and do not significantly change the\nlearned representation, and sometimes even learn a representation of lower\nquality than a naive baseline. However, we observe that these algorithms can\nstill achieve high test accuracy because they enable a model to learn a\nclassifier that closely resembles an estimated linear classifier trained for\nlinear probing. Furthermore, the base model learned in the first task, which\ninvolves single-task learning, exhibits varying levels of representation\nquality across different algorithms, and this variance impacts the final\nperformance of CIL algorithms. Therefore, we suggest that the\nrepresentation-level evaluation should be considered as an additional recipe\nfor more diverse evaluation for CIL algorithms.\n", "date": "2022-06-16", "categories": []}
{"id": "2304.00933", "title": "Knowledge Accumulation in Continually Learned Representations and the\n  Issue of Feature Forgetting", "abstract": "  Continual learning research has shown that neural networks suffer from\ncatastrophic forgetting \"at the output level\", but it is debated whether this\nis also the case at the level of learned representations. Multiple recent\nstudies ascribe representations a certain level of innate robustness against\nforgetting -- that they only forget minimally in comparison with forgetting at\nthe output level. We revisit and expand upon the experiments that revealed this\ndifference in forgetting and illustrate the coexistence of two phenomena that\naffect the quality of continually learned representations: knowledge\naccumulation and feature forgetting. Taking both aspects into account, we show\nthat, even though forgetting in the representation (i.e. feature forgetting)\ncan be small in absolute terms, when measuring relative to how much was learned\nduring a task, forgetting in the representation tends to be just as\ncatastrophic as forgetting at the output level. Next we show that this feature\nforgetting is problematic as it substantially slows down the incremental\nlearning of good general representations (i.e. knowledge accumulation).\nFinally, we study how feature forgetting and knowledge accumulation are\naffected by different types of continual learning methods.\n", "date": "2023-04-03", "categories": []}
{"id": "2406.15766", "title": "Continual Learning with Diffusion-based Generative Replay for Industrial\n  Streaming Data", "abstract": "  The Industrial Internet of Things (IIoT) integrates interconnected sensors\nand devices to support industrial applications, but its dynamic environments\npose challenges related to data drift. Considering the limited resources and\nthe need to effectively adapt models to new data distributions, this paper\nintroduces a Continual Learning (CL) approach, i.e., Distillation-based\nSelf-Guidance (DSG), to address challenges presented by industrial streaming\ndata via a novel generative replay mechanism. DSG utilizes knowledge\ndistillation to transfer knowledge from the previous diffusion-based generator\nto the updated one, improving both the stability of the generator and the\nquality of reproduced data, thereby enhancing the mitigation of catastrophic\nforgetting. Experimental results on CWRU, DSA, and WISDM datasets demonstrate\nthe effectiveness of DSG. DSG outperforms the state-of-the-art baseline in\naccuracy, demonstrating improvements ranging from 2.9% to 5.0% on key datasets,\nshowcasing its potential for practical industrial applications.\n", "date": "2024-06-22", "categories": []}
{"id": "2406.15972", "title": "EVCL: Elastic Variational Continual Learning with Weight Consolidation", "abstract": "  Continual learning aims to allow models to learn new tasks without forgetting\nwhat has been learned before. This work introduces Elastic Variational\nContinual Learning with Weight Consolidation (EVCL), a novel hybrid model that\nintegrates the variational posterior approximation mechanism of Variational\nContinual Learning (VCL) with the regularization-based parameter-protection\nstrategy of Elastic Weight Consolidation (EWC). By combining the strengths of\nboth methods, EVCL effectively mitigates catastrophic forgetting and enables\nbetter capture of dependencies between model parameters and task-specific data.\nEvaluated on five discriminative tasks, EVCL consistently outperforms existing\nbaselines in both domain-incremental and task-incremental learning scenarios\nfor deep discriminative models.\n", "date": "2024-06-23", "categories": []}
{"id": "2406.16231", "title": "Gradual Divergence for Seamless Adaptation: A Novel Domain Incremental\n  Learning Method", "abstract": "  Domain incremental learning (DIL) poses a significant challenge in real-world\nscenarios, as models need to be sequentially trained on diverse domains over\ntime, all the while avoiding catastrophic forgetting. Mitigating representation\ndrift, which refers to the phenomenon of learned representations undergoing\nchanges as the model adapts to new tasks, can help alleviate catastrophic\nforgetting. In this study, we propose a novel DIL method named DARE, featuring\na three-stage training process: Divergence, Adaptation, and REfinement. This\nprocess gradually adapts the representations associated with new tasks into the\nfeature space spanned by samples from previous tasks, simultaneously\nintegrating task-specific decision boundaries. Additionally, we introduce a\nnovel strategy for buffer sampling and demonstrate the effectiveness of our\nproposed method, combined with this sampling strategy, in reducing\nrepresentation drift within the feature encoder. This contribution effectively\nalleviates catastrophic forgetting across multiple DIL benchmarks. Furthermore,\nour approach prevents sudden representation drift at task boundaries, resulting\nin a well-calibrated DIL model that maintains the performance on previous\ntasks.\n", "date": "2024-06-23", "categories": []}
{"id": "2406.16615", "title": "The Championship-Winning Solution for the 5th CLVISION Challenge 2024", "abstract": "  In this paper, we introduce our approach to the 5th CLVision Challenge, which\npresents distinctive challenges beyond traditional class incremental learning.\nUnlike standard settings, this competition features the recurrence of\npreviously encountered classes and includes unlabeled data that may contain\nOut-of-Distribution (OOD) categories. Our approach is based on Winning\nSubnetworks to allocate independent parameter spaces for each task addressing\nthe catastrophic forgetting problem in class incremental learning and employ\nthree training strategies: supervised classification learning, unsupervised\ncontrastive learning, and pseudo-label classification learning to fully utilize\nthe information in both labeled and unlabeled data, enhancing the\nclassification performance of each subnetwork. Furthermore, during the\ninference stage, we have devised an interaction strategy between subnetworks,\nwhere the prediction for a specific class of a particular sample is the average\nlogits across different subnetworks corresponding to that class, leveraging the\nknowledge learned from different subnetworks on recurring classes to improve\nclassification accuracy. These strategies can be simultaneously applied to the\nthree scenarios of the competition, effectively solving the difficulties in the\ncompetition scenarios. Experimentally, our method ranks first in both the\npre-selection and final evaluation stages, with an average accuracy of 0.4535\nduring the preselection stage and an average accuracy of 0.4805 during the\nfinal evaluation stage.\n", "date": "2024-06-24", "categories": []}
{"id": "2308.08378", "title": "Advancing continual lifelong learning in neural information retrieval:\n  definition, dataset, framework, and empirical evaluation", "abstract": "  Continual learning refers to the capability of a machine learning model to\nlearn and adapt to new information, without compromising its performance on\npreviously learned tasks. Although several studies have investigated continual\nlearning methods for information retrieval tasks, a well-defined task\nformulation is still lacking, and it is unclear how typical learning strategies\nperform in this context. To address this challenge, a systematic task\nformulation of continual neural information retrieval is presented, along with\na multiple-topic dataset that simulates continuous information retrieval. A\ncomprehensive continual neural information retrieval framework consisting of\ntypical retrieval models and continual learning strategies is then proposed.\nEmpirical evaluations illustrate that the proposed framework can successfully\nprevent catastrophic forgetting in neural information retrieval and enhance\nperformance on previously learned tasks. The results indicate that\nembedding-based retrieval models experience a decline in their continual\nlearning performance as the topic shift distance and dataset volume of new\ntasks increase. In contrast, pretraining-based models do not show any such\ncorrelation. Adopting suitable learning strategies can mitigate the effects of\ntopic shift and data augmentation.\n", "date": "2023-08-16", "categories": []}
{"id": "2310.04741", "title": "Keep Moving: identifying task-relevant subspaces to maximise plasticity\n  for newly learned tasks", "abstract": "  Continual learning algorithms strive to acquire new knowledge while\npreserving prior information. Often, these algorithms emphasise stability and\nrestrict network updates upon learning new tasks. In many cases, such\nrestrictions come at a cost to the model's plasticity, i.e. the model's ability\nto adapt to the requirements of a new task. But is all change detrimental?\nHere, we approach this question by proposing that activation spaces in neural\nnetworks can be decomposed into two subspaces: a readout range in which change\naffects prior tasks and a null space in which change does not alter prior\nperformance. Based on experiments with this novel technique, we show that,\nindeed, not all activation change is associated with forgetting. Instead, only\nchange in the subspace visible to the readout of a task can lead to decreased\nstability, while restricting change outside of this subspace is associated only\nwith a loss of plasticity. Analysing various commonly used algorithms, we show\nthat regularisation-based techniques do not fully disentangle the two spaces\nand, as a result, restrict plasticity more than need be. We expand our results\nby investigating a linear model in which we can manipulate learning in the two\nsubspaces directly and thus causally link activation changes to stability and\nplasticity. For hierarchical, nonlinear cases, we present an approximation that\nenables us to estimate functionally relevant subspaces at every layer of a deep\nnonlinear network, corroborating our previous insights. Together, this work\nprovides novel means to derive insights into the mechanisms behind stability\nand plasticity in continual learning and may serve as a diagnostic tool to\nguide developments of future continual learning algorithms that stabilise\ninference while allowing maximal space for learning.\n", "date": "2023-10-07", "categories": []}
{"id": "2403.06398", "title": "On the Diminishing Returns of Width for Continual Learning", "abstract": "  While deep neural networks have demonstrated groundbreaking performance in\nvarious settings, these models often suffer from \\emph{catastrophic forgetting}\nwhen trained on new tasks in sequence. Several works have empirically\ndemonstrated that increasing the width of a neural network leads to a decrease\nin catastrophic forgetting but have yet to characterize the exact relationship\nbetween width and continual learning. We design one of the first frameworks to\nanalyze Continual Learning Theory and prove that width is directly related to\nforgetting in Feed-Forward Networks (FFN). Specifically, we demonstrate that\nincreasing network widths to reduce forgetting yields diminishing returns. We\nempirically verify our claims at widths hitherto unexplored in prior studies\nwhere the diminishing returns are clearly observed as predicted by our theory.\n", "date": "2024-03-11", "categories": []}
{"id": "2403.07015", "title": "Adaptive Hyperparameter Optimization for Continual Learning Scenarios", "abstract": "  Hyperparameter selection in continual learning scenarios is a challenging and\nunderexplored aspect, especially in practical non-stationary environments.\nTraditional approaches, such as grid searches with held-out validation data\nfrom all tasks, are unrealistic for building accurate lifelong learning\nsystems. This paper aims to explore the role of hyperparameter selection in\ncontinual learning and the necessity of continually and automatically tuning\nthem according to the complexity of the task at hand. Hence, we propose\nleveraging the nature of sequence task learning to improve Hyperparameter\nOptimization efficiency. By using the functional analysis of variance-based\ntechniques, we identify the most crucial hyperparameters that have an impact on\nperformance. We demonstrate empirically that this approach, agnostic to\ncontinual scenarios and strategies, allows us to speed up hyperparameters\noptimization continually across tasks and exhibit robustness even in the face\nof varying sequential task orders. We believe that our findings can contribute\nto the advancement of continual learning methodologies towards more efficient,\nrobust and adaptable models for real-world applications.\n", "date": "2024-03-09", "categories": []}
{"id": "2406.13583", "title": "Low-Rank Mixture-of-Experts for Continual Medical Image Segmentation", "abstract": "  The primary goal of continual learning (CL) task in medical image\nsegmentation field is to solve the \"catastrophic forgetting\" problem, where the\nmodel totally forgets previously learned features when it is extended to new\ncategories (class-level) or tasks (task-level). Due to the privacy protection,\nthe historical data labels are inaccessible. Prevalent continual learning\nmethods primarily focus on generating pseudo-labels for old datasets to force\nthe model to memorize the learned features. However, the incorrect\npseudo-labels may corrupt the learned feature and lead to a new problem that\nthe better the model is trained on the old task, the poorer the model performs\non the new tasks. To avoid this problem, we propose a network by introducing\nthe data-specific Mixture of Experts (MoE) structure to handle the new tasks or\ncategories, ensuring that the network parameters of previous tasks are\nunaffected or only minimally impacted. To further overcome the tremendous\nmemory costs caused by introducing additional structures, we propose a Low-Rank\nstrategy which significantly reduces memory cost. We validate our method on\nboth class-level and task-level continual learning challenges. Extensive\nexperiments on multiple datasets show our model outperforms all other methods.\n", "date": "2024-06-19", "categories": []}
{"id": "2303.10263", "title": "Fixed Design Analysis of Regularization-Based Continual Learning", "abstract": "  We consider a continual learning (CL) problem with two linear regression\ntasks in the fixed design setting, where the feature vectors are assumed fixed\nand the labels are assumed to be random variables. We consider an\n$\\ell_2$-regularized CL algorithm, which computes an Ordinary Least Squares\nparameter to fit the first dataset, then computes another parameter that fits\nthe second dataset under an $\\ell_2$-regularization penalizing its deviation\nfrom the first parameter, and outputs the second parameter. For this algorithm,\nwe provide tight bounds on the average risk over the two tasks. Our risk bounds\nreveal a provable trade-off between forgetting and intransigence of the\n$\\ell_2$-regularized CL algorithm: with a large regularization parameter, the\nalgorithm output forgets less information about the first task but is\nintransigent to extract new information from the second task; and vice versa.\nOur results suggest that catastrophic forgetting could happen for CL with\ndissimilar tasks (under a precise similarity measurement) and that a well-tuned\n$\\ell_2$-regularization can partially mitigate this issue by introducing\nintransigence.\n", "date": "2023-03-17", "categories": []}
{"id": "2312.03177", "title": "Using Curiosity for an Even Representation of Tasks in Continual Offline\n  Reinforcement Learning", "abstract": "  In this work, we investigate the means of using curiosity on replay buffers\nto improve offline multi-task continual reinforcement learning when tasks,\nwhich are defined by the non-stationarity in the environment, are non labeled\nand not evenly exposed to the learner in time. In particular, we investigate\nthe use of curiosity both as a tool for task boundary detection and as a\npriority metric when it comes to retaining old transition tuples, which we\nrespectively use to propose two different buffers. Firstly, we propose a Hybrid\nReservoir Buffer with Task Separation (HRBTS), where curiosity is used to\ndetect task boundaries that are not known due to the task agnostic nature of\nthe problem. Secondly, by using curiosity as a priority metric when it comes to\nretaining old transition tuples, a Hybrid Curious Buffer (HCB) is proposed. We\nultimately show that these buffers, in conjunction with regular reinforcement\nlearning algorithms, can be used to alleviate the catastrophic forgetting issue\nsuffered by the state of the art on replay buffers when the agent's exposure to\ntasks is not equal along time. We evaluate catastrophic forgetting and the\nefficiency of our proposed buffers against the latest works such as the Hybrid\nReservoir Buffer (HRB) and the Multi-Time Scale Replay Buffer (MTR) in three\ndifferent continual reinforcement learning settings. Experiments were done on\nclassical control tasks and Metaworld environment. Experiments show that our\nproposed replay buffers display better immunity to catastrophic forgetting\ncompared to existing works in most of the settings.\n", "date": "2023-12-05", "categories": []}
{"id": "2402.08526", "title": "Can LLMs Learn New Concepts Incrementally without Forgetting?", "abstract": "  Large Language Models (LLMs) have achieved remarkable success across various\ntasks, yet their ability to learn incrementally without forgetting remains\nunderexplored. Incremental learning (IL) is crucial as it enables models to\nacquire new knowledge while retaining previously learned information, akin to\nhuman learning. Existing benchmarks for IL are insufficient due to data leakage\nissues and the overqualification of LLMs. To address these challenges, we\nintroduce Concept-1K, a novel dataset comprising 1,023 recently emerged\nconcepts across diverse domains. The concepts in Concept-1K are discrete,\ninterpretable units of knowledge that allow for fine-grained analysis of\nlearning and forgetting processes. Using Concept-1K as a testbed, we aim to\nanswer the question: ``Can LLMs learn new concepts incrementally without\nforgetting like humans?'' Our investigation reveals that LLMs still suffer from\ncatastrophic forgetting and that LoRA, despite fine-tuning fewer parameters,\nmay lead to more forgetting on training data. Additionally, we explore the\nroles of in-context learning, model scale, buffer size, and pretraining in IL\nperformance. These findings highlight the strengths and limitations of LLMs in\nIL scenarios and provide a robust benchmark for future research.\n", "date": "2024-02-13", "categories": []}
{"id": "2401.02576", "title": "t-DGR: A Trajectory-Based Deep Generative Replay Method for Continual\n  Learning in Decision Making", "abstract": "  Deep generative replay has emerged as a promising approach for continual\nlearning in decision-making tasks. This approach addresses the problem of\ncatastrophic forgetting by leveraging the generation of trajectories from\npreviously encountered tasks to augment the current dataset. However, existing\ndeep generative replay methods for continual learning rely on autoregressive\nmodels, which suffer from compounding errors in the generated trajectories. In\nthis paper, we propose a simple, scalable, and non-autoregressive method for\ncontinual learning in decision-making tasks using a generative model that\ngenerates task samples conditioned on the trajectory timestep. We evaluate our\nmethod on Continual World benchmarks and find that our approach achieves\nstate-of-the-art performance on the average success rate metric among continual\nlearning methods. Code is available at https://github.com/WilliamYue37/t-DGR.\n", "date": "2024-01-04", "categories": []}
{"id": "2401.18054", "title": "Benchmarking Sensitivity of Continual Graph Learning for Skeleton-Based\n  Action Recognition", "abstract": "  Continual learning (CL) is the research field that aims to build machine\nlearning models that can accumulate knowledge continuously over different tasks\nwithout retraining from scratch. Previous studies have shown that pre-training\ngraph neural networks (GNN) may lead to negative transfer (Hu et al., 2020)\nafter fine-tuning, a setting which is closely related to CL. Thus, we focus on\nstudying GNN in the continual graph learning (CGL) setting. We propose the\nfirst continual graph learning benchmark for spatio-temporal graphs and use it\nto benchmark well-known CGL methods in this novel setting. The benchmark is\nbased on the N-UCLA and NTU-RGB+D datasets for skeleton-based action\nrecognition. Beyond benchmarking for standard performance metrics, we study the\nclass and task-order sensitivity of CGL methods, i.e., the impact of learning\norder on each class/task's performance, and the architectural sensitivity of\nCGL methods with backbone GNN at various widths and depths. We reveal that\ntask-order robust methods can still be class-order sensitive and observe\nresults that contradict previous empirical observations on architectural\nsensitivity in CL.\n", "date": "2024-01-31", "categories": []}
{"id": "2406.09052", "title": "Data-Free Generative Replay for Class-Incremental Learning on Imbalanced\n  Data", "abstract": "  Continual learning is a challenging problem in machine learning, especially\nfor image classification tasks with imbalanced datasets. It becomes even more\nchallenging when it involves learning new classes incrementally. One method for\nincremental class learning, addressing dataset imbalance, is rehearsal using\npreviously stored data. In rehearsal-based methods, access to previous data is\nrequired for either training the classifier or the generator, but it may not be\nfeasible due to storage, legal, or data access constraints. Although there are\nmany rehearsal-free alternatives for class incremental learning, such as\nparameter or loss regularization, knowledge distillation, and dynamic\narchitectures, they do not consistently achieve good results, especially on\nimbalanced data. This paper proposes a new approach called Data-Free Generative\nReplay (DFGR) for class incremental learning, where the generator is trained\nwithout access to real data. In addition, DFGR also addresses dataset imbalance\nin continual learning of an image classifier. Instead of using training data,\nDFGR trains a generator using mean and variance statistics of batch-norm and\nfeature maps derived from a pre-trained classification model. The results of\nour experiments demonstrate that DFGR performs significantly better than other\ndata-free methods and reveal the performance impact of specific parameter\nsettings. DFGR achieves up to 88.5% and 46.6% accuracy on MNIST and\nFashionMNIST datasets, respectively. Our code is available at\nhttps://github.com/2younis/DFGR\n", "date": "2024-06-07", "categories": []}
{"id": "2405.08969", "title": "Wearable Sensor-Based Few-Shot Continual Learning on Hand Gestures for\n  Motor-Impaired Individuals via Latent Embedding Exploitation", "abstract": "  Hand gestures can provide a natural means of human-computer interaction and\nenable people who cannot speak to communicate efficiently. Existing hand\ngesture recognition methods heavily depend on pre-defined gestures, however,\nmotor-impaired individuals require new gestures tailored to each individual's\ngesture motion and style. Gesture samples collected from different persons have\ndistribution shifts due to their health conditions, the severity of the\ndisability, motion patterns of the arms, etc. In this paper, we introduce the\nLatent Embedding Exploitation (LEE) mechanism in our replay-based Few-Shot\nContinual Learning (FSCL) framework that significantly improves the performance\nof fine-tuning a model for out-of-distribution data. Our method produces a\ndiversified latent feature space by leveraging a preserved latent embedding\nknown as gesture prior knowledge, along with intra-gesture divergence derived\nfrom two additional embeddings. Thus, the model can capture latent statistical\nstructure in highly variable gestures with limited samples. We conduct an\nexperimental evaluation using the SmartWatch Gesture and the Motion Gesture\ndatasets. The proposed method results in an average test accuracy of 57.0%,\n64.6%, and 69.3% by using one, three, and five samples for six different\ngestures. Our method helps motor-impaired persons leverage wearable devices,\nand their unique styles of movement can be learned and applied in\nhuman-computer interaction and social communication. Code is available at:\nhttps://github.com/riyadRafiq/wearable-latent-embedding-exploitation\n", "date": "2024-05-14", "categories": []}
{"id": "2404.12766", "title": "Continual Learning on a Diet: Learning from Sparsely Labeled Streams\n  Under Constrained Computation", "abstract": "  We propose and study a realistic Continual Learning (CL) setting where\nlearning algorithms are granted a restricted computational budget per time step\nwhile training. We apply this setting to large-scale semi-supervised Continual\nLearning scenarios with sparse label rates. Previous proficient CL methods\nperform very poorly in this challenging setting. Overfitting to the sparse\nlabeled data and insufficient computational budget are the two main culprits\nfor such a poor performance. Our new setting encourages learning methods to\neffectively and efficiently utilize the unlabeled data during training. To that\nend, we propose a simple but highly effective baseline, DietCL, which utilizes\nboth unlabeled and labeled data jointly. DietCL meticulously allocates\ncomputational budget for both types of data. We validate our baseline, at\nscale, on several datasets, e.g., CLOC, ImageNet10K, and CGLM, under constraint\nbudget setups. DietCL outperforms, by a large margin, all existing supervised\nCL algorithms as well as more recent continual semi-supervised methods. Our\nextensive analysis and ablations demonstrate that DietCL is stable under a full\nspectrum of label sparsity, computational budget, and various other ablations.\n", "date": "2024-04-19", "categories": []}
{"id": "2405.09637", "title": "CLASSP: a Biologically-Inspired Approach to Continual Learning through\n  Adjustment Suppression and Sparsity Promotion", "abstract": "  This paper introduces a new biologically-inspired training method named\nContinual Learning through Adjustment Suppression and Sparsity Promotion\n(CLASSP). CLASSP is based on two main principles observed in neuroscience,\nparticularly in the context of synaptic transmission and Long-Term Potentiation\n(LTP). The first principle is a decay rate over the weight adjustment, which is\nimplemented as a generalization of the AdaGrad optimization algorithm. This\nmeans that weights that have received many updates should have lower learning\nrates as they likely encode important information about previously seen data.\nHowever, this principle results in a diffuse distribution of updates throughout\nthe model, as it promotes updates for weights that haven't been previously\nupdated, while a sparse update distribution is preferred to leave weights\nunassigned for future tasks. Therefore, the second principle introduces a\nthreshold on the loss gradient. This promotes sparse learning by updating a\nweight only if the loss gradient with respect to that weight is above a certain\nthreshold, i.e. only updating weights with a significant impact on the current\nloss. Both principles reflect phenomena observed in LTP, where a threshold\neffect and a gradual saturation of potentiation have been observed. CLASSP is\nimplemented in a Python/PyTorch class, making it applicable to any model. When\ncompared with Elastic Weight Consolidation (EWC) using Computer Vision and\nsentiment analysis datasets, CLASSP demonstrates superior performance in terms\nof accuracy and memory footprint.\n", "date": "2024-04-29", "categories": []}
{"id": "2405.14623", "title": "U-TELL: Unsupervised Task Expert Lifelong Learning", "abstract": "  Continual learning (CL) models are designed to learn new tasks arriving\nsequentially without re-training the network. However, real-world ML\napplications have very limited label information and these models suffer from\ncatastrophic forgetting. To address these issues, we propose an unsupervised CL\nmodel with task experts called Unsupervised Task Expert Lifelong Learning\n(U-TELL) to continually learn the data arriving in a sequence addressing\ncatastrophic forgetting. During training of U-TELL, we introduce a new expert\non arrival of a new task. Our proposed architecture has task experts, a\nstructured data generator and a task assigner. Each task expert is composed of\n3 blocks; i) a variational autoencoder to capture the task distribution and\nperform data abstraction, ii) a k-means clustering module, and iii) a structure\nextractor to preserve latent task data signature. During testing, task assigner\nselects a suitable expert to perform clustering. U-TELL does not store or\nreplay task samples, instead, we use generated structured samples to train the\ntask assigner. We compared U-TELL with five SOTA unsupervised CL methods.\nU-TELL outperformed all baselines on seven benchmarks and one industry dataset\nfor various CL scenarios with a training time over 6 times faster than the best\nperforming baseline.\n", "date": "2024-05-23", "categories": []}
{"id": "2406.05631", "title": "CCSI: Continual Class-Specific Impression for Data-free Class\n  Incremental Learning", "abstract": "  In real-world clinical settings, traditional deep learning-based\nclassification methods struggle with diagnosing newly introduced disease types\nbecause they require samples from all disease classes for offline training.\nClass incremental learning offers a promising solution by adapting a deep\nnetwork trained on specific disease classes to handle new diseases. However,\ncatastrophic forgetting occurs, decreasing the performance of earlier classes\nwhen adapting the model to new data. Prior proposed methodologies to overcome\nthis require perpetual storage of previous samples, posing potential practical\nconcerns regarding privacy and storage regulations in healthcare. To this end,\nwe propose a novel data-free class incremental learning framework that utilizes\ndata synthesis on learned classes instead of data storage from previous\nclasses. Our key contributions include acquiring synthetic data known as\nContinual Class-Specific Impression (CCSI) for previously inaccessible trained\nclasses and presenting a methodology to effectively utilize this data for\nupdating networks when introducing new classes. We obtain CCSI by employing\ndata inversion over gradients of the trained classification model on previous\nclasses starting from the mean image of each class inspired by common landmarks\nshared among medical images and utilizing continual normalization layers\nstatistics as a regularizer in this pixel-wise optimization process.\nSubsequently, we update the network by combining the synthesized data with new\nclass data and incorporate several losses, including an intra-domain\ncontrastive loss to generalize the deep network trained on the synthesized data\nto real data, a margin loss to increase separation among previous classes and\nnew ones, and a cosine-normalized cross-entropy loss to alleviate the adverse\neffects of imbalanced distributions in training data.\n", "date": "2024-06-09", "categories": []}
{"id": "2406.06213", "title": "A Statistical Theory of Regularization-Based Continual Learning", "abstract": "  We provide a statistical analysis of regularization-based continual learning\non a sequence of linear regression tasks, with emphasis on how different\nregularization terms affect the model performance. We first derive the\nconvergence rate for the oracle estimator obtained as if all data were\navailable simultaneously. Next, we consider a family of generalized\n$\\ell_2$-regularization algorithms indexed by matrix-valued hyperparameters,\nwhich includes the minimum norm estimator and continual ridge regression as\nspecial cases. As more tasks are introduced, we derive an iterative update\nformula for the estimation error of generalized $\\ell_2$-regularized\nestimators, from which we determine the hyperparameters resulting in the\noptimal algorithm. Interestingly, the choice of hyperparameters can effectively\nbalance the trade-off between forward and backward knowledge transfer and\nadjust for data heterogeneity. Moreover, the estimation error of the optimal\nalgorithm is derived explicitly, which is of the same order as that of the\noracle estimator. In contrast, our lower bounds for the minimum norm estimator\nand continual ridge regression show their suboptimality. A byproduct of our\ntheoretical analysis is the equivalence between early stopping and generalized\n$\\ell_2$-regularization in continual learning, which may be of independent\ninterest. Finally, we conduct experiments to complement our theory.\n", "date": "2024-06-10", "categories": []}
{"id": "2406.06391", "title": "Towards Lifelong Learning of Large Language Models: A Survey", "abstract": "  As the applications of large language models (LLMs) expand across diverse\nfields, the ability of these models to adapt to ongoing changes in data, tasks,\nand user preferences becomes crucial. Traditional training methods, relying on\nstatic datasets, are increasingly inadequate for coping with the dynamic nature\nof real-world information. Lifelong learning, also known as continual or\nincremental learning, addresses this challenge by enabling LLMs to learn\ncontinuously and adaptively over their operational lifetime, integrating new\nknowledge while retaining previously learned information and preventing\ncatastrophic forgetting. This survey delves into the sophisticated landscape of\nlifelong learning, categorizing strategies into two primary groups: Internal\nKnowledge and External Knowledge. Internal Knowledge includes continual\npretraining and continual finetuning, each enhancing the adaptability of LLMs\nin various scenarios. External Knowledge encompasses retrieval-based and\ntool-based lifelong learning, leveraging external data sources and\ncomputational tools to extend the model's capabilities without modifying core\nparameters. The key contributions of our survey are: (1) Introducing a novel\ntaxonomy categorizing the extensive literature of lifelong learning into 12\nscenarios; (2) Identifying common techniques across all lifelong learning\nscenarios and classifying existing literature into various technique groups\nwithin each scenario; (3) Highlighting emerging techniques such as model\nexpansion and data selection, which were less explored in the pre-LLM era.\nThrough a detailed examination of these groups and their respective categories,\nthis survey aims to enhance the adaptability, reliability, and overall\nperformance of LLMs in real-world applications.\n", "date": "2024-06-10", "categories": []}
{"id": "2306.05101", "title": "Regularizing with Pseudo-Negatives for Continual Self-Supervised\n  Learning", "abstract": "  We introduce a novel Pseudo-Negative Regularization (PNR) framework for\neffective continual self-supervised learning (CSSL). Our PNR leverages\npseudo-negatives obtained through model-based augmentation in a way that newly\nlearned representations may not contradict what has been learned in the past.\nSpecifically, for the InfoNCE-based contrastive learning methods, we define\nsymmetric pseudo-negatives obtained from current and previous models and use\nthem in both main and regularization loss terms. Furthermore, we extend this\nidea to non-contrastive learning methods which do not inherently rely on\nnegatives. For these methods, a pseudo-negative is defined as the output from\nthe previous model for a differently augmented version of the anchor sample and\nis asymmetrically applied to the regularization term. Extensive experimental\nresults demonstrate that our PNR framework achieves state-of-the-art\nperformance in representation learning during CSSL by effectively balancing the\ntrade-off between plasticity and stability.\n", "date": "2023-06-08", "categories": []}
{"id": "2406.01229", "title": "AGALE: A Graph-Aware Continual Learning Evaluation Framework", "abstract": "  In recent years, continual learning (CL) techniques have made significant\nprogress in learning from streaming data while preserving knowledge across\nsequential tasks, particularly in the realm of euclidean data. To foster fair\nevaluation and recognize challenges in CL settings, several evaluation\nframeworks have been proposed, focusing mainly on the single- and multi-label\nclassification task on euclidean data. However, these evaluation frameworks are\nnot trivially applicable when the input data is graph-structured, as they do\nnot consider the topological structure inherent in graphs. Existing continual\ngraph learning (CGL) evaluation frameworks have predominantly focussed on\nsingle-label scenarios in the node classification (NC) task. This focus has\noverlooked the complexities of multi-label scenarios, where nodes may exhibit\naffiliations with multiple labels, simultaneously participating in multiple\ntasks. We develop a graph-aware evaluation (\\agale) framework that accommodates\nboth single-labeled and multi-labeled nodes, addressing the limitations of\nprevious evaluation frameworks. In particular, we define new incremental\nsettings and devise data partitioning algorithms tailored to CGL datasets. We\nperform extensive experiments comparing methods from the domains of continual\nlearning, continual graph learning, and dynamic graph learning (DGL). We\ntheoretically analyze \\agale and provide new insights about the role of\nhomophily in the performance of compared methods. We release our framework at\nhttps://github.com/Tianqi-py/AGALE.\n", "date": "2024-06-03", "categories": []}
{"id": "2406.04484", "title": "Step Out and Seek Around: On Warm-Start Training with Incremental Data", "abstract": "  Data often arrives in sequence over time in real-world deep learning\napplications such as autonomous driving. When new training data is available,\ntraining the model from scratch undermines the benefit of leveraging the\nlearned knowledge, leading to significant training costs. Warm-starting from a\npreviously trained checkpoint is the most intuitive way to retain knowledge and\nadvance learning. However, existing literature suggests that this warm-starting\ndegrades generalization. In this paper, we advocate for warm-starting but\nstepping out of the previous converging point, thus allowing a better\nadaptation to new data without compromising previous knowledge. We propose\nKnowledge Consolidation and Acquisition (CKCA), a continuous model improvement\nalgorithm with two novel components. First, a novel feature regularization\n(FeatReg) to retain and refine knowledge from existing checkpoints; Second, we\npropose adaptive knowledge distillation (AdaKD), a novel approach to forget\nmitigation and knowledge transfer. We tested our method on ImageNet using\nmultiple splits of the training data. Our approach achieves up to $8.39\\%$\nhigher top1 accuracy than the vanilla warm-starting and consistently\noutperforms the prior art with a large margin.\n", "date": "2024-06-06", "categories": []}
{"id": "2406.04584", "title": "CLoG: Benchmarking Continual Learning of Image Generation Models", "abstract": "  Continual Learning (CL) poses a significant challenge in Artificial\nIntelligence, aiming to mirror the human ability to incrementally acquire\nknowledge and skills. While extensive research has focused on CL within the\ncontext of classification tasks, the advent of increasingly powerful generative\nmodels necessitates the exploration of Continual Learning of Generative models\n(CLoG). This paper advocates for shifting the research focus from\nclassification-based CL to CLoG. We systematically identify the unique\nchallenges presented by CLoG compared to traditional classification-based CL.\nWe adapt three types of existing CL methodologies, replay-based,\nregularization-based, and parameter-isolation-based methods to generative tasks\nand introduce comprehensive benchmarks for CLoG that feature great diversity\nand broad task coverage. Our benchmarks and results yield intriguing insights\nthat can be valuable for developing future CLoG methods. Additionally, we will\nrelease a codebase designed to facilitate easy benchmarking and experimentation\nin CLoG publicly at https://github.com/linhaowei1/CLoG. We believe that\nshifting the research focus to CLoG will benefit the continual learning\ncommunity and illuminate the path for next-generation AI-generated content\n(AIGC) in a lifelong learning paradigm.\n", "date": "2024-06-07", "categories": []}
{"id": "2406.05114", "title": "The Expanding Scope of the Stability Gap: Unveiling its Presence in\n  Joint Incremental Learning of Homogeneous Tasks", "abstract": "  Recent research identified a temporary performance drop on previously learned\ntasks when transitioning to a new one. This drop is called the stability gap\nand has great consequences for continual learning: it complicates the direct\nemployment of continually learning since the worse-case performance at\ntask-boundaries is dramatic, it limits its potential as an energy-efficient\ntraining paradigm, and finally, the stability drop could result in a reduced\nfinal performance of the algorithm. In this paper, we show that the stability\ngap also occurs when applying joint incremental training of homogeneous tasks.\nIn this scenario, the learner continues training on the same data distribution\nand has access to all data from previous tasks. In addition, we show that in\nthis scenario, there exists a low-loss linear path to the next minima, but that\nSGD optimization does not choose this path. We perform further analysis\nincluding a finer batch-wise analysis which could provide insights towards\npotential solution directions.\n", "date": "2024-06-07", "categories": []}
{"id": "2401.08295", "title": "SAPT: A Shared Attention Framework for Parameter-Efficient Continual\n  Learning of Large Language Models", "abstract": "  The continual learning (CL) ability is vital for deploying large language\nmodels (LLMs) in the dynamic world. Existing methods devise the learning module\nto acquire task-specific knowledge with parameter-efficient tuning (PET) block\nand the selection module to pick out the corresponding one for the testing\ninput, aiming at handling the challenges of catastrophic forgetting and\nknowledge transfer in CL. However, these methods tend to address only one of\nthe challenges, ignoring the potential of aligning the two modules to\neffectively address catastrophic forgetting and knowledge transfer\nsimultaneously. To this end, we propose a novel Shared Attention Framework\n(SAPT), to align the PET learning and selection via the Shared Attentive\nLearning \\& Selection module. Extensive Experiments on two CL benchmarks\ndemonstrate the superiority of SAPT. Moreover, SAPT consistently demonstrates\nits superiority when we scale it to different model sizes (from 770M to 13B),\ndifferent model architectures (T5 and LLaMA-2) and unseen tasks.\n", "date": "2024-01-16", "categories": []}
{"id": "2309.02870", "title": "Rethinking Momentum Knowledge Distillation in Online Continual Learning", "abstract": "  Online Continual Learning (OCL) addresses the problem of training neural\nnetworks on a continuous data stream where multiple classification tasks emerge\nin sequence. In contrast to offline Continual Learning, data can be seen only\nonce in OCL, which is a very severe constraint. In this context, replay-based\nstrategies have achieved impressive results and most state-of-the-art\napproaches heavily depend on them. While Knowledge Distillation (KD) has been\nextensively used in offline Continual Learning, it remains under-exploited in\nOCL, despite its high potential. In this paper, we analyze the challenges in\napplying KD to OCL and give empirical justifications. We introduce a direct yet\neffective methodology for applying Momentum Knowledge Distillation (MKD) to\nmany flagship OCL methods and demonstrate its capabilities to enhance existing\napproaches. In addition to improving existing state-of-the-art accuracy by more\nthan $10\\%$ points on ImageNet100, we shed light on MKD internal mechanics and\nimpacts during training in OCL. We argue that similar to replay, MKD should be\nconsidered a central component of OCL. The code is available at\n\\url{https://github.com/Nicolas1203/mkd_ocl}.\n", "date": "2023-09-06", "categories": []}
{"id": "2406.03065", "title": "Decision Boundary-aware Knowledge Consolidation Generates Better\n  Instance-Incremental Learner", "abstract": "  Instance-incremental learning (IIL) focuses on learning continually with data\nof the same classes. Compared to class-incremental learning (CIL), the IIL is\nseldom explored because IIL suffers less from catastrophic forgetting (CF).\nHowever, besides retaining knowledge, in real-world deployment scenarios where\nthe class space is always predefined, continual and cost-effective model\npromotion with the potential unavailability of previous data is a more\nessential demand. Therefore, we first define a new and more practical IIL\nsetting as promoting the model's performance besides resisting CF with only new\nobservations. Two issues have to be tackled in the new IIL setting: 1) the\nnotorious catastrophic forgetting because of no access to old data, and 2)\nbroadening the existing decision boundary to new observations because of\nconcept drift. To tackle these problems, our key insight is to moderately\nbroaden the decision boundary to fail cases while retain old boundary. Hence,\nwe propose a novel decision boundary-aware distillation method with\nconsolidating knowledge to teacher to ease the student learning new knowledge.\nWe also establish the benchmarks on existing datasets Cifar-100 and ImageNet.\nNotably, extensive experiments demonstrate that the teacher model can be a\nbetter incremental learner than the student model, which overturns previous\nknowledge distillation-based methods treating student as the main role.\n", "date": "2024-06-05", "categories": []}
{"id": "2406.03216", "title": "Choice of PEFT Technique in Continual Learning: Prompt Tuning is Not All\n  You Need", "abstract": "  Recent Continual Learning (CL) methods have combined pretrained Transformers\nwith prompt tuning, a parameter-efficient fine-tuning (PEFT) technique. We\nargue that the choice of prompt tuning in prior works was an undefended and\nunablated decision, which has been uncritically adopted by subsequent research,\nbut warrants further research to understand its implications. In this paper, we\nconduct this research and find that the choice of prompt tuning as a PEFT\nmethod hurts the overall performance of the CL system. To illustrate this, we\nreplace prompt tuning with LoRA in two state-of-the-art continual learning\nmethods: Learning to Prompt and S-Prompts. These variants consistently achieve\nhigher accuracy across a wide range of domain-incremental and class-incremental\nbenchmarks, while being competitive in inference speed. Our work highlights a\ncrucial argument: unexamined choices can hinder progress in the field, and\nrigorous ablations, such as the PEFT method, are required to drive meaningful\nadoption of CL techniques in real-world applications.\n", "date": "2024-06-05", "categories": []}
{"id": "2406.01646", "title": "iKAN: Global Incremental Learning with KAN for Human Activity\n  Recognition Across Heterogeneous Datasets", "abstract": "  This work proposes an incremental learning (IL) framework for wearable sensor\nhuman activity recognition (HAR) that tackles two challenges simultaneously:\ncatastrophic forgetting and non-uniform inputs. The scalable framework, iKAN,\npioneers IL with Kolmogorov-Arnold Networks (KAN) to replace multi-layer\nperceptrons as the classifier that leverages the local plasticity and global\nstability of splines. To adapt KAN for HAR, iKAN uses task-specific feature\nbranches and a feature redistribution layer. Unlike existing IL methods that\nprimarily adjust the output dimension or the number of classifier nodes to\nadapt to new tasks, iKAN focuses on expanding the feature extraction branches\nto accommodate new inputs from different sensor modalities while maintaining\nconsistent dimensions and the number of classifier outputs. Continual learning\nacross six public HAR datasets demonstrated the iKAN framework's incremental\nlearning performance, with a last performance of 84.9\\% (weighted F1 score) and\nan average incremental performance of 81.34\\%, which significantly outperforms\nthe two existing incremental learning methods, such as EWC (51.42\\%) and\nexperience replay (59.92\\%).\n", "date": "2024-06-03", "categories": []}
{"id": "2406.02015", "title": "Parameterizing Federated Continual Learning for Reproducible Research", "abstract": "  Federated Learning (FL) systems evolve in heterogeneous and ever-evolving\nenvironments that challenge their performance. Under real deployments, the\nlearning tasks of clients can also evolve with time, which calls for the\nintegration of methodologies such as Continual Learning. To enable research\nreproducibility, we propose a set of experimental best practices that precisely\ncapture and emulate complex learning scenarios. Our framework, Freddie, is the\nfirst entirely configurable framework for Federated Continual Learning (FCL),\nand it can be seamlessly deployed on a large number of machines thanks to the\nuse of Kubernetes and containerization. We demonstrate the effectiveness of\nFreddie on two use cases, (i) large-scale FL on CIFAR100 and (ii) heterogeneous\ntask sequence on FCL, which highlight unaddressed performance challenges in FCL\nscenarios.\n", "date": "2024-06-04", "categories": []}
{"id": "2406.02428", "title": "Harnessing Neural Unit Dynamics for Effective and Scalable\n  Class-Incremental Learning", "abstract": "  Class-incremental learning (CIL) aims to train a model to learn new classes\nfrom non-stationary data streams without forgetting old ones. In this paper, we\npropose a new kind of connectionist model by tailoring neural unit dynamics\nthat adapt the behavior of neural networks for CIL. In each training session,\nit introduces a supervisory mechanism to guide network expansion whose growth\nsize is compactly commensurate with the intrinsic complexity of a newly\narriving task. This constructs a near-minimal network while allowing the model\nto expand its capacity when cannot sufficiently hold new classes. At inference\ntime, it automatically reactivates the required neural units to retrieve\nknowledge and leaves the remaining inactivated to prevent interference. We name\nour model AutoActivator, which is effective and scalable. To gain insights into\nthe neural unit dynamics, we theoretically analyze the model's convergence\nproperty via a universal approximation theorem on learning sequential mappings,\nwhich is under-explored in the CIL community. Experiments show that our method\nachieves strong CIL performance in rehearsal-free and minimal-expansion\nsettings with different backbones.\n", "date": "2024-06-04", "categories": []}
{"id": "2210.03869", "title": "TAME: Task Agnostic Continual Learning using Multiple Experts", "abstract": "  The goal of lifelong learning is to continuously learn from non-stationary\ndistributions, where the non-stationarity is typically imposed by a sequence of\ndistinct tasks. Prior works have mostly considered idealistic settings, where\nthe identity of tasks is known at least at training. In this paper we focus on\na fundamentally harder, so-called task-agnostic setting where the task\nidentities are not known and the learning machine needs to infer them from the\nobservations. Our algorithm, which we call TAME (Task-Agnostic continual\nlearning using Multiple Experts), automatically detects the shift in data\ndistributions and switches between task expert networks in an online manner. At\ntraining, the strategy for switching between tasks hinges on an extremely\nsimple observation that for each new coming task there occurs a\nstatistically-significant deviation in the value of the loss function that\nmarks the onset of this new task. At inference, the switching between experts\nis governed by the selector network that forwards the test sample to its\nrelevant expert network. The selector network is trained on a small subset of\ndata drawn uniformly at random. We control the growth of the task expert\nnetworks as well as selector network by employing online pruning. Our\nexperimental results show the efficacy of our approach on benchmark continual\nlearning data sets, outperforming the previous task-agnostic methods and even\nthe techniques that admit task identities at both training and testing, while\nat the same time using a comparable model size.\n", "date": "2022-10-08", "categories": []}
{"id": "2403.05890", "title": "Towards Efficient Replay in Federated Incremental Learning", "abstract": "  In Federated Learning (FL), the data in each client is typically assumed\nfixed or static. However, data often comes in an incremental manner in\nreal-world applications, where the data domain may increase dynamically. In\nthis work, we study catastrophic forgetting with data heterogeneity in\nFederated Incremental Learning (FIL) scenarios where edge clients may lack\nenough storage space to retain full data. We propose to employ a simple,\ngeneric framework for FIL named Re-Fed, which can coordinate each client to\ncache important samples for replay. More specifically, when a new task arrives,\neach client first caches selected previous samples based on their global and\nlocal importance. Then, the client trains the local model with both the cached\nsamples and the samples from the new task. Theoretically, we analyze the\nability of Re-Fed to discover important samples for replay thus alleviating the\ncatastrophic forgetting problem. Moreover, we empirically show that Re-Fed\nachieves competitive performance compared to state-of-the-art methods.\n", "date": "2024-03-09", "categories": []}
{"id": "2403.11549", "title": "Boosting Continual Learning of Vision-Language Models via\n  Mixture-of-Experts Adapters", "abstract": "  Continual learning can empower vision-language models to continuously acquire\nnew knowledge, without the need for access to the entire historical dataset.\nHowever, mitigating the performance degradation in large-scale models is\nnon-trivial due to (i) parameter shifts throughout lifelong learning and (ii)\nsignificant computational burdens associated with full-model tuning. In this\nwork, we present a parameter-efficient continual learning framework to\nalleviate long-term forgetting in incremental learning with vision-language\nmodels. Our approach involves the dynamic expansion of a pre-trained CLIP\nmodel, through the integration of Mixture-of-Experts (MoE) adapters in response\nto new tasks. To preserve the zero-shot recognition capability of\nvision-language models, we further introduce a Distribution Discriminative\nAuto-Selector (DDAS) that automatically routes in-distribution and\nout-of-distribution inputs to the MoE Adapter and the original CLIP,\nrespectively. Through extensive experiments across various settings, our\nproposed method consistently outperforms previous state-of-the-art approaches\nwhile concurrently reducing parameter training burdens by 60%. Our code locates\nat https://github.com/JiazuoYu/MoE-Adapters4CL\n", "date": "2024-03-18", "categories": []}
{"id": "2406.01054", "title": "Confidence-Based Task Prediction in Continual Disease Classification\n  Using Probability Distribution", "abstract": "  Deep learning models are widely recognized for their effectiveness in\nidentifying medical image findings in disease classification. However, their\nlimitations become apparent in the dynamic and ever-changing clinical\nenvironment, characterized by the continuous influx of newly annotated medical\ndata from diverse sources. In this context, the need for continual learning\nbecomes particularly paramount, not only to adapt to evolving medical scenarios\nbut also to ensure the privacy of healthcare data. In our research, we\nemphasize the utilization of a network comprising expert classifiers, where a\nnew expert classifier is added each time a new task is introduced. We present\nCTP, a task-id predictor that utilizes confidence scores, leveraging the\nprobability distribution (logits) of the classifier to accurately determine the\ntask-id at inference time. Logits are adjusted to ensure that classifiers yield\na high-entropy distribution for data associated with tasks other than their\nown. By defining a noise region in the distribution and computing confidence\nscores, CTP achieves superior performance when compared to other relevant\ncontinual learning methods. Additionally, the performance of CTP can be further\nimproved by providing it with a continuum of data at the time of inference.\n", "date": "2024-06-03", "categories": []}
{"id": "2310.00154", "title": "Primal Dual Continual Learning: Balancing Stability and Plasticity\n  through Adaptive Memory Allocation", "abstract": "  Continual learning is inherently a constrained learning problem. The goal is\nto learn a predictor under a no-forgetting requirement. Although several prior\nstudies formulate it as such, they do not solve the constrained problem\nexplicitly. In this work, we show that it is both possible and beneficial to\nundertake the constrained optimization problem directly. To do this, we\nleverage recent results in constrained learning through Lagrangian duality. We\nfocus on memory-based methods, where a small subset of samples from previous\ntasks can be stored in a replay buffer. In this setting, we analyze two\nversions of the continual learning problem: a coarse approach with constraints\nat the task level and a fine approach with constraints at the sample level. We\nshow that dual variables indicate the sensitivity of the optimal value of the\ncontinual learning problem with respect to constraint perturbations. We then\nleverage this result to partition the buffer in the coarse approach, allocating\nmore resources to harder tasks, and to populate the buffer in the fine\napproach, including only impactful samples. We derive a deviation bound on dual\nvariables as sensitivity indicators, and empirically corroborate this result in\ndiverse continual learning benchmarks. We also discuss the limitations of these\nmethods with respect to the amount of memory available and the expressiveness\nof the parametrization.\n", "date": "2023-09-29", "categories": []}
{"id": "2311.03732", "title": "Learning to Learn for Few-shot Continual Active Learning", "abstract": "  Continual learning strives to ensure stability in solving previously seen\ntasks while demonstrating plasticity in a novel domain. Recent advances in\ncontinual learning are mostly confined to a supervised learning setting,\nespecially in NLP domain. In this work, we consider a few-shot continual active\nlearning setting where labeled data are inadequate, and unlabeled data are\nabundant but with a limited annotation budget. We exploit meta-learning and\npropose a method, called Meta-Continual Active Learning. This method\nsequentially queries the most informative examples from a pool of unlabeled\ndata for annotation to enhance task-specific performance and tackle continual\nlearning problems through meta-objective. Specifically, we employ meta-learning\nand experience replay to address inter-task confusion and catastrophic\nforgetting. We further incorporate textual augmentations to avoid memory\nover-fitting caused by experience replay and sample queries, thereby ensuring\ngeneralization. We conduct extensive experiments on benchmark text\nclassification datasets from diverse domains to validate the feasibility and\neffectiveness of meta-continual active learning. We also analyze the impact of\ndifferent active learning strategies on various meta continual learning models.\nThe experimental results demonstrate that introducing randomness into sample\nselection is the best default strategy for maintaining generalization in\nmeta-continual learning framework.\n", "date": "2023-11-07", "categories": []}
{"id": "2403.03938", "title": "GUIDE: Guidance-based Incremental Learning with Diffusion Models", "abstract": "  We introduce GUIDE, a novel continual learning approach that directs\ndiffusion models to rehearse samples at risk of being forgotten. Existing\ngenerative strategies combat catastrophic forgetting by randomly sampling\nrehearsal examples from a generative model. Such an approach contradicts\nbuffer-based approaches where sampling strategy plays an important role. We\npropose to bridge this gap by incorporating classifier guidance into the\ndiffusion process to produce rehearsal examples specifically targeting\ninformation forgotten by a continuously trained model. This approach enables\nthe generation of samples from preceding task distributions, which are more\nlikely to be misclassified in the context of recently encountered classes. Our\nexperimental results show that GUIDE significantly reduces catastrophic\nforgetting, outperforming conventional random sampling approaches and\nsurpassing recent state-of-the-art methods in continual learning with\ngenerative replay.\n", "date": "2024-03-06", "categories": []}
{"id": "2405.20600", "title": "Multi-label Class Incremental Emotion Decoding with Augmented Emotional\n  Semantics Learning", "abstract": "  Emotion decoding plays an important role in affective human-computer\ninteraction. However, previous studies ignored the dynamic real-world scenario,\nwhere human experience a blend of multiple emotions which are incrementally\nintegrated into the model, leading to the multi-label class incremental\nlearning (MLCIL) problem. Existing methods have difficulty in solving MLCIL\nissue due to notorious catastrophic forgetting caused by partial label problem\nand inadequate label semantics mining. In this paper, we propose an augmented\nemotional semantics learning framework for multi-label class incremental\nemotion decoding. Specifically, we design an augmented emotional relation graph\nmodule with label disambiguation to handle the past-missing partial label\nproblem. Then, we leverage domain knowledge from affective dimension space to\nalleviate future-missing partial label problem by knowledge distillation.\nBesides, an emotional semantics learning module is constructed with a graph\nautoencoder to obtain emotion embeddings in order to guide the\nsemantic-specific feature decoupling for better multi-label learning. Extensive\nexperiments on three datasets show the superiority of our method for improving\nemotion decoding performance and mitigating forgetting on MLCIL problem.\n", "date": "2024-05-31", "categories": []}
{"id": "2310.11952", "title": "Recasting Continual Learning as Sequence Modeling", "abstract": "  In this work, we aim to establish a strong connection between two significant\nbodies of machine learning research: continual learning and sequence modeling.\nThat is, we propose to formulate continual learning as a sequence modeling\nproblem, allowing advanced sequence models to be utilized for continual\nlearning. Under this formulation, the continual learning process becomes the\nforward pass of a sequence model. By adopting the meta-continual learning (MCL)\nframework, we can train the sequence model at the meta-level, on multiple\ncontinual learning episodes. As a specific example of our new formulation, we\ndemonstrate the application of Transformers and their efficient variants as MCL\nmethods. Our experiments on seven benchmarks, covering both classification and\nregression, show that sequence models can be an attractive solution for general\nMCL.\n", "date": "2023-10-18", "categories": []}
{"id": "2402.03917", "title": "Elastic Feature Consolidation for Cold Start Exemplar-Free Incremental\n  Learning", "abstract": "  Exemplar-Free Class Incremental Learning (EFCIL) aims to learn from a\nsequence of tasks without having access to previous task data. In this paper,\nwe consider the challenging Cold Start scenario in which insufficient data is\navailable in the first task to learn a high-quality backbone. This is\nespecially challenging for EFCIL since it requires high plasticity, which\nresults in feature drift which is difficult to compensate for in the\nexemplar-free setting. To address this problem, we propose a simple and\neffective approach that consolidates feature representations by regularizing\ndrift in directions highly relevant to previous tasks and employs prototypes to\nreduce task-recency bias. Our method, called Elastic Feature Consolidation\n(EFC), exploits a tractable second-order approximation of feature drift based\non an Empirical Feature Matrix (EFM). The EFM induces a pseudo-metric in\nfeature space which we use to regularize feature drift in important directions\nand to update Gaussian prototypes used in a novel asymmetric cross entropy loss\nwhich effectively balances prototype rehearsal with data from new tasks.\nExperimental results on CIFAR-100, Tiny-ImageNet, ImageNet-Subset and\nImageNet-1K demonstrate that Elastic Feature Consolidation is better able to\nlearn new tasks by maintaining model plasticity and significantly outperform\nthe state-of-the-art.\n", "date": "2024-02-06", "categories": []}
{"id": "2405.19005", "title": "Auto-selected Knowledge Adapters for Lifelong Person Re-identification", "abstract": "  Lifelong Person Re-Identification (LReID) extends traditional ReID by\nrequiring systems to continually learn from non-overlapping datasets across\ndifferent times and locations, adapting to new identities while preserving\nknowledge of previous ones. Existing approaches, either rehearsal-free or\nrehearsal-based, still suffer from the problem of catastrophic forgetting since\nthey try to cram diverse knowledge into one fixed model. To overcome this\nlimitation, we introduce a novel framework AdalReID, that adopts knowledge\nadapters and a parameter-free auto-selection mechanism for lifelong learning.\nConcretely, we incrementally build distinct adapters to learn domain-specific\nknowledge at each step, which can effectively learn and preserve knowledge\nacross different datasets. Meanwhile, the proposed auto-selection strategy\nadaptively calculates the knowledge similarity between the input set and the\nadapters. On the one hand, the appropriate adapters are selected for the inputs\nto process ReID, and on the other hand, the knowledge interaction and fusion\nbetween adapters are enhanced to improve the generalization ability of the\nmodel. Extensive experiments are conducted to demonstrate the superiority of\nour AdalReID, which significantly outperforms SOTAs by about 10$\\sim$20\\% mAP\non both seen and unseen domains.\n", "date": "2024-05-29", "categories": []}
{"id": "2405.19568", "title": "Organizing Background to Explore Latent Classes for Incremental Few-shot\n  Semantic Segmentation", "abstract": "  The goal of incremental Few-shot Semantic Segmentation (iFSS) is to extend\npre-trained segmentation models to new classes via few annotated images without\naccess to old training data. During incrementally learning novel classes, the\ndata distribution of old classes will be destroyed, leading to catastrophic\nforgetting. Meanwhile, the novel classes have only few samples, making models\nimpossible to learn the satisfying representations of novel classes. For the\niFSS problem, we propose a network called OINet, i.e., the background embedding\nspace \\textbf{O}rganization and prototype \\textbf{I}nherit Network.\nSpecifically, when training base classes, OINet uses multiple classification\nheads for the background and sets multiple sub-class prototypes to reserve\nembedding space for the latent novel classes. During incrementally learning\nnovel classes, we propose a strategy to select the sub-class prototypes that\nbest match the current learning novel classes and make the novel classes\ninherit the selected prototypes' embedding space. This operation allows the\nnovel classes to be registered in the embedding space using few samples without\naffecting the distribution of the base classes. Results on Pascal-VOC and COCO\nshow that OINet achieves a new state of the art.\n", "date": "2024-05-29", "categories": []}
{"id": "2405.20236", "title": "Disentangling and Mitigating the Impact of Task Similarity for Continual\n  Learning", "abstract": "  Continual learning of partially similar tasks poses a challenge for\nartificial neural networks, as task similarity presents both an opportunity for\nknowledge transfer and a risk of interference and catastrophic forgetting.\nHowever, it remains unclear how task similarity in input features and readout\npatterns influences knowledge transfer and forgetting, as well as how they\ninteract with common algorithms for continual learning. Here, we develop a\nlinear teacher-student model with latent structure and show analytically that\nhigh input feature similarity coupled with low readout similarity is\ncatastrophic for both knowledge transfer and retention. Conversely, the\nopposite scenario is relatively benign. Our analysis further reveals that\ntask-dependent activity gating improves knowledge retention at the expense of\ntransfer, while task-dependent plasticity gating does not affect either\nretention or transfer performance at the over-parameterized limit. In contrast,\nweight regularization based on the Fisher information metric significantly\nimproves retention, regardless of task similarity, without compromising\ntransfer performance. Nevertheless, its diagonal approximation and\nregularization in the Euclidean space are much less robust against task\nsimilarity. We demonstrate consistent results in a permuted MNIST task with\nlatent variables. Overall, this work provides insights into when continual\nlearning is difficult and how to mitigate it.\n", "date": "2024-05-30", "categories": []}
{"id": "2405.18663", "title": "Lifelong Learning and Selective Forgetting via Contrastive Strategy", "abstract": "  Lifelong learning aims to train a model with good performance for new tasks\nwhile retaining the capacity of previous tasks. However, some practical\nscenarios require the system to forget undesirable knowledge due to privacy\nissues, which is called selective forgetting. The joint task of the two is\ndubbed Learning with Selective Forgetting (LSF). In this paper, we propose a\nnew framework based on contrastive strategy for LSF. Specifically, for the\npreserved classes (tasks), we make features extracted from different samples\nwithin a same class compacted. And for the deleted classes, we make the\nfeatures from different samples of a same class dispersed and irregular, i.e.,\nthe network does not have any regular response to samples from a specific\ndeleted class as if the network has no training at all. Through maintaining or\ndisturbing the feature distribution, the forgetting and memory of different\nclasses can be or independent of each other. Experiments are conducted on four\nbenchmark datasets, and our method acieves new state-of-the-art.\n", "date": "2024-05-28", "categories": []}
{"id": "2405.18756", "title": "Provable Contrastive Continual Learning", "abstract": "  Continual learning requires learning incremental tasks with dynamic data\ndistributions. So far, it has been observed that employing a combination of\ncontrastive loss and distillation loss for training in continual learning\nyields strong performance. To the best of our knowledge, however, this\ncontrastive continual learning framework lacks convincing theoretical\nexplanations. In this work, we fill this gap by establishing theoretical\nperformance guarantees, which reveal how the performance of the model is\nbounded by training losses of previous tasks in the contrastive continual\nlearning framework. Our theoretical explanations further support the idea that\npre-training can benefit continual learning. Inspired by our theoretical\nanalysis of these guarantees, we propose a novel contrastive continual learning\nalgorithm called CILA, which uses adaptive distillation coefficients for\ndifferent tasks. These distillation coefficients are easily computed by the\nratio between average distillation losses and average contrastive losses from\nprevious tasks. Our method shows great improvement on standard benchmarks and\nachieves new state-of-the-art performance.\n", "date": "2024-05-29", "categories": []}
{"id": "2405.18758", "title": "Learning to Continually Learn with the Bayesian Principle", "abstract": "  In the present era of deep learning, continual learning research is mainly\nfocused on mitigating forgetting when training a neural network with stochastic\ngradient descent on a non-stationary stream of data. On the other hand, in the\nmore classical literature of statistical machine learning, many models have\nsequential Bayesian update rules that yield the same learning outcome as the\nbatch training, i.e., they are completely immune to catastrophic forgetting.\nHowever, they are often overly simple to model complex real-world data. In this\nwork, we adopt the meta-learning paradigm to combine the strong\nrepresentational power of neural networks and simple statistical models'\nrobustness to forgetting. In our novel meta-continual learning framework,\ncontinual learning takes place only in statistical models via ideal sequential\nBayesian update rules, while neural networks are meta-learned to bridge the raw\ndata and the statistical models. Since the neural networks remain fixed during\ncontinual learning, they are protected from catastrophic forgetting. This\napproach not only achieves significantly improved performance but also exhibits\nexcellent scalability. Since our approach is domain-agnostic and\nmodel-agnostic, it can be applied to a wide range of problems and easily\nintegrated with existing model architectures.\n", "date": "2024-05-29", "categories": []}
{"id": "2405.19074", "title": "Resurrecting Old Classes with New Data for Exemplar-Free Continual\n  Learning", "abstract": "  Continual learning methods are known to suffer from catastrophic forgetting,\na phenomenon that is particularly hard to counter for methods that do not store\nexemplars of previous tasks. Therefore, to reduce potential drift in the\nfeature extractor, existing exemplar-free methods are typically evaluated in\nsettings where the first task is significantly larger than subsequent tasks.\nTheir performance drops drastically in more challenging settings starting with\na smaller first task. To address this problem of feature drift estimation for\nexemplar-free methods, we propose to adversarially perturb the current samples\nsuch that their embeddings are close to the old class prototypes in the old\nmodel embedding space. We then estimate the drift in the embedding space from\nthe old to the new model using the perturbed images and compensate the\nprototypes accordingly. We exploit the fact that adversarial samples are\ntransferable from the old to the new feature space in a continual learning\nsetting. The generation of these images is simple and computationally cheap. We\ndemonstrate in our experiments that the proposed approach better tracks the\nmovement of prototypes in embedding space and outperforms existing methods on\nseveral standard continual learning benchmarks as well as on fine-grained\ndatasets. Code is available at https://github.com/dipamgoswami/ADC.\n", "date": "2024-05-29", "categories": []}
{"id": "2405.19234", "title": "Forward-Backward Knowledge Distillation for Continual Clustering", "abstract": "  Unsupervised Continual Learning (UCL) is a burgeoning field in machine\nlearning, focusing on enabling neural networks to sequentially learn tasks\nwithout explicit label information. Catastrophic Forgetting (CF), where models\nforget previously learned tasks upon learning new ones, poses a significant\nchallenge in continual learning, especially in UCL, where labeled information\nof data is not accessible. CF mitigation strategies, such as knowledge\ndistillation and replay buffers, often face memory inefficiency and privacy\nissues. Although current research in UCL has endeavored to refine data\nrepresentations and address CF in streaming data contexts, there is a\nnoticeable lack of algorithms specifically designed for unsupervised\nclustering. To fill this gap, in this paper, we introduce the concept of\nUnsupervised Continual Clustering (UCC). We propose Forward-Backward Knowledge\nDistillation for unsupervised Continual Clustering (FBCC) to counteract CF\nwithin the context of UCC. FBCC employs a single continual learner (the\n``teacher'') with a cluster projector, along with multiple student models, to\naddress the CF issue. The proposed method consists of two phases: Forward\nKnowledge Distillation, where the teacher learns new clusters while retaining\nknowledge from previous tasks with guidance from specialized student models,\nand Backward Knowledge Distillation, where a student model mimics the teacher's\nbehavior to retain task-specific knowledge, aiding the teacher in subsequent\ntasks. FBCC marks a pioneering approach to UCC, demonstrating enhanced\nperformance and memory efficiency in clustering across various tasks,\noutperforming the application of clustering algorithms to the latent space of\nstate-of-the-art UCL algorithms.\n", "date": "2024-05-29", "categories": []}
{"id": "2310.08204", "title": "STELLA: Continual Audio-Video Pre-training with Spatio-Temporal\n  Localized Alignment", "abstract": "  Continuously learning a variety of audio-video semantics over time is crucial\nfor audio-related reasoning tasks in our ever-evolving world. However, this is\na nontrivial problem and poses two critical challenges: sparse spatio-temporal\ncorrelation between audio-video pairs and multimodal correlation overwriting\nthat forgets audio-video relations. To tackle this problem, we propose a new\ncontinual audio-video pre-training method with two novel ideas: (1) Localized\nPatch Importance Scoring: we introduce a multimodal encoder to determine the\nimportance score for each patch, emphasizing semantically intertwined\naudio-video patches. (2) Replay-guided Correlation Assessment: to reduce the\ncorruption of previously learned audiovisual knowledge due to drift, we propose\nto assess the correlation of the current patches on the past steps to identify\nthe patches exhibiting high correlations with the past steps. Based on the\nresults from the two ideas, we perform probabilistic patch selection for\neffective continual audio-video pre-training. Experimental validation on\nmultiple benchmarks shows that our method achieves a 3.69%p of relative\nperformance gain in zero-shot retrieval tasks compared to strong continual\nlearning baselines, while reducing memory consumption by ~45%.\n", "date": "2023-10-12", "categories": []}
{"id": "2405.17457", "title": "Data-Free Federated Class Incremental Learning with Diffusion-Based\n  Generative Memory", "abstract": "  Federated Class Incremental Learning (FCIL) is a critical yet largely\nunderexplored issue that deals with the dynamic incorporation of new classes\nwithin federated learning (FL). Existing methods often employ generative\nadversarial networks (GANs) to produce synthetic images to address privacy\nconcerns in FL. However, GANs exhibit inherent instability and high\nsensitivity, compromising the effectiveness of these methods. In this paper, we\nintroduce a novel data-free federated class incremental learning framework with\ndiffusion-based generative memory (DFedDGM) to mitigate catastrophic forgetting\nby generating stable, high-quality images through diffusion models. We design a\nnew balanced sampler to help train the diffusion models to alleviate the common\nnon-IID problem in FL, and introduce an entropy-based sample filtering\ntechnique from an information theory perspective to enhance the quality of\ngenerative samples. Finally, we integrate knowledge distillation with a\nfeature-based regularization term for better knowledge transfer. Our framework\ndoes not incur additional communication costs compared to the baseline FedAvg\nmethod. Extensive experiments across multiple datasets demonstrate that our\nmethod significantly outperforms existing baselines, e.g., over a 4%\nimprovement in average accuracy on the Tiny-ImageNet dataset.\n", "date": "2024-05-22", "categories": []}
{"id": "2405.17583", "title": "Understanding Forgetting in Continual Learning with Linear Regression", "abstract": "  Continual learning, focused on sequentially learning multiple tasks, has\ngained significant attention recently. Despite the tremendous progress made in\nthe past, the theoretical understanding, especially factors contributing to\ncatastrophic forgetting, remains relatively unexplored. In this paper, we\nprovide a general theoretical analysis of forgetting in the linear regression\nmodel via Stochastic Gradient Descent (SGD) applicable to both\nunderparameterized and overparameterized regimes. Our theoretical framework\nreveals some interesting insights into the intricate relationship between task\nsequence and algorithmic parameters, an aspect not fully captured in previous\nstudies due to their restrictive assumptions. Specifically, we demonstrate\nthat, given a sufficiently large data size, the arrangement of tasks in a\nsequence, where tasks with larger eigenvalues in their population data\ncovariance matrices are trained later, tends to result in increased forgetting.\nAdditionally, our findings highlight that an appropriate choice of step size\nwill help mitigate forgetting in both underparameterized and overparameterized\nsettings. To validate our theoretical analysis, we conducted simulation\nexperiments on both linear regression models and Deep Neural Networks (DNNs).\nResults from these simulations substantiate our theoretical findings.\n", "date": "2024-05-27", "categories": []}
{"id": "2305.08698", "title": "Continual Multimodal Knowledge Graph Construction", "abstract": "  Current Multimodal Knowledge Graph Construction (MKGC) models struggle with\nthe real-world dynamism of continuously emerging entities and relations, often\nsuccumbing to catastrophic forgetting-loss of previously acquired knowledge.\nThis study introduces benchmarks aimed at fostering the development of the\ncontinual MKGC domain. We further introduce MSPT framework, designed to\nsurmount the shortcomings of existing MKGC approaches during multimedia data\nprocessing. MSPT harmonizes the retention of learned knowledge (stability) and\nthe integration of new data (plasticity), outperforming current continual\nlearning and multimodal methods. Our results confirm MSPT's superior\nperformance in evolving knowledge environments, showcasing its capacity to\nnavigate balance between stability and plasticity.\n", "date": "2023-05-15", "categories": []}
{"id": "2402.10447", "title": "Incremental Sequence Labeling: A Tale of Two Shifts", "abstract": "  The incremental sequence labeling task involves continuously learning new\nclasses over time while retaining knowledge of the previous ones. Our\ninvestigation identifies two significant semantic shifts: E2O (where the model\nmislabels an old entity as a non-entity) and O2E (where the model labels a\nnon-entity or old entity as a new entity). Previous research has predominantly\nfocused on addressing the E2O problem, neglecting the O2E issue. This\nnegligence results in a model bias towards classifying new data samples as\nbelonging to the new class during the learning process. To address these\nchallenges, we propose a novel framework, Incremental Sequential Labeling\nwithout Semantic Shifts (IS3). Motivated by the identified semantic shifts (E2O\nand O2E), IS3 aims to mitigate catastrophic forgetting in models. As for the\nE2O problem, we use knowledge distillation to maintain the model's\ndiscriminative ability for old entities. Simultaneously, to tackle the O2E\nproblem, we alleviate the model's bias towards new entities through debiased\nloss and optimization levels. Our experimental evaluation, conducted on three\ndatasets with various incremental settings, demonstrates the superior\nperformance of IS3 compared to the previous state-of-the-art method by a\nsignificant margin.The data, code, and scripts are publicly available at\nhttps://github.com/zzz47zzz/codebase-for-incremental-learning-with-llm.\n", "date": "2024-02-16", "categories": []}
{"id": "2403.01244", "title": "Mitigating Catastrophic Forgetting in Large Language Models with\n  Self-Synthesized Rehearsal", "abstract": "  Large language models (LLMs) suffer from catastrophic forgetting during\ncontinual learning. Conventional rehearsal-based methods rely on previous\ntraining data to retain the model's ability, which may not be feasible in\nreal-world applications. When conducting continual learning based on a\npublicly-released LLM checkpoint, the availability of the original training\ndata may be non-existent. To address this challenge, we propose a framework\ncalled Self-Synthesized Rehearsal (SSR) that uses the LLM to generate synthetic\ninstances for rehearsal. Concretely, we first employ the base LLM for\nin-context learning to generate synthetic instances. Subsequently, we utilize\nthe latest LLM to refine the instance outputs based on the synthetic inputs,\npreserving its acquired ability. Finally, we select diverse high-quality\nsynthetic instances for rehearsal in future stages. Experimental results\ndemonstrate that SSR achieves superior or comparable performance compared to\nconventional rehearsal-based approaches while being more data-efficient.\nBesides, SSR effectively preserves the generalization capabilities of LLMs in\ngeneral domains.\n", "date": "2024-03-02", "categories": []}
{"id": "2405.16113", "title": "Enabling On-Device Learning via Experience Replay with Efficient Dataset\n  Condensation", "abstract": "  Upon deployment to edge devices, it is often desirable for a model to further\nlearn from streaming data to improve accuracy. However, extracting\nrepresentative features from such data is challenging because it is typically\nunlabeled, non-independent and identically distributed (non-i.i.d), and is seen\nonly once. To mitigate this issue, a common strategy is to maintain a small\ndata buffer on the edge device to hold the most representative data for further\nlearning. As most data is either never stored or quickly discarded, identifying\nthe most representative data to avoid significant information loss becomes\ncritical. In this paper, we propose an on-device framework that addresses this\nissue by condensing incoming data into more informative samples. Specifically,\nto effectively handle unlabeled incoming data, we propose a pseudo-labeling\ntechnique designed for unlabeled on-device learning environments. Additionally,\nwe develop a dataset condensation technique that only requires little\ncomputation resources. To counteract the effects of noisy labels during the\ncondensation process, we further utilize a contrastive learning objective to\nimprove the purity of class data within the buffer. Our empirical results\nindicate substantial improvements over existing methods, particularly when\nbuffer capacity is severely restricted. For instance, with a buffer capacity of\njust one sample per class, our method achieves an accuracy that outperforms the\nbest existing baseline by 58.4% on the CIFAR-10 dataset.\n", "date": "2024-05-25", "categories": []}
{"id": "2405.17022", "title": "Compositional Few-Shot Class-Incremental Learning", "abstract": "  Few-shot class-incremental learning (FSCIL) is proposed to continually learn\nfrom novel classes with only a few samples after the (pre-)training on base\nclasses with sufficient data. However, this remains a challenge. In contrast,\nhumans can easily recognize novel classes with a few samples. Cognitive science\ndemonstrates that an important component of such human capability is\ncompositional learning. This involves identifying visual primitives from\nlearned knowledge and then composing new concepts using these transferred\nprimitives, making incremental learning both effective and interpretable. To\nimitate human compositional learning, we propose a cognitive-inspired method\nfor the FSCIL task. We define and build a compositional model based on set\nsimilarities, and then equip it with a primitive composition module and a\nprimitive reuse module. In the primitive composition module, we propose to\nutilize the Centered Kernel Alignment (CKA) similarity to approximate the\nsimilarity between primitive sets, allowing the training and evaluation based\non primitive compositions. In the primitive reuse module, we enhance primitive\nreusability by classifying inputs based on primitives replaced with the closest\nprimitives from other classes. Experiments on three datasets validate our\nmethod, showing it outperforms current state-of-the-art methods with improved\ninterpretability. Our code is available at\nhttps://github.com/Zoilsen/Comp-FSCIL.\n", "date": "2024-05-27", "categories": []}
{"id": "2405.17054", "title": "Improving Data-aware and Parameter-aware Robustness for Continual\n  Learning", "abstract": "  The goal of Continual Learning (CL) task is to continuously learn multiple\nnew tasks sequentially while achieving a balance between the plasticity and\nstability of new and old knowledge. This paper analyzes that this insufficiency\narises from the ineffective handling of outliers, leading to abnormal gradients\nand unexpected model updates. To address this issue, we enhance the data-aware\nand parameter-aware robustness of CL, proposing a Robust Continual Learning\n(RCL) method. From the data perspective, we develop a contrastive loss based on\nthe concepts of uniformity and alignment, forming a feature distribution that\nis more applicable to outliers. From the parameter perspective, we present a\nforward strategy for worst-case perturbation and apply robust gradient\nprojection to the parameters. The experimental results on three benchmarks show\nthat the proposed method effectively maintains robustness and achieves new\nstate-of-the-art (SOTA) results. The code is available at:\nhttps://github.com/HanxiXiao/RCL\n", "date": "2024-05-27", "categories": []}
{"id": "2310.00113", "title": "HyperMask: Adaptive Hypernetwork-based Masks for Continual Learning", "abstract": "  Artificial neural networks suffer from catastrophic forgetting when they are\nsequentially trained on multiple tasks. Many continual learning (CL) strategies\nare trying to overcome this problem. One of the most effective is the\nhypernetwork-based approach. The hypernetwork generates the weights of a target\nmodel based on the task's identity. The model's main limitation is that, in\npractice, the hypernetwork can produce completely different architectures for\nsubsequent tasks. To solve such a problem, we use the lottery ticket\nhypothesis, which postulates the existence of sparse subnetworks, named winning\ntickets, that preserve the performance of a whole network. In the paper, we\npropose a method called HyperMask, which dynamically filters a target network\ndepending on the CL task. The hypernetwork produces semi-binary masks to obtain\ndedicated target subnetworks. Moreover, due to the lottery ticket hypothesis,\nwe can use a single network with weighted subnets. Depending on the task, the\nimportance of some weights may be dynamically enhanced while others may be\nweakened. HyperMask achieves competitive results in several CL datasets and, in\nsome scenarios, goes beyond the state-of-the-art scores, both with derived and\nunknown task identities.\n", "date": "2023-09-29", "categories": []}
{"id": "2405.15155", "title": "CLIP model is an Efficient Online Lifelong Learner", "abstract": "  Online Lifelong Learning (OLL) addresses the challenge of learning from\ncontinuous and non-stationary data streams. Existing online lifelong learning\nmethods based on image classification models often require preset conditions\nsuch as the total number of classes or maximum memory capacity, which hinders\nthe realization of real never-ending learning and renders them impractical for\nreal-world scenarios. In this work, we propose that vision-language models,\nsuch as Contrastive Language-Image Pretraining (CLIP), are more suitable\ncandidates for online lifelong learning. We discover that maintaining symmetry\nbetween image and text is crucial during Parameter-Efficient Tuning (PET) for\nCLIP model in online lifelong learning. To this end, we introduce the Symmetric\nImage-Text (SIT) tuning strategy. We conduct extensive experiments on multiple\nlifelong learning benchmark datasets and elucidate the effectiveness of SIT\nthrough gradient analysis. Additionally, we assess the impact of lifelong\nlearning on generalizability of CLIP and found that tuning the image encoder is\nbeneficial for lifelong learning, while tuning the text encoder aids in\nzero-shot learning.\n", "date": "2024-05-24", "categories": []}
{"id": "2405.15157", "title": "Rethinking Class-Incremental Learning from a Dynamic Imbalanced Learning\n  Perspective", "abstract": "  Deep neural networks suffer from catastrophic forgetting when continually\nlearning new concepts. In this paper, we analyze this problem from a data\nimbalance point of view. We argue that the imbalance between old task and new\ntask data contributes to forgetting of the old tasks. Moreover, the increasing\nimbalance ratio during incremental learning further aggravates the problem. To\naddress the dynamic imbalance issue, we propose Uniform Prototype Contrastive\nLearning (UPCL), where uniform and compact features are learned. Specifically,\nwe generate a set of non-learnable uniform prototypes before each task starts.\nThen we assign these uniform prototypes to each class and guide the feature\nlearning through prototype contrastive learning. We also dynamically adjust the\nrelative margin between old and new classes so that the feature distribution\nwill be maintained balanced and compact. Finally, we demonstrate through\nextensive experiments that the proposed method achieves state-of-the-art\nperformance on several benchmark datasets including CIFAR100, ImageNet100 and\nTinyImageNet.\n", "date": "2024-05-24", "categories": []}
{"id": "2405.15633", "title": "Less is more: Summarizing Patch Tokens for efficient Multi-Label\n  Class-Incremental Learning", "abstract": "  Prompt tuning has emerged as an effective rehearsal-free technique for\nclass-incremental learning (CIL) that learns a tiny set of task-specific\nparameters (or prompts) to instruct a pre-trained transformer to learn on a\nsequence of tasks. Albeit effective, prompt tuning methods do not lend well in\nthe multi-label class incremental learning (MLCIL) scenario (where an image\ncontains multiple foreground classes) due to the ambiguity in selecting the\ncorrect prompt(s) corresponding to different foreground objects belonging to\nmultiple tasks. To circumvent this issue we propose to eliminate the prompt\nselection mechanism by maintaining task-specific pathways, which allow us to\nlearn representations that do not interact with the ones from the other tasks.\nSince independent pathways in truly incremental scenarios will result in an\nexplosion of computation due to the quadratically complex multi-head\nself-attention (MSA) operation in prompt tuning, we propose to reduce the\noriginal patch token embeddings into summarized tokens. Prompt tuning is then\napplied to these fewer summarized tokens to compute the final representation.\nOur proposed method Multi-Label class incremental learning via summarising\npAtch tokeN Embeddings (MULTI-LANE) enables learning disentangled task-specific\nrepresentations in MLCIL while ensuring fast inference. We conduct experiments\nin common benchmarks and demonstrate that our MULTI-LANE achieves a new\nstate-of-the-art in MLCIL. Additionally, we show that MULTI-LANE is also\ncompetitive in the CIL setting. Source code available at\nhttps://github.com/tdemin16/multi-lane\n", "date": "2024-05-24", "categories": []}
{"id": "2402.01262", "title": "Class incremental learning with probability dampening and cascaded gated\n  classifier", "abstract": "  Humans are capable of acquiring new knowledge and transferring learned\nknowledge into different domains, incurring a small forgetting. The same\nability, called Continual Learning, is challenging to achieve when operating\nwith neural networks due to the forgetting affecting past learned tasks when\nlearning new ones. This forgetting can be mitigated by replaying stored samples\nfrom past tasks, but a large memory size may be needed for long sequences of\ntasks; moreover, this could lead to overfitting on saved samples. In this\npaper, we propose a novel regularisation approach and a novel incremental\nclassifier called, respectively, Margin Dampening and Cascaded Scaling\nClassifier. The first combines a soft constraint and a knowledge distillation\napproach to preserve past learned knowledge while allowing the model to learn\nnew patterns effectively. The latter is a gated incremental classifier, helping\nthe model modify past predictions without directly interfering with them. This\nis achieved by modifying the output of the model with auxiliary scaling\nfunctions. We empirically show that our approach performs well on multiple\nbenchmarks against well-established baselines, and we also study each component\nof our proposal and how the combinations of such components affect the final\nresults.\n", "date": "2024-02-02", "categories": []}
{"id": "2405.13632", "title": "Task agnostic continual learning with Pairwise layer architecture", "abstract": "  Most of the dominant approaches to continual learning are based on either\nmemory replay, parameter isolation, or regularization techniques that require\ntask boundaries to calculate task statistics. We propose a static\narchitecture-based method that doesn't use any of these. We show that we can\nimprove the continual learning performance by replacing the final layer of our\nnetworks with our pairwise interaction layer. The pairwise interaction layer\nuses sparse representations from a Winner-take-all style activation function to\nfind the relevant correlations in the hidden layer representations. The\nnetworks using this architecture show competitive performance in MNIST and\nFashionMNIST-based continual image classification experiments. We demonstrate\nthis in an online streaming continual learning setup where the learning system\ncannot access task labels or boundaries.\n", "date": "2024-05-22", "categories": []}
{"id": "2405.13978", "title": "Mitigating Interference in the Knowledge Continuum through\n  Attention-Guided Incremental Learning", "abstract": "  Continual learning (CL) remains a significant challenge for deep neural\nnetworks, as it is prone to forgetting previously acquired knowledge. Several\napproaches have been proposed in the literature, such as experience rehearsal,\nregularization, and parameter isolation, to address this problem. Although\nalmost zero forgetting can be achieved in task-incremental learning,\nclass-incremental learning remains highly challenging due to the problem of\ninter-task class separation. Limited access to previous task data makes it\ndifficult to discriminate between classes of current and previous tasks. To\naddress this issue, we propose `Attention-Guided Incremental Learning' (AGILE),\na novel rehearsal-based CL approach that incorporates compact task attention to\neffectively reduce interference between tasks. AGILE utilizes lightweight,\nlearnable task projection vectors to transform the latent representations of a\nshared task attention module toward task distribution. Through extensive\nempirical evaluation, we show that AGILE significantly improves generalization\nperformance by mitigating task interference and outperforming rehearsal-based\napproaches in several CL scenarios. Furthermore, AGILE can scale well to a\nlarge number of tasks with minimal overhead while remaining well-calibrated\nwith reduced task-recency bias.\n", "date": "2024-05-22", "categories": []}
{"id": "2405.14291", "title": "Variational Bayes for Federated Continual Learning", "abstract": "  Federated continual learning (FCL) has received increasing attention due to\nits potential in handling real-world streaming data, characterized by evolving\ndata distributions and varying client classes over time. The constraints of\nstorage limitations and privacy concerns confine local models to exclusively\naccess the present data within each learning cycle. Consequently, this\nrestriction induces performance degradation in model training on previous data,\ntermed \"catastrophic forgetting\". However, existing FCL approaches need to\nidentify or know changes in data distribution, which is difficult in the real\nworld. To release these limitations, this paper directs attention to a broader\ncontinuous framework. Within this framework, we introduce Federated Bayesian\nNeural Network (FedBNN), a versatile and efficacious framework employing a\nvariational Bayesian neural network across all clients. Our method continually\nintegrates knowledge from local and historical data distributions into a single\nmodel, adeptly learning from new data distributions while retaining performance\non historical distributions. We rigorously evaluate FedBNN's performance\nagainst prevalent methods in federated learning and continual learning using\nvarious metrics. Experimental analyses across diverse datasets demonstrate that\nFedBNN achieves state-of-the-art results in mitigating forgetting.\n", "date": "2024-05-23", "categories": []}
{"id": "2405.10992", "title": "Overcoming Catastrophic Forgetting by Exemplar Selection in\n  Task-oriented Dialogue System", "abstract": "  Intelligent task-oriented dialogue systems (ToDs) are expected to\ncontinuously acquire new knowledge, also known as Continual Learning (CL),\nwhich is crucial to fit ever-changing user needs. However, catastrophic\nforgetting dramatically degrades the model performance in face of a long\nstreamed curriculum. In this paper, we aim to overcome the forgetting problem\nin ToDs and propose a method (HESIT) with hyper-gradient-based exemplar\nstrategy, which samples influential exemplars for periodic retraining. Instead\nof unilaterally observing data or models, HESIT adopts a profound exemplar\nselection strategy that considers the general performance of the trained model\nwhen selecting exemplars for each task domain. Specifically, HESIT analyzes the\ntraining data influence by tracing their hyper-gradient in the optimization\nprocess. Furthermore, HESIT avoids estimating Hessian to make it compatible for\nToDs with a large pre-trained model. Experimental results show that HESIT\neffectively alleviates catastrophic forgetting by exemplar selection, and\nachieves state-of-the-art performance on the largest CL benchmark of ToDs in\nterms of all metrics.\n", "date": "2024-05-16", "categories": []}
{"id": "2405.11822", "title": "FeTT: Continual Class Incremental Learning via Feature Transformation\n  Tuning", "abstract": "  Continual learning (CL) aims to extend deep models from static and enclosed\nenvironments to dynamic and complex scenarios, enabling systems to continuously\nacquire new knowledge of novel categories without forgetting previously learned\nknowledge. Recent CL models have gradually shifted towards the utilization of\npre-trained models (PTMs) with parameter-efficient fine-tuning (PEFT)\nstrategies. However, continual fine-tuning still presents a serious challenge\nof catastrophic forgetting due to the absence of previous task data.\nAdditionally, the fine-tune-then-frozen mechanism suffers from performance\nlimitations due to feature channels suppression and insufficient training data\nin the first CL task. To this end, this paper proposes feature transformation\ntuning (FeTT) model to non-parametrically fine-tune backbone features across\nall tasks, which not only operates independently of CL training data but also\nsmooths feature channels to prevent excessive suppression. Then, the extended\nensemble strategy incorporating different PTMs with FeTT model facilitates\nfurther performance improvement. We further elaborate on the discussions of the\nfine-tune-then-frozen paradigm and the FeTT model from the perspectives of\ndiscrepancy in class marginal distributions and feature channels. Extensive\nexperiments on CL benchmarks validate the effectiveness of our proposed method.\n", "date": "2024-05-20", "categories": []}
{"id": "2405.11829", "title": "Adversarially Diversified Rehearsal Memory (ADRM): Mitigating Memory\n  Overfitting Challenge in Continual Learning", "abstract": "  Continual learning focuses on learning non-stationary data distribution\nwithout forgetting previous knowledge. Rehearsal-based approaches are commonly\nused to combat catastrophic forgetting. However, these approaches suffer from a\nproblem called \"rehearsal memory overfitting, \" where the model becomes too\nspecialized on limited memory samples and loses its ability to generalize\neffectively. As a result, the effectiveness of the rehearsal memory\nprogressively decays, ultimately resulting in catastrophic forgetting of the\nlearned tasks.\n  We introduce the Adversarially Diversified Rehearsal Memory (ADRM) to address\nthe memory overfitting challenge. This novel method is designed to enrich\nmemory sample diversity and bolster resistance against natural and adversarial\nnoise disruptions. ADRM employs the FGSM attacks to introduce adversarially\nmodified memory samples, achieving two primary objectives: enhancing memory\ndiversity and fostering a robust response to continual feature drifts in memory\nsamples.\n  Our contributions are as follows: Firstly, ADRM addresses overfitting in\nrehearsal memory by employing FGSM to diversify and increase the complexity of\nthe memory buffer. Secondly, we demonstrate that ADRM mitigates memory\noverfitting and significantly improves the robustness of CL models, which is\ncrucial for safety-critical applications. Finally, our detailed analysis of\nfeatures and visualization demonstrates that ADRM mitigates feature drifts in\nCL memory samples, significantly reducing catastrophic forgetting and resulting\nin a more resilient CL model. Additionally, our in-depth t-SNE visualizations\nof feature distribution and the quantification of the feature similarity\nfurther enrich our understanding of feature representation in existing CL\napproaches. Our code is publically available at\nhttps://github.com/hikmatkhan/ADRM.\n", "date": "2024-05-20", "categories": []}
{"id": "2404.07518", "title": "Remembering Transformer for Continual Learning", "abstract": "  Neural networks encounter the challenge of Catastrophic Forgetting (CF) in\ncontinual learning, where new task learning interferes with previously learned\nknowledge. Existing data fine-tuning and regularization methods necessitate\ntask identity information during inference and cannot eliminate interference\namong different tasks, while soft parameter sharing approaches encounter the\nproblem of an increasing model parameter size. To tackle these challenges, we\npropose the Remembering Transformer, inspired by the brain's Complementary\nLearning Systems (CLS). Remembering Transformer employs a mixture-of-adapters\narchitecture and a generative model-based novelty detection mechanism in a\npretrained Transformer to alleviate CF. Remembering Transformer dynamically\nroutes task data to the most relevant adapter with enhanced parameter\nefficiency based on knowledge distillation. We conducted extensive experiments,\nincluding ablation studies on the novelty detection mechanism and model\ncapacity of the mixture-of-adapters, in a broad range of class-incremental\nsplit tasks and permutation tasks. Our approach demonstrated SOTA performance\nsurpassing the second-best method by 15.90% in the split tasks, reducing the\nmemory footprint from 11.18M to 0.22M in the five splits CIFAR10 task.\n", "date": "2024-04-11", "categories": []}
{"id": "2405.09820", "title": "Densely Distilling Cumulative Knowledge for Continual Learning", "abstract": "  Continual learning, involving sequential training on diverse tasks, often\nfaces catastrophic forgetting. While knowledge distillation-based approaches\nexhibit notable success in preventing forgetting, we pinpoint a limitation in\ntheir ability to distill the cumulative knowledge of all the previous tasks. To\nremedy this, we propose Dense Knowledge Distillation (DKD). DKD uses a task\npool to track the model's capabilities. It partitions the output logits of the\nmodel into dense groups, each corresponding to a task in the task pool. It then\ndistills all tasks' knowledge using all groups. However, using all the groups\ncan be computationally expensive, we also suggest random group selection in\neach optimization step. Moreover, we propose an adaptive weighting scheme,\nwhich balances the learning of new classes and the retention of old classes,\nbased on the count and similarity of the classes. Our DKD outperforms recent\nstate-of-the-art baselines across diverse benchmarks and scenarios. Empirical\nanalysis underscores DKD's ability to enhance model stability, promote flatter\nminima for improved generalization, and remains robust across various memory\nbudgets and task orders. Moreover, it seamlessly integrates with other CL\nmethods to boost performance and proves versatile in offline scenarios like\nmodel compression.\n", "date": "2024-05-16", "categories": []}
{"id": "2405.09133", "title": "Overcoming Domain Drift in Online Continual Learning", "abstract": "  Online Continual Learning (OCL) empowers machine learning models to acquire\nnew knowledge online across a sequence of tasks. However, OCL faces a\nsignificant challenge: catastrophic forgetting, wherein the model learned in\nprevious tasks is substantially overwritten upon encountering new tasks,\nleading to a biased forgetting of prior knowledge. Moreover, the continual\ndoman drift in sequential learning tasks may entail the gradual displacement of\nthe decision boundaries in the learned feature space, rendering the learned\nknowledge susceptible to forgetting. To address the above problem, in this\npaper, we propose a novel rehearsal strategy, termed Drift-Reducing Rehearsal\n(DRR), to anchor the domain of old tasks and reduce the negative transfer\neffects. First, we propose to select memory for more representative samples\nguided by constructed centroids in a data stream. Then, to keep the model from\ndomain chaos in drifting, a two-level angular cross-task Contrastive Margin\nLoss (CML) is proposed, to encourage the intra-class and intra-task\ncompactness, and increase the inter-class and inter-task discrepancy. Finally,\nto further suppress the continual domain drift, we present an optional Centorid\nDistillation Loss (CDL) on the rehearsal memory to anchor the knowledge in\nfeature space for each previous old task. Extensive experimental results on\nfour benchmark datasets validate that the proposed DRR can effectively mitigate\nthe continual domain drift and achieve the state-of-the-art (SOTA) performance\nin OCL.\n", "date": "2024-05-15", "categories": []}
{"id": "2405.09492", "title": "MGSER-SAM: Memory-Guided Soft Experience Replay with Sharpness-Aware\n  Optimization for Enhanced Continual Learning", "abstract": "  Deep neural networks suffer from the catastrophic forgetting problem in the\nfield of continual learning (CL). To address this challenge, we propose\nMGSER-SAM, a novel memory replay-based algorithm specifically engineered to\nenhance the generalization capabilities of CL models. We first intergrate the\nSAM optimizer, a component designed for optimizing flatness, which seamlessly\nfits into well-known Experience Replay frameworks such as ER and DER++. Then,\nMGSER-SAM distinctively addresses the complex challenge of reconciling\nconflicts in weight perturbation directions between ongoing tasks and\npreviously stored memories, which is underexplored in the SAM optimizer. This\nis effectively accomplished by the strategic integration of soft logits and the\nalignment of memory gradient directions, where the regularization terms\nfacilitate the concurrent minimization of various training loss terms integral\nto the CL process. Through rigorous experimental analysis conducted across\nmultiple benchmarks, MGSER-SAM has demonstrated a consistent ability to\noutperform existing baselines in all three CL scenarios. Comparing to the\nrepresentative memory replay-based baselines ER and DER++, MGSER-SAM not only\nimproves the testing accuracy by $24.4\\%$ and $17.6\\%$ respectively, but also\nachieves the lowest forgetting on each benchmark.\n", "date": "2024-05-15", "categories": []}
{"id": "2405.08015", "title": "A Methodology-Oriented Study of Catastrophic Forgetting in Incremental\n  Deep Neural Networks", "abstract": "  Human being and different species of animals having the skills to gather,\ntransferring knowledge, processing, fine-tune and generating information\nthroughout their lifetime. The ability of learning throughout their lifespan is\nreferred as continuous learning which is using neurocognition mechanism.\nConsequently, in real world computational system of incremental learning\nautonomous agents also needs such continuous learning mechanism which provide\nretrieval of information and long-term memory consolidation. However, the main\nchallenge in artificial intelligence is that the incremental learning of the\nautonomous agent when new data confronted. In such scenarios, the main concern\nis catastrophic forgetting(CF), i.e., while learning the sequentially, neural\nnetwork underfits the old data when it confronted with new data. To tackle this\nCF problem many numerous studied have been proposed, however it is very\ndifficult to compare their performance due to dissimilarity in their evaluation\nmechanism. Here we focus on the comparison of all algorithms which are having\nsimilar type of evaluation mechanism. Here we are comparing three types of\nincremental learning methods: (1) Exemplar based methods, (2) Memory based\nmethods, and (3) Network based method. In this survey paper, methodology\noriented study for catastrophic forgetting in incremental deep neural network\nis addressed. Furthermore, it contains the mathematical overview of impact-full\nmethods which can be help researchers to deal with CF.\n", "date": "2024-05-11", "categories": []}
{"id": "2405.08038", "title": "Feature Expansion and enhanced Compression for Class Incremental\n  Learning", "abstract": "  Class incremental learning consists in training discriminative models to\nclassify an increasing number of classes over time. However, doing so using\nonly the newly added class data leads to the known problem of catastrophic\nforgetting of the previous classes. Recently, dynamic deep learning\narchitectures have been shown to exhibit a better stability-plasticity\ntrade-off by dynamically adding new feature extractors to the model in order to\nlearn new classes followed by a compression step to scale the model back to its\noriginal size, thus avoiding a growing number of parameters. In this context,\nwe propose a new algorithm that enhances the compression of previous class\nknowledge by cutting and mixing patches of previous class samples with the new\nimages during compression using our Rehearsal-CutMix method. We show that this\nnew data augmentation reduces catastrophic forgetting by specifically targeting\npast class information and improving its compression. Extensive experiments\nperformed on the CIFAR and ImageNet datasets under diverse incremental learning\nevaluation protocols demonstrate that our approach consistently outperforms the\nstate-of-the-art . The code will be made available upon publication of our\nwork.\n", "date": "2024-05-13", "categories": []}
{"id": "2405.08533", "title": "Dynamic Feature Learning and Matching for Class-Incremental Learning", "abstract": "  Class-incremental learning (CIL) has emerged as a means to learn new classes\nincrementally without catastrophic forgetting of previous classes. Recently,\nCIL has undergone a paradigm shift towards dynamic architectures due to their\nsuperior performance. However, these models are still limited by the following\naspects: (i) Data augmentation (DA), which are tightly coupled with CIL,\nremains under-explored in dynamic architecture scenarios. (ii) Feature\nrepresentation. The discriminativeness of dynamic feature are sub-optimal and\npossess potential for refinement. (iii) Classifier. The misalignment between\ndynamic feature and classifier constrains the capabilities of the model. To\ntackle the aforementioned drawbacks, we propose the Dynamic Feature Learning\nand Matching (DFLM) model in this paper from above three perspectives.\nSpecifically, we firstly introduce class weight information and non-stationary\nfunctions to extend the mix DA method for dynamically adjusting the focus on\nmemory during training. Then, von Mises-Fisher (vMF) classifier is employed to\neffectively model the dynamic feature distribution and implicitly learn their\ndiscriminative properties. Finally, the matching loss is proposed to facilitate\nthe alignment between the learned dynamic features and the classifier by\nminimizing the distribution distance. Extensive experiments on CIL benchmarks\nvalidate that our proposed model achieves significant performance improvements\nover existing methods.\n", "date": "2024-05-14", "categories": []}
{"id": "2308.03810", "title": "AdaER: An Adaptive Experience Replay Approach for Continual Lifelong\n  Learning", "abstract": "  Continual lifelong learning is an machine learning framework inspired by\nhuman learning, where learners are trained to continuously acquire new\nknowledge in a sequential manner. However, the non-stationary nature of\nstreaming training data poses a significant challenge known as catastrophic\nforgetting, which refers to the rapid forgetting of previously learned\nknowledge when new tasks are introduced. While some approaches, such as\nexperience replay (ER), have been proposed to mitigate this issue, their\nperformance remains limited, particularly in the class-incremental scenario\nwhich is considered natural and highly challenging. In this paper, we present a\nnovel algorithm, called adaptive-experience replay (AdaER), to address the\nchallenge of continual lifelong learning. AdaER consists of two stages: memory\nreplay and memory update. In the memory replay stage, AdaER introduces a\ncontextually-cued memory recall (C-CMR) strategy, which selectively replays\nmemories that are most conflicting with the current input data in terms of both\ndata and task. Additionally, AdaER incorporates an entropy-balanced reservoir\nsampling (E-BRS) strategy to enhance the performance of the memory buffer by\nmaximizing information entropy. To evaluate the effectiveness of AdaER, we\nconduct experiments on established supervised continual lifelong learning\nbenchmarks, specifically focusing on class-incremental learning scenarios. The\nresults demonstrate that AdaER outperforms existing continual lifelong learning\nbaselines, highlighting its efficacy in mitigating catastrophic forgetting and\nimproving learning performance.\n", "date": "2023-08-07", "categories": []}
{"id": "2405.07142", "title": "Cross-Domain Continual Learning via CLAMP", "abstract": "  Artificial neural networks, celebrated for their human-like cognitive\nlearning abilities, often encounter the well-known catastrophic forgetting (CF)\nproblem, where the neural networks lose the proficiency in previously acquired\nknowledge. Despite numerous efforts to mitigate CF, it remains the significant\nchallenge particularly in complex changing environments. This challenge is even\nmore pronounced in cross-domain adaptation following the continual learning\n(CL) setting, which is a more challenging and realistic scenario that is\nunder-explored. To this end, this article proposes a cross-domain CL approach\nmaking possible to deploy a single model in such environments without\nadditional labelling costs. Our approach, namely continual learning approach\nfor many processes (CLAMP), integrates a class-aware adversarial domain\nadaptation strategy to align a source domain and a target domain. An\nassessor-guided learning process is put forward to navigate the learning\nprocess of a base model assigning a set of weights to every sample controlling\nthe influence of every sample and the interactions of each loss function in\nsuch a way to balance the stability and plasticity dilemma thus preventing the\nCF problem. The first assessor focuses on the negative transfer problem\nrejecting irrelevant samples of the source domain while the second assessor\nprevents noisy pseudo labels of the target domain. Both assessors are trained\nin the meta-learning approach using random transformation techniques and\nsimilar samples of the source domain. Theoretical analysis and extensive\nnumerical validations demonstrate that CLAMP significantly outperforms\nestablished baseline algorithms across all experiments by at least $10\\%$\nmargin.\n", "date": "2024-05-12", "categories": []}
{"id": "2404.10758", "title": "Watch Your Step: Optimal Retrieval for Continual Learning at Scale", "abstract": "  In continual learning, a model learns incrementally over time while\nminimizing interference between old and new tasks. One of the most widely used\napproaches in continual learning is referred to as replay. Replay methods\nsupport interleaved learning by storing past experiences in a replay buffer.\nAlthough there are methods for selectively constructing the buffer and\nreprocessing its contents, there is limited exploration of the problem of\nselectively retrieving samples from the buffer. Current solutions have been\ntested in limited settings and, more importantly, in isolation. Existing work\nhas also not explored the impact of duplicate replays on performance. In this\nwork, we propose a framework for evaluating selective retrieval strategies,\ncategorized by simple, independent class- and sample-selective primitives. We\nevaluated several combinations of existing strategies for selective retrieval\nand present their performances. Furthermore, we propose a set of strategies to\nprevent duplicate replays and explore whether new samples with low loss values\ncan be learned without replay. In an effort to match our problem setting to a\nrealistic continual learning pipeline, we restrict our experiments to a setting\ninvolving a large, pre-trained, open vocabulary object detection model, which\nis fully fine-tuned on a sequence of 15 datasets.\n", "date": "2024-04-16", "categories": []}
{"id": "2405.05984", "title": "Few-Shot Class Incremental Learning via Robust Transformer Approach", "abstract": "  Few-Shot Class-Incremental Learning presents an extension of the Class\nIncremental Learning problem where a model is faced with the problem of data\nscarcity while addressing the catastrophic forgetting problem. This problem\nremains an open problem because all recent works are built upon the\nconvolutional neural networks performing sub-optimally compared to the\ntransformer approaches. Our paper presents Robust Transformer Approach built\nupon the Compact Convolution Transformer. The issue of overfitting due to few\nsamples is overcome with the notion of the stochastic classifier, where the\nclassifier's weights are sampled from a distribution with mean and variance\nvectors, thus increasing the likelihood of correct classifications, and the\nbatch-norm layer to stabilize the training process. The issue of CF is dealt\nwith the idea of delta parameters, small task-specific trainable parameters\nwhile keeping the backbone networks frozen. A non-parametric approach is\ndeveloped to infer the delta parameters for the model's predictions. The\nprototype rectification approach is applied to avoid biased prototype\ncalculations due to the issue of data scarcity. The advantage of ROBUSTA is\ndemonstrated through a series of experiments in the benchmark problems where it\nis capable of outperforming prior arts with big margins without any data\naugmentation protocols.\n", "date": "2024-05-08", "categories": []}
{"id": "2405.06296", "title": "Fast Evaluation of DNN for Past Dataset in Incremental Learning", "abstract": "  During the operation of a system including a deep neural network (DNN), new\ninput values that were not included in the training dataset are given to the\nDNN. In such a case, the DNN may be incrementally trained with the new input\nvalues; however, that training may reduce the accuracy of the DNN in regard to\nthe dataset that was previously obtained and used for the past training. It is\nnecessary to evaluate the effect of the additional training on the accuracy for\nthe past dataset. However, evaluation by testing all the input values included\nin the past dataset takes time. Therefore, we propose a new method to quickly\nevaluate the effect on the accuracy for the past dataset. In the proposed\nmethod, the gradient of the parameter values (such as weight and bias) for the\npast dataset is extracted by running the DNN before the training. Then, after\nthe training, its effect on the accuracy with respect to the past dataset is\ncalculated from the gradient and update differences of the parameter values. To\nshow the usefulness of the proposed method, we present experimental results\nwith several datasets. The results show that the proposed method can estimate\nthe accuracy change by additional training in a constant time.\n", "date": "2024-05-10", "categories": []}
{"id": "2405.06389", "title": "Continual Novel Class Discovery via Feature Enhancement and Adaptation", "abstract": "  Continual Novel Class Discovery (CNCD) aims to continually discover novel\nclasses without labels while maintaining the recognition capability for\npreviously learned classes. The main challenges faced by CNCD include the\nfeature-discrepancy problem, the inter-session confusion problem, etc. In this\npaper, we propose a novel Feature Enhancement and Adaptation method for the\nCNCD to tackle the above challenges, which consists of a guide-to-novel\nframework, a centroid-to-samples similarity constraint (CSS), and a\nboundary-aware prototype constraint (BAP). More specifically, the\nguide-to-novel framework is established to continually discover novel classes\nunder the guidance of prior distribution. Afterward, the CSS is designed to\nconstrain the relationship between centroid-to-samples similarities of\ndifferent classes, thereby enhancing the distinctiveness of features among\nnovel classes. Finally, the BAP is proposed to keep novel class features aware\nof the positions of other class prototypes during incremental sessions, and\nbetter adapt novel class features to the shared feature space. Experimental\nresults on three benchmark datasets demonstrate the superiority of our method,\nespecially in more challenging protocols with more incremental sessions.\n", "date": "2024-05-10", "categories": []}
{"id": "2403.05346", "title": "VLM-PL: Advanced Pseudo Labeling Approach for Class Incremental Object\n  Detection via Vision-Language Model", "abstract": "  In the field of Class Incremental Object Detection (CIOD), creating models\nthat can continuously learn like humans is a major challenge. Pseudo-labeling\nmethods, although initially powerful, struggle with multi-scenario incremental\nlearning due to their tendency to forget past knowledge. To overcome this, we\nintroduce a new approach called Vision-Language Model assisted Pseudo-Labeling\n(VLM-PL). This technique uses Vision-Language Model (VLM) to verify the\ncorrectness of pseudo ground-truths (GTs) without requiring additional model\ntraining. VLM-PL starts by deriving pseudo GTs from a pre-trained detector.\nThen, we generate custom queries for each pseudo GT using carefully designed\nprompt templates that combine image and text features. This allows the VLM to\nclassify the correctness through its responses. Furthermore, VLM-PL integrates\nrefined pseudo and real GTs from upcoming training, effectively combining new\nand old knowledge. Extensive experiments conducted on the Pascal VOC and MS\nCOCO datasets not only highlight VLM-PL's exceptional performance in\nmulti-scenario but also illuminate its effectiveness in dual-scenario by\nachieving state-of-the-art results in both.\n", "date": "2024-03-08", "categories": []}
{"id": "2405.05496", "title": "Boosting Large Language Models with Continual Learning for Aspect-based\n  Sentiment Analysis", "abstract": "  Aspect-based sentiment analysis (ABSA) is an important subtask of sentiment\nanalysis, which aims to extract the aspects and predict their sentiments. Most\nexisting studies focus on improving the performance of the target domain by\nfine-tuning domain-specific models (trained on source domains) based on the\ntarget domain dataset. Few works propose continual learning tasks for ABSA,\nwhich aim to learn the target domain's ability while maintaining the history\ndomains' abilities. In this paper, we propose a Large Language Model-based\nContinual Learning (\\texttt{LLM-CL}) model for ABSA. First, we design a domain\nknowledge decoupling module to learn a domain-invariant adapter and separate\ndomain-variant adapters dependently with an orthogonal constraint. Then, we\nintroduce a domain knowledge warmup strategy to align the representation\nbetween domain-invariant and domain-variant knowledge. In the test phase, we\nindex the corresponding domain-variant knowledge via domain positioning to not\nrequire each sample's domain ID. Extensive experiments over 19 datasets\nindicate that our \\texttt{LLM-CL} model obtains new state-of-the-art\nperformance.\n", "date": "2024-05-09", "categories": []}
{"id": "2305.19076", "title": "Approximate Bayesian Class-Conditional Models under Continuous\n  Representation Shift", "abstract": "  For models consisting of a classifier in some representation space, learning\nonline from a non-stationary data stream often necessitates changes in the\nrepresentation. So, the question arises of what is the best way to adapt the\nclassifier to shifts in representation. Current methods only slowly change the\nclassifier to representation shift, introducing noise into learning as the\nclassifier is misaligned to the representation. We propose DeepCCG, an\nempirical Bayesian approach to solve this problem. DeepCCG works by updating\nthe posterior of a class conditional Gaussian classifier such that the\nclassifier adapts in one step to representation shift. The use of a class\nconditional Gaussian classifier also enables DeepCCG to use a log conditional\nmarginal likelihood loss to update the representation. To perform the update to\nthe classifier and representation, DeepCCG maintains a fixed number of examples\nin memory and so a key part of DeepCCG is selecting what examples to store,\nchoosing the subset that minimises the KL divergence between the true posterior\nand the posterior induced by the subset. We explore the behaviour of DeepCCG in\nonline continual learning (CL), demonstrating that it performs well against a\nspectrum of online CL methods and that it reduces the change in performance due\nto representation shift.\n", "date": "2023-05-30", "categories": []}
{"id": "2402.17323", "title": "SDDGR: Stable Diffusion-based Deep Generative Replay for Class\n  Incremental Object Detection", "abstract": "  In the field of class incremental learning (CIL), generative replay has\nbecome increasingly prominent as a method to mitigate the catastrophic\nforgetting, alongside the continuous improvements in generative models.\nHowever, its application in class incremental object detection (CIOD) has been\nsignificantly limited, primarily due to the complexities of scenes involving\nmultiple labels. In this paper, we propose a novel approach called stable\ndiffusion deep generative replay (SDDGR) for CIOD. Our method utilizes a\ndiffusion-based generative model with pre-trained text-to-diffusion networks to\ngenerate realistic and diverse synthetic images. SDDGR incorporates an\niterative refinement strategy to produce high-quality images encompassing old\nclasses. Additionally, we adopt an L2 knowledge distillation technique to\nimprove the retention of prior knowledge in synthetic images. Furthermore, our\napproach includes pseudo-labeling for old objects within new task images,\npreventing misclassification as background elements. Extensive experiments on\nthe COCO 2017 dataset demonstrate that SDDGR significantly outperforms existing\nalgorithms, achieving a new state-of-the-art in various CIOD scenarios. The\nsource code will be made available to the public.\n", "date": "2024-02-27", "categories": []}
{"id": "2405.02685", "title": "FedProK: Trustworthy Federated Class-Incremental Learning via\n  Prototypical Feature Knowledge Transfer", "abstract": "  Federated Class-Incremental Learning (FCIL) focuses on continually\ntransferring the previous knowledge to learn new classes in dynamic Federated\nLearning (FL). However, existing methods do not consider the trustworthiness of\nFCIL, i.e., improving continual utility, privacy, and efficiency\nsimultaneously, which is greatly influenced by catastrophic forgetting and data\nheterogeneity among clients. To address this issue, we propose FedProK\n(Federated Prototypical Feature Knowledge Transfer), leveraging prototypical\nfeature as a novel representation of knowledge to perform spatial-temporal\nknowledge transfer. Specifically, FedProK consists of two components: (1)\nfeature translation procedure on the client side by temporal knowledge transfer\nfrom the learned classes and (2) prototypical knowledge fusion on the server\nside by spatial knowledge transfer among clients. Extensive experiments\nconducted in both synchronous and asynchronous settings demonstrate that our\nFedProK outperforms the other state-of-the-art methods in three perspectives of\ntrustworthiness, validating its effectiveness in selectively transferring\nspatial-temporal knowledge.\n", "date": "2024-05-04", "categories": []}
{"id": "2405.02766", "title": "Beyond Unimodal Learning: The Importance of Integrating Multiple\n  Modalities for Lifelong Learning", "abstract": "  While humans excel at continual learning (CL), deep neural networks (DNNs)\nexhibit catastrophic forgetting. A salient feature of the brain that allows\neffective CL is that it utilizes multiple modalities for learning and\ninference, which is underexplored in DNNs. Therefore, we study the role and\ninteractions of multiple modalities in mitigating forgetting and introduce a\nbenchmark for multimodal continual learning. Our findings demonstrate that\nleveraging multiple views and complementary information from multiple\nmodalities enables the model to learn more accurate and robust representations.\nThis makes the model less vulnerable to modality-specific regularities and\nconsiderably mitigates forgetting. Furthermore, we observe that individual\nmodalities exhibit varying degrees of robustness to distribution shift.\nFinally, we propose a method for integrating and aligning the information from\ndifferent modalities by utilizing the relational structural similarities\nbetween the data points in each modality. Our method sets a strong baseline\nthat enables both single- and multimodal inference. Our study provides a\npromising case for further exploring the role of multiple modalities in\nenabling CL and provides a standard benchmark for future research.\n", "date": "2024-05-04", "categories": []}
{"id": "2311.18763", "title": "Continual Diffusion with STAMINA: STack-And-Mask INcremental Adapters", "abstract": "  Recent work has demonstrated a remarkable ability to customize text-to-image\ndiffusion models to multiple, fine-grained concepts in a sequential (i.e.,\ncontinual) manner while only providing a few example images for each concept.\nThis setting is known as continual diffusion. Here, we ask the question: Can we\nscale these methods to longer concept sequences without forgetting? Although\nprior work mitigates the forgetting of previously learned concepts, we show\nthat its capacity to learn new tasks reaches saturation over longer sequences.\nWe address this challenge by introducing a novel method, STack-And-Mask\nINcremental Adapters (STAMINA), which is composed of low-ranked\nattention-masked adapters and customized MLP tokens. STAMINA is designed to\nenhance the robust fine-tuning properties of LoRA for sequential concept\nlearning via learnable hard-attention masks parameterized with low rank MLPs,\nenabling precise, scalable learning via sparse adaptation. Notably, all\nintroduced trainable parameters can be folded back into the model after\ntraining, inducing no additional inference parameter costs. We show that\nSTAMINA outperforms the prior SOTA for the setting of text-to-image continual\ncustomization on a 50-concept benchmark composed of landmarks and human faces,\nwith no stored replay data. Additionally, we extended our method to the setting\nof continual learning for image classification, demonstrating that our gains\nalso translate to state-of-the-art performance in this standard benchmark.\n", "date": "2023-11-30", "categories": []}
{"id": "2210.01422", "title": "Time-Varying Propensity Score to Bridge the Gap between the Past and\n  Present", "abstract": "  Real-world deployment of machine learning models is challenging because data\nevolves over time. While no model can work when data evolves in an arbitrary\nfashion, if there is some pattern to these changes, we might be able to design\nmethods to address it. This paper addresses situations when data evolves\ngradually. We introduce a time-varying propensity score that can detect gradual\nshifts in the distribution of data which allows us to selectively sample past\ndata to update the model -- not just similar data from the past like that of a\nstandard propensity score but also data that evolved in a similar fashion in\nthe past. The time-varying propensity score is quite general: we demonstrate\ndifferent ways of implementing it and evaluate it on a variety of problems\nranging from supervised learning (e.g., image classification problems) where\ndata undergoes a sequence of gradual shifts, to reinforcement learning tasks\n(e.g., robotic manipulation and continuous control) where data shifts as the\npolicy or the task changes.\n", "date": "2022-10-04", "categories": []}
{"id": "2304.06027", "title": "Continual Diffusion: Continual Customization of Text-to-Image Diffusion\n  with C-LoRA", "abstract": "  Recent works demonstrate a remarkable ability to customize text-to-image\ndiffusion models while only providing a few example images. What happens if you\ntry to customize such models using multiple, fine-grained concepts in a\nsequential (i.e., continual) manner? In our work, we show that recent\nstate-of-the-art customization of text-to-image models suffer from catastrophic\nforgetting when new concepts arrive sequentially. Specifically, when adding a\nnew concept, the ability to generate high quality images of past, similar\nconcepts degrade. To circumvent this forgetting, we propose a new method,\nC-LoRA, composed of a continually self-regularized low-rank adaptation in cross\nattention layers of the popular Stable Diffusion model. Furthermore, we use\ncustomization prompts which do not include the word of the customized object\n(i.e., \"person\" for a human face dataset) and are initialized as completely\nrandom embeddings. Importantly, our method induces only marginal additional\nparameter costs and requires no storage of user data for replay. We show that\nC-LoRA not only outperforms several baselines for our proposed setting of\ntext-to-image continual customization, which we refer to as Continual\nDiffusion, but that we achieve a new state-of-the-art in the well-established\nrehearsal-free continual learning setting for image classification. The high\nachieving performance of C-LoRA in two separate domains positions it as a\ncompelling solution for a wide range of applications, and we believe it has\nsignificant potential for practical impact. Project page:\nhttps://jamessealesmith.github.io/continual-diffusion/\n", "date": "2023-04-12", "categories": []}
{"id": "2306.06545", "title": "A Probabilistic Framework for Modular Continual Learning", "abstract": "  Modular approaches that use a different composition of modules for each\nproblem are a promising direction in continual learning (CL). However,\nsearching through the large, discrete space of module compositions is\nchallenging, especially because evaluating a composition's performance requires\na round of neural network training. We address this challenge through a modular\nCL framework, PICLE, that uses a probabilistic model to cheaply compute the\nfitness of each composition, allowing PICLE to achieve both perceptual,\nfew-shot and latent transfer. The model combines prior knowledge about good\nmodule compositions with dataset-specific information. We evaluate PICLE using\ntwo benchmark suites designed to assess different desiderata of CL techniques.\nComparing to a wide range of approaches, we show that PICLE is the first\nmodular CL algorithm to achieve perceptual, few-shot and latent transfer while\nscaling well to large search spaces, outperforming previous state-of-the-art\nmodular CL approaches on long problem sequences.\n", "date": "2023-06-11", "categories": []}
{"id": "2308.13646", "title": "GRASP: A Rehearsal Policy for Efficient Online Continual Learning", "abstract": "  Continual learning (CL) in deep neural networks (DNNs) involves incrementally\naccumulating knowledge in a DNN from a growing data stream. A major challenge\nin CL is that non-stationary data streams cause catastrophic forgetting of\npreviously learned abilities. A popular solution is rehearsal: storing past\nobservations in a buffer and then sampling the buffer to update the DNN.\nUniform sampling in a class-balanced manner is highly effective, and better\nsample selection policies have been elusive. Here, we propose a new sample\nselection policy called GRASP that selects the most prototypical (easy) samples\nfirst and then gradually selects less prototypical (harder) examples. GRASP has\nlittle additional compute or memory overhead compared to uniform selection,\nenabling it to scale to large datasets. Compared to 17 other rehearsal\npolicies, GRASP achieves higher accuracy in CL experiments on ImageNet.\nCompared to uniform balanced sampling, GRASP achieves the same performance with\n40% fewer updates. We also show that GRASP is effective for CL on five text\nclassification datasets.\n", "date": "2023-08-25", "categories": []}
{"id": "2404.00781", "title": "Addressing Loss of Plasticity and Catastrophic Forgetting in Continual\n  Learning", "abstract": "  Deep representation learning methods struggle with continual learning,\nsuffering from both catastrophic forgetting of useful units and loss of\nplasticity, often due to rigid and unuseful units. While many methods address\nthese two issues separately, only a few currently deal with both\nsimultaneously. In this paper, we introduce Utility-based Perturbed Gradient\nDescent (UPGD) as a novel approach for the continual learning of\nrepresentations. UPGD combines gradient updates with perturbations, where it\napplies smaller modifications to more useful units, protecting them from\nforgetting, and larger modifications to less useful units, rejuvenating their\nplasticity. We use a challenging streaming learning setup where continual\nlearning problems have hundreds of non-stationarities and unknown task\nboundaries. We show that many existing methods suffer from at least one of the\nissues, predominantly manifested by their decreasing accuracy over tasks. On\nthe other hand, UPGD continues to improve performance and surpasses or is\ncompetitive with all methods in all problems. Finally, in extended\nreinforcement learning experiments with PPO, we show that while Adam exhibits a\nperformance drop after initial learning, UPGD avoids it by addressing both\ncontinual learning issues.\n", "date": "2024-03-31", "categories": []}
{"id": "2404.19462", "title": "Continual Model-based Reinforcement Learning for Data Efficient Wireless\n  Network Optimisation", "abstract": "  We present a method that addresses the pain point of long lead-time required\nto deploy cell-level parameter optimisation policies to new wireless network\nsites. Given a sequence of action spaces represented by overlapping subsets of\ncell-level configuration parameters provided by domain experts, we formulate\nthroughput optimisation as Continual Reinforcement Learning of control\npolicies. Simulation results suggest that the proposed system is able to\nshorten the end-to-end deployment lead-time by two-fold compared to a\nreinitialise-and-retrain baseline without any drop in optimisation gain.\n", "date": "2024-04-30", "categories": []}
{"id": "2303.09347", "title": "CSSL-MHTR: Continual Self-Supervised Learning for Scalable Multi-script\n  Handwritten Text Recognition", "abstract": "  Self-supervised learning has recently emerged as a strong alternative in\ndocument analysis. These approaches are now capable of learning high-quality\nimage representations and overcoming the limitations of supervised methods,\nwhich require a large amount of labeled data. However, these methods are unable\nto capture new knowledge in an incremental fashion, where data is presented to\nthe model sequentially, which is closer to the realistic scenario. In this\npaper, we explore the potential of continual self-supervised learning to\nalleviate the catastrophic forgetting problem in handwritten text recognition,\nas an example of sequence recognition. Our method consists in adding\nintermediate layers called adapters for each task, and efficiently distilling\nknowledge from the previous model while learning the current task. Our proposed\nframework is efficient in both computation and memory complexity. To\ndemonstrate its effectiveness, we evaluate our method by transferring the\nlearned model to diverse text recognition downstream tasks, including Latin and\nnon-Latin scripts. As far as we know, this is the first application of\ncontinual self-supervised learning for handwritten text recognition. We attain\nstate-of-the-art performance on English, Italian and Russian scripts, whilst\nadding only a few parameters per task. The code and trained models will be\npublicly available.\n", "date": "2023-03-16", "categories": []}
{"id": "2404.17651", "title": "Hard ASH: Sparsity and the right optimizer make a continual learner", "abstract": "  In class incremental learning, neural networks typically suffer from\ncatastrophic forgetting. We show that an MLP featuring a sparse activation\nfunction and an adaptive learning rate optimizer can compete with established\nregularization techniques in the Split-MNIST task. We highlight the\neffectiveness of the Adaptive SwisH (ASH) activation function in this context\nand introduce a novel variant, Hard Adaptive SwisH (Hard ASH) to further\nenhance the learning retention.\n", "date": "2024-04-26", "categories": []}
{"id": "2404.18060", "title": "Prompt Customization for Continual Learning", "abstract": "  Contemporary continual learning approaches typically select prompts from a\npool, which function as supplementary inputs to a pre-trained model. However,\nthis strategy is hindered by the inherent noise of its selection approach when\nhandling increasing tasks. In response to these challenges, we reformulate the\nprompting approach for continual learning and propose the prompt customization\n(PC) method. PC mainly comprises a prompt generation module (PGM) and a prompt\nmodulation module (PMM). In contrast to conventional methods that employ hard\nprompt selection, PGM assigns different coefficients to prompts from a\nfixed-sized pool of prompts and generates tailored prompts. Moreover, PMM\nfurther modulates the prompts by adaptively assigning weights according to the\ncorrelations between input data and corresponding prompts. We evaluate our\nmethod on four benchmark datasets for three diverse settings, including the\nclass, domain, and task-agnostic incremental learning tasks. Experimental\nresults demonstrate consistent improvement (by up to 16.2\\%), yielded by the\nproposed method, over the state-of-the-art (SOTA) techniques.\n", "date": "2024-04-28", "categories": []}
{"id": "2404.18161", "title": "IMEX-Reg: Implicit-Explicit Regularization in the Function Space for\n  Continual Learning", "abstract": "  Continual learning (CL) remains one of the long-standing challenges for deep\nneural networks due to catastrophic forgetting of previously acquired\nknowledge. Although rehearsal-based approaches have been fairly successful in\nmitigating catastrophic forgetting, they suffer from overfitting on buffered\nsamples and prior information loss, hindering generalization under low-buffer\nregimes. Inspired by how humans learn using strong inductive biases, we propose\nIMEX-Reg to improve the generalization performance of experience rehearsal in\nCL under low buffer regimes. Specifically, we employ a two-pronged\nimplicit-explicit regularization approach using contrastive representation\nlearning (CRL) and consistency regularization. To further leverage the global\nrelationship between representations learned using CRL, we propose a\nregularization strategy to guide the classifier toward the activation\ncorrelations in the unit hypersphere of the CRL. Our results show that IMEX-Reg\nsignificantly improves generalization performance and outperforms\nrehearsal-based approaches in several CL scenarios. It is also robust to\nnatural and adversarial corruptions with less task-recency bias. Additionally,\nwe provide theoretical insights to support our design decisions further.\n", "date": "2024-04-28", "categories": []}
{"id": "2312.00923", "title": "Label Delay in Online Continual Learning", "abstract": "  Online continual learning, the process of training models on streaming data,\nhas gained increasing attention in recent years. However, a critical aspect\noften overlooked is the label delay, where new data may not be labeled due to\nslow and costly annotation processes. We introduce a new continual learning\nframework with explicit modeling of the label delay between data and label\nstreams over time steps. In each step, the framework reveals both unlabeled\ndata from the current time step $t$ and labels delayed with $d$ steps, from the\ntime step $t-d$. In our extensive experiments amounting to 1060 GPU days, we\nshow that merely augmenting the computational resources is insufficient to\ntackle this challenge. Our findings underline a notable performance decline\nwhen solely relying on labeled data when the label delay becomes significant.\nMore surprisingly, when using state-of-the-art SSL and TTA techniques to\nutilize the newer, unlabeled data, they fail to surpass the performance of a\nna\\\"ive method that simply trains on the delayed supervised stream. To this\nend, we introduce a simple, efficient baseline that rehearses from the labeled\nmemory samples that are most similar to the new unlabeled samples. This method\nbridges the accuracy gap caused by label delay without significantly increasing\ncomputational complexity. We show experimentally that our method is the least\naffected by the label delay factor and in some cases successfully recovers the\naccuracy of the non-delayed counterpart. We conduct various ablations and\nsensitivity experiments, demonstrating the effectiveness of our approach.\n", "date": "2023-12-01", "categories": []}
{"id": "2311.11908", "title": "Continual Learning: Applications and the Road Forward", "abstract": "  Continual learning is a subfield of machine learning, which aims to allow\nmachine learning models to continuously learn on new data, by accumulating\nknowledge without forgetting what was learned in the past. In this work, we\ntake a step back, and ask: \"Why should one care about continual learning in the\nfirst place?\". We set the stage by examining recent continual learning papers\npublished at four major machine learning conferences, and show that\nmemory-constrained settings dominate the field. Then, we discuss five open\nproblems in machine learning, and even though they might seem unrelated to\ncontinual learning at first sight, we show that continual learning will\ninevitably be part of their solution. These problems are model editing,\npersonalization and specialization, on-device learning, faster (re-)training\nand reinforcement learning. Finally, by comparing the desiderata from these\nunsolved problems and the current assumptions in continual learning, we\nhighlight and discuss four future directions for continual learning research.\nWe hope that this work offers an interesting perspective on the future of\ncontinual learning, while displaying its potential value and the paths we have\nto pursue in order to make it successful. This work is the result of the many\ndiscussions the authors had at the Dagstuhl seminar on Deep Continual Learning,\nin March 2023.\n", "date": "2023-11-20", "categories": []}
{"id": "2401.16386", "title": "Continual Learning with Pre-Trained Models: A Survey", "abstract": "  Nowadays, real-world applications often face streaming data, which requires\nthe learning system to absorb new knowledge as data evolves. Continual Learning\n(CL) aims to achieve this goal and meanwhile overcome the catastrophic\nforgetting of former knowledge when learning new ones. Typical CL methods build\nthe model from scratch to grow with incoming data. However, the advent of the\npre-trained model (PTM) era has sparked immense research interest, particularly\nin leveraging PTMs' robust representational capabilities. This paper presents a\ncomprehensive survey of the latest advancements in PTM-based CL. We categorize\nexisting methodologies into three distinct groups, providing a comparative\nanalysis of their similarities, differences, and respective advantages and\ndisadvantages. Additionally, we offer an empirical study contrasting various\nstate-of-the-art methods to highlight concerns regarding fairness in\ncomparisons. The source code to reproduce these evaluations is available at:\nhttps://github.com/sun-hailong/LAMDA-PILOT\n", "date": "2024-01-29", "categories": []}
{"id": "2403.08755", "title": "DAM: Dynamic Adapter Merging for Continual Video QA Learning", "abstract": "  We present a parameter-efficient method for continual video\nquestion-answering (VidQA) learning. Our method, named DAM, uses the proposed\nDynamic Adapter Merging to (i) mitigate catastrophic forgetting, (ii) enable\nefficient adaptation to continually arriving datasets, (iii) handle inputs from\nunknown datasets during inference, and (iv) enable knowledge sharing across\nsimilar dataset domains. Given a set of continually streaming VidQA datasets,\nwe sequentially train dataset-specific adapters for each dataset while freezing\nthe parameters of a large pretrained video-language backbone. During inference,\ngiven a video-question sample from an unknown domain, our method first uses the\nproposed non-parametric router function to compute a probability for each\nadapter, reflecting how relevant that adapter is to the current video-question\ninput instance. Subsequently, the proposed dynamic adapter merging scheme\naggregates all the adapter weights into a new adapter instance tailored for\nthat particular test sample to compute the final VidQA prediction, mitigating\nthe impact of inaccurate router predictions and facilitating knowledge sharing\nacross domains. Our DAM model outperforms prior state-of-the-art continual\nlearning approaches by 9.1% while exhibiting 1.9% less forgetting on 6 VidQA\ndatasets spanning various domains. We further extend DAM to continual image\nclassification and image QA and outperform prior methods by a large margin. The\ncode is publicly available at: https://github.com/klauscc/DAM\n", "date": "2024-03-13", "categories": []}
{"id": "2404.14588", "title": "Brain-Inspired Continual Learning-Robust Feature Distillation and\n  Re-Consolidation for Class Incremental Learning", "abstract": "  Artificial intelligence (AI) and neuroscience share a rich history, with\nadvancements in neuroscience shaping the development of AI systems capable of\nhuman-like knowledge retention. Leveraging insights from neuroscience and\nexisting research in adversarial and continual learning, we introduce a novel\nframework comprising two core concepts: feature distillation and\nre-consolidation. Our framework, named Robust Rehearsal, addresses the\nchallenge of catastrophic forgetting inherent in continual learning (CL)\nsystems by distilling and rehearsing robust features. Inspired by the mammalian\nbrain's memory consolidation process, Robust Rehearsal aims to emulate the\nrehearsal of distilled experiences during learning tasks. Additionally, it\nmimics memory re-consolidation, where new experiences influence the integration\nof past experiences to mitigate forgetting. Extensive experiments conducted on\nCIFAR10, CIFAR100, and real-world helicopter attitude datasets showcase the\nsuperior performance of CL models trained with Robust Rehearsal compared to\nbaseline methods. Furthermore, examining different optimization training\nobjectives-joint, continual, and adversarial learning-we highlight the crucial\nrole of feature learning in model performance. This underscores the\nsignificance of rehearsing CL-robust samples in mitigating catastrophic\nforgetting. In conclusion, aligning CL approaches with neuroscience insights\noffers promising solutions to the challenge of catastrophic forgetting, paving\nthe way for more robust and human-like AI systems.\n", "date": "2024-04-22", "categories": []}
{"id": "2404.14607", "title": "Q-Tuning: Queue-based Prompt Tuning for Lifelong Few-shot Language\n  Learning", "abstract": "  This paper introduces \\textbf{Q-tuning}, a novel approach for continual\nprompt tuning that enables the lifelong learning of a pre-trained language\nmodel. When learning a new task, Q-tuning trains a task-specific prompt by\nadding it to a prompt queue consisting of the prompts from older tasks. To\nbetter transfer the knowledge of old tasks, we design an adaptive knowledge\naggregation technique that reweighs previous prompts in the queue with a\nlearnable low-rank matrix. Once the prompt queue reaches its maximum capacity,\nwe leverage a PCA-based eviction rule to reduce the queue's size, allowing the\nnewly trained prompt to be added while preserving the primary knowledge of old\ntasks. In order to mitigate the accumulation of information loss caused by the\neviction, we additionally propose a globally shared prefix prompt and a memory\nretention regularization based on information theory. Extensive experiments\ndemonstrate that our approach outperforms the state-of-the-art methods\nsubstantially on continual prompt tuning benchmarks. Moreover, our approach\nenables lifelong learning on linearly growing task sequences while requiring\nconstant complexity for training and inference.\n", "date": "2024-04-22", "categories": []}
{"id": "2305.13804", "title": "OER: Offline Experience Replay for Continual Offline Reinforcement\n  Learning", "abstract": "  The capability of continuously learning new skills via a sequence of\npre-collected offline datasets is desired for an agent. However, consecutively\nlearning a sequence of offline tasks likely leads to the catastrophic\nforgetting issue under resource-limited scenarios. In this paper, we formulate\na new setting, continual offline reinforcement learning (CORL), where an agent\nlearns a sequence of offline reinforcement learning tasks and pursues good\nperformance on all learned tasks with a small replay buffer without exploring\nany of the environments of all the sequential tasks. For consistently learning\non all sequential tasks, an agent requires acquiring new knowledge and\nmeanwhile preserving old knowledge in an offline manner. To this end, we\nintroduced continual learning algorithms and experimentally found experience\nreplay (ER) to be the most suitable algorithm for the CORL problem. However, we\nobserve that introducing ER into CORL encounters a new distribution shift\nproblem: the mismatch between the experiences in the replay buffer and\ntrajectories from the learned policy. To address such an issue, we propose a\nnew model-based experience selection (MBES) scheme to build the replay buffer,\nwhere a transition model is learned to approximate the state distribution. This\nmodel is used to bridge the distribution bias between the replay buffer and the\nlearned model by filtering the data from offline data that most closely\nresembles the learned model for storage. Moreover, in order to enhance the\nability on learning new tasks, we retrofit the experience replay method with a\nnew dual behavior cloning (DBC) architecture to avoid the disturbance of\nbehavior-cloning loss on the Q-learning process. In general, we call our\nalgorithm offline experience replay (OER). Extensive experiments demonstrate\nthat our OER method outperforms SOTA baselines in widely-used Mujoco\nenvironments.\n", "date": "2023-05-23", "categories": []}
{"id": "2311.08106", "title": "Carpe Diem: On the Evaluation of World Knowledge in Lifelong Language\n  Models", "abstract": "  The dynamic nature of knowledge in an ever-changing world presents challenges\nfor language models trained on static data; the model in the real world often\nrequires not only acquiring new knowledge but also overwriting outdated\ninformation into updated ones. To study the ability of language models for\nthese time-dependent dynamics in human language, we introduce a novel task,\nEvolvingQA, a temporally evolving question-answering benchmark designed for\ntraining and evaluating LMs on an evolving Wikipedia database. The construction\nof EvolvingQA is automated with our pipeline using large language models. We\nuncover that existing continual learning baselines suffer from updating and\nremoving outdated knowledge. Our analysis suggests that models fail to rectify\nknowledge due to small weight gradients. In addition, we elucidate that\nlanguage models particularly struggle to reflect the change of numerical or\ntemporal information. Our work aims to model the dynamic nature of real-world\ninformation, suggesting faithful evaluations of the evolution-adaptability of\nlanguage models.\n", "date": "2023-11-14", "categories": []}
{"id": "2404.00257", "title": "YOLOOC: YOLO-based Open-Class Incremental Object Detection with Novel\n  Class Discovery", "abstract": "  Because of its use in practice, open-world object detection (OWOD) has gotten\na lot of attention recently. The challenge is how can a model detect novel\nclasses and then incrementally learn them without forgetting previously known\nclasses. Previous approaches hinge on strongly-supervised or weakly-supervised\nnovel-class data for novel-class detection, which may not apply to real\napplications. We construct a new benchmark that novel classes are only\nencountered at the inference stage. And we propose a new OWOD detector YOLOOC,\nbased on the YOLO architecture yet for the Open-Class setup. We introduce label\nsmoothing to prevent the detector from over-confidently mapping novel classes\nto known classes and to discover novel classes. Extensive experiments conducted\non our more realistic setup demonstrate the effectiveness of our method for\ndiscovering novel classes in our new benchmark.\n", "date": "2024-03-30", "categories": []}
{"id": "2404.13148", "title": "BACS: Background Aware Continual Semantic Segmentation", "abstract": "  Semantic segmentation plays a crucial role in enabling comprehensive scene\nunderstanding for robotic systems. However, generating annotations is\nchallenging, requiring labels for every pixel in an image. In scenarios like\nautonomous driving, there's a need to progressively incorporate new classes as\nthe operating environment of the deployed agent becomes more complex. For\nenhanced annotation efficiency, ideally, only pixels belonging to new classes\nwould be annotated. This approach is known as Continual Semantic Segmentation\n(CSS). Besides the common problem of classical catastrophic forgetting in the\ncontinual learning setting, CSS suffers from the inherent ambiguity of the\nbackground, a phenomenon we refer to as the \"background shift'', since pixels\nlabeled as background could correspond to future classes (forward background\nshift) or previous classes (backward background shift). As a result, continual\nlearning approaches tend to fail. This paper proposes a Backward Background\nShift Detector (BACS) to detect previously observed classes based on their\ndistance in the latent space from the foreground centroids of previous steps.\nMoreover, we propose a modified version of the cross-entropy loss function,\nincorporating the BACS detector to down-weight background pixels associated\nwith formerly observed classes. To combat catastrophic forgetting, we employ\nmasked feature distillation alongside dark experience replay. Additionally, our\napproach includes a transformer decoder capable of adjusting to new classes\nwithout necessitating an additional classification head. We validate BACS's\nsuperior performance over existing state-of-the-art methods on standard CSS\nbenchmarks.\n", "date": "2024-04-19", "categories": []}
{"id": "2404.14099", "title": "DynaMMo: Dynamic Model Merging for Efficient Class Incremental Learning\n  for Medical Images", "abstract": "  Continual learning, the ability to acquire knowledge from new data while\nretaining previously learned information, is a fundamental challenge in machine\nlearning. Various approaches, including memory replay, knowledge distillation,\nmodel regularization, and dynamic network expansion, have been proposed to\naddress this issue. Thus far, dynamic network expansion methods have achieved\nstate-of-the-art performance at the cost of incurring significant computational\noverhead. This is due to the need for additional model buffers, which makes it\nless feasible in resource-constrained settings, particularly in the medical\ndomain. To overcome this challenge, we propose Dynamic Model Merging, DynaMMo,\na method that merges multiple networks at different stages of model training to\nachieve better computational efficiency. Specifically, we employ lightweight\nlearnable modules for each task and combine them into a unified model to\nminimize computational overhead. DynaMMo achieves this without compromising\nperformance, offering a cost-effective solution for continual learning in\nmedical applications. We evaluate DynaMMo on three publicly available datasets,\ndemonstrating its effectiveness compared to existing approaches. DynaMMo offers\naround 10-fold reduction in GFLOPS with a small drop of 2.76 in average\naccuracy when compared to state-of-the-art dynamic-based approaches. The code\nimplementation of this work will be available upon the acceptance of this work\nat https://github.com/BioMedIA-MBZUAI/DynaMMo.\n", "date": "2024-04-22", "categories": []}
{"id": "2308.12462", "title": "Overcoming Generic Knowledge Loss with Selective Parameter Update", "abstract": "  Foundation models encompass an extensive knowledge base and offer remarkable\ntransferability. However, this knowledge becomes outdated or insufficient over\ntime. The challenge lies in continuously updating foundation models to\naccommodate novel information while retaining their original capabilities.\nLeveraging the fact that foundation models have initial knowledge on various\ntasks and domains, we propose a novel approach that, instead of updating all\nparameters equally, localizes the updates to a sparse set of parameters\nrelevant to the task being learned. We strike a balance between efficiency and\nnew task performance, while maintaining the transferability and\ngeneralizability of foundation models. We extensively evaluate our method on\nfoundational vision-language models with a diverse spectrum of continual\nlearning tasks. Our method achieves improvements on the accuracy of the newly\nlearned tasks up to 7% while preserving the pretraining knowledge with a\nnegligible decrease of 0.9% on a representative control set accuracy.\n", "date": "2023-08-23", "categories": []}
{"id": "2404.11981", "title": "Tendency-driven Mutual Exclusivity for Weakly Supervised Incremental\n  Semantic Segmentation", "abstract": "  Weakly Incremental Learning for Semantic Segmentation (WILSS) leverages a\npre-trained segmentation model to segment new classes using cost-effective and\nreadily available image-level labels. A prevailing way to solve WILSS is the\ngeneration of seed areas for each new class, serving as a form of pixel-level\nsupervision. However, a scenario usually arises where a pixel is concurrently\npredicted as an old class by the pre-trained segmentation model and a new class\nby the seed areas. Such a scenario becomes particularly problematic in WILSS,\nas the lack of pixel-level annotations on new classes makes it intractable to\nascertain whether the pixel pertains to the new class or not. To surmount this\nissue, we propose an innovative, tendency-driven relationship of mutual\nexclusivity, meticulously tailored to govern the behavior of the seed areas and\nthe predictions generated by the pre-trained segmentation model. This\nrelationship stipulates that predictions for the new and old classes must not\nconflict whilst prioritizing the preservation of predictions for the old\nclasses, which not only addresses the conflicting prediction issue but also\neffectively mitigates the inherent challenge of incremental learning -\ncatastrophic forgetting. Furthermore, under the auspices of this\ntendency-driven mutual exclusivity relationship, we generate pseudo masks for\nthe new classes, allowing for concurrent execution with model parameter\nupdating via the resolution of a bi-level optimization problem. Extensive\nexperiments substantiate the effectiveness of our framework, resulting in the\nestablishment of new benchmarks and paving the way for further research in this\nfield.\n", "date": "2024-04-18", "categories": []}
{"id": "2404.12526", "title": "Adaptive Memory Replay for Continual Learning", "abstract": "  Foundation Models (FMs) have become the hallmark of modern AI, however, these\nmodels are trained on massive data, leading to financially expensive training.\nUpdating FMs as new data becomes available is important, however, can lead to\n`catastrophic forgetting', where models underperform on tasks related to data\nsub-populations observed too long ago. This continual learning (CL) phenomenon\nhas been extensively studied, but primarily in a setting where only a small\namount of past data can be stored. We advocate for the paradigm where memory is\nabundant, allowing us to keep all previous data, but computational resources\nare limited. In this setting, traditional replay-based CL approaches are\noutperformed by a simple baseline which replays past data selected uniformly at\nrandom, indicating that this setting necessitates a new approach. We address\nthis by introducing a framework of adaptive memory replay for continual\nlearning, where sampling of past data is phrased as a multi-armed bandit\nproblem. We utilize Bolzmann sampling to derive a method which dynamically\nselects past data for training conditioned on the current task, assuming full\ndata access and emphasizing training efficiency. Through extensive evaluations\non both vision and language pre-training tasks, we demonstrate the\neffectiveness of our approach, which maintains high performance while reducing\nforgetting by up to 10% at no training efficiency cost.\n", "date": "2024-04-18", "categories": []}
{"id": "2404.12710", "title": "FedMeS: Personalized Federated Continual Learning Leveraging Local\n  Memory", "abstract": "  We focus on the problem of Personalized Federated Continual Learning (PFCL):\na group of distributed clients, each with a sequence of local tasks on\narbitrary data distributions, collaborate through a central server to train a\npersonalized model at each client, with the model expected to achieve good\nperformance on all local tasks. We propose a novel PFCL framework called\nFederated Memory Strengthening FedMeS to address the challenges of client drift\nand catastrophic forgetting. In FedMeS, each client stores samples from\nprevious tasks using a small amount of local memory, and leverages this\ninformation to both 1) calibrate gradient updates in training process; and 2)\nperform KNN-based Gaussian inference to facilitate personalization. FedMeS is\ndesigned to be task-oblivious, such that the same inference process is applied\nto samples from all tasks to achieve good performance. FedMeS is analyzed\ntheoretically and evaluated experimentally. It is shown to outperform all\nbaselines in average accuracy and forgetting rate, over various combinations of\ndatasets, task distributions, and client numbers.\n", "date": "2024-04-19", "categories": []}
{"id": "2404.10662", "title": "Continual Offline Reinforcement Learning via Diffusion-based Dual\n  Generative Replay", "abstract": "  We study continual offline reinforcement learning, a practical paradigm that\nfacilitates forward transfer and mitigates catastrophic forgetting to tackle\nsequential offline tasks. We propose a dual generative replay framework that\nretains previous knowledge by concurrent replay of generated pseudo-data.\nFirst, we decouple the continual learning policy into a diffusion-based\ngenerative behavior model and a multi-head action evaluation model, allowing\nthe policy to inherit distributional expressivity for encompassing a\nprogressive range of diverse behaviors. Second, we train a task-conditioned\ndiffusion model to mimic state distributions of past tasks. Generated states\nare paired with corresponding responses from the behavior generator to\nrepresent old tasks with high-fidelity replayed samples. Finally, by\ninterleaving pseudo samples with real ones of the new task, we continually\nupdate the state and behavior generators to model progressively diverse\nbehaviors, and regularize the multi-head critic via behavior cloning to\nmitigate forgetting. Experiments demonstrate that our method achieves better\nforward transfer with less forgetting, and closely approximates the results of\nusing previous ground-truth data due to its high-fidelity replay of the sample\nspace. Our code is available at\n\\href{https://github.com/NJU-RL/CuGRO}{https://github.com/NJU-RL/CuGRO}.\n", "date": "2024-04-16", "categories": []}
{"id": "2401.00622", "title": "Federated Class-Incremental Learning with New-Class Augmented\n  Self-Distillation", "abstract": "  Federated Learning (FL) enables collaborative model training among\nparticipants while guaranteeing the privacy of raw data. Mainstream FL\nmethodologies overlook the dynamic nature of real-world data, particularly its\ntendency to grow in volume and diversify in classes over time. This oversight\nresults in FL methods suffering from catastrophic forgetting, where the trained\nmodels inadvertently discard previously learned information upon assimilating\nnew data. In response to this challenge, we propose a novel Federated\nClass-Incremental Learning (FCIL) method, named \\underline{Fed}erated\n\\underline{C}lass-Incremental \\underline{L}earning with New-Class\n\\underline{A}ugmented \\underline{S}elf-Di\\underline{S}tillation (FedCLASS). The\ncore of FedCLASS is to enrich the class scores of historical models with new\nclass scores predicted by current models and utilize the combined knowledge for\nself-distillation, enabling a more sufficient and precise knowledge transfer\nfrom historical models to current models. Theoretical analyses demonstrate that\nFedCLASS stands on reliable foundations, considering scores of old classes\npredicted by historical models as conditional probabilities in the absence of\nnew classes, and the scores of new classes predicted by current models as the\nconditional probabilities of class scores derived from historical models.\nEmpirical experiments demonstrate the superiority of FedCLASS over four\nbaseline algorithms in reducing average forgetting rate and boosting global\naccuracy.\n", "date": "2024-01-01", "categories": []}
{"id": "2404.05555", "title": "On the Convergence of Continual Learning with Adaptive Methods", "abstract": "  One of the objectives of continual learning is to prevent catastrophic\nforgetting in learning multiple tasks sequentially, and the existing solutions\nhave been driven by the conceptualization of the plasticity-stability dilemma.\nHowever, the convergence of continual learning for each sequential task is less\nstudied so far. In this paper, we provide a convergence analysis of\nmemory-based continual learning with stochastic gradient descent and empirical\nevidence that training current tasks causes the cumulative degradation of\nprevious tasks. We propose an adaptive method for nonconvex continual learning\n(NCCL), which adjusts step sizes of both previous and current tasks with the\ngradients. The proposed method can achieve the same convergence rate as the SGD\nmethod when the catastrophic forgetting term which we define in the paper is\nsuppressed at each iteration. Further, we demonstrate that the proposed\nalgorithm improves the performance of continual learning over existing methods\nfor several image classification tasks.\n", "date": "2024-04-08", "categories": []}
{"id": "2404.09339", "title": "Towards Practical Tool Usage for Continually Learning LLMs", "abstract": "  Large language models (LLMs) show an innate skill for solving language based\ntasks. But insights have suggested an inability to adjust for information or\ntask-solving skills becoming outdated, as their knowledge, stored directly\nwithin their parameters, remains static in time. Tool use helps by offloading\nwork to systems that the LLM can access through an interface, but LLMs that use\nthem still must adapt to nonstationary environments for prolonged use, as new\ntools can emerge and existing tools can change. Nevertheless, tools require\nless specialized knowledge, therefore we hypothesize they are better suited for\ncontinual learning (CL) as they rely less on parametric memory for solving\ntasks and instead focus on learning when to apply pre-defined tools. To verify\nthis, we develop a synthetic benchmark and follow this by aggregating existing\nNLP tasks to form a more realistic testing scenario. While we demonstrate\nscaling model size is not a solution, regardless of tool usage, continual\nlearning techniques can enable tool LLMs to both adapt faster while forgetting\nless, highlighting their potential as continual learners.\n", "date": "2024-04-14", "categories": []}
{"id": "2404.07817", "title": "Calibration of Continual Learning Models", "abstract": "  Continual Learning (CL) focuses on maximizing the predictive performance of a\nmodel across a non-stationary stream of data. Unfortunately, CL models tend to\nforget previous knowledge, thus often underperforming when compared with an\noffline model trained jointly on the entire data stream. Given that any CL\nmodel will eventually make mistakes, it is of crucial importance to build\ncalibrated CL models: models that can reliably tell their confidence when\nmaking a prediction. Model calibration is an active research topic in machine\nlearning, yet to be properly investigated in CL. We provide the first empirical\nstudy of the behavior of calibration approaches in CL, showing that CL\nstrategies do not inherently learn calibrated models. To mitigate this issue,\nwe design a continual calibration approach that improves the performance of\npost-processing calibration methods over a wide range of different benchmarks\nand CL strategies. CL does not necessarily need perfect predictive models, but\nrather it can benefit from reliable predictive models. We believe our study on\ncontinual calibration represents a first step towards this direction.\n", "date": "2024-04-11", "categories": []}
{"id": "2404.07470", "title": "Scalable Language Model with Generalized Continual Learning", "abstract": "  Continual learning has gained increasing importance as it facilitates the\nacquisition and refinement of scalable knowledge and skills in language models.\nHowever, existing methods typically encounter strict limitations and challenges\nin real-world scenarios, such as reliance on experience replay, optimization\nconstraints, and inference task-ID. In this study, we introduce the Scalable\nLanguage Model (SLM) to overcome these limitations within a more challenging\nand generalized setting, representing a significant advancement toward\npractical applications for continual learning. Specifically, we propose the\nJoint Adaptive Re-Parameterization (JARe), integrated with Dynamic Task-related\nKnowledge Retrieval (DTKR), to enable adaptive adjustment of language models\nbased on specific downstream tasks. This approach leverages the task\ndistribution within the vector space, aiming to achieve a smooth and effortless\ncontinual learning process. Our method demonstrates state-of-the-art\nperformance on diverse backbones and benchmarks, achieving effective continual\nlearning in both full-set and few-shot scenarios with minimal forgetting.\nMoreover, while prior research primarily focused on a single task type such as\nclassification, our study goes beyond, with the large language model, i.e.,\nLLaMA-2, to explore the effects across diverse domains and task types, such\nthat a single language model can be decently scaled to broader applications.\n", "date": "2024-04-11", "categories": []}
{"id": "2404.07507", "title": "Learning to Classify New Foods Incrementally Via Compressed Exemplars", "abstract": "  Food image classification systems play a crucial role in health monitoring\nand diet tracking through image-based dietary assessment techniques. However,\nexisting food recognition systems rely on static datasets characterized by a\npre-defined fixed number of food classes. This contrasts drastically with the\nreality of food consumption, which features constantly changing data.\nTherefore, food image classification systems should adapt to and manage data\nthat continuously evolves. This is where continual learning plays an important\nrole. A challenge in continual learning is catastrophic forgetting, where ML\nmodels tend to discard old knowledge upon learning new information. While\nmemory-replay algorithms have shown promise in mitigating this problem by\nstoring old data as exemplars, they are hampered by the limited capacity of\nmemory buffers, leading to an imbalance between new and previously learned\ndata. To address this, our work explores the use of neural image compression to\nextend buffer size and enhance data diversity. We introduced the concept of\ncontinuously learning a neural compression model to adaptively improve the\nquality of compressed data and optimize the bitrates per pixel (bpp) to store\nmore exemplars. Our extensive experiments, including evaluations on\nfood-specific datasets including Food-101 and VFN-74, as well as the general\ndataset ImageNet-100, demonstrate improvements in classification accuracy. This\nprogress is pivotal in advancing more realistic food recognition systems that\nare capable of adapting to continually evolving data. Moreover, the principles\nand methodologies we've developed hold promise for broader applications,\nextending their benefits to other domains of continual machine learning\nsystems.\n", "date": "2024-04-11", "categories": []}
{"id": "2404.07729", "title": "Realistic Continual Learning Approach using Pre-trained Models", "abstract": "  Continual learning (CL) is crucial for evaluating adaptability in learning\nsolutions to retain knowledge. Our research addresses the challenge of\ncatastrophic forgetting, where models lose proficiency in previously learned\ntasks as they acquire new ones. While numerous solutions have been proposed,\nexisting experimental setups often rely on idealized class-incremental learning\nscenarios. We introduce Realistic Continual Learning (RealCL), a novel CL\nparadigm where class distributions across tasks are random, departing from\nstructured setups.\n  We also present CLARE (Continual Learning Approach with pRE-trained models\nfor RealCL scenarios), a pre-trained model-based solution designed to integrate\nnew knowledge while preserving past learning. Our contributions include\npioneering RealCL as a generalization of traditional CL setups, proposing CLARE\nas an adaptable approach for RealCL tasks, and conducting extensive experiments\ndemonstrating its effectiveness across various RealCL scenarios. Notably, CLARE\noutperforms existing models on RealCL benchmarks, highlighting its versatility\nand robustness in unpredictable learning environments.\n", "date": "2024-04-11", "categories": []}
{"id": "2306.13812", "title": "Maintaining Plasticity in Deep Continual Learning", "abstract": "  Modern deep-learning systems are specialized to problem settings in which\ntraining occurs once and then never again, as opposed to continual-learning\nsettings in which training occurs continually. If deep-learning systems are\napplied in a continual learning setting, then it is well known that they may\nfail to remember earlier examples. More fundamental, but less well known, is\nthat they may also lose their ability to learn on new examples, a phenomenon\ncalled loss of plasticity. We provide direct demonstrations of loss of\nplasticity using the MNIST and ImageNet datasets repurposed for continual\nlearning as sequences of tasks. In ImageNet, binary classification performance\ndropped from 89% accuracy on an early task down to 77%, about the level of a\nlinear network, on the 2000th task. Loss of plasticity occurred with a wide\nrange of deep network architectures, optimizers, activation functions, batch\nnormalization, dropout, but was substantially eased by L2-regularization,\nparticularly when combined with weight perturbation. Further, we introduce a\nnew algorithm -- continual backpropagation -- which slightly modifies\nconventional backpropagation to reinitialize a small fraction of less-used\nunits after each example and appears to maintain plasticity indefinitely.\n", "date": "2023-06-23", "categories": []}
{"id": "2404.06622", "title": "Calibrating Higher-Order Statistics for Few-Shot Class-Incremental\n  Learning with Pre-trained Vision Transformers", "abstract": "  Few-shot class-incremental learning (FSCIL) aims to adapt the model to new\nclasses from very few data (5 samples) without forgetting the previously\nlearned classes. Recent works in many-shot CIL (MSCIL) (using all available\ntraining data) exploited pre-trained models to reduce forgetting and achieve\nbetter plasticity. In a similar fashion, we use ViT models pre-trained on\nlarge-scale datasets for few-shot settings, which face the critical issue of\nlow plasticity. FSCIL methods start with a many-shot first task to learn a very\ngood feature extractor and then move to the few-shot setting from the second\ntask onwards. While the focus of most recent studies is on how to learn the\nmany-shot first task so that the model generalizes to all future few-shot\ntasks, we explore in this work how to better model the few-shot data using\npre-trained models, irrespective of how the first task is trained. Inspired by\nrecent works in MSCIL, we explore how using higher-order feature statistics can\ninfluence the classification of few-shot classes. We identify the main\nchallenge of obtaining a good covariance matrix from few-shot data and propose\nto calibrate the covariance matrix for new classes based on semantic similarity\nto the many-shot base classes. Using the calibrated feature statistics in\ncombination with existing methods significantly improves few-shot continual\nclassification on several FSCIL benchmarks. Code is available at\nhttps://github.com/dipamgoswami/FSCIL-Calibration.\n", "date": "2024-04-09", "categories": []}
{"id": "2404.06972", "title": "Toward industrial use of continual learning : new metrics proposal for\n  class incremental learning", "abstract": "  In this paper, we investigate continual learning performance metrics used in\nclass incremental learning strategies for continual learning (CL) using some\nhigh performing methods. We investigate especially mean task accuracy. First,\nwe show that it lacks of expressiveness through some simple experiments to\ncapture performance. We show that monitoring average tasks performance is over\noptimistic and can lead to misleading conclusions for future real life\nindustrial uses. Then, we propose first a simple metric, Minimal Incremental\nClass Accuracy (MICA) which gives a fair and more useful evaluation of\ndifferent continual learning methods. Moreover, in order to provide a simple\nway to easily compare different methods performance in continual learning, we\nderive another single scalar metric that take into account the learning\nperformance variation as well as our newly introduced metric.\n", "date": "2024-04-10", "categories": []}
{"id": "2309.05334", "title": "MultIOD: Rehearsal-free Multihead Incremental Object Detector", "abstract": "  Class-Incremental learning (CIL) refers to the ability of artificial agents\nto integrate new classes as they appear in a stream. It is particularly\ninteresting in evolving environments where agents have limited access to memory\nand computational resources. The main challenge of incremental learning is\ncatastrophic forgetting, the inability of neural networks to retain past\nknowledge when learning a new one. Unfortunately, most existing\nclass-incremental methods for object detection are applied to two-stage\nalgorithms such as Faster-RCNN, and rely on rehearsal memory to retain past\nknowledge. We argue that those are not suitable in resource-limited\nenvironments, and more effort should be dedicated to anchor-free and\nrehearsal-free object detection. In this paper, we propose MultIOD, a\nclass-incremental object detector based on CenterNet. Our contributions are:\n(1) we propose a multihead feature pyramid and multihead detection architecture\nto efficiently separate class representations, (2) we employ transfer learning\nbetween classes learned initially and those learned incrementally to tackle\ncatastrophic forgetting, and (3) we use a class-wise non-max-suppression as a\npost-processing technique to remove redundant boxes. Results show that our\nmethod outperforms state-of-the-art methods on two Pascal VOC datasets, while\nonly saving the model in its current state, contrary to other\ndistillation-based counterparts.\n", "date": "2023-09-11", "categories": []}
{"id": "2402.01348", "title": "CORE: Mitigating Catastrophic Forgetting in Continual Learning through\n  Cognitive Replay", "abstract": "  This paper introduces a novel perspective to significantly mitigate\ncatastrophic forgetting in continuous learning (CL), which emphasizes models'\ncapacity to preserve existing knowledge and assimilate new information. Current\nreplay-based methods treat every task and data sample equally and thus can not\nfully exploit the potential of the replay buffer. In response, we propose\nCOgnitive REplay (CORE), which draws inspiration from human cognitive review\nprocesses. CORE includes two key strategies: Adaptive Quantity Allocation and\nQuality-Focused Data Selection. The former adaptively modulates the replay\nbuffer allocation for each task based on its forgetting rate, while the latter\nguarantees the inclusion of representative data that best encapsulates the\ncharacteristics of each task within the buffer. Our approach achieves an\naverage accuracy of 37.95% on split-CIFAR10, surpassing the best baseline\nmethod by 6.52%. Additionally, it significantly enhances the accuracy of the\npoorest-performing task by 6.30% compared to the top baseline. Code is\navailable at https://github.com/sterzhang/CORE.\n", "date": "2024-02-02", "categories": []}
{"id": "2404.04002", "title": "Continual Learning with Weight Interpolation", "abstract": "  Continual learning poses a fundamental challenge for modern machine learning\nsystems, requiring models to adapt to new tasks while retaining knowledge from\nprevious ones. Addressing this challenge necessitates the development of\nefficient algorithms capable of learning from data streams and accumulating\nknowledge over time. This paper proposes a novel approach to continual learning\nutilizing the weight consolidation method. Our method, a simple yet powerful\ntechnique, enhances robustness against catastrophic forgetting by interpolating\nbetween old and new model weights after each novel task, effectively merging\ntwo models to facilitate exploration of local minima emerging after arrival of\nnew concepts. Moreover, we demonstrate that our approach can complement\nexisting rehearsal-based replay approaches, improving their accuracy and\nfurther mitigating the forgetting phenomenon. Additionally, our method provides\nan intuitive mechanism for controlling the stability-plasticity trade-off.\nExperimental results showcase the significant performance enhancement to\nstate-of-the-art experience replay algorithms the proposed weight consolidation\napproach offers. Our algorithm can be downloaded from\nhttps://github.com/jedrzejkozal/weight-interpolation-cl.\n", "date": "2024-04-05", "categories": []}
{"id": "2401.08478", "title": "Solving Continual Offline Reinforcement Learning with Decision\n  Transformer", "abstract": "  Continuous offline reinforcement learning (CORL) combines continuous and\noffline reinforcement learning, enabling agents to learn multiple tasks from\nstatic datasets without forgetting prior tasks. However, CORL faces challenges\nin balancing stability and plasticity. Existing methods, employing Actor-Critic\nstructures and experience replay (ER), suffer from distribution shifts, low\nefficiency, and weak knowledge-sharing. We aim to investigate whether Decision\nTransformer (DT), another offline RL paradigm, can serve as a more suitable\noffline continuous learner to address these issues. We first compare AC-based\noffline algorithms with DT in the CORL framework. DT offers advantages in\nlearning efficiency, distribution shift mitigation, and zero-shot\ngeneralization but exacerbates the forgetting problem during supervised\nparameter updates. We introduce multi-head DT (MH-DT) and low-rank adaptation\nDT (LoRA-DT) to mitigate DT's forgetting problem. MH-DT stores task-specific\nknowledge using multiple heads, facilitating knowledge sharing with common\ncomponents. It employs distillation and selective rehearsal to enhance current\ntask learning when a replay buffer is available. In buffer-unavailable\nscenarios, LoRA-DT merges less influential weights and fine-tunes DT's decisive\nMLP layer to adapt to the current task. Extensive experiments on MoJuCo and\nMeta-World benchmarks demonstrate that our methods outperform SOTA CORL\nbaselines and showcase enhanced learning capabilities and superior memory\nefficiency.\n", "date": "2024-01-16", "categories": []}
{"id": "2404.04476", "title": "DELTA: Decoupling Long-Tailed Online Continual Learning", "abstract": "  A significant challenge in achieving ubiquitous Artificial Intelligence is\nthe limited ability of models to rapidly learn new information in real-world\nscenarios where data follows long-tailed distributions, all while avoiding\nforgetting previously acquired knowledge. In this work, we study the\nunder-explored problem of Long-Tailed Online Continual Learning (LTOCL), which\naims to learn new tasks from sequentially arriving class-imbalanced data\nstreams. Each data is observed only once for training without knowing the task\ndata distribution. We present DELTA, a decoupled learning approach designed to\nenhance learning representations and address the substantial imbalance in\nLTOCL. We enhance the learning process by adapting supervised contrastive\nlearning to attract similar samples and repel dissimilar (out-of-class)\nsamples. Further, by balancing gradients during training using an equalization\nloss, DELTA significantly enhances learning outcomes and successfully mitigates\ncatastrophic forgetting. Through extensive evaluation, we demonstrate that\nDELTA improves the capacity for incremental learning, surpassing existing OCL\nmethods. Our results suggest considerable promise for applying OCL in\nreal-world applications.\n", "date": "2024-04-06", "categories": []}
{"id": "2404.04871", "title": "Data Stream Sampling with Fuzzy Task Boundaries and Noisy Labels", "abstract": "  In the realm of continual learning, the presence of noisy labels within data\nstreams represents a notable obstacle to model reliability and fairness. We\nfocus on the data stream scenario outlined in pertinent literature,\ncharacterized by fuzzy task boundaries and noisy labels. To address this\nchallenge, we introduce a novel and intuitive sampling method called Noisy Test\nDebiasing (NTD) to mitigate noisy labels in evolving data streams and establish\na fair and robust continual learning algorithm. NTD is straightforward to\nimplement, making it feasible across various scenarios. Our experiments\nbenchmark four datasets, including two synthetic noise datasets (CIFAR10 and\nCIFAR100) and real-world noise datasets (mini-WebVision and Food-101N). The\nresults validate the efficacy of NTD for online continual learning in scenarios\nwith noisy labels in data streams. Compared to the previous leading approach,\nNTD achieves a training speedup enhancement over two times while maintaining or\nsurpassing accuracy levels. Moreover, NTD utilizes less than one-fifth of the\nGPU memory resources compared to previous leading methods.\n", "date": "2024-04-07", "categories": []}
{"id": "2306.05401", "title": "RDumb: A simple approach that questions our progress in continual\n  test-time adaptation", "abstract": "  Test-Time Adaptation (TTA) allows to update pre-trained models to changing\ndata distributions at deployment time. While early work tested these algorithms\nfor individual fixed distribution shifts, recent work proposed and applied\nmethods for continual adaptation over long timescales. To examine the reported\nprogress in the field, we propose the Continually Changing Corruptions (CCC)\nbenchmark to measure asymptotic performance of TTA techniques. We find that\neventually all but one state-of-the-art methods collapse and perform worse than\na non-adapting model, including models specifically proposed to be robust to\nperformance collapse. In addition, we introduce a simple baseline, \"RDumb\",\nthat periodically resets the model to its pretrained state. RDumb performs\nbetter or on par with the previously proposed state-of-the-art in all\nconsidered benchmarks. Our results show that previous TTA approaches are\nneither effective at regularizing adaptation to avoid collapse nor able to\noutperform a simplistic resetting strategy.\n", "date": "2023-06-08", "categories": []}
{"id": "2404.00228", "title": "InfLoRA: Interference-Free Low-Rank Adaptation for Continual Learning", "abstract": "  Continual learning requires the model to learn multiple tasks sequentially.\nIn continual learning, the model should possess the ability to maintain its\nperformance on old tasks (stability) and the ability to adapt to new tasks\ncontinuously (plasticity). Recently, parameter-efficient fine-tuning (PEFT),\nwhich involves freezing a pre-trained model and injecting a small number of\nlearnable parameters to adapt to downstream tasks, has gained increasing\npopularity in continual learning. Although existing continual learning methods\nbased on PEFT have demonstrated superior performance compared to those not\nbased on PEFT, most of them do not consider how to eliminate the interference\nof the new task on the old tasks, which inhibits the model from making a good\ntrade-off between stability and plasticity. In this work, we propose a new PEFT\nmethod, called interference-free low-rank adaptation (InfLoRA), for continual\nlearning. InfLoRA injects a small number of parameters to reparameterize the\npre-trained weights and shows that fine-tuning these injected parameters is\nequivalent to fine-tuning the pre-trained weights within a subspace.\nFurthermore, InfLoRA designs this subspace to eliminate the interference of the\nnew task on the old tasks, making a good trade-off between stability and\nplasticity. Experimental results show that InfLoRA outperforms existing\nstate-of-the-art continual learning methods on multiple datasets.\n", "date": "2024-03-30", "categories": []}
{"id": "2404.02507", "title": "Lifelong Event Detection with Embedding Space Separation and Compaction", "abstract": "  To mitigate forgetting, existing lifelong event detection methods typically\nmaintain a memory module and replay the stored memory data during the learning\nof a new task. However, the simple combination of memory data and new-task\nsamples can still result in substantial forgetting of previously acquired\nknowledge, which may occur due to the potential overlap between the feature\ndistribution of new data and the previously learned embedding space. Moreover,\nthe model suffers from overfitting on the few memory samples rather than\neffectively remembering learned patterns. To address the challenges of\nforgetting and overfitting, we propose a novel method based on embedding space\nseparation and compaction. Our method alleviates forgetting of previously\nlearned tasks by forcing the feature distribution of new data away from the\nprevious embedding space. It also mitigates overfitting by a memory calibration\nmechanism that encourages memory data to be close to its prototype to enhance\nintra-class compactness. In addition, the learnable parameters of the new task\nare initialized by drawing upon acquired knowledge from the previously learned\ntask to facilitate forward knowledge transfer. With extensive experiments, we\ndemonstrate that our method can significantly outperform previous\nstate-of-the-art approaches.\n", "date": "2024-04-03", "categories": []}
{"id": "2404.02719", "title": "Can We Understand Plasticity Through Neural Collapse?", "abstract": "  This paper explores the connection between two recently identified phenomena\nin deep learning: plasticity loss and neural collapse. We analyze their\ncorrelation in different scenarios, revealing a significant association during\nthe initial training phase on the first task. Additionally, we introduce a\nregularization approach to mitigate neural collapse, demonstrating its\neffectiveness in alleviating plasticity loss in this specific setting.\n", "date": "2024-04-03", "categories": []}
{"id": "2404.02754", "title": "Continual Learning of Numerous Tasks from Long-tail Distributions", "abstract": "  Continual learning, an important aspect of artificial intelligence and\nmachine learning research, focuses on developing models that learn and adapt to\nnew tasks while retaining previously acquired knowledge. Existing continual\nlearning algorithms usually involve a small number of tasks with uniform sizes\nand may not accurately represent real-world learning scenarios. In this paper,\nwe investigate the performance of continual learning algorithms with a large\nnumber of tasks drawn from a task distribution that is long-tail in terms of\ntask sizes. We design one synthetic dataset and two real-world continual\nlearning datasets to evaluate the performance of existing algorithms in such a\nsetting. Moreover, we study an overlooked factor in continual learning, the\noptimizer states, e.g. first and second moments in the Adam optimizer, and\ninvestigate how it can be used to improve continual learning performance. We\npropose a method that reuses the optimizer states in Adam by maintaining a\nweighted average of the second moments from previous tasks. We demonstrate that\nour method, compatible with most existing continual learning algorithms,\neffectively reduces forgetting with only a small amount of additional\ncomputational or memory costs, and provides further improvements on existing\ncontinual learning algorithms, particularly in a long-tail task sequence.\n", "date": "2024-04-03", "categories": []}
{"id": "2403.16003", "title": "Diverse Representation Embedding for Lifelong Person Re-Identification", "abstract": "  Lifelong Person Re-Identification (LReID) aims to continuously learn from\nsuccessive data streams, matching individuals across multiple cameras. The key\nchallenge for LReID is how to effectively preserve old knowledge while\nincrementally learning new information, which is caused by task-level domain\ngaps and limited old task datasets. Existing methods based on CNN backbone are\ninsufficient to explore the representation of each instance from different\nperspectives, limiting model performance on limited old task datasets and new\ntask datasets. Unlike these methods, we propose a Diverse Representations\nEmbedding (DRE) framework that first explores a pure transformer for LReID. The\nproposed DRE preserves old knowledge while adapting to new information based on\ninstance-level and task-level layout. Concretely, an Adaptive Constraint Module\n(ACM) is proposed to implement integration and push away operations between\nmultiple overlapping representations generated by transformer-based backbone,\nobtaining rich and discriminative representations for each instance to improve\nadaptive ability of LReID. Based on the processed diverse representations, we\npropose Knowledge Update (KU) and Knowledge Preservation (KP) strategies at the\ntask-level layout by introducing the adjustment model and the learner model. KU\nstrategy enhances the adaptive learning ability of learner models for new\ninformation under the adjustment model prior, and KP strategy preserves old\nknowledge operated by representation-level alignment and logit-level\nsupervision in limited old task datasets while guaranteeing the adaptive\nlearning information capacity of the LReID model. Compared to state-of-the-art\nmethods, our method achieves significantly improved performance in holistic,\nlarge-scale, and occluded datasets.\n", "date": "2024-03-24", "categories": []}
{"id": "2404.01317", "title": "Intelligent Learning Rate Distribution to reduce Catastrophic Forgetting\n  in Transformers", "abstract": "  Pretraining language models on large text corpora is a common practice in\nnatural language processing. Fine-tuning of these models is then performed to\nachieve the best results on a variety of tasks. In this paper, we investigate\nthe problem of catastrophic forgetting in transformer neural networks and\nquestion the common practice of fine-tuning with a flat learning rate for the\nentire network in this context. We perform a hyperparameter optimization\nprocess to find learning rate distributions that are better than a flat\nlearning rate. We combine the learning rate distributions thus found and show\nthat they generalize to better performance with respect to the problem of\ncatastrophic forgetting. We validate these learning rate distributions with a\nvariety of NLP benchmarks from the GLUE dataset.\n", "date": "2024-03-27", "categories": []}
{"id": "2404.01628", "title": "Learning Equi-angular Representations for Online Continual Learning", "abstract": "  Online continual learning suffers from an underfitted solution due to\ninsufficient training for prompt model update (e.g., single-epoch training). To\naddress the challenge, we propose an efficient online continual learning method\nusing the neural collapse phenomenon. In particular, we induce neural collapse\nto form a simplex equiangular tight frame (ETF) structure in the representation\nspace so that the continuously learned model with a single epoch can better fit\nto the streamed data by proposing preparatory data training and residual\ncorrection in the representation space. With an extensive set of empirical\nvalidations using CIFAR-10/100, TinyImageNet, ImageNet-200, and ImageNet-1K, we\nshow that our proposed method outperforms state-of-the-art methods by a\nnoticeable margin in various online continual learning scenarios such as\ndisjoint and Gaussian scheduled continuous (i.e., boundary-free) data setups.\n", "date": "2024-04-02", "categories": []}
{"id": "2404.01828", "title": "Defense without Forgetting: Continual Adversarial Defense with\n  Anisotropic & Isotropic Pseudo Replay", "abstract": "  Deep neural networks have demonstrated susceptibility to adversarial attacks.\nAdversarial defense techniques often focus on one-shot setting to maintain\nrobustness against attack. However, new attacks can emerge in sequences in\nreal-world deployment scenarios. As a result, it is crucial for a defense model\nto constantly adapt to new attacks, but the adaptation process can lead to\ncatastrophic forgetting of previously defended against attacks. In this paper,\nwe discuss for the first time the concept of continual adversarial defense\nunder a sequence of attacks, and propose a lifelong defense baseline called\nAnisotropic \\& Isotropic Replay (AIR), which offers three advantages: (1)\nIsotropic replay ensures model consistency in the neighborhood distribution of\nnew data, indirectly aligning the output preference between old and new tasks.\n(2) Anisotropic replay enables the model to learn a compromise data manifold\nwith fresh mixed semantics for further replay constraints and potential future\nattacks. (3) A straightforward regularizer mitigates the 'plasticity-stability'\ntrade-off by aligning model output between new and old tasks. Experiment\nresults demonstrate that AIR can approximate or even exceed the empirical\nperformance upper bounds achieved by Joint Training.\n", "date": "2024-04-02", "categories": []}
{"id": "2404.02117", "title": "Pre-trained Vision and Language Transformers Are Few-Shot Incremental\n  Learners", "abstract": "  Few-Shot Class Incremental Learning (FSCIL) is a task that requires a model\nto learn new classes incrementally without forgetting when only a few samples\nfor each class are given. FSCIL encounters two significant challenges:\ncatastrophic forgetting and overfitting, and these challenges have driven prior\nstudies to primarily rely on shallow models, such as ResNet-18. Even though\ntheir limited capacity can mitigate both forgetting and overfitting issues, it\nleads to inadequate knowledge transfer during few-shot incremental sessions. In\nthis paper, we argue that large models such as vision and language transformers\npre-trained on large datasets can be excellent few-shot incremental learners.\nTo this end, we propose a novel FSCIL framework called PriViLege, Pre-trained\nVision and Language transformers with prompting functions and knowledge\ndistillation. Our framework effectively addresses the challenges of\ncatastrophic forgetting and overfitting in large models through new pre-trained\nknowledge tuning (PKT) and two losses: entropy-based divergence loss and\nsemantic knowledge distillation loss. Experimental results show that the\nproposed PriViLege significantly outperforms the existing state-of-the-art\nmethods with a large margin, e.g., +9.38% in CUB200, +20.58% in CIFAR-100, and\n+13.36% in miniImageNet. Our implementation code is available at\nhttps://github.com/KHU-AGI/PriViLege.\n", "date": "2024-04-02", "categories": []}
{"id": "2312.00600", "title": "Improving Plasticity in Online Continual Learning via Collaborative\n  Learning", "abstract": "  Online Continual Learning (CL) solves the problem of learning the\never-emerging new classification tasks from a continuous data stream. Unlike\nits offline counterpart, in online CL, the training data can only be seen once.\nMost existing online CL research regards catastrophic forgetting (i.e., model\nstability) as almost the only challenge. In this paper, we argue that the\nmodel's capability to acquire new knowledge (i.e., model plasticity) is another\nchallenge in online CL. While replay-based strategies have been shown to be\neffective in alleviating catastrophic forgetting, there is a notable gap in\nresearch attention toward improving model plasticity. To this end, we propose\nCollaborative Continual Learning (CCL), a collaborative learning based strategy\nto improve the model's capability in acquiring new concepts. Additionally, we\nintroduce Distillation Chain (DC), a collaborative learning scheme to boost the\ntraining of the models. We adapt CCL-DC to existing representative online CL\nworks. Extensive experiments demonstrate that even if the learners are\nwell-trained with state-of-the-art online CL methods, our strategy can still\nimprove model plasticity dramatically, and thereby improve the overall\nperformance by a large margin. The source code of our work is available at\nhttps://github.com/maorong-wang/CCL-DC.\n", "date": "2023-12-01", "categories": []}
{"id": "2402.18528", "title": "Gradient Reweighting: Towards Imbalanced Class-Incremental Learning", "abstract": "  Class-Incremental Learning (CIL) trains a model to continually recognize new\nclasses from non-stationary data while retaining learned knowledge. A major\nchallenge of CIL arises when applying to real-world data characterized by\nnon-uniform distribution, which introduces a dual imbalance problem involving\n(i) disparities between stored exemplars of old tasks and new class data\n(inter-phase imbalance), and (ii) severe class imbalances within each\nindividual task (intra-phase imbalance). We show that this dual imbalance issue\ncauses skewed gradient updates with biased weights in FC layers, thus inducing\nover/under-fitting and catastrophic forgetting in CIL. Our method addresses it\nby reweighting the gradients towards balanced optimization and unbiased\nclassifier learning. Additionally, we observe imbalanced forgetting where\nparadoxically the instance-rich classes suffer higher performance degradation\nduring CIL due to a larger amount of training data becoming unavailable in\nsubsequent learning phases. To tackle this, we further introduce a\ndistribution-aware knowledge distillation loss to mitigate forgetting by\naligning output logits proportionally with the distribution of lost training\ndata. We validate our method on CIFAR-100, ImageNetSubset, and Food101 across\nvarious evaluation protocols and demonstrate consistent improvements compared\nto existing works, showing great potential to apply CIL in real-world scenarios\nwith enhanced robustness and effectiveness.\n", "date": "2024-02-28", "categories": []}
{"id": "2403.03477", "title": "Continual Segmentation with Disentangled Objectness Learning and Class\n  Recognition", "abstract": "  Most continual segmentation methods tackle the problem as a per-pixel\nclassification task. However, such a paradigm is very challenging, and we find\nquery-based segmenters with built-in objectness have inherent advantages\ncompared with per-pixel ones, as objectness has strong transfer ability and\nforgetting resistance. Based on these findings, we propose CoMasTRe by\ndisentangling continual segmentation into two stages: forgetting-resistant\ncontinual objectness learning and well-researched continual classification.\nCoMasTRe uses a two-stage segmenter learning class-agnostic mask proposals at\nthe first stage and leaving recognition to the second stage. During continual\nlearning, a simple but effective distillation is adopted to strengthen\nobjectness. To further mitigate the forgetting of old classes, we design a\nmulti-label class distillation strategy suited for segmentation. We assess the\neffectiveness of CoMasTRe on PASCAL VOC and ADE20K. Extensive experiments show\nthat our method outperforms per-pixel and query-based methods on both datasets.\nCode will be available at https://github.com/jordangong/CoMasTRe.\n", "date": "2024-03-06", "categories": []}
{"id": "2404.00360", "title": "Reusable Architecture Growth for Continual Stereo Matching", "abstract": "  The remarkable performance of recent stereo depth estimation models benefits\nfrom the successful use of convolutional neural networks to regress dense\ndisparity. Akin to most tasks, this needs gathering training data that covers a\nnumber of heterogeneous scenes at deployment time. However, training samples\nare typically acquired continuously in practical applications, making the\ncapability to learn new scenes continually even more crucial. For this purpose,\nwe propose to perform continual stereo matching where a model is tasked to 1)\ncontinually learn new scenes, 2) overcome forgetting previously learned scenes,\nand 3) continuously predict disparities at inference. We achieve this goal by\nintroducing a Reusable Architecture Growth (RAG) framework. RAG leverages\ntask-specific neural unit search and architecture growth to learn new scenes\ncontinually in both supervised and self-supervised manners. It can maintain\nhigh reusability during growth by reusing previous units while obtaining good\nperformance. Additionally, we present a Scene Router module to adaptively\nselect the scene-specific architecture path at inference. Comprehensive\nexperiments on numerous datasets show that our framework performs impressively\nin various weather, road, and city circumstances and surpasses the\nstate-of-the-art methods in more challenging cross-dataset settings. Further\nexperiments also demonstrate the adaptability of our method to unseen scenes,\nwhich can facilitate end-to-end stereo architecture learning and practical\ndeployment.\n", "date": "2024-03-30", "categories": []}
{"id": "2404.00417", "title": "Orchestrate Latent Expertise: Advancing Online Continual Learning with\n  Multi-Level Supervision and Reverse Self-Distillation", "abstract": "  To accommodate real-world dynamics, artificial intelligence systems need to\ncope with sequentially arriving content in an online manner. Beyond regular\nContinual Learning (CL) attempting to address catastrophic forgetting with\noffline training of each task, Online Continual Learning (OCL) is a more\nchallenging yet realistic setting that performs CL in a one-pass data stream.\nCurrent OCL methods primarily rely on memory replay of old training samples.\nHowever, a notable gap from CL to OCL stems from the additional\noverfitting-underfitting dilemma associated with the use of rehearsal buffers:\nthe inadequate learning of new training samples (underfitting) and the repeated\nlearning of a few old training samples (overfitting). To this end, we introduce\na novel approach, Multi-level Online Sequential Experts (MOSE), which\ncultivates the model as stacked sub-experts, integrating multi-level\nsupervision and reverse self-distillation. Supervision signals across multiple\nstages facilitate appropriate convergence of the new task while gathering\nvarious strengths from experts by knowledge distillation mitigates the\nperformance decline of old tasks. MOSE demonstrates remarkable efficacy in\nlearning new samples and preserving past knowledge through multi-level experts,\nthereby significantly advancing OCL performance over state-of-the-art baselines\n(e.g., up to 7.3% on Split CIFAR-100 and 6.1% on Split Tiny-ImageNet).\n", "date": "2024-03-30", "categories": []}
{"id": "2404.00418", "title": "Continual Learning for Autonomous Robots: A Prototype-based Approach", "abstract": "  Humans and animals learn throughout their lives from limited amounts of\nsensed data, both with and without supervision. Autonomous, intelligent robots\nof the future are often expected to do the same. The existing continual\nlearning (CL) methods are usually not directly applicable to robotic settings:\nthey typically require buffering and a balanced replay of training data. A\nfew-shot online continual learning (FS-OCL) setting has been proposed to\naddress more realistic scenarios where robots must learn from a non-repeated\nsparse data stream. To enable truly autonomous life-long learning, an\nadditional challenge of detecting novelties and learning new items without\nsupervision needs to be addressed. We address this challenge with our new\nprototype-based approach called Continually Learning Prototypes (CLP). In\naddition to being capable of FS-OCL learning, CLP also detects novel objects\nand learns them without supervision. To mitigate forgetting, CLP utilizes a\nnovel metaplasticity mechanism that adapts the learning rate individually per\nprototype. CLP is rehearsal-free, hence does not require a memory buffer, and\nis compatible with neuromorphic hardware, characterized by ultra-low power\nconsumption, real-time processing abilities, and on-chip learning. Indeed, we\nhave open-sourced a simple version of CLP in the neuromorphic software\nframework Lava, targetting Intel's neuromorphic chip Loihi 2. We evaluate CLP\non a robotic vision dataset, OpenLORIS. In a low-instance FS-OCL scenario, CLP\nshows state-of-the-art results. In the open world, CLP detects novelties with\nsuperior precision and recall and learns features of the detected novel classes\nwithout supervision, achieving a strong baseline of 99% base class and 65%/76%\n(5-shot/10-shot) novel class accuracy.\n", "date": "2024-03-30", "categories": []}
{"id": "2404.00790", "title": "Rehearsal-Free Modular and Compositional Continual Learning for Language\n  Models", "abstract": "  Continual learning aims at incrementally acquiring new knowledge while not\nforgetting existing knowledge. To overcome catastrophic forgetting, methods are\neither rehearsal-based, i.e., store data examples from previous tasks for data\nreplay, or isolate parameters dedicated to each task. However, rehearsal-based\nmethods raise privacy and memory issues, and parameter-isolation continual\nlearning does not consider interaction between tasks, thus hindering knowledge\ntransfer. In this work, we propose MoCL, a rehearsal-free Modular and\nCompositional Continual Learning framework which continually adds new modules\nto language models and composes them with existing modules. Experiments on\nvarious benchmarks show that MoCL outperforms state of the art and effectively\nfacilitates knowledge transfer.\n", "date": "2024-03-31", "categories": []}
{"id": "2404.00901", "title": "Slightly Shift New Classes to Remember Old Classes for Video\n  Class-Incremental Learning", "abstract": "  Recent video class-incremental learning usually excessively pursues the\naccuracy of the newly seen classes and relies on memory sets to mitigate\ncatastrophic forgetting of the old classes. However, limited storage only\nallows storing a few representative videos. So we propose SNRO, which slightly\nshifts the features of new classes to remember old classes. Specifically, SNRO\ncontains Examples Sparse(ES) and Early Break(EB). ES decimates at a lower\nsample rate to build memory sets and uses interpolation to align those sparse\nframes in the future. By this, SNRO stores more examples under the same memory\nconsumption and forces the model to focus on low-semantic features which are\nharder to be forgotten. EB terminates the training at a small epoch, preventing\nthe model from overstretching into the high-semantic space of the current task.\nExperiments on UCF101, HMDB51, and UESTC-MMEA-CL datasets show that SNRO\nperforms better than other approaches while consuming the same memory\nconsumption.\n", "date": "2024-04-01", "categories": []}
{"id": "2404.00983", "title": "Continual Learning for Smart City: A Survey", "abstract": "  With the digitization of modern cities, large data volumes and powerful\ncomputational resources facilitate the rapid update of intelligent models\ndeployed in smart cities. Continual learning (CL) is a novel machine learning\nparadigm that constantly updates models to adapt to changing environments,\nwhere the learning tasks, data, and distributions can vary over time. Our\nsurvey provides a comprehensive review of continual learning methods that are\nwidely used in smart city development. The content consists of three parts: 1)\nMethodology-wise. We categorize a large number of basic CL methods and advanced\nCL frameworks in combination with other learning paradigms including graph\nlearning, spatial-temporal learning, multi-modal learning, and federated\nlearning. 2) Application-wise. We present numerous CL applications covering\ntransportation, environment, public health, safety, networks, and associated\ndatasets related to urban computing. 3) Challenges. We discuss current problems\nand challenges and envision several promising research directions. We believe\nthis survey can help relevant researchers quickly familiarize themselves with\nthe current state of continual learning research used in smart city development\nand direct them to future research trends.\n", "date": "2024-04-01", "categories": []}
{"id": "2403.20086", "title": "Selective Attention-based Modulation for Continual Learning", "abstract": "  We present SAM, a biologically-plausible selective attention-driven\nmodulation approach to enhance classification models in a continual learning\nsetting. Inspired by neurophysiological evidence that the primary visual cortex\ndoes not contribute to object manifold untangling for categorization and that\nprimordial attention biases are still embedded in the modern brain, we propose\nto employ auxiliary saliency prediction features as a modulation signal to\ndrive and stabilize the learning of a sequence of non-i.i.d. classification\ntasks. Experimental results confirm that SAM effectively enhances the\nperformance (in some cases up to about twenty percent points) of\nstate-of-the-art continual learning methods, both in class-incremental and\ntask-incremental settings. Moreover, we show that attention-based modulation\nsuccessfully encourages the learning of features that are more robust to the\npresence of spurious features and to adversarial attacks than baseline methods.\nCode is available at: https://github.com/perceivelab/SAM.\n", "date": "2024-03-29", "categories": []}
{"id": "2403.20126", "title": "ECLIPSE: Efficient Continual Learning in Panoptic Segmentation with\n  Visual Prompt Tuning", "abstract": "  Panoptic segmentation, combining semantic and instance segmentation, stands\nas a cutting-edge computer vision task. Despite recent progress with deep\nlearning models, the dynamic nature of real-world applications necessitates\ncontinual learning, where models adapt to new classes (plasticity) over time\nwithout forgetting old ones (catastrophic forgetting). Current continual\nsegmentation methods often rely on distillation strategies like knowledge\ndistillation and pseudo-labeling, which are effective but result in increased\ntraining complexity and computational overhead. In this paper, we introduce a\nnovel and efficient method for continual panoptic segmentation based on Visual\nPrompt Tuning, dubbed ECLIPSE. Our approach involves freezing the base model\nparameters and fine-tuning only a small set of prompt embeddings, addressing\nboth catastrophic forgetting and plasticity and significantly reducing the\ntrainable parameters. To mitigate inherent challenges such as error propagation\nand semantic drift in continual segmentation, we propose logit manipulation to\neffectively leverage common knowledge across the classes. Experiments on ADE20K\ncontinual panoptic segmentation benchmark demonstrate the superiority of\nECLIPSE, notably its robustness against catastrophic forgetting and its\nreasonable plasticity, achieving a new state-of-the-art. The code is available\nat https://github.com/clovaai/ECLIPSE.\n", "date": "2024-03-29", "categories": []}
{"id": "2403.20317", "title": "Convolutional Prompting meets Language Models for Continual Learning", "abstract": "  Continual Learning (CL) enables machine learning models to learn from\ncontinuously shifting new training data in absence of data from old tasks.\nRecently, pretrained vision transformers combined with prompt tuning have shown\npromise for overcoming catastrophic forgetting in CL. These approaches rely on\na pool of learnable prompts which can be inefficient in sharing knowledge\nacross tasks leading to inferior performance. In addition, the lack of\nfine-grained layer specific prompts does not allow these to fully express the\nstrength of the prompts for CL. We address these limitations by proposing\nConvPrompt, a novel convolutional prompt creation mechanism that maintains\nlayer-wise shared embeddings, enabling both layer-specific learning and better\nconcept transfer across tasks. The intelligent use of convolution enables us to\nmaintain a low parameter overhead without compromising performance. We further\nleverage Large Language Models to generate fine-grained text descriptions of\neach category which are used to get task similarity and dynamically decide the\nnumber of prompts to be learned. Extensive experiments demonstrate the\nsuperiority of ConvPrompt and improves SOTA by ~3% with significantly less\nparameter overhead. We also perform strong ablation over various modules to\ndisentangle the importance of different components.\n", "date": "2024-03-29", "categories": []}
{"id": "2212.08251", "title": "Task-Adaptive Saliency Guidance for Exemplar-free Class Incremental\n  Learning", "abstract": "  Exemplar-free Class Incremental Learning (EFCIL) aims to sequentially learn\ntasks with access only to data from the current one. EFCIL is of interest\nbecause it mitigates concerns about privacy and long-term storage of data,\nwhile at the same time alleviating the problem of catastrophic forgetting in\nincremental learning. In this work, we introduce task-adaptive saliency for\nEFCIL and propose a new framework, which we call Task-Adaptive Saliency\nSupervision (TASS), for mitigating the negative effects of saliency drift\nbetween different tasks. We first apply boundary-guided saliency to maintain\ntask adaptivity and \\textit{plasticity} on model attention. Besides, we\nintroduce task-agnostic low-level signals as auxiliary supervision to increase\nthe \\textit{stability} of model attention. Finally, we introduce a module for\ninjecting and recovering saliency noise to increase the robustness of saliency\npreservation. Our experiments demonstrate that our method can better preserve\nsaliency maps across tasks and achieve state-of-the-art results on the\nCIFAR-100, Tiny-ImageNet, and ImageNet-Subset EFCIL benchmarks. Code is\navailable at \\url{https://github.com/scok30/tass}.\n", "date": "2022-12-16", "categories": []}
{"id": "2403.18132", "title": "Recommendation of data-free class-incremental learning algorithms by\n  simulating future data", "abstract": "  Class-incremental learning deals with sequential data streams composed of\nbatches of classes. Various algorithms have been proposed to address the\nchallenging case where samples from past classes cannot be stored. However,\nselecting an appropriate algorithm for a user-defined setting is an open\nproblem, as the relative performance of these algorithms depends on the\nincremental settings. To solve this problem, we introduce an algorithm\nrecommendation method that simulates the future data stream. Given an initial\nset of classes, it leverages generative models to simulate future classes from\nthe same visual domain. We evaluate recent algorithms on the simulated stream\nand recommend the one which performs best in the user-defined incremental\nsetting. We illustrate the effectiveness of our method on three large datasets\nusing six algorithms and six incremental settings. Our method outperforms\ncompetitive baselines, and performance is close to that of an oracle choosing\nthe best algorithm in each setting. This work contributes to facilitate the\npractical deployment of incremental learning.\n", "date": "2024-03-26", "categories": []}
{"id": "2403.18258", "title": "Enhancing Generative Class Incremental Learning Performance with Model\n  Forgetting Approach", "abstract": "  This study presents a novel approach to Generative Class Incremental Learning\n(GCIL) by introducing the forgetting mechanism, aimed at dynamically managing\nclass information for better adaptation to streaming data. GCIL is one of the\nhot topics in the field of computer vision, and this is considered one of the\ncrucial tasks in society, specifically the continual learning of generative\nmodels. The ability to forget is a crucial brain function that facilitates\ncontinual learning by selectively discarding less relevant information for\nhumans. However, in the field of machine learning models, the concept of\nintentionally forgetting has not been extensively investigated. In this study\nwe aim to bridge this gap by incorporating the forgetting mechanisms into GCIL,\nthereby examining their impact on the models' ability to learn in continual\nlearning. Through our experiments, we have found that integrating the\nforgetting mechanisms significantly enhances the models' performance in\nacquiring new knowledge, underscoring the positive role that strategic\nforgetting plays in the process of continual learning.\n", "date": "2024-03-27", "categories": []}
{"id": "2403.18266", "title": "Branch-Tuning: Balancing Stability and Plasticity for Continual\n  Self-Supervised Learning", "abstract": "  Self-supervised learning (SSL) has emerged as an effective paradigm for\nderiving general representations from vast amounts of unlabeled data. However,\nas real-world applications continually integrate new content, the high\ncomputational and resource demands of SSL necessitate continual learning rather\nthan complete retraining. This poses a challenge in striking a balance between\nstability and plasticity when adapting to new information. In this paper, we\nemploy Centered Kernel Alignment for quantitatively analyzing model stability\nand plasticity, revealing the critical roles of batch normalization layers for\nstability and convolutional layers for plasticity. Motivated by this, we\npropose Branch-tuning, an efficient and straightforward method that achieves a\nbalance between stability and plasticity in continual SSL. Branch-tuning\nconsists of branch expansion and compression, and can be easily applied to\nvarious SSL methods without the need of modifying the original methods,\nretaining old data or models. We validate our method through incremental\nexperiments on various benchmark datasets, demonstrating its effectiveness and\npractical value in real-world scenarios. We hope our work offers new insights\nfor future continual self-supervised learning research. The code will be made\npublicly available.\n", "date": "2024-03-27", "categories": []}
{"id": "2403.18291", "title": "Towards Non-Exemplar Semi-Supervised Class-Incremental Learning", "abstract": "  Deep neural networks perform remarkably well in close-world scenarios.\nHowever, novel classes emerged continually in real applications, making it\nnecessary to learn incrementally. Class-incremental learning (CIL) aims to\ngradually recognize new classes while maintaining the discriminability of old\nones. Existing CIL methods have two limitations: a heavy reliance on preserving\nold data for forgetting mitigation and the need for vast labeled data for\nknowledge adaptation. To overcome these issues, we propose a non-exemplar\nsemi-supervised CIL framework with contrastive learning and semi-supervised\nincremental prototype classifier (Semi-IPC). On the one hand, contrastive\nlearning helps the model learn rich representations, easing the trade-off\nbetween learning representations of new classes and forgetting that of old\nclasses. On the other hand, Semi-IPC learns a prototype for each class with\nunsupervised regularization, enabling the model to incrementally learn from\npartially labeled new data while maintaining the knowledge of old classes.\nExperiments on benchmark datasets demonstrate the strong performance of our\nmethod: without storing any old samples and only using less than 1% of labels,\nSemi-IPC outperforms advanced exemplar-based methods. We hope our work offers\nnew insights for future CIL research. The code will be made publicly available.\n", "date": "2024-03-27", "categories": []}
{"id": "2403.18383", "title": "Generative Multi-modal Models are Good Class-Incremental Learners", "abstract": "  In class-incremental learning (CIL) scenarios, the phenomenon of catastrophic\nforgetting caused by the classifier's bias towards the current task has long\nposed a significant challenge. It is mainly caused by the characteristic of\ndiscriminative models. With the growing popularity of the generative\nmulti-modal models, we would explore replacing discriminative models with\ngenerative ones for CIL. However, transitioning from discriminative to\ngenerative models requires addressing two key challenges. The primary challenge\nlies in transferring the generated textual information into the classification\nof distinct categories. Additionally, it requires formulating the task of CIL\nwithin a generative framework. To this end, we propose a novel generative\nmulti-modal model (GMM) framework for class-incremental learning. Our approach\ndirectly generates labels for images using an adapted generative model. After\nobtaining the detailed text, we use a text encoder to extract text features and\nemploy feature matching to determine the most similar label as the\nclassification prediction. In the conventional CIL settings, we achieve\nsignificantly better results in long-sequence task scenarios. Under the\nFew-shot CIL setting, we have improved by at least 14\\% accuracy over all the\ncurrent state-of-the-art methods with significantly less forgetting. Our code\nis available at \\url{https://github.com/DoubleClass/GMM}.\n", "date": "2024-03-27", "categories": []}
{"id": "2403.18550", "title": "OrCo: Towards Better Generalization via Orthogonality and Contrast for\n  Few-Shot Class-Incremental Learning", "abstract": "  Few-Shot Class-Incremental Learning (FSCIL) introduces a paradigm in which\nthe problem space expands with limited data. FSCIL methods inherently face the\nchallenge of catastrophic forgetting as data arrives incrementally, making\nmodels susceptible to overwriting previously acquired knowledge. Moreover,\ngiven the scarcity of labeled samples available at any given time, models may\nbe prone to overfitting and find it challenging to strike a balance between\nextensive pretraining and the limited incremental data. To address these\nchallenges, we propose the OrCo framework built on two core principles:\nfeatures' orthogonality in the representation space, and contrastive learning.\nIn particular, we improve the generalization of the embedding space by\nemploying a combination of supervised and self-supervised contrastive losses\nduring the pretraining phase. Additionally, we introduce OrCo loss to address\nchallenges arising from data limitations during incremental sessions. Through\nfeature space perturbations and orthogonality between classes, the OrCo loss\nmaximizes margins and reserves space for the following incremental data. This,\nin turn, ensures the accommodation of incoming classes in the feature space\nwithout compromising previously acquired knowledge. Our experimental results\nshowcase state-of-the-art performance across three benchmark datasets,\nincluding mini-ImageNet, CIFAR100, and CUB datasets. Code is available at\nhttps://github.com/noorahmedds/OrCo\n", "date": "2024-03-27", "categories": []}
{"id": "2310.15694", "title": "COPR: Continual Learning Human Preference through Optimal Policy\n  Regularization", "abstract": "  The technique of Reinforcement Learning from Human Feedback (RLHF) is a\ncommonly employed method to improve pre-trained Language Models (LM), enhancing\ntheir ability to conform to human preferences. Nevertheless, the current\nRLHF-based LMs necessitate full retraining each time novel queries or feedback\nare introduced, which becomes a challenging task because human preferences can\nvary between different domains or tasks. Retraining LMs poses practical\ndifficulties in many real-world situations due to the significant time and\ncomputational resources required, along with concerns related to data privacy.\nTo address this limitation, we propose a new method called Continual Optimal\nPolicy Regularization (COPR), in which we compute the distribution of optimal\npolicy bypassing the partition function and then regularize the current policy\nbased on the historically optimal distribution to mitigate Catastrophic\nForgetting (CF). COPR involves a single learning phase and doesn't necessitate\ncomplex reinforcement learning. Importantly, it shares the capability with RLHF\nto learn from unlabeled data by maintaining a scoring module, similar to reward\nmodel, making it flexible for continually learning without human feedback. Our\nexperimental results show that COPR outperforms strong Continuous Learning (CL)\nbaselines when it comes to consistently aligning with human preferences on\nincremental tasks and domains.\n", "date": "2023-10-24", "categories": []}
{"id": "2403.17503", "title": "DS-AL: A Dual-Stream Analytic Learning for Exemplar-Free\n  Class-Incremental Learning", "abstract": "  Class-incremental learning (CIL) under an exemplar-free constraint has\npresented a significant challenge. Existing methods adhering to this constraint\nare prone to catastrophic forgetting, far more so than replay-based techniques\nthat retain access to past samples. In this paper, to solve the exemplar-free\nCIL problem, we propose a Dual-Stream Analytic Learning (DS-AL) approach. The\nDS-AL contains a main stream offering an analytical (i.e., closed-form) linear\nsolution, and a compensation stream improving the inherent under-fitting\nlimitation due to adopting linear mapping. The main stream redefines the CIL\nproblem into a Concatenated Recursive Least Squares (C-RLS) task, allowing an\nequivalence between the CIL and its joint-learning counterpart. The\ncompensation stream is governed by a Dual-Activation Compensation (DAC) module.\nThis module re-activates the embedding with a different activation function\nfrom the main stream one, and seeks fitting compensation by projecting the\nembedding to the null space of the main stream's linear mapping. Empirical\nresults demonstrate that the DS-AL, despite being an exemplar-free technique,\ndelivers performance comparable with or better than that of replay-based\nmethods across various datasets, including CIFAR-100, ImageNet-100 and\nImageNet-Full. Additionally, the C-RLS' equivalent property allows the DS-AL to\nexecute CIL in a phase-invariant manner. This is evidenced by a\nnever-before-seen 500-phase CIL ImageNet task, which performs on a level\nidentical to a 5-phase one. Our codes are available at\nhttps://github.com/ZHUANGHP/Analytic-continual-learning.\n", "date": "2024-03-26", "categories": []}
{"id": "2403.17733", "title": "Continual Few-shot Event Detection via Hierarchical Augmentation\n  Networks", "abstract": "  Traditional continual event detection relies on abundant labeled data for\ntraining, which is often impractical to obtain in real-world applications. In\nthis paper, we introduce continual few-shot event detection (CFED), a more\ncommonly encountered scenario when a substantial number of labeled samples are\nnot accessible. The CFED task is challenging as it involves memorizing previous\nevent types and learning new event types with few-shot samples. To mitigate\nthese challenges, we propose a memory-based framework: Hierarchical\nAugmentation Networks (HANet). To memorize previous event types with limited\nmemory, we incorporate prototypical augmentation into the memory set. For the\nissue of learning new event types in few-shot scenarios, we propose a\ncontrastive augmentation module for token representations. Despite comparing\nwith previous state-of-the-art methods, we also conduct comparisons with\nChatGPT. Experiment results demonstrate that our method significantly\noutperforms all of these methods in multiple continual few-shot event detection\ntasks.\n", "date": "2024-03-26", "categories": []}
{"id": "2304.11959", "title": "A Forward and Backward Compatible Framework for Few-shot\n  Class-incremental Pill Recognition", "abstract": "  Automatic Pill Recognition (APR) systems are crucial for enhancing hospital\nefficiency, assisting visually impaired individuals, and preventing\ncross-infection. However, most existing deep learning-based pill recognition\nsystems can only perform classification on classes with sufficient training\ndata. In practice, the high cost of data annotation and the continuous increase\nin new pill classes necessitate the development of a few-shot class-incremental\npill recognition system. This paper introduces the first few-shot\nclass-incremental pill recognition framework, named Discriminative and\nBidirectional Compatible Few-Shot Class-Incremental Learning (DBC-FSCIL). It\nencompasses forward-compatible and backward-compatible learning components. In\nforward-compatible learning, we propose an innovative virtual class synthesis\nstrategy and a Center-Triplet (CT) loss to enhance discriminative feature\nlearning. These virtual classes serve as placeholders in the feature space for\nfuture class updates, providing diverse semantic knowledge for model training.\nFor backward-compatible learning, we develop a strategy to synthesize reliable\npseudo-features of old classes using uncertainty quantification, facilitating\nData Replay (DR) and Knowledge Distillation (KD). This approach allows for the\nflexible synthesis of features and effectively reduces additional storage\nrequirements for samples and models. Additionally, we construct a new pill\nimage dataset for FSCIL and assess various mainstream FSCIL methods,\nestablishing new benchmarks. Our experimental results demonstrate that our\nframework surpasses existing State-of-the-art (SOTA) methods. The code is\navailable at https://github.com/zhang-jinghua/DBC-FSCIL.\n", "date": "2023-04-24", "categories": []}
{"id": "2403.16124", "title": "Enhancing Visual Continual Learning with Language-Guided Supervision", "abstract": "  Continual learning (CL) aims to empower models to learn new tasks without\nforgetting previously acquired knowledge. Most prior works concentrate on the\ntechniques of architectures, replay data, regularization, \\etc. However, the\ncategory name of each class is largely neglected. Existing methods commonly\nutilize the one-hot labels and randomly initialize the classifier head. We\nargue that the scarce semantic information conveyed by the one-hot labels\nhampers the effective knowledge transfer across tasks. In this paper, we\nrevisit the role of the classifier head within the CL paradigm and replace the\nclassifier with semantic knowledge from pretrained language models (PLMs).\nSpecifically, we use PLMs to generate semantic targets for each class, which\nare frozen and serve as supervision signals during training. Such targets fully\nconsider the semantic correlation between all classes across tasks. Empirical\nstudies show that our approach mitigates forgetting by alleviating\nrepresentation drifting and facilitating knowledge transfer across tasks. The\nproposed method is simple to implement and can seamlessly be plugged into\nexisting methods with negligible adjustments. Extensive experiments based on\neleven mainstream baselines demonstrate the effectiveness and generalizability\nof our approach to various protocols. For example, under the class-incremental\nlearning setting on ImageNet-100, our method significantly improves the Top-1\naccuracy by 3.2\\% to 6.1\\% while reducing the forgetting rate by 2.6\\% to\n13.1\\%.\n", "date": "2024-03-24", "categories": []}
{"id": "2403.16221", "title": "Exemplar-Free Class Incremental Learning via Incremental Representation", "abstract": "  Exemplar-Free Class Incremental Learning (efCIL) aims to continuously\nincorporate the knowledge from new classes while retaining previously learned\ninformation, without storing any old-class exemplars (i.e., samples). For this\npurpose, various efCIL methods have been proposed over the past few years,\ngenerally with elaborately constructed old pseudo-features, increasing the\ndifficulty of model development and interpretation. In contrast, we propose a\n\\textbf{simple Incremental Representation (IR) framework} for efCIL without\nconstructing old pseudo-features. IR utilizes dataset augmentation to cover a\nsuitable feature space and prevents the model from forgetting by using a single\nL2 space maintenance loss. We discard the transient classifier trained on each\none of the sequence tasks and instead replace it with a 1-near-neighbor\nclassifier for inference, ensuring the representation is incrementally updated\nduring CIL. Extensive experiments demonstrate that our proposed IR achieves\ncomparable performance while significantly preventing the model from forgetting\non CIFAR100, TinyImageNet, and ImageNetSubset datasets.\n", "date": "2024-03-24", "categories": []}
{"id": "2312.06710", "title": "Class-Prototype Conditional Diffusion Model with Gradient Projection for\n  Continual Learning", "abstract": "  Mitigating catastrophic forgetting is a key hurdle in continual learning.\nDeep Generative Replay (GR) provides techniques focused on generating samples\nfrom prior tasks to enhance the model's memory capabilities using generative AI\nmodels ranging from Generative Adversarial Networks (GANs) to the more recent\nDiffusion Models (DMs). A major issue is the deterioration in the quality of\ngenerated data compared to the original, as the generator continuously\nself-learns from its outputs. This degradation can lead to the potential risk\nof catastrophic forgetting (CF) occurring in the classifier. To address this,\nwe propose the Gradient Projection Class-Prototype Conditional Diffusion Model\n(GPPDM), a GR-based approach for continual learning that enhances image quality\nin generators and thus reduces the CF in classifiers. The cornerstone of GPPDM\nis a learnable class prototype that captures the core characteristics of images\nin a given class. This prototype, integrated into the diffusion model's\ndenoising process, ensures the generation of high-quality images of the old\ntasks, hence reducing the risk of CF in classifiers. Moreover, to further\nmitigate the CF of diffusion models, we propose a gradient projection technique\ntailored for the cross-attention layer of diffusion models to maximally\nmaintain and preserve the representations of old task data in the current task\nas close as possible to their representations when they first arrived. Our\nempirical studies on diverse datasets demonstrate that our proposed method\nsignificantly outperforms existing state-of-the-art models, highlighting its\nsatisfactory ability to preserve image quality and enhance the model's memory\nretention.\n", "date": "2023-12-10", "categories": []}
{"id": "2403.14679", "title": "Continual Learning by Three-Phase Consolidation", "abstract": "  TPC (Three-Phase Consolidation) is here introduced as a simple but effective\napproach to continually learn new classes (and/or instances of known classes)\nwhile controlling forgetting of previous knowledge. Each experience (a.k.a.\ntask) is learned in three phases characterized by different rules and learning\ndynamics, aimed at removing the class-bias problem (due to class unbalancing)\nand limiting gradient-based corrections to prevent forgetting of\nunderrepresented classes. Several experiments on complex datasets demonstrate\nits accuracy and efficiency advantages over competitive existing approaches.\nThe algorithm and all the results presented in this paper are fully\nreproducible thanks to its publication on the Avalanche open framework for\ncontinual learning.\n", "date": "2024-03-12", "categories": []}
{"id": "2403.14910", "title": "Defying Imbalanced Forgetting in Class Incremental Learning", "abstract": "  We observe a high level of imbalance in the accuracy of different classes in\nthe same old task for the first time. This intriguing phenomenon, discovered in\nreplay-based Class Incremental Learning (CIL), highlights the imbalanced\nforgetting of learned classes, as their accuracy is similar before the\noccurrence of catastrophic forgetting. This discovery remains previously\nunidentified due to the reliance on average incremental accuracy as the\nmeasurement for CIL, which assumes that the accuracy of classes within the same\ntask is similar. However, this assumption is invalid in the face of\ncatastrophic forgetting. Further empirical studies indicate that this\nimbalanced forgetting is caused by conflicts in representation between\nsemantically similar old and new classes. These conflicts are rooted in the\ndata imbalance present in replay-based CIL methods. Building on these insights,\nwe propose CLass-Aware Disentanglement (CLAD) to predict the old classes that\nare more likely to be forgotten and enhance their accuracy. Importantly, CLAD\ncan be seamlessly integrated into existing CIL methods. Extensive experiments\ndemonstrate that CLAD consistently improves current replay-based methods,\nresulting in performance gains of up to 2.56%.\n", "date": "2024-03-22", "categories": []}
{"id": "2310.16226", "title": "TiC-CLIP: Continual Training of CLIP Models", "abstract": "  Keeping large foundation models up to date on latest data is inherently\nexpensive. To avoid the prohibitive costs of constantly retraining, it is\nimperative to continually train these models. This problem is exacerbated by\nthe lack of any large scale continual learning benchmarks or baselines. We\nintroduce the first set of web-scale Time-Continual (TiC) benchmarks for\ntraining vision-language models: TiC-DataComp, TiC-YFCC, and TiC-Redcaps.\nTiC-DataComp, our largest dataset, contains over 12.7B timestamped image-text\npairs spanning 9 years (2014-2022). We first use our benchmarks to curate\nvarious dynamic evaluations to measure temporal robustness of existing models.\nWe show OpenAI's CLIP (trained on data up to 2020) loses $\\approx 8\\%$\nzero-shot accuracy on our curated retrieval task from 2021-2022 compared with\nmore recently trained models in OpenCLIP repository. We then study how to\nefficiently train models on time-continuous data. We demonstrate that a simple\nrehearsal-based approach that continues training from the last checkpoint and\nreplays old data reduces compute by $2.5\\times$ when compared to the standard\npractice of retraining from scratch. Code is available at\nhttps://github.com/apple/ml-tic-clip.\n", "date": "2023-10-24", "categories": []}
{"id": "2401.07208", "title": "Enhanced Few-Shot Class-Incremental Learning via Ensemble Models", "abstract": "  Few-shot class-incremental learning (FSCIL) aims to continually fit new\nclasses with limited training data, while maintaining the performance of\npreviously learned classes. The main challenges are overfitting the rare new\ntraining samples and forgetting old classes. While catastrophic forgetting has\nbeen extensively studied, the overfitting problem has attracted less attention\nin FSCIL. To tackle overfitting challenge, we design a new ensemble model\nframework cooperated with data augmentation to boost generalization. In this\nway, the enhanced model works as a library storing abundant features to\nguarantee fast adaptation to downstream tasks. Specifically, the multi-input\nmulti-output ensemble structure is applied with a spatial-aware data\naugmentation strategy, aiming at diversifying the feature extractor and\nalleviating overfitting in incremental sessions. Moreover, self-supervised\nlearning is also integrated to further improve the model generalization.\nComprehensive experimental results show that the proposed method can indeed\nmitigate the overfitting problem in FSCIL, and outperform the state-of-the-art\nmethods.\n", "date": "2024-01-14", "categories": []}
{"id": "2403.13845", "title": "Learning to better see the unseen: Broad-Deep Mixed Anti-Forgetting\n  Framework for Incremental Zero-Shot Fault Diagnosis", "abstract": "  Zero-shot fault diagnosis (ZSFD) is capable of identifying unseen faults via\npredicting fault attributes labeled by human experts. We first recognize the\ndemand of ZSFD to deal with continuous changes in industrial processes, i.e.,\nthe model's ability to adapt to new fault categories and attributes while\navoiding forgetting the diagnosis ability learned previously. To overcome the\nissue that the existing ZSFD paradigm cannot learn from evolving streams of\ntraining data in industrial scenarios, the incremental ZSFD (IZSFD) paradigm is\nproposed for the first time, which incorporates category increment and\nattribute increment for both traditional ZSFD and generalized ZSFD paradigms.\nTo achieve IZSFD, we present a broad-deep mixed anti-forgetting framework\n(BDMAFF) that aims to learn from new fault categories and attributes. To tackle\nthe issue of forgetting, BDMAFF effectively accumulates previously acquired\nknowledge from two perspectives: features and attribute prototypes. The feature\nmemory is established through a deep generative model that employs\nanti-forgetting training strategies, ensuring the generation quality of\nhistorical categories is supervised and maintained. The diagnosis model SEEs\nthe UNSEEN faults with the help of generated samples from the generative model.\nThe attribute prototype memory is established through a diagnosis model\ninspired by the broad learning system. Unlike traditional incremental learning\nalgorithms, BDMAFF introduces a memory-driven iterative update strategy for the\ndiagnosis model, which allows the model to learn new faults and attributes\nwithout requiring the storage of all historical training samples. The\neffectiveness of the proposed method is verified by a real hydraulic system and\nthe Tennessee-Eastman benchmark process.\n", "date": "2024-03-18", "categories": []}
{"id": "2403.14101", "title": "Text-Enhanced Data-free Approach for Federated Class-Incremental\n  Learning", "abstract": "  Federated Class-Incremental Learning (FCIL) is an underexplored yet pivotal\nissue, involving the dynamic addition of new classes in the context of\nfederated learning. In this field, Data-Free Knowledge Transfer (DFKT) plays a\ncrucial role in addressing catastrophic forgetting and data privacy problems.\nHowever, prior approaches lack the crucial synergy between DFKT and the model\ntraining phases, causing DFKT to encounter difficulties in generating\nhigh-quality data from a non-anchored latent space of the old task model. In\nthis paper, we introduce LANDER (Label Text Centered Data-Free Knowledge\nTransfer) to address this issue by utilizing label text embeddings (LTE)\nproduced by pretrained language models. Specifically, during the model training\nphase, our approach treats LTE as anchor points and constrains the feature\nembeddings of corresponding training samples around them, enriching the\nsurrounding area with more meaningful information. In the DFKT phase, by using\nthese LTE anchors, LANDER can synthesize more meaningful samples, thereby\neffectively addressing the forgetting problem. Additionally, instead of tightly\nconstraining embeddings toward the anchor, the Bounding Loss is introduced to\nencourage sample embeddings to remain flexible within a defined radius. This\napproach preserves the natural differences in sample embeddings and mitigates\nthe embedding overlap caused by heterogeneous federated settings. Extensive\nexperiments conducted on CIFAR100, Tiny-ImageNet, and ImageNet demonstrate that\nLANDER significantly outperforms previous methods and achieves state-of-the-art\nperformance in FCIL. The code is available at\nhttps://github.com/tmtuan1307/lander.\n", "date": "2024-03-21", "categories": []}
{"id": "2305.11488", "title": "AttriCLIP: A Non-Incremental Learner for Incremental Knowledge Learning", "abstract": "  Continual learning aims to enable a model to incrementally learn knowledge\nfrom sequentially arrived data. Previous works adopt the conventional\nclassification architecture, which consists of a feature extractor and a\nclassifier. The feature extractor is shared across sequentially arrived tasks\nor classes, but one specific group of weights of the classifier corresponding\nto one new class should be incrementally expanded. Consequently, the parameters\nof a continual learner gradually increase. Moreover, as the classifier contains\nall historical arrived classes, a certain size of the memory is usually\nrequired to store rehearsal data to mitigate classifier bias and catastrophic\nforgetting. In this paper, we propose a non-incremental learner, named\nAttriCLIP, to incrementally extract knowledge of new classes or tasks.\nSpecifically, AttriCLIP is built upon the pre-trained visual-language model\nCLIP. Its image encoder and text encoder are fixed to extract features from\nboth images and text. Text consists of a category name and a fixed number of\nlearnable parameters which are selected from our designed attribute word bank\nand serve as attributes. As we compute the visual and textual similarity for\nclassification, AttriCLIP is a non-incremental learner. The attribute prompts,\nwhich encode the common knowledge useful for classification, can effectively\nmitigate the catastrophic forgetting and avoid constructing a replay memory. We\nevaluate our AttriCLIP and compare it with CLIP-based and previous\nstate-of-the-art continual learning methods in realistic settings with\ndomain-shift and long-sequence learning. The results show that our method\nperforms favorably against previous state-of-the-arts. The implementation code\ncan be available at https://github.com/bhrqw/AttriCLIP.\n", "date": "2023-05-19", "categories": []}
{"id": "2403.13130", "title": "Self-generated Replay Memories for Continual Neural Machine Translation", "abstract": "  Modern Neural Machine Translation systems exhibit strong performance in\nseveral different languages and are constantly improving. Their ability to\nlearn continuously is, however, still severely limited by the catastrophic\nforgetting issue. In this work, we leverage a key property of encoder-decoder\nTransformers, i.e. their generative ability, to propose a novel approach to\ncontinually learning Neural Machine Translation systems. We show how this can\neffectively learn on a stream of experiences comprising different languages, by\nleveraging a replay memory populated by using the model itself as a generator\nof parallel sentences. We empirically demonstrate that our approach can\ncounteract catastrophic forgetting without requiring explicit memorization of\ntraining data. Code will be publicly available upon publication. Code:\nhttps://github.com/m-resta/sg-rep\n", "date": "2024-03-19", "categories": []}
{"id": "2403.13249", "title": "A Unified and General Framework for Continual Learning", "abstract": "  Continual Learning (CL) focuses on learning from dynamic and changing data\ndistributions while retaining previously acquired knowledge. Various methods\nhave been developed to address the challenge of catastrophic forgetting,\nincluding regularization-based, Bayesian-based, and memory-replay-based\ntechniques. However, these methods lack a unified framework and common\nterminology for describing their approaches. This research aims to bridge this\ngap by introducing a comprehensive and overarching framework that encompasses\nand reconciles these existing methodologies. Notably, this new framework is\ncapable of encompassing established CL approaches as special instances within a\nunified and general optimization objective. An intriguing finding is that\ndespite their diverse origins, these methods share common mathematical\nstructures. This observation highlights the compatibility of these seemingly\ndistinct techniques, revealing their interconnectedness through a shared\nunderlying optimization objective. Moreover, the proposed general framework\nintroduces an innovative concept called refresh learning, specifically designed\nto enhance the CL performance. This novel approach draws inspiration from\nneuroscience, where the human brain often sheds outdated information to improve\nthe retention of crucial knowledge and facilitate the acquisition of new\ninformation. In essence, refresh learning operates by initially unlearning\ncurrent data and subsequently relearning it. It serves as a versatile plug-in\nthat seamlessly integrates with existing CL methods, offering an adaptable and\neffective enhancement to the learning process. Extensive experiments on CL\nbenchmarks and theoretical analysis demonstrate the effectiveness of the\nproposed refresh learning. Code is available at\n\\url{https://github.com/joey-wang123/CL-refresh-learning}.\n", "date": "2024-03-20", "categories": []}
{"id": "2310.18639", "title": "Towards Plastic and Stable Exemplar-Free Incremental Learning: A\n  Dual-Learner Framework with Cumulative Parameter Averaging", "abstract": "  The dilemma between plasticity and stability presents a significant challenge\nin Incremental Learning (IL), especially in the exemplar-free scenario where\naccessing old-task samples is strictly prohibited during the learning of a new\ntask. A straightforward solution to this issue is learning and storing an\nindependent model for each task, known as Single Task Learning (STL). Despite\nthe linear growth in model storage with the number of tasks in STL, we\nempirically discover that averaging these model parameters can potentially\npreserve knowledge across all tasks. Inspired by this observation, we propose a\nDual-Learner framework with Cumulative Parameter Averaging (DLCPA). DLCPA\nemploys a dual-learner design: a plastic learner focused on acquiring new-task\nknowledge and a stable learner responsible for accumulating all learned\nknowledge. The knowledge from the plastic learner is transferred to the stable\nlearner via cumulative parameter averaging. Additionally, several task-specific\nclassifiers work in cooperation with the stable learner to yield the final\nprediction. Specifically, when learning a new task, these modules are updated\nin a cyclic manner: i) the plastic learner is initially optimized using a\nself-supervised loss besides the supervised loss to enhance the feature\nextraction robustness; ii) the stable learner is then updated with respect to\nthe plastic learner in a cumulative parameter averaging manner to maintain its\ntask-wise generalization; iii) the task-specific classifier is accordingly\noptimized to align with the stable learner. Experimental results on CIFAR-100\nand Tiny-ImageNet show that DLCPA outperforms several state-of-the-art\nexemplar-free baselines in both Task-IL and Class-IL settings.\n", "date": "2023-10-28", "categories": []}
{"id": "2401.10191", "title": "Divide and not forget: Ensemble of selectively trained experts in\n  Continual Learning", "abstract": "  Class-incremental learning is becoming more popular as it helps models widen\ntheir applicability while not forgetting what they already know. A trend in\nthis area is to use a mixture-of-expert technique, where different models work\ntogether to solve the task. However, the experts are usually trained all at\nonce using whole task data, which makes them all prone to forgetting and\nincreasing computational burden. To address this limitation, we introduce a\nnovel approach named SEED. SEED selects only one, the most optimal expert for a\nconsidered task, and uses data from this task to fine-tune only this expert.\nFor this purpose, each expert represents each class with a Gaussian\ndistribution, and the optimal expert is selected based on the similarity of\nthose distributions. Consequently, SEED increases diversity and heterogeneity\nwithin the experts while maintaining the high stability of this ensemble\nmethod. The extensive experiments demonstrate that SEED achieves\nstate-of-the-art performance in exemplar-free settings across various\nscenarios, showing the potential of expert diversification through data in\ncontinual learning.\n", "date": "2024-01-18", "categories": []}
{"id": "2403.02628", "title": "Interactive Continual Learning: Fast and Slow Thinking", "abstract": "  Advanced life forms, sustained by the synergistic interaction of neural\ncognitive mechanisms, continually acquire and transfer knowledge throughout\ntheir lifespan. In contrast, contemporary machine learning paradigms exhibit\nlimitations in emulating the facets of continual learning (CL). Nonetheless,\nthe emergence of large language models (LLMs) presents promising avenues for\nrealizing CL via interactions with these models. Drawing on Complementary\nLearning System theory, this paper presents a novel Interactive Continual\nLearning (ICL) framework, enabled by collaborative interactions among models of\nvarious sizes. Specifically, we assign the ViT model as System1 and multimodal\nLLM as System2. To enable the memory module to deduce tasks from class\ninformation and enhance Set2Set retrieval, we propose the Class-Knowledge-Task\nMulti-Head Attention (CKT-MHA). Additionally, to improve memory retrieval in\nSystem1 through enhanced geometric representation, we introduce the CL-vMF\nmechanism, based on the von Mises-Fisher (vMF) distribution. Meanwhile, we\nintroduce the von Mises-Fisher Outlier Detection and Interaction (vMF-ODI)\nstrategy to identify hard examples, thus enhancing collaboration between\nSystem1 and System2 for complex reasoning realization. Comprehensive evaluation\nof our proposed ICL demonstrates significant resistance to forgetting and\nsuperior performance relative to existing methods. Code is available at\ngithub.com/ICL.\n", "date": "2024-03-05", "categories": []}
{"id": "2308.15236", "title": "Rotation Augmented Distillation for Exemplar-Free Class Incremental\n  Learning with Detailed Analysis", "abstract": "  Class incremental learning (CIL) aims to recognize both the old and new\nclasses along the increment tasks. Deep neural networks in CIL suffer from\ncatastrophic forgetting and some approaches rely on saving exemplars from\nprevious tasks, known as the exemplar-based setting, to alleviate this problem.\nOn the contrary, this paper focuses on the Exemplar-Free setting with no old\nclass sample preserved. Balancing the plasticity and stability in deep feature\nlearning with only supervision from new classes is more challenging. Most\nexisting Exemplar-Free CIL methods report the overall performance only and lack\nfurther analysis. In this work, different methods are examined with\ncomplementary metrics in greater detail. Moreover, we propose a simple CIL\nmethod, Rotation Augmented Distillation (RAD), which achieves one of the\ntop-tier performances under the Exemplar-Free setting. Detailed analysis shows\nour RAD benefits from the superior balance between plasticity and stability.\nFinally, more challenging exemplar-free settings with fewer initial classes are\nundertaken for further demonstrations and comparisons among the\nstate-of-the-art methods.\n", "date": "2023-08-29", "categories": []}
{"id": "2403.10022", "title": "Lifelong Person Re-Identification with Backward-Compatibility", "abstract": "  Lifelong person re-identification (LReID) assumes a practical scenario where\nthe model is sequentially trained on continuously incoming datasets while\nalleviating the catastrophic forgetting in the old datasets. However, not only\nthe training datasets but also the gallery images are incrementally\naccumulated, that requires a huge amount of computational complexity and\nstorage space to extract the features at the inference phase. In this paper, we\naddress the above mentioned problem by incorporating the backward-compatibility\nto LReID for the first time. We train the model using the continuously incoming\ndatasets while maintaining the model's compatibility toward the previously\ntrained old models without re-computing the features of the old gallery images.\nTo this end, we devise the cross-model compatibility loss based on the\ncontrastive learning with respect to the replay features across all the old\ndatasets. Moreover, we also develop the knowledge consolidation method based on\nthe part classification to learn the shared representation across different\ndatasets for the backward-compatibility. We suggest a more practical\nmethodology for performance evaluation as well where all the gallery and query\nimages are considered together. Experimental results demonstrate that the\nproposed method achieves a significantly higher performance of the\nbackward-compatibility compared with the existing methods. It is a promising\ntool for more practical scenarios of LReID.\n", "date": "2024-03-15", "categories": []}
{"id": "2403.10894", "title": "Towards Robustness and Diversity: Continual Learning in Dialog\n  Generation with Text-Mixup and Batch Nuclear-Norm Maximization", "abstract": "  In our dynamic world where data arrives in a continuous stream, continual\nlearning enables us to incrementally add new tasks/domains without the need to\nretrain from scratch. A major challenge in continual learning of language model\nis catastrophic forgetting, the tendency of models to forget knowledge from\npreviously trained tasks/domains when training on new ones. This paper studies\ndialog generation under the continual learning setting. We propose a novel\nmethod that 1) uses \\textit{Text-Mixup} as data augmentation to avoid model\noverfitting on replay memory and 2) leverages Batch-Nuclear Norm Maximization\n(BNNM) to alleviate the problem of mode collapse. Experiments on a $37$-domain\ntask-oriented dialog dataset and DailyDialog (a $10$-domain chitchat dataset)\ndemonstrate that our proposed approach outperforms the state-of-the-art in\ncontinual learning.\n", "date": "2024-03-16", "categories": []}
{"id": "2403.10929", "title": "Function-space Parameterization of Neural Networks for Sequential\n  Learning", "abstract": "  Sequential learning paradigms pose challenges for gradient-based deep\nlearning due to difficulties incorporating new data and retaining prior\nknowledge. While Gaussian processes elegantly tackle these problems, they\nstruggle with scalability and handling rich inputs, such as images. To address\nthese issues, we introduce a technique that converts neural networks from\nweight space to function space, through a dual parameterization. Our\nparameterization offers: (i) a way to scale function-space methods to large\ndata sets via sparsification, (ii) retention of prior knowledge when access to\npast data is limited, and (iii) a mechanism to incorporate new data without\nretraining. Our experiments demonstrate that we can retain knowledge in\ncontinual learning and incorporate new data efficiently. We further show its\nstrengths in uncertainty quantification and guiding exploration in model-based\nRL. Further information and code is available on the project website.\n", "date": "2024-03-16", "categories": []}
{"id": "2403.11070", "title": "Controllable Relation Disentanglement for Few-Shot Class-Incremental\n  Learning", "abstract": "  In this paper, we propose to tackle Few-Shot Class-Incremental Learning\n(FSCIL) from a new perspective, i.e., relation disentanglement, which means\nenhancing FSCIL via disentangling spurious relation between categories. The\nchallenge of disentangling spurious correlations lies in the poor\ncontrollability of FSCIL. On one hand, an FSCIL model is required to be trained\nin an incremental manner and thus it is very hard to directly control\nrelationships between categories of different sessions. On the other hand,\ntraining samples per novel category are only in the few-shot setting, which\nincreases the difficulty of alleviating spurious relation issues as well. To\novercome this challenge, in this paper, we propose a new simple-yet-effective\nmethod, called ConTrollable Relation-disentangLed Few-Shot Class-Incremental\nLearning (CTRL-FSCIL). Specifically, during the base session, we propose to\nanchor base category embeddings in feature space and construct disentanglement\nproxies to bridge gaps between the learning for category representations in\ndifferent sessions, thereby making category relation controllable. During\nincremental learning, the parameters of the backbone network are frozen in\norder to relieve the negative impact of data scarcity. Moreover, a\ndisentanglement loss is designed to effectively guide a relation\ndisentanglement controller to disentangle spurious correlations between the\nembeddings encoded by the backbone. In this way, the spurious correlation issue\nin FSCIL can be suppressed. Extensive experiments on CIFAR-100, mini-ImageNet,\nand CUB-200 datasets demonstrate the effectiveness of our CTRL-FSCIL method.\n", "date": "2024-03-17", "categories": []}
{"id": "2403.11373", "title": "Reconstruct before Query: Continual Missing Modality Learning with\n  Decomposed Prompt Collaboration", "abstract": "  Pre-trained large multi-modal models (LMMs) exploit fine-tuning to adapt\ndiverse user applications. Nevertheless, fine-tuning may face challenges due to\ndeactivated sensors (e.g., cameras turned off for privacy or technical issues),\nyielding modality-incomplete data and leading to inconsistency in training data\nand the data for inference. Additionally, continuous training leads to\ncatastrophic forgetting, diluting the knowledge in pre-trained LMMs. To\novercome these challenges, we introduce a novel task, Continual Missing\nModality Learning (CMML), to investigate how models can generalize when data of\ncertain modalities is missing during continual fine-tuning. Our preliminary\nbenchmarks reveal that existing methods suffer from a significant performance\ndrop in CMML, even with the aid of advanced continual learning techniques.\nTherefore, we devise a framework termed Reconstruct before Query (RebQ). It\ndecomposes prompts into modality-specific ones and breaks them into components\nstored in pools accessible via a key-query mechanism, which facilitates\nParameterEfficient Fine-Tuning and enhances knowledge transferability for\nsubsequent tasks. Meanwhile, our RebQ leverages extensive multi-modal knowledge\nfrom pre-trained LMMs to reconstruct the data of missing modality.\nComprehensive experiments demonstrate that RebQ effectively reconstructs the\nmissing modality information and retains pre-trained knowledge. Specifically,\ncompared with the baseline, RebQ improves average precision from 20.00 to 50.92\nand decreases average forgetting from 75.95 to 8.56. Code and datasets are\navailable on https://github.com/Tree-Shu-Zhao/RebQ.pytorch\n", "date": "2024-03-17", "categories": []}
{"id": "2403.11435", "title": "InsCL: A Data-efficient Continual Learning Paradigm for Fine-tuning\n  Large Language Models with Instructions", "abstract": "  Instruction tuning effectively optimizes Large Language Models (LLMs) for\ndownstream tasks. Due to the changing environment in real-life applications,\nLLMs necessitate continual task-specific adaptation without catastrophic\nforgetting. Considering the heavy computational cost, replay-based Continual\nLearning (CL) methods are the simplest and most widely used for LLMs to address\nthe forgetting issue. However, traditional replay-based methods do not fully\nutilize instructions to customize the replay strategy. In this work, we propose\na novel paradigm called Instruction-based Continual Learning (InsCL). InsCL\ndynamically replays previous data based on task similarity, calculated by\nWasserstein Distance with instructions. Moreover, we further introduce an\nInstruction Information Metric (InsInfo) to quantify the complexity and\ndiversity of instructions. According to InsInfo, InsCL guides the replay\nprocess more inclined to high-quality data. We conduct extensive experiments\nover 16 tasks with different training orders, observing consistent performance\nimprovements of InsCL. When all tasks have been trained, InsCL achieves\nperformance gains of 3.0 Relative Gain compared with Random Replay, and 27.96\nRelative Gain compared with No Replay.\n", "date": "2024-03-18", "categories": []}
{"id": "2403.11537", "title": "Semantic Prompting with Image-Token for Continual Learning", "abstract": "  Continual learning aims to refine model parameters for new tasks while\nretaining knowledge from previous tasks. Recently, prompt-based learning has\nemerged to leverage pre-trained models to be prompted to learn subsequent tasks\nwithout the reliance on the rehearsal buffer. Although this approach has\ndemonstrated outstanding results, existing methods depend on preceding\ntask-selection process to choose appropriate prompts. However, imperfectness in\ntask-selection may lead to negative impacts on the performance particularly in\nthe scenarios where the number of tasks is large or task distributions are\nimbalanced. To address this issue, we introduce I-Prompt, a task-agnostic\napproach focuses on the visual semantic information of image tokens to\neliminate task prediction. Our method consists of semantic prompt matching,\nwhich determines prompts based on similarities between tokens, and image\ntoken-level prompting, which applies prompts directly to image tokens in the\nintermediate layers. Consequently, our method achieves competitive performance\non four benchmarks while significantly reducing training time compared to\nstate-of-the-art methods. Moreover, we demonstrate the superiority of our\nmethod across various scenarios through extensive experiments.\n", "date": "2024-03-18", "categories": []}
{"id": "2403.12030", "title": "Expandable Subspace Ensemble for Pre-Trained Model-Based\n  Class-Incremental Learning", "abstract": "  Class-Incremental Learning (CIL) requires a learning system to continually\nlearn new classes without forgetting. Despite the strong performance of\nPre-Trained Models (PTMs) in CIL, a critical issue persists: learning new\nclasses often results in the overwriting of old ones. Excessive modification of\nthe network causes forgetting, while minimal adjustments lead to an inadequate\nfit for new classes. As a result, it is desired to figure out a way of\nefficient model updating without harming former knowledge. In this paper, we\npropose ExpAndable Subspace Ensemble (EASE) for PTM-based CIL. To enable model\nupdating without conflict, we train a distinct lightweight adapter module for\neach new task, aiming to create task-specific subspaces. These adapters span a\nhigh-dimensional feature space, enabling joint decision-making across multiple\nsubspaces. As data evolves, the expanding subspaces render the old class\nclassifiers incompatible with new-stage spaces. Correspondingly, we design a\nsemantic-guided prototype complement strategy that synthesizes old classes' new\nfeatures without using any old class instance. Extensive experiments on seven\nbenchmark datasets verify EASE's state-of-the-art performance. Code is\navailable at: https://github.com/sun-hailong/CVPR24-Ease\n", "date": "2024-03-18", "categories": []}
{"id": "2208.14602", "title": "Continuous QA Learning with Structured Prompts", "abstract": "  QA models with lifelong learning (LL) abilities are important for practical\nQA applications, and architecture-based LL methods are reported to be an\neffective implementation for these models. However, it is non-trivial to extend\nprevious approaches to QA tasks since they either require access to task\nidentities in the testing phase or do not explicitly model samples from unseen\ntasks. In this paper, we propose Diana: a dynamic architecture-based lifelong\nQA model that tries to learn a sequence of QA tasks with a prompt enhanced\nlanguage model. Four types of hierarchically organized prompts are used in\nDiana to capture QA knowledge from different granularities. Specifically, we\ndedicate task-level prompts to capture task-specific knowledge to retain high\nLL performances and maintain instance-level prompts to learn knowledge shared\nacross different input samples to improve the model's generalization\nperformance. Moreover, we dedicate separate prompts to explicitly model unseen\ntasks and introduce a set of prompt key vectors to facilitate knowledge sharing\nbetween tasks. Extensive experiments demonstrate that Diana outperforms\nstate-of-the-art lifelong QA models, especially in handling unseen tasks.\n", "date": "2022-08-31", "categories": []}
{"id": "2403.10245", "title": "CoLeCLIP: Open-Domain Continual Learning via Joint Task Prompt and\n  Vocabulary Learning", "abstract": "  This paper explores the problem of continual learning (CL) of vision-language\nmodels (VLMs) in open domains, where the models need to perform continual\nupdating and inference on a streaming of datasets from diverse seen and unseen\ndomains with novel classes. Such a capability is crucial for various\napplications in open environments, e.g., AI assistants, autonomous driving\nsystems, and robotics. Current CL studies mostly focus on closed-set scenarios\nin a single domain with known classes. Large pre-trained VLMs like CLIP have\ndemonstrated superior zero-shot recognition ability, and a number of recent\nstudies leverage this ability to mitigate catastrophic forgetting in CL, but\nthey focus on closed-set CL in a single domain dataset. Open-domain CL of large\nVLMs is significantly more challenging due to 1) large class correlations and\ndomain gaps across the datasets and 2) the forgetting of zero-shot knowledge in\nthe pre-trained VLMs in addition to the knowledge learned from the newly\nadapted datasets. In this work we introduce a novel approach, termed CoLeCLIP,\nthat learns an open-domain CL model based on CLIP. It addresses these\nchallenges by a joint learning of a set of task prompts and a cross-domain\nclass vocabulary. Extensive experiments on 11 domain datasets show that\nCoLeCLIP outperforms state-of-the-art methods for open-domain CL under both\ntask- and class-incremental learning settings.\n", "date": "2024-03-15", "categories": []}
{"id": "2403.10253", "title": "Open Continual Feature Selection via Granular-Ball Knowledge Transfer", "abstract": "  This paper presents a novel framework for continual feature selection (CFS)\nin data preprocessing, particularly in the context of an open and dynamic\nenvironment where unknown classes may emerge. CFS encounters two primary\nchallenges: the discovery of unknown knowledge and the transfer of known\nknowledge. To this end, the proposed CFS method combines the strengths of\ncontinual learning (CL) with granular-ball computing (GBC), which focuses on\nconstructing a granular-ball knowledge base to detect unknown classes and\nfacilitate the transfer of previously learned knowledge for further feature\nselection. CFS consists of two stages: initial learning and open learning. The\nformer aims to establish an initial knowledge base through multi-granularity\nrepresentation using granular-balls. The latter utilizes prior granular-ball\nknowledge to identify unknowns, updates the knowledge base for granular-ball\nknowledge transfer, reinforces old knowledge, and integrates new knowledge.\nSubsequently, we devise an optimal feature subset mechanism that incorporates\nminimal new features into the existing optimal subset, often yielding superior\nresults during each period. Extensive experimental results on public benchmark\ndatasets demonstrate our method's superiority in terms of both effectiveness\nand efficiency compared to state-of-the-art feature selection methods.\n", "date": "2024-03-15", "categories": []}
{"id": "2403.08568", "title": "Consistent Prompting for Rehearsal-Free Continual Learning", "abstract": "  Continual learning empowers models to adapt autonomously to the ever-changing\nenvironment or data streams without forgetting old knowledge. Prompt-based\napproaches are built on frozen pre-trained models to learn the task-specific\nprompts and classifiers efficiently. Existing prompt-based methods are\ninconsistent between training and testing, limiting their effectiveness. Two\ntypes of inconsistency are revealed. Test predictions are made from all\nclassifiers while training only focuses on the current task classifier without\nholistic alignment, leading to Classifier inconsistency. Prompt inconsistency\nindicates that the prompt selected during testing may not correspond to the one\nassociated with this task during training. In this paper, we propose a novel\nprompt-based method, Consistent Prompting (CPrompt), for more aligned training\nand testing. Specifically, all existing classifiers are exposed to prompt\ntraining, resulting in classifier consistency learning. In addition, prompt\nconsistency learning is proposed to enhance prediction robustness and boost\nprompt selection accuracy. Our Consistent Prompting surpasses its prompt-based\ncounterparts and achieves state-of-the-art performance on multiple continual\nlearning benchmarks. Detailed analysis shows that improvements come from more\nconsistent training and testing.\n", "date": "2024-03-13", "categories": []}
{"id": "2309.15048", "title": "Class Incremental Learning via Likelihood Ratio Based Task Prediction", "abstract": "  Class incremental learning (CIL) is a challenging setting of continual\nlearning, which learns a series of tasks sequentially. Each task consists of a\nset of unique classes. The key feature of CIL is that no task identifier (or\ntask-id) is provided at test time. Predicting the task-id for each test sample\nis a challenging problem. An emerging theory-guided approach (called TIL+OOD)\nis to train a task-specific model for each task in a shared network for all\ntasks based on a task-incremental learning (TIL) method to deal with\ncatastrophic forgetting. The model for each task is an out-of-distribution\n(OOD) detector rather than a conventional classifier. The OOD detector can\nperform both within-task (in-distribution (IND)) class prediction and OOD\ndetection. The OOD detection capability is the key to task-id prediction during\ninference. However, this paper argues that using a traditional OOD detector for\ntask-id prediction is sub-optimal because additional information (e.g., the\nreplay data and the learned tasks) available in CIL can be exploited to design\na better and principled method for task-id prediction. We call the new method\nTPL (Task-id Prediction based on Likelihood Ratio). TPL markedly outperforms\nstrong CIL baselines and has negligible catastrophic forgetting. The code of\nTPL is publicly available at https://github.com/linhaowei1/TPL.\n", "date": "2023-09-26", "categories": []}
{"id": "2402.18086", "title": "Generalizable Two-Branch Framework for Image Class-Incremental Learning", "abstract": "  Deep neural networks often severely forget previously learned knowledge when\nlearning new knowledge. Various continual learning (CL) methods have been\nproposed to handle such a catastrophic forgetting issue from different\nperspectives and achieved substantial improvements. In this paper, a novel\ntwo-branch continual learning framework is proposed to further enhance most\nexisting CL methods. Specifically, the main branch can be any existing CL model\nand the newly introduced side branch is a lightweight convolutional network.\nThe output of each main branch block is modulated by the output of the\ncorresponding side branch block. Such a simple two-branch model can then be\neasily implemented and learned with the vanilla optimization setting without\nwhistles and bells. Extensive experiments with various settings on multiple\nimage datasets show that the proposed framework yields consistent improvements\nover state-of-the-art methods.\n", "date": "2024-02-28", "categories": []}
{"id": "2403.07548", "title": "Online Continual Learning For Interactive Instruction Following Agents", "abstract": "  In learning an embodied agent executing daily tasks via language directives,\nthe literature largely assumes that the agent learns all training data at the\nbeginning. We argue that such a learning scenario is less realistic since a\nrobotic agent is supposed to learn the world continuously as it explores and\nperceives it. To take a step towards a more realistic embodied agent learning\nscenario, we propose two continual learning setups for embodied agents;\nlearning new behaviors (Behavior Incremental Learning, Behavior-IL) and new\nenvironments (Environment Incremental Learning, Environment-IL) For the tasks,\nprevious 'data prior' based continual learning methods maintain logits for the\npast tasks. However, the stored information is often insufficiently learned\ninformation and requires task boundary information, which might not always be\navailable. Here, we propose to update them based on confidence scores without\ntask boundary information during training (i.e., task-free) in a moving average\nfashion, named Confidence-Aware Moving Average (CAMA). In the proposed\nBehavior-IL and Environment-IL setups, our simple CAMA outperforms prior state\nof the art in our empirical validations by noticeable margins. The project page\nincluding codes is https://github.com/snumprlab/cl-alfred.\n", "date": "2024-03-12", "categories": []}
{"id": "2403.06670", "title": "CEAT: Continual Expansion and Absorption Transformer for Non-Exemplar\n  Class-Incremental Learning", "abstract": "  In real-world applications, dynamic scenarios require the models to possess\nthe capability to learn new tasks continuously without forgetting the old\nknowledge. Experience-Replay methods store a subset of the old images for joint\ntraining. In the scenario of more strict privacy protection, storing the old\nimages becomes infeasible, which leads to a more severe plasticity-stability\ndilemma and classifier bias. To meet the above challenges, we propose a new\narchitecture, named continual expansion and absorption transformer~(CEAT). The\nmodel can learn the novel knowledge by extending the expanded-fusion layers in\nparallel with the frozen previous parameters. After the task ends, we\nlosslessly absorb the extended parameters into the backbone to ensure that the\nnumber of parameters remains constant. To improve the learning ability of the\nmodel, we designed a novel prototype contrastive loss to reduce the overlap\nbetween old and new classes in the feature space. Besides, to address the\nclassifier bias towards the new classes, we propose a novel approach to\ngenerate the pseudo-features to correct the classifier. We experiment with our\nmethods on three standard Non-Exemplar Class-Incremental Learning~(NECIL)\nbenchmarks. Extensive experiments demonstrate that our model gets a significant\nimprovement compared with the previous works and achieves 5.38%, 5.20%, and\n4.92% improvement on CIFAR-100, TinyImageNet, and ImageNet-Subset.\n", "date": "2024-03-11", "categories": []}
{"id": "2403.07292", "title": "Continual All-in-One Adverse Weather Removal with Knowledge Replay on a\n  Unified Network Structure", "abstract": "  In real-world applications, image degeneration caused by adverse weather is\nalways complex and changes with different weather conditions from days and\nseasons. Systems in real-world environments constantly encounter adverse\nweather conditions that are not previously observed. Therefore, it practically\nrequires adverse weather removal models to continually learn from incrementally\ncollected data reflecting various degeneration types. Existing adverse weather\nremoval approaches, for either single or multiple adverse weathers, are mainly\ndesigned for a static learning paradigm, which assumes that the data of all\ntypes of degenerations to handle can be finely collected at one time before a\nsingle-phase learning process. They thus cannot directly handle the incremental\nlearning requirements. To address this issue, we made the earliest effort to\ninvestigate the continual all-in-one adverse weather removal task, in a setting\ncloser to real-world applications. Specifically, we develop a novel continual\nlearning framework with effective knowledge replay (KR) on a unified network\nstructure. Equipped with a principal component projection and an effective\nknowledge distillation mechanism, the proposed KR techniques are tailored for\nthe all-in-one weather removal task. It considers the characteristics of the\nimage restoration task with multiple degenerations in continual learning, and\nthe knowledge for different degenerations can be shared and accumulated in the\nunified network structure. Extensive experimental results demonstrate the\neffectiveness of the proposed method to deal with this challenging task, which\nperforms competitively to existing dedicated or joint training image\nrestoration methods. Our code is available at\nhttps://github.com/xiaojihh/CL_all-in-one.\n", "date": "2024-03-12", "categories": []}
{"id": "2403.07356", "title": "Premonition: Using Generative Models to Preempt Future Data Changes in\n  Continual Learning", "abstract": "  Continual learning requires a model to adapt to ongoing changes in the data\ndistribution, and often to the set of tasks to be performed. It is rare,\nhowever, that the data and task changes are completely unpredictable. Given a\ndescription of an overarching goal or data theme, which we call a realm, humans\ncan often guess what concepts are associated with it. We show here that the\ncombination of a large language model and an image generation model can\nsimilarly provide useful premonitions as to how a continual learning challenge\nmight develop over time. We use the large language model to generate text\ndescriptions of semantically related classes that might potentially appear in\nthe data stream in future. These descriptions are then rendered using Stable\nDiffusion to generate new labelled image samples. The resulting synthetic\ndataset is employed for supervised pre-training, but is discarded prior to\ncommencing continual learning, along with the pre-training classification head.\nWe find that the backbone of our pre-trained networks can learn representations\nuseful for the downstream continual learning problem, thus becoming a valuable\ninput to any existing continual learning method. Although there are\ncomplexities arising from the domain gap between real and synthetic images, we\nshow that pre-training models in this manner improves multiple Class Incremenal\nLearning (CIL) methods on fine-grained image classification benchmarks.\nSupporting code can be found at https://github.com/cl-premonition/premonition.\n", "date": "2024-03-12", "categories": []}
{"id": "2403.07406", "title": "FeTrIL++: Feature Translation for Exemplar-Free Class-Incremental\n  Learning with Hill-Climbing", "abstract": "  Exemplar-free class-incremental learning (EFCIL) poses significant\nchallenges, primarily due to catastrophic forgetting, necessitating a delicate\nbalance between stability and plasticity to accurately recognize both new and\nprevious classes. Traditional EFCIL approaches typically skew towards either\nmodel plasticity through successive fine-tuning or stability by employing a\nfixed feature extractor beyond the initial incremental state. Building upon the\nfoundational FeTrIL framework, our research extends into novel experimental\ndomains to examine the efficacy of various oversampling techniques and dynamic\noptimization strategies across multiple challenging datasets and incremental\nsettings. We specifically explore how oversampling impacts accuracy relative to\nfeature availability and how different optimization methodologies, including\ndynamic recalibration and feature pool diversification, influence incremental\nlearning outcomes. The results from these comprehensive experiments, conducted\non CIFAR100, Tiny-ImageNet, and an ImageNet-Subset, under-score the superior\nperformance of FeTrIL in balancing accuracy for both new and past classes\nagainst ten contemporary methods. Notably, our extensions reveal the nuanced\nimpacts of oversampling and optimization on EFCIL, contributing to a more\nrefined understanding of feature-space manipulation for class incremental\nlearning. FeTrIL and its extended analysis in this paper FeTrIL++ pave the way\nfor more adaptable and efficient EFCIL methodologies, promising significant\nimprovements in handling catastrophic forgetting without the need for\nexemplars.\n", "date": "2024-03-12", "categories": []}
{"id": "2403.07586", "title": "Federated Learning of Socially Appropriate Agent Behaviours in Simulated\n  Home Environments", "abstract": "  As social robots become increasingly integrated into daily life, ensuring\ntheir behaviours align with social norms is crucial. For their widespread\nopen-world application, it is important to explore Federated Learning (FL)\nsettings where individual robots can learn about their unique environments\nwhile also learning from each others' experiences. In this paper, we present a\nnovel FL benchmark that evaluates different strategies, using multi-label\nregression objectives, where each client individually learns to predict the\nsocial appropriateness of different robot actions while also sharing their\nlearning with others. Furthermore, splitting the training data by different\ncontexts such that each client incrementally learns across contexts, we present\na novel Federated Continual Learning (FCL) benchmark that adapts FL-based\nmethods to use state-of-the-art Continual Learning (CL) methods to continually\nlearn socially appropriate agent behaviours under different contextual\nsettings. Federated Averaging (FedAvg) of weights emerges as a robust FL\nstrategy while rehearsal-based FCL enables incrementally learning the social\nappropriateness of robot actions, across contextual splits.\n", "date": "2024-03-12", "categories": []}
{"id": "2403.07851", "title": "12 mJ per Class On-Device Online Few-Shot Class-Incremental Learning", "abstract": "  Few-Shot Class-Incremental Learning (FSCIL) enables machine learning systems\nto expand their inference capabilities to new classes using only a few labeled\nexamples, without forgetting the previously learned classes. Classical\nbackpropagation-based learning and its variants are often unsuitable for\nbattery-powered, memory-constrained systems at the extreme edge. In this work,\nwe introduce Online Few-Shot Class-Incremental Learning (O-FSCIL), based on a\nlightweight model consisting of a pretrained and metalearned feature extractor\nand an expandable explicit memory storing the class prototypes. The\narchitecture is pretrained with a novel feature orthogonality regularization\nand metalearned with a multi-margin loss. For learning a new class, our\napproach extends the explicit memory with novel class prototypes, while the\nremaining architecture is kept frozen. This allows learning previously unseen\nclasses based on only a few examples with one single pass (hence online).\nO-FSCIL obtains an average accuracy of 68.62% on the FSCIL CIFAR100 benchmark,\nachieving state-of-the-art results. Tailored for ultra-low-power platforms, we\nimplement O-FSCIL on the 60 mW GAP9 microcontroller, demonstrating online\nlearning capabilities within just 12 mJ per new class.\n", "date": "2024-03-12", "categories": []}
{"id": "2305.18806", "title": "Prediction Error-based Classification for Class-Incremental Learning", "abstract": "  Class-incremental learning (CIL) is a particularly challenging variant of\ncontinual learning, where the goal is to learn to discriminate between all\nclasses presented in an incremental fashion. Existing approaches often suffer\nfrom excessive forgetting and imbalance of the scores assigned to classes that\nhave not been seen together during training. In this study, we introduce a\nnovel approach, Prediction Error-based Classification (PEC), which differs from\ntraditional discriminative and generative classification paradigms. PEC\ncomputes a class score by measuring the prediction error of a model trained to\nreplicate the outputs of a frozen random neural network on data from that\nclass. The method can be interpreted as approximating a classification rule\nbased on Gaussian Process posterior variance. PEC offers several practical\nadvantages, including sample efficiency, ease of tuning, and effectiveness even\nwhen data are presented one class at a time. Our empirical results show that\nPEC performs strongly in single-pass-through-data CIL, outperforming other\nrehearsal-free baselines in all cases and rehearsal-based methods with moderate\nreplay buffer size in most cases across multiple benchmarks.\n", "date": "2023-05-30", "categories": []}
{"id": "2403.06102", "title": "Coherent Temporal Synthesis for Incremental Action Segmentation", "abstract": "  Data replay is a successful incremental learning technique for images. It\nprevents catastrophic forgetting by keeping a reservoir of previous data,\noriginal or synthesized, to ensure the model retains past knowledge while\nadapting to novel concepts. However, its application in the video domain is\nrudimentary, as it simply stores frame exemplars for action recognition. This\npaper presents the first exploration of video data replay techniques for\nincremental action segmentation, focusing on action temporal modeling. We\npropose a Temporally Coherent Action (TCA) model, which represents actions\nusing a generative model instead of storing individual frames. The integration\nof a conditioning variable that captures temporal coherence allows our model to\nunderstand the evolution of action features over time. Therefore, action\nsegments generated by TCA for replay are diverse and temporally coherent. In a\n10-task incremental setup on the Breakfast dataset, our approach achieves\nsignificant increases in accuracy for up to 22% compared to the baselines.\n", "date": "2024-03-10", "categories": []}
{"id": "2403.06288", "title": "Probing Image Compression For Class-Incremental Learning", "abstract": "  Image compression emerges as a pivotal tool in the efficient handling and\ntransmission of digital images. Its ability to substantially reduce file size\nnot only facilitates enhanced data storage capacity but also potentially brings\nadvantages to the development of continual machine learning (ML) systems, which\nlearn new knowledge incrementally from sequential data. Continual ML systems\noften rely on storing representative samples, also known as exemplars, within a\nlimited memory constraint to maintain the performance on previously learned\ndata. These methods are known as memory replay-based algorithms and have proven\neffective at mitigating the detrimental effects of catastrophic forgetting.\nNonetheless, the limited memory buffer size often falls short of adequately\nrepresenting the entire data distribution. In this paper, we explore the use of\nimage compression as a strategy to enhance the buffer's capacity, thereby\nincreasing exemplar diversity. However, directly using compressed exemplars\nintroduces domain shift during continual ML, marked by a discrepancy between\ncompressed training data and uncompressed testing data. Additionally, it is\nessential to determine the appropriate compression algorithm and select the\nmost effective rate for continual ML systems to balance the trade-off between\nexemplar quality and quantity. To this end, we introduce a new framework to\nincorporate image compression for continual ML including a pre-processing data\ncompression step and an efficient compression rate/algorithm selection method.\nWe conduct extensive experiments on CIFAR-100 and ImageNet datasets and show\nthat our method significantly improves image classification accuracy in\ncontinual ML settings.\n", "date": "2024-03-10", "categories": []}
{"id": "2403.06295", "title": "A streamlined Approach to Multimodal Few-Shot Class Incremental Learning\n  for Fine-Grained Datasets", "abstract": "  Few-shot Class-Incremental Learning (FSCIL) poses the challenge of retaining\nprior knowledge while learning from limited new data streams, all without\noverfitting. The rise of Vision-Language models (VLMs) has unlocked numerous\napplications, leveraging their existing knowledge to fine-tune on custom data.\nHowever, training the whole model is computationally prohibitive, and VLMs\nwhile being versatile in general domains still struggle with fine-grained\ndatasets crucial for many applications. We tackle these challenges with two\nproposed simple modules. The first, Session-Specific Prompts (SSP), enhances\nthe separability of image-text embeddings across sessions. The second,\nHyperbolic distance, compresses representations of image-text pairs within the\nsame class while expanding those from different classes, leading to better\nrepresentations. Experimental results demonstrate an average 10-point increase\ncompared to baselines while requiring at least 8 times fewer trainable\nparameters. This improvement is further underscored on our three newly\nintroduced fine-grained datasets.\n", "date": "2024-03-10", "categories": []}
{"id": "2211.12971", "title": "Cooperative data-driven modeling", "abstract": "  Data-driven modeling in mechanics is evolving rapidly based on recent machine\nlearning advances, especially on artificial neural networks. As the field\nmatures, new data and models created by different groups become available,\nopening possibilities for cooperative modeling. However, artificial neural\nnetworks suffer from catastrophic forgetting, i.e. they forget how to perform\nan old task when trained on a new one. This hinders cooperation because\nadapting an existing model for a new task affects the performance on a previous\ntask trained by someone else. The authors developed a continual learning method\nthat addresses this issue, applying it here for the first time to solid\nmechanics. In particular, the method is applied to recurrent neural networks to\npredict history-dependent plasticity behavior, although it can be used on any\nother architecture (feedforward, convolutional, etc.) and to predict other\nphenomena. This work intends to spawn future developments on continual learning\nthat will foster cooperative strategies among the mechanics community to solve\nincreasingly challenging problems. We show that the chosen continual learning\nstrategy can sequentially learn several constitutive laws without forgetting\nthem, using less data to achieve the same error as standard (non-cooperative)\ntraining of one law per model.\n", "date": "2022-11-23", "categories": []}
{"id": "2304.04646", "title": "ECG-CL: A Comprehensive Electrocardiogram Interpretation Method Based on\n  Continual Learning", "abstract": "  Electrocardiogram (ECG) monitoring is one of the most powerful technique of\ncardiovascular disease (CVD) early identification, and the introduction of\nintelligent wearable ECG devices has enabled daily monitoring. However, due to\nthe need for professional expertise in the ECGs interpretation, general public\naccess has once again been restricted, prompting the need for the development\nof advanced diagnostic algorithms. Classic rule-based algorithms are now\ncompletely outperformed by deep learning based methods. But the advancement of\nsmart diagnostic algorithms is hampered by issues like small dataset,\ninconsistent data labeling, inefficient use of local and global ECG\ninformation, memory and inference time consuming deployment of multiple models,\nand lack of information transfer between tasks. We propose a multi-resolution\nmodel that can sustain high-resolution low-level semantic information\nthroughout, with the help of the development of low-resolution high-level\nsemantic information, by capitalizing on both local morphological information\nand global rhythm information. From the perspective of effective data leverage\nand inter-task knowledge transfer, we develop a parameter isolation based ECG\ncontinual learning (ECG-CL) approach. We evaluated our model's performance on\nfour open-access datasets by designing segmentation-to-classification for\ncross-domain incremental learning, minority-to-majority class for category\nincremental learning, and small-to-large sample for task incremental learning.\nOur approach is shown to successfully extract informative morphological and\nrhythmic features from ECG segmentation, leading to higher quality\nclassification results. From the perspective of intelligent wearable\napplications, the possibility of a comprehensive ECG interpretation algorithm\nbased on single-lead ECGs is also confirmed.\n", "date": "2023-04-10", "categories": []}
{"id": "2403.04140", "title": "Contrastive Augmented Graph2Graph Memory Interaction for Few Shot\n  Continual Learning", "abstract": "  Few-Shot Class-Incremental Learning (FSCIL) has gained considerable attention\nin recent years for its pivotal role in addressing continuously arriving\nclasses. However, it encounters additional challenges. The scarcity of samples\nin new sessions intensifies overfitting, causing incompatibility between the\noutput features of new and old classes, thereby escalating catastrophic\nforgetting. A prevalent strategy involves mitigating catastrophic forgetting\nthrough the Explicit Memory (EM), which comprise of class prototypes. However,\ncurrent EM-based methods retrieves memory globally by performing\nVector-to-Vector (V2V) interaction between features corresponding to the input\nand prototypes stored in EM, neglecting the geometric structure of local\nfeatures. This hinders the accurate modeling of their positional relationships.\nTo incorporate information of local geometric structure, we extend the V2V\ninteraction to Graph-to-Graph (G2G) interaction. For enhancing local structures\nfor better G2G alignment and the prevention of local feature collapse, we\npropose the Local Graph Preservation (LGP) mechanism. Additionally, to address\nsample scarcity in classes from new sessions, the Contrast-Augmented G2G\n(CAG2G) is introduced to promote the aggregation of same class features thus\nhelps few-shot learning. Extensive comparisons on CIFAR100, CUB200, and the\nchallenging ImageNet-R dataset demonstrate the superiority of our method over\nexisting methods.\n", "date": "2024-03-07", "categories": []}
{"id": "2403.04599", "title": "Contrastive Continual Learning with Importance Sampling and\n  Prototype-Instance Relation Distillation", "abstract": "  Recently, because of the high-quality representations of contrastive learning\nmethods, rehearsal-based contrastive continual learning has been proposed to\nexplore how to continually learn transferable representation embeddings to\navoid the catastrophic forgetting issue in traditional continual settings.\nBased on this framework, we propose Contrastive Continual Learning via\nImportance Sampling (CCLIS) to preserve knowledge by recovering previous data\ndistributions with a new strategy for Replay Buffer Selection (RBS), which\nminimize estimated variance to save hard negative samples for representation\nlearning with high quality. Furthermore, we present the Prototype-instance\nRelation Distillation (PRD) loss, a technique designed to maintain the\nrelationship between prototypes and sample representations using a\nself-distillation process. Experiments on standard continual learning\nbenchmarks reveal that our method notably outperforms existing baselines in\nterms of knowledge preservation and thereby effectively counteracts\ncatastrophic forgetting in online contexts. The code is available at\nhttps://github.com/lijy373/CCLIS.\n", "date": "2024-03-07", "categories": []}
{"id": "2403.04759", "title": "Lifelong Intelligence Beyond the Edge using Hyperdimensional Computing", "abstract": "  On-device learning has emerged as a prevailing trend that avoids the slow\nresponse time and costly communication of cloud-based learning. The ability to\nlearn continuously and indefinitely in a changing environment, and with\nresource constraints, is critical for real sensor deployments. However,\nexisting designs are inadequate for practical scenarios with (i) streaming data\ninput, (ii) lack of supervision and (iii) limited on-board resources. In this\npaper, we design and deploy the first on-device lifelong learning system called\nLifeHD for general IoT applications with limited supervision. LifeHD is\ndesigned based on a novel neurally-inspired and lightweight learning paradigm\ncalled Hyperdimensional Computing (HDC). We utilize a two-tier associative\nmemory organization to intelligently store and manage high-dimensional,\nlow-precision vectors, which represent the historical patterns as cluster\ncentroids. We additionally propose two variants of LifeHD to cope with scarce\nlabeled inputs and power constraints. We implement LifeHD on off-the-shelf edge\nplatforms and perform extensive evaluations across three scenarios. Our\nmeasurements show that LifeHD improves the unsupervised clustering accuracy by\nup to 74.8% compared to the state-of-the-art NN-based unsupervised lifelong\nlearning baselines with as much as 34.3x better energy efficiency. Our code is\navailable at https://github.com/Orienfish/LifeHD.\n", "date": "2024-03-07", "categories": []}
{"id": "2403.03382", "title": "Adaptive Discovering and Merging for Incremental Novel Class Discovery", "abstract": "  One important desideratum of lifelong learning aims to discover novel classes\nfrom unlabelled data in a continuous manner. The central challenge is twofold:\ndiscovering and learning novel classes while mitigating the issue of\ncatastrophic forgetting of established knowledge. To this end, we introduce a\nnew paradigm called Adaptive Discovering and Merging (ADM) to discover novel\ncategories adaptively in the incremental stage and integrate novel knowledge\ninto the model without affecting the original knowledge. To discover novel\nclasses adaptively, we decouple representation learning and novel class\ndiscovery, and use Triple Comparison (TC) and Probability Regularization (PR)\nto constrain the probability discrepancy and diversity for adaptive category\nassignment. To merge the learned novel knowledge adaptively, we propose a\nhybrid structure with base and novel branches named Adaptive Model Merging\n(AMM), which reduces the interference of the novel branch on the old classes to\npreserve the previous knowledge, and merges the novel branch to the base model\nwithout performance loss and parameter growth. Extensive experiments on several\ndatasets show that ADM significantly outperforms existing class-incremental\nNovel Class Discovery (class-iNCD) approaches. Moreover, our AMM also benefits\nthe class-incremental Learning (class-IL) task by alleviating the catastrophic\nforgetting problem.\n", "date": "2024-03-06", "categories": []}
{"id": "2403.03462", "title": "Interactive Continual Learning Architecture for Long-Term\n  Personalization of Home Service Robots", "abstract": "  For robots to perform assistive tasks in unstructured home environments, they\nmust learn and reason on the semantic knowledge of the environments. Despite a\nresurgence in the development of semantic reasoning architectures, these\nmethods assume that all the training data is available a priori. However, each\nuser's environment is unique and can continue to change over time, which makes\nthese methods unsuitable for personalized home service robots. Although\nresearch in continual learning develops methods that can learn and adapt over\ntime, most of these methods are tested in the narrow context of object\nclassification on static image datasets. In this paper, we combine ideas from\ncontinual learning, semantic reasoning, and interactive machine learning\nliterature and develop a novel interactive continual learning architecture for\ncontinual learning of semantic knowledge in a home environment through\nhuman-robot interaction. The architecture builds on core cognitive principles\nof learning and memory for efficient and real-time learning of new knowledge\nfrom humans. We integrate our architecture with a physical mobile manipulator\nrobot and perform extensive system evaluations in a laboratory environment over\ntwo months. Our results demonstrate the effectiveness of our architecture to\nallow a physical robot to continually adapt to the changes in the environment\nfrom limited data provided by the users (experimenters), and use the learned\nknowledge to perform object fetching tasks.\n", "date": "2024-03-06", "categories": []}
{"id": "2403.02718", "title": "DP-CRE: Continual Relation Extraction via Decoupled Contrastive Learning\n  and Memory Structure Preservation", "abstract": "  Continuous Relation Extraction (CRE) aims to incrementally learn relation\nknowledge from a non-stationary stream of data. Since the introduction of new\nrelational tasks can overshadow previously learned information, catastrophic\nforgetting becomes a significant challenge in this domain. Current replay-based\ntraining paradigms prioritize all data uniformly and train memory samples\nthrough multiple rounds, which would result in overfitting old tasks and\npronounced bias towards new tasks because of the imbalances of the replay set.\nTo handle the problem, we introduce the DecouPled CRE (DP-CRE) framework that\ndecouples the process of prior information preservation and new knowledge\nacquisition. This framework examines alterations in the embedding space as new\nrelation classes emerge, distinctly managing the preservation and acquisition\nof knowledge. Extensive experiments show that DP-CRE significantly outperforms\nother CRE baselines across two datasets.\n", "date": "2024-03-05", "categories": []}
{"id": "2403.03082", "title": "Recall-Oriented Continual Learning with Generative Adversarial\n  Meta-Model", "abstract": "  The stability-plasticity dilemma is a major challenge in continual learning,\nas it involves balancing the conflicting objectives of maintaining performance\non previous tasks while learning new tasks. In this paper, we propose the\nrecall-oriented continual learning framework to address this challenge.\nInspired by the human brain's ability to separate the mechanisms responsible\nfor stability and plasticity, our framework consists of a two-level\narchitecture where an inference network effectively acquires new knowledge and\na generative network recalls past knowledge when necessary. In particular, to\nmaximize the stability of past knowledge, we investigate the complexity of\nknowledge depending on different representations, and thereby introducing\ngenerative adversarial meta-model (GAMM) that incrementally learns\ntask-specific parameters instead of input data samples of the task. Through our\nexperiments, we show that our framework not only effectively learns new\nknowledge without any disruption but also achieves high stability of previous\nknowledge in both task-aware and task-agnostic learning scenarios. Our code is\navailable at: https://github.com/bigdata-inha/recall-oriented-cl-framework.\n", "date": "2024-03-05", "categories": []}
{"id": "2309.15135", "title": "Contrastive Continual Multi-view Clustering with Filtered Structural\n  Fusion", "abstract": "  Multi-view clustering thrives in applications where views are collected in\nadvance by extracting consistent and complementary information among views.\nHowever, it overlooks scenarios where data views are collected sequentially,\ni.e., real-time data. Due to privacy issues or memory burden, previous views\nare not available with time in these situations. Some methods are proposed to\nhandle it but are trapped in a stability-plasticity dilemma. In specific, these\nmethods undergo a catastrophic forgetting of prior knowledge when a new view is\nattained. Such a catastrophic forgetting problem (CFP) would cause the\nconsistent and complementary information hard to get and affect the clustering\nperformance. To tackle this, we propose a novel method termed Contrastive\nContinual Multi-view Clustering with Filtered Structural Fusion (CCMVC-FSF).\nPrecisely, considering that data correlations play a vital role in clustering\nand prior knowledge ought to guide the clustering process of a new view, we\ndevelop a data buffer with fixed size to store filtered structural information\nand utilize it to guide the generation of a robust partition matrix via\ncontrastive learning. Furthermore, we theoretically connect CCMVC-FSF with\nsemi-supervised learning and knowledge distillation. Extensive experiments\nexhibit the excellence of the proposed method.\n", "date": "2023-09-26", "categories": []}
{"id": "2402.13711", "title": "DSLR: Diversity Enhancement and Structure Learning for Rehearsal-based\n  Graph Continual Learning", "abstract": "  We investigate the replay buffer in rehearsal-based approaches for graph\ncontinual learning (GCL) methods. Existing rehearsal-based GCL methods select\nthe most representative nodes for each class and store them in a replay buffer\nfor later use in training subsequent tasks. However, we discovered that\nconsidering only the class representativeness of each replayed node makes the\nreplayed nodes to be concentrated around the center of each class, incurring a\npotential risk of overfitting to nodes residing in those regions, which\naggravates catastrophic forgetting. Moreover, as the rehearsal-based approach\nheavily relies on a few replayed nodes to retain knowledge obtained from\nprevious tasks, involving the replayed nodes that have irrelevant neighbors in\nthe model training may have a significant detrimental impact on model\nperformance. In this paper, we propose a GCL model named DSLR, specifically, we\ndevise a coverage-based diversity (CD) approach to consider both the class\nrepresentativeness and the diversity within each class of the replayed nodes.\nMoreover, we adopt graph structure learning (GSL) to ensure that the replayed\nnodes are connected to truly informative neighbors. Extensive experimental\nresults demonstrate the effectiveness and efficiency of DSLR. Our source code\nis available at https://github.com/seungyoon-Choi/DSLR_official.\n", "date": "2024-02-21", "categories": []}
{"id": "2403.01554", "title": "Transformers for Supervised Online Continual Learning", "abstract": "  Transformers have become the dominant architecture for sequence modeling\ntasks such as natural language processing or audio processing, and they are now\neven considered for tasks that are not naturally sequential such as image\nclassification. Their ability to attend to and to process a set of tokens as\ncontext enables them to develop in-context few-shot learning abilities.\nHowever, their potential for online continual learning remains relatively\nunexplored. In online continual learning, a model must adapt to a\nnon-stationary stream of data, minimizing the cumulative nextstep prediction\nloss. We focus on the supervised online continual learning setting, where we\nlearn a predictor $x_t \\rightarrow y_t$ for a sequence of examples $(x_t,\ny_t)$. Inspired by the in-context learning capabilities of transformers and\ntheir connection to meta-learning, we propose a method that leverages these\nstrengths for online continual learning. Our approach explicitly conditions a\ntransformer on recent observations, while at the same time online training it\nwith stochastic gradient descent, following the procedure introduced with\nTransformer-XL. We incorporate replay to maintain the benefits of multi-epoch\ntraining while adhering to the sequential protocol. We hypothesize that this\ncombination enables fast adaptation through in-context learning and sustained\nlongterm improvement via parametric learning. Our method demonstrates\nsignificant improvements over previous state-of-the-art results on CLOC, a\nchallenging large-scale real-world benchmark for image geo-localization.\n", "date": "2024-03-03", "categories": []}
{"id": "2403.02334", "title": "Gradient Correlation Subspace Learning against Catastrophic Forgetting", "abstract": "  Efficient continual learning techniques have been a topic of significant\nresearch over the last few years. A fundamental problem with such learning is\nsevere degradation of performance on previously learned tasks, known also as\ncatastrophic forgetting. This paper introduces a novel method to reduce\ncatastrophic forgetting in the context of incremental class learning called\nGradient Correlation Subspace Learning (GCSL). The method detects a subspace of\nthe weights that is least affected by previous tasks and projects the weights\nto train for the new task into said subspace. The method can be applied to one\nor more layers of a given network architectures and the size of the subspace\nused can be altered from layer to layer and task to task. Code will be\navailable at\n\\href{https://github.com/vgthengane/GCSL}{https://github.com/vgthengane/GCSL}\n", "date": "2024-03-04", "categories": []}
{"id": "2312.16475", "title": "Federated Continual Learning via Knowledge Fusion: A Survey", "abstract": "  Data privacy and silos are nontrivial and greatly challenging in many\nreal-world applications. Federated learning is a decentralized approach to\ntraining models across multiple local clients without the exchange of raw data\nfrom client devices to global servers. However, existing works focus on a\nstatic data environment and ignore continual learning from streaming data with\nincremental tasks. Federated Continual Learning (FCL) is an emerging paradigm\nto address model learning in both federated and continual learning\nenvironments. The key objective of FCL is to fuse heterogeneous knowledge from\ndifferent clients and retain knowledge of previous tasks while learning on new\nones. In this work, we delineate federated learning and continual learning\nfirst and then discuss their integration, i.e., FCL, and particular FCL via\nknowledge fusion. In summary, our motivations are four-fold: we (1) raise a\nfundamental problem called ''spatial-temporal catastrophic forgetting'' and\nevaluate its impact on the performance using a well-known method called\nfederated averaging (FedAvg), (2) integrate most of the existing FCL methods\ninto two generic frameworks, namely synchronous FCL and asynchronous FCL, (3)\ncategorize a large number of methods according to the mechanism involved in\nknowledge fusion, and finally (4) showcase an outlook on the future work of\nFCL.\n", "date": "2023-12-27", "categories": []}
{"id": "2402.18865", "title": "Analyzing and Reducing Catastrophic Forgetting in Parameter Efficient\n  Tuning", "abstract": "  Existing research has shown that large language models (LLMs) exhibit\nremarkable performance in language understanding and generation. However, when\nLLMs are continuously fine-tuned on complex and diverse domain-specific\ndownstream tasks, the inference performance on historical tasks decreases\ndramatically, which is known as a catastrophic forgetting problem. A trade-off\nneeds to be kept between learning plasticity and memory stability. Plenty of\nexisting works have explored strategies like memory replay, regularization and\nparameter isolation, but little is known about the geometric connection of\nvarious adjacent minima in the continual LLMs fine-tuning scenarios. In this\nwork, we investigate the geometric connections of different minima through the\nlens of mode connectivity, which means different minima can be connected by a\nlow-loss valley. Through extensive experiments, we uncover the mode\nconnectivity phenomenon in the LLMs continual learning scenario and find that\nit can strike a balance between plasticity and stability. Building upon these\nfindings, we propose a simple yet effective method called Interpolation-based\nLoRA (I-LoRA), which constructs a dual-memory experience replay framework based\non LoRA parameter interpolations. Extensive experiments and analysis on eight\ndomain-specific CL benchmarks demonstrate that I-LoRA consistently show\nsignificant improvement over the previous state-of-the-art approaches with up\nto $11\\%$ performance gains, providing a strong baseline and insights for\nfuture research on the large language model continual learning problem. Our\ncode is available at \\url{https://github.com/which47/LLMCL}.\n", "date": "2024-02-29", "categories": []}
{"id": "2402.17792", "title": "EGNN-C+: Interpretable Evolving Granular Neural Network and Application\n  in Classification of Weakly-Supervised EEG Data Streams", "abstract": "  We introduce a modified incremental learning algorithm for evolving Granular\nNeural Network Classifiers (eGNN-C+). We use double-boundary hyper-boxes to\nrepresent granules, and customize the adaptation procedures to enhance the\nrobustness of outer boxes for data coverage and noise suppression, while\nensuring that inner boxes remain flexible to capture drifts. The classifier\nevolves from scratch, incorporates new classes on the fly, and performs local\nincremental feature weighting. As an application, we focus on the\nclassification of emotion-related patterns within electroencephalogram (EEG)\nsignals. Emotion recognition is crucial for enhancing the realism and\ninteractivity of computer systems. We extract features from the Fourier\nspectrum of EEG signals obtained from 28 individuals engaged in playing\ncomputer games -- a public dataset. Each game elicits a different predominant\nemotion: boredom, calmness, horror, or joy. We analyze individual electrodes,\ntime window lengths, and frequency bands to assess the accuracy and\ninterpretability of resulting user-independent neural models. The findings\nindicate that both brain hemispheres assist classification, especially\nelectrodes on the temporal (T8) and parietal (P7) areas, alongside\ncontributions from frontal and occipital electrodes. While patterns may\nmanifest in any band, the Alpha (8-13Hz), Delta (1-4Hz), and Theta (4-8Hz)\nbands, in this order, exhibited higher correspondence with the emotion classes.\nThe eGNN-C+ demonstrates effectiveness in learning EEG data. It achieves an\naccuracy of 81.7% and a 0.0029 II interpretability using 10-second time\nwindows, even in face of a highly-stochastic time-varying 4-class\nclassification problem.\n", "date": "2024-02-26", "categories": []}
{"id": "2402.18449", "title": "HOP to the Next Tasks and Domains for Continual Learning in NLP", "abstract": "  Continual Learning (CL) aims to learn a sequence of problems (i.e., tasks and\ndomains) by transferring knowledge acquired on previous problems, whilst\navoiding forgetting of past ones. Different from previous approaches which\nfocused on CL for one NLP task or domain in a specific use-case, in this paper,\nwe address a more general CL setting to learn from a sequence of problems in a\nunique framework. Our method, HOP, permits to hop across tasks and domains by\naddressing the CL problem along three directions: (i) we employ a set of\nadapters to generalize a large pre-trained model to unseen problems, (ii) we\ncompute high-order moments over the distribution of embedded representations to\ndistinguish independent and correlated statistics across different tasks and\ndomains, (iii) we process this enriched information with auxiliary heads\nspecialized for each end problem. Extensive experimental campaign on 4 NLP\napplications, 5 benchmarks and 2 CL setups demonstrates the effectiveness of\nour HOP.\n", "date": "2024-02-28", "categories": []}
{"id": "2304.01899", "title": "Cross-Class Feature Augmentation for Class Incremental Learning", "abstract": "  We propose a novel class incremental learning approach by incorporating a\nfeature augmentation technique motivated by adversarial attacks. We employ a\nclassifier learned in the past to complement training examples rather than\nsimply play a role as a teacher for knowledge distillation towards subsequent\nmodels. The proposed approach has a unique perspective to utilize the previous\nknowledge in class incremental learning since it augments features of arbitrary\ntarget classes using examples in other classes via adversarial attacks on a\npreviously learned classifier. By allowing the cross-class feature\naugmentations, each class in the old tasks conveniently populates samples in\nthe feature space, which alleviates the collapse of the decision boundaries\ncaused by sample deficiency for the previous tasks, especially when the number\nof stored exemplars is small. This idea can be easily incorporated into\nexisting class incremental learning algorithms without any architecture\nmodification. Extensive experiments on the standard benchmarks show that our\nmethod consistently outperforms existing class incremental learning methods by\nsignificant margins in various scenarios, especially under an environment with\nan extremely limited memory budget.\n", "date": "2023-04-04", "categories": []}
{"id": "2402.16189", "title": "One-stage Prompt-based Continual Learning", "abstract": "  Prompt-based Continual Learning (PCL) has gained considerable attention as a\npromising continual learning solution as it achieves state-of-the-art\nperformance while preventing privacy violation and memory overhead issues.\nNonetheless, existing PCL approaches face significant computational burdens\nbecause of two Vision Transformer (ViT) feed-forward stages; one is for the\nquery ViT that generates a prompt query to select prompts inside a prompt pool;\nthe other one is a backbone ViT that mixes information between selected prompts\nand image tokens. To address this, we introduce a one-stage PCL framework by\ndirectly using the intermediate layer's token embedding as a prompt query. This\ndesign removes the need for an additional feed-forward stage for query ViT,\nresulting in ~50% computational cost reduction for both training and inference\nwith marginal accuracy drop < 1%. We further introduce a Query-Pool\nRegularization (QR) loss that regulates the relationship between the prompt\nquery and the prompt pool to improve representation power. The QR loss is only\napplied during training time, so there is no computational overhead at\ninference from the QR loss. With the QR loss, our approach maintains ~ 50%\ncomputational cost reduction during inference as well as outperforms the prior\ntwo-stage PCL methods by ~1.4% on public class-incremental continual learning\nbenchmarks including CIFAR-100, ImageNet-R, and DomainNet.\n", "date": "2024-02-25", "categories": []}
{"id": "2402.16674", "title": "ConSept: Continual Semantic Segmentation via Adapter-based Vision\n  Transformer", "abstract": "  In this paper, we delve into the realm of vision transformers for continual\nsemantic segmentation, a problem that has not been sufficiently explored in\nprevious literature. Empirical investigations on the adaptation of existing\nframeworks to vanilla ViT reveal that incorporating visual adapters into ViTs\nor fine-tuning ViTs with distillation terms is advantageous for enhancing the\nsegmentation capability of novel classes. These findings motivate us to propose\nContinual semantic Segmentation via Adapter-based ViT, namely ConSept. Within\nthe simplified architecture of ViT with linear segmentation head, ConSept\nintegrates lightweight attention-based adapters into vanilla ViTs. Capitalizing\non the feature adaptation abilities of these adapters, ConSept not only retains\nsuperior segmentation ability for old classes, but also attains promising\nsegmentation quality for novel classes. To further harness the intrinsic\nanti-catastrophic forgetting ability of ConSept and concurrently enhance the\nsegmentation capabilities for both old and new classes, we propose two key\nstrategies: distillation with a deterministic old-classes boundary for improved\nanti-catastrophic forgetting, and dual dice losses to regularize segmentation\nmaps, thereby improving overall segmentation performance. Extensive experiments\nshow the effectiveness of ConSept on multiple continual semantic segmentation\nbenchmarks under overlapped or disjoint settings. Code will be publicly\navailable at \\url{https://github.com/DongSky/ConSept}.\n", "date": "2024-02-26", "categories": []}
{"id": "2402.16767", "title": "CorpusBrain++: A Continual Generative Pre-Training Framework for\n  Knowledge-Intensive Language Tasks", "abstract": "  Knowledge-intensive language tasks (KILTs) typically require retrieving\nrelevant documents from trustworthy corpora, e.g., Wikipedia, to produce\nspecific answers. Very recently, a pre-trained generative retrieval model for\nKILTs, named CorpusBrain, was proposed and reached new state-of-the-art\nretrieval performance. However, most existing research on KILTs, including\nCorpusBrain, has predominantly focused on a static document collection,\noverlooking the dynamic nature of real-world scenarios, where new documents are\ncontinuously being incorporated into the source corpus. To address this gap, it\nis crucial to explore the capability of retrieval models to effectively handle\nthe dynamic retrieval scenario inherent in KILTs.\n  In this work, we first introduce the continual document learning (CDL) task\nfor KILTs and build a novel benchmark dataset named KILT++ based on the\noriginal KILT dataset for evaluation. Then, we conduct a comprehensive study\nover the use of pre-trained CorpusBrain on KILT++. Unlike the promising results\nin the stationary scenario, CorpusBrain is prone to catastrophic forgetting in\nthe dynamic scenario, hence hampering the retrieval performance. To alleviate\nthis issue, we propose CorpusBrain++, a continual generative pre-training\nframework. Empirical results demonstrate the significant effectiveness and\nremarkable efficiency of CorpusBrain++ in comparison to both traditional and\ngenerative IR methods.\n", "date": "2024-02-26", "categories": []}
{"id": "2402.15227", "title": "Fixed Random Classifier Rearrangement for Continual Learning", "abstract": "  With the explosive growth of data, continual learning capability is\nincreasingly important for neural networks. Due to catastrophic forgetting,\nneural networks inevitably forget the knowledge of old tasks after learning new\nones. In visual classification scenario, a common practice of alleviating the\nforgetting is to constrain the backbone. However, the impact of classifiers is\nunderestimated. In this paper, we analyze the variation of model predictions in\nsequential binary classification tasks and find that the norm of the equivalent\none-class classifiers significantly affects the forgetting level. Based on this\nconclusion, we propose a two-stage continual learning algorithm named Fixed\nRandom Classifier Rearrangement (FRCR). In first stage, FRCR replaces the\nlearnable classifiers with fixed random classifiers, constraining the norm of\nthe equivalent one-class classifiers without affecting the performance of the\nnetwork. In second stage, FRCR rearranges the entries of new classifiers to\nimplicitly reduce the drift of old latent representations. The experimental\nresults on multiple datasets show that FRCR significantly mitigates the model\nforgetting; subsequent experimental analyses further validate the effectiveness\nof the algorithm.\n", "date": "2024-02-23", "categories": []}
{"id": "2402.14155", "title": "Can Similarity-Based Domain-Ordering Reduce Catastrophic Forgetting for\n  Intent Recognition?", "abstract": "  Task-oriented dialogue systems are expected to handle a constantly expanding\nset of intents and domains even after they have been deployed to support more\nand more functionalities. To live up to this expectation, it becomes critical\nto mitigate the catastrophic forgetting problem (CF) that occurs in continual\nlearning (CL) settings for a task such as intent recognition. While existing\ndialogue systems research has explored replay-based and regularization-based\nmethods to this end, the effect of domain ordering on the CL performance of\nintent recognition models remains unexplored. If understood well, domain\nordering has the potential to be an orthogonal technique that can be leveraged\nalongside existing techniques such as experience replay. Our work fills this\ngap by comparing the impact of three domain-ordering strategies (min-sum path,\nmax-sum path, random) on the CL performance of a generative intent recognition\nmodel. Our findings reveal that the min-sum path strategy outperforms the\nothers in reducing catastrophic forgetting when training on the 220M T5-Base\nmodel. However, this advantage diminishes with the larger 770M T5-Large model.\nThese results underscores the potential of domain ordering as a complementary\nstrategy for mitigating catastrophic forgetting in continually learning intent\nrecognition models, particularly in resource-constrained scenarios.\n", "date": "2024-02-21", "categories": []}
{"id": "2402.12465", "title": "Neuro-mimetic Task-free Unsupervised Online Learning with Continual\n  Self-Organizing Maps", "abstract": "  An intelligent system capable of continual learning is one that can process\nand extract knowledge from potentially infinitely long streams of pattern\nvectors. The major challenge that makes crafting such a system difficult is\nknown as catastrophic forgetting - an agent, such as one based on artificial\nneural networks (ANNs), struggles to retain previously acquired knowledge when\nlearning from new samples. Furthermore, ensuring that knowledge is preserved\nfor previous tasks becomes more challenging when input is not supplemented with\ntask boundary information. Although forgetting in the context of ANNs has been\nstudied extensively, there still exists far less work investigating it in terms\nof unsupervised architectures such as the venerable self-organizing map (SOM),\na neural model often used in clustering and dimensionality reduction. While the\ninternal mechanisms of SOMs could, in principle, yield sparse representations\nthat improve memory retention, we observe that, when a fixed-size SOM processes\ncontinuous data streams, it experiences concept drift. In light of this, we\npropose a generalization of the SOM, the continual SOM (CSOM), which is capable\nof online unsupervised learning under a low memory budget. Our results, on\nbenchmarks including MNIST, Kuzushiji-MNIST, and Fashion-MNIST, show almost a\ntwo times increase in accuracy, and CIFAR-10 demonstrates a state-of-the-art\nresult when tested on (online) unsupervised class incremental learning setting.\n", "date": "2024-02-19", "categories": []}
{"id": "2402.12490", "title": "Towards Cross-Domain Continual Learning", "abstract": "  Continual learning is a process that involves training learning agents to\nsequentially master a stream of tasks or classes without revisiting past data.\nThe challenge lies in leveraging previously acquired knowledge to learn new\ntasks efficiently, while avoiding catastrophic forgetting. Existing methods\nprimarily focus on single domains, restricting their applicability to specific\nproblems.\n  In this work, we introduce a novel approach called Cross-Domain Continual\nLearning (CDCL) that addresses the limitations of being limited to single\nsupervised domains. Our method combines inter- and intra-task cross-attention\nmechanisms within a compact convolutional network. This integration enables the\nmodel to maintain alignment with features from previous tasks, thereby delaying\nthe data drift that may occur between tasks, while performing unsupervised\ncross-domain (UDA) between related domains. By leveraging an\nintra-task-specific pseudo-labeling method, we ensure accurate input pairs for\nboth labeled and unlabeled samples, enhancing the learning process. To validate\nour approach, we conduct extensive experiments on public UDA datasets,\nshowcasing its positive performance on cross-domain continual learning\nchallenges. Additionally, our work introduces incremental ideas that contribute\nto the advancement of this field.\n  We make our code and models available to encourage further exploration and\nreproduction of our results: \\url{https://github.com/Ivsucram/CDCL}\n", "date": "2024-02-19", "categories": []}
{"id": "2402.12624", "title": "Efficient Parameter Mining and Freezing for Continual Object Detection", "abstract": "  Continual Object Detection is essential for enabling intelligent agents to\ninteract proactively with humans in real-world settings. While\nparameter-isolation strategies have been extensively explored in the context of\ncontinual learning for classification, they have yet to be fully harnessed for\nincremental object detection scenarios. Drawing inspiration from prior research\nthat focused on mining individual neuron responses and integrating insights\nfrom recent developments in neural pruning, we proposed efficient ways to\nidentify which layers are the most important for a network to maintain the\nperformance of a detector across sequential updates. The presented findings\nhighlight the substantial advantages of layer-level parameter isolation in\nfacilitating incremental learning within object detection models, offering\npromising avenues for future research and application in real-world scenarios.\n", "date": "2024-02-20", "categories": []}
{"id": "2402.12722", "title": "Structural Knowledge Informed Continual Multivariate Time Series\n  Forecasting", "abstract": "  Recent studies in multivariate time series (MTS) forecasting reveal that\nexplicitly modeling the hidden dependencies among different time series can\nyield promising forecasting performance and reliable explanations. However,\nmodeling variable dependencies remains underexplored when MTS is continuously\naccumulated under different regimes (stages). Due to the potential distribution\nand dependency disparities, the underlying model may encounter the catastrophic\nforgetting problem, i.e., it is challenging to memorize and infer different\ntypes of variable dependencies across different regimes while maintaining\nforecasting performance. To address this issue, we propose a novel Structural\nKnowledge Informed Continual Learning (SKI-CL) framework to perform MTS\nforecasting within a continual learning paradigm, which leverages structural\nknowledge to steer the forecasting model toward identifying and adapting to\ndifferent regimes, and selects representative MTS samples from each regime for\nmemory replay. Specifically, we develop a forecasting model based on graph\nstructure learning, where a consistency regularization scheme is imposed\nbetween the learned variable dependencies and the structural knowledge while\noptimizing the forecasting objective over the MTS data. As such, MTS\nrepresentations learned in each regime are associated with distinct structural\nknowledge, which helps the model memorize a variety of conceivable scenarios\nand results in accurate forecasts in the continual learning context. Meanwhile,\nwe develop a representation-matching memory replay scheme that maximizes the\ntemporal coverage of MTS data to efficiently preserve the underlying temporal\ndynamics and dependency structures of each regime. Thorough empirical studies\non synthetic and real-world benchmarks validate SKI-CL's efficacy and\nadvantages over the state-of-the-art for continual MTS forecasting tasks.\n", "date": "2024-02-20", "categories": []}
{"id": "2402.12987", "title": "Towards Robust Graph Incremental Learning on Evolving Graphs", "abstract": "  Incremental learning is a machine learning approach that involves training a\nmodel on a sequence of tasks, rather than all tasks at once. This ability to\nlearn incrementally from a stream of tasks is crucial for many real-world\napplications. However, incremental learning is a challenging problem on\ngraph-structured data, as many graph-related problems involve prediction tasks\nfor each individual node, known as Node-wise Graph Incremental Learning (NGIL).\nThis introduces non-independent and non-identically distributed characteristics\nin the sample data generation process, making it difficult to maintain the\nperformance of the model as new tasks are added. In this paper, we focus on the\ninductive NGIL problem, which accounts for the evolution of graph structure\n(structural shift) induced by emerging tasks. We provide a formal formulation\nand analysis of the problem, and propose a novel regularization-based technique\ncalled Structural-Shift-Risk-Mitigation (SSRM) to mitigate the impact of the\nstructural shift on catastrophic forgetting of the inductive NGIL problem. We\nshow that the structural shift can lead to a shift in the input distribution\nfor the existing tasks, and further lead to an increased risk of catastrophic\nforgetting. Through comprehensive empirical studies with several benchmark\ndatasets, we demonstrate that our proposed method,\nStructural-Shift-Risk-Mitigation (SSRM), is flexible and easy to adapt to\nimprove the performance of state-of-the-art GNN incremental learning frameworks\nin the inductive setting.\n", "date": "2024-02-20", "categories": []}
{"id": "2309.08227", "title": "VERSE: Virtual-Gradient Aware Streaming Lifelong Learning with Anytime\n  Inference", "abstract": "  Lifelong learning or continual learning is the problem of training an AI\nagent continuously while also preventing it from forgetting its previously\nacquired knowledge. Streaming lifelong learning is a challenging setting of\nlifelong learning with the goal of continuous learning in a dynamic\nnon-stationary environment without forgetting. We introduce a novel approach to\nlifelong learning, which is streaming (observes each training example only\nonce), requires a single pass over the data, can learn in a class-incremental\nmanner, and can be evaluated on-the-fly (anytime inference). To accomplish\nthese, we propose a novel \\emph{virtual gradients} based approach for continual\nrepresentation learning which adapts to each new example while also\ngeneralizing well on past data to prevent catastrophic forgetting. Our approach\nalso leverages an exponential-moving-average-based semantic memory to further\nenhance performance. Experiments on diverse datasets with temporally correlated\nobservations demonstrate our method's efficacy and superior performance over\nexisting methods.\n", "date": "2023-09-15", "categories": []}
{"id": "2309.10479", "title": "RECALL+: Adversarial Web-based Replay for Continual Learning in Semantic\n  Segmentation", "abstract": "  Catastrophic forgetting of previous knowledge is a critical issue in\ncontinual learning typically handled through various regularization strategies.\nHowever, existing methods struggle especially when several incremental steps\nare performed. In this paper, we extend our previous approach (RECALL) and\ntackle forgetting by exploiting unsupervised web-crawled data to retrieve\nexamples of old classes from online databases. In contrast to the original\nmethodology, which did not incorporate an assessment of web-based data, the\npresent work proposes two advanced techniques: an adversarial approach and an\nadaptive threshold strategy. These methods are utilized to meticulously choose\nsamples from web data that exhibit strong statistical congruence with the no\nlonger available training data. Furthermore, we improved the pseudo-labeling\nscheme to achieve a more accurate labeling of web data that also considers\nclasses being learned in the current step. Experimental results show that this\nenhanced approach achieves remarkable results, particularly when the\nincremental scenario spans multiple steps.\n", "date": "2023-09-19", "categories": []}
{"id": "2312.13027", "title": "Doubly Perturbed Task Free Continual Learning", "abstract": "  Task Free online continual learning (TF-CL) is a challenging problem where\nthe model incrementally learns tasks without explicit task information.\nAlthough training with entire data from the past, present as well as future is\nconsidered as the gold standard, naive approaches in TF-CL with the current\nsamples may be conflicted with learning with samples in the future, leading to\ncatastrophic forgetting and poor plasticity. Thus, a proactive consideration of\nan unseen future sample in TF-CL becomes imperative. Motivated by this\nintuition, we propose a novel TF-CL framework considering future samples and\nshow that injecting adversarial perturbations on both input data and\ndecision-making is effective. Then, we propose a novel method named Doubly\nPerturbed Continual Learning (DPCL) to efficiently implement these input and\ndecision-making perturbations. Specifically, for input perturbation, we propose\nan approximate perturbation method that injects noise into the input data as\nwell as the feature vector and then interpolates the two perturbed samples. For\ndecision-making process perturbation, we devise multiple stochastic\nclassifiers. We also investigate a memory management scheme and learning rate\nscheduling reflecting our proposed double perturbations. We demonstrate that\nour proposed method outperforms the state-of-the-art baseline methods by large\nmargins on various TF-CL benchmarks.\n", "date": "2023-12-20", "categories": []}
{"id": "2402.11565", "title": "Continual Learning on Graphs: Challenges, Solutions, and Opportunities", "abstract": "  Continual learning on graph data has recently attracted paramount attention\nfor its aim to resolve the catastrophic forgetting problem on existing tasks\nwhile adapting the sequentially updated model to newly emerged graph tasks.\nWhile there have been efforts to summarize progress on continual learning\nresearch over Euclidean data, e.g., images and texts, a systematic review of\nprogress in continual learning on graphs, a.k.a, continual graph learning (CGL)\nor lifelong graph learning, is still demanding. Graph data are far more complex\nin terms of data structures and application scenarios, making CGL task\nsettings, model designs, and applications extremely challenging. To bridge the\ngap, we provide a comprehensive review of existing continual graph learning\n(CGL) algorithms by elucidating the different task settings and categorizing\nthe existing methods based on their characteristics. We compare the CGL methods\nwith traditional continual learning techniques and analyze the applicability of\nthe traditional continual learning techniques to CGL tasks. Additionally, we\nreview the benchmark works that are crucial to CGL research. Finally, we\ndiscuss the remaining challenges and propose several future directions. We will\nmaintain an up-to-date GitHub repository featuring a comprehensive list of CGL\nalgorithms, accessible at\nhttps://github.com/UConn-DSIS/Survey-of-Continual-Learning-on-Graphs.\n", "date": "2024-02-18", "categories": []}
{"id": "2306.03364", "title": "Learning Representations on the Unit Sphere: Investigating Angular\n  Gaussian and von Mises-Fisher Distributions for Online Continual Learning", "abstract": "  We use the maximum a posteriori estimation principle for learning\nrepresentations distributed on the unit sphere. We propose to use the angular\nGaussian distribution, which corresponds to a Gaussian projected on the\nunit-sphere and derive the associated loss function. We also consider the von\nMises-Fisher distribution, which is the conditional of a Gaussian in the\nunit-sphere. The learned representations are pushed toward fixed directions,\nwhich are the prior means of the Gaussians; allowing for a learning strategy\nthat is resilient to data drift. This makes it suitable for online continual\nlearning, which is the problem of training neural networks on a continuous data\nstream, where multiple classification tasks are presented sequentially so that\ndata from past tasks are no longer accessible, and data from the current task\ncan be seen only once. To address this challenging scenario, we propose a\nmemory-based representation learning technique equipped with our new loss\nfunctions. Our approach does not require negative data or knowledge of task\nboundaries and performs well with smaller batch sizes while being\ncomputationally efficient. We demonstrate with extensive experiments that the\nproposed method outperforms the current state-of-the-art methods on both\nstandard evaluation scenarios and realistic scenarios with blurry task\nboundaries. For reproducibility, we use the same training pipeline for every\ncompared method and share the code at https://github.com/Nicolas1203/ocl-fd.\n", "date": "2023-06-06", "categories": []}
{"id": "2402.10427", "title": "Evaluating and Improving Continual Learning in Spoken Language\n  Understanding", "abstract": "  Continual learning has emerged as an increasingly important challenge across\nvarious tasks, including Spoken Language Understanding (SLU). In SLU, its\nobjective is to effectively handle the emergence of new concepts and evolving\nenvironments. The evaluation of continual learning algorithms typically\ninvolves assessing the model's stability, plasticity, and generalizability as\nfundamental aspects of standards. However, existing continual learning metrics\nprimarily focus on only one or two of the properties. They neglect the overall\nperformance across all tasks, and do not adequately disentangle the plasticity\nversus stability/generalizability trade-offs within the model. In this work, we\npropose an evaluation methodology that provides a unified evaluation on\nstability, plasticity, and generalizability in continual learning. By employing\nthe proposed metric, we demonstrate how introducing various knowledge\ndistillations can improve different aspects of these three properties of the\nSLU model. We further show that our proposed metric is more sensitive in\ncapturing the impact of task ordering in continual learning, making it better\nsuited for practical use-case scenarios.\n", "date": "2024-02-16", "categories": []}
{"id": "2211.16234", "title": "SimCS: Simulation for Domain Incremental Online Continual Segmentation", "abstract": "  Continual Learning is a step towards lifelong intelligence where models\ncontinuously learn from recently collected data without forgetting previous\nknowledge. Existing continual learning approaches mostly focus on image\nclassification in the class-incremental setup with clear task boundaries and\nunlimited computational budget. This work explores the problem of Online\nDomain-Incremental Continual Segmentation (ODICS), where the model is\ncontinually trained over batches of densely labeled images from different\ndomains, with limited computation and no information about the task boundaries.\nODICS arises in many practical applications. In autonomous driving, this may\ncorrespond to the realistic scenario of training a segmentation model over time\non a sequence of cities. We analyze several existing continual learning methods\nand show that they perform poorly in this setting despite working well in\nclass-incremental segmentation. We propose SimCS, a parameter-free method\ncomplementary to existing ones that uses simulated data to regularize continual\nlearning. Experiments show that SimCS provides consistent improvements when\ncombined with different CL methods.\n", "date": "2022-11-29", "categories": []}
{"id": "2402.10063", "title": "Balancing the Causal Effects in Class-Incremental Learning", "abstract": "  Class-Incremental Learning (CIL) is a practical and challenging problem for\nachieving general artificial intelligence. Recently, Pre-Trained Models (PTMs)\nhave led to breakthroughs in both visual and natural language processing tasks.\nDespite recent studies showing PTMs' potential ability to learn sequentially, a\nplethora of work indicates the necessity of alleviating the catastrophic\nforgetting of PTMs. Through a pilot study and a causal analysis of CIL, we\nreveal that the crux lies in the imbalanced causal effects between new and old\ndata. Specifically, the new data encourage models to adapt to new classes while\nhindering the adaptation of old classes. Similarly, the old data encourages\nmodels to adapt to old classes while hindering the adaptation of new classes.\nIn other words, the adaptation process between new and old classes conflicts\nfrom the causal perspective. To alleviate this problem, we propose Balancing\nthe Causal Effects (BaCE) in CIL. Concretely, BaCE proposes two objectives for\nbuilding causal paths from both new and old data to the prediction of new and\nclasses, respectively. In this way, the model is encouraged to adapt to all\nclasses with causal effects from both new and old data and thus alleviates the\ncausal imbalance problem. We conduct extensive experiments on continual image\nclassification, continual text classification, and continual named entity\nrecognition. Empirical results show that BaCE outperforms a series of CIL\nmethods on different tasks and settings.\n", "date": "2024-02-15", "categories": []}
{"id": "2402.10130", "title": "Is Continual Learning Ready for Real-world Challenges?", "abstract": "  Despite continual learning's long and well-established academic history, its\napplication in real-world scenarios remains rather limited. This paper contends\nthat this gap is attributable to a misalignment between the actual challenges\nof continual learning and the evaluation protocols in use, rendering proposed\nsolutions ineffective for addressing the complexities of real-world setups. We\nvalidate our hypothesis and assess progress to date, using a new 3D semantic\nsegmentation benchmark, OCL-3DSS. We investigate various continual learning\nschemes from the literature by utilizing more realistic protocols that\nnecessitate online and continual learning for dynamic, real-world scenarios\n(eg., in robotics and 3D vision applications). The outcomes are sobering: all\nconsidered methods perform poorly, significantly deviating from the upper bound\nof joint offline training. This raises questions about the applicability of\nexisting methods in realistic settings. Our paper aims to initiate a paradigm\nshift, advocating for the adoption of continual learning methods through new\nexperimental protocols that better emulate real-world conditions to facilitate\nbreakthroughs in the field.\n", "date": "2024-02-15", "categories": []}
{"id": "2402.08255", "title": "Distal Interference: Exploring the Limits of Model-Based Continual\n  Learning", "abstract": "  Continual learning is the sequential learning of different tasks by a machine\nlearning model. Continual learning is known to be hindered by catastrophic\ninterference or forgetting, i.e. rapid unlearning of earlier learned tasks when\nnew tasks are learned. Despite their practical success, artificial neural\nnetworks (ANNs) are prone to catastrophic interference. This study analyses how\ngradient descent and overlapping representations between distant input points\nlead to distal interference and catastrophic interference. Distal interference\nrefers to the phenomenon where training a model on a subset of the domain leads\nto non-local changes on other subsets of the domain. This study shows that\nuniformly trainable models without distal interference must be exponentially\nlarge. A novel antisymmetric bounded exponential layer B-spline ANN\narchitecture named ABEL-Spline is proposed that can approximate any continuous\nfunction, is uniformly trainable, has polynomial computational complexity, and\nprovides some guarantees for distal interference. Experiments are presented to\ndemonstrate the theoretical properties of ABEL-Splines. ABEL-Splines are also\nevaluated on benchmark regression problems. It is concluded that the weaker\ndistal interference guarantees in ABEL-Splines are insufficient for model-only\ncontinual learning. It is conjectured that continual learning with polynomial\ncomplexity models requires augmentation of the training data or algorithm.\n", "date": "2024-02-13", "categories": []}
{"id": "2402.07216", "title": "A novel spatial-frequency domain network for zero-shot incremental\n  learning", "abstract": "  Zero-shot incremental learning aims to enable the model to generalize to new\nclasses without forgetting previously learned classes. However, the semantic\ngap between old and new sample classes can lead to catastrophic forgetting.\nAdditionally, existing algorithms lack capturing significant information from\neach sample image domain, impairing models' classification performance.\nTherefore, this paper proposes a novel Spatial-Frequency Domain Network\n(SFDNet) which contains a Spatial-Frequency Feature Extraction (SFFE) module\nand Attention Feature Alignment (AFA) module to improve the Zero-Shot\nTranslation for Class Incremental algorithm. Firstly, SFFE module is designed\nwhich contains a dual attention mechanism for obtaining salient\nspatial-frequency feature information. Secondly, a novel feature fusion module\nis conducted for obtaining fused spatial-frequency domain features. Thirdly,\nthe Nearest Class Mean classifier is utilized to select the most suitable\ncategory. Finally, iteration between tasks is performed using the Zero-Shot\nTranslation model. The proposed SFDNet has the ability to effectively extract\nspatial-frequency feature representation from input images, improve the\naccuracy of image classification, and fundamentally alleviate catastrophic\nforgetting. Extensive experiments on the CUB 200-2011 and CIFAR100 datasets\ndemonstrate that our proposed algorithm outperforms state-of-the-art\nincremental learning algorithms.\n", "date": "2024-02-11", "categories": []}
{"id": "2402.06330", "title": "Continual Learning on Graphs: A Survey", "abstract": "  Recently, continual graph learning has been increasingly adopted for diverse\ngraph-structured data processing tasks in non-stationary environments. Despite\nits promising learning capability, current studies on continual graph learning\nmainly focus on mitigating the catastrophic forgetting problem while ignoring\ncontinuous performance improvement. To bridge this gap, this article aims to\nprovide a comprehensive survey of recent efforts on continual graph learning.\nSpecifically, we introduce a new taxonomy of continual graph learning from the\nperspective of overcoming catastrophic forgetting. Moreover, we systematically\nanalyze the challenges of applying these continual graph learning methods in\nimproving performance continuously and then discuss the possible solutions.\nFinally, we present open issues and future directions pertaining to the\ndevelopment of continual graph learning and discuss how they impact continuous\nperformance improvement.\n", "date": "2024-02-09", "categories": []}
{"id": "2402.05860", "title": "Privacy-Preserving Synthetic Continual Semantic Segmentation for Robotic\n  Surgery", "abstract": "  Deep Neural Networks (DNNs) based semantic segmentation of the robotic\ninstruments and tissues can enhance the precision of surgical activities in\nrobot-assisted surgery. However, in biological learning, DNNs cannot learn\nincremental tasks over time and exhibit catastrophic forgetting, which refers\nto the sharp decline in performance on previously learned tasks after learning\na new one. Specifically, when data scarcity is the issue, the model shows a\nrapid drop in performance on previously learned instruments after learning new\ndata with new instruments. The problem becomes worse when it limits releasing\nthe dataset of the old instruments for the old model due to privacy concerns\nand the unavailability of the data for the new or updated version of the\ninstruments for the continual learning model. For this purpose, we develop a\nprivacy-preserving synthetic continual semantic segmentation framework by\nblending and harmonizing (i) open-source old instruments foreground to the\nsynthesized background without revealing real patient data in public and (ii)\nnew instruments foreground to extensively augmented real background. To boost\nthe balanced logit distillation from the old model to the continual learning\nmodel, we design overlapping class-aware temperature normalization (CAT) by\ncontrolling model learning utility. We also introduce multi-scale\nshifted-feature distillation (SD) to maintain long and short-range spatial\nrelationships among the semantic objects where conventional short-range spatial\nfeatures with limited information reduce the power of feature distillation. We\ndemonstrate the effectiveness of our framework on the EndoVis 2017 and 2018\ninstrument segmentation dataset with a generalized continual learning setting.\nCode is available at~\\url{https://github.com/XuMengyaAmy/Synthetic_CAT_SD}.\n", "date": "2024-02-08", "categories": []}
{"id": "2303.17235", "title": "Kaizen: Practical Self-supervised Continual Learning with Continual\n  Fine-tuning", "abstract": "  Self-supervised learning (SSL) has shown remarkable performance in computer\nvision tasks when trained offline. However, in a Continual Learning (CL)\nscenario where new data is introduced progressively, models still suffer from\ncatastrophic forgetting. Retraining a model from scratch to adapt to newly\ngenerated data is time-consuming and inefficient. Previous approaches suggested\nre-purposing self-supervised objectives with knowledge distillation to mitigate\nforgetting across tasks, assuming that labels from all tasks are available\nduring fine-tuning. In this paper, we generalize self-supervised continual\nlearning in a practical setting where available labels can be leveraged in any\nstep of the SSL process. With an increasing number of continual tasks, this\noffers more flexibility in the pre-training and fine-tuning phases. With\nKaizen, we introduce a training architecture that is able to mitigate\ncatastrophic forgetting for both the feature extractor and classifier with a\ncarefully designed loss function. By using a set of comprehensive evaluation\nmetrics reflecting different aspects of continual learning, we demonstrated\nthat Kaizen significantly outperforms previous SSL models in competitive vision\nbenchmarks, with up to 16.5% accuracy improvement on split CIFAR-100. Kaizen is\nable to balance the trade-off between knowledge retention and learning from new\ndata with an end-to-end model, paving the way for practical deployment of\ncontinual learning systems.\n", "date": "2023-03-30", "categories": []}
{"id": "2305.18441", "title": "DeCoR: Defy Knowledge Forgetting by Predicting Earlier Audio Codes", "abstract": "  Lifelong audio feature extraction involves learning new sound classes\nincrementally, which is essential for adapting to new data distributions over\ntime. However, optimizing the model only on new data can lead to catastrophic\nforgetting of previously learned tasks, which undermines the model's ability to\nperform well over the long term. This paper introduces a new approach to\ncontinual audio representation learning called DeCoR. Unlike other methods that\nstore previous data, features, or models, DeCoR indirectly distills knowledge\nfrom an earlier model to the latest by predicting quantization indices from a\ndelayed codebook. We demonstrate that DeCoR improves acoustic scene\nclassification accuracy and integrates well with continual self-supervised\nrepresentation learning. Our approach introduces minimal storage and\ncomputation overhead, making it a lightweight and efficient solution for\ncontinual learning.\n", "date": "2023-05-29", "categories": []}
{"id": "2402.01364", "title": "Continual Learning for Large Language Models: A Survey", "abstract": "  Large language models (LLMs) are not amenable to frequent re-training, due to\nhigh training costs arising from their massive scale. However, updates are\nnecessary to endow LLMs with new skills and keep them up-to-date with rapidly\nevolving human knowledge. This paper surveys recent works on continual learning\nfor LLMs. Due to the unique nature of LLMs, we catalog continue learning\ntechniques in a novel multi-staged categorization scheme, involving continual\npretraining, instruction tuning, and alignment. We contrast continual learning\nfor LLMs with simpler adaptation methods used in smaller models, as well as\nwith other enhancement strategies like retrieval-augmented generation and model\nediting. Moreover, informed by a discussion of benchmarks and evaluation, we\nidentify several challenges and future work directions for this crucial task.\n", "date": "2024-02-02", "categories": []}
{"id": "2402.04596", "title": "Towards Improved Imbalance Robustness in Continual Multi-Label Learning\n  with Dual Output Spiking Architecture (DOSA)", "abstract": "  Algorithms designed for addressing typical supervised classification problems\ncan only learn from a fixed set of samples and labels, making them unsuitable\nfor the real world, where data arrives as a stream of samples often associated\nwith multiple labels over time. This motivates the study of task-agnostic\ncontinual multi-label learning problems. While algorithms using deep learning\napproaches for continual multi-label learning have been proposed in the recent\nliterature, they tend to be computationally heavy. Although spiking neural\nnetworks (SNNs) offer a computationally efficient alternative to artificial\nneural networks, existing literature has not used SNNs for continual\nmulti-label learning. Also, accurately determining multiple labels with SNNs is\nstill an open research problem. This work proposes a dual output spiking\narchitecture (DOSA) to bridge these research gaps. A novel imbalance-aware loss\nfunction is also proposed, improving the multi-label classification performance\nof the model by making it more robust to data imbalance. A modified F1 score is\npresented to evaluate the effectiveness of the proposed loss function in\nhandling imbalance. Experiments on several benchmark multi-label datasets show\nthat DOSA trained with the proposed loss function shows improved robustness to\ndata imbalance and obtains better continual multi-label learning performance\nthan CIFDM, a previous state-of-the-art algorithm.\n", "date": "2024-02-07", "categories": []}
{"id": "2302.00487", "title": "A Comprehensive Survey of Continual Learning: Theory, Method and\n  Application", "abstract": "  To cope with real-world dynamics, an intelligent system needs to\nincrementally acquire, update, accumulate, and exploit knowledge throughout its\nlifetime. This ability, known as continual learning, provides a foundation for\nAI systems to develop themselves adaptively. In a general sense, continual\nlearning is explicitly limited by catastrophic forgetting, where learning a new\ntask usually results in a dramatic performance degradation of the old tasks.\nBeyond this, increasingly numerous advances have emerged in recent years that\nlargely extend the understanding and application of continual learning. The\ngrowing and widespread interest in this direction demonstrates its realistic\nsignificance as well as complexity. In this work, we present a comprehensive\nsurvey of continual learning, seeking to bridge the basic settings, theoretical\nfoundations, representative methods, and practical applications. Based on\nexisting theoretical and empirical results, we summarize the general objectives\nof continual learning as ensuring a proper stability-plasticity trade-off and\nan adequate intra/inter-task generalizability in the context of resource\nefficiency. Then we provide a state-of-the-art and elaborated taxonomy,\nextensively analyzing how representative methods address continual learning,\nand how they are adapted to particular challenges in realistic applications.\nThrough an in-depth discussion of promising directions, we believe that such a\nholistic perspective can greatly facilitate subsequent exploration in this\nfield and beyond.\n", "date": "2023-01-31", "categories": []}
{"id": "2307.09653", "title": "HAT-CL: A Hard-Attention-to-the-Task PyTorch Library for Continual\n  Learning", "abstract": "  Catastrophic forgetting, the phenomenon in which a neural network loses\npreviously obtained knowledge during the learning of new tasks, poses a\nsignificant challenge in continual learning. The Hard-Attention-to-the-Task\n(HAT) mechanism has shown potential in mitigating this problem, but its\npractical implementation has been complicated by issues of usability and\ncompatibility, and a lack of support for existing network reuse. In this paper,\nwe introduce HAT-CL, a user-friendly, PyTorch-compatible redesign of the HAT\nmechanism. HAT-CL not only automates gradient manipulation but also streamlines\nthe transformation of PyTorch modules into HAT modules. It achieves this by\nproviding a comprehensive suite of modules that can be seamlessly integrated\ninto existing architectures. Additionally, HAT-CL offers ready-to-use HAT\nnetworks that are smoothly integrated with the TIMM library. Beyond the\nredesign and reimplementation of HAT, we also introduce novel mask manipulation\ntechniques for HAT, which have consistently shown improvements across various\nexperiments. Our work paves the way for a broader application of the HAT\nmechanism, opening up new possibilities in continual learning across diverse\nmodels and applications.\n", "date": "2023-07-18", "categories": []}
{"id": "2402.01201", "title": "Few-Shot Class-Incremental Learning with Prior Knowledge", "abstract": "  To tackle the issues of catastrophic forgetting and overfitting in few-shot\nclass-incremental learning (FSCIL), previous work has primarily concentrated on\npreserving the memory of old knowledge during the incremental phase. The role\nof pre-trained model in shaping the effectiveness of incremental learning is\nfrequently underestimated in these studies. Therefore, to enhance the\ngeneralization ability of the pre-trained model, we propose Learning with Prior\nKnowledge (LwPK) by introducing nearly free prior knowledge from a few\nunlabeled data of subsequent incremental classes. We cluster unlabeled\nincremental class samples to produce pseudo-labels, then jointly train these\nwith labeled base class samples, effectively allocating embedding space for\nboth old and new class data. Experimental results indicate that LwPK\neffectively enhances the model resilience against catastrophic forgetting, with\ntheoretical analysis based on empirical risk minimization and class distance\nmeasurement corroborating its operational principles. The source code of LwPK\nis publicly available at: \\url{https://github.com/StevenJ308/LwPK}.\n", "date": "2024-02-02", "categories": []}
{"id": "2402.00481", "title": "Bias Mitigating Few-Shot Class-Incremental Learning", "abstract": "  Few-shot class-incremental learning (FSCIL) aims at recognizing novel classes\ncontinually with limited novel class samples. A mainstream baseline for FSCIL\nis first to train the whole model in the base session, then freeze the feature\nextractor in the incremental sessions. Despite achieving high overall accuracy,\nmost methods exhibit notably low accuracy for incremental classes. Some recent\nmethods somewhat alleviate the accuracy imbalance between base and incremental\nclasses by fine-tuning the feature extractor in the incremental sessions, but\nthey further cause the accuracy imbalance between past and current incremental\nclasses. In this paper, we study the causes of such classification accuracy\nimbalance for FSCIL, and abstract them into a unified model bias problem. Based\non the analyses, we propose a novel method to mitigate model bias of the FSCIL\nproblem during training and inference processes, which includes mapping ability\nstimulation, separately dual-feature classification, and self-optimizing\nclassifiers. Extensive experiments on three widely-used FSCIL benchmark\ndatasets show that our method significantly mitigates the model bias problem\nand achieves state-of-the-art performance.\n", "date": "2024-02-01", "categories": []}
{"id": "2402.00580", "title": "Continuous Unsupervised Domain Adaptation Using Stabilized\n  Representations and Experience Replay", "abstract": "  We introduce an algorithm for tackling the problem of unsupervised domain\nadaptation (UDA) in continual learning (CL) scenarios. The primary objective is\nto maintain model generalization under domain shift when new domains arrive\ncontinually through updating a base model when only unlabeled data is\naccessible in subsequent tasks. While there are many existing UDA algorithms,\nthey typically require access to both the source and target domain datasets\nsimultaneously. Conversely, existing CL approaches can handle tasks that all\nhave labeled data. Our solution is based on stabilizing the learned internal\ndistribution to enhances the model generalization on new domains. The internal\ndistribution is modeled by network responses in hidden layer. We model this\ninternal distribution using a Gaussian mixture model (GMM ) and update the\nmodel by matching the internally learned distribution of new domains to the\nestimated GMM. Additionally, we leverage experience replay to overcome the\nproblem of catastrophic forgetting, where the model loses previously acquired\nknowledge when learning new tasks. We offer theoretical analysis to explain why\nour algorithm would work. We also offer extensive comparative and analytic\nexperiments to demonstrate that our method is effective. We perform experiments\non four benchmark datasets to demonstrate that our approach is effective.\n", "date": "2024-01-31", "categories": []}
{"id": "2401.17401", "title": "Step-size Optimization for Continual Learning", "abstract": "  In continual learning, a learner has to keep learning from the data over its\nwhole life time. A key issue is to decide what knowledge to keep and what\nknowledge to let go. In a neural network, this can be implemented by using a\nstep-size vector to scale how much gradient samples change network weights.\nCommon algorithms, like RMSProp and Adam, use heuristics, specifically\nnormalization, to adapt this step-size vector. In this paper, we show that\nthose heuristics ignore the effect of their adaptation on the overall objective\nfunction, for example by moving the step-size vector away from better step-size\nvectors. On the other hand, stochastic meta-gradient descent algorithms, like\nIDBD (Sutton, 1992), explicitly optimize the step-size vector with respect to\nthe overall objective function. On simple problems, we show that IDBD is able\nto consistently improve step-size vectors, where RMSProp and Adam do not. We\nexplain the differences between the two approaches and their respective\nlimitations. We conclude by suggesting that combining both approaches could be\na promising future direction to improve the performance of neural networks in\ncontinual learning.\n", "date": "2024-01-30", "categories": []}
{"id": "2305.03923", "title": "Active Continual Learning: On Balancing Knowledge Retention and\n  Learnability", "abstract": "  Acquiring new knowledge without forgetting what has been learned in a\nsequence of tasks is the central focus of continual learning (CL). While tasks\narrive sequentially, the training data are often prepared and annotated\nindependently, leading to the CL of incoming supervised learning tasks. This\npaper considers the under-explored problem of active continual learning (ACL)\nfor a sequence of active learning (AL) tasks, where each incoming task includes\na pool of unlabelled data and an annotation budget. We investigate the\neffectiveness and interplay between several AL and CL algorithms in the domain,\nclass and task-incremental scenarios. Our experiments reveal the trade-off\nbetween two contrasting goals of not forgetting the old knowledge and the\nability to quickly learn new knowledge in CL and AL, respectively. While\nconditioning the AL query strategy on the annotations collected for the\nprevious tasks leads to improved task performance on the domain and task\nincremental learning, our proposed forgetting-learning profile suggests a gap\nin balancing the effect of AL and CL for the class-incremental scenario.\n", "date": "2023-05-06", "categories": []}
{"id": "2401.15275", "title": "Dynamic Transformer Architecture for Continual Learning of Multimodal\n  Tasks", "abstract": "  Transformer neural networks are increasingly replacing prior architectures in\na wide range of applications in different data modalities. The increasing size\nand computational demands of fine-tuning large pre-trained transformer neural\nnetworks pose significant challenges for the widespread adoption of these\nmodels for applications that demand on-edge computing. To tackle this\nchallenge, continual learning (CL) emerges as a solution by facilitating the\ntransfer of knowledge across tasks that arrive sequentially for an autonomously\nlearning agent. However, current CL methods mainly focus on learning tasks that\nare exclusively vision-based or language-based. We propose a transformer-based\nCL framework focusing on learning tasks that involve both vision and language,\nknown as Vision-and-Language (VaL) tasks. Due to the success of transformers in\nother modalities, our architecture has the potential to be used in multimodal\nlearning settings. In our framework, we benefit from introducing extra\nparameters to a base transformer to specialize the network for each task. As a\nresult, we enable dynamic model expansion to learn several tasks in a sequence.\nWe also use knowledge distillation to benefit from relevant past experiences to\nlearn the current task more efficiently. Our proposed method, Task Attentive\nMultimodal Continual Learning (TAM-CL), allows for the exchange of information\nbetween tasks while mitigating the problem of catastrophic forgetting. Notably,\nour approach is scalable, incurring minimal memory and time overhead. TAM-CL\nachieves state-of-the-art (SOTA) performance on challenging multimodal tasks\n", "date": "2024-01-27", "categories": []}
{"id": "2401.15973", "title": "Sample Weight Estimation Using Meta-Updates for Online Continual\n  Learning", "abstract": "  The loss function plays an important role in optimizing the performance of a\nlearning system. A crucial aspect of the loss function is the assignment of\nsample weights within a mini-batch during loss computation. In the context of\ncontinual learning (CL), most existing strategies uniformly treat samples when\ncalculating the loss value, thereby assigning equal weights to each sample.\nWhile this approach can be effective in certain standard benchmarks, its\noptimal effectiveness, particularly in more complex scenarios, remains\nunderexplored. This is particularly pertinent in training \"in the wild,\" such\nas with self-training, where labeling is automated using a reference model.\nThis paper introduces the Online Meta-learning for Sample Importance (OMSI)\nstrategy that approximates sample weights for a mini-batch in an online CL\nstream using an inner- and meta-update mechanism. This is done by first\nestimating sample weight parameters for each sample in the mini-batch, then,\nupdating the model with the adapted sample weights. We evaluate OMSI in two\ndistinct experimental settings. First, we show that OMSI enhances both learning\nand retained accuracy in a controlled noisy-labeled data stream. Then, we test\nthe strategy in three standard benchmarks and compare it with other popular\nreplay-based strategies. This research aims to foster the ongoing exploration\nin the area of self-adaptive CL.\n", "date": "2024-01-29", "categories": []}
{"id": "2309.14808", "title": "Revisiting Softmax Masking: Stop Gradient for Enhancing Stability in\n  Replay-based Continual Learning", "abstract": "  In replay-based methods for continual learning, replaying input samples in\nepisodic memory has shown its effectiveness in alleviating catastrophic\nforgetting. However, the potential key factor of cross-entropy loss with\nsoftmax in causing catastrophic forgetting has been underexplored. In this\npaper, we analyze the effect of softmax and revisit softmax masking with\nnegative infinity to shed light on its ability to mitigate catastrophic\nforgetting. Based on the analyses, it is found that negative infinity masked\nsoftmax is not always compatible with dark knowledge. To improve the\ncompatibility, we propose a general masked softmax that controls the stability\nby adjusting the gradient scale to old and new classes. We demonstrate that\nutilizing our method on other replay-based methods results in better\nperformance, primarily by enhancing model stability in continual learning\nbenchmarks, even when the buffer size is set to an extremely small value.\n", "date": "2023-09-26", "categories": []}
{"id": "2401.12617", "title": "The Joint Effect of Task Similarity and Overparameterization on\n  Catastrophic Forgetting -- An Analytical Model", "abstract": "  In continual learning, catastrophic forgetting is affected by multiple\naspects of the tasks. Previous works have analyzed separately how forgetting is\naffected by either task similarity or overparameterization. In contrast, our\npaper examines how task similarity and overparameterization jointly affect\nforgetting in an analyzable model. Specifically, we focus on two-task continual\nlinear regression, where the second task is a random orthogonal transformation\nof an arbitrary first task (an abstraction of random permutation tasks). We\nderive an exact analytical expression for the expected forgetting - and uncover\na nuanced pattern. In highly overparameterized models, intermediate task\nsimilarity causes the most forgetting. However, near the interpolation\nthreshold, forgetting decreases monotonically with the expected task\nsimilarity. We validate our findings with linear regression on synthetic data,\nand with neural networks on established permutation task benchmarks.\n", "date": "2024-01-23", "categories": []}
{"id": "2210.00044", "title": "Task Formulation Matters When Learning Continually: A Case Study in\n  Visual Question Answering", "abstract": "  Continual learning aims to train a model incrementally on a sequence of tasks\nwithout forgetting previous knowledge. Although continual learning has been\nwidely studied in computer vision, its application to Vision+Language tasks is\nnot that straightforward, as settings can be parameterized in multiple ways\naccording to their input modalities. In this paper, we present a detailed study\nof how different settings affect performance for Visual Question Answering. We\nfirst propose three plausible task formulations and demonstrate their impact on\nthe performance of continual learning algorithms. We break down several factors\nof task similarity, showing that performance and sensitivity to task order\nhighly depend on the shift of the output distribution. We also investigate the\npotential of pretrained models and compare the robustness of transformer models\nwith different visual embeddings. Finally, we provide an analysis interpreting\nmodel representations and their impact on forgetting. Our results highlight the\nimportance of stabilizing visual representations in deeper layers.\n", "date": "2022-09-30", "categories": []}
{"id": "2401.11544", "title": "Hierarchical Prompts for Rehearsal-free Continual Learning", "abstract": "  Continual learning endeavors to equip the model with the capability to\nintegrate current task knowledge while mitigating the forgetting of past task\nknowledge. Inspired by prompt tuning, prompt-based methods maintain a frozen\nbackbone and train with slight learnable prompts to minimize the catastrophic\nforgetting that arises due to updating a large number of backbone parameters.\nNonetheless, these learnable prompts tend to concentrate on the discriminatory\nknowledge of the current task while ignoring past task knowledge, leading to\nthat learnable prompts still suffering from catastrophic forgetting. This paper\nintroduces a novel rehearsal-free paradigm for continual learning termed\nHierarchical Prompts (H-Prompts), comprising three categories of prompts --\nclass prompt, task prompt, and general prompt. To effectively depict the\nknowledge of past classes, class prompt leverages Bayesian Distribution\nAlignment to model the distribution of classes in each task. To reduce the\nforgetting of past task knowledge, task prompt employs Cross-task Knowledge\nExcavation to amalgamate the knowledge encapsulated in the learned class\nprompts of past tasks and current task knowledge. Furthermore, general prompt\nutilizes Generalized Knowledge Exploration to deduce highly generalized\nknowledge in a self-supervised manner. Evaluations on two benchmarks\nsubstantiate the efficacy of the proposed H-Prompts, exemplified by an average\naccuracy of 87.8% in Split CIFAR-100 and 70.6% in Split ImageNet-R.\n", "date": "2024-01-21", "categories": []}
{"id": "2401.09916", "title": "Enabling On-device Continual Learning with Binary Neural Networks", "abstract": "  On-device learning remains a formidable challenge, especially when dealing\nwith resource-constrained devices that have limited computational capabilities.\nThis challenge is primarily rooted in two key issues: first, the memory\navailable on embedded devices is typically insufficient to accommodate the\nmemory-intensive back-propagation algorithm, which often relies on\nfloating-point precision. Second, the development of learning algorithms on\nmodels with extreme quantization levels, such as Binary Neural Networks (BNNs),\nis critical due to the drastic reduction in bit representation. In this study,\nwe propose a solution that combines recent advancements in the field of\nContinual Learning (CL) and Binary Neural Networks to enable on-device training\nwhile maintaining competitive performance. Specifically, our approach leverages\nbinary latent replay (LR) activations and a novel quantization scheme that\nsignificantly reduces the number of bits required for gradient computation. The\nexperimental validation demonstrates a significant accuracy improvement in\ncombination with a noticeable reduction in memory requirement, confirming the\nsuitability of our approach in expanding the practical applications of deep\nlearning in real-world scenarios.\n", "date": "2024-01-18", "categories": []}
{"id": "2401.08940", "title": "CEL: A Continual Learning Model for Disease Outbreak Prediction by\n  Leveraging Domain Adaptation via Elastic Weight Consolidation", "abstract": "  Continual learning, the ability of a model to learn over time without\nforgetting previous knowledge and, therefore, be adaptive to new data, is\nparamount in dynamic fields such as disease outbreak prediction. Deep neural\nnetworks, i.e., LSTM, are prone to error due to catastrophic forgetting. This\nstudy introduces a novel CEL model for continual learning by leveraging domain\nadaptation via Elastic Weight Consolidation (EWC). This model aims to mitigate\nthe catastrophic forgetting phenomenon in a domain incremental setting. The\nFisher Information Matrix (FIM) is constructed with EWC to develop a\nregularization term that penalizes changes to important parameters, namely, the\nimportant previous knowledge. CEL's performance is evaluated on three distinct\ndiseases, Influenza, Mpox, and Measles, with different metrics. The high\nR-squared values during evaluation and reevaluation outperform the other\nstate-of-the-art models in several contexts, indicating that CEL adapts to\nincremental data well. CEL's robustness and reliability are underscored by its\nminimal 65% forgetting rate and 18% higher memory stability compared to\nexisting benchmark studies. This study highlights CEL's versatility in disease\noutbreak prediction, addressing evolving data with temporal patterns. It offers\na valuable model for proactive disease control with accurate, timely\npredictions.\n", "date": "2024-01-17", "categories": []}
{"id": "2401.09067", "title": "Towards Continual Learning Desiderata via HSIC-Bottleneck\n  Orthogonalization and Equiangular Embedding", "abstract": "  Deep neural networks are susceptible to catastrophic forgetting when trained\non sequential tasks. Various continual learning (CL) methods often rely on\nexemplar buffers or/and network expansion for balancing model stability and\nplasticity, which, however, compromises their practical value due to privacy\nand memory concerns. Instead, this paper considers a strict yet realistic\nsetting, where the training data from previous tasks is unavailable and the\nmodel size remains relatively constant during sequential training. To achieve\nsuch desiderata, we propose a conceptually simple yet effective method that\nattributes forgetting to layer-wise parameter overwriting and the resulting\ndecision boundary distortion. This is achieved by the synergy between two key\ncomponents: HSIC-Bottleneck Orthogonalization (HBO) implements non-overwritten\nparameter updates mediated by Hilbert-Schmidt independence criterion in an\northogonal space and EquiAngular Embedding (EAE) enhances decision boundary\nadaptation between old and new tasks with predefined basis vectors. Extensive\nexperiments demonstrate that our method achieves competitive accuracy\nperformance, even with absolute superiority of zero exemplar buffer and 1.02x\nthe base model.\n", "date": "2024-01-17", "categories": []}
{"id": "2301.05246", "title": "Online Class-Incremental Learning For Real-World Food Image\n  Classification", "abstract": "  Food image classification is essential for monitoring health and tracking\ndietary in image-based dietary assessment methods. However, conventional\nsystems often rely on static datasets with fixed classes and uniform\ndistribution. In contrast, real-world food consumption patterns, shaped by\ncultural, economic, and personal influences, involve dynamic and evolving data.\nThus, require the classification system to cope with continuously evolving\ndata. Online Class Incremental Learning (OCIL) addresses the challenge of\nlearning continuously from a single-pass data stream while adapting to the new\nknowledge and reducing catastrophic forgetting. Experience Replay (ER) based\nOCIL methods store a small portion of previous data and have shown encouraging\nperformance. However, most existing OCIL works assume that the distribution of\nencountered data is perfectly balanced, which rarely happens in real-world\nscenarios. In this work, we explore OCIL for real-world food image\nclassification by first introducing a probabilistic framework to simulate\nrealistic food consumption scenarios. Subsequently, we present an attachable\nDynamic Model Update (DMU) module designed for existing ER methods, which\nenables the selection of relevant images for model training, addressing\nchallenges arising from data repetition and imbalanced sample occurrences\ninherent in realistic food consumption patterns within the OCIL framework. Our\nperformance evaluation demonstrates significant enhancements compared to\nestablished ER methods, showing great potential for lifelong learning in\nreal-world food image classification scenarios. The code of our method is\npublicly accessible at\nhttps://gitlab.com/viper-purdue/OCIL-real-world-food-image-classification\n", "date": "2023-01-12", "categories": []}
{"id": "2307.02251", "title": "RanPAC: Random Projections and Pre-trained Models for Continual Learning", "abstract": "  Continual learning (CL) aims to incrementally learn different tasks (such as\nclassification) in a non-stationary data stream without forgetting old ones.\nMost CL works focus on tackling catastrophic forgetting under a\nlearning-from-scratch paradigm. However, with the increasing prominence of\nfoundation models, pre-trained models equipped with informative representations\nhave become available for various downstream requirements. Several CL methods\nbased on pre-trained models have been explored, either utilizing pre-extracted\nfeatures directly (which makes bridging distribution gaps challenging) or\nincorporating adaptors (which may be subject to forgetting). In this paper, we\npropose a concise and effective approach for CL with pre-trained models. Given\nthat forgetting occurs during parameter updating, we contemplate an alternative\napproach that exploits training-free random projectors and class-prototype\naccumulation, which thus bypasses the issue. Specifically, we inject a frozen\nRandom Projection layer with nonlinear activation between the pre-trained\nmodel's feature representations and output head, which captures interactions\nbetween features with expanded dimensionality, providing enhanced linear\nseparability for class-prototype-based CL. We also demonstrate the importance\nof decorrelating the class-prototypes to reduce the distribution disparity when\nusing pre-trained representations. These techniques prove to be effective and\ncircumvent the problem of forgetting for both class- and domain-incremental\ncontinual learning. Compared to previous methods applied to pre-trained\nViT-B/16 models, we reduce final error rates by between 20% and 62% on seven\nclass-incremental benchmarks, despite not using any rehearsal memory. We\nconclude that the full potential of pre-trained models for simple, effective,\nand fast CL has not hitherto been fully tapped. Code is at\ngithub.com/RanPAC/RanPAC.\n", "date": "2023-07-05", "categories": []}
{"id": "2303.13199", "title": "First Session Adaptation: A Strong Replay-Free Baseline for\n  Class-Incremental Learning", "abstract": "  In Class-Incremental Learning (CIL) an image classification system is exposed\nto new classes in each learning session and must be updated incrementally.\nMethods approaching this problem have updated both the classification head and\nthe feature extractor body at each session of CIL. In this work, we develop a\nbaseline method, First Session Adaptation (FSA), that sheds light on the\nefficacy of existing CIL approaches and allows us to assess the relative\nperformance contributions from head and body adaption. FSA adapts a pre-trained\nneural network body only on the first learning session and fixes it thereafter;\na head based on linear discriminant analysis (LDA), is then placed on top of\nthe adapted body, allowing exact updates through CIL. FSA is replay-free\ni.e.~it does not memorize examples from previous sessions of continual\nlearning. To empirically motivate FSA, we first consider a diverse selection of\n22 image-classification datasets, evaluating different heads and body\nadaptation techniques in high/low-shot offline settings. We find that the LDA\nhead performs well and supports CIL out-of-the-box. We also find that\nFeaturewise Layer Modulation (FiLM) adapters are highly effective in the\nfew-shot setting, and full-body adaption in the high-shot setting. Second, we\nempirically investigate various CIL settings including high-shot CIL and\nfew-shot CIL, including settings that have previously been used in the\nliterature. We show that FSA significantly improves over the state-of-the-art\nin 15 of the 16 settings considered. FSA with FiLM adapters is especially\nperformant in the few-shot setting. These results indicate that current\napproaches to continuous body adaptation are not working as expected. Finally,\nwe propose a measure that can be applied to a set of unlabelled inputs which is\npredictive of the benefits of body adaptation.\n", "date": "2023-03-23", "categories": []}
{"id": "2309.14062", "title": "FeCAM: Exploiting the Heterogeneity of Class Distributions in\n  Exemplar-Free Continual Learning", "abstract": "  Exemplar-free class-incremental learning (CIL) poses several challenges since\nit prohibits the rehearsal of data from previous tasks and thus suffers from\ncatastrophic forgetting. Recent approaches to incrementally learning the\nclassifier by freezing the feature extractor after the first task have gained\nmuch attention. In this paper, we explore prototypical networks for CIL, which\ngenerate new class prototypes using the frozen feature extractor and classify\nthe features based on the Euclidean distance to the prototypes. In an analysis\nof the feature distributions of classes, we show that classification based on\nEuclidean metrics is successful for jointly trained features. However, when\nlearning from non-stationary data, we observe that the Euclidean metric is\nsuboptimal and that feature distributions are heterogeneous. To address this\nchallenge, we revisit the anisotropic Mahalanobis distance for CIL. In\naddition, we empirically show that modeling the feature covariance relations is\nbetter than previous attempts at sampling features from normal distributions\nand training a linear classifier. Unlike existing methods, our approach\ngeneralizes to both many- and few-shot CIL settings, as well as to\ndomain-incremental settings. Interestingly, without updating the backbone\nnetwork, our method obtains state-of-the-art results on several standard\ncontinual learning benchmarks. Code is available at\nhttps://github.com/dipamgoswami/FeCAM.\n", "date": "2023-09-25", "categories": []}
{"id": "2312.00840", "title": "Towards Redundancy-Free Sub-networks in Continual Learning", "abstract": "  Catastrophic Forgetting (CF) is a prominent issue in continual learning.\nParameter isolation addresses this challenge by masking a sub-network for each\ntask to mitigate interference with old tasks. However, these sub-networks are\nconstructed relying on weight magnitude, which does not necessarily correspond\nto the importance of weights, resulting in maintaining unimportant weights and\nconstructing redundant sub-networks. To overcome this limitation, inspired by\ninformation bottleneck, which removes redundancy between adjacent network\nlayers, we propose \\textbf{\\underline{I}nformation \\underline{B}ottleneck\n\\underline{M}asked sub-network (IBM)} to eliminate redundancy within\nsub-networks. Specifically, IBM accumulates valuable information into essential\nweights to construct redundancy-free sub-networks, not only effectively\nmitigating CF by freezing the sub-networks but also facilitating new tasks\ntraining through the transfer of valuable knowledge. Additionally, IBM\ndecomposes hidden representations to automate the construction process and make\nit flexible. Extensive experiments demonstrate that IBM consistently\noutperforms state-of-the-art methods. Notably, IBM surpasses the\nstate-of-the-art parameter isolation method with a 70\\% reduction in the number\nof parameters within sub-networks and an 80\\% decrease in training time.\n", "date": "2023-12-01", "categories": []}
{"id": "2401.05667", "title": "EsaCL: Efficient Continual Learning of Sparse Models", "abstract": "  A key challenge in the continual learning setting is to efficiently learn a\nsequence of tasks without forgetting how to perform previously learned tasks.\nMany existing approaches to this problem work by either retraining the model on\nprevious tasks or by expanding the model to accommodate new tasks. However,\nthese approaches typically suffer from increased storage and computational\nrequirements, a problem that is worsened in the case of sparse models due to\nneed for expensive re-training after sparsification. To address this challenge,\nwe propose a new method for efficient continual learning of sparse models\n(EsaCL) that can automatically prune redundant parameters without adversely\nimpacting the model's predictive power, and circumvent the need of retraining.\nWe conduct a theoretical analysis of loss landscapes with parameter pruning,\nand design a directional pruning (SDP) strategy that is informed by the\nsharpness of the loss function with respect to the model parameters. SDP\nensures model with minimal loss of predictive accuracy, accelerating the\nlearning of sparse models at each stage. To accelerate model update, we\nintroduce an intelligent data selection (IDS) strategy that can identify\ncritical instances for estimating loss landscape, yielding substantially\nimproved data efficiency. The results of our experiments show that EsaCL\nachieves performance that is competitive with the state-of-the-art methods on\nthree continual learning benchmarks, while using substantially reduced memory\nand computational resources.\n", "date": "2024-01-11", "categories": []}
{"id": "2309.10149", "title": "Analysis of the Memorization and Generalization Capabilities of AI\n  Agents: Are Continual Learners Robust?", "abstract": "  In continual learning (CL), an AI agent (e.g., autonomous vehicles or\nrobotics) learns from non-stationary data streams under dynamic environments.\nFor the practical deployment of such applications, it is important to guarantee\nrobustness to unseen environments while maintaining past experiences. In this\npaper, a novel CL framework is proposed to achieve robust generalization to\ndynamic environments while retaining past knowledge. The considered CL agent\nuses a capacity-limited memory to save previously observed environmental\ninformation to mitigate forgetting issues. Then, data points are sampled from\nthe memory to estimate the distribution of risks over environmental change so\nas to obtain predictors that are robust with unseen changes. The generalization\nand memorization performance of the proposed framework are theoretically\nanalyzed. This analysis showcases the tradeoff between memorization and\ngeneralization with the memory size. Experiments show that the proposed\nalgorithm outperforms memory-based CL baselines across all environments while\nsignificantly improving the generalization performance on unseen target\nenvironments.\n", "date": "2023-09-18", "categories": []}
{"id": "2305.16645", "title": "Summarizing Stream Data for Memory-Constrained Online Continual Learning", "abstract": "  Replay-based methods have proved their effectiveness on online continual\nlearning by rehearsing past samples from an auxiliary memory. With many efforts\nmade on improving training schemes based on the memory, however, the\ninformation carried by each sample in the memory remains under-investigated.\nUnder circumstances with restricted storage space, the informativeness of the\nmemory becomes critical for effective replay. Although some works design\nspecific strategies to select representative samples, by only employing a small\nnumber of original images, the storage space is still not well utilized. To\nthis end, we propose to Summarize the knowledge from the Stream Data (SSD) into\nmore informative samples by distilling the training characteristics of real\nimages. Through maintaining the consistency of training gradients and\nrelationship to the past tasks, the summarized samples are more representative\nfor the stream data compared to the original images. Extensive experiments are\nconducted on multiple online continual learning benchmarks to support that the\nproposed SSD method significantly enhances the replay effects. We demonstrate\nthat with limited extra computational overhead, SSD provides more than 3%\naccuracy boost for sequential CIFAR-100 under extremely restricted memory\nbuffer. Code in https://github.com/vimar-gu/SSD.\n", "date": "2023-05-26", "categories": []}
{"id": "2207.11005", "title": "AdaptCL: Adaptive Continual Learning for Tackling Heterogeneity in\n  Sequential Datasets", "abstract": "  Managing heterogeneous datasets that vary in complexity, size, and similarity\nin continual learning presents a significant challenge. Task-agnostic continual\nlearning is necessary to address this challenge, as datasets with varying\nsimilarity pose difficulties in distinguishing task boundaries. Conventional\ntask-agnostic continual learning practices typically rely on rehearsal or\nregularization techniques. However, rehearsal methods may struggle with varying\ndataset sizes and regulating the importance of old and new data due to rigid\nbuffer sizes. Meanwhile, regularization methods apply generic constraints to\npromote generalization but can hinder performance when dealing with dissimilar\ndatasets lacking shared features, necessitating a more adaptive approach. In\nthis paper, we propose AdaptCL, a novel adaptive continual learning method to\ntackle heterogeneity in sequential datasets. AdaptCL employs fine-grained\ndata-driven pruning to adapt to variations in data complexity and dataset size.\nIt also utilizes task-agnostic parameter isolation to mitigate the impact of\nvarying degrees of catastrophic forgetting caused by differences in data\nsimilarity. Through a two-pronged case study approach, we evaluate AdaptCL on\nboth datasets of MNIST Variants and DomainNet, as well as datasets from\ndifferent domains. The latter include both large-scale, diverse binary-class\ndatasets and few-shot, multi-class datasets. Across all these scenarios,\nAdaptCL consistently exhibits robust performance, demonstrating its flexibility\nand general applicability in handling heterogeneous datasets.\n", "date": "2022-07-22", "categories": []}
{"id": "2303.04751", "title": "Multimodal Parameter-Efficient Few-Shot Class Incremental Learning", "abstract": "  Few-Shot Class Incremental Learning (FSCIL) is a challenging continual\nlearning task, where limited training examples are available during several\nlearning sessions. To succeed in this task, it is necessary to avoid\nover-fitting new classes caused by biased distributions in the few-shot\ntraining sets. The general approach to address this issue involves enhancing\nthe representational capability of a pre-defined backbone architecture by\nadding special modules for backward compatibility with older classes. However,\nthis approach has not yet solved the dilemma of ensuring high classification\naccuracy over time while reducing the gap between the performance obtained on\nlarger training sets and the smaller ones. In this work, we propose an\nalternative approach called Continual Parameter-Efficient CLIP (CPE-CLIP) to\nreduce the loss of information between different learning sessions. Instead of\nadapting additional modules to address information loss, we leverage the vast\nknowledge acquired by CLIP in large-scale pre-training and its effectiveness in\ngeneralizing to new concepts. Our approach is multimodal and\nparameter-efficient, relying on learnable prompts for both the language and\nvision encoders to enable transfer learning across sessions. We also introduce\nprompt regularization to improve performance and prevent forgetting. Our\nexperimental results demonstrate that CPE-CLIP significantly improves FSCIL\nperformance compared to state-of-the-art proposals while also drastically\nreducing the number of learnable parameters and training costs.\n", "date": "2023-03-08", "categories": []}
{"id": "2401.03077", "title": "A Topology-aware Graph Coarsening Framework for Continual Graph Learning", "abstract": "  Continual learning on graphs tackles the problem of training a graph neural\nnetwork (GNN) where graph data arrive in a streaming fashion and the model\ntends to forget knowledge from previous tasks when updating with new data.\nTraditional continual learning strategies such as Experience Replay can be\nadapted to streaming graphs, however, these methods often face challenges such\nas inefficiency in preserving graph topology and incapability of capturing the\ncorrelation between old and new tasks. To address these challenges, we propose\nTA$\\mathbb{CO}$, a (t)opology-(a)ware graph (co)arsening and (co)ntinual\nlearning framework that stores information from previous tasks as a reduced\ngraph. At each time period, this reduced graph expands by combining with a new\ngraph and aligning shared nodes, and then it undergoes a \"zoom out\" process by\nreduction to maintain a stable size. We design a graph coarsening algorithm\nbased on node representation proximities to efficiently reduce a graph and\npreserve topological information. We empirically demonstrate the learning\nprocess on the reduced graph can approximate that of the original graph. Our\nexperiments validate the effectiveness of the proposed framework on three\nreal-world datasets using different backbone GNN models.\n", "date": "2024-01-05", "categories": []}
{"id": "2401.03129", "title": "Examining Forgetting in Continual Pre-training of Aligned Large Language\n  Models", "abstract": "  Recent advances in Large Language Models (LLMs) have exhibited remarkable\nproficiency across various tasks. Given the potent applications of LLMs in\nnumerous fields, there has been a surge in LLM development. In developing LLMs,\na common practice involves continual pre-training on previously fine-tuned\nmodels. However, this can lead to catastrophic forgetting. In our work, we\ninvestigate the phenomenon of forgetting that occurs during continual\npre-training on an existing fine-tuned LLM. We evaluate the impact of\ncontinuous pre-training on the fine-tuned LLM across various dimensions,\nincluding output format, knowledge, and reliability. Experiment results\nhighlight the non-trivial challenge of addressing catastrophic forgetting\nduring continual pre-training, especially the repetition issue.\n", "date": "2024-01-06", "categories": []}
{"id": "2401.02457", "title": "eCIL-MU: Embedding based Class Incremental Learning and Machine\n  Unlearning", "abstract": "  New categories may be introduced over time, or existing categories may need\nto be reclassified. Class incremental learning (CIL) is employed for the\ngradual acquisition of knowledge about new categories while preserving\ninformation about previously learned ones in such dynamic environments. It\nmight also be necessary to also eliminate the influence of related categories\non the model to adapt to reclassification. We thus introduce class-level\nmachine unlearning (MU) within CIL. Typically, MU methods tend to be\ntime-consuming and can potentially harm the model's performance. A continuous\nstream of unlearning requests could lead to catastrophic forgetting. To address\nthese issues, we propose a non-destructive eCIL-MU framework based on embedding\ntechniques to map data into vectors and then be stored in vector databases. Our\napproach exploits the overlap between CIL and MU tasks for acceleration.\nExperiments demonstrate the capability of achieving unlearning effectiveness\nand orders of magnitude (upto $\\sim 278\\times$) of acceleration.\n", "date": "2024-01-04", "categories": []}
{"id": "2305.04106", "title": "On the Usage of Continual Learning for Out-of-Distribution\n  Generalization in Pre-trained Language Models of Code", "abstract": "  Pre-trained language models (PLMs) have become a prevalent technique in deep\nlearning for code, utilizing a two-stage pre-training and fine-tuning procedure\nto acquire general knowledge about code and specialize in a variety of\ndownstream tasks. However, the dynamic nature of software codebases poses a\nchallenge to the effectiveness and robustness of PLMs. In particular,\nworld-realistic scenarios potentially lead to significant differences between\nthe distribution of the pre-training and test data, i.e., distribution shift,\nresulting in a degradation of the PLM's performance on downstream tasks. In\nthis paper, we stress the need for adapting PLMs of code to software data whose\ndistribution changes over time, a crucial problem that has been overlooked in\nprevious works. The motivation of this work is to consider the PLM in a\nnon-stationary environment, where fine-tuning data evolves over time according\nto a software evolution scenario. Specifically, we design a scenario where the\nmodel needs to learn from a stream of programs containing new, unseen APIs over\ntime. We study two widely used PLM architectures, i.e., a GPT2 decoder and a\nRoBERTa encoder, on two downstream tasks, API call and API usage prediction. We\ndemonstrate that the most commonly used fine-tuning technique from prior work\nis not robust enough to handle the dynamic nature of APIs, leading to the loss\nof previously acquired knowledge i.e., catastrophic forgetting. To address\nthese issues, we implement five continual learning approaches, including\nreplay-based and regularization-based methods. Our findings demonstrate that\nutilizing these straightforward methods effectively mitigates catastrophic\nforgetting in PLMs across both downstream tasks while achieving comparable or\nsuperior performance.\n", "date": "2023-05-06", "categories": []}
{"id": "2401.02255", "title": "Balancing Continual Learning and Fine-tuning for Human Activity\n  Recognition", "abstract": "  Wearable-based Human Activity Recognition (HAR) is a key task in\nhuman-centric machine learning due to its fundamental understanding of human\nbehaviours. Due to the dynamic nature of human behaviours, continual learning\npromises HAR systems that are tailored to users' needs. However, because of the\ndifficulty in collecting labelled data with wearable sensors, existing\napproaches that focus on supervised continual learning have limited\napplicability, while unsupervised continual learning methods only handle\nrepresentation learning while delaying classifier training to a later stage.\nThis work explores the adoption and adaptation of CaSSLe, a continual\nself-supervised learning model, and Kaizen, a semi-supervised continual\nlearning model that balances representation learning and down-stream\nclassification, for the task of wearable-based HAR. These schemes re-purpose\ncontrastive learning for knowledge retention and, Kaizen combines that with\nself-training in a unified scheme that can leverage unlabelled and labelled\ndata for continual learning. In addition to comparing state-of-the-art\nself-supervised continual learning schemes, we further investigated the\nimportance of different loss terms and explored the trade-off between knowledge\nretention and learning from new tasks. In particular, our extensive evaluation\ndemonstrated that the use of a weighting factor that reflects the ratio between\nlearned and new classes achieves the best overall trade-off in continual\nlearning.\n", "date": "2024-01-04", "categories": []}
{"id": "2401.01054", "title": "Elastic Multi-Gradient Descent for Parallel Continual Learning", "abstract": "  The goal of Continual Learning (CL) is to continuously learn from new data\nstreams and accomplish the corresponding tasks. Previously studied CL assumes\nthat data are given in sequence nose-to-tail for different tasks, thus indeed\nbelonging to Serial Continual Learning (SCL). This paper studies the novel\nparadigm of Parallel Continual Learning (PCL) in dynamic multi-task scenarios,\nwhere a diverse set of tasks is encountered at different time points. PCL\npresents challenges due to the training of an unspecified number of tasks with\nvarying learning progress, leading to the difficulty of guaranteeing effective\nmodel updates for all encountered tasks. In our previous conference work, we\nfocused on measuring and reducing the discrepancy among gradients in a\nmulti-objective optimization problem, which, however, may still contain\nnegative transfers in every model update. To address this issue, in the dynamic\nmulti-objective optimization problem, we introduce task-specific elastic\nfactors to adjust the descent direction towards the Pareto front. The proposed\nmethod, called Elastic Multi-Gradient Descent (EMGD), ensures that each update\nfollows an appropriate Pareto descent direction, minimizing any negative impact\non previously learned tasks. To balance the training between old and new tasks,\nwe also propose a memory editing mechanism guided by the gradient computed\nusing EMGD. This editing process updates the stored data points, reducing\ninterference in the Pareto descent direction from previous tasks. Experiments\non public datasets validate the effectiveness of our EMGD in the PCL setting.\n", "date": "2024-01-02", "categories": []}
{"id": "2312.17210", "title": "Continual Learning via Sequential Function-Space Variational Inference", "abstract": "  Sequential Bayesian inference over predictive functions is a natural\nframework for continual learning from streams of data. However, applying it to\nneural networks has proved challenging in practice. Addressing the drawbacks of\nexisting techniques, we propose an optimization objective derived by\nformulating continual learning as sequential function-space variational\ninference. In contrast to existing methods that regularize neural network\nparameters directly, this objective allows parameters to vary widely during\ntraining, enabling better adaptation to new tasks. Compared to objectives that\ndirectly regularize neural network predictions, the proposed objective allows\nfor more flexible variational distributions and more effective regularization.\nWe demonstrate that, across a range of task sequences, neural networks trained\nvia sequential function-space variational inference achieve better predictive\naccuracy than networks trained with related methods while depending less on\nmaintaining a set of representative points from previous tasks.\n", "date": "2023-12-28", "categories": []}
{"id": "2312.14990", "title": "Learning to Prompt Knowledge Transfer for Open-World Continual Learning", "abstract": "  This paper studies the problem of continual learning in an open-world\nscenario, referred to as Open-world Continual Learning (OwCL). OwCL is\nincreasingly rising while it is highly challenging in two-fold: i) learning a\nsequence of tasks without forgetting knowns in the past, and ii) identifying\nunknowns (novel objects/classes) in the future. Existing OwCL methods suffer\nfrom the adaptability of task-aware boundaries between knowns and unknowns, and\ndo not consider the mechanism of knowledge transfer. In this work, we propose\nPro-KT, a novel prompt-enhanced knowledge transfer model for OwCL. Pro-KT\nincludes two key components: (1) a prompt bank to encode and transfer both\ntask-generic and task-specific knowledge, and (2) a task-aware open-set\nboundary to identify unknowns in the new tasks. Experimental results using two\nreal-world datasets demonstrate that the proposed Pro-KT outperforms the\nstate-of-the-art counterparts in both the detection of unknowns and the\nclassification of knowns markedly.\n", "date": "2023-12-22", "categories": []}
{"id": "2310.15571", "title": "Visually Grounded Continual Language Learning with Selective\n  Specialization", "abstract": "  A desirable trait of an artificial agent acting in the visual world is to\ncontinually learn a sequence of language-informed tasks while striking a\nbalance between sufficiently specializing in each task and building a\ngeneralized knowledge for transfer. Selective specialization, i.e., a careful\nselection of model components to specialize in each task, is a strategy to\nprovide control over this trade-off. However, the design of selection\nstrategies requires insights on the role of each model component in learning\nrather specialized or generalizable representations, which poses a gap in\ncurrent research. Thus, our aim with this work is to provide an extensive\nanalysis of selection strategies for visually grounded continual language\nlearning. Due to the lack of suitable benchmarks for this purpose, we introduce\ntwo novel diagnostic datasets that provide enough control and flexibility for a\nthorough model analysis. We assess various heuristics for module specialization\nstrategies as well as quantifiable measures for two different types of model\narchitectures. Finally, we design conceptually simple approaches based on our\nanalysis that outperform common continual learning baselines. Our results\ndemonstrate the need for further efforts towards better aligning continual\nlearning algorithms with the learning behaviors of individual model parts.\n", "date": "2023-10-24", "categories": []}
{"id": "2312.13500", "title": "Federated Continual Novel Class Learning", "abstract": "  In a privacy-focused era, Federated Learning (FL) has emerged as a promising\nmachine learning technique. However, most existing FL studies assume that the\ndata distribution remains nearly fixed over time, while real-world scenarios\noften involve dynamic and continual changes. To equip FL systems with continual\nmodel evolution capabilities, we focus on an important problem called Federated\nContinual Novel Class Learning (FedCN) in this work. The biggest challenge in\nFedCN is to merge and align novel classes that are discovered and learned by\ndifferent clients without compromising privacy. To address this, we propose a\nGlobal Alignment Learning (GAL) framework that can accurately estimate the\nglobal novel class number and provide effective guidance for local training\nfrom a global perspective, all while maintaining privacy protection.\nSpecifically, GAL first locates high-density regions in the representation\nspace through a bi-level clustering mechanism to estimate the novel class\nnumber, with which the global prototypes corresponding to novel classes can be\nconstructed. Then, GAL uses a novel semantic weighted loss to capture all\npossible correlations between these prototypes and the training data for\nmitigating the impact of pseudo-label noise and data heterogeneity. Extensive\nexperiments on various datasets demonstrate GAL's superior performance over\nstate-of-the-art novel class discovery methods. In particular, GAL achieves\nsignificant improvements in novel-class performance, increasing the accuracy by\n5.1% to 10.6% in the case of one novel class learning stage and by 7.8% to\n17.9% in the case of two novel class learning stages, without sacrificing\nknown-class performance. Moreover, GAL is shown to be effective in equipping a\nvariety of different mainstream FL algorithms with novel class discovery and\nlearning capability, highlighting its potential for many real-world\napplications.\n", "date": "2023-12-21", "categories": []}
{"id": "2312.13699", "title": "Adapt & Align: Continual Learning with Generative Models Latent Space\n  Alignment", "abstract": "  In this work, we introduce Adapt & Align, a method for continual learning of\nneural networks by aligning latent representations in generative models. Neural\nNetworks suffer from abrupt loss in performance when retrained with additional\ntraining data from different distributions. At the same time, training with\nadditional data without access to the previous examples rarely improves the\nmodel's performance. In this work, we propose a new method that mitigates those\nproblems by employing generative models and splitting the process of their\nupdate into two parts. In the first one, we train a local generative model\nusing only data from a new task. In the second phase, we consolidate latent\nrepresentations from the local model with a global one that encodes knowledge\nof all past experiences. We introduce our approach with Variational\nAuteoncoders and Generative Adversarial Networks. Moreover, we show how we can\nuse those generative models as a general method for continual knowledge\nconsolidation that can be used in downstream tasks such as classification.\n", "date": "2023-12-21", "categories": []}
{"id": "2312.02916", "title": "MIND: Multi-Task Incremental Network Distillation", "abstract": "  The recent surge of pervasive devices that generate dynamic data streams has\nunderscored the necessity for learning systems to adapt continually to data\ndistributional shifts. To tackle this challenge, the research community has put\nforth a spectrum of methodologies, including the demanding pursuit of\nclass-incremental learning without replay data. In this study, we present MIND,\na parameter isolation method that aims to significantly enhance the performance\nof replay-free solutions and achieve state-of-the-art results on several widely\nstudied datasets. Our approach introduces two main contributions: two\nalternative distillation procedures that significantly improve the efficiency\nof MIND increasing the accumulated knowledge of each sub-network, and the\noptimization of the BachNorm layers across tasks inside the sub-networks.\nOverall, MIND outperforms all the state-of-the-art methods for rehearsal-free\nClass-Incremental learning (with an increment in classification accuracy of\napprox. +6% on CIFAR-100/10 and +10% on TinyImageNet/10) reaching up to approx.\n+40% accuracy in Domain-Incremental scenarios. Moreover, we ablated each\ncontribution to demonstrate its impact on performance improvement. Our results\nshowcase the superior performance of MIND indicating its potential for\naddressing the challenges posed by Class-incremental and Domain-Incremental\nlearning in resource-constrained environments.\n", "date": "2023-12-05", "categories": []}
{"id": "2312.12722", "title": "Fine-Grained Knowledge Selection and Restoration for Non-Exemplar Class\n  Incremental Learning", "abstract": "  Non-exemplar class incremental learning aims to learn both the new and old\ntasks without accessing any training data from the past. This strict\nrestriction enlarges the difficulty of alleviating catastrophic forgetting\nsince all techniques can only be applied to current task data. Considering this\nchallenge, we propose a novel framework of fine-grained knowledge selection and\nrestoration. The conventional knowledge distillation-based methods place too\nstrict constraints on the network parameters and features to prevent\nforgetting, which limits the training of new tasks. To loose this constraint,\nwe proposed a novel fine-grained selective patch-level distillation to\nadaptively balance plasticity and stability. Some task-agnostic patches can be\nused to preserve the decision boundary of the old task. While some patches\ncontaining the important foreground are favorable for learning the new task.\n  Moreover, we employ a task-agnostic mechanism to generate more realistic\nprototypes of old tasks with the current task sample for reducing classifier\nbias for fine-grained knowledge restoration. Extensive experiments on CIFAR100,\nTinyImageNet and ImageNet-Subset demonstrate the effectiveness of our method.\nCode is available at https://github.com/scok30/vit-cil.\n", "date": "2023-12-20", "categories": []}
{"id": "2312.13026", "title": "FusDom: Combining In-Domain and Out-of-Domain Knowledge for Continuous\n  Self-Supervised Learning", "abstract": "  Continued pre-training (CP) offers multiple advantages, like target domain\nadaptation and the potential to exploit the continuous stream of unlabeled data\navailable online. However, continued pre-training on out-of-domain\ndistributions often leads to catastrophic forgetting of previously acquired\nknowledge, leading to sub-optimal ASR performance. This paper presents FusDom,\na simple and novel methodology for SSL-based continued pre-training. FusDom\nlearns speech representations that are robust and adaptive yet not forgetful of\nconcepts seen in the past. Instead of solving the SSL pre-text task on the\noutput representations of a single model, FusDom leverages two identical\npre-trained SSL models, a teacher and a student, with a modified pre-training\nhead to solve the CP SSL pre-text task. This head employs a cross-attention\nmechanism between the representations of both models while only the student\nreceives gradient updates and the teacher does not. Finally, the student is\nfine-tuned for ASR. In practice, FusDom outperforms all our baselines across\nsettings significantly, with WER improvements in the range of 0.2 WER - 7.3 WER\nin the target domain while retaining the performance in the earlier domain.\n", "date": "2023-12-20", "categories": []}
{"id": "2302.01242", "title": "Neuro-Symbolic Continual Learning: Knowledge, Reasoning Shortcuts and\n  Concept Rehearsal", "abstract": "  We introduce Neuro-Symbolic Continual Learning, where a model has to solve a\nsequence of neuro-symbolic tasks, that is, it has to map sub-symbolic inputs to\nhigh-level concepts and compute predictions by reasoning consistently with\nprior knowledge. Our key observation is that neuro-symbolic tasks, although\ndifferent, often share concepts whose semantics remains stable over time.\nTraditional approaches fall short: existing continual strategies ignore\nknowledge altogether, while stock neuro-symbolic architectures suffer from\ncatastrophic forgetting. We show that leveraging prior knowledge by combining\nneuro-symbolic architectures with continual strategies does help avoid\ncatastrophic forgetting, but also that doing so can yield models affected by\nreasoning shortcuts. These undermine the semantics of the acquired concepts,\neven when detailed prior knowledge is provided upfront and inference is exact,\nand in turn continual performance. To overcome these issues, we introduce COOL,\na COncept-level cOntinual Learning strategy tailored for neuro-symbolic\ncontinual problems that acquires high-quality concepts and remembers them over\ntime. Our experiments on three novel benchmarks highlights how COOL attains\nsustained high performance on neuro-symbolic continual learning tasks in which\nother strategies fail.\n", "date": "2023-02-02", "categories": []}
{"id": "2311.09506", "title": "Investigating the Impact of Weight Sharing Decisions on Knowledge\n  Transfer in Continual Learning", "abstract": "  Continual Learning (CL) has generated attention as a method of avoiding\nCatastrophic Forgetting (CF) in the sequential training of neural networks,\nimproving network efficiency and adaptability to different tasks. Additionally,\nCL serves as an ideal setting for studying network behavior and Forward\nKnowledge Transfer (FKT) between tasks. Pruning methods for CL train\nsubnetworks to handle the sequential tasks which allows us to take a structured\napproach to investigating FKT. Sharing prior subnetworks' weights leverages\npast knowledge for the current task through FKT. Understanding which weights to\nshare is important as sharing all weights can yield sub-optimal accuracy. This\npaper investigates how different sharing decisions affect the FKT between\ntasks. Through this lens we demonstrate how task complexity and similarity\ninfluence the optimal weight sharing decisions, giving insights into the\nrelationships between tasks and helping inform decision making in similar CL\nmethods. We implement three sequential datasets designed to emphasize variation\nin task complexity and similarity, reporting results for both ResNet-18 and\nVGG-16. By sharing in accordance with the decisions supported by our findings,\nwe show that we can improve task accuracy compared to other sharing decisions.\n", "date": "2023-11-16", "categories": []}
{"id": "2312.11669", "title": "Prediction and Control in Continual Reinforcement Learning", "abstract": "  Temporal difference (TD) learning is often used to update the estimate of the\nvalue function which is used by RL agents to extract useful policies. In this\npaper, we focus on value function estimation in continual reinforcement\nlearning. We propose to decompose the value function into two components which\nupdate at different timescales: a permanent value function, which holds general\nknowledge that persists over time, and a transient value function, which allows\nquick adaptation to new situations. We establish theoretical results showing\nthat our approach is well suited for continual learning and draw connections to\nthe complementary learning systems (CLS) theory from neuroscience. Empirically,\nthis approach improves performance significantly on both prediction and control\nproblems.\n", "date": "2023-12-18", "categories": []}
{"id": "2312.10153", "title": "Bayesian Metaplasticity from Synaptic Uncertainty", "abstract": "  Catastrophic forgetting remains a challenge for neural networks, especially\nin lifelong learning scenarios. In this study, we introduce MEtaplasticity from\nSynaptic Uncertainty (MESU), inspired by metaplasticity and Bayesian inference\nprinciples. MESU harnesses synaptic uncertainty to retain information over\ntime, with its update rule closely approximating the diagonal Newton's method\nfor synaptic updates. Through continual learning experiments on permuted MNIST\ntasks, we demonstrate MESU's remarkable capability to maintain learning\nperformance across 100 tasks without the need of explicit task boundaries.\n", "date": "2023-12-15", "categories": []}
{"id": "2312.10549", "title": "Catastrophic Forgetting in Deep Learning: A Comprehensive Taxonomy", "abstract": "  Deep Learning models have achieved remarkable performance in tasks such as\nimage classification or generation, often surpassing human accuracy. However,\nthey can struggle to learn new tasks and update their knowledge without access\nto previous data, leading to a significant loss of accuracy known as\nCatastrophic Forgetting (CF). This phenomenon was first observed by McCloskey\nand Cohen in 1989 and remains an active research topic. Incremental learning\nwithout forgetting is widely recognized as a crucial aspect in building better\nAI systems, as it allows models to adapt to new tasks without losing the\nability to perform previously learned ones. This article surveys recent studies\nthat tackle CF in modern Deep Learning models that use gradient descent as\ntheir learning algorithm. Although several solutions have been proposed, a\ndefinitive solution or consensus on assessing CF is yet to be established. The\narticle provides a comprehensive review of recent solutions, proposes a\ntaxonomy to organize them, and identifies research gaps in this area.\n", "date": "2023-12-16", "categories": []}
{"id": "2303.10891", "title": "Non-Exemplar Online Class-incremental Continual Learning via\n  Dual-prototype Self-augment and Refinement", "abstract": "  This paper investigates a new, practical, but challenging problem named\nNon-exemplar Online Class-incremental continual Learning (NO-CL), which aims to\npreserve the discernibility of base classes without buffering data examples and\nefficiently learn novel classes continuously in a single-pass (i.e., online)\ndata stream. The challenges of this task are mainly two-fold: (1) Both base and\nnovel classes suffer from severe catastrophic forgetting as no previous samples\nare available for replay. (2) As the online data can only be observed once,\nthere is no way to fully re-train the whole model, e.g., re-calibrate the\ndecision boundaries via prototype alignment or feature distillation. In this\npaper, we propose a novel Dual-prototype Self-augment and Refinement method\n(DSR) for NO-CL problem, which consists of two strategies: 1) Dual class\nprototypes: vanilla and high-dimensional prototypes are exploited to utilize\nthe pre-trained information and obtain robust quasi-orthogonal representations\nrather than example buffers for both privacy preservation and memory reduction.\n2) Self-augment and refinement: Instead of updating the whole network, we\noptimize high-dimensional prototypes alternatively with the extra projection\nmodule based on self-augment vanilla prototypes, through a bi-level\noptimization problem. Extensive experiments demonstrate the effectiveness and\nsuperiority of the proposed DSR in NO-CL.\n", "date": "2023-03-20", "categories": []}
{"id": "2312.09352", "title": "PBES: PCA Based Exemplar Sampling Algorithm for Continual Learning", "abstract": "  We propose a novel exemplar selection approach based on Principal Component\nAnalysis (PCA) and median sampling, and a neural network training regime in the\nsetting of class-incremental learning. This approach avoids the pitfalls due to\noutliers in the data and is both simple to implement and use across various\nincremental machine learning models. It also has independent usage as a\nsampling algorithm. We achieve better performance compared to state-of-the-art\nmethods.\n", "date": "2023-12-14", "categories": []}
{"id": "2312.09357", "title": "DSS: A Diverse Sample Selection Method to Preserve Knowledge in\n  Class-Incremental Learning", "abstract": "  Rehearsal-based techniques are commonly used to mitigate catastrophic\nforgetting (CF) in Incremental learning (IL). The quality of the exemplars\nselected is important for this purpose and most methods do not ensure the\nappropriate diversity of the selected exemplars. We propose a new technique\n\"DSS\" -- Diverse Selection of Samples from the input data stream in the\nClass-incremental learning (CIL) setup under both disjoint and fuzzy task\nboundary scenarios. Our method outperforms state-of-the-art methods and is much\nsimpler to understand and implement.\n", "date": "2023-12-14", "categories": []}
{"id": "2312.09361", "title": "RTRA: Rapid Training of Regularization-based Approaches in Continual\n  Learning", "abstract": "  Catastrophic forgetting(CF) is a significant challenge in continual learning\n(CL). In regularization-based approaches to mitigate CF, modifications to\nimportant training parameters are penalized in subsequent tasks using an\nappropriate loss function. We propose the RTRA, a modification to the widely\nused Elastic Weight Consolidation (EWC) regularization scheme, using the\nNatural Gradient for loss function optimization. Our approach improves the\ntraining of regularization-based methods without sacrificing test-data\nperformance. We compare the proposed RTRA approach against EWC using the\niFood251 dataset. We show that RTRA has a clear edge over the state-of-the-art\napproaches.\n", "date": "2023-12-14", "categories": []}
{"id": "2312.09651", "title": "What to Remember: Self-Adaptive Continual Learning for Audio Deepfake\n  Detection", "abstract": "  The rapid evolution of speech synthesis and voice conversion has raised\nsubstantial concerns due to the potential misuse of such technology, prompting\na pressing need for effective audio deepfake detection mechanisms. Existing\ndetection models have shown remarkable success in discriminating known deepfake\naudio, but struggle when encountering new attack types. To address this\nchallenge, one of the emergent effective approaches is continual learning. In\nthis paper, we propose a continual learning approach called Radian Weight\nModification (RWM) for audio deepfake detection. The fundamental concept\nunderlying RWM involves categorizing all classes into two groups: those with\ncompact feature distributions across tasks, such as genuine audio, and those\nwith more spread-out distributions, like various types of fake audio. These\ndistinctions are quantified by means of the in-class cosine distance, which\nsubsequently serves as the basis for RWM to introduce a trainable gradient\nmodification direction for distinct data types. Experimental evaluations\nagainst mainstream continual learning methods reveal the superiority of RWM in\nterms of knowledge acquisition and mitigating forgetting in audio deepfake\ndetection. Furthermore, RWM's applicability extends beyond audio deepfake\ndetection, demonstrating its potential significance in diverse machine learning\ndomains such as image recognition.\n", "date": "2023-12-15", "categories": []}
{"id": "2312.08677", "title": "Adaptive Shortcut Debiasing for Online Continual Learning", "abstract": "  We propose a novel framework DropTop that suppresses the shortcut bias in\nonline continual learning (OCL) while being adaptive to the varying degree of\nthe shortcut bias incurred by continuously changing environment. By the\nobserved high-attention property of the shortcut bias, highly-activated\nfeatures are considered candidates for debiasing. More importantly, resolving\nthe limitation of the online environment where prior knowledge and auxiliary\ndata are not ready, two novel techniques -- feature map fusion and adaptive\nintensity shifting -- enable us to automatically determine the appropriate\nlevel and proportion of the candidate shortcut features to be dropped.\nExtensive experiments on five benchmark datasets demonstrate that, when\ncombined with various OCL algorithms, DropTop increases the average accuracy by\nup to 10.4% and decreases the forgetting by up to 63.2%.\n", "date": "2023-12-14", "categories": []}
{"id": "2312.08740", "title": "Learning a Low-Rank Feature Representation: Achieving Better Trade-Off\n  between Stability and Plasticity in Continual Learning", "abstract": "  In continual learning, networks confront a trade-off between stability and\nplasticity when trained on a sequence of tasks. To bolster plasticity without\nsacrificing stability, we propose a novel training algorithm called LRFR. This\napproach optimizes network parameters in the null space of the past tasks'\nfeature representation matrix to guarantee the stability. Concurrently, we\njudiciously select only a subset of neurons in each layer of the network while\ntraining individual tasks to learn the past tasks' feature representation\nmatrix in low-rank. This increases the null space dimension when designing\nnetwork parameters for subsequent tasks, thereby enhancing the plasticity.\nUsing CIFAR-100 and TinyImageNet as benchmark datasets for continual learning,\nthe proposed approach consistently outperforms state-of-the-art methods.\n", "date": "2023-12-14", "categories": []}
{"id": "2305.06408", "title": "Accelerating Batch Active Learning Using Continual Learning Techniques", "abstract": "  A major problem with Active Learning (AL) is high training costs since models\nare typically retrained from scratch after every query round. We start by\ndemonstrating that standard AL on neural networks with warm starting fails,\nboth to accelerate training and to avoid catastrophic forgetting when using\nfine-tuning over AL query rounds. We then develop a new class of techniques,\ncircumventing this problem, by biasing further training towards previously\nlabeled sets. We accomplish this by employing existing, and developing novel,\nreplay-based Continual Learning (CL) algorithms that are effective at quickly\nlearning the new without forgetting the old, especially when data comes from an\nevolving distribution. We call this paradigm Continual Active Learning (CAL).\nWe show CAL achieves significant speedups using a plethora of replay schemes\nthat use model distillation and that select diverse, uncertain points from the\nhistory. We conduct experiments across many data domains, including natural\nlanguage, vision, medical imaging, and computational biology, each with\ndifferent neural architectures and dataset sizes. CAL consistently provides a\n3x reduction in training time, while retaining performance.\n", "date": "2023-05-10", "categories": []}
{"id": "2211.13108", "title": "Integral Continual Learning Along the Tangent Vector Field of Tasks", "abstract": "  We propose a lightweight continual learning method which incorporates\ninformation from specialized datasets incrementally, by integrating it along\nthe vector field of \"generalist\" models. The tangent plane to the specialist\nmodel acts as a generalist guide and avoids the kind of over-fitting that leads\nto catastrophic forgetting, while exploiting the convexity of the optimization\nlandscape in the tangent plane. It maintains a small fixed-size memory buffer,\nas low as 0.4% of the source datasets, which is updated by simple resampling.\nOur method achieves strong performance across various buffer sizes for\ndifferent datasets. Specifically, in the class-incremental setting we\noutperform the existing methods that do not require distillation by an average\nof 18.77% and 28.48%, for Seq-CIFAR-10 and Seq-TinyImageNet respectively. Our\nmethod can easily be used in conjunction with existing replay-based continual\nlearning methods. When memory buffer constraints are relaxed to allow storage\nof metadata such as logits, we attain an error reduction of 17.84% towards the\nparagon performance on Seq-CIFAR-10.\n", "date": "2022-11-23", "categories": []}
{"id": "2312.07082", "title": "Continual Learning through Networks Splitting and Merging with\n  Dreaming-Meta-Weighted Model Fusion", "abstract": "  It's challenging to balance the networks stability and plasticity in\ncontinual learning scenarios, considering stability suffers from the update of\nmodel and plasticity benefits from it. Existing works usually focus more on the\nstability and restrict the learning plasticity of later tasks to avoid\ncatastrophic forgetting of learned knowledge. Differently, we propose a\ncontinual learning method named Split2MetaFusion which can achieve better\ntrade-off by employing a two-stage strategy: splitting and meta-weighted\nfusion. In this strategy, a slow model with better stability, and a fast model\nwith better plasticity are learned sequentially at the splitting stage. Then\nstability and plasticity are both kept by fusing the two models in an adaptive\nmanner. Towards this end, we design an optimizer named Task-Preferred Null\nSpace Projector(TPNSP) to the slow learning process for narrowing the fusion\ngap. To achieve better model fusion, we further design a Dreaming-Meta-Weighted\nfusion policy for better maintaining the old and new knowledge simultaneously,\nwhich doesn't require to use the previous datasets. Experimental results and\nanalysis reported in this work demonstrate the superiority of the proposed\nmethod for maintaining networks stability and keeping its plasticity. Our code\nwill be released.\n", "date": "2023-12-12", "categories": []}
{"id": "2306.00393", "title": "Teacher Agent: A Knowledge Distillation-Free Framework for\n  Rehearsal-based Video Incremental Learning", "abstract": "  Rehearsal-based video incremental learning often employs knowledge\ndistillation to mitigate catastrophic forgetting of previously learned data.\nHowever, this method faces two major challenges for video task: substantial\ncomputing resources from loading teacher model and limited replay capability\nfrom performance-limited teacher model. To address these problems, we first\npropose a knowledge distillation-free framework for rehearsal-based video\nincremental learning called \\textit{Teacher Agent}. Instead of loading\nparameter-heavy teacher networks, we introduce an agent generator that is\neither parameter-free or uses only a few parameters to obtain accurate and\nreliable soft labels. This method not only greatly reduces the computing\nrequirement but also circumvents the problem of knowledge misleading caused by\ninaccurate predictions of the teacher model. Moreover, we put forward a\nself-correction loss which provides an effective regularization signal for the\nreview of old knowledge, which in turn alleviates the problem of catastrophic\nforgetting. Further, to ensure that the samples in the memory buffer are\nmemory-efficient and representative, we introduce a unified sampler for\nrehearsal-based video incremental learning to mine fixed-length key video\nframes. Interestingly, based on the proposed strategies, the network exhibits a\nhigh level of robustness against spatial resolution reduction when compared to\nthe baseline. Extensive experiments demonstrate the advantages of our method,\nyielding significant performance improvements while utilizing only half the\nspatial resolution of video clips as network inputs in the incremental phases.\n", "date": "2023-06-01", "categories": []}
{"id": "2212.09744", "title": "DSI++: Updating Transformer Memory with New Documents", "abstract": "  Differentiable Search Indices (DSIs) encode a corpus of documents in model\nparameters and use the same model to answer user queries directly. Despite the\nstrong performance of DSI models, deploying them in situations where the corpus\nchanges over time is computationally expensive because reindexing the corpus\nrequires re-training the model. In this work, we introduce DSI++, a continual\nlearning challenge for DSI to incrementally index new documents while being\nable to answer queries related to both previously and newly indexed documents.\nAcross different model scales and document identifier representations, we show\nthat continual indexing of new documents leads to considerable forgetting of\npreviously indexed documents. We also hypothesize and verify that the model\nexperiences forgetting events during training, leading to unstable learning. To\nmitigate these issues, we investigate two approaches. The first focuses on\nmodifying the training dynamics. Flatter minima implicitly alleviate\nforgetting, so we optimize for flatter loss basins and show that the model\nstably memorizes more documents ($+12\\%$). Next, we introduce a generative\nmemory to sample pseudo-queries for documents and supplement them during\ncontinual indexing to prevent forgetting for the retrieval task. Extensive\nexperiments on novel continual indexing benchmarks based on Natural Questions\n(NQ) and MS MARCO demonstrate that our proposed solution mitigates forgetting\nsignificantly. Concretely, it improves the average Hits@10 by $+21.1\\%$ over\ncompetitive baselines for NQ and requires $6$ times fewer model updates\ncompared to re-training the DSI model for incrementally indexing five corpora\nin a sequence.\n", "date": "2022-12-19", "categories": []}
{"id": "2304.14310", "title": "Incremental Generalized Category Discovery", "abstract": "  We explore the problem of Incremental Generalized Category Discovery (IGCD).\nThis is a challenging category incremental learning setting where the goal is\nto develop models that can correctly categorize images from previously seen\ncategories, in addition to discovering novel ones. Learning is performed over a\nseries of time steps where the model obtains new labeled and unlabeled data,\nand discards old data, at each iteration. The difficulty of the problem is\ncompounded in our generalized setting as the unlabeled data can contain images\nfrom categories that may or may not have been observed before. We present a new\nmethod for IGCD which combines non-parametric categorization with efficient\nimage sampling to mitigate catastrophic forgetting. To quantify performance, we\npropose a new benchmark dataset named iNatIGCD that is motivated by a\nreal-world fine-grained visual categorization task. In our experiments we\noutperform existing related methods\n", "date": "2023-04-27", "categories": []}
{"id": "2309.02833", "title": "Image-Object-Specific Prompt Learning for Few-Shot Class-Incremental\n  Learning", "abstract": "  While many FSCIL studies have been undertaken, achieving satisfactory\nperformance, especially during incremental sessions, has remained challenging.\nOne prominent challenge is that the encoder, trained with an ample base session\ntraining set, often underperforms in incremental sessions. In this study, we\nintroduce a novel training framework for FSCIL, capitalizing on the\ngeneralizability of the Contrastive Language-Image Pre-training (CLIP) model to\nunseen classes. We achieve this by formulating image-object-specific (IOS)\nclassifiers for the input images. Here, an IOS classifier refers to one that\ntargets specific attributes (like wings or wheels) of class objects rather than\nthe image's background. To create these IOS classifiers, we encode a bias\nprompt into the classifiers using our specially designed module, which\nharnesses key-prompt pairs to pinpoint the IOS features of classes in each\nsession. From an FSCIL standpoint, our framework is structured to retain\nprevious knowledge and swiftly adapt to new sessions without forgetting or\noverfitting. This considers the updatability of modules in each session and\nsome tricks empirically found for fast convergence. Our approach consistently\ndemonstrates superior performance compared to state-of-the-art methods across\nthe miniImageNet, CIFAR100, and CUB200 datasets. Further, we provide additional\nexperiments to validate our learned model's ability to achieve IOS classifiers.\nWe also conduct ablation studies to analyze the impact of each module within\nthe architecture.\n", "date": "2023-09-06", "categories": []}
{"id": "2312.05100", "title": "Continual learning for surface defect segmentation by subnetwork\n  creation and selection", "abstract": "  We introduce a new continual (or lifelong) learning algorithm called LDA-CP&S\nthat performs segmentation tasks without undergoing catastrophic forgetting.\nThe method is applied to two different surface defect segmentation problems\nthat are learned incrementally, i.e. providing data about one type of defect at\na time, while still being capable of predicting every defect that was seen\npreviously. Our method creates a defect-related subnetwork for each defect type\nvia iterative pruning and trains a classifier based on linear discriminant\nanalysis (LDA). At the inference stage, we first predict the defect type with\nLDA and then predict the surface defects using the selected subnetwork. We\ncompare our method with other continual learning methods showing a significant\nimprovement -- mean Intersection over Union better by a factor of two when\ncompared to existing methods on both datasets. Importantly, our approach shows\ncomparable results with joint training when all the training data (all defects)\nare seen simultaneously\n", "date": "2023-12-08", "categories": []}
{"id": "2312.05229", "title": "Few-Shot Class-Incremental Learning via Training-Free Prototype\n  Calibration", "abstract": "  Real-world scenarios are usually accompanied by continuously appearing\nclasses with scare labeled samples, which require the machine learning model to\nincrementally learn new classes and maintain the knowledge of base classes. In\nthis Few-Shot Class-Incremental Learning (FSCIL) scenario, existing methods\neither introduce extra learnable components or rely on a frozen feature\nextractor to mitigate catastrophic forgetting and overfitting problems.\nHowever, we find a tendency for existing methods to misclassify the samples of\nnew classes into base classes, which leads to the poor performance of new\nclasses. In other words, the strong discriminability of base classes distracts\nthe classification of new classes. To figure out this intriguing phenomenon, we\nobserve that although the feature extractor is only trained on base classes, it\ncan surprisingly represent the semantic similarity between the base and unseen\nnew classes. Building upon these analyses, we propose a simple yet effective\nTraining-frEE calibratioN (TEEN) strategy to enhance the discriminability of\nnew classes by fusing the new prototypes (i.e., mean features of a class) with\nweighted base prototypes. In addition to standard benchmarks in FSCIL, TEEN\ndemonstrates remarkable performance and consistent improvements over baseline\nmethods in the few-shot learning scenario. Code is available at:\nhttps://github.com/wangkiw/TEEN\n", "date": "2023-12-08", "categories": []}
{"id": "2312.04225", "title": "TLCE: Transfer-Learning Based Classifier Ensembles for Few-Shot\n  Class-Incremental Learning", "abstract": "  Few-shot class-incremental learning (FSCIL) struggles to incrementally\nrecognize novel classes from few examples without catastrophic forgetting of\nold classes or overfitting to new classes. We propose TLCE, which ensembles\nmultiple pre-trained models to improve separation of novel and old classes.\nTLCE minimizes interference between old and new classes by mapping old class\nimages to quasi-orthogonal prototypes using episodic training. It then\nensembles diverse pre-trained models to better adapt to novel classes despite\ndata imbalance. Extensive experiments on various datasets demonstrate that our\ntransfer learning ensemble approach outperforms state-of-the-art FSCIL methods.\n", "date": "2023-12-07", "categories": []}
{"id": "2312.03309", "title": "Benchmarking Continual Learning from Cognitive Perspectives", "abstract": "  Continual learning addresses the problem of continuously acquiring and\ntransferring knowledge without catastrophic forgetting of old concepts. While\nhumans achieve continual learning via diverse neurocognitive mechanisms, there\nis a mismatch between cognitive properties and evaluation methods of continual\nlearning models. First, the measurement of continual learning models mostly\nrelies on evaluation metrics at a micro-level, which cannot characterize\ncognitive capacities of the model. Second, the measurement is method-specific,\nemphasizing model strengths in one aspect while obscuring potential weaknesses\nin other respects. To address these issues, we propose to integrate model\ncognitive capacities and evaluation metrics into a unified evaluation paradigm.\nWe first characterize model capacities via desiderata derived from cognitive\nproperties supporting human continual learning. The desiderata concern (1)\nadaptability in varying lengths of task sequence; (2) sensitivity to dynamic\ntask variations; and (3) efficiency in memory usage and training time\nconsumption. Then we design evaluation protocols for each desideratum to assess\ncognitive capacities of recent continual learning models. Experimental results\nshow that no method we consider has satisfied all the desiderata and is still\nfar away from realizing truly continual learning. Although some methods exhibit\nsome degree of adaptability and efficiency, no method is able to identify task\nrelationships when encountering dynamic task variations, or achieve a trade-off\nin learning similarities and differences between tasks. Inspired by these\nresults, we discuss possible factors that influence model performance in these\ndesiderata and provide guidance for the improvement of continual learning\nmodels.\n", "date": "2023-12-06", "categories": []}
{"id": "2211.16994", "title": "Continual Learning with Distributed Optimization: Does CoCoA Forget?", "abstract": "  We focus on the continual learning problem where the tasks arrive\nsequentially and the aim is to perform well on the newly arrived task without\nperformance degradation on the previously seen tasks. In contrast to the\ncontinual learning literature focusing on the centralized setting, we\ninvestigate the distributed estimation framework. We consider the\nwell-established distributed learning algorithm COCOA. We derive closed form\nexpressions for the iterations for the overparametrized case. We illustrate the\nconvergence and the error performance of the algorithm based on the\nover/under-parameterization of the problem. Our results show that depending on\nthe problem dimensions and data generation assumptions, COCOA can perform\ncontinual learning over a sequence of tasks, i.e., it can learn a new task\nwithout forgetting previously learned tasks, with access only to one task at a\ntime.\n", "date": "2022-11-30", "categories": []}
{"id": "2308.06053", "title": "Cost-effective On-device Continual Learning over Memory Hierarchy with\n  Miro", "abstract": "  Continual learning (CL) trains NN models incrementally from a continuous\nstream of tasks. To remember previously learned knowledge, prior studies store\nold samples over a memory hierarchy and replay them when new tasks arrive. Edge\ndevices that adopt CL to preserve data privacy are typically energy-sensitive\nand thus require high model accuracy while not compromising energy efficiency,\ni.e., cost-effectiveness. Our work is the first to explore the design space of\nhierarchical memory replay-based CL to gain insights into achieving\ncost-effectiveness on edge devices. We present Miro, a novel system runtime\nthat carefully integrates our insights into the CL framework by enabling it to\ndynamically configure the CL system based on resource states for the best\ncost-effectiveness. To reach this goal, Miro also performs online profiling on\nparameters with clear accuracy-energy trade-offs and adapts to optimal values\nwith low overhead. Extensive evaluations show that Miro significantly\noutperforms baseline systems we build for comparison, consistently achieving\nhigher cost-effectiveness.\n", "date": "2023-08-11", "categories": []}
{"id": "2305.13622", "title": "Continual Learning with Strong Experience Replay", "abstract": "  Continual Learning (CL) aims at incrementally learning new tasks without\nforgetting the knowledge acquired from old ones. Experience Replay (ER) is a\nsimple and effective rehearsal-based strategy, which optimizes the model with\ncurrent training data and a subset of old samples stored in a memory buffer. To\nfurther reduce forgetting, recent approaches extend ER with various techniques,\nsuch as model regularization and memory sampling. However, the prediction\nconsistency between the new model and the old one on current training data has\nbeen seldom explored, resulting in less knowledge preserved when few previous\nsamples are available. To address this issue, we propose a CL method with\nStrong Experience Replay (SER), which additionally utilizes future experiences\nmimicked on the current training data, besides distilling past experience from\nthe memory buffer. In our method, the updated model will produce approximate\noutputs as its original ones, which can effectively preserve the acquired\nknowledge. Experimental results on multiple image classification datasets show\nthat our SER method surpasses the state-of-the-art methods by a noticeable\nmargin.\n", "date": "2023-05-23", "categories": []}
{"id": "2308.14831", "title": "Continual Learning with Dynamic Sparse Training: Exploring Algorithms\n  for Effective Model Updates", "abstract": "  Continual learning (CL) refers to the ability of an intelligent system to\nsequentially acquire and retain knowledge from a stream of data with as little\ncomputational overhead as possible. To this end; regularization, replay,\narchitecture, and parameter isolation approaches were introduced to the\nliterature. Parameter isolation using a sparse network which enables to\nallocate distinct parts of the neural network to different tasks and also\nallows to share of parameters between tasks if they are similar. Dynamic Sparse\nTraining (DST) is a prominent way to find these sparse networks and isolate\nthem for each task. This paper is the first empirical study investigating the\neffect of different DST components under the CL paradigm to fill a critical\nresearch gap and shed light on the optimal configuration of DST for CL if it\nexists. Therefore, we perform a comprehensive study in which we investigate\nvarious DST components to find the best topology per task on well-known\nCIFAR100 and miniImageNet benchmarks in a task-incremental CL setup since our\nprimary focus is to evaluate the performance of various DST criteria, rather\nthan the process of mask selection. We found that, at a low sparsity level,\nErdos-R\\'enyi Kernel (ERK) initialization utilizes the backbone more\nefficiently and allows to effectively learn increments of tasks. At a high\nsparsity level, unless it is extreme, uniform initialization demonstrates a\nmore reliable and robust performance. In terms of growth strategy; performance\nis dependent on the defined initialization strategy and the extent of sparsity.\nFinally, adaptivity within DST components is a promising way for better\ncontinual learners.\n", "date": "2023-08-28", "categories": []}
{"id": "2312.01188", "title": "Efficient Expansion and Gradient Based Task Inference for Replay Free\n  Incremental Learning", "abstract": "  This paper proposes a simple but highly efficient expansion-based model for\ncontinual learning. The recent feature transformation, masking and\nfactorization-based methods are efficient, but they grow the model only over\nthe global or shared parameter. Therefore, these approaches do not fully\nutilize the previously learned information because the same task-specific\nparameter forgets the earlier knowledge. Thus, these approaches show limited\ntransfer learning ability. Moreover, most of these models have constant\nparameter growth for all tasks, irrespective of the task complexity. Our work\nproposes a simple filter and channel expansion based method that grows the\nmodel over the previous task parameters and not just over the global parameter.\nTherefore, it fully utilizes all the previously learned information without\nforgetting, which results in better knowledge transfer. The growth rate in our\nproposed model is a function of task complexity; therefore for a simple task,\nthe model has a smaller parameter growth while for complex tasks, the model\nrequires more parameters to adapt to the current task. Recent expansion based\nmodels show promising results for task incremental learning (TIL). However, for\nclass incremental learning (CIL), prediction of task id is a crucial challenge;\nhence, their results degrade rapidly as the number of tasks increase. In this\nwork, we propose a robust task prediction method that leverages entropy\nweighted data augmentations and the models gradient using pseudo labels. We\nevaluate our model on various datasets and architectures in the TIL, CIL and\ngenerative continual learning settings. The proposed approach shows\nstate-of-the-art results in all these settings. Our extensive ablation studies\nshow the efficacy of the proposed components.\n", "date": "2023-12-02", "categories": []}
{"id": "2312.01795", "title": "Distributed Continual Learning with CoCoA in High-dimensional Linear\n  Regression", "abstract": "  We consider estimation under scenarios where the signals of interest exhibit\nchange of characteristics over time. In particular, we consider the continual\nlearning problem where different tasks, e.g., data with different\ndistributions, arrive sequentially and the aim is to perform well on the newly\narrived task without performance degradation on the previously seen tasks. In\ncontrast to the continual learning literature focusing on the centralized\nsetting, we investigate the problem from a distributed estimation perspective.\nWe consider the well-established distributed learning algorithm COCOA, which\ndistributes the model parameters and the corresponding features over the\nnetwork. We provide exact analytical characterization for the generalization\nerror of COCOA under continual learning for linear regression in a range of\nscenarios, where overparameterization is of particular interest. These\nanalytical results characterize how the generalization error depends on the\nnetwork structure, the task similarity and the number of tasks, and show how\nthese dependencies are intertwined. In particular, our results show that the\ngeneralization error can be significantly reduced by adjusting the network\nsize, where the most favorable network size depends on task similarity and the\nnumber of tasks. We present numerical results verifying the theoretical\nanalysis and illustrate the continual learning performance of COCOA with a\ndigit classification task.\n", "date": "2023-12-04", "categories": []}
{"id": "2307.11046", "title": "A Definition of Continual Reinforcement Learning", "abstract": "  In a standard view of the reinforcement learning problem, an agent's goal is\nto efficiently identify a policy that maximizes long-term reward. However, this\nperspective is based on a restricted view of learning as finding a solution,\nrather than treating learning as endless adaptation. In contrast, continual\nreinforcement learning refers to the setting in which the best agents never\nstop learning. Despite the importance of continual reinforcement learning, the\ncommunity lacks a simple definition of the problem that highlights its\ncommitments and makes its primary concepts precise and clear. To this end, this\npaper is dedicated to carefully defining the continual reinforcement learning\nproblem. We formalize the notion of agents that \"never stop learning\" through a\nnew mathematical language for analyzing and cataloging agents. Using this new\nlanguage, we define a continual learning agent as one that can be understood as\ncarrying out an implicit search process indefinitely, and continual\nreinforcement learning as the setting in which the best agents are all\ncontinual learning agents. We provide two motivating examples, illustrating\nthat traditional views of multi-task reinforcement learning and continual\nsupervised learning are special cases of our definition. Collectively, these\ndefinitions and perspectives formalize many intuitive concepts at the heart of\nlearning, and open new research pathways surrounding continual learning agents.\n", "date": "2023-07-20", "categories": []}
{"id": "2312.00237", "title": "Negotiated Representations to Prevent Forgetting in Machine Learning\n  Applications", "abstract": "  Catastrophic forgetting is a significant challenge in the field of machine\nlearning, particularly in neural networks. When a neural network learns to\nperform well on a new task, it often forgets its previously acquired knowledge\nor experiences. This phenomenon occurs because the network adjusts its weights\nand connections to minimize the loss on the new task, which can inadvertently\noverwrite or disrupt the representations that were crucial for the previous\ntasks. As a result, the the performance of the network on earlier tasks\ndeteriorates, limiting its ability to learn and adapt to a sequence of tasks.\nIn this paper, we propose a novel method for preventing catastrophic forgetting\nin machine learning applications, specifically focusing on neural networks. Our\napproach aims to preserve the knowledge of the network across multiple tasks\nwhile still allowing it to learn new information effectively. We demonstrate\nthe effectiveness of our method by conducting experiments on various benchmark\ndatasets, including Split MNIST, Split CIFAR10, Split Fashion MNIST, and Split\nCIFAR100. These datasets are created by dividing the original datasets into\nseparate, non overlapping tasks, simulating a continual learning scenario where\nthe model needs to learn multiple tasks sequentially without forgetting the\nprevious ones. Our proposed method tackles the catastrophic forgetting problem\nby incorporating negotiated representations into the learning process, which\nallows the model to maintain a balance between retaining past experiences and\nadapting to new tasks. By evaluating our method on these challenging datasets,\nwe aim to showcase its potential for addressing catastrophic forgetting and\nimproving the performance of neural networks in continual learning settings.\n", "date": "2023-11-30", "categories": []}
{"id": "2311.18630", "title": "SATHUR: Self Augmenting Task Hallucinal Unified Representation for\n  Generalized Class Incremental Learning", "abstract": "  Class Incremental Learning (CIL) is inspired by the human ability to learn\nnew classes without forgetting previous ones. CIL becomes more challenging in\nreal-world scenarios when the samples in each incremental step are imbalanced.\nThis creates another branch of problem, called Generalized Class Incremental\nLearning (GCIL) where each incremental step is structured more realistically.\nGrow When Required (GWR) network, a type of Self-Organizing Map (SOM),\ndynamically create and remove nodes and edges for adaptive learning. GWR\nperforms incremental learning from feature vectors extracted by a Convolutional\nNeural Network (CNN), which acts as a feature extractor. The inherent ability\nof GWR to form distinct clusters, each corresponding to a class in the feature\nvector space, regardless of the order of samples or class imbalances, is well\nsuited to achieving GCIL. To enhance GWR's classification performance, a\nhigh-quality feature extractor is required. However, when the convolutional\nlayers are adapted at each incremental step, the GWR nodes corresponding to\nprior knowledge are subject to near-invalidation. This work introduces the Self\nAugmenting Task Hallucinal Unified Representation (SATHUR), which\nre-initializes the GWR network at each incremental step, aligning it with the\ncurrent feature extractor. Comprehensive experimental results demonstrate that\nour proposed method significantly outperforms other state-of-the-art GCIL\nmethods on CIFAR-100 and CORe50 datasets.\n", "date": "2023-08-13", "categories": []}
{"id": "2306.02027", "title": "Evolving Knowledge Mining for Class Incremental Segmentation", "abstract": "  Class Incremental Semantic Segmentation (CISS) has been a trend recently due\nto its great significance in real-world applications. Although the existing\nCISS methods demonstrate remarkable performance, they either leverage the\nhigh-level knowledge (feature) only while neglecting the rich and diverse\nknowledge in the low-level features, leading to poor old knowledge preservation\nand weak new knowledge exploration; or use multi-level features for knowledge\ndistillation by retraining a heavy backbone, which is computationally\nintensive. In this paper, we for the first time investigate the efficient\nmulti-grained knowledge reuse for CISS, and propose a novel method, Evolving\nkNowleDge minING (ENDING), employing a frozen backbone. ENDING incorporates two\nkey modules: evolving fusion and semantic enhancement, for dynamic and\ncomprehensive exploration of multi-grained knowledge. Evolving fusion is\ntailored to extract knowledge from individual low-level feature using a\npersonalized lightweight network, which is generated from a meta-net, taking\nthe high-level feature as input. This design enables the evolution of knowledge\nmining and fusing when applied to incremental new classes. In contrast,\nsemantic enhancement is specifically crafted to aggregate prototype-based\nsemantics from multi-level features, contributing to an enhanced\nrepresentation. We evaluate our method on two widely used benchmarks and\nconsistently demonstrate new state-of-the-art performance. The code is\navailable at https://github.com/zhiheLu/ENDING_ISS.\n", "date": "2023-06-03", "categories": []}
{"id": "2311.17450", "title": "Continual Learning for Image Segmentation with Dynamic Query", "abstract": "  Image segmentation based on continual learning exhibits a critical drop of\nperformance, mainly due to catastrophic forgetting and background shift, as\nthey are required to incorporate new classes continually. In this paper, we\npropose a simple, yet effective Continual Image Segmentation method with\nincremental Dynamic Query (CISDQ), which decouples the representation learning\nof both old and new knowledge with lightweight query embedding. CISDQ mainly\nincludes three contributions: 1) We define dynamic queries with adaptive\nbackground class to exploit past knowledge and learn future classes naturally.\n2) CISDQ proposes a class/instance-aware Query Guided Knowledge Distillation\nstrategy to overcome catastrophic forgetting by capturing the inter-class\ndiversity and intra-class identity. 3) Apart from semantic segmentation, CISDQ\nintroduce the continual learning for instance segmentation in which\ninstance-wise labeling and supervision are considered. Extensive experiments on\nthree datasets for two tasks (i.e., continual semantic and instance\nsegmentation are conducted to demonstrate that CISDQ achieves the\nstate-of-the-art performance, specifically, obtaining 4.4% and 2.9% mIoU\nimprovements for the ADE 100-10 (6 steps) setting and ADE 100-5 (11 steps)\nsetting.\n", "date": "2023-11-29", "categories": []}
{"id": "2311.17601", "title": "Continual Learning with Low Rank Adaptation", "abstract": "  Recent work using pretrained transformers has shown impressive performance\nwhen fine-tuned with data from the downstream problem of interest. However,\nthey struggle to retain that performance when the data characteristics changes.\nIn this paper, we focus on continual learning, where a pre-trained transformer\nis updated to perform well on new data, while retaining its performance on data\nit was previously trained on. Earlier works have tackled this primarily through\nmethods inspired from prompt tuning. We question this choice, and investigate\nthe applicability of Low Rank Adaptation (LoRA) to continual learning. On a\nrange of domain-incremental learning benchmarks, our LoRA-based solution,\nCoLoR, yields state-of-the-art performance, while still being as parameter\nefficient as the prompt tuning based methods.\n", "date": "2023-11-29", "categories": []}
{"id": "2311.17608", "title": "Adversarial Robust Memory-Based Continual Learner", "abstract": "  Despite the remarkable advances that have been made in continual learning,\nthe adversarial vulnerability of such methods has not been fully discussed. We\ndelve into the adversarial robustness of memory-based continual learning\nalgorithms and observe limited robustness improvement by directly applying\nadversarial training techniques. Preliminary studies reveal the twin challenges\nfor building adversarial robust continual learners: accelerated forgetting in\ncontinual learning and gradient obfuscation in adversarial robustness. In this\nstudy, we put forward a novel adversarial robust memory-based continual learner\nthat adjusts data logits to mitigate the forgetting of pasts caused by\nadversarial samples. Furthermore, we devise a gradient-based data selection\nmechanism to overcome the gradient obfuscation caused by limited stored data.\nThe proposed approach can widely integrate with existing memory-based continual\nlearning as well as adversarial training algorithms in a plug-and-play way.\nExtensive experiments on Split-CIFAR10/100 and Split-Tiny-ImageNet demonstrate\nthe effectiveness of our approach, achieving up to 8.13% higher accuracy for\nadversarial data.\n", "date": "2023-11-29", "categories": []}
{"id": "2211.13131", "title": "FeTrIL: Feature Translation for Exemplar-Free Class-Incremental Learning", "abstract": "  Exemplar-free class-incremental learning is very challenging due to the\nnegative effect of catastrophic forgetting. A balance between stability and\nplasticity of the incremental process is needed in order to obtain good\naccuracy for past as well as new classes. Existing exemplar-free\nclass-incremental methods focus either on successive fine tuning of the model,\nthus favoring plasticity, or on using a feature extractor fixed after the\ninitial incremental state, thus favoring stability. We introduce a method which\ncombines a fixed feature extractor and a pseudo-features generator to improve\nthe stability-plasticity balance. The generator uses a simple yet effective\ngeometric translation of new class features to create representations of past\nclasses, made of pseudo-features. The translation of features only requires the\nstorage of the centroid representations of past classes to produce their\npseudo-features. Actual features of new classes and pseudo-features of past\nclasses are fed into a linear classifier which is trained incrementally to\ndiscriminate between all classes. The incremental process is much faster with\nthe proposed method compared to mainstream ones which update the entire deep\nmodel. Experiments are performed with three challenging datasets, and different\nincremental settings. A comparison with ten existing methods shows that our\nmethod outperforms the others in most cases.\n", "date": "2022-11-23", "categories": []}
{"id": "2311.16206", "title": "Continual Instruction Tuning for Large Multimodal Models", "abstract": "  Instruction tuning is now a widely adopted approach to aligning large\nmultimodal models (LMMs) to follow human intent. It unifies the data format of\nvision-language tasks, enabling multi-task joint training. However,\nvision-language tasks are constantly being created in practice. Instead of\nalways re-training LMMs when new tasks arrive, continual learning offers\nflexibility for models to continually and efficiently exploit the evolving\ndata. This work aims to explore the following two questions: 1) Do LMMs still\nsuffer from catastrophic forgetting in continual instruction tuning? 2) Are the\nexisting three classes of continual learning methods still applicable to the\ncontinual instruction tuning of LMMs? An extensive study is conducted to\naddress the above questions. First, we establish the first benchmark in this\nsetting and reveal that catastrophic forgetting is still observed when\ncontinually instruction-tuning LMMs. However, the multi-task joint instruction\ntuning can facilitate the model's continual learning ability and mitigate\nforgetting. Second, we integrate and adapt classic continual learning methods\nto our context, demonstrating the efficacy of data replay and model expansion\nstrategies across diverse scenarios. In contrast, regularization-based methods\nonly perform well on models that have been jointly instruction-tuned on\nmultiple tasks. Third, we delve into the correlation and forgetting dynamics\nbetween vision-language task pairs and propose task-similarity-informed\nregularization and model expansion methods for continual instruction tuning of\nLMMs. Experimental results show that our approach consistently boosts the\nmodel's performance.\n", "date": "2023-11-27", "categories": []}
{"id": "2311.16485", "title": "Class-Adaptive Sampling Policy for Efficient Continual Learning", "abstract": "  Continual learning (CL) aims to acquire new knowledge while preserving\ninformation from previous experiences without forgetting. Though buffer-based\nmethods (i.e., retaining samples from previous tasks) have achieved acceptable\nperformance, determining how to allocate the buffer remains a critical\nchallenge. Most recent research focuses on refining these methods but often\nfails to sufficiently consider the varying influence of samples on the learning\nprocess, and frequently overlooks the complexity of the classes/concepts being\nlearned. Generally, these methods do not directly take into account the\ncontribution of individual classes. However, our investigation indicates that\nmore challenging classes necessitate preserving a larger number of samples\ncompared to less challenging ones. To address this issue, we propose a novel\nmethod and policy named 'Class-Adaptive Sampling Policy' (CASP), which\ndynamically allocates storage space within the buffer. By utilizing concepts of\nclass contribution and difficulty, CASP adaptively manages buffer space,\nallowing certain classes to occupy a larger portion of the buffer while\nreducing storage for others. This approach significantly improves the\nefficiency of knowledge retention and utilization. CASP provides a versatile\nsolution to boost the performance and efficiency of CL. It meets the demand for\ndynamic buffer allocation, accommodating the varying contributions of different\nclasses and their learning complexities over time.\n", "date": "2023-11-27", "categories": []}
{"id": "2311.14905", "title": "Class Gradient Projection For Continual Learning", "abstract": "  Catastrophic forgetting is one of the most critical challenges in Continual\nLearning (CL). Recent approaches tackle this problem by projecting the gradient\nupdate orthogonal to the gradient subspace of existing tasks. While the results\nare remarkable, those approaches ignore the fact that these calculated\ngradients are not guaranteed to be orthogonal to the gradient subspace of each\nclass due to the class deviation in tasks, e.g., distinguishing \"Man\" from\n\"Sea\" v.s. differentiating \"Boy\" from \"Girl\". Therefore, this strategy may\nstill cause catastrophic forgetting for some classes. In this paper, we propose\nClass Gradient Projection (CGP), which calculates the gradient subspace from\nindividual classes rather than tasks. Gradient update orthogonal to the\ngradient subspace of existing classes can be effectively utilized to minimize\ninterference from other classes. To improve the generalization and efficiency,\nwe further design a Base Refining (BR) algorithm to combine similar classes and\nrefine class bases dynamically. Moreover, we leverage a contrastive learning\nmethod to improve the model's ability to handle unseen tasks. Extensive\nexperiments on benchmark datasets demonstrate the effectiveness of our proposed\napproach. It improves the previous methods by 2.0% on the CIFAR-100 dataset.\n", "date": "2023-11-25", "categories": []}
{"id": "2311.14911", "title": "CUCL: Codebook for Unsupervised Continual Learning", "abstract": "  The focus of this study is on Unsupervised Continual Learning (UCL), as it\npresents an alternative to Supervised Continual Learning which needs\nhigh-quality manual labeled data. The experiments under the UCL paradigm\nindicate a phenomenon where the results on the first few tasks are suboptimal.\nThis phenomenon can render the model inappropriate for practical applications.\nTo address this issue, after analyzing the phenomenon and identifying the lack\nof diversity as a vital factor, we propose a method named Codebook for\nUnsupervised Continual Learning (CUCL) which promotes the model to learn\ndiscriminative features to complete the class boundary. Specifically, we first\nintroduce a Product Quantization to inject diversity into the representation\nand apply a cross quantized contrastive loss between the original\nrepresentation and the quantized one to capture discriminative information.\nThen, based on the quantizer, we propose an effective Codebook Rehearsal to\naddress catastrophic forgetting. This study involves conducting extensive\nexperiments on CIFAR100, TinyImageNet, and MiniImageNet benchmark datasets. Our\nmethod significantly boosts the performances of supervised and unsupervised\nmethods. For instance, on TinyImageNet, our method led to a relative\nimprovement of 12.76% and 7% when compared with Simsiam and BYOL, respectively.\n", "date": "2023-11-25", "categories": []}
{"id": "2311.15276", "title": "Efficient Rehearsal Free Zero Forgetting Continual Learning using\n  Adaptive Weight Modulation", "abstract": "  Artificial neural networks encounter a notable challenge known as continual\nlearning, which involves acquiring knowledge of multiple tasks over an extended\nperiod. This challenge arises due to the tendency of previously learned weights\nto be adjusted to suit the objectives of new tasks, resulting in a phenomenon\ncalled catastrophic forgetting. Most approaches to this problem seek a balance\nbetween maximizing performance on the new tasks and minimizing the forgetting\nof previous tasks. In contrast, our approach attempts to maximize the\nperformance of the new task, while ensuring zero forgetting. This is\naccomplished by creating a task-specific modulation parameters for each task.\nOnly these would be learnable parameters during learning of consecutive tasks.\nThrough comprehensive experimental evaluations, our model demonstrates superior\nperformance in acquiring and retaining novel tasks that pose difficulties for\nother multi-task models. This emphasizes the efficacy of our approach in\npreventing catastrophic forgetting while accommodating the acquisition of new\ntasks\n", "date": "2023-11-26", "categories": []}
{"id": "2311.11995", "title": "BrainWash: A Poisoning Attack to Forget in Continual Learning", "abstract": "  Continual learning has gained substantial attention within the deep learning\ncommunity, offering promising solutions to the challenging problem of\nsequential learning. Yet, a largely unexplored facet of this paradigm is its\nsusceptibility to adversarial attacks, especially with the aim of inducing\nforgetting. In this paper, we introduce \"BrainWash,\" a novel data poisoning\nmethod tailored to impose forgetting on a continual learner. By adding the\nBrainWash noise to a variety of baselines, we demonstrate how a trained\ncontinual learner can be induced to forget its previously learned tasks\ncatastrophically, even when using these continual learning baselines. An\nimportant feature of our approach is that the attacker requires no access to\nprevious tasks' data and is armed merely with the model's current parameters\nand the data belonging to the most recent task. Our extensive experiments\nhighlight the efficacy of BrainWash, showcasing degradation in performance\nacross various regularization-based continual learning methods.\n", "date": "2023-11-20", "categories": []}
{"id": "2311.13623", "title": "Density Distribution-based Learning Framework for Addressing Online\n  Continual Learning Challenges", "abstract": "  In this paper, we address the challenges of online Continual Learning (CL) by\nintroducing a density distribution-based learning framework. CL, especially the\nClass Incremental Learning, enables adaptation to new test distributions while\ncontinuously learning from a single-pass training data stream, which is more in\nline with the practical application requirements of real-world scenarios.\nHowever, existing CL methods often suffer from catastrophic forgetting and\nhigher computing costs due to complex algorithm designs, limiting their\npractical use. Our proposed framework overcomes these limitations by achieving\nsuperior average accuracy and time-space efficiency, bridging the performance\ngap between CL and classical machine learning. Specifically, we adopt an\nindependent Generative Kernel Density Estimation (GKDE) model for each CL task.\nDuring the testing stage, the GKDEs utilize a self-reported max probability\ndensity value to determine which one is responsible for predicting incoming\ntest instances. A GKDE-based learning objective can ensure that samples with\nthe same label are grouped together, while dissimilar instances are pushed\nfarther apart. Extensive experiments conducted on multiple CL datasets validate\nthe effectiveness of our proposed framework. Our method outperforms popular CL\napproaches by a significant margin, while maintaining competitive time-space\nefficiency, making our framework suitable for real-world applications. Code\nwill be available at https://github.com/xxxx/xxxx.\n", "date": "2023-11-22", "categories": []}
{"id": "2310.09886", "title": "Lifelong Sequence Generation with Dynamic Module Expansion and\n  Adaptation", "abstract": "  Lifelong sequence generation (LSG), a problem in continual learning, aims to\ncontinually train a model on a sequence of generation tasks to learn constantly\nemerging new generation patterns while avoiding the forgetting of previous\nknowledge. Existing LSG methods mainly focus on maintaining old knowledge while\npaying little attention to knowledge transfer across tasks. In contrast, humans\ncan better learn new tasks by leveraging previously acquired knowledge from\nsimilar tasks. Inspired by the learning paradigm of humans, we propose Dynamic\nModule Expansion and Adaptation (DMEA), which enables the model to dynamically\ndetermine the architecture for acquiring new knowledge based on task\ncorrelation and select the most similar previous tasks to facilitate adaptation\nto new tasks. In addition, as the learning process can easily be biased towards\nthe current task which might cause more severe forgetting of previously learned\nknowledge, we propose dynamic gradient scaling to balance the learning of the\ncurrent task and replayed tasks. With extensive experiments, we demonstrate\nthat DMEA can consistently outperform existing methods in different LSG\nsettings.\n", "date": "2023-10-15", "categories": []}
{"id": "2209.08660", "title": "Learn the Time to Learn: Replay Scheduling in Continual Learning", "abstract": "  Replay methods are known to be successful at mitigating catastrophic\nforgetting in continual learning scenarios despite having limited access to\nhistorical data. However, storing historical data is cheap in many real-world\nsettings, yet replaying all historical data is often prohibited due to\nprocessing time constraints. In such settings, we propose that continual\nlearning systems should learn the time to learn and schedule which tasks to\nreplay at different time steps. We first demonstrate the benefits of our\nproposal by using Monte Carlo tree search to find a proper replay schedule, and\nshow that the found replay schedules can outperform fixed scheduling policies\nwhen combined with various replay methods in different continual learning\nsettings. Additionally, we propose a framework for learning replay scheduling\npolicies with reinforcement learning. We show that the learned policies can\ngeneralize better in new continual learning scenarios compared to equally\nreplaying all seen tasks, without added computational cost. Our study reveals\nthe importance of learning the time to learn in continual learning, which\nbrings current research closer to real-world needs.\n", "date": "2022-09-18", "categories": []}
{"id": "2302.10396", "title": "Assessing Domain Gap for Continual Domain Adaptation in Object Detection", "abstract": "  To ensure reliable object detection in autonomous systems, the detector must\nbe able to adapt to changes in appearance caused by environmental factors such\nas time of day, weather, and seasons. Continually adapting the detector to\nincorporate these changes is a promising solution, but it can be\ncomputationally costly. Our proposed approach is to selectively adapt the\ndetector only when necessary, using new data that does not have the same\ndistribution as the current training data. To this end, we investigate three\npopular metrics for domain gap evaluation and find that there is a correlation\nbetween the domain gap and detection accuracy. Therefore, we apply the domain\ngap as a criterion to decide when to adapt the detector. Our experiments show\nthat our approach has the potential to improve the efficiency of the detector's\noperation in real-world scenarios, where environmental conditions change in a\ncyclical manner, without sacrificing the overall performance of the detector.\nOur code is publicly available at https://github.com/dadung/DGE-CDA.\n", "date": "2023-02-21", "categories": []}
{"id": "2311.07784", "title": "A Data-Free Approach to Mitigate Catastrophic Forgetting in Federated\n  Class Incremental Learning for Vision Tasks", "abstract": "  Deep learning models often suffer from forgetting previously learned\ninformation when trained on new data. This problem is exacerbated in federated\nlearning (FL), where the data is distributed and can change independently for\neach user. Many solutions are proposed to resolve this catastrophic forgetting\nin a centralized setting. However, they do not apply directly to FL because of\nits unique complexities, such as privacy concerns and resource limitations. To\novercome these challenges, this paper presents a framework for\n$\\textbf{federated class incremental learning}$ that utilizes a generative\nmodel to synthesize samples from past distributions. This data can be later\nexploited alongside the training data to mitigate catastrophic forgetting. To\npreserve privacy, the generative model is trained on the server using data-free\nmethods at the end of each task without requesting data from clients. Moreover,\nour solution does not demand the users to store old data or models, which gives\nthem the freedom to join/leave the training at any time. Additionally, we\nintroduce SuperImageNet, a new regrouping of the ImageNet dataset specifically\ntailored for federated continual learning. We demonstrate significant\nimprovements compared to existing baselines through extensive experiments on\nmultiple datasets.\n", "date": "2023-11-13", "categories": []}
{"id": "2311.10787", "title": "Assurance for Deployed Continual Learning Systems", "abstract": "  The future success of the Navy will depend, in part, on artificial\nintelligence. In practice, many artificially intelligent algorithms, and in\nparticular deep learning models, rely on continual learning to maintain\nperformance in dynamic environments. The software requires adaptation to\nmaintain its initial level of performance in unseen situations. However, if not\nmonitored properly, continual learning may lead to several issues including\ncatastrophic forgetting in which a trained model forgets previously learned\ntasks when being retrained on new data. The authors created a new framework for\nsafely performing continual learning with the goal of pairing this safety\nframework with a deep learning computer vision algorithm to allow for safe and\nhigh-performing automatic deck tracking on carriers and amphibious assault\nships. The safety framework includes several features, such as an ensemble of\nconvolutional neural networks to perform image classification, a manager to\nrecord confidences and determine the best answer from the ensemble, a model of\nthe environment to predict when the system may fail to meet minimum performance\nmetrics, a performance monitor to log system and domain performance and check\nagainst requirements, and a retraining component to update the ensemble and\nmanager to maintain performance. The authors validated the proposed method\nusing extensive simulation studies based on dynamic image classification. The\nauthors showed the safety framework could probabilistically detect out of\ndistribution data. The results also show the framework can detect when the\nsystem is no longer performing safely and can significantly extend the working\nenvelope of an image classifier.\n", "date": "2023-11-16", "categories": []}
{"id": "2311.10919", "title": "PACOL: Poisoning Attacks Against Continual Learners", "abstract": "  Continual learning algorithms are typically exposed to untrusted sources that\ncontain training data inserted by adversaries and bad actors. An adversary can\ninsert a small number of poisoned samples, such as mislabeled samples from\npreviously learned tasks, or intentional adversarial perturbed samples, into\nthe training datasets, which can drastically reduce the model's performance. In\nthis work, we demonstrate that continual learning systems can be manipulated by\nmalicious misinformation and present a new category of data poisoning attacks\nspecific for continual learners, which we refer to as {\\em Poisoning Attacks\nAgainst Continual Learners} (PACOL). The effectiveness of labeling flipping\nattacks inspires PACOL; however, PACOL produces attack samples that do not\nchange the sample's label and produce an attack that causes catastrophic\nforgetting. A comprehensive set of experiments shows the vulnerability of\ncommonly used generative replay and regularization-based continual learning\napproaches against attack methods. We evaluate the ability of label-flipping\nand a new adversarial poison attack, namely PACOL proposed in this work, to\nforce the continual learning system to forget the knowledge of a learned\ntask(s). More specifically, we compared the performance degradation of\ncontinual learning systems trained on benchmark data streams with and without\npoisoning attacks. Moreover, we discuss the stealthiness of the attacks in\nwhich we test the success rate of data sanitization defense and other outlier\ndetection-based defenses for filtering out adversarial samples.\n", "date": "2023-11-18", "categories": []}
{"id": "2311.11420", "title": "LifeLearner: Hardware-Aware Meta Continual Learning System for Embedded\n  Computing Platforms", "abstract": "  Continual Learning (CL) allows applications such as user personalization and\nhousehold robots to learn on the fly and adapt to context. This is an important\nfeature when context, actions, and users change. However, enabling CL on\nresource-constrained embedded systems is challenging due to the limited labeled\ndata, memory, and computing capacity. In this paper, we propose LifeLearner, a\nhardware-aware meta continual learning system that drastically optimizes system\nresources (lower memory, latency, energy consumption) while ensuring high\naccuracy. Specifically, we (1) exploit meta-learning and rehearsal strategies\nto explicitly cope with data scarcity issues and ensure high accuracy, (2)\neffectively combine lossless and lossy compression to significantly reduce the\nresource requirements of CL and rehearsal samples, and (3) developed\nhardware-aware system on embedded and IoT platforms considering the hardware\ncharacteristics. As a result, LifeLearner achieves near-optimal CL performance,\nfalling short by only 2.8% on accuracy compared to an Oracle baseline. With\nrespect to the state-of-the-art (SOTA) Meta CL method, LifeLearner drastically\nreduces the memory footprint (by 178.7x), end-to-end latency by 80.8-94.2%, and\nenergy consumption by 80.9-94.2%. In addition, we successfully deployed\nLifeLearner on two edge devices and a microcontroller unit, thereby enabling\nefficient CL on resource-constrained platforms where it would be impractical to\nrun SOTA methods and the far-reaching deployment of adaptable CL in a\nubiquitous manner. Code is available at\nhttps://github.com/theyoungkwon/LifeLearner.\n", "date": "2023-11-19", "categories": []}
{"id": "2311.11557", "title": "Replay-enhanced Continual Reinforcement Learning", "abstract": "  Replaying past experiences has proven to be a highly effective approach for\naverting catastrophic forgetting in supervised continual learning. However,\nsome crucial factors are still largely ignored, making it vulnerable to serious\nfailure, when used as a solution to forgetting in continual reinforcement\nlearning, even in the context of perfect memory where all data of previous\ntasks are accessible in the current task. On the one hand, since most\nreinforcement learning algorithms are not invariant to the reward scale, the\npreviously well-learned tasks (with high rewards) may appear to be more salient\nto the current learning process than the current task (with small initial\nrewards). This causes the agent to concentrate on those salient tasks at the\nexpense of generality on the current task. On the other hand, offline learning\non replayed tasks while learning a new task may induce a distributional shift\nbetween the dataset and the learned policy on old tasks, resulting in\nforgetting. In this paper, we introduce RECALL, a replay-enhanced method that\ngreatly improves the plasticity of existing replay-based methods on new tasks\nwhile effectively avoiding the recurrence of catastrophic forgetting in\ncontinual reinforcement learning. RECALL leverages adaptive normalization on\napproximate targets and policy distillation on old tasks to enhance generality\nand stability, respectively. Extensive experiments on the Continual World\nbenchmark show that RECALL performs significantly better than purely perfect\nmemory replay, and achieves comparable or better overall performance against\nstate-of-the-art continual learning methods.\n", "date": "2023-11-20", "categories": []}
{"id": "2311.10271", "title": "Prompt Pool based Class-Incremental Continual Learning for Dialog State\n  Tracking", "abstract": "  Continual learning is crucial for dialog state tracking (DST) in dialog\nsystems, since requirements from users for new functionalities are often\nencountered. However, most of existing continual learning methods for DST\nrequire task identities during testing, which is a severe limit in real-world\napplications. In this paper, we aim to address continual learning of DST in the\nclass-incremental scenario (namely the task identity is unknown in testing).\nInspired by the recently emerging prompt tuning method that performs well on\ndialog systems, we propose to use the prompt pool method, where we maintain a\npool of key-value paired prompts and select prompts from the pool according to\nthe distance between the dialog history and the prompt keys. The proposed\nmethod can automatically identify tasks and select appropriate prompts during\ntesting. We conduct experiments on Schema-Guided Dialog dataset (SGD) and\nanother dataset collected from a real-world dialog application. Experiment\nresults show that the prompt pool method achieves much higher joint goal\naccuracy than the baseline. After combining with a rehearsal buffer, the model\nperformance can be further improved.\n", "date": "2023-11-17", "categories": []}
{"id": "2302.00903", "title": "No One Left Behind: Real-World Federated Class-Incremental Learning", "abstract": "  Federated learning (FL) is a hot collaborative training framework via\naggregating model parameters of decentralized local clients. However, most FL\nmethods unreasonably assume data categories of FL framework are known and fixed\nin advance. Moreover, some new local clients that collect novel categories\nunseen by other clients may be introduced to FL training irregularly. These\nissues render global model to undergo catastrophic forgetting on old\ncategories, when local clients receive new categories consecutively under\nlimited memory of storing old categories. To tackle the above issues, we\npropose a novel Local-Global Anti-forgetting (LGA) model. It ensures no local\nclients are left behind as they learn new classes continually, by addressing\nlocal and global catastrophic forgetting. Specifically, considering tackling\nclass imbalance of local client to surmount local forgetting, we develop a\ncategory-balanced gradient-adaptive compensation loss and a category\ngradient-induced semantic distillation loss. They can balance heterogeneous\nforgetting speeds of hard-to-forget and easy-to-forget old categories, while\nensure consistent class-relations within different tasks. Moreover, a proxy\nserver is designed to tackle global forgetting caused by Non-IID class\nimbalance between different clients. It augments perturbed prototype images of\nnew categories collected from local clients via self-supervised prototype\naugmentation, thus improving robustness to choose the best old global model for\nlocal-side semantic distillation loss. Experiments on representative datasets\nverify superior performance of our model against comparison methods. The code\nis available at https://github.com/JiahuaDong/LGA.\n", "date": "2023-02-02", "categories": []}
{"id": "2311.09632", "title": "Online Continual Knowledge Learning for Language Models", "abstract": "  Large Language Models (LLMs) serve as repositories of extensive world\nknowledge, enabling them to perform tasks such as question-answering and\nfact-checking. However, this knowledge can become obsolete as global contexts\nchange. In this paper, we introduce a novel problem in the realm of continual\nlearning: Online Continual Knowledge Learning (OCKL). This problem formulation\naims to manage the dynamic nature of world knowledge in LMs under real-time\nconstraints. We propose a new benchmark and evaluation metric designed to\nmeasure both the rate of new knowledge acquisition and the retention of\npreviously learned knowledge. Our empirical evaluation, conducted using a\nvariety of state-of-the-art methods, establishes robust base-lines for OCKL.\nOur results reveal that existing continual learning approaches are\nunfortunately insufficient for tackling the unique challenges posed by OCKL. We\nidentify key factors that influence the trade-off between knowledge acquisition\nand retention, thereby advancing our understanding of how to train LMs in a\ncontinually evolving environment.\n", "date": "2023-11-16", "categories": []}
{"id": "2311.08764", "title": "Combining Past, Present and Future: A Self-Supervised Approach for Class\n  Incremental Learning", "abstract": "  Class Incremental Learning (CIL) aims to handle the scenario where data of\nnovel classes occur continuously and sequentially. The model should recognize\nthe sequential novel classes while alleviating the catastrophic forgetting. In\nthe self-supervised manner, it becomes more challenging to avoid the conflict\nbetween the feature embedding spaces of novel classes and old ones without any\nclass labels. To address the problem, we propose a self-supervised CIL\nframework CPPF, meaning Combining Past, Present and Future. In detail, CPPF\nconsists of a prototype clustering module (PC), an embedding space reserving\nmodule (ESR) and a multi-teacher distillation module (MTD). 1) The PC and the\nESR modules reserve embedding space for subsequent phases at the prototype\nlevel and the feature level respectively to prepare for knowledge learned in\nthe future. 2) The MTD module maintains the representations of the current\nphase without the interference of past knowledge. One of the teacher networks\nretains the representations of the past phases, and the other teacher network\ndistills relation information of the current phase to the student network.\nExtensive experiments on CIFAR100 and ImageNet100 datasets demonstrate that our\nproposed method boosts the performance of self-supervised class incremental\nlearning. We will release code in the near future.\n", "date": "2023-11-15", "categories": []}
{"id": "2303.09447", "title": "Steering Prototypes with Prompt-tuning for Rehearsal-free Continual\n  Learning", "abstract": "  In the context of continual learning, prototypes-as representative class\nembeddings-offer advantages in memory conservation and the mitigation of\ncatastrophic forgetting. However, challenges related to semantic drift and\nprototype interference persist. In this study, we introduce the Contrastive\nPrototypical Prompt (CPP) approach. Through task-specific prompt-tuning,\nunderpinned by a contrastive learning objective, we effectively address both\naforementioned challenges. Our evaluations on four challenging\nclass-incremental benchmarks reveal that CPP achieves a significant 4% to 6%\nimprovement over state-of-the-art methods. Importantly, CPP operates without a\nrehearsal buffer and narrows the performance divergence between continual and\noffline joint-learning, suggesting an innovative scheme for Transformer-based\ncontinual learning systems.\n", "date": "2023-03-16", "categories": []}
{"id": "2311.06828", "title": "Towards Continual Reinforcement Learning for Quadruped Robots", "abstract": "  Quadruped robots have emerged as an evolving technology that currently\nleverages simulators to develop a robust controller capable of functioning in\nthe real-world without the need for further training. However, since it is\nimpossible to predict all possible real-world situations, our research explores\nthe possibility of enabling them to continue learning even after their\ndeployment. To this end, we designed two continual learning scenarios,\nsequentially training the robot on different environments while simultaneously\nevaluating its performance across all of them. Our approach sheds light on the\nextent of both forward and backward skill transfer, as well as the degree to\nwhich the robot might forget previously acquired skills. By addressing these\nfactors, we hope to enhance the adaptability and performance of quadruped\nrobots in real-world scenarios.\n", "date": "2023-11-12", "categories": []}
{"id": "2311.06921", "title": "Concept Matching: Clustering-based Federated Continual Learning", "abstract": "  Federated Continual Learning (FCL) has emerged as a promising paradigm that\ncombines Federated Learning (FL) and Continual Learning (CL). To achieve good\nmodel accuracy, FCL needs to tackle catastrophic forgetting due to concept\ndrift over time in CL, and to overcome the potential interference among clients\nin FL. We propose Concept Matching (CM), a clustering-based framework for FCL\nto address these challenges. The CM framework groups the client models into\nconcept model clusters, and then builds different global models to capture\ndifferent concepts in FL over time. In each round, the server sends the global\nconcept models to the clients. To avoid catastrophic forgetting, each client\nselects the concept model best-matching the concept of the current data for\nfurther fine-tuning. To avoid interference among client models with different\nconcepts, the server clusters the models representing the same concept,\naggregates the model weights in each cluster, and updates the global concept\nmodel with the cluster model of the same concept. Since the server does not\nknow the concepts captured by the aggregated cluster models, we propose a novel\nserver concept matching algorithm that effectively updates a global concept\nmodel with a matching cluster model. The CM framework provides flexibility to\nuse different clustering, aggregation, and concept matching algorithms. The\nevaluation demonstrates that CM outperforms state-of-the-art systems and scales\nwell with the number of clients and the model size.\n", "date": "2023-11-12", "categories": []}
{"id": "2308.14991", "title": "Incorporating Neuro-Inspired Adaptability for Continual Learning in\n  Artificial Intelligence", "abstract": "  Continual learning aims to empower artificial intelligence (AI) with strong\nadaptability to the real world. For this purpose, a desirable solution should\nproperly balance memory stability with learning plasticity, and acquire\nsufficient compatibility to capture the observed distributions. Existing\nadvances mainly focus on preserving memory stability to overcome catastrophic\nforgetting, but remain difficult to flexibly accommodate incremental changes as\nbiological intelligence (BI) does. By modeling a robust Drosophila learning\nsystem that actively regulates forgetting with multiple learning modules, here\nwe propose a generic approach that appropriately attenuates old memories in\nparameter distributions to improve learning plasticity, and accordingly\ncoordinates a multi-learner architecture to ensure solution compatibility.\nThrough extensive theoretical and empirical validation, our approach not only\nclearly enhances the performance of continual learning, especially over\nsynaptic regularization methods in task-incremental settings, but also\npotentially advances the understanding of neurological adaptive mechanisms,\nserving as a novel paradigm to progress AI and BI together.\n", "date": "2023-08-29", "categories": []}
{"id": "2311.04301", "title": "Class-Incremental Continual Learning for General Purpose Healthcare\n  Models", "abstract": "  Healthcare clinics regularly encounter dynamic data that changes due to\nvariations in patient populations, treatment policies, medical devices, and\nemerging disease patterns. Deep learning models can suffer from catastrophic\nforgetting when fine-tuned in such scenarios, causing poor performance on\npreviously learned tasks. Continual learning allows learning on new tasks\nwithout performance drop on previous tasks. In this work, we investigate the\nperformance of continual learning models on four different medical imaging\nscenarios involving ten classification datasets from diverse modalities,\nclinical specialties, and hospitals. We implement various continual learning\napproaches and evaluate their performance in these scenarios. Our results\ndemonstrate that a single model can sequentially learn new tasks from different\nspecialties and achieve comparable performance to naive methods. These findings\nindicate the feasibility of recycling or sharing models across the same or\ndifferent medical specialties, offering another step towards the development of\ngeneral-purpose medical imaging AI that can be shared across institutions.\n", "date": "2023-11-07", "categories": []}
{"id": "2305.12030", "title": "Learning Continually on a Sequence of Graphs -- The Dynamical System Way", "abstract": "  Continual learning~(CL) is a field concerned with learning a series of\ninter-related task with the tasks typically defined in the sense of either\nregression or classification. In recent years, CL has been studied extensively\nwhen these tasks are defined using Euclidean data -- data, such as images, that\ncan be described by a set of vectors in an n-dimensional real space. However,\nthe literature is quite sparse, when the data corresponding to a CL task is\nnonEuclidean -- data , such as graphs, point clouds or manifold, where the\nnotion of similarity in the sense of Euclidean metric does not hold. For\ninstance, a graph is described by a tuple of vertices and edges and\nsimilarities between two graphs is not well defined through a Euclidean metric.\nDue to this fundamental nature of the data, developing CL for nonEuclidean data\npresents several theoretical and methodological challenges. In particular, CL\nfor graphs requires explicit modelling of nonstationary behavior of vertices\nand edges and their effects on the learning problem. Therefore, in this work,\nwe develop a adaptive dynamic programming viewpoint for CL with graphs. In this\nwork, we formulate a two-player sequential game between the act of learning new\ntasks~(generalization) and remembering previously learned tasks~(forgetting).\nWe prove mathematically the existence of a solution to the game and demonstrate\nconvergence to the solution of the game. Finally, we demonstrate the efficacy\nof our method on a number of graph benchmarks with a comprehensive ablation\nstudy while establishing state-of-the-art performance.\n", "date": "2023-05-19", "categories": []}
{"id": "2308.09544", "title": "Adapt Your Teacher: Improving Knowledge Distillation for Exemplar-free\n  Continual Learning", "abstract": "  In this work, we investigate exemplar-free class incremental learning (CIL)\nwith knowledge distillation (KD) as a regularization strategy, aiming to\nprevent forgetting. KD-based methods are successfully used in CIL, but they\noften struggle to regularize the model without access to exemplars of the\ntraining data from previous tasks. Our analysis reveals that this issue\noriginates from substantial representation shifts in the teacher network when\ndealing with out-of-distribution data. This causes large errors in the KD loss\ncomponent, leading to performance degradation in CIL models. Inspired by recent\ntest-time adaptation methods, we introduce Teacher Adaptation (TA), a method\nthat concurrently updates the teacher and the main models during incremental\ntraining. Our method seamlessly integrates with KD-based CIL approaches and\nallows for consistent enhancement of their performance across multiple\nexemplar-free CIL benchmarks. The source code for our method is available at\nhttps://github.com/fszatkowski/cl-teacher-adaptation.\n", "date": "2023-08-18", "categories": []}
{"id": "2311.00548", "title": "Continual atlas-based segmentation of prostate MRI", "abstract": "  Continual learning (CL) methods designed for natural image classification\noften fail to reach basic quality standards for medical image segmentation.\nAtlas-based segmentation, a well-established approach in medical imaging,\nincorporates domain knowledge on the region of interest, leading to\nsemantically coherent predictions. This is especially promising for CL, as it\nallows us to leverage structural information and strike an optimal balance\nbetween model rigidity and plasticity over time. When combined with\nprivacy-preserving prototypes, this process offers the advantages of\nrehearsal-based CL without compromising patient privacy. We propose Atlas\nReplay, an atlas-based segmentation approach that uses prototypes to generate\nhigh-quality segmentation masks through image registration that maintain\nconsistency even as the training distribution changes. We explore how our\nproposed method performs compared to state-of-the-art CL methods in terms of\nknowledge transferability across seven publicly available prostate segmentation\ndatasets. Prostate segmentation plays a vital role in diagnosing prostate\ncancer, however, it poses challenges due to substantial anatomical variations,\nbenign structural differences in older age groups, and fluctuating acquisition\nparameters. Our results show that Atlas Replay is both robust and generalizes\nwell to yet-unseen domains while being able to maintain knowledge, unlike\nend-to-end segmentation methods. Our code base is available under\nhttps://github.com/MECLabTUDA/Atlas-Replay.\n", "date": "2023-11-01", "categories": []}
{"id": "2311.02428", "title": "Task Arithmetic with LoRA for Continual Learning", "abstract": "  Continual learning refers to the problem where the training data is available\nin sequential chunks, termed \"tasks\". The majority of progress in continual\nlearning has been stunted by the problem of catastrophic forgetting, which is\ncaused by sequential training of the model on streams of data. Moreover, it\nbecomes computationally expensive to sequentially train large models multiple\ntimes. To mitigate both of these problems at once, we propose a novel method to\ncontinually train transformer-based vision models using low-rank adaptation and\ntask arithmetic. Our method completely bypasses the problem of catastrophic\nforgetting, as well as reducing the computational requirement for training\nmodels on each task. When aided with a small memory of 10 samples per class,\nour method achieves performance close to full-set finetuning. We present\nrigorous ablations to support the prowess of our method.\n", "date": "2023-11-04", "categories": []}
{"id": "2311.01617", "title": "Look-Ahead Selective Plasticity for Continual Learning of Visual Tasks", "abstract": "  Contrastive representation learning has emerged as a promising technique for\ncontinual learning as it can learn representations that are robust to\ncatastrophic forgetting and generalize well to unseen future tasks. Previous\nwork in continual learning has addressed forgetting by using previous task data\nand trained models. Inspired by event models created and updated in the brain,\nwe propose a new mechanism that takes place during task boundaries, i.e., when\none task finishes and another starts. By observing the redundancy-inducing\nability of contrastive loss on the output of a neural network, our method\nleverages the first few samples of the new task to identify and retain\nparameters contributing most to the transfer ability of the neural network,\nfreeing up the remaining parts of the network to learn new features. We\nevaluate the proposed methods on benchmark computer vision datasets including\nCIFAR10 and TinyImagenet and demonstrate state-of-the-art performance in the\ntask-incremental, class-incremental, and domain-incremental continual learning\nscenarios.\n", "date": "2023-11-02", "categories": []}
{"id": "2303.10725", "title": "SIESTA: Efficient Online Continual Learning with Sleep", "abstract": "  In supervised continual learning, a deep neural network (DNN) is updated with\nan ever-growing data stream. Unlike the offline setting where data is shuffled,\nwe cannot make any distributional assumptions about the data stream. Ideally,\nonly one pass through the dataset is needed for computational efficiency.\nHowever, existing methods are inadequate and make many assumptions that cannot\nbe made for real-world applications, while simultaneously failing to improve\ncomputational efficiency. In this paper, we propose a novel continual learning\nmethod, SIESTA based on wake/sleep framework for training, which is well\naligned to the needs of on-device learning. The major goal of SIESTA is to\nadvance compute efficient continual learning so that DNNs can be updated\nefficiently using far less time and energy. The principal innovations of SIESTA\nare: 1) rapid online updates using a rehearsal-free, backpropagation-free, and\ndata-driven network update rule during its wake phase, and 2) expedited memory\nconsolidation using a compute-restricted rehearsal policy during its sleep\nphase. For memory efficiency, SIESTA adapts latent rehearsal using memory\nindexing from REMIND. Compared to REMIND and prior arts, SIESTA is far more\ncomputationally efficient, enabling continual learning on ImageNet-1K in under\n2 hours on a single GPU; moreover, in the augmentation-free setting it matches\nthe performance of the offline learner, a milestone critical to driving\nadoption of continual learning in real-world applications.\n", "date": "2023-03-19", "categories": []}
{"id": "2305.09253", "title": "Online Continual Learning Without the Storage Constraint", "abstract": "  Traditional online continual learning (OCL) research has primarily focused on\nmitigating catastrophic forgetting with fixed and limited storage allocation\nthroughout an agent's lifetime. However, a broad range of real-world\napplications are primarily constrained by computational costs rather than\nstorage limitations. In this paper, we target such applications, investigating\nthe online continual learning problem under relaxed storage constraints and\nlimited computational budgets. We contribute a simple algorithm, which updates\na kNN classifier continually along with a fixed, pretrained feature extractor.\nWe selected this algorithm due to its exceptional suitability for online\ncontinual learning. It can adapt to rapidly changing streams, has zero\nstability gap, operates within tiny computational budgets, has low storage\nrequirements by only storing features, and has a consistency property: It never\nforgets previously seen data. These attributes yield significant improvements,\nallowing our proposed algorithm to outperform existing methods by over 20% in\naccuracy on two large-scale OCL datasets: Continual LOCalization (CLOC) with\n39M images and 712 classes and Continual Google Landmarks V2 (CGLM) with 580K\nimages and 10,788 classes, even when existing methods retain all previously\nseen images. Furthermore, we achieve this superior performance with\nconsiderably reduced computational and storage expenses. We provide code to\nreproduce our results at github.com/drimpossible/ACM.\n", "date": "2023-05-16", "categories": []}
{"id": "2307.10943", "title": "Proxy Anchor-based Unsupervised Learning for Continuous Generalized\n  Category Discovery", "abstract": "  Recent advances in deep learning have significantly improved the performance\nof various computer vision applications. However, discovering novel categories\nin an incremental learning scenario remains a challenging problem due to the\nlack of prior knowledge about the number and nature of new categories. Existing\nmethods for novel category discovery are limited by their reliance on labeled\ndatasets and prior knowledge about the number of novel categories and the\nproportion of novel samples in the batch. To address the limitations and more\naccurately reflect real-world scenarios, in this paper, we propose a novel\nunsupervised class incremental learning approach for discovering novel\ncategories on unlabeled sets without prior knowledge. The proposed method\nfine-tunes the feature extractor and proxy anchors on labeled sets, then splits\nsamples into old and novel categories and clusters on the unlabeled dataset.\nFurthermore, the proxy anchors-based exemplar generates representative category\nvectors to mitigate catastrophic forgetting. Experimental results demonstrate\nthat our proposed approach outperforms the state-of-the-art methods on\nfine-grained datasets under real-world scenarios.\n", "date": "2023-07-20", "categories": []}
{"id": "2311.01227", "title": "Robust Feature Learning and Global Variance-Driven Classifier Alignment\n  for Long-Tail Class Incremental Learning", "abstract": "  This paper introduces a two-stage framework designed to enhance long-tail\nclass incremental learning, enabling the model to progressively learn new\nclasses, while mitigating catastrophic forgetting in the context of long-tailed\ndata distributions. Addressing the challenge posed by the under-representation\nof tail classes in long-tail class incremental learning, our approach achieves\nclassifier alignment by leveraging global variance as an informative measure\nand class prototypes in the second stage. This process effectively captures\nclass properties and eliminates the need for data balancing or additional layer\ntuning. Alongside traditional class incremental learning losses in the first\nstage, the proposed approach incorporates mixup classes to learn robust feature\nrepresentations, ensuring smoother boundaries. The proposed framework can\nseamlessly integrate as a module with any class incremental learning method to\neffectively handle long-tail class incremental learning scenarios. Extensive\nexperimentation on the CIFAR-100 and ImageNet-Subset datasets validates the\napproach's efficacy, showcasing its superiority over state-of-the-art\ntechniques across various long-tail CIL settings.\n", "date": "2023-11-02", "categories": []}
{"id": "2310.20052", "title": "Look At Me, No Replay! SurpriseNet: Anomaly Detection Inspired Class\n  Incremental Learning", "abstract": "  Continual learning aims to create artificial neural networks capable of\naccumulating knowledge and skills through incremental training on a sequence of\ntasks. The main challenge of continual learning is catastrophic interference,\nwherein new knowledge overrides or interferes with past knowledge, leading to\nforgetting. An associated issue is the problem of learning \"cross-task\nknowledge,\" where models fail to acquire and retain knowledge that helps\ndifferentiate classes across task boundaries. A common solution to both\nproblems is \"replay,\" where a limited buffer of past instances is utilized to\nlearn cross-task knowledge and mitigate catastrophic interference. However, a\nnotable drawback of these methods is their tendency to overfit the limited\nreplay buffer. In contrast, our proposed solution, SurpriseNet, addresses\ncatastrophic interference by employing a parameter isolation method and\nlearning cross-task knowledge using an auto-encoder inspired by anomaly\ndetection. SurpriseNet is applicable to both structured and unstructured data,\nas it does not rely on image-specific inductive biases. We have conducted\nempirical experiments demonstrating the strengths of SurpriseNet on various\ntraditional vision continual-learning benchmarks, as well as on structured data\ndatasets. Source code made available at https://doi.org/10.5281/zenodo.8247906\nand https://github.com/tachyonicClock/SurpriseNet-CIKM-23\n", "date": "2023-10-30", "categories": []}
{"id": "2310.20268", "title": "Constructing Sample-to-Class Graph for Few-Shot Class-Incremental\n  Learning", "abstract": "  Few-shot class-incremental learning (FSCIL) aims to build machine learning\nmodel that can continually learn new concepts from a few data samples, without\nforgetting knowledge of old classes.\n  The challenges of FSCIL lies in the limited data of new classes, which not\nonly lead to significant overfitting issues but also exacerbates the notorious\ncatastrophic forgetting problems. As proved in early studies, building sample\nrelationships is beneficial for learning from few-shot samples. In this paper,\nwe promote the idea to the incremental scenario, and propose a Sample-to-Class\n(S2C) graph learning method for FSCIL.\n  Specifically, we propose a Sample-level Graph Network (SGN) that focuses on\nanalyzing sample relationships within a single session. This network helps\naggregate similar samples, ultimately leading to the extraction of more refined\nclass-level features.\n  Then, we present a Class-level Graph Network (CGN) that establishes\nconnections across class-level features of both new and old classes. This\nnetwork plays a crucial role in linking the knowledge between different\nsessions and helps improve overall learning in the FSCIL scenario. Moreover, we\ndesign a multi-stage strategy for training S2C model, which mitigates the\ntraining challenges posed by limited data in the incremental process.\n  The multi-stage training strategy is designed to build S2C graph from base to\nfew-shot stages, and improve the capacity via an extra pseudo-incremental\nstage. Experiments on three popular benchmark datasets show that our method\nclearly outperforms the baselines and sets new state-of-the-art results in\nFSCIL.\n", "date": "2023-10-31", "categories": []}
{"id": "2310.20348", "title": "Class Incremental Learning with Pre-trained Vision-Language Models", "abstract": "  With the advent of large-scale pre-trained models, interest in adapting and\nexploiting them for continual learning scenarios has grown.\n  In this paper, we propose an approach to exploiting pre-trained\nvision-language models (e.g. CLIP) that enables further adaptation instead of\nonly using zero-shot learning of new tasks. We augment a pre-trained CLIP model\nwith additional layers after the Image Encoder or before the Text Encoder. We\ninvestigate three different strategies: a Linear Adapter, a Self-attention\nAdapter, each operating on the image embedding, and Prompt Tuning which instead\nmodifies prompts input to the CLIP text encoder. We also propose a method for\nparameter retention in the adapter layers that uses a measure of parameter\nimportance to better maintain stability and plasticity during incremental\nlearning. Our experiments demonstrate that the simplest solution -- a single\nLinear Adapter layer with parameter retention -- produces the best results.\nExperiments on several conventional benchmarks consistently show a significant\nmargin of improvement over the current state-of-the-art.\n", "date": "2023-10-31", "categories": []}
{"id": "2309.08353", "title": "Continual Learning with Deep Streaming Regularized Discriminant Analysis", "abstract": "  Continual learning is increasingly sought after in real world machine\nlearning applications, as it enables learning in a more human-like manner.\nConventional machine learning approaches fail to achieve this, as incrementally\nupdating the model with non-identically distributed data leads to catastrophic\nforgetting, where existing representations are overwritten. Although\ntraditional continual learning methods have mostly focused on batch learning,\nwhich involves learning from large collections of labeled data sequentially,\nthis approach is not well-suited for real-world applications where we would\nlike new data to be integrated directly. This necessitates a paradigm shift\ntowards streaming learning. In this paper, we propose a streaming version of\nregularized discriminant analysis as a solution to this challenge. We combine\nour algorithm with a convolutional neural network and demonstrate that it\noutperforms both batch learning and existing streaming learning algorithms on\nthe ImageNet ILSVRC-2012 dataset.\n", "date": "2023-09-15", "categories": []}
{"id": "2310.18885", "title": "A foundational neural operator that continuously learns without\n  forgetting", "abstract": "  Machine learning has witnessed substantial growth, leading to the development\nof advanced artificial intelligence models crafted to address a wide range of\nreal-world challenges spanning various domains, such as computer vision,\nnatural language processing, and scientific computing. Nevertheless, the\ncreation of custom models for each new task remains a resource-intensive\nundertaking, demanding considerable computational time and memory resources. In\nthis study, we introduce the concept of the Neural Combinatorial Wavelet Neural\nOperator (NCWNO) as a foundational model for scientific computing. This model\nis specifically designed to excel in learning from a diverse spectrum of\nphysics and continuously adapt to the solution operators associated with\nparametric partial differential equations (PDEs). The NCWNO leverages a gated\nstructure that employs local wavelet experts to acquire shared features across\nmultiple physical systems, complemented by a memory-based ensembling approach\namong these local wavelet experts. This combination enables rapid adaptation to\nnew challenges. The proposed foundational model offers two key advantages: (i)\nit can simultaneously learn solution operators for multiple parametric PDEs,\nand (ii) it can swiftly generalize to new parametric PDEs with minimal\nfine-tuning. The proposed NCWNO is the first foundational operator learning\nalgorithm distinguished by its (i) robustness against catastrophic forgetting,\n(ii) the maintenance of positive transfer for new parametric PDEs, and (iii)\nthe facilitation of knowledge transfer across dissimilar tasks. Through an\nextensive set of benchmark examples, we demonstrate that the NCWNO can\noutperform task-specific baseline operator learning frameworks with minimal\nhyperparameter tuning at the prediction stage. We also show that with minimal\nfine-tuning, the NCWNO performs accurate combinatorial learning of new\nparametric PDEs.\n", "date": "2023-10-29", "categories": []}
{"id": "2310.19272", "title": "NPCL: Neural Processes for Uncertainty-Aware Continual Learning", "abstract": "  Continual learning (CL) aims to train deep neural networks efficiently on\nstreaming data while limiting the forgetting caused by new tasks. However,\nlearning transferable knowledge with less interference between tasks is\ndifficult, and real-world deployment of CL models is limited by their inability\nto measure predictive uncertainties. To address these issues, we propose\nhandling CL tasks with neural processes (NPs), a class of meta-learners that\nencode different tasks into probabilistic distributions over functions all\nwhile providing reliable uncertainty estimates. Specifically, we propose an\nNP-based CL approach (NPCL) with task-specific modules arranged in a\nhierarchical latent variable model. We tailor regularizers on the learned\nlatent distributions to alleviate forgetting. The uncertainty estimation\ncapabilities of the NPCL can also be used to handle the task head/module\ninference challenge in CL. Our experiments show that the NPCL outperforms\nprevious CL approaches. We validate the effectiveness of uncertainty estimation\nin the NPCL for identifying novel data and evaluating instance-level model\nconfidence. Code is available at \\url{https://github.com/srvCodes/NPCL}.\n", "date": "2023-10-30", "categories": []}
{"id": "2306.14884", "title": "Learning to Modulate pre-trained Models in RL", "abstract": "  Reinforcement Learning (RL) has been successful in various domains like\nrobotics, game playing, and simulation. While RL agents have shown impressive\ncapabilities in their specific tasks, they insufficiently adapt to new tasks.\nIn supervised learning, this adaptation problem is addressed by large-scale\npre-training followed by fine-tuning to new down-stream tasks. Recently,\npre-training on multiple tasks has been gaining traction in RL. However,\nfine-tuning a pre-trained model often suffers from catastrophic forgetting.\nThat is, the performance on the pre-training tasks deteriorates when\nfine-tuning on new tasks. To investigate the catastrophic forgetting\nphenomenon, we first jointly pre-train a model on datasets from two benchmark\nsuites, namely Meta-World and DMControl. Then, we evaluate and compare a\nvariety of fine-tuning methods prevalent in natural language processing, both\nin terms of performance on new tasks, and how well performance on pre-training\ntasks is retained. Our study shows that with most fine-tuning approaches, the\nperformance on pre-training tasks deteriorates significantly. Therefore, we\npropose a novel method, Learning-to-Modulate (L2M), that avoids the degradation\nof learned skills by modulating the information flow of the frozen pre-trained\nmodel via a learnable modulation pool. Our method achieves state-of-the-art\nperformance on the Continual-World benchmark, while retaining performance on\nthe pre-training tasks. Finally, to aid future research in this area, we\nrelease a dataset encompassing 50 Meta-World and 16 DMControl tasks.\n", "date": "2023-06-26", "categories": []}
{"id": "2310.16931", "title": "CL-MASR: A Continual Learning Benchmark for Multilingual ASR", "abstract": "  Modern multilingual automatic speech recognition (ASR) systems like Whisper\nhave made it possible to transcribe audio in multiple languages with a single\nmodel. However, current state-of-the-art ASR models are typically evaluated on\nindividual languages or in a multi-task setting, overlooking the challenge of\ncontinually learning new languages. There is insufficient research on how to\nadd new languages without losing valuable information from previous data.\nFurthermore, existing continual learning benchmarks focus mostly on vision and\nlanguage tasks, leaving continual learning for multilingual ASR largely\nunexplored. To bridge this gap, we propose CL-MASR, a benchmark designed for\nstudying multilingual ASR in a continual learning setting. CL-MASR provides a\ndiverse set of continual learning methods implemented on top of large-scale\npretrained ASR models, along with common metrics to assess the effectiveness of\nlearning new languages while addressing the issue of catastrophic forgetting.\nTo the best of our knowledge, CL-MASR is the first continual learning benchmark\nfor the multilingual ASR task. The code is available at\nhttps://github.com/speechbrain/benchmarks.\n", "date": "2023-10-25", "categories": []}
{"id": "2310.16996", "title": "Towards Continually Learning Application Performance Models", "abstract": "  Machine learning-based performance models are increasingly being used to\nbuild critical job scheduling and application optimization decisions.\nTraditionally, these models assume that data distribution does not change as\nmore samples are collected over time. However, owing to the complexity and\nheterogeneity of production HPC systems, they are susceptible to hardware\ndegradation, replacement, and/or software patches, which can lead to drift in\nthe data distribution that can adversely affect the performance models. To this\nend, we develop continually learning performance models that account for the\ndistribution drift, alleviate catastrophic forgetting, and improve\ngeneralizability. Our best model was able to retain accuracy, regardless of\nhaving to learn the new distribution of data inflicted by system changes, while\ndemonstrating a 2x improvement in the prediction accuracy of the whole data\nsequence in comparison to the naive approach.\n", "date": "2023-10-25", "categories": []}
{"id": "2310.16115", "title": "Wakening Past Concepts without Past Data: Class-Incremental Learning\n  from Online Placebos", "abstract": "  Not forgetting old class knowledge is a key challenge for class-incremental\nlearning (CIL) when the model continuously adapts to new classes. A common\ntechnique to address this is knowledge distillation (KD), which penalizes\nprediction inconsistencies between old and new models. Such prediction is made\nwith almost new class data, as old class data is extremely scarce due to the\nstrict memory limitation in CIL. In this paper, we take a deep dive into KD\nlosses and find that \"using new class data for KD\" not only hinders the model\nadaption (for learning new classes) but also results in low efficiency for\npreserving old class knowledge. We address this by \"using the placebos of old\nclasses for KD\", where the placebos are chosen from a free image stream, such\nas Google Images, in an automatical and economical fashion. To this end, we\ntrain an online placebo selection policy to quickly evaluate the quality of\nstreaming images (good or bad placebos) and use only good ones for one-time\nfeed-forward computation of KD. We formulate the policy training process as an\nonline Markov Decision Process (MDP), and introduce an online learning\nalgorithm to solve this MDP problem without causing much computation costs. In\nexperiments, we show that our method 1) is surprisingly effective even when\nthere is no class overlap between placebos and original old class data, 2) does\nnot require any additional supervision or memory budget, and 3) significantly\noutperforms a number of top-performing CIL methods, in particular when using\nlower memory budgets for old class exemplars, e.g., five exemplars per class.\n", "date": "2023-10-24", "categories": []}
{"id": "2304.08130", "title": "A Survey on Few-Shot Class-Incremental Learning", "abstract": "  Large deep learning models are impressive, but they struggle when real-time\ndata is not available. Few-shot class-incremental learning (FSCIL) poses a\nsignificant challenge for deep neural networks to learn new tasks from just a\nfew labeled samples without forgetting the previously learned ones. This setup\neasily leads to catastrophic forgetting and overfitting problems, severely\naffecting model performance. Studying FSCIL helps overcome deep learning model\nlimitations on data volume and acquisition time, while improving practicality\nand adaptability of machine learning models. This paper provides a\ncomprehensive survey on FSCIL. Unlike previous surveys, we aim to synthesize\nfew-shot learning and incremental learning, focusing on introducing FSCIL from\ntwo perspectives, while reviewing over 30 theoretical research studies and more\nthan 20 applied research studies. From the theoretical perspective, we provide\na novel categorization approach that divides the field into five subcategories,\nincluding traditional machine learning methods, meta-learning based methods,\nfeature and feature space-based methods, replay-based methods, and dynamic\nnetwork structure-based methods. We also evaluate the performance of recent\ntheoretical research on benchmark datasets of FSCIL. From the application\nperspective, FSCIL has achieved impressive achievements in various fields of\ncomputer vision such as image classification, object detection, and image\nsegmentation, as well as in natural language processing and graph. We summarize\nthe important applications. Finally, we point out potential future research\ndirections, including applications, problem setups, and theory development.\nOverall, this paper offers a comprehensive analysis of the latest advances in\nFSCIL from a methodological, performance, and application perspective.\n", "date": "2023-04-17", "categories": []}
{"id": "2310.15470", "title": "Continual Event Extraction with Semantic Confusion Rectification", "abstract": "  We study continual event extraction, which aims to extract incessantly\nemerging event information while avoiding forgetting. We observe that the\nsemantic confusion on event types stems from the annotations of the same text\nbeing updated over time. The imbalance between event types even aggravates this\nissue. This paper proposes a novel continual event extraction model with\nsemantic confusion rectification. We mark pseudo labels for each sentence to\nalleviate semantic confusion. We transfer pivotal knowledge between current and\nprevious models to enhance the understanding of event types. Moreover, we\nencourage the model to focus on the semantics of long-tailed event types by\nleveraging other associated types. Experimental results show that our model\noutperforms state-of-the-art baselines and is proficient in imbalanced\ndatasets.\n", "date": "2023-10-24", "categories": []}
{"id": "2310.15974", "title": "Minimax Forward and Backward Learning of Evolving Tasks with Performance\n  Guarantees", "abstract": "  For a sequence of classification tasks that arrive over time, it is common\nthat tasks are evolving in the sense that consecutive tasks often have a higher\nsimilarity. The incremental learning of a growing sequence of tasks holds\npromise to enable accurate classification even with few samples per task by\nleveraging information from all the tasks in the sequence (forward and backward\nlearning). However, existing techniques developed for continual learning and\nconcept drift adaptation are either designed for tasks with time-independent\nsimilarities or only aim to learn the last task in the sequence. This paper\npresents incremental minimax risk classifiers (IMRCs) that effectively exploit\nforward and backward learning and account for evolving tasks. In addition, we\nanalytically characterize the performance improvement provided by forward and\nbackward learning in terms of the tasks' expected quadratic change and the\nnumber of tasks. The experimental evaluation shows that IMRCs can result in a\nsignificant performance improvement, especially for reduced sample sizes.\n", "date": "2023-10-24", "categories": []}
{"id": "2310.13977", "title": "Continual Invariant Risk Minimization", "abstract": "  Empirical risk minimization can lead to poor generalization behavior on\nunseen environments if the learned model does not capture invariant feature\nrepresentations. Invariant risk minimization (IRM) is a recent proposal for\ndiscovering environment-invariant representations. IRM was introduced by\nArjovsky et al. (2019) and extended by Ahuja et al. (2020). IRM assumes that\nall environments are available to the learning system at the same time. With\nthis work, we generalize the concept of IRM to scenarios where environments are\nobserved sequentially. We show that existing approaches, including those\ndesigned for continual learning, fail to identify the invariant features and\nmodels across sequentially presented environments. We extend IRM under a\nvariational Bayesian and bilevel framework, creating a general approach to\ncontinual invariant risk minimization. We also describe a strategy to solve the\noptimization problems using a variant of the alternating direction method of\nmultiplier (ADMM). We show empirically using multiple datasets and with\nmultiple sequential environments that the proposed methods outperform or is\ncompetitive with prior approaches.\n", "date": "2023-10-21", "categories": []}
{"id": "2310.14152", "title": "Orthogonal Subspace Learning for Language Model Continual Learning", "abstract": "  Benefiting from massive corpora and advanced hardware, large language models\n(LLMs) exhibit remarkable capabilities in language understanding and\ngeneration. However, their performance degrades in scenarios where multiple\ntasks are encountered sequentially, also known as catastrophic forgetting. In\nthis paper, we propose orthogonal low-rank adaptation (O-LoRA), a simple and\nefficient approach for continual learning in language models, effectively\nmitigating catastrophic forgetting while learning new tasks. Specifically,\nO-LoRA learns tasks in different (low-rank) vector subspaces that are kept\northogonal to each other in order to minimize interference. Our method induces\nonly marginal additional parameter costs and requires no user data storage for\nreplay. Experimental results on continual learning benchmarks show that our\nmethod outperforms state-of-the-art methods. Furthermore, compared to previous\napproaches, our method excels in preserving the generalization ability of LLMs\non unseen tasks.\n", "date": "2023-10-22", "categories": []}
{"id": "2310.14248", "title": "From Static to Dynamic: A Continual Learning Framework for Large\n  Language Models", "abstract": "  The vast number of parameters in large language models (LLMs) endows them\nwith remarkable capabilities, allowing them to excel in a variety of natural\nlanguage processing tasks. However, this complexity also presents challenges,\nmaking LLMs difficult to train and inhibiting their ability to continuously\nassimilate new knowledge, which may lead to inaccuracies in their outputs. To\nmitigate these issues, this paper presents DynaMind, a novel continual learning\nframework designed for LLMs. DynaMind incorporates memory mechanisms to\nassimilate new knowledge and modular operators to enhance the model inference\nprocess with the newly assimilated knowledge, consequently improving the\naccuracies of LLMs' outputs. Benchmark experiments demonstrate DynaMind's\neffectiveness in overcoming these challenges. The code and demo of DynaMind are\navailable on GitHub: https://github.com/Elfsong/DynaMind.\n", "date": "2023-10-22", "categories": []}
{"id": "2310.14510", "title": "CITB: A Benchmark for Continual Instruction Tuning", "abstract": "  Continual learning (CL) is a paradigm that aims to replicate the human\nability to learn and accumulate knowledge continually without forgetting\nprevious knowledge and transferring it to new tasks. Recent instruction tuning\n(IT) involves fine-tuning models to make them more adaptable to solving NLP\ntasks in general. However, it is still uncertain how instruction tuning works\nin the context of CL tasks. This challenging yet practical problem is\nformulated as Continual Instruction Tuning (CIT). In this work, we establish a\nCIT benchmark consisting of learning and evaluation protocols. We curate two\nlong dialogue task streams of different types, InstrDialog and InstrDialog++,\nto study various CL methods systematically. Our experiments show that existing\nCL methods do not effectively leverage the rich natural language instructions,\nand fine-tuning an instruction-tuned model sequentially can yield similar or\nbetter results. We further explore different aspects that might affect the\nlearning of CIT. We hope this benchmark will facilitate more research in this\ndirection.\n", "date": "2023-10-23", "categories": []}
{"id": "2310.14541", "title": "Continual Named Entity Recognition without Catastrophic Forgetting", "abstract": "  Continual Named Entity Recognition (CNER) is a burgeoning area, which\ninvolves updating an existing model by incorporating new entity types\nsequentially. Nevertheless, continual learning approaches are often severely\nafflicted by catastrophic forgetting. This issue is intensified in CNER due to\nthe consolidation of old entity types from previous steps into the non-entity\ntype at each step, leading to what is known as the semantic shift problem of\nthe non-entity type. In this paper, we introduce a pooled feature distillation\nloss that skillfully navigates the trade-off between retaining knowledge of old\nentity types and acquiring new ones, thereby more effectively mitigating the\nproblem of catastrophic forgetting. Additionally, we develop a confidence-based\npseudo-labeling for the non-entity type, \\emph{i.e.,} predicting entity types\nusing the old model to handle the semantic shift of the non-entity type.\nFollowing the pseudo-labeling process, we suggest an adaptive re-weighting\ntype-balanced learning strategy to handle the issue of biased type\ndistribution. We carried out comprehensive experiments on ten CNER settings\nusing three different datasets. The results illustrate that our method\nsignificantly outperforms prior state-of-the-art approaches, registering an\naverage improvement of $6.3$\\% and $8.0$\\% in Micro and Macro F1 scores,\nrespectively.\n", "date": "2023-10-23", "categories": []}
{"id": "2310.15054", "title": "Coordinated Replay Sample Selection for Continual Federated Learning", "abstract": "  Continual Federated Learning (CFL) combines Federated Learning (FL), the\ndecentralized learning of a central model on a number of client devices that\nmay not communicate their data, and Continual Learning (CL), the learning of a\nmodel from a continual stream of data without keeping the entire history. In\nCL, the main challenge is \\textit{forgetting} what was learned from past data.\nWhile replay-based algorithms that keep a small pool of past training data are\neffective to reduce forgetting, only simple replay sample selection strategies\nhave been applied to CFL in prior work, and no previous work has explored\ncoordination among clients for better sample selection. To bridge this gap, we\nadapt a replay sample selection objective based on loss gradient diversity to\nCFL and propose a new relaxation-based selection of samples to optimize the\nobjective. Next, we propose a practical algorithm to coordinate gradient-based\nreplay sample selection across clients without communicating private data. We\nbenchmark our coordinated and uncoordinated replay sample selection algorithms\nagainst random sampling-based baselines with language models trained on a large\nscale de-identified real-world text dataset. We show that gradient-based sample\nselection methods both boost performance and reduce forgetting compared to\nrandom sampling methods, with our coordination method showing gains early in\nthe low replay size regime (when the budget for storing past data is small).\n", "date": "2023-10-23", "categories": []}
{"id": "2310.13024", "title": "Towards Anytime Fine-tuning: Continually Pre-trained Language Models\n  with Hypernetwork Prompt", "abstract": "  Continual pre-training has been urgent for adapting a pre-trained model to a\nmultitude of domains and tasks in the fast-evolving world. In practice, a\ncontinually pre-trained model is expected to demonstrate not only greater\ncapacity when fine-tuned on pre-trained domains but also a non-decreasing\nperformance on unseen ones. In this work, we first investigate such anytime\nfine-tuning effectiveness of existing continual pre-training approaches,\nconcluding with unanimously decreased performance on unseen domains. To this\nend, we propose a prompt-guided continual pre-training method, where we train a\nhypernetwork to generate domain-specific prompts by both agreement and\ndisagreement losses. The agreement loss maximally preserves the generalization\nof a pre-trained model to new domains, and the disagreement one guards the\nexclusiveness of the generated hidden states for each domain. Remarkably,\nprompts by the hypernetwork alleviate the domain identity when fine-tuning and\npromote knowledge transfer across domains. Our method achieved improvements of\n3.57% and 3.4% on two real-world datasets (including domain shift and temporal\nshift), respectively, demonstrating its efficacy.\n", "date": "2023-10-19", "categories": []}
{"id": "2302.08243", "title": "New Insights on Relieving Task-Recency Bias for Online Class Incremental\n  Learning", "abstract": "  To imitate the ability of keeping learning of human, continual learning which\ncan learn from a never-ending data stream has attracted more interests\nrecently. In all settings, the online class incremental learning (OCIL), where\nincoming samples from data stream can be used only once, is more challenging\nand can be encountered more frequently in real world. Actually, all continual\nlearning models face a stability-plasticity dilemma, where the stability means\nthe ability to preserve old knowledge while the plasticity denotes the ability\nto incorporate new knowledge. Although replay-based methods have shown\nexceptional promise, most of them concentrate on the strategy for updating and\nretrieving memory to keep stability at the expense of plasticity. To strike a\npreferable trade-off between stability and plasticity, we propose an Adaptive\nFocus Shifting algorithm (AFS), which dynamically adjusts focus to ambiguous\nsamples and non-target logits in model learning. Through a deep analysis of the\ntask-recency bias caused by class imbalance, we propose a revised focal loss to\nmainly keep stability. \\Rt{By utilizing a new weight function, the revised\nfocal loss will pay more attention to current ambiguous samples, which are the\npotentially valuable samples to make model progress quickly.} To promote\nplasticity, we introduce a virtual knowledge distillation. By designing a\nvirtual teacher, it assigns more attention to non-target classes, which can\nsurmount overconfidence and encourage model to focus on inter-class\ninformation. Extensive experiments on three popular datasets for OCIL have\nshown the effectiveness of AFS. The code will be available at\n\\url{https://github.com/czjghost/AFS}.\n", "date": "2023-02-16", "categories": []}
{"id": "2310.12244", "title": "A Unified Approach to Domain Incremental Learning with Memory: Theory\n  and Algorithm", "abstract": "  Domain incremental learning aims to adapt to a sequence of domains with\naccess to only a small subset of data (i.e., memory) from previous domains.\nVarious methods have been proposed for this problem, but it is still unclear\nhow they are related and when practitioners should choose one method over\nanother. In response, we propose a unified framework, dubbed Unified Domain\nIncremental Learning (UDIL), for domain incremental learning with memory. Our\nUDIL **unifies** various existing methods, and our theoretical analysis shows\nthat UDIL always achieves a tighter generalization error bound compared to\nthese methods. The key insight is that different existing methods correspond to\nour bound with different **fixed** coefficients; based on insights from this\nunification, our UDIL allows **adaptive** coefficients during training, thereby\nalways achieving the tightest bound. Empirical results show that our UDIL\noutperforms the state-of-the-art domain incremental learning methods on both\nsynthetic and real-world datasets. Code will be available at\nhttps://github.com/Wang-ML-Lab/unified-continual-learning.\n", "date": "2023-10-18", "categories": []}
{"id": "2308.11063", "title": "MetaGCD: Learning to Continually Learn in Generalized Category Discovery", "abstract": "  In this paper, we consider a real-world scenario where a model that is\ntrained on pre-defined classes continually encounters unlabeled data that\ncontains both known and novel classes. The goal is to continually discover\nnovel classes while maintaining the performance in known classes. We name the\nsetting Continual Generalized Category Discovery (C-GCD). Existing methods for\nnovel class discovery cannot directly handle the C-GCD setting due to some\nunrealistic assumptions, such as the unlabeled data only containing novel\nclasses. Furthermore, they fail to discover novel classes in a continual\nfashion. In this work, we lift all these assumptions and propose an approach,\ncalled MetaGCD, to learn how to incrementally discover with less forgetting.\nOur proposed method uses a meta-learning framework and leverages the offline\nlabeled data to simulate the testing incremental learning process. A\nmeta-objective is defined to revolve around two conflicting learning objectives\nto achieve novel class discovery without forgetting. Furthermore, a soft\nneighborhood-based contrastive network is proposed to discriminate uncorrelated\nimages while attracting correlated images. We build strong baselines and\nconduct extensive experiments on three widely used benchmarks to demonstrate\nthe superiority of our method.\n", "date": "2023-08-21", "categories": []}
{"id": "2309.01289", "title": "Federated Orthogonal Training: Mitigating Global Catastrophic Forgetting\n  in Continual Federated Learning", "abstract": "  Federated Learning (FL) has gained significant attraction due to its ability\nto enable privacy-preserving training over decentralized data. Current\nliterature in FL mostly focuses on single-task learning. However, over time,\nnew tasks may appear in the clients and the global model should learn these\ntasks without forgetting previous tasks. This real-world scenario is known as\nContinual Federated Learning (CFL). The main challenge of CFL is Global\nCatastrophic Forgetting, which corresponds to the fact that when the global\nmodel is trained on new tasks, its performance on old tasks decreases. There\nhave been a few recent works on CFL to propose methods that aim to address the\nglobal catastrophic forgetting problem. However, these works either have\nunrealistic assumptions on the availability of past data samples or violate the\nprivacy principles of FL. We propose a novel method, Federated Orthogonal\nTraining (FOT), to overcome these drawbacks and address the global catastrophic\nforgetting in CFL. Our algorithm extracts the global input subspace of each\nlayer for old tasks and modifies the aggregated updates of new tasks such that\nthey are orthogonal to the global principal subspace of old tasks for each\nlayer. This decreases the interference between tasks, which is the main cause\nfor forgetting. We empirically show that FOT outperforms state-of-the-art\ncontinual learning methods in the CFL setting, achieving an average accuracy\ngain of up to 15% with 27% lower forgetting while only incurring a minimal\ncomputation and communication cost.\n", "date": "2023-09-03", "categories": []}
{"id": "2310.11341", "title": "Dual Cognitive Architecture: Incorporating Biases and Multi-Memory\n  Systems for Lifelong Learning", "abstract": "  Artificial neural networks (ANNs) exhibit a narrow scope of expertise on\nstationary independent data. However, the data in the real world is continuous\nand dynamic, and ANNs must adapt to novel scenarios while also retaining the\nlearned knowledge to become lifelong learners. The ability of humans to excel\nat these tasks can be attributed to multiple factors ranging from cognitive\ncomputational structures, cognitive biases, and the multi-memory systems in the\nbrain. We incorporate key concepts from each of these to design a novel\nframework, Dual Cognitive Architecture (DUCA), which includes multiple\nsub-systems, implicit and explicit knowledge representation dichotomy,\ninductive bias, and a multi-memory system. The inductive bias learner within\nDUCA is instrumental in encoding shape information, effectively countering the\ntendency of ANNs to learn local textures. Simultaneously, the inclusion of a\nsemantic memory submodule facilitates the gradual consolidation of knowledge,\nreplicating the dynamics observed in fast and slow learning systems,\nreminiscent of the principles underpinning the complementary learning system in\nhuman cognition. DUCA shows improvement across different settings and datasets,\nand it also exhibits reduced task recency bias, without the need for extra\ninformation. To further test the versatility of lifelong learning methods on a\nchallenging distribution shift, we introduce a novel domain-incremental dataset\nDN4IL. In addition to improving performance on existing benchmarks, DUCA also\ndemonstrates superior performance on this complex dataset.\n", "date": "2023-10-17", "categories": []}
{"id": "2306.03310", "title": "LIBERO: Benchmarking Knowledge Transfer for Lifelong Robot Learning", "abstract": "  Lifelong learning offers a promising paradigm of building a generalist agent\nthat learns and adapts over its lifespan. Unlike traditional lifelong learning\nproblems in image and text domains, which primarily involve the transfer of\ndeclarative knowledge of entities and concepts, lifelong learning in\ndecision-making (LLDM) also necessitates the transfer of procedural knowledge,\nsuch as actions and behaviors. To advance research in LLDM, we introduce\nLIBERO, a novel benchmark of lifelong learning for robot manipulation.\nSpecifically, LIBERO highlights five key research topics in LLDM: 1) how to\nefficiently transfer declarative knowledge, procedural knowledge, or the\nmixture of both; 2) how to design effective policy architectures and 3)\neffective algorithms for LLDM; 4) the robustness of a lifelong learner with\nrespect to task ordering; and 5) the effect of model pretraining for LLDM. We\ndevelop an extendible procedural generation pipeline that can in principle\ngenerate infinitely many tasks. For benchmarking purpose, we create four task\nsuites (130 tasks in total) that we use to investigate the above-mentioned\nresearch topics. To support sample-efficient learning, we provide high-quality\nhuman-teleoperated demonstration data for all tasks. Our extensive experiments\npresent several insightful or even unexpected discoveries: sequential\nfinetuning outperforms existing lifelong learning methods in forward transfer,\nno single visual encoder architecture excels at all types of knowledge\ntransfer, and naive supervised pretraining can hinder agents' performance in\nthe subsequent LLDM. Check the website at https://libero-project.github.io for\nthe code and the datasets.\n", "date": "2023-06-05", "categories": []}
{"id": "2308.02084", "title": "Efficient Model Adaptation for Continual Learning at the Edge", "abstract": "  Most machine learning (ML) systems assume stationary and matching data\ndistributions during training and deployment. This is often a false assumption.\nWhen ML models are deployed on real devices, data distributions often shift\nover time due to changes in environmental factors, sensor characteristics, and\ntask-of-interest. While it is possible to have a human-in-the-loop to monitor\nfor distribution shifts and engineer new architectures in response to these\nshifts, such a setup is not cost-effective. Instead, non-stationary automated\nML (AutoML) models are needed. This paper presents the\nEncoder-Adaptor-Reconfigurator (EAR) framework for efficient continual learning\nunder domain shifts. The EAR framework uses a fixed deep neural network (DNN)\nfeature encoder and trains shallow networks on top of the encoder to handle\nnovel data. The EAR framework is capable of 1) detecting when new data is\nout-of-distribution (OOD) by combining DNNs with hyperdimensional computing\n(HDC), 2) identifying low-parameter neural adaptors to adapt the model to the\nOOD data using zero-shot neural architecture search (ZS-NAS), and 3) minimizing\ncatastrophic forgetting on previous tasks by progressively growing the neural\narchitecture as needed and dynamically routing data through the appropriate\nadaptors and reconfigurators for handling domain-incremental and\nclass-incremental continual learning. We systematically evaluate our approach\non several benchmark datasets for domain adaptation and demonstrate strong\nperformance compared to state-of-the-art algorithms for OOD detection and\nfew-/zero-shot NAS.\n", "date": "2023-08-03", "categories": []}
{"id": "2308.11073", "title": "Audio-Visual Class-Incremental Learning", "abstract": "  In this paper, we introduce audio-visual class-incremental learning, a\nclass-incremental learning scenario for audio-visual video recognition. We\ndemonstrate that joint audio-visual modeling can improve class-incremental\nlearning, but current methods fail to preserve semantic similarity between\naudio and visual features as incremental step grows. Furthermore, we observe\nthat audio-visual correlations learned in previous tasks can be forgotten as\nincremental steps progress, leading to poor performance. To overcome these\nchallenges, we propose AV-CIL, which incorporates Dual-Audio-Visual Similarity\nConstraint (D-AVSC) to maintain both instance-aware and class-aware semantic\nsimilarity between audio-visual modalities and Visual Attention Distillation\n(VAD) to retain previously learned audio-guided visual attentive ability. We\ncreate three audio-visual class-incremental datasets, AVE-Class-Incremental\n(AVE-CI), Kinetics-Sounds-Class-Incremental (K-S-CI), and\nVGGSound100-Class-Incremental (VS100-CI) based on the AVE, Kinetics-Sounds, and\nVGGSound datasets, respectively. Our experiments on AVE-CI, K-S-CI, and\nVS100-CI demonstrate that AV-CIL significantly outperforms existing\nclass-incremental learning methods in audio-visual class-incremental learning.\nCode and data are available at: https://github.com/weiguoPian/AV-CIL_ICCV2023.\n", "date": "2023-08-21", "categories": []}
{"id": "2310.09436", "title": "Sub-network Discovery and Soft-masking for Continual Learning of Mixed\n  Tasks", "abstract": "  Continual learning (CL) has two main objectives: preventing catastrophic\nforgetting (CF) and encouraging knowledge transfer (KT). The existing\nliterature mainly focused on overcoming CF. Some work has also been done on KT\nwhen the tasks are similar. To our knowledge, only one method has been proposed\nto learn a sequence of mixed tasks. However, these techniques still suffer from\nCF and/or limited KT. This paper proposes a new CL method to achieve both. It\novercomes CF by isolating the knowledge of each task via discovering a\nsubnetwork for it. A soft-masking mechanism is also proposed to preserve the\nprevious knowledge and to enable the new task to leverage the past knowledge to\nachieve KT. Experiments using classification, generation, information\nextraction, and their mixture (i.e., heterogeneous tasks) show that the\nproposed method consistently outperforms strong baselines.\n", "date": "2023-10-13", "categories": []}
{"id": "2310.10184", "title": "Continual Generalized Intent Discovery: Marching Towards Dynamic and\n  Open-world Intent Recognition", "abstract": "  In a practical dialogue system, users may input out-of-domain (OOD) queries.\nThe Generalized Intent Discovery (GID) task aims to discover OOD intents from\nOOD queries and extend them to the in-domain (IND) classifier. However, GID\nonly considers one stage of OOD learning, and needs to utilize the data in all\nprevious stages for joint training, which limits its wide application in\nreality. In this paper, we introduce a new task, Continual Generalized Intent\nDiscovery (CGID), which aims to continuously and automatically discover OOD\nintents from dynamic OOD data streams and then incrementally add them to the\nclassifier with almost no previous data, thus moving towards dynamic intent\nrecognition in an open world. Next, we propose a method called Prototype-guided\nLearning with Replay and Distillation (PLRD) for CGID, which bootstraps new\nintent discovery through class prototypes and balances new and old intents\nthrough data replay and feature distillation. Finally, we conduct detailed\nexperiments and analysis to verify the effectiveness of PLRD and understand the\nkey challenges of CGID for future research.\n", "date": "2023-10-16", "categories": []}
{"id": "2310.10417", "title": "Prior-Free Continual Learning with Unlabeled Data in the Wild", "abstract": "  Continual Learning (CL) aims to incrementally update a trained model on new\ntasks without forgetting the acquired knowledge of old ones. Existing CL\nmethods usually reduce forgetting with task priors, \\ie using task identity or\na subset of previously seen samples for model training. However, these methods\nwould be infeasible when such priors are unknown in real-world applications. To\naddress this fundamental but seldom-studied problem, we propose a Prior-Free\nContinual Learning (PFCL) method, which learns new tasks without knowing the\ntask identity or any previous data. First, based on a fixed single-head\narchitecture, we eliminate the need for task identity to select the\ntask-specific output head. Second, we employ a regularization-based strategy\nfor consistent predictions between the new and old models, avoiding revisiting\nprevious samples. However, using this strategy alone often performs poorly in\nclass-incremental scenarios, particularly for a long sequence of tasks. By\nanalyzing the effectiveness and limitations of conventional\nregularization-based methods, we propose enhancing model consistency with an\nauxiliary unlabeled dataset additionally. Moreover, since some auxiliary data\nmay degrade the performance, we further develop a reliable sample selection\nstrategy to obtain consistent performance improvement. Extensive experiments on\nmultiple image classification benchmark datasets show that our PFCL method\nsignificantly mitigates forgetting in all three learning scenarios.\nFurthermore, when compared to the most recent rehearsal-based methods that\nreplay a limited number of previous samples, PFCL achieves competitive\naccuracy. Our code is available at: https://github.com/visiontao/pfcl\n", "date": "2023-10-16", "categories": []}
{"id": "2303.15833", "title": "Complementary Domain Adaptation and Generalization for Unsupervised\n  Continual Domain Shift Learning", "abstract": "  Continual domain shift poses a significant challenge in real-world\napplications, particularly in situations where labeled data is not available\nfor new domains. The challenge of acquiring knowledge in this problem setting\nis referred to as unsupervised continual domain shift learning. Existing\nmethods for domain adaptation and generalization have limitations in addressing\nthis issue, as they focus either on adapting to a specific domain or\ngeneralizing to unseen domains, but not both. In this paper, we propose\nComplementary Domain Adaptation and Generalization (CoDAG), a simple yet\neffective learning framework that combines domain adaptation and generalization\nin a complementary manner to achieve three major goals of unsupervised\ncontinual domain shift learning: adapting to a current domain, generalizing to\nunseen domains, and preventing forgetting of previously seen domains. Our\napproach is model-agnostic, meaning that it is compatible with any existing\ndomain adaptation and generalization algorithms. We evaluate CoDAG on several\nbenchmark datasets and demonstrate that our model outperforms state-of-the-art\nmodels in all datasets and evaluation metrics, highlighting its effectiveness\nand robustness in handling unsupervised continual domain shift learning.\n", "date": "2023-03-28", "categories": []}
{"id": "2306.09675", "title": "Multi-View Class Incremental Learning", "abstract": "  Multi-view learning (MVL) has gained great success in integrating information\nfrom multiple perspectives of a dataset to improve downstream task performance.\nTo make MVL methods more practical in an open-ended environment, this paper\ninvestigates a novel paradigm called multi-view class incremental learning\n(MVCIL), where a single model incrementally classifies new classes from a\ncontinual stream of views, requiring no access to earlier views of data.\nHowever, MVCIL is challenged by the catastrophic forgetting of old information\nand the interference with learning new concepts. To address this, we first\ndevelop a randomization-based representation learning technique serving for\nfeature extraction to guarantee their separate view-optimal working states,\nduring which multiple views belonging to a class are presented sequentially;\nThen, we integrate them one by one in the orthogonality fusion subspace spanned\nby the extracted features; Finally, we introduce selective weight consolidation\nfor learning-without-forgetting decision-making while encountering new classes.\nExtensive experiments on synthetic and real-world datasets validate the\neffectiveness of our approach.\n", "date": "2023-06-16", "categories": []}
{"id": "2310.08855", "title": "Overcoming Recency Bias of Normalization Statistics in Continual\n  Learning: Balance and Adaptation", "abstract": "  Continual learning entails learning a sequence of tasks and balancing their\nknowledge appropriately. With limited access to old training samples, much of\nthe current work in deep neural networks has focused on overcoming catastrophic\nforgetting of old tasks in gradient-based optimization. However, the\nnormalization layers provide an exception, as they are updated interdependently\nby the gradient and statistics of currently observed training samples, which\nrequire specialized strategies to mitigate recency bias. In this work, we focus\non the most popular Batch Normalization (BN) and provide an in-depth\ntheoretical analysis of its sub-optimality in continual learning. Our analysis\ndemonstrates the dilemma between balance and adaptation of BN statistics for\nincremental tasks, which potentially affects training stability and\ngeneralization. Targeting on these particular challenges, we propose Adaptive\nBalance of BN (AdaB$^2$N), which incorporates appropriately a Bayesian-based\nstrategy to adapt task-wise contributions and a modified momentum to balance BN\nstatistics, corresponding to the training and testing stages. By implementing\nBN in a continual learning fashion, our approach achieves significant\nperformance gains across a wide range of benchmarks, particularly for the\nchallenging yet realistic online scenarios (e.g., up to 7.68%, 6.86% and 4.26%\non Split CIFAR-10, Split CIFAR-100 and Split Mini-ImageNet, respectively). Our\ncode is available at https://github.com/lvyilin/AdaB2N.\n", "date": "2023-10-13", "categories": []}
{"id": "2310.08038", "title": "Continual Learning via Manifold Expansion Replay", "abstract": "  In continual learning, the learner learns multiple tasks in sequence, with\ndata being acquired only once for each task. Catastrophic forgetting is a major\nchallenge to continual learning. To reduce forgetting, some existing\nrehearsal-based methods use episodic memory to replay samples of previous\ntasks. However, in the process of knowledge integration when learning a new\ntask, this strategy also suffers from catastrophic forgetting due to an\nimbalance between old and new knowledge. To address this problem, we propose a\nnovel replay strategy called Manifold Expansion Replay (MaER). We argue that\nexpanding the implicit manifold of the knowledge representation in the episodic\nmemory helps to improve the robustness and expressiveness of the model. To this\nend, we propose a greedy strategy to keep increasing the diameter of the\nimplicit manifold represented by the knowledge in the buffer during memory\nmanagement. In addition, we introduce Wasserstein distance instead of cross\nentropy as distillation loss to preserve previous knowledge. With extensive\nexperimental validation on MNIST, CIFAR10, CIFAR100, and TinyImageNet, we show\nthat the proposed method significantly improves the accuracy in continual\nlearning setup, outperforming the state of the arts.\n", "date": "2023-10-12", "categories": []}
{"id": "2310.08217", "title": "TriRE: A Multi-Mechanism Learning Paradigm for Continual Knowledge\n  Retention and Promotion", "abstract": "  Continual learning (CL) has remained a persistent challenge for deep neural\nnetworks due to catastrophic forgetting (CF) of previously learned tasks.\nSeveral techniques such as weight regularization, experience rehearsal, and\nparameter isolation have been proposed to alleviate CF. Despite their relative\nsuccess, these research directions have predominantly remained orthogonal and\nsuffer from several shortcomings, while missing out on the advantages of\ncompeting strategies. On the contrary, the brain continually learns,\naccommodates, and transfers knowledge across tasks by simultaneously leveraging\nseveral neurophysiological processes, including neurogenesis, active\nforgetting, neuromodulation, metaplasticity, experience rehearsal, and\ncontext-dependent gating, rarely resulting in CF. Inspired by how the brain\nexploits multiple mechanisms concurrently, we propose TriRE, a novel CL\nparadigm that encompasses retaining the most prominent neurons for each task,\nrevising and solidifying the extracted knowledge of current and past tasks, and\nactively promoting less active neurons for subsequent tasks through rewinding\nand relearning. Across CL settings, TriRE significantly reduces task\ninterference and surpasses different CL approaches considered in isolation.\n", "date": "2023-10-12", "categories": []}
{"id": "2310.07234", "title": "Hierarchical Decomposition of Prompt-Based Continual Learning:\n  Rethinking Obscured Sub-optimality", "abstract": "  Prompt-based continual learning is an emerging direction in leveraging\npre-trained knowledge for downstream continual learning, and has almost reached\nthe performance pinnacle under supervised pre-training. However, our empirical\nresearch reveals that the current strategies fall short of their full potential\nunder the more realistic self-supervised pre-training, which is essential for\nhandling vast quantities of unlabeled data in practice. This is largely due to\nthe difficulty of task-specific knowledge being incorporated into instructed\nrepresentations via prompt parameters and predicted by uninstructed\nrepresentations at test time. To overcome the exposed sub-optimality, we\nconduct a theoretical analysis of the continual learning objective in the\ncontext of pre-training, and decompose it into hierarchical components:\nwithin-task prediction, task-identity inference, and task-adaptive prediction.\nFollowing these empirical and theoretical insights, we propose Hierarchical\nDecomposition (HiDe-)Prompt, an innovative approach that explicitly optimizes\nthe hierarchical components with an ensemble of task-specific prompts and\nstatistics of both uninstructed and instructed representations, further with\nthe coordination of a contrastive regularization strategy. Our extensive\nexperiments demonstrate the superior performance of HiDe-Prompt and its\nrobustness to pre-training paradigms in continual learning (e.g., up to 15.01%\nand 9.61% lead on Split CIFAR-100 and Split ImageNet-R, respectively). Our code\nis available at \\url{https://github.com/thu-ml/HiDe-Prompt}.\n", "date": "2023-10-11", "categories": []}
{"id": "2303.07925", "title": "Deep incremental learning models for financial temporal tabular datasets\n  with distribution shifts", "abstract": "  We present a robust deep incremental learning framework for regression tasks\non financial temporal tabular datasets which is built upon the incremental use\nof commonly available tabular and time series prediction models to adapt to\ndistributional shifts typical of financial datasets. The framework uses a\nsimple basic building block (decision trees) to build self-similar models of\nany required complexity to deliver robust performance under adverse situations\nsuch as regime changes, fat-tailed distributions, and low signal-to-noise\nratios. As a detailed study, we demonstrate our scheme using XGBoost models\ntrained on the Numerai dataset and show that a two layer deep ensemble of\nXGBoost models over different model snapshots delivers high quality predictions\nunder different market regimes. We also show that the performance of XGBoost\nmodels with different number of boosting rounds in three scenarios (small,\nstandard and large) is monotonically increasing with respect to model size and\nconverges towards the generalisation upper bound. We also evaluate the\nrobustness of the model under variability of different hyperparameters, such as\nmodel complexity and data sampling settings. Our model has low hardware\nrequirements as no specialised neural architectures are used and each base\nmodel can be independently trained in parallel.\n", "date": "2023-03-14", "categories": []}
{"id": "2306.17560", "title": "Class-Incremental Learning using Diffusion Model for Distillation and\n  Replay", "abstract": "  Class-incremental learning aims to learn new classes in an incremental\nfashion without forgetting the previously learned ones. Several research works\nhave shown how additional data can be used by incremental models to help\nmitigate catastrophic forgetting. In this work, following the recent\nbreakthrough in text-to-image generative models and their wide distribution, we\npropose the use of a pretrained Stable Diffusion model as a source of\nadditional data for class-incremental learning. Compared to competitive methods\nthat rely on external, often unlabeled, datasets of real images, our approach\ncan generate synthetic samples belonging to the same classes as the previously\nencountered images. This allows us to use those additional data samples not\nonly in the distillation loss but also for replay in the classification loss.\nExperiments on the competitive benchmarks CIFAR100, ImageNet-Subset, and\nImageNet demonstrate how this new approach can be used to further improve the\nperformance of state-of-the-art methods for class-incremental learning on large\nscale datasets.\n", "date": "2023-06-30", "categories": []}
{"id": "2310.02995", "title": "IBCL: Zero-shot Model Generation for Task Trade-offs in Continual\n  Learning", "abstract": "  Like generic multi-task learning, continual learning has the nature of\nmulti-objective optimization, and therefore faces a trade-off between the\nperformance of different tasks. That is, to optimize for the current task\ndistribution, it may need to compromise performance on some previous tasks.\nThis means that there exist multiple models that are Pareto-optimal at\ndifferent times, each addressing a distinct task performance trade-off.\nResearchers have discussed how to train particular models to address specific\ntrade-off preferences. However, existing algorithms require training overheads\nproportional to the number of preferences -- a large burden when there are\nmultiple, possibly infinitely many, preferences. As a response, we propose\nImprecise Bayesian Continual Learning (IBCL). Upon a new task, IBCL (1) updates\na knowledge base in the form of a convex hull of model parameter distributions\nand (2) obtains particular models to address task trade-off preferences with\nzero-shot. That is, IBCL does not require any additional training overhead to\ngenerate preference-addressing models from its knowledge base. We show that\nmodels obtained by IBCL have guarantees in identifying the Pareto optimal\nparameters. Moreover, experiments on standard image classification and NLP\ntasks support this guarantee. Statistically, IBCL improves average per-task\naccuracy by at most 23\\% and peak per-task accuracy by at most 15\\% with\nrespect to the baseline methods, with steadily near-zero or positive backward\ntransfer. Most importantly, IBCL significantly reduces the training overhead\nfrom training 1 model per preference to at most 3 models for all preferences.\n", "date": "2023-10-04", "categories": []}
{"id": "2310.06362", "title": "InfoCL: Alleviating Catastrophic Forgetting in Continual Text\n  Classification from An Information Theoretic Perspective", "abstract": "  Continual learning (CL) aims to constantly learn new knowledge over time\nwhile avoiding catastrophic forgetting on old tasks. We focus on continual text\nclassification under the class-incremental setting. Recent CL studies have\nidentified the severe performance decrease on analogous classes as a key factor\nfor catastrophic forgetting. In this paper, through an in-depth exploration of\nthe representation learning process in CL, we discover that the compression\neffect of the information bottleneck leads to confusion on analogous classes.\nTo enable the model learn more sufficient representations, we propose a novel\nreplay-based continual text classification method, InfoCL. Our approach\nutilizes fast-slow and current-past contrastive learning to perform mutual\ninformation maximization and better recover the previously learned\nrepresentations. In addition, InfoCL incorporates an adversarial memory\naugmentation strategy to alleviate the overfitting problem of replay.\nExperimental results demonstrate that InfoCL effectively mitigates forgetting\nand achieves state-of-the-art performance on three text classification tasks.\nThe code is publicly available at https://github.com/Yifan-Song793/InfoCL.\n", "date": "2023-10-10", "categories": []}
{"id": "2310.06368", "title": "CoinSeg: Contrast Inter- and Intra- Class Representations for\n  Incremental Segmentation", "abstract": "  Class incremental semantic segmentation aims to strike a balance between the\nmodel's stability and plasticity by maintaining old knowledge while adapting to\nnew concepts. However, most state-of-the-art methods use the freeze strategy\nfor stability, which compromises the model's plasticity.In contrast, releasing\nparameter training for plasticity could lead to the best performance for all\ncategories, but this requires discriminative feature representation.Therefore,\nwe prioritize the model's plasticity and propose the Contrast inter- and\nintra-class representations for Incremental Segmentation (CoinSeg), which\npursues discriminative representations for flexible parameter tuning. Inspired\nby the Gaussian mixture model that samples from a mixture of Gaussian\ndistributions, CoinSeg emphasizes intra-class diversity with multiple\ncontrastive representation centroids. Specifically, we use mask proposals to\nidentify regions with strong objectness that are likely to be diverse\ninstances/centroids of a category. These mask proposals are then used for\ncontrastive representations to reinforce intra-class diversity. Meanwhile, to\navoid bias from intra-class diversity, we also apply category-level\npseudo-labels to enhance category-level consistency and inter-category\ndiversity. Additionally, CoinSeg ensures the model's stability and alleviates\nforgetting through a specific flexible tuning strategy. We validate CoinSeg on\nPascal VOC 2012 and ADE20K datasets with multiple incremental scenarios and\nachieve superior results compared to previous state-of-the-art methods,\nespecially in more challenging and realistic long-term scenarios. Code is\navailable at https://github.com/zkzhang98/CoinSeg.\n", "date": "2023-10-10", "categories": []}
{"id": "2310.06547", "title": "Rationale-Enhanced Language Models are Better Continual Relation\n  Learners", "abstract": "  Continual relation extraction (CRE) aims to solve the problem of catastrophic\nforgetting when learning a sequence of newly emerging relations. Recent CRE\nstudies have found that catastrophic forgetting arises from the model's lack of\nrobustness against future analogous relations. To address the issue, we\nintroduce rationale, i.e., the explanations of relation classification results\ngenerated by large language models (LLM), into CRE task. Specifically, we\ndesign the multi-task rationale tuning strategy to help the model learn current\nrelations robustly. We also conduct contrastive rationale replay to further\ndistinguish analogous relations. Experimental results on two standard\nbenchmarks demonstrate that our method outperforms the state-of-the-art CRE\nmodels.\n", "date": "2023-10-10", "categories": []}
{"id": "2310.06762", "title": "TRACE: A Comprehensive Benchmark for Continual Learning in Large\n  Language Models", "abstract": "  Aligned large language models (LLMs) demonstrate exceptional capabilities in\ntask-solving, following instructions, and ensuring safety. However, the\ncontinual learning aspect of these aligned LLMs has been largely overlooked.\nExisting continual learning benchmarks lack sufficient challenge for leading\naligned LLMs, owing to both their simplicity and the models' potential exposure\nduring instruction tuning. In this paper, we introduce TRACE, a novel benchmark\ndesigned to evaluate continual learning in LLMs. TRACE consists of 8 distinct\ndatasets spanning challenging tasks including domain-specific tasks,\nmultilingual capabilities, code generation, and mathematical reasoning. All\ndatasets are standardized into a unified format, allowing for effortless\nautomatic evaluation of LLMs. Our experiments show that after training on\nTRACE, aligned LLMs exhibit significant declines in both general ability and\ninstruction-following capabilities. For example, the accuracy of llama2-chat\n13B on gsm8k dataset declined precipitously from 28.8\\% to 2\\% after training\non our datasets. This highlights the challenge of finding a suitable tradeoff\nbetween achieving performance on specific tasks while preserving the original\nprowess of LLMs. Empirical findings suggest that tasks inherently equipped with\nreasoning paths contribute significantly to preserving certain capabilities of\nLLMs against potential declines. Motivated by this, we introduce the\nReasoning-augmented Continual Learning (RCL) approach. RCL integrates\ntask-specific cues with meta-rationales, effectively reducing catastrophic\nforgetting in LLMs while expediting convergence on novel tasks.\n", "date": "2023-10-10", "categories": []}
{"id": "2207.05225", "title": "Susceptibility of Continual Learning Against Adversarial Attacks", "abstract": "  Recent continual learning approaches have primarily focused on mitigating\ncatastrophic forgetting. Nevertheless, two critical areas have remained\nrelatively unexplored: 1) evaluating the robustness of proposed methods and 2)\nensuring the security of learned tasks. This paper investigates the\nsusceptibility of continually learned tasks, including current and previously\nacquired tasks, to adversarial attacks. Specifically, we have observed that any\nclass belonging to any task can be easily targeted and misclassified as the\ndesired target class of any other task. Such susceptibility or vulnerability of\nlearned tasks to adversarial attacks raises profound concerns regarding data\nintegrity and privacy. To assess the robustness of continual learning\napproaches, we consider continual learning approaches in all three scenarios,\ni.e., task-incremental learning, domain-incremental learning, and\nclass-incremental learning. In this regard, we explore the robustness of three\nregularization-based methods, three replay-based approaches, and one hybrid\ntechnique that combines replay and exemplar approaches. We empirically\ndemonstrated that in any setting of continual learning, any class, whether\nbelonging to the current or previously learned tasks, is susceptible to\nmisclassification. Our observations identify potential limitations of continual\nlearning approaches against adversarial attacks and highlight that current\ncontinual learning algorithms could not be suitable for deployment in\nreal-world settings.\n", "date": "2022-07-11", "categories": []}
{"id": "2303.05118", "title": "SLCA: Slow Learner with Classifier Alignment for Continual Learning on a\n  Pre-trained Model", "abstract": "  The goal of continual learning is to improve the performance of recognition\nmodels in learning sequentially arrived data. Although most existing works are\nestablished on the premise of learning from scratch, growing efforts have been\ndevoted to incorporating the benefits of pre-training. However, how to\nadaptively exploit the pre-trained knowledge for each incremental task while\nmaintaining its generalizability remains an open question. In this work, we\npresent an extensive analysis for continual learning on a pre-trained model\n(CLPM), and attribute the key challenge to a progressive overfitting problem.\nObserving that selectively reducing the learning rate can almost resolve this\nissue in the representation layer, we propose a simple but extremely effective\napproach named Slow Learner with Classifier Alignment (SLCA), which further\nimproves the classification layer by modeling the class-wise distributions and\naligning the classification layers in a post-hoc fashion. Across a variety of\nscenarios, our proposal provides substantial improvements for CLPM (e.g., up to\n49.76%, 50.05%, 44.69% and 40.16% on Split CIFAR-100, Split ImageNet-R, Split\nCUB-200 and Split Cars-196, respectively), and thus outperforms\nstate-of-the-art approaches by a large margin. Based on such a strong baseline,\ncritical factors and promising directions are analyzed in-depth to facilitate\nsubsequent research. Code has been made available at:\nhttps://github.com/GengDavid/SLCA.\n", "date": "2023-03-09", "categories": []}
{"id": "2308.13957", "title": "Differentiable Weight Masks for Domain Transfer", "abstract": "  One of the major drawbacks of deep learning models for computer vision has\nbeen their inability to retain multiple sources of information in a modular\nfashion. For instance, given a network that has been trained on a source task,\nwe would like to re-train this network on a similar, yet different, target task\nwhile maintaining its performance on the source task. Simultaneously,\nresearchers have extensively studied modularization of network weights to\nlocalize and identify the set of weights culpable for eliciting the observed\nperformance on a given task. One set of works studies the modularization\ninduced in the weights of a neural network by learning and analysing weight\nmasks. In this work, we combine these fields to study three such weight masking\nmethods and analyse their ability to mitigate \"forgetting'' on the source task\nwhile also allowing for efficient finetuning on the target task. We find that\ndifferent masking techniques have trade-offs in retaining knowledge in the\nsource task without adversely affecting target task performance.\n", "date": "2023-08-26", "categories": []}
{"id": "2310.04801", "title": "Parameterizing Context: Unleashing the Power of Parameter-Efficient\n  Fine-Tuning and In-Context Tuning for Continual Table Semantic Parsing", "abstract": "  Continual table semantic parsing aims to train a parser on a sequence of\ntasks, where each task requires the parser to translate natural language into\nSQL based on task-specific tables but only offers limited training examples.\nConventional methods tend to suffer from overfitting with limited supervision,\nas well as catastrophic forgetting due to parameter updates. Despite recent\nadvancements that partially alleviate these issues through semi-supervised data\naugmentation and retention of a few past examples, the performance is still\nlimited by the volume of unsupervised data and stored examples. To overcome\nthese challenges, this paper introduces a novel method integrating\n\\textit{parameter-efficient fine-tuning} (PEFT) and \\textit{in-context tuning}\n(ICT) for training a continual table semantic parser. Initially, we present a\ntask-adaptive PEFT framework capable of fully circumventing catastrophic\nforgetting, which is achieved by freezing the pre-trained model backbone and\nfine-tuning small-scale prompts. Building on this, we propose a teacher-student\nframework-based solution. The teacher addresses the few-shot problem using ICT,\nwhich procures contextual information by demonstrating a few training examples.\nIn turn, the student leverages the proposed PEFT framework to learn from the\nteacher's output distribution, and subsequently compresses and saves the\ncontextual information to the prompts, eliminating the need to store any\ntraining examples. Experimental evaluations on two benchmarks affirm the\nsuperiority of our method over prevalent few-shot and continual learning\nbaselines across various metrics.\n", "date": "2023-10-07", "categories": []}
{"id": "2310.05644", "title": "Diagnosing Catastrophe: Large parts of accuracy loss in continual\n  learning can be accounted for by readout misalignment", "abstract": "  Unlike primates, training artificial neural networks on changing data\ndistributions leads to a rapid decrease in performance on old tasks. This\nphenomenon is commonly referred to as catastrophic forgetting. In this paper,\nwe investigate the representational changes that underlie this performance\ndecrease and identify three distinct processes that together account for the\nphenomenon. The largest component is a misalignment between hidden\nrepresentations and readout layers. Misalignment occurs due to learning on\nadditional tasks and causes internal representations to shift. Representational\ngeometry is partially conserved under this misalignment and only a small part\nof the information is irrecoverably lost. All types of representational changes\nscale with the dimensionality of hidden representations. These insights have\nimplications for deep learning applications that need to be continuously\nupdated, but may also aid aligning ANN models to the rather robust biological\nvision.\n", "date": "2023-10-09", "categories": []}
{"id": "2309.16905", "title": "Towards a Unified Framework for Adaptable Problematic Content Detection\n  via Continual Learning", "abstract": "  Detecting problematic content, such as hate speech, is a multifaceted and\never-changing task, influenced by social dynamics, user populations, diversity\nof sources, and evolving language. There has been significant efforts, both in\nacademia and in industry, to develop annotated resources that capture various\naspects of problematic content. Due to researchers' diverse objectives, the\nannotations are inconsistent and hence, reports of progress on detection of\nproblematic content are fragmented. This pattern is expected to persist unless\nwe consolidate resources considering the dynamic nature of the problem. We\npropose integrating the available resources, and leveraging their dynamic\nnature to break this pattern. In this paper, we introduce a continual learning\nbenchmark and framework for problematic content detection comprising over 84\nrelated tasks encompassing 15 annotation schemas from 8 sources. Our benchmark\ncreates a novel measure of progress: prioritizing the adaptability of\nclassifiers to evolving tasks over excelling in specific tasks. To ensure the\ncontinuous relevance of our framework, we designed it so that new tasks can\neasily be integrated into the benchmark. Our baseline results demonstrate the\npotential of continual learning in capturing the evolving content and adapting\nto novel manifestations of problematic content.\n", "date": "2023-09-29", "categories": []}
{"id": "2310.03848", "title": "OpenIncrement: A Unified Framework for Open Set Recognition and Deep\n  Class-Incremental Learning", "abstract": "  In most works on deep incremental learning research, it is assumed that novel\nsamples are pre-identified for neural network retraining. However, practical\ndeep classifiers often misidentify these samples, leading to erroneous\npredictions. Such misclassifications can degrade model performance. Techniques\nlike open set recognition offer a means to detect these novel samples,\nrepresenting a significant area in the machine learning domain.\n  In this paper, we introduce a deep class-incremental learning framework\nintegrated with open set recognition. Our approach refines class-incrementally\nlearned features to adapt them for distance-based open set recognition.\nExperimental results validate that our method outperforms state-of-the-art\nincremental learning techniques and exhibits superior performance in open set\nrecognition compared to baseline methods.\n", "date": "2023-10-05", "categories": []}
{"id": "2310.03898", "title": "Class-Incremental Learning Using Generative Experience Replay Based on\n  Time-aware Regularization", "abstract": "  Learning new tasks accumulatively without forgetting remains a critical\nchallenge in continual learning. Generative experience replay addresses this\nchallenge by synthesizing pseudo-data points for past learned tasks and later\nreplaying them for concurrent training along with the new tasks' data.\nGenerative replay is the best strategy for continual learning under a strict\nclass-incremental setting when certain constraints need to be met: (i) constant\nmodel size, (ii) no pre-training dataset, and (iii) no memory buffer for\nstoring past tasks' data. Inspired by the biological nervous system mechanisms,\nwe introduce a time-aware regularization method to dynamically fine-tune the\nthree training objective terms used for generative replay: supervised learning,\nlatent regularization, and data reconstruction. Experimental results on major\nbenchmarks indicate that our method pushes the limit of brain-inspired\ncontinual learners under such strict settings, improves memory retention, and\nincreases the average performance over continually arriving tasks.\n", "date": "2023-10-05", "categories": []}
{"id": "2310.04334", "title": "Saliency-Guided Hidden Associative Replay for Continual Learning", "abstract": "  Continual Learning is a burgeoning domain in next-generation AI, focusing on\ntraining neural networks over a sequence of tasks akin to human learning. While\nCL provides an edge over traditional supervised learning, its central challenge\nremains to counteract catastrophic forgetting and ensure the retention of prior\ntasks during subsequent learning. Amongst various strategies to tackle this,\nreplay based methods have emerged as preeminent, echoing biological memory\nmechanisms. However, these methods are memory intensive, often preserving\nentire data samples, an approach inconsistent with humans selective memory\nretention of salient experiences. While some recent works have explored the\nstorage of only significant portions of data in episodic memory, the inherent\nnature of partial data necessitates innovative retrieval mechanisms. Current\nsolutions, like inpainting, approximate full data reconstruction from partial\ncues, a method that diverges from genuine human memory processes. Addressing\nthese nuances, this paper presents the Saliency Guided Hidden Associative\nReplay for Continual Learning. This novel framework synergizes associative\nmemory with replay-based strategies. SHARC primarily archives salient data\nsegments via sparse memory encoding. Importantly, by harnessing associative\nmemory paradigms, it introduces a content focused memory retrieval mechanism,\npromising swift and near-perfect recall, bringing CL a step closer to authentic\nhuman memory processes. Extensive experimental results demonstrate the\neffectiveness of our proposed method for various continual learning tasks.\n", "date": "2023-10-06", "categories": []}
{"id": "2310.03675", "title": "Hadamard Domain Training with Integers for Class Incremental Quantized\n  Learning", "abstract": "  Continual learning is a desirable feature in many modern machine learning\napplications, which allows in-field adaptation and updating, ranging from\naccommodating distribution shift, to fine-tuning, and to learning new tasks.\nFor applications with privacy and low latency requirements, the compute and\nmemory demands imposed by continual learning can be cost-prohibitive for\nresource-constraint edge platforms. Reducing computational precision through\nfully quantized training (FQT) simultaneously reduces memory footprint and\nincreases compute efficiency for both training and inference. However,\naggressive quantization especially integer FQT typically degrades model\naccuracy to unacceptable levels. In this paper, we propose a technique that\nleverages inexpensive Hadamard transforms to enable low-precision training with\nonly integer matrix multiplications. We further determine which tensors need\nstochastic rounding and propose tiled matrix multiplication to enable low-bit\nwidth accumulators. We demonstrate the effectiveness of our technique on\nseveral human activity recognition datasets and CIFAR100 in a class incremental\nlearning setting. We achieve less than 0.5% and 3% accuracy degradation while\nwe quantize all matrix multiplications inputs down to 4-bits with 8-bit\naccumulators.\n", "date": "2023-10-05", "categories": []}
{"id": "2207.01145", "title": "Memory Population in Continual Learning via Outlier Elimination", "abstract": "  Catastrophic forgetting, the phenomenon of forgetting previously learned\ntasks when learning a new one, is a major hurdle in developing continual\nlearning algorithms. A popular method to alleviate forgetting is to use a\nmemory buffer, which stores a subset of previously learned task examples for\nuse during training on new tasks. The de facto method of filling memory is by\nrandomly selecting previous examples. However, this process could introduce\noutliers or noisy samples that could hurt the generalization of the model. This\npaper introduces Memory Outlier Elimination (MOE), a method for identifying and\neliminating outliers in the memory buffer by choosing samples from\nlabel-homogeneous subpopulations. We show that a space with a high homogeneity\nis related to a feature space that is more representative of the class\ndistribution. In practice, MOE removes a sample if it is surrounded by samples\nfrom different labels. We demonstrate the effectiveness of MOE on CIFAR-10,\nCIFAR-100, and CORe50, outperforming previous well-known memory population\nmethods.\n", "date": "2022-07-04", "categories": []}
{"id": "2305.15700", "title": "Fairness Continual Learning Approach to Semantic Scene Understanding in\n  Open-World Environments", "abstract": "  Continual semantic segmentation aims to learn new classes while maintaining\nthe information from the previous classes. Although prior studies have shown\nimpressive progress in recent years, the fairness concern in the continual\nsemantic segmentation needs to be better addressed. Meanwhile, fairness is one\nof the most vital factors in deploying the deep learning model, especially in\nhuman-related or safety applications. In this paper, we present a novel\nFairness Continual Learning approach to the semantic segmentation problem. In\nparticular, under the fairness objective, a new fairness continual learning\nframework is proposed based on class distributions. Then, a novel Prototypical\nContrastive Clustering loss is proposed to address the significant challenges\nin continual learning, i.e., catastrophic forgetting and background shift. Our\nproposed loss has also been proven as a novel, generalized learning paradigm of\nknowledge distillation commonly used in continual learning. Moreover, the\nproposed Conditional Structural Consistency loss further regularized the\nstructural constraint of the predicted segmentation. Our proposed approach has\nachieved State-of-the-Art performance on three standard scene understanding\nbenchmarks, i.e., ADE20K, Cityscapes, and Pascal VOC, and promoted the fairness\nof the segmentation model.\n", "date": "2023-05-25", "categories": []}
{"id": "2307.08114", "title": "Tangent Model Composition for Ensembling and Continual Fine-tuning", "abstract": "  Tangent Model Composition (TMC) is a method to combine component models\nindependently fine-tuned around a pre-trained point. Component models are\ntangent vectors to the pre-trained model that can be added, scaled, or\nsubtracted to support incremental learning, ensembling, or unlearning.\nComponent models are composed at inference time via scalar combination,\nreducing the cost of ensembling to that of a single model. TMC improves\naccuracy by 4.2% compared to ensembling non-linearly fine-tuned models at a\n2.5x to 10x reduction of inference cost, growing linearly with the number of\ncomponent models. Each component model can be forgotten at zero cost, with no\nresidual effect on the resulting inference. When used for continual\nfine-tuning, TMC is not constrained by sequential bias and can be executed in\nparallel on federated data. TMC outperforms recently published continual\nfine-tuning methods almost uniformly on each setting -- task-incremental,\nclass-incremental, and data-incremental -- on a total of 13 experiments across\n3 benchmark datasets, despite not using any replay buffer. TMC is designed for\ncomposing models that are local to a pre-trained embedding, but could be\nextended to more general settings. The code is available at:\nhttps://github.com/tianyu139/tangent-model-composition\n", "date": "2023-07-16", "categories": []}
{"id": "2308.07939", "title": "Ada-QPacknet -- adaptive pruning with bit width reduction as an\n  efficient continual learning method without forgetting", "abstract": "  Continual Learning (CL) is a process in which there is still huge gap between\nhuman and deep learning model efficiency. Recently, many CL algorithms were\ndesigned. Most of them have many problems with learning in dynamic and complex\nenvironments. In this work new architecture based approach Ada-QPacknet is\ndescribed. It incorporates the pruning for extracting the sub-network for each\ntask. The crucial aspect in architecture based CL methods is theirs capacity.\nIn presented method the size of the model is reduced by efficient linear and\nnonlinear quantisation approach. The method reduces the bit-width of the\nweights format. The presented results shows that low bit quantisation achieves\nsimilar accuracy as floating-point sub-network on a well-know CL scenarios. To\nour knowledge it is the first CL strategy which incorporates both compression\ntechniques pruning and quantisation for generating task sub-networks. The\npresented algorithm was tested on well-known episode combinations and compared\nwith most popular algorithms. Results show that proposed approach outperforms\nmost of the CL strategies in task and class incremental scenarios.\n", "date": "2023-08-14", "categories": []}
{"id": "2310.01165", "title": "Towards guarantees for parameter isolation in continual learning", "abstract": "  Deep learning has proved to be a successful paradigm for solving many\nchallenges in machine learning. However, deep neural networks fail when trained\nsequentially on multiple tasks, a shortcoming known as catastrophic forgetting\nin the continual learning literature. Despite a recent flourish of learning\nalgorithms successfully addressing this problem, we find that provable\nguarantees against catastrophic forgetting are lacking. In this work, we study\nthe relationship between learning and forgetting by looking at the geometry of\nneural networks' loss landscape. We offer a unifying perspective on a family of\ncontinual learning algorithms, namely methods based on parameter isolation, and\nwe establish guarantees on catastrophic forgetting for some of them.\n", "date": "2023-10-02", "categories": []}
{"id": "2310.01365", "title": "Elephant Neural Networks: Born to Be a Continual Learner", "abstract": "  Catastrophic forgetting remains a significant challenge to continual learning\nfor decades. While recent works have proposed effective methods to mitigate\nthis problem, they mainly focus on the algorithmic side. Meanwhile, we do not\nfully understand what architectural properties of neural networks lead to\ncatastrophic forgetting. This study aims to fill this gap by studying the role\nof activation functions in the training dynamics of neural networks and their\nimpact on catastrophic forgetting. Our study reveals that, besides sparse\nrepresentations, the gradient sparsity of activation functions also plays an\nimportant role in reducing forgetting. Based on this insight, we propose a new\nclass of activation functions, elephant activation functions, that can generate\nboth sparse representations and sparse gradients. We show that by simply\nreplacing classical activation functions with elephant activation functions, we\ncan significantly improve the resilience of neural networks to catastrophic\nforgetting. Our method has broad applicability and benefits for continual\nlearning in regression, class incremental learning, and reinforcement learning\ntasks. Specifically, we achieves excellent performance on Split MNIST dataset\nin just one single pass, without using replay buffer, task boundary\ninformation, or pre-training.\n", "date": "2023-10-02", "categories": []}
{"id": "2309.15413", "title": "Inherit with Distillation and Evolve with Contrast: Exploring Class\n  Incremental Semantic Segmentation Without Exemplar Memory", "abstract": "  As a front-burner problem in incremental learning, class incremental semantic\nsegmentation (CISS) is plagued by catastrophic forgetting and semantic drift.\nAlthough recent methods have utilized knowledge distillation to transfer\nknowledge from the old model, they are still unable to avoid pixel confusion,\nwhich results in severe misclassification after incremental steps due to the\nlack of annotations for past and future classes. Meanwhile data-replay-based\napproaches suffer from storage burdens and privacy concerns. In this paper, we\npropose to address CISS without exemplar memory and resolve catastrophic\nforgetting as well as semantic drift synchronously. We present Inherit with\nDistillation and Evolve with Contrast (IDEC), which consists of a Dense\nKnowledge Distillation on all Aspects (DADA) manner and an Asymmetric\nRegion-wise Contrastive Learning (ARCL) module. Driven by the devised dynamic\nclass-specific pseudo-labelling strategy, DADA distils intermediate-layer\nfeatures and output-logits collaboratively with more emphasis on\nsemantic-invariant knowledge inheritance. ARCL implements region-wise\ncontrastive learning in the latent space to resolve semantic drift among known\nclasses, current classes, and unknown classes. We demonstrate the effectiveness\nof our method on multiple CISS tasks by state-of-the-art performance, including\nPascal VOC 2012, ADE20K and ISPRS datasets. Our method also shows superior\nanti-forgetting ability, particularly in multi-step CISS tasks.\n", "date": "2023-09-27", "categories": []}
{"id": "2309.17192", "title": "A Survey of Incremental Transfer Learning: Combining Peer-to-Peer\n  Federated Learning and Domain Incremental Learning for Multicenter\n  Collaboration", "abstract": "  Due to data privacy constraints, data sharing among multiple clinical centers\nis restricted, which impedes the development of high performance deep learning\nmodels from multicenter collaboration. Naive weight transfer methods share\nintermediate model weights without raw data and hence can bypass data privacy\nrestrictions. However, performance drops are typically observed when the model\nis transferred from one center to the next because of the forgetting problem.\nIncremental transfer learning, which combines peer-to-peer federated learning\nand domain incremental learning, can overcome the data privacy issue and\nmeanwhile preserve model performance by using continual learning techniques. In\nthis work, a conventional domain/task incremental learning framework is adapted\nfor incremental transfer learning. A comprehensive survey on the efficacy of\ndifferent regularization-based continual learning methods for multicenter\ncollaboration is performed. The influences of data heterogeneity, classifier\nhead setting, network optimizer, model initialization, center order, and weight\ntransfer type have been investigated thoroughly. Our framework is publicly\naccessible to the research community for further development.\n", "date": "2023-09-29", "categories": []}
{"id": "2308.11677", "title": "An Analysis of Initial Training Strategies for Exemplar-Free\n  Class-Incremental Learning", "abstract": "  Class-Incremental Learning (CIL) aims to build classification models from\ndata streams. At each step of the CIL process, new classes must be integrated\ninto the model. Due to catastrophic forgetting, CIL is particularly challenging\nwhen examples from past classes cannot be stored, the case on which we focus\nhere. To date, most approaches are based exclusively on the target dataset of\nthe CIL process. However, the use of models pre-trained in a self-supervised\nway on large amounts of data has recently gained momentum. The initial model of\nthe CIL process may only use the first batch of the target dataset, or also use\npre-trained weights obtained on an auxiliary dataset. The choice between these\ntwo initial learning strategies can significantly influence the performance of\nthe incremental learning model, but has not yet been studied in depth.\nPerformance is also influenced by the choice of the CIL algorithm, the neural\narchitecture, the nature of the target task, the distribution of classes in the\nstream and the number of examples available for learning. We conduct a\ncomprehensive experimental study to assess the roles of these factors. We\npresent a statistical analysis framework that quantifies the relative\ncontribution of each factor to incremental performance. Our main finding is\nthat the initial training strategy is the dominant factor influencing the\naverage incremental accuracy, but that the choice of CIL algorithm is more\nimportant in preventing forgetting. Based on this analysis, we propose\npractical recommendations for choosing the right initial training strategy for\na given incremental learning use case. These recommendations are intended to\nfacilitate the practical deployment of incremental learning.\n", "date": "2023-08-22", "categories": []}
{"id": "2308.10328", "title": "A Comprehensive Empirical Evaluation on Online Continual Learning", "abstract": "  Online continual learning aims to get closer to a live learning experience by\nlearning directly on a stream of data with temporally shifting distribution and\nby storing a minimum amount of data from that stream. In this empirical\nevaluation, we evaluate various methods from the literature that tackle online\ncontinual learning. More specifically, we focus on the class-incremental\nsetting in the context of image classification, where the learner must learn\nnew classes incrementally from a stream of data. We compare these methods on\nthe Split-CIFAR100 and Split-TinyImagenet benchmarks, and measure their average\naccuracy, forgetting, stability, and quality of the representations, to\nevaluate various aspects of the algorithm at the end but also during the whole\ntraining period. We find that most methods suffer from stability and\nunderfitting issues. However, the learned representations are comparable to\ni.i.d. training under the same computational budget. No clear winner emerges\nfrom the results and basic experience replay, when properly tuned and\nimplemented, is a very strong baseline. We release our modular and extensible\ncodebase at https://github.com/AlbinSou/ocl_survey based on the avalanche\nframework to reproduce our results and encourage future research.\n", "date": "2023-08-20", "categories": []}
{"id": "2309.13563", "title": "Multivariate Prototype Representation for Domain-Generalized Incremental\n  Learning", "abstract": "  Deep learning models suffer from catastrophic forgetting when being\nfine-tuned with samples of new classes. This issue becomes even more pronounced\nwhen faced with the domain shift between training and testing data. In this\npaper, we study the critical and less explored Domain-Generalized\nClass-Incremental Learning (DGCIL). We design a DGCIL approach that remembers\nold classes, adapts to new classes, and can classify reliably objects from\nunseen domains. Specifically, our loss formulation maintains classification\nboundaries and suppresses the domain-specific information of each class. With\nno old exemplars stored, we use knowledge distillation and estimate old class\nprototype drift as incremental training advances. Our prototype representations\nare based on multivariate Normal distributions whose means and covariances are\nconstantly adapted to changing model features to represent old classes well by\nadapting to the feature space drift. For old classes, we sample pseudo-features\nfrom the adapted Normal distributions with the help of Cholesky decomposition.\nIn contrast to previous pseudo-feature sampling strategies that rely solely on\naverage mean prototypes, our method excels at capturing varying semantic\ninformation. Experiments on several benchmarks validate our claims.\n", "date": "2023-09-24", "categories": []}
{"id": "2309.12078", "title": "Clustering-based Domain-Incremental Learning", "abstract": "  We consider the problem of learning multiple tasks in a continual learning\nsetting in which data from different tasks is presented to the learner in a\nstreaming fashion. A key challenge in this setting is the so-called\n\"catastrophic forgetting problem\", in which the performance of the learner in\nan \"old task\" decreases when subsequently trained on a \"new task\". Existing\ncontinual learning methods, such as Averaged Gradient Episodic Memory (A-GEM)\nand Orthogonal Gradient Descent (OGD), address catastrophic forgetting by\nminimizing the loss for the current task without increasing the loss for\nprevious tasks. However, these methods assume the learner knows when the task\nchanges, which is unrealistic in practice. In this paper, we alleviate the need\nto provide the algorithm with information about task changes by using an online\nclustering-based approach on a dynamically updated finite pool of samples or\ngradients. We thereby successfully counteract catastrophic forgetting in one of\nthe hardest settings, namely: domain-incremental learning, a setting for which\nthe problem was previously unsolved. We showcase the benefits of our approach\nby applying these ideas to projection-based methods, such as A-GEM and OGD,\nwhich lead to task-agnostic versions of them. Experiments on real datasets\ndemonstrate the effectiveness of the proposed strategy and its promising\nperformance compared to state-of-the-art methods.\n", "date": "2023-09-21", "categories": []}
{"id": "2309.11305", "title": "Create and Find Flatness: Building Flat Training Spaces in Advance for\n  Continual Learning", "abstract": "  Catastrophic forgetting remains a critical challenge in the field of\ncontinual learning, where neural networks struggle to retain prior knowledge\nwhile assimilating new information. Most existing studies emphasize mitigating\nthis issue only when encountering new tasks, overlooking the significance of\nthe pre-task phase. Therefore, we shift the attention to the current task\nlearning stage, presenting a novel framework, C&F (Create and Find Flatness),\nwhich builds a flat training space for each task in advance. Specifically,\nduring the learning of the current task, our framework adaptively creates a\nflat region around the minimum in the loss landscape. Subsequently, it finds\nthe parameters' importance to the current task based on their flatness degrees.\nWhen adapting the model to a new task, constraints are applied according to the\nflatness and a flat space is simultaneously prepared for the impending task. We\ntheoretically demonstrate the consistency between the created and found\nflatness. In this manner, our framework not only accommodates ample parameter\nspace for learning new tasks but also preserves the preceding knowledge of\nearlier tasks. Experimental results exhibit C&F's state-of-the-art performance\nas a standalone continual learning approach and its efficacy as a framework\nincorporating other methods. Our work is available at\nhttps://github.com/Eric8932/Create-and-Find-Flatness.\n", "date": "2023-09-20", "categories": []}
{"id": "2309.09455", "title": "CaT: Balanced Continual Graph Learning with Graph Condensation", "abstract": "  Continual graph learning (CGL) is purposed to continuously update a graph\nmodel with graph data being fed in a streaming manner. Since the model easily\nforgets previously learned knowledge when training with new-coming data, the\ncatastrophic forgetting problem has been the major focus in CGL. Recent\nreplay-based methods intend to solve this problem by updating the model using\nboth (1) the entire new-coming data and (2) a sampling-based memory bank that\nstores replayed graphs to approximate the distribution of historical data.\nAfter updating the model, a new replayed graph sampled from the incoming graph\nwill be added to the existing memory bank. Despite these methods are intuitive\nand effective for the CGL, two issues are identified in this paper. Firstly,\nmost sampling-based methods struggle to fully capture the historical\ndistribution when the storage budget is tight. Secondly, a significant data\nimbalance exists in terms of the scales of the complex new-coming graph data\nand the lightweight memory bank, resulting in unbalanced training. To solve\nthese issues, a Condense and Train (CaT) framework is proposed in this paper.\nPrior to each model update, the new-coming graph is condensed to a small yet\ninformative synthesised replayed graph, which is then stored in a Condensed\nGraph Memory with historical replay graphs. In the continual learning phase, a\nTraining in Memory scheme is used to update the model directly with the\nCondensed Graph Memory rather than the whole new-coming graph, which alleviates\nthe data imbalance problem. Extensive experiments conducted on four benchmark\ndatasets successfully demonstrate superior performances of the proposed CaT\nframework in terms of effectiveness and efficiency. The code has been released\non https://github.com/superallen13/CaT-CGL.\n", "date": "2023-09-18", "categories": []}
{"id": "2309.10012", "title": "Looking through the past: better knowledge retention for generative\n  replay in continual learning", "abstract": "  In this work, we improve the generative replay in a continual learning\nsetting to perform well on challenging scenarios. Current generative rehearsal\nmethods are usually benchmarked on small and simple datasets as they are not\npowerful enough to generate more complex data with a greater number of classes.\nWe notice that in VAE-based generative replay, this could be attributed to the\nfact that the generated features are far from the original ones when mapped to\nthe latent space. Therefore, we propose three modifications that allow the\nmodel to learn and generate complex data. More specifically, we incorporate the\ndistillation in latent space between the current and previous models to reduce\nfeature drift. Additionally, a latent matching for the reconstruction and\noriginal data is proposed to improve generated features alignment. Further,\nbased on the observation that the reconstructions are better for preserving\nknowledge, we add the cycling of generations through the previously trained\nmodel to make them closer to the original data. Our method outperforms other\ngenerative replay methods in various scenarios. Code available at\nhttps://github.com/valeriya-khan/looking-through-the-past.\n", "date": "2023-09-18", "categories": []}
{"id": "2309.09122", "title": "FDCNet: Feature Drift Compensation Network for Class-Incremental Weakly\n  Supervised Object Localization", "abstract": "  This work addresses the task of class-incremental weakly supervised object\nlocalization (CI-WSOL). The goal is to incrementally learn object localization\nfor novel classes using only image-level annotations while retaining the\nability to localize previously learned classes. This task is important because\nannotating bounding boxes for every new incoming data is expensive, although\nobject localization is crucial in various applications. To the best of our\nknowledge, we are the first to address this task. Thus, we first present a\nstrong baseline method for CI-WSOL by adapting the strategies of\nclass-incremental classifiers to mitigate catastrophic forgetting. These\nstrategies include applying knowledge distillation, maintaining a small data\nset from previous tasks, and using cosine normalization. We then propose the\nfeature drift compensation network to compensate for the effects of feature\ndrifts on class scores and localization maps. Since updating network parameters\nto learn new tasks causes feature drifts, compensating for the final outputs is\nnecessary. Finally, we evaluate our proposed method by conducting experiments\non two publicly available datasets (ImageNet-100 and CUB-200). The experimental\nresults demonstrate that the proposed method outperforms other baseline\nmethods.\n", "date": "2023-09-17", "categories": []}
{"id": "2309.06896", "title": "Domain-Aware Augmentations for Unsupervised Online General Continual\n  Learning", "abstract": "  Continual Learning has been challenging, especially when dealing with\nunsupervised scenarios such as Unsupervised Online General Continual Learning\n(UOGCL), where the learning agent has no prior knowledge of class boundaries or\ntask change information. While previous research has focused on reducing\nforgetting in supervised setups, recent studies have shown that self-supervised\nlearners are more resilient to forgetting. This paper proposes a novel approach\nthat enhances memory usage for contrastive learning in UOGCL by defining and\nusing stream-dependent data augmentations together with some implementation\ntricks. Our proposed method is simple yet effective, achieves state-of-the-art\nresults compared to other unsupervised approaches in all considered setups, and\nreduces the gap between supervised and unsupervised continual learning. Our\ndomain-aware augmentation procedure can be adapted to other replay-based\nmethods, making it a promising strategy for continual learning.\n", "date": "2023-09-13", "categories": []}
{"id": "2309.06917", "title": "Continual Learning with Dirichlet Generative-based Rehearsal", "abstract": "  Recent advancements in data-driven task-oriented dialogue systems (ToDs)\nstruggle with incremental learning due to computational constraints and\ntime-consuming issues. Continual Learning (CL) attempts to solve this by\navoiding intensive pre-training, but it faces the problem of catastrophic\nforgetting (CF). While generative-based rehearsal CL methods have made\nsignificant strides, generating pseudo samples that accurately reflect the\nunderlying task-specific distribution is still a challenge. In this paper, we\npresent Dirichlet Continual Learning (DCL), a novel generative-based rehearsal\nstrategy for CL. Unlike the traditionally used Gaussian latent variable in the\nConditional Variational Autoencoder (CVAE), DCL leverages the flexibility and\nversatility of the Dirichlet distribution to model the latent prior variable.\nThis enables it to efficiently capture sentence-level features of previous\ntasks and effectively guide the generation of pseudo samples. In addition, we\nintroduce Jensen-Shannon Knowledge Distillation (JSKD), a robust logit-based\nknowledge distillation method that enhances knowledge transfer during pseudo\nsample generation. Our experiments confirm the efficacy of our approach in both\nintent detection and slot-filling tasks, outperforming state-of-the-art\nmethods.\n", "date": "2023-09-13", "categories": []}
{"id": "2303.05911", "title": "Lifelong Machine Learning Potentials", "abstract": "  Machine learning potentials (MLPs) trained on accurate quantum chemical data\ncan retain the high accuracy, while inflicting little computational demands. On\nthe downside, they need to be trained for each individual system. In recent\nyears, a vast number of MLPs has been trained from scratch because learning\nadditional data typically requires to train again on all data to not forget\npreviously acquired knowledge. Additionally, most common structural descriptors\nof MLPs cannot represent efficiently a large number of different chemical\nelements. In this work, we tackle these problems by introducing\nelement-embracing atom-centered symmetry functions (eeACSFs) which combine\nstructural properties and element information from the periodic table. These\neeACSFs are a key for our development of a lifelong machine learning potential\n(lMLP). Uncertainty quantification can be exploited to transgress a fixed,\npre-trained MLP to arrive at a continuously adapting lMLP, because a predefined\nlevel of accuracy can be ensured. To extend the applicability of an lMLP to new\nsystems, we apply continual learning strategies to enable autonomous and\non-the-fly training on a continuous stream of new data. For the training of\ndeep neural networks, we propose the continual resilient (CoRe) optimizer and\nincremental learning strategies relying on rehearsal of data, regularization of\nparameters, and the architecture of the model.\n", "date": "2023-03-10", "categories": []}
{"id": "2309.06086", "title": "Plasticity-Optimized Complementary Networks for Unsupervised Continual\n  Learning", "abstract": "  Continuous unsupervised representation learning (CURL) research has greatly\nbenefited from improvements in self-supervised learning (SSL) techniques. As a\nresult, existing CURL methods using SSL can learn high-quality representations\nwithout any labels, but with a notable performance drop when learning on a\nmany-tasks data stream. We hypothesize that this is caused by the\nregularization losses that are imposed to prevent forgetting, leading to a\nsuboptimal plasticity-stability trade-off: they either do not adapt fully to\nthe incoming data (low plasticity), or incur significant forgetting when\nallowed to fully adapt to a new SSL pretext-task (low stability). In this work,\nwe propose to train an expert network that is relieved of the duty of keeping\nthe previous knowledge and can focus on performing optimally on the new tasks\n(optimizing plasticity). In the second phase, we combine this new knowledge\nwith the previous network in an adaptation-retrospection phase to avoid\nforgetting and initialize a new expert with the knowledge of the old network.\nWe perform several experiments showing that our proposed approach outperforms\nother CURL exemplar-free methods in few- and many-task split settings.\nFurthermore, we show how to adapt our approach to semi-supervised continual\nlearning (Semi-SCL) and show that we surpass the accuracy of other\nexemplar-free Semi-SCL methods and reach the results of some others that use\nexemplars.\n", "date": "2023-09-12", "categories": []}
{"id": "2212.05102", "title": "A soft nearest-neighbor framework for continual semi-supervised learning", "abstract": "  Despite significant advances, the performance of state-of-the-art continual\nlearning approaches hinges on the unrealistic scenario of fully labeled data.\nIn this paper, we tackle this challenge and propose an approach for continual\nsemi-supervised learning--a setting where not all the data samples are labeled.\nA primary issue in this scenario is the model forgetting representations of\nunlabeled data and overfitting the labeled samples. We leverage the power of\nnearest-neighbor classifiers to nonlinearly partition the feature space and\nflexibly model the underlying data distribution thanks to its non-parametric\nnature. This enables the model to learn a strong representation for the current\ntask, and distill relevant information from previous tasks. We perform a\nthorough experimental evaluation and show that our method outperforms all the\nexisting approaches by large margins, setting a solid state of the art on the\ncontinual semi-supervised learning paradigm. For example, on CIFAR-100 we\nsurpass several others even when using at least 30 times less supervision (0.8%\nvs. 25% of annotations). Finally, our method works well on both low and high\nresolution images and scales seamlessly to more complex datasets such as\nImageNet-100. The code is publicly available on\nhttps://github.com/kangzhiq/NNCSL\n", "date": "2022-12-09", "categories": []}
{"id": "2303.06015", "title": "Dynamic Y-KD: A Hybrid Approach to Continual Instance Segmentation", "abstract": "  Despite the success of deep learning models on instance segmentation, current\nmethods still suffer from catastrophic forgetting in continual learning\nscenarios. In this paper, our contributions for continual instance segmentation\nare threefold. First, we propose the Y-knowledge distillation (Y-KD), a\ntechnique that shares a common feature extractor between the teacher and\nstudent networks. As the teacher is also updated with new data in Y-KD, the\nincreased plasticity results in new modules that are specialized on new\nclasses. Second, our Y-KD approach is supported by a dynamic architecture\nmethod that trains task-specific modules with a unique instance segmentation\nhead, thereby significantly reducing forgetting. Third, we complete our\napproach by leveraging checkpoint averaging as a simple method to manually\nbalance the trade-off between performance on the various sets of classes, thus\nincreasing control over the model's behavior without any additional cost. These\ncontributions are united in our model that we name the Dynamic Y-KD network.\n  We perform extensive experiments on several single-step and multi-steps\nincremental learning scenarios, and we show that our approach outperforms\nprevious methods both on past and new classes. For instance, compared to recent\nwork, our method obtains +2.1% mAP on old classes in 15-1, +7.6% mAP on new\nclasses in 19-1 and reaches 91.5% of the mAP obtained by joint-training on all\nclasses in 15-5.\n", "date": "2023-03-10", "categories": []}
{"id": "2304.06619", "title": "Class-Incremental Learning of Plant and Disease Detection: Growing\n  Branches with Knowledge Distillation", "abstract": "  This paper investigates the problem of class-incremental object detection for\nagricultural applications where a model needs to learn new plant species and\ndiseases incrementally without forgetting the previously learned ones. We adapt\ntwo public datasets to include new categories over time, simulating a more\nrealistic and dynamic scenario. We then compare three class-incremental\nlearning methods that leverage different forms of knowledge distillation to\nmitigate catastrophic forgetting. Our experiments show that all three methods\nsuffer from catastrophic forgetting, but the Dynamic Y-KD approach, which\nadditionally uses a dynamic architecture that grows new branches to learn new\ntasks, outperforms ILOD and Faster-ILOD in most settings both on new and old\nclasses.\n  These results highlight the challenges and opportunities of continual object\ndetection for agricultural applications. In particular, we hypothesize that the\nlarge intra-class and small inter-class variability that is typical of plant\nimages exacerbate the difficulty of learning new categories without interfering\nwith previous knowledge. We publicly release our code to encourage future work.\n", "date": "2023-04-13", "categories": []}
{"id": "2309.05281", "title": "Class-Incremental Grouping Network for Continual Audio-Visual Learning", "abstract": "  Continual learning is a challenging problem in which models need to be\ntrained on non-stationary data across sequential tasks for class-incremental\nlearning. While previous methods have focused on using either regularization or\nrehearsal-based frameworks to alleviate catastrophic forgetting in image\nclassification, they are limited to a single modality and cannot learn compact\nclass-aware cross-modal representations for continual audio-visual learning. To\naddress this gap, we propose a novel class-incremental grouping network (CIGN)\nthat can learn category-wise semantic features to achieve continual\naudio-visual learning. Our CIGN leverages learnable audio-visual class tokens\nand audio-visual grouping to continually aggregate class-aware features.\nAdditionally, it utilizes class tokens distillation and continual grouping to\nprevent forgetting parameters learned from previous tasks, thereby improving\nthe model's ability to capture discriminative audio-visual categories. We\nconduct extensive experiments on VGGSound-Instruments, VGGSound-100, and\nVGG-Sound Sources benchmarks. Our experimental results demonstrate that the\nCIGN achieves state-of-the-art audio-visual class-incremental learning\nperformance. Code is available at https://github.com/stoneMo/CIGN.\n", "date": "2023-09-11", "categories": []}
{"id": "2309.04081", "title": "UER: A Heuristic Bias Addressing Approach for Online Continual Learning", "abstract": "  Online continual learning aims to continuously train neural networks from a\ncontinuous data stream with a single pass-through data. As the most effective\napproach, the rehearsal-based methods replay part of previous data. Commonly\nused predictors in existing methods tend to generate biased dot-product logits\nthat prefer to the classes of current data, which is known as a bias issue and\na phenomenon of forgetting. Many approaches have been proposed to overcome the\nforgetting problem by correcting the bias; however, they still need to be\nimproved in online fashion. In this paper, we try to address the bias issue by\na more straightforward and more efficient method. By decomposing the\ndot-product logits into an angle factor and a norm factor, we empirically find\nthat the bias problem mainly occurs in the angle factor, which can be used to\nlearn novel knowledge as cosine logits. On the contrary, the norm factor\nabandoned by existing methods helps remember historical knowledge. Based on\nthis observation, we intuitively propose to leverage the norm factor to balance\nthe new and old knowledge for addressing the bias. To this end, we develop a\nheuristic approach called unbias experience replay (UER). UER learns current\nsamples only by the angle factor and further replays previous samples by both\nthe norm and angle factors. Extensive experiments on three datasets show that\nUER achieves superior performance over various state-of-the-art methods. The\ncode is in https://github.com/FelixHuiweiLin/UER.\n", "date": "2023-09-08", "categories": []}
{"id": "2308.04014", "title": "Continual Pre-Training of Large Language Models: How to (re)warm your\n  model?", "abstract": "  Large language models (LLMs) are routinely pre-trained on billions of tokens,\nonly to restart the process over again once new data becomes available. A much\ncheaper and more efficient solution would be to enable the continual\npre-training of these models, i.e. updating pre-trained models with new data\ninstead of re-training them from scratch. However, the distribution shift\ninduced by novel data typically results in degraded performance on past data.\nTaking a step towards efficient continual pre-training, in this work, we\nexamine the effect of different warm-up strategies. Our hypothesis is that the\nlearning rate must be re-increased to improve compute efficiency when training\non a new dataset. We study the warmup phase of models pre-trained on the Pile\n(upstream data, 300B tokens) as we continue to pre-train on SlimPajama\n(downstream data, 297B tokens), following a linear warmup and cosine decay\nschedule. We conduct all experiments on the Pythia 410M language model\narchitecture and evaluate performance through validation perplexity. We\nexperiment with different pre-training checkpoints, various maximum learning\nrates, and various warmup lengths. Our results show that while rewarming models\nfirst increases the loss on upstream and downstream data, in the longer run it\nimproves the downstream performance, outperforming models trained from\nscratch$\\unicode{x2013}$even for a large downstream dataset.\n", "date": "2023-08-08", "categories": []}
{"id": "2307.04869", "title": "Fed-CPrompt: Contrastive Prompt for Rehearsal-Free Federated Continual\n  Learning", "abstract": "  Federated continual learning (FCL) learns incremental tasks over time from\nconfidential datasets distributed across clients. This paper focuses on\nrehearsal-free FCL, which has severe forgetting issues when learning new tasks\ndue to the lack of access to historical task data. To address this issue, we\npropose Fed-CPrompt based on prompt learning techniques to obtain task-specific\nprompts in a communication-efficient way. Fed-CPrompt introduces two key\ncomponents, asynchronous prompt learning, and contrastive continual loss, to\nhandle asynchronous task arrival and heterogeneous data distributions in FCL,\nrespectively. Extensive experiments demonstrate the effectiveness of\nFed-CPrompt in achieving SOTA rehearsal-free FCL performance.\n", "date": "2023-07-10", "categories": []}
{"id": "2309.02551", "title": "Continual Improvement of Threshold-Based Novelty Detection", "abstract": "  When evaluated in dynamic, open-world situations, neural networks struggle to\ndetect unseen classes. This issue complicates the deployment of continual\nlearners in realistic environments where agents are not explicitly informed\nwhen novel categories are encountered. A common family of techniques for\ndetecting novelty relies on thresholds of similarity between observed data\npoints and the data used for training. However, these methods often require\nmanually specifying (ahead of time) the value of these thresholds, and are\ntherefore incapable of adapting to the nature of the data. We propose a new\nmethod for automatically selecting these thresholds utilizing a linear search\nand leave-one-out cross-validation on the ID classes. We demonstrate that this\nnovel method for selecting thresholds results in improved total accuracy on\nMNIST, Fashion MNIST, and CIFAR-10.\n", "date": "2023-09-05", "categories": []}
{"id": "2309.02995", "title": "Continual Evidential Deep Learning for Out-of-Distribution Detection", "abstract": "  Uncertainty-based deep learning models have attracted a great deal of\ninterest for their ability to provide accurate and reliable predictions.\nEvidential deep learning stands out achieving remarkable performance in\ndetecting out-of-distribution (OOD) data with a single deterministic neural\nnetwork. Motivated by this fact, in this paper we propose the integration of an\nevidential deep learning method into a continual learning framework in order to\nperform simultaneously incremental object classification and OOD detection.\nMoreover, we analyze the ability of vacuity and dissonance to differentiate\nbetween in-distribution data belonging to old classes and OOD data. The\nproposed method, called CEDL, is evaluated on CIFAR-100 considering two\nsettings consisting of 5 and 10 tasks, respectively. From the obtained results,\nwe could appreciate that the proposed method, in addition to provide comparable\nresults in object classification with respect to the baseline, largely\noutperforms OOD detection compared to several posthoc methods on three\nevaluation metrics: AUROC, AUPR and FPR95.\n", "date": "2023-09-06", "categories": []}
{"id": "2211.02947", "title": "Prototypical quadruplet for few-shot class incremental learning", "abstract": "  Scarcity of data and incremental learning of new tasks pose two major\nbottlenecks for many modern computer vision algorithms. The phenomenon of\ncatastrophic forgetting, i.e., the model's inability to classify previously\nlearned data after training with new batches of data, is a major challenge.\nConventional methods address catastrophic forgetting while compromising the\ncurrent session's training. Generative replay-based approaches, such as\ngenerative adversarial networks (GANs), have been proposed to mitigate\ncatastrophic forgetting, but training GANs with few samples may lead to\ninstability. To address these challenges, we propose a novel method that\nimproves classification robustness by identifying a better embedding space\nusing an improved contrasting loss. Our approach retains previously acquired\nknowledge in the embedding space, even when trained with new classes, by\nupdating previous session class prototypes to represent the true class mean,\nwhich is crucial for our nearest class mean classification strategy. We\ndemonstrate the effectiveness of our method by showing that the embedding space\nremains intact after training the model with new classes and outperforms\nexisting state-of-the-art algorithms in terms of accuracy across different\nsessions.\n", "date": "2022-11-05", "categories": []}
{"id": "2308.11917", "title": "LFS-GAN: Lifelong Few-Shot Image Generation", "abstract": "  We address a challenging lifelong few-shot image generation task for the\nfirst time. In this situation, a generative model learns a sequence of tasks\nusing only a few samples per task. Consequently, the learned model encounters\nboth catastrophic forgetting and overfitting problems at a time. Existing\nstudies on lifelong GANs have proposed modulation-based methods to prevent\ncatastrophic forgetting. However, they require considerable additional\nparameters and cannot generate high-fidelity and diverse images from limited\ndata. On the other hand, the existing few-shot GANs suffer from severe\ncatastrophic forgetting when learning multiple tasks. To alleviate these\nissues, we propose a framework called Lifelong Few-Shot GAN (LFS-GAN) that can\ngenerate high-quality and diverse images in lifelong few-shot image generation\ntask. Our proposed framework learns each task using an efficient task-specific\nmodulator - Learnable Factorized Tensor (LeFT). LeFT is rank-constrained and\nhas a rich representation ability due to its unique reconstruction technique.\nFurthermore, we propose a novel mode seeking loss to improve the diversity of\nour model in low-data circumstances. Extensive experiments demonstrate that the\nproposed LFS-GAN can generate high-fidelity and diverse images without any\nforgetting and mode collapse in various domains, achieving state-of-the-art in\nlifelong few-shot image generation task. Surprisingly, we find that our LFS-GAN\neven outperforms the existing few-shot GANs in the few-shot image generation\ntask. The code is available at Github.\n", "date": "2023-08-23", "categories": []}
{"id": "2309.00688", "title": "Jointly Exploring Client Drift and Catastrophic Forgetting in Dynamic\n  Learning", "abstract": "  Federated and Continual Learning have emerged as potential paradigms for the\nrobust and privacy-aware use of Deep Learning in dynamic environments. However,\nClient Drift and Catastrophic Forgetting are fundamental obstacles to\nguaranteeing consistent performance. Existing work only addresses these\nproblems separately, which neglects the fact that the root cause behind both\nforms of performance deterioration is connected. We propose a unified analysis\nframework for building a controlled test environment for Client Drift -- by\nperturbing a defined ratio of clients -- and Catastrophic Forgetting -- by\nshifting all clients with a particular strength. Our framework further\nleverages this new combined analysis by generating a 3D landscape of the\ncombined performance impact from both. We demonstrate that the performance drop\nthrough Client Drift, caused by a certain share of shifted clients, is\ncorrelated to the drop from Catastrophic Forgetting resulting from a\ncorresponding shift strength. Correlation tests between both problems for\nComputer Vision (CelebA) and Medical Imaging (PESO) support this new\nperspective, with an average Pearson rank correlation coefficient of over 0.94.\nOur framework's novel ability of combined spatio-temporal shift analysis allows\nus to investigate how both forms of distribution shift behave in mixed\nscenarios, opening a new pathway for better generalization. We show that a\ncombination of moderate Client Drift and Catastrophic Forgetting can even\nimprove the performance of the resulting model (causing a \"Generalization\nBump\") compared to when only one of the shifts occurs individually. We apply a\nsimple and commonly used method from Continual Learning in the federated\nsetting and observe this phenomenon to be reoccurring, leveraging the ability\nof our framework to analyze existing and novel methods for Federated and\nContinual Learning.\n", "date": "2023-09-01", "categories": []}
{"id": "2309.00862", "title": "Big-model Driven Few-shot Continual Learning", "abstract": "  Few-shot continual learning (FSCL) has attracted intensive attention and\nachieved some advances in recent years, but now it is difficult to again make a\nbig stride in accuracy due to the limitation of only few-shot incremental\nsamples. Inspired by distinctive human cognition ability in life learning, in\nthis work, we propose a novel Big-model driven Few-shot Continual Learning\n(B-FSCL) framework to gradually evolve the model under the traction of the\nworld's big-models (like human accumulative knowledge). Specifically, we\nperform the big-model driven transfer learning to leverage the powerful\nencoding capability of these existing big-models, which can adapt the continual\nmodel to a few of newly added samples while avoiding the over-fitting problem.\nConsidering that the big-model and the continual model may have different\nperceived results for the identical images, we introduce an instance-level\nadaptive decision mechanism to provide the high-level flexibility cognitive\nsupport adjusted to varying samples. In turn, the adaptive decision can be\nfurther adopted to optimize the parameters of the continual model, performing\nthe adaptive distillation of big-model's knowledge information. Experimental\nresults of our proposed B-FSCL on three popular datasets (including CIFAR100,\nminilmageNet and CUB200) completely surpass all state-of-the-art FSCL methods.\n", "date": "2023-09-02", "categories": []}
{"id": "2309.01050", "title": "Efficient Curriculum based Continual Learning with Informative Subset\n  Selection for Remote Sensing Scene Classification", "abstract": "  We tackle the problem of class incremental learning (CIL) in the realm of\nlandcover classification from optical remote sensing (RS) images in this paper.\nThe paradigm of CIL has recently gained much prominence given the fact that\ndata are generally obtained in a sequential manner for real-world phenomenon.\nHowever, CIL has not been extensively considered yet in the domain of RS\nirrespective of the fact that the satellites tend to discover new classes at\ndifferent geographical locations temporally. With this motivation, we propose a\nnovel CIL framework inspired by the recent success of replay-memory based\napproaches and tackling two of their shortcomings. In order to reduce the\neffect of catastrophic forgetting of the old classes when a new stream arrives,\nwe learn a curriculum of the new classes based on their similarity with the old\nclasses. This is found to limit the degree of forgetting substantially. Next\nwhile constructing the replay memory, instead of randomly selecting samples\nfrom the old streams, we propose a sample selection strategy which ensures the\nselection of highly confident samples so as to reduce the effects of noise. We\nobserve a sharp improvement in the CIL performance with the proposed\ncomponents. Experimental results on the benchmark NWPU-RESISC45, PatternNet,\nand EuroSAT datasets confirm that our method offers improved\nstability-plasticity trade-off than the literature.\n", "date": "2023-09-03", "categories": []}
{"id": "2309.02210", "title": "Continual Cross-Dataset Adaptation in Road Surface Classification", "abstract": "  Accurate road surface classification is crucial for autonomous vehicles (AVs)\nto optimize driving conditions, enhance safety, and enable advanced road\nmapping. However, deep learning models for road surface classification suffer\nfrom poor generalization when tested on unseen datasets. To update these models\nwith new information, also the original training dataset must be taken into\naccount, in order to avoid catastrophic forgetting. This is, however,\ninefficient if not impossible, e.g., when the data is collected in streams or\nlarge amounts. To overcome this limitation and enable fast and efficient\ncross-dataset adaptation, we propose to employ continual learning finetuning\nmethods designed to retain past knowledge while adapting to new data, thus\neffectively avoiding forgetting. Experimental results demonstrate the\nsuperiority of this approach over naive finetuning, achieving performance close\nto fresh retraining. While solving this known problem, we also provide a\ngeneral description of how the same technique can be adopted in other AV\nscenarios. We highlight the potential computational and economic benefits that\na continual-based adaptation can bring to the AV industry, while also reducing\ngreenhouse emissions due to unnecessary joint retraining.\n", "date": "2023-09-05", "categories": []}
{"id": "2309.00462", "title": "New metrics for analyzing continual learners", "abstract": "  Deep neural networks have shown remarkable performance when trained on\nindependent and identically distributed data from a fixed set of classes.\nHowever, in real-world scenarios, it can be desirable to train models on a\ncontinuous stream of data where multiple classification tasks are presented\nsequentially. This scenario, known as Continual Learning (CL) poses challenges\nto standard learning algorithms which struggle to maintain knowledge of old\ntasks while learning new ones. This stability-plasticity dilemma remains\ncentral to CL and multiple metrics have been proposed to adequately measure\nstability and plasticity separately. However, none considers the increasing\ndifficulty of the classification task, which inherently results in performance\nloss for any model. In that sense, we analyze some limitations of current\nmetrics and identify the presence of setup-induced forgetting. Therefore, we\npropose new metrics that account for the task's increasing difficulty. Through\nexperiments on benchmark datasets, we demonstrate that our proposed metrics can\nprovide new insights into the stability-plasticity trade-off achieved by models\nin the continual learning environment.\n", "date": "2023-09-01", "categories": []}
{"id": "2308.16567", "title": "ScrollNet: Dynamic Weight Importance for Continual Learning", "abstract": "  The principle underlying most existing continual learning (CL) methods is to\nprioritize stability by penalizing changes in parameters crucial to old tasks,\nwhile allowing for plasticity in other parameters. The importance of weights\nfor each task can be determined either explicitly through learning a\ntask-specific mask during training (e.g., parameter isolation-based approaches)\nor implicitly by introducing a regularization term (e.g., regularization-based\napproaches). However, all these methods assume that the importance of weights\nfor each task is unknown prior to data exposure. In this paper, we propose\nScrollNet as a scrolling neural network for continual learning. ScrollNet can\nbe seen as a dynamic network that assigns the ranking of weight importance for\neach task before data exposure, thus achieving a more favorable\nstability-plasticity tradeoff during sequential task learning by reassigning\nthis ranking for different tasks. Additionally, we demonstrate that ScrollNet\ncan be combined with various CL methods, including regularization-based and\nreplay-based approaches. Experimental results on CIFAR100 and TinyImagenet\ndatasets show the effectiveness of our proposed method. We release our code at\nhttps://github.com/FireFYF/ScrollNet.git.\n", "date": "2023-08-31", "categories": []}
{"id": "2308.15827", "title": "Introducing Language Guidance in Prompt-based Continual Learning", "abstract": "  Continual Learning aims to learn a single model on a sequence of tasks\nwithout having access to data from previous tasks. The biggest challenge in the\ndomain still remains catastrophic forgetting: a loss in performance on seen\nclasses of earlier tasks. Some existing methods rely on an expensive replay\nbuffer to store a chunk of data from previous tasks. This, while promising,\nbecomes expensive when the number of tasks becomes large or data can not be\nstored for privacy reasons. As an alternative, prompt-based methods have been\nproposed that store the task information in a learnable prompt pool. This\nprompt pool instructs a frozen image encoder on how to solve each task. While\nthe model faces a disjoint set of classes in each task in this setting, we\nargue that these classes can be encoded to the same embedding space of a\npre-trained language encoder. In this work, we propose Language Guidance for\nPrompt-based Continual Learning (LGCL) as a plug-in for prompt-based methods.\nLGCL is model agnostic and introduces language guidance at the task level in\nthe prompt pool and at the class level on the output feature of the vision\nencoder. We show with extensive experimentation that LGCL consistently improves\nthe performance of prompt-based continual learning methods to set a new\nstate-of-the art. LGCL achieves these performance improvements without needing\nany additional learnable parameters.\n", "date": "2023-08-30", "categories": []}
{"id": "2308.15308", "title": "On-Device Learning with Binary Neural Networks", "abstract": "  Existing Continual Learning (CL) solutions only partially address the\nconstraints on power, memory and computation of the deep learning models when\ndeployed on low-power embedded CPUs. In this paper, we propose a CL solution\nthat embraces the recent advancements in CL field and the efficiency of the\nBinary Neural Networks (BNN), that use 1-bit for weights and activations to\nefficiently execute deep learning models. We propose a hybrid quantization of\nCWR* (an effective CL approach) that considers differently forward and backward\npass in order to retain more precision during gradient update step and at the\nsame time minimizing the latency overhead. The choice of a binary network as\nbackbone is essential to meet the constraints of low power devices and, to the\nbest of authors' knowledge, this is the first attempt to prove on-device\nlearning with BNN. The experimental validation carried out confirms the\nvalidity and the suitability of the proposed method.\n", "date": "2023-08-29", "categories": []}
{"id": "2209.06767", "title": "Parameter-Efficient Finetuning for Robust Continual Multilingual\n  Learning", "abstract": "  We introduce and study the problem of Continual Multilingual Learning (CML)\nwhere a previously trained multilingual model is periodically updated using new\ndata arriving in stages. If the new data is present only in a subset of\nlanguages, we find that the resulting model shows improved performance only on\nthe languages included in the latest update (and a few closely related\nlanguages) while its performance on all the remaining languages degrade\nsignificantly. We address this challenge by proposing LAFT-URIEL, a\nparameter-efficient finetuning strategy which aims to increase the number of\nlanguages on which the model improves after an update, while reducing the\nmagnitude of loss in performance for the remaining languages. LAFT-URIEL uses\nlinguistic knowledge to balance overfitting and knowledge sharing across\nlanguages, allowing for an additional 25% of task languages to see an\nimprovement in performance after an update, while also reducing the average\nmagnitude of losses on the remaining languages by 78% relative.\n", "date": "2022-09-14", "categories": []}
{"id": "2308.13982", "title": "Universal Graph Continual Learning", "abstract": "  We address catastrophic forgetting issues in graph learning as incoming data\ntransits from one to another graph distribution. Whereas prior studies\nprimarily tackle one setting of graph continual learning such as incremental\nnode classification, we focus on a universal approach wherein each data point\nin a task can be a node or a graph, and the task varies from node to graph\nclassification. We propose a novel method that enables graph neural networks to\nexcel in this universal setting. Our approach perseveres knowledge about past\ntasks through a rehearsal mechanism that maintains local and global structure\nconsistency across the graphs. We benchmark our method against various\ncontinual learning baselines in real-world graph datasets and achieve\nsignificant improvement in average performance and forgetting across tasks.\n", "date": "2023-08-27", "categories": []}
{"id": "2308.14100", "title": "Rethinking Exemplars for Continual Semantic Segmentation in Endoscopy\n  Scenes: Entropy-based Mini-Batch Pseudo-Replay", "abstract": "  Endoscopy is a widely used technique for the early detection of diseases or\nrobotic-assisted minimally invasive surgery (RMIS). Numerous deep learning\n(DL)-based research works have been developed for automated diagnosis or\nprocessing of endoscopic view. However, existing DL models may suffer from\ncatastrophic forgetting. When new target classes are introduced over time or\ncross institutions, the performance of old classes may suffer severe\ndegradation. More seriously, data privacy and storage issues may lead to the\nunavailability of old data when updating the model. Therefore, it is necessary\nto develop a continual learning (CL) methodology to solve the problem of\ncatastrophic forgetting in endoscopic image segmentation. To tackle this, we\npropose a Endoscopy Continual Semantic Segmentation (EndoCSS) framework that\ndoes not involve the storage and privacy issues of exemplar data. The framework\nincludes a mini-batch pseudo-replay (MB-PR) mechanism and a self-adaptive noisy\ncross-entropy (SAN-CE) loss. The MB-PR strategy circumvents privacy and storage\nissues by generating pseudo-replay images through a generative model.\nMeanwhile, the MB-PR strategy can also correct the model deviation to the\nreplay data and current training data, which is aroused by the significant\ndifference in the amount of current and replay images. Therefore, the model can\nperform effective representation learning on both new and old tasks. SAN-CE\nloss can help model fitting by adjusting the model's output logits, and also\nimprove the robustness of training. Extensive continual semantic segmentation\n(CSS) experiments on public datasets demonstrate that our method can robustly\nand effectively address the catastrophic forgetting brought by class increment\nin endoscopy scenes. The results show that our framework holds excellent\npotential for real-world deployment in a streaming learning manner.\n", "date": "2023-08-27", "categories": []}
{"id": "2308.14374", "title": "Online Continual Learning on Hierarchical Label Expansion", "abstract": "  Continual learning (CL) enables models to adapt to new tasks and environments\nwithout forgetting previously learned knowledge. While current CL setups have\nignored the relationship between labels in the past task and the new task with\nor without small task overlaps, real-world scenarios often involve hierarchical\nrelationships between old and new tasks, posing another challenge for\ntraditional CL approaches. To address this challenge, we propose a novel\nmulti-level hierarchical class incremental task configuration with an online\nlearning constraint, called hierarchical label expansion (HLE). Our\nconfiguration allows a network to first learn coarse-grained classes, with data\nlabels continually expanding to more fine-grained classes in various hierarchy\ndepths. To tackle this new setup, we propose a rehearsal-based method that\nutilizes hierarchy-aware pseudo-labeling to incorporate hierarchical class\ninformation. Additionally, we propose a simple yet effective memory management\nand sampling strategy that selectively adopts samples of newly encountered\nclasses. Our experiments demonstrate that our proposed method can effectively\nuse hierarchy on our HLE setup to improve classification accuracy across all\nlevels of hierarchies, regardless of depth and class imbalance ratio,\noutperforming prior state-of-the-art works by significant margins while also\noutperforming them on the conventional disjoint, blurry and i-Blurry CL setups.\n", "date": "2023-08-28", "categories": []}
{"id": "2304.07707", "title": "Non-exemplar Class-incremental Learning by Random Auxiliary Classes\n  Augmentation and Mixed Features", "abstract": "  Non-exemplar class-incremental learning refers to classifying new and old\nclasses without storing samples of old classes. Since only new class samples\nare available for optimization, it often occurs catastrophic forgetting of old\nknowledge. To alleviate this problem, many new methods are proposed such as\nmodel distillation, class augmentation. In this paper, we propose an effective\nnon-exemplar method called RAMF consisting of Random Auxiliary classes\naugmentation and Mixed Feature. On the one hand, we design a novel random\nauxiliary classes augmentation method, where one augmentation is randomly\nselected from three augmentations and applied on the input to generate\naugmented samples and extra class labels. By extending data and label space, it\nallows the model to learn more diverse representations, which can prevent the\nmodel from being biased towards learning task-specific features. When learning\nnew tasks, it will reduce the change of feature space and improve model\ngeneralization. On the other hand, we employ mixed feature to replace the new\nfeatures since only using new feature to optimize the model will affect the\nrepresentation that was previously embedded in the feature space. Instead, by\nmixing new and old features, old knowledge can be retained without increasing\nthe computational complexity. Extensive experiments on three benchmarks\ndemonstrate the superiority of our approach, which outperforms the\nstate-of-the-art non-exemplar methods and is comparable to high-performance\nreplay-based methods.\n", "date": "2023-04-16", "categories": []}
{"id": "2308.13305", "title": "Dynamic Residual Classifier for Class Incremental Learning", "abstract": "  The rehearsal strategy is widely used to alleviate the catastrophic\nforgetting problem in class incremental learning (CIL) by preserving limited\nexemplars from previous tasks. With imbalanced sample numbers between old and\nnew classes, the classifier learning can be biased. Existing CIL methods\nexploit the long-tailed (LT) recognition techniques, e.g., the adjusted losses\nand the data re-sampling methods, to handle the data imbalance issue within\neach increment task. In this work, the dynamic nature of data imbalance in CIL\nis shown and a novel Dynamic Residual Classifier (DRC) is proposed to handle\nthis challenging scenario. Specifically, DRC is built upon a recent advance\nresidual classifier with the branch layer merging to handle the model-growing\nproblem. Moreover, DRC is compatible with different CIL pipelines and\nsubstantially improves them. Combining DRC with the model adaptation and fusion\n(MAF) pipeline, this method achieves state-of-the-art results on both the\nconventional CIL and the LT-CIL benchmarks. Extensive experiments are also\nconducted for a detailed analysis. The code is publicly available.\n", "date": "2023-08-25", "categories": []}
{"id": "2308.03374", "title": "Heterogeneous Forgetting Compensation for Class-Incremental Learning", "abstract": "  Class-incremental learning (CIL) has achieved remarkable successes in\nlearning new classes consecutively while overcoming catastrophic forgetting on\nold categories. However, most existing CIL methods unreasonably assume that all\nold categories have the same forgetting pace, and neglect negative influence of\nforgetting heterogeneity among different old classes on forgetting\ncompensation. To surmount the above challenges, we develop a novel\nHeterogeneous Forgetting Compensation (HFC) model, which can resolve\nheterogeneous forgetting of easy-to-forget and hard-to-forget old categories\nfrom both representation and gradient aspects. Specifically, we design a\ntask-semantic aggregation block to alleviate heterogeneous forgetting from\nrepresentation aspect. It aggregates local category information within each\ntask to learn task-shared global representations. Moreover, we develop two\nnovel plug-and-play losses: a gradient-balanced forgetting compensation loss\nand a gradient-balanced relation distillation loss to alleviate forgetting from\ngradient aspect. They consider gradient-balanced compensation to rectify\nforgetting heterogeneity of old categories and heterogeneous relation\nconsistency. Experiments on several representative datasets illustrate\neffectiveness of our HFC model. The code is available at\nhttps://github.com/JiahuaDong/HFC.\n", "date": "2023-08-07", "categories": []}
{"id": "2308.12366", "title": "Continual Zero-Shot Learning through Semantically Guided Generative\n  Random Walks", "abstract": "  Learning novel concepts, remembering previous knowledge, and adapting it to\nfuture tasks occur simultaneously throughout a human's lifetime. To model such\ncomprehensive abilities, continual zero-shot learning (CZSL) has recently been\nintroduced. However, most existing methods overused unseen semantic information\nthat may not be continually accessible in realistic settings. In this paper, we\naddress the challenge of continual zero-shot learning where unseen information\nis not provided during training, by leveraging generative modeling. The heart\nof the generative-based methods is to learn quality representations from seen\nclasses to improve the generative understanding of the unseen visual space.\nMotivated by this, we introduce generalization-bound tools and provide the\nfirst theoretical explanation for the benefits of generative modeling to CZSL\ntasks. Guided by the theoretical analysis, we then propose our learning\nalgorithm that employs a novel semantically guided Generative Random Walk (GRW)\nloss. The GRW loss augments the training by continually encouraging the model\nto generate realistic and characterized samples to represent the unseen space.\nOur algorithm achieves state-of-the-art performance on AWA1, AWA2, CUB, and SUN\ndatasets, surpassing existing CZSL methods by 3-7\\%. The code has been made\navailable here \\url{https://github.com/wx-zhang/IGCZSL}\n", "date": "2023-08-23", "categories": []}
{"id": "2308.12510", "title": "Masked Autoencoders are Efficient Class Incremental Learners", "abstract": "  Class Incremental Learning (CIL) aims to sequentially learn new classes while\navoiding catastrophic forgetting of previous knowledge. We propose to use\nMasked Autoencoders (MAEs) as efficient learners for CIL. MAEs were originally\ndesigned to learn useful representations through reconstructive unsupervised\nlearning, and they can be easily integrated with a supervised loss for\nclassification. Moreover, MAEs can reliably reconstruct original input images\nfrom randomly selected patches, which we use to store exemplars from past tasks\nmore efficiently for CIL. We also propose a bilateral MAE framework to learn\nfrom image-level and embedding-level fusion, which produces better-quality\nreconstructed images and more stable representations. Our experiments confirm\nthat our approach performs better than the state-of-the-art on CIFAR-100,\nImageNet-Subset, and ImageNet-Full. The code is available at\nhttps://github.com/scok30/MAE-CIL .\n", "date": "2023-08-24", "categories": []}
{"id": "2308.12679", "title": "A Continual Learning Approach for Cross-Domain White Blood Cell\n  Classification", "abstract": "  Accurate classification of white blood cells in peripheral blood is essential\nfor diagnosing hematological diseases. Due to constantly evolving clinical\nsettings, data sources, and disease classifications, it is necessary to update\nmachine learning classification models regularly for practical real-world use.\nSuch models significantly benefit from sequentially learning from incoming data\nstreams without forgetting previously acquired knowledge. However, models can\nsuffer from catastrophic forgetting, causing a drop in performance on previous\ntasks when fine-tuned on new data. Here, we propose a rehearsal-based continual\nlearning approach for class incremental and domain incremental scenarios in\nwhite blood cell classification. To choose representative samples from previous\ntasks, we employ exemplar set selection based on the model's predictions. This\ninvolves selecting the most confident samples and the most challenging samples\nidentified through uncertainty estimation of the model. We thoroughly evaluated\nour proposed approach on three white blood cell classification datasets that\ndiffer in color, resolution, and class composition, including scenarios where\nnew domains or new classes are introduced to the model with every task. We also\ntest a long class incremental experiment with both new domains and new classes.\nOur results demonstrate that our approach outperforms established baselines in\ncontinual learning, including existing iCaRL and EWC methods for classifying\nwhite blood cells in cross-domain environments.\n", "date": "2023-08-24", "categories": []}
{"id": "2206.00309", "title": "Label-Efficient Online Continual Object Detection in Streaming Video", "abstract": "  Humans can watch a continuous video stream and effortlessly perform continual\nacquisition and transfer of new knowledge with minimal supervision yet\nretaining previously learnt experiences. In contrast, existing continual\nlearning (CL) methods require fully annotated labels to effectively learn from\nindividual frames in a video stream. Here, we examine a more realistic and\nchallenging problem$\\unicode{x2014}$Label-Efficient Online Continual Object\nDetection (LEOCOD) in streaming video. We propose a plug-and-play module,\nEfficient-CLS, that can be easily inserted into and improve existing continual\nlearners for object detection in video streams with reduced data annotation\ncosts and model retraining time. We show that our method has achieved\nsignificant improvement with minimal forgetting across all supervision levels\non two challenging CL benchmarks for streaming real-world videos. Remarkably,\nwith only 25% annotated video frames, our method still outperforms the base CL\nlearners, which are trained with 100% annotations on all video frames. The data\nand source code will be publicly available at\nhttps://github.com/showlab/Efficient-CLS.\n", "date": "2022-06-01", "categories": []}
{"id": "2306.01792", "title": "Task Relation-aware Continual User Representation Learning", "abstract": "  User modeling, which learns to represent users into a low-dimensional\nrepresentation space based on their past behaviors, got a surge of interest\nfrom the industry for providing personalized services to users. Previous\nefforts in user modeling mainly focus on learning a task-specific user\nrepresentation that is designed for a single task. However, since learning\ntask-specific user representations for every task is infeasible, recent studies\nintroduce the concept of universal user representation, which is a more\ngeneralized representation of a user that is relevant to a variety of tasks.\nDespite their effectiveness, existing approaches for learning universal user\nrepresentations are impractical in real-world applications due to the data\nrequirement, catastrophic forgetting and the limited learning capability for\ncontinually added tasks. In this paper, we propose a novel continual user\nrepresentation learning method, called TERACON, whose learning capability is\nnot limited as the number of learned tasks increases while capturing the\nrelationship between the tasks. The main idea is to introduce an embedding for\neach task, i.e., task embedding, which is utilized to generate task-specific\nsoft masks that not only allow the entire model parameters to be updated until\nthe end of training sequence, but also facilitate the relationship between the\ntasks to be captured. Moreover, we introduce a novel knowledge retention module\nwith pseudo-labeling strategy that successfully alleviates the long-standing\nproblem of continual learning, i.e., catastrophic forgetting. Extensive\nexperiments on public and proprietary real-world datasets demonstrate the\nsuperiority and practicality of TERACON. Our code is available at\nhttps://github.com/Sein-Kim/TERACON.\n", "date": "2023-06-01", "categories": []}
{"id": "2308.11801", "title": "Variational Density Propagation Continual Learning", "abstract": "  Deep Neural Networks (DNNs) deployed to the real world are regularly subject\nto out-of-distribution (OoD) data, various types of noise, and shifting\nconceptual objectives. This paper proposes a framework for adapting to data\ndistribution drift modeled by benchmark Continual Learning datasets. We develop\nand evaluate a method of Continual Learning that leverages uncertainty\nquantification from Bayesian Inference to mitigate catastrophic forgetting. We\nexpand on previous approaches by removing the need for Monte Carlo sampling of\nthe model weights to sample the predictive distribution. We optimize a\nclosed-form Evidence Lower Bound (ELBO) objective approximating the predictive\ndistribution by propagating the first two moments of a distribution, i.e. mean\nand covariance, through all network layers. Catastrophic forgetting is\nmitigated by using the closed-form ELBO to approximate the Minimum Description\nLength (MDL) Principle, inherently penalizing changes in the model likelihood\nby minimizing the KL Divergence between the variational posterior for the\ncurrent task and the previous task's variational posterior acting as the prior.\nLeveraging the approximation of the MDL principle, we aim to initially learn a\nsparse variational posterior and then minimize additional model complexity\nlearned for subsequent tasks. Our approach is evaluated for the task\nincremental learning scenario using density propagated versions of\nfully-connected and convolutional neural networks across multiple sequential\nbenchmark datasets with varying task sequence lengths. Ultimately, this\nprocedure produces a minimally complex network over a series of tasks\nmitigating catastrophic forgetting.\n", "date": "2023-08-22", "categories": []}
{"id": "2308.11357", "title": "Exemplar-Free Continual Transformer with Convolutions", "abstract": "  Continual Learning (CL) involves training a machine learning model in a\nsequential manner to learn new information while retaining previously learned\ntasks without the presence of previous training data. Although there has been\nsignificant interest in CL, most recent CL approaches in computer vision have\nfocused on convolutional architectures only. However, with the recent success\nof vision transformers, there is a need to explore their potential for CL.\nAlthough there have been some recent CL approaches for vision transformers,\nthey either store training instances of previous tasks or require a task\nidentifier during test time, which can be limiting. This paper proposes a new\nexemplar-free approach for class/task incremental learning called ConTraCon,\nwhich does not require task-id to be explicitly present during inference and\navoids the need for storing previous training instances. The proposed approach\nleverages the transformer architecture and involves re-weighting the key,\nquery, and value weights of the multi-head self-attention layers of a\ntransformer trained on a similar task. The re-weighting is done using\nconvolution, which enables the approach to maintain low parameter requirements\nper task. Additionally, an image augmentation-based entropic task\nidentification approach is used to predict tasks without requiring task-ids\nduring inference. Experiments on four benchmark datasets demonstrate that the\nproposed approach outperforms several competitive approaches while requiring\nfewer parameters.\n", "date": "2023-08-22", "categories": []}
{"id": "2211.14963", "title": "Neural Architecture for Online Ensemble Continual Learning", "abstract": "  Continual learning with an increasing number of classes is a challenging\ntask. The difficulty rises when each example is presented exactly once, which\nrequires the model to learn online. Recent methods with classic parameter\noptimization procedures have been shown to struggle in such setups or have\nlimitations like non-differentiable components or memory buffers. For this\nreason, we present the fully differentiable ensemble method that allows us to\nefficiently train an ensemble of neural networks in the end-to-end regime. The\nproposed technique achieves SOTA results without a memory buffer and clearly\noutperforms the reference methods. The conducted experiments have also shown a\nsignificant increase in the performance for small ensembles, which demonstrates\nthe capability of obtaining relatively high classification accuracy with a\nreduced number of classifiers.\n", "date": "2022-11-27", "categories": []}
{"id": "2303.10070", "title": "A Unified Continual Learning Framework with General Parameter-Efficient\n  Tuning", "abstract": "  The \"pre-training $\\rightarrow$ downstream adaptation\" presents both new\nopportunities and challenges for Continual Learning (CL). Although the recent\nstate-of-the-art in CL is achieved through Parameter-Efficient-Tuning (PET)\nadaptation paradigm, only prompt has been explored, limiting its application to\nTransformers only. In this paper, we position prompting as one instantiation of\nPET, and propose a unified CL framework with general PET, dubbed as\nLearning-Accumulation-Ensemble (LAE). PET, e.g., using Adapter, LoRA, or\nPrefix, can adapt a pre-trained model to downstream tasks with fewer parameters\nand resources. Given a PET method, our LAE framework incorporates it for CL\nwith three novel designs. 1) Learning: the pre-trained model adapts to the new\ntask by tuning an online PET module, along with our adaptation speed\ncalibration to align different PET modules, 2) Accumulation: the task-specific\nknowledge learned by the online PET module is accumulated into an offline PET\nmodule through momentum update, 3) Ensemble: During inference, we respectively\nconstruct two experts with online/offline PET modules (which are favored by the\nnovel/historical tasks) for prediction ensemble. We show that LAE is compatible\nwith a battery of PET methods and gains strong CL capability. For example, LAE\nwith Adaptor PET surpasses the prior state-of-the-art by 1.3% and 3.6% in\nlast-incremental accuracy on CIFAR100 and ImageNet-R datasets, respectively.\nCode is available at \\url{https://github.com/gqk/LAE}.\n", "date": "2023-03-17", "categories": []}
{"id": "2308.10445", "title": "When Prompt-based Incremental Learning Does Not Meet Strong Pretraining", "abstract": "  Incremental learning aims to overcome catastrophic forgetting when learning\ndeep networks from sequential tasks. With impressive learning efficiency and\nperformance, prompt-based methods adopt a fixed backbone to sequential tasks by\nlearning task-specific prompts. However, existing prompt-based methods heavily\nrely on strong pretraining (typically trained on ImageNet-21k), and we find\nthat their models could be trapped if the potential gap between the pretraining\ntask and unknown future tasks is large. In this work, we develop a learnable\nAdaptive Prompt Generator (APG). The key is to unify the prompt retrieval and\nprompt learning processes into a learnable prompt generator. Hence, the whole\nprompting process can be optimized to reduce the negative effects of the gap\nbetween tasks effectively. To make our APG avoid learning ineffective\nknowledge, we maintain a knowledge pool to regularize APG with the feature\ndistribution of each class. Extensive experiments show that our method\nsignificantly outperforms advanced methods in exemplar-free incremental\nlearning without (strong) pretraining. Besides, under strong retraining, our\nmethod also has comparable performance to existing prompt-based models, showing\nthat our method can still benefit from pretraining. Codes can be found at\nhttps://github.com/TOM-tym/APG\n", "date": "2023-08-21", "categories": []}
{"id": "2308.10846", "title": "Real World Time Series Benchmark Datasets with Distribution Shifts:\n  Global Crude Oil Price and Volatility", "abstract": "  The scarcity of task-labeled time-series benchmarks in the financial domain\nhinders progress in continual learning. Addressing this deficit would foster\ninnovation in this area. Therefore, we present COB, Crude Oil Benchmark\ndatasets. COB includes 30 years of asset prices that exhibit significant\ndistribution shifts and optimally generates corresponding task (i.e., regime)\nlabels based on these distribution shifts for the three most important crude\noils in the world. Our contributions include creating real-world benchmark\ndatasets by transforming asset price data into volatility proxies, fitting\nmodels using expectation-maximization (EM), generating contextual task labels\nthat align with real-world events, and providing these labels as well as the\ngeneral algorithm to the public. We show that the inclusion of these task\nlabels universally improves performance on four continual learning algorithms,\nsome state-of-the-art, over multiple forecasting horizons. We hope these\nbenchmarks accelerate research in handling distribution shifts in real-world\ndata, especially due to the global importance of the assets considered. We've\nmade the (1) raw price data, (2) task labels generated by our approach, (3) and\ncode for our algorithm available at https://oilpricebenchmarks.github.io.\n", "date": "2023-08-21", "categories": []}
{"id": "2211.15470", "title": "Learning to Learn: How to Continuously Teach Humans and Machines", "abstract": "  Curriculum design is a fundamental component of education. For example, when\nwe learn mathematics at school, we build upon our knowledge of addition to\nlearn multiplication. These and other concepts must be mastered before our\nfirst algebra lesson, which also reinforces our addition and multiplication\nskills. Designing a curriculum for teaching either a human or a machine shares\nthe underlying goal of maximizing knowledge transfer from earlier to later\ntasks, while also minimizing forgetting of learned tasks. Prior research on\ncurriculum design for image classification focuses on the ordering of training\nexamples during a single offline task. Here, we investigate the effect of the\norder in which multiple distinct tasks are learned in a sequence. We focus on\nthe online class-incremental continual learning setting, where algorithms or\nhumans must learn image classes one at a time during a single pass through a\ndataset. We find that curriculum consistently influences learning outcomes for\nhumans and for multiple continual machine learning algorithms across several\nbenchmark datasets. We introduce a novel-object recognition dataset for human\ncurriculum learning experiments and observe that curricula that are effective\nfor humans are highly correlated with those that are effective for machines. As\nan initial step towards automated curriculum design for online\nclass-incremental learning, we propose a novel algorithm, dubbed Curriculum\nDesigner (CD), that designs and ranks curricula based on inter-class feature\nsimilarities. We find significant overlap between curricula that are\nempirically highly effective and those that are highly ranked by our CD. Our\nstudy establishes a framework for further research on teaching humans and\nmachines to learn continuously using optimized curricula.\n", "date": "2022-11-28", "categories": []}
{"id": "2303.06937", "title": "TARGET: Federated Class-Continual Learning via Exemplar-Free\n  Distillation", "abstract": "  This paper focuses on an under-explored yet important problem: Federated\nClass-Continual Learning (FCCL), where new classes are dynamically added in\nfederated learning. Existing FCCL works suffer from various limitations, such\nas requiring additional datasets or storing the private data from previous\ntasks. In response, we first demonstrate that non-IID data exacerbates\ncatastrophic forgetting issue in FL. Then we propose a novel method called\nTARGET (federat\\textbf{T}ed cl\\textbf{A}ss-continual lea\\textbf{R}nin\\textbf{G}\nvia \\textbf{E}xemplar-free dis\\textbf{T}illation), which alleviates\ncatastrophic forgetting in FCCL while preserving client data privacy. Our\nproposed method leverages the previously trained global model to transfer\nknowledge of old tasks to the current task at the model level. Moreover, a\ngenerator is trained to produce synthetic data to simulate the global\ndistribution of data on each client at the data level. Compared to previous\nFCCL methods, TARGET does not require any additional datasets or storing real\ndata from previous tasks, which makes it ideal for data-sensitive scenarios.\n", "date": "2023-03-13", "categories": []}
{"id": "2308.08793", "title": "Task Relation Distillation and Prototypical Pseudo Label for Incremental\n  Named Entity Recognition", "abstract": "  Incremental Named Entity Recognition (INER) involves the sequential learning\nof new entity types without accessing the training data of previously learned\ntypes. However, INER faces the challenge of catastrophic forgetting specific\nfor incremental learning, further aggravated by background shift (i.e., old and\nfuture entity types are labeled as the non-entity type in the current task). To\naddress these challenges, we propose a method called task Relation Distillation\nand Prototypical pseudo label (RDP) for INER. Specifically, to tackle\ncatastrophic forgetting, we introduce a task relation distillation scheme that\nserves two purposes: 1) ensuring inter-task semantic consistency across\ndifferent incremental learning tasks by minimizing inter-task relation\ndistillation loss, and 2) enhancing the model's prediction confidence by\nminimizing intra-task self-entropy loss. Simultaneously, to mitigate background\nshift, we develop a prototypical pseudo label strategy that distinguishes old\nentity types from the current non-entity type using the old model. This\nstrategy generates high-quality pseudo labels by measuring the distances\nbetween token embeddings and type-wise prototypes. We conducted extensive\nexperiments on ten INER settings of three benchmark datasets (i.e., CoNLL2003,\nI2B2, and OntoNotes5). The results demonstrate that our method achieves\nsignificant improvements over the previous state-of-the-art methods, with an\naverage increase of 6.08% in Micro F1 score and 7.71% in Macro F1 score.\n", "date": "2023-08-17", "categories": []}
{"id": "2308.09297", "title": "NAPA-VQ: Neighborhood Aware Prototype Augmentation with Vector\n  Quantization for Continual Learning", "abstract": "  Catastrophic forgetting; the loss of old knowledge upon acquiring new\nknowledge, is a pitfall faced by deep neural networks in real-world\napplications. Many prevailing solutions to this problem rely on storing\nexemplars (previously encountered data), which may not be feasible in\napplications with memory limitations or privacy constraints. Therefore, the\nrecent focus has been on Non-Exemplar based Class Incremental Learning (NECIL)\nwhere a model incrementally learns about new classes without using any past\nexemplars. However, due to the lack of old data, NECIL methods struggle to\ndiscriminate between old and new classes causing their feature representations\nto overlap. We propose NAPA-VQ: Neighborhood Aware Prototype Augmentation with\nVector Quantization, a framework that reduces this class overlap in NECIL. We\ndraw inspiration from Neural Gas to learn the topological relationships in the\nfeature space, identifying the neighboring classes that are most likely to get\nconfused with each other. This neighborhood information is utilized to enforce\nstrong separation between the neighboring classes as well as to generate old\nclass representative prototypes that can better aid in obtaining a\ndiscriminative decision boundary between old and new classes. Our comprehensive\nexperiments on CIFAR-100, TinyImageNet, and ImageNet-Subset demonstrate that\nNAPA-VQ outperforms the State-of-the-art NECIL methods by an average\nimprovement of 5%, 2%, and 4% in accuracy and 10%, 3%, and 9% in forgetting\nrespectively. Our code can be found in https://github.com/TamashaM/NAPA-VQ.git.\n", "date": "2023-08-18", "categories": []}
{"id": "2308.09303", "title": "Online Class Incremental Learning on Stochastic Blurry Task Boundary via\n  Mask and Visual Prompt Tuning", "abstract": "  Continual learning aims to learn a model from a continuous stream of data,\nbut it mainly assumes a fixed number of data and tasks with clear task\nboundaries. However, in real-world scenarios, the number of input data and\ntasks is constantly changing in a statistical way, not a static way. Although\nrecently introduced incremental learning scenarios having blurry task\nboundaries somewhat address the above issues, they still do not fully reflect\nthe statistical properties of real-world situations because of the fixed ratio\nof disjoint and blurry samples. In this paper, we propose a new Stochastic\nincremental Blurry task boundary scenario, called Si-Blurry, which reflects the\nstochastic properties of the real-world. We find that there are two major\nchallenges in the Si-Blurry scenario: (1) inter- and intra-task forgettings and\n(2) class imbalance problem. To alleviate them, we introduce Mask and Visual\nPrompt tuning (MVP). In MVP, to address the inter- and intra-task forgetting\nissues, we propose a novel instance-wise logit masking and contrastive visual\nprompt tuning loss. Both of them help our model discern the classes to be\nlearned in the current batch. It results in consolidating the previous\nknowledge. In addition, to alleviate the class imbalance problem, we introduce\na new gradient similarity-based focal loss and adaptive feature scaling to ease\noverfitting to the major classes and underfitting to the minor classes.\nExtensive experiments show that our proposed MVP significantly outperforms the\nexisting state-of-the-art methods in our challenging Si-Blurry scenario.\n", "date": "2023-08-18", "categories": []}
{"id": "2308.09610", "title": "On the Effectiveness of LayerNorm Tuning for Continual Learning in\n  Vision Transformers", "abstract": "  State-of-the-art rehearsal-free continual learning methods exploit the\npeculiarities of Vision Transformers to learn task-specific prompts,\ndrastically reducing catastrophic forgetting. However, there is a tradeoff\nbetween the number of learned parameters and the performance, making such\nmodels computationally expensive. In this work, we aim to reduce this cost\nwhile maintaining competitive performance. We achieve this by revisiting and\nextending a simple transfer learning idea: learning task-specific normalization\nlayers. Specifically, we tune the scale and bias parameters of LayerNorm for\neach continual learning task, selecting them at inference time based on the\nsimilarity between task-specific keys and the output of the pre-trained model.\nTo make the classifier robust to incorrect selection of parameters during\ninference, we introduce a two-stage training procedure, where we first optimize\nthe task-specific parameters and then train the classifier with the same\nselection procedure of the inference time. Experiments on ImageNet-R and\nCIFAR-100 show that our method achieves results that are either superior or on\npar with {the state of the art} while being computationally cheaper.\n", "date": "2023-08-18", "categories": []}
{"id": "2308.08259", "title": "Graph Relation Aware Continual Learning", "abstract": "  Continual graph learning (CGL) studies the problem of learning from an\ninfinite stream of graph data, consolidating historical knowledge, and\ngeneralizing it to the future task. At once, only current graph data are\navailable. Although some recent attempts have been made to handle this task, we\nstill face two potential challenges: 1) most of existing works only manipulate\non the intermediate graph embedding and ignore intrinsic properties of graphs.\nIt is non-trivial to differentiate the transferred information across graphs.\n2) recent attempts take a parameter-sharing policy to transfer knowledge across\ntime steps or progressively expand new architecture given shifted graph\ndistribution. Learning a single model could loss discriminative information for\neach graph task while the model expansion scheme suffers from high model\ncomplexity. In this paper, we point out that latent relations behind graph\nedges can be attributed as an invariant factor for the evolving graphs and the\nstatistical information of latent relations evolves. Motivated by this, we\ndesign a relation-aware adaptive model, dubbed as RAM-CG, that consists of a\nrelation-discovery modular to explore latent relations behind edges and a\ntask-awareness masking classifier to accounts for the shifted. Extensive\nexperiments show that RAM-CG provides significant 2.2%, 6.9% and 6.6% accuracy\nimprovements over the state-of-the-art results on CitationNet, OGBN-arxiv and\nTWITCH dataset, respective.\n", "date": "2023-08-16", "categories": []}
{"id": "2308.00924", "title": "Continual Domain Adaptation on Aerial Images under Gradually Degrading\n  Weather", "abstract": "  Domain adaptation (DA) strives to mitigate the domain gap between the source\ndomain where a model is trained, and the target domain where the model is\ndeployed. When a deep learning model is deployed on an aerial platform, it may\nface gradually degrading weather conditions during operation, leading to\nwidening domain gaps between the training data and the encountered evaluation\ndata. We synthesize two such gradually worsening weather conditions on real\nimages from two existing aerial imagery datasets, generating a total of four\nbenchmark datasets. Under the continual, or test-time adaptation setting, we\nevaluate three DA models on our datasets: a baseline standard DA model and two\ncontinual DA models. In such setting, the models can access only one small\nportion, or one batch of the target data at a time, and adaptation takes place\ncontinually, and over only one epoch of the data. The combination of the\nconstraints of continual adaptation, and gradually deteriorating weather\nconditions provide the practical DA scenario for aerial deployment. Among the\nevaluated models, we consider both convolutional and transformer architectures\nfor comparison. We discover stability issues during adaptation for existing\nbuffer-fed continual DA methods, and offer gradient normalization as a simple\nsolution to curb training instability.\n", "date": "2023-08-02", "categories": []}
{"id": "2302.13334", "title": "Knowledge Restore and Transfer for Multi-label Class-Incremental\n  Learning", "abstract": "  Current class-incremental learning research mainly focuses on single-label\nclassification tasks while multi-label class-incremental learning (MLCIL) with\nmore practical application scenarios is rarely studied. Although there have\nbeen many anti-forgetting methods to solve the problem of catastrophic\nforgetting in class-incremental learning, these methods have difficulty in\nsolving the MLCIL problem due to label absence and information dilution. In\nthis paper, we propose a knowledge restore and transfer (KRT) framework for\nMLCIL, which includes a dynamic pseudo-label (DPL) module to restore the old\nclass knowledge and an incremental cross-attention(ICA) module to save\nsession-specific knowledge and transfer old class knowledge to the new model\nsufficiently. Besides, we propose a token loss to jointly optimize the\nincremental cross-attention module. Experimental results on MS-COCO and PASCAL\nVOC datasets demonstrate the effectiveness of our method for improving\nrecognition performance and mitigating forgetting on multi-label\nclass-incremental learning tasks.\n", "date": "2023-02-26", "categories": []}
{"id": "2308.06925", "title": "CBA: Improving Online Continual Learning via Continual Bias Adaptor", "abstract": "  Online continual learning (CL) aims to learn new knowledge and consolidate\npreviously learned knowledge from non-stationary data streams. Due to the\ntime-varying training setting, the model learned from a changing distribution\neasily forgets the previously learned knowledge and biases toward the newly\nreceived task. To address this problem, we propose a Continual Bias Adaptor\n(CBA) module to augment the classifier network to adapt to catastrophic\ndistribution change during training, such that the classifier network is able\nto learn a stable consolidation of previously learned tasks. In the testing\nstage, CBA can be removed which introduces no additional computation cost and\nmemory overhead. We theoretically reveal the reason why the proposed method can\neffectively alleviate catastrophic distribution shifts, and empirically\ndemonstrate its effectiveness through extensive experiments based on four\nrehearsal-based baselines and three public continual learning benchmarks.\n", "date": "2023-08-14", "categories": []}
{"id": "2308.07146", "title": "CTP: Towards Vision-Language Continual Pretraining via Compatible\n  Momentum Contrast and Topology Preservation", "abstract": "  Vision-Language Pretraining (VLP) has shown impressive results on diverse\ndownstream tasks by offline training on large-scale datasets. Regarding the\ngrowing nature of real-world data, such an offline training paradigm on\never-expanding data is unsustainable, because models lack the continual\nlearning ability to accumulate knowledge constantly. However, most continual\nlearning studies are limited to uni-modal classification and existing\nmulti-modal datasets cannot simulate continual non-stationary data stream\nscenarios. To support the study of Vision-Language Continual Pretraining\n(VLCP), we first contribute a comprehensive and unified benchmark dataset P9D\nwhich contains over one million product image-text pairs from 9 industries. The\ndata from each industry as an independent task supports continual learning and\nconforms to the real-world long-tail nature to simulate pretraining on web\ndata. We comprehensively study the characteristics and challenges of VLCP, and\npropose a new algorithm: Compatible momentum contrast with Topology\nPreservation, dubbed CTP. The compatible momentum model absorbs the knowledge\nof the current and previous-task models to flexibly update the modal feature.\nMoreover, Topology Preservation transfers the knowledge of embedding across\ntasks while preserving the flexibility of feature adjustment. The experimental\nresults demonstrate our method not only achieves superior performance compared\nwith other baselines but also does not bring an expensive training burden.\nDataset and codes are available at https://github.com/KevinLight831/CTP.\n", "date": "2023-08-14", "categories": []}
{"id": "2303.06628", "title": "Preventing Zero-Shot Transfer Degradation in Continual Learning of\n  Vision-Language Models", "abstract": "  Continual learning (CL) can help pre-trained vision-language models\nefficiently adapt to new or under-trained data distributions without\nre-training. Nevertheless, during the continual training of the Contrastive\nLanguage-Image Pre-training (CLIP) model, we observe that the model's zero-shot\ntransfer ability significantly degrades due to catastrophic forgetting.\nExisting CL methods can mitigate forgetting by replaying previous data.\nHowever, since the CLIP dataset is private, replay methods cannot access the\npre-training dataset. In addition, replaying data of previously learned\ndownstream tasks can enhance their performance but comes at the cost of\nsacrificing zero-shot performance. To address this challenge, we propose a\nnovel method ZSCL to prevent zero-shot transfer degradation in the continual\nlearning of vision-language models in both feature and parameter space. In the\nfeature space, a reference dataset is introduced for distillation between the\ncurrent and initial models. The reference dataset should have semantic\ndiversity but no need to be labeled, seen in pre-training, or matched\nimage-text pairs. In parameter space, we prevent a large parameter shift by\naveraging weights during the training. We propose a more challenging\nMulti-domain Task Incremental Learning (MTIL) benchmark to evaluate different\nmethods, where tasks are from various domains instead of class-separated in a\nsingle dataset. Our method outperforms other methods in the traditional\nclass-incremental learning setting and the MTIL by 9.7% average score. Our code\nlocates at https://github.com/Thunderbeee/ZSCL.\n", "date": "2023-03-12", "categories": []}
{"id": "2308.04539", "title": "Improving Performance in Continual Learning Tasks using Bio-Inspired\n  Architectures", "abstract": "  The ability to learn continuously from an incoming data stream without\ncatastrophic forgetting is critical to designing intelligent systems. Many\napproaches to continual learning rely on stochastic gradient descent and its\nvariants that employ global error updates, and hence need to adopt strategies\nsuch as memory buffers or replay to circumvent its stability, greed, and\nshort-term memory limitations. To address this limitation, we have developed a\nbiologically inspired lightweight neural network architecture that incorporates\nsynaptic plasticity mechanisms and neuromodulation and hence learns through\nlocal error signals to enable online continual learning without stochastic\ngradient descent.\n  Our approach leads to superior online continual learning performance on\nSplit-MNIST, Split-CIFAR-10, and Split-CIFAR-100 datasets compared to other\nmemory-constrained learning approaches and matches that of the state-of-the-art\nmemory-intensive replay-based approaches. We further demonstrate the\neffectiveness of our approach by integrating key design concepts into other\nbackpropagation-based continual learning algorithms, significantly improving\ntheir accuracy. Our results provide compelling evidence for the importance of\nincorporating biological principles into machine learning models and offer\ninsights into how we can leverage them to design more efficient and robust\nsystems for online continual learning.\n", "date": "2023-08-08", "categories": []}
{"id": "2304.09042", "title": "Adapter Learning in Pretrained Feature Extractor for Continual Learning\n  of Diseases", "abstract": "  Currently intelligent diagnosis systems lack the ability of continually\nlearning to diagnose new diseases once deployed, under the condition of\npreserving old disease knowledge. In particular, updating an intelligent\ndiagnosis system with training data of new diseases would cause catastrophic\nforgetting of old disease knowledge. To address the catastrophic forgetting\nissue, an Adapter-based Continual Learning framework called ACL is proposed to\nhelp effectively learn a set of new diseases at each round (or task) of\ncontinual learning, without changing the shared feature extractor. The\nlearnable lightweight task-specific adapter(s) can be flexibly designed (e.g.,\ntwo convolutional layers) and then added to the pretrained and fixed feature\nextractor. Together with a specially designed task-specific head which absorbs\nall previously learned old diseases as a single \"out-of-distribution\" category,\ntask-specific adapter(s) can help the pretrained feature extractor more\neffectively extract discriminative features between diseases. In addition, a\nsimple yet effective fine-tuning is applied to collaboratively fine-tune\nmultiple task-specific heads such that outputs from different heads are\ncomparable and consequently the appropriate classifier head can be more\naccurately selected during model inference. Extensive empirical evaluations on\nthree image datasets demonstrate the superior performance of ACL in continual\nlearning of new diseases. The source code is available at\nhttps://github.com/GiantJun/CL_Pytorch.\n", "date": "2023-04-18", "categories": []}
{"id": "2308.02790", "title": "Few-shot Class-Incremental Semantic Segmentation via Pseudo-Labeling and\n  Knowledge Distillation", "abstract": "  We address the problem of learning new classes for semantic segmentation\nmodels from few examples, which is challenging because of the following two\nreasons. Firstly, it is difficult to learn from limited novel data to capture\nthe underlying class distribution. Secondly, it is challenging to retain\nknowledge for existing classes and to avoid catastrophic forgetting. For\nlearning from limited data, we propose a pseudo-labeling strategy to augment\nthe few-shot training annotations in order to learn novel classes more\neffectively. Given only one or a few images labeled with the novel classes and\na much larger set of unlabeled images, we transfer the knowledge from labeled\nimages to unlabeled images with a coarse-to-fine pseudo-labeling approach in\ntwo steps. Specifically, we first match each labeled image to its nearest\nneighbors in the unlabeled image set at the scene level, in order to obtain\nimages with a similar scene layout. This is followed by obtaining pseudo-labels\nwithin this neighborhood by applying classifiers learned on the few-shot\nannotations. In addition, we use knowledge distillation on both labeled and\nunlabeled data to retain knowledge on existing classes. We integrate the above\nsteps into a single convolutional neural network with a unified learning\nobjective. Extensive experiments on the Cityscapes and KITTI datasets validate\nthe efficacy of the proposed approach in the self-driving domain. Code is\navailable from https://github.com/ChasonJiang/FSCILSS.\n", "date": "2023-08-05", "categories": []}
{"id": "2308.02346", "title": "Class Incremental Learning with Self-Supervised Pre-Training and\n  Prototype Learning", "abstract": "  Deep Neural Network (DNN) has achieved great success on datasets of closed\nclass set. However, new classes, like new categories of social media topics,\nare continuously added to the real world, making it necessary to incrementally\nlearn. This is hard for DNN because it tends to focus on fitting to new classes\nwhile ignoring old classes, a phenomenon known as catastrophic forgetting.\nState-of-the-art methods rely on knowledge distillation and data replay\ntechniques but still have limitations. In this work, we analyze the causes of\ncatastrophic forgetting in class incremental learning, which owes to three\nfactors: representation drift, representation confusion, and classifier\ndistortion. Based on this view, we propose a two-stage learning framework with\na fixed encoder and an incrementally updated prototype classifier. The encoder\nis trained with self-supervised learning to generate a feature space with high\nintrinsic dimensionality, thus improving its transferability and generality.\nThe classifier incrementally learns new prototypes while retaining the\nprototypes of previously learned data, which is crucial in preserving the\ndecision boundary.Our method does not rely on preserved samples of old classes,\nis thus a non-exemplar based CIL method. Experiments on public datasets show\nthat our method can significantly outperform state-of-the-art exemplar-based\nmethods when they reserved 5 examplers per class, under the incremental setting\nof 10 phases, by 18.24% on CIFAR-100 and 9.37% on ImageNet100.\n", "date": "2023-08-04", "categories": []}
{"id": "2308.01698", "title": "Balanced Destruction-Reconstruction Dynamics for Memory-replay Class\n  Incremental Learning", "abstract": "  Class incremental learning (CIL) aims to incrementally update a trained model\nwith the new classes of samples (plasticity) while retaining previously learned\nability (stability). To address the most challenging issue in this goal, i.e.,\ncatastrophic forgetting, the mainstream paradigm is memory-replay CIL, which\nconsolidates old knowledge by replaying a small number of old classes of\nsamples saved in the memory. Despite effectiveness, the inherent\ndestruction-reconstruction dynamics in memory-replay CIL are an intrinsic\nlimitation: if the old knowledge is severely destructed, it will be quite hard\nto reconstruct the lossless counterpart. Our theoretical analysis shows that\nthe destruction of old knowledge can be effectively alleviated by balancing the\ncontribution of samples from the current phase and those saved in the memory.\nMotivated by this theoretical finding, we propose a novel Balanced\nDestruction-Reconstruction module (BDR) for memory-replay CIL, which can\nachieve better knowledge reconstruction by reducing the degree of maximal\ndestruction of old knowledge. Specifically, to achieve a better balance between\nold knowledge and new classes, the proposed BDR module takes into account two\nfactors: the variance in training status across different classes and the\nquantity imbalance of samples from the current phase and memory. By dynamically\nmanipulating the gradient during training based on these factors, BDR can\neffectively alleviate knowledge destruction and improve knowledge\nreconstruction. Extensive experiments on a range of CIL benchmarks have shown\nthat as a lightweight plug-and-play module, BDR can significantly improve the\nperformance of existing state-of-the-art methods with good generalization.\n", "date": "2023-08-03", "categories": []}
{"id": "2308.01746", "title": "Neural Collapse Terminus: A Unified Solution for Class Incremental\n  Learning and Its Variants", "abstract": "  How to enable learnability for new classes while keeping the capability well\non old classes has been a crucial challenge for class incremental learning.\nBeyond the normal case, long-tail class incremental learning and few-shot class\nincremental learning are also proposed to consider the data imbalance and data\nscarcity, respectively, which are common in real-world implementations and\nfurther exacerbate the well-known problem of catastrophic forgetting. Existing\nmethods are specifically proposed for one of the three tasks. In this paper, we\noffer a unified solution to the misalignment dilemma in the three tasks.\nConcretely, we propose neural collapse terminus that is a fixed structure with\nthe maximal equiangular inter-class separation for the whole label space. It\nserves as a consistent target throughout the incremental training to avoid\ndividing the feature space incrementally. For CIL and LTCIL, we further propose\na prototype evolving scheme to drive the backbone features into our neural\ncollapse terminus smoothly. Our method also works for FSCIL with only minor\nadaptations. Theoretical analysis indicates that our method holds the neural\ncollapse optimality in an incremental fashion regardless of data imbalance or\ndata scarcity. We also design a generalized case where we do not know the total\nnumber of classes and whether the data distribution is normal, long-tail, or\nfew-shot for each coming session, to test the generalizability of our method.\nExtensive experiments with multiple datasets are conducted to demonstrate the\neffectiveness of our unified solution to all the three tasks and the\ngeneralized case.\n", "date": "2023-08-03", "categories": []}
{"id": "2308.01895", "title": "Improving Replay Sample Selection and Storage for Less Forgetting in\n  Continual Learning", "abstract": "  Continual learning seeks to enable deep learners to train on a series of\ntasks of unknown length without suffering from the catastrophic forgetting of\nprevious tasks. One effective solution is replay, which involves storing few\nprevious experiences in memory and replaying them when learning the current\ntask. However, there is still room for improvement when it comes to selecting\nthe most informative samples for storage and determining the optimal number of\nsamples to be stored. This study aims to address these issues with a novel\ncomparison of the commonly used reservoir sampling to various alternative\npopulation strategies and providing a novel detailed analysis of how to find\nthe optimal number of stored samples.\n", "date": "2023-08-03", "categories": []}
{"id": "2308.01127", "title": "DiffusePast: Diffusion-based Generative Replay for Class Incremental\n  Semantic Segmentation", "abstract": "  The Class Incremental Semantic Segmentation (CISS) extends the traditional\nsegmentation task by incrementally learning newly added classes. Previous work\nhas introduced generative replay, which involves replaying old class samples\ngenerated from a pre-trained GAN, to address the issues of catastrophic\nforgetting and privacy concerns. However, the generated images lack semantic\nprecision and exhibit out-of-distribution characteristics, resulting in\ninaccurate masks that further degrade the segmentation performance. To tackle\nthese challenges, we propose DiffusePast, a novel framework featuring a\ndiffusion-based generative replay module that generates semantically accurate\nimages with more reliable masks guided by different instructions (e.g., text\nprompts or edge maps). Specifically, DiffusePast introduces a dual-generator\nparadigm, which focuses on generating old class images that align with the\ndistribution of downstream datasets while preserving the structure and layout\nof the original images, enabling more precise masks. To adapt to the novel\nvisual concepts of newly added classes continuously, we incorporate class-wise\ntoken embedding when updating the dual-generator. Moreover, we assign adequate\npseudo-labels of old classes to the background pixels in the new step images,\nfurther mitigating the forgetting of previously learned knowledge. Through\ncomprehensive experiments, our method demonstrates competitive performance\nacross mainstream benchmarks, striking a better balance between the performance\nof old and novel classes.\n", "date": "2023-08-02", "categories": []}
{"id": "2212.11110", "title": "Lifelong Reinforcement Learning with Modulating Masks", "abstract": "  Lifelong learning aims to create AI systems that continuously and\nincrementally learn during a lifetime, similar to biological learning. Attempts\nso far have met problems, including catastrophic forgetting, interference among\ntasks, and the inability to exploit previous knowledge. While considerable\nresearch has focused on learning multiple supervised classification tasks that\ninvolve changes in the input distribution, lifelong reinforcement learning\n(LRL) must deal with variations in the state and transition distributions, and\nin the reward functions. Modulating masks with a fixed backbone network,\nrecently developed for classification, are particularly suitable to deal with\nsuch a large spectrum of task variations. In this paper, we adapted modulating\nmasks to work with deep LRL, specifically PPO and IMPALA agents. The comparison\nwith LRL baselines in both discrete and continuous RL tasks shows superior\nperformance. We further investigated the use of a linear combination of\npreviously learned masks to exploit previous knowledge when learning new tasks:\nnot only is learning faster, the algorithm solves tasks that we could not\notherwise solve from scratch due to extremely sparse rewards. The results\nsuggest that RL with modulating masks is a promising approach to lifelong\nlearning, to the composition of knowledge to learn increasingly complex tasks,\nand to knowledge reuse for efficient and faster learning.\n", "date": "2022-12-21", "categories": []}
{"id": "2307.16419", "title": "Subspace Distillation for Continual Learning", "abstract": "  An ultimate objective in continual learning is to preserve knowledge learned\nin preceding tasks while learning new tasks. To mitigate forgetting prior\nknowledge, we propose a novel knowledge distillation technique that takes into\nthe account the manifold structure of the latent/output space of a neural\nnetwork in learning novel tasks. To achieve this, we propose to approximate the\ndata manifold up-to its first order, hence benefiting from linear subspaces to\nmodel the structure and maintain the knowledge of a neural network while\nlearning novel concepts. We demonstrate that the modeling with subspaces\nprovides several intriguing properties, including robustness to noise and\ntherefore effective for mitigating Catastrophic Forgetting in continual\nlearning. We also discuss and show how our proposed method can be adopted to\naddress both classification and segmentation problems. Empirically, we observe\nthat our proposed method outperforms various continual learning methods on\nseveral challenging datasets including Pascal VOC, and Tiny-Imagenet.\nFurthermore, we show how the proposed method can be seamlessly combined with\nexisting learning approaches to improve their performances. The codes of this\narticle will be available at https://github.com/csiro-robotics/SDCL.\n", "date": "2023-07-31", "categories": []}
{"id": "2307.16459", "title": "L3DMC: Lifelong Learning using Distillation via Mixed-Curvature Space", "abstract": "  The performance of a lifelong learning (L3) model degrades when it is trained\non a series of tasks, as the geometrical formation of the embedding space\nchanges while learning novel concepts sequentially. The majority of existing L3\napproaches operate on a fixed-curvature (e.g., zero-curvature Euclidean) space\nthat is not necessarily suitable for modeling the complex geometric structure\nof data. Furthermore, the distillation strategies apply constraints directly on\nlow-dimensional embeddings, discouraging the L3 model from learning new\nconcepts by making the model highly stable. To address the problem, we propose\na distillation strategy named L3DMC that operates on mixed-curvature spaces to\npreserve the already-learned knowledge by modeling and maintaining complex\ngeometrical structures. We propose to embed the projected low dimensional\nembedding of fixed-curvature spaces (Euclidean and hyperbolic) to\nhigher-dimensional Reproducing Kernel Hilbert Space (RKHS) using a\npositive-definite kernel function to attain rich representation. Afterward, we\noptimize the L3 model by minimizing the discrepancies between the new sample\nrepresentation and the subspace constructed using the old representation in\nRKHS. L3DMC is capable of adapting new knowledge better without forgetting old\nknowledge as it combines the representation power of multiple fixed-curvature\nspaces and is performed on higher-dimensional RKHS. Thorough experiments on\nthree benchmarks demonstrate the effectiveness of our proposed distillation\nstrategy for medical image classification in L3 settings. Our code\nimplementation is publicly available at\nhttps://github.com/csiro-robotics/L3DMC.\n", "date": "2023-07-31", "categories": []}
{"id": "2308.00199", "title": "CBCL-PR: A Cognitively Inspired Model for Class-Incremental Learning in\n  Robotics", "abstract": "  For most real-world applications, robots need to adapt and learn continually\nwith limited data in their environments. In this paper, we consider the problem\nof Few-Shot class Incremental Learning (FSIL), in which an AI agent is required\nto learn incrementally from a few data samples without forgetting the data it\nhas previously learned. To solve this problem, we present a novel framework\ninspired by theories of concept learning in the hippocampus and the neocortex.\nOur framework represents object classes in the form of sets of clusters and\nstores them in memory. The framework replays data generated by the clusters of\nthe old classes, to avoid forgetting when learning new classes. Our approach is\nevaluated on two object classification datasets resulting in state-of-the-art\n(SOTA) performance for class-incremental learning and FSIL. We also evaluate\nour framework for FSIL on a robot demonstrating that the robot can continually\nlearn to classify a large set of household objects with limited human\nassistance.\n", "date": "2023-07-31", "categories": []}
{"id": "2308.00301", "title": "Online Prototype Learning for Online Continual Learning", "abstract": "  Online continual learning (CL) studies the problem of learning continuously\nfrom a single-pass data stream while adapting to new data and mitigating\ncatastrophic forgetting. Recently, by storing a small subset of old data,\nreplay-based methods have shown promising performance. Unlike previous methods\nthat focus on sample storage or knowledge distillation against catastrophic\nforgetting, this paper aims to understand why the online learning models fail\nto generalize well from a new perspective of shortcut learning. We identify\nshortcut learning as the key limiting factor for online CL, where the learned\nfeatures may be biased, not generalizable to new tasks, and may have an adverse\nimpact on knowledge distillation. To tackle this issue, we present the online\nprototype learning (OnPro) framework for online CL. First, we propose online\nprototype equilibrium to learn representative features against shortcut\nlearning and discriminative features to avoid class confusion, ultimately\nachieving an equilibrium status that separates all seen classes well while\nlearning new classes. Second, with the feedback of online prototypes, we devise\na novel adaptive prototypical feedback mechanism to sense the classes that are\neasily misclassified and then enhance their boundaries. Extensive experimental\nresults on widely-used benchmark datasets demonstrate the superior performance\nof OnPro over the state-of-the-art baseline methods. Source code is available\nat https://github.com/weilllllls/OnPro.\n", "date": "2023-08-01", "categories": []}
{"id": "2303.07811", "title": "ICICLE: Interpretable Class Incremental Continual Learning", "abstract": "  Continual learning enables incremental learning of new tasks without\nforgetting those previously learned, resulting in positive knowledge transfer\nthat can enhance performance on both new and old tasks. However, continual\nlearning poses new challenges for interpretability, as the rationale behind\nmodel predictions may change over time, leading to interpretability concept\ndrift. We address this problem by proposing Interpretable Class-InCremental\nLEarning (ICICLE), an exemplar-free approach that adopts a prototypical\npart-based approach. It consists of three crucial novelties: interpretability\nregularization that distills previously learned concepts while preserving\nuser-friendly positive reasoning; proximity-based prototype initialization\nstrategy dedicated to the fine-grained setting; and task-recency bias\ncompensation devoted to prototypical parts. Our experimental results\ndemonstrate that ICICLE reduces the interpretability concept drift and\noutperforms the existing exemplar-free methods of common class-incremental\nlearning when applied to concept-based models.\n", "date": "2023-03-14", "categories": []}
{"id": "2305.14758", "title": "MRN: Multiplexed Routing Network for Incremental Multilingual Text\n  Recognition", "abstract": "  Multilingual text recognition (MLTR) systems typically focus on a fixed set\nof languages, which makes it difficult to handle newly added languages or adapt\nto ever-changing data distribution. In this paper, we propose the Incremental\nMLTR (IMLTR) task in the context of incremental learning (IL), where different\nlanguages are introduced in batches. IMLTR is particularly challenging due to\nrehearsal-imbalance, which refers to the uneven distribution of sample\ncharacters in the rehearsal set, used to retain a small amount of old data as\npast memories. To address this issue, we propose a Multiplexed Routing Network\n(MRN). MRN trains a recognizer for each language that is currently seen.\nSubsequently, a language domain predictor is learned based on the rehearsal set\nto weigh the recognizers. Since the recognizers are derived from the original\ndata, MRN effectively reduces the reliance on older data and better fights\nagainst catastrophic forgetting, the core issue in IL. We extensively evaluate\nMRN on MLT17 and MLT19 datasets. It outperforms existing general-purpose IL\nmethods by large margins, with average accuracy improvements ranging from 10.3%\nto 35.8% under different settings. Code is available at\nhttps://github.com/simplify23/MRN.\n", "date": "2023-05-24", "categories": []}
{"id": "2212.02846", "title": "Statistical mechanics of continual learning: variational principle and\n  mean-field potential", "abstract": "  An obstacle to artificial general intelligence is set by continual learning\nof multiple tasks of different nature. Recently, various heuristic tricks, both\nfrom machine learning and from neuroscience angles, were proposed, but they\nlack a unified theory ground. Here, we focus on continual learning in\nsingle-layered and multi-layered neural networks of binary weights. A\nvariational Bayesian learning setting is thus proposed, where the neural\nnetworks are trained in a field-space, rather than gradient-ill-defined\ndiscrete-weight space, and furthermore, weight uncertainty is naturally\nincorporated, and modulates synaptic resources among tasks. From a physics\nperspective, we translate the variational continual learning into Franz-Parisi\nthermodynamic potential framework, where previous task knowledge acts as a\nprior and a reference as well. We thus interpret the continual learning of the\nbinary perceptron in a teacher-student setting as a Franz-Parisi potential\ncomputation. The learning performance can then be analytically studied with\nmean-field order parameters, whose predictions coincide with numerical\nexperiments using stochastic gradient descent methods. Based on the variational\nprinciple and Gaussian field approximation of internal preactivations in hidden\nlayers, we also derive the learning algorithm considering weight uncertainty,\nwhich solves the continual learning with binary weights using multi-layered\nneural networks, and performs better than the currently available\nmetaplasticity algorithm. Our proposed principled frameworks also connect to\nelastic weight consolidation, weight-uncertainty modulated learning, and\nneuroscience inspired metaplasticity, providing a theory-grounded method for\nthe real-world multi-task learning with deep networks.\n", "date": "2022-12-06", "categories": []}
{"id": "2307.15105", "title": "Detecting Morphing Attacks via Continual Incremental Training", "abstract": "  Scenarios in which restrictions in data transfer and storage limit the\npossibility to compose a single dataset -- also exploiting different data\nsources -- to perform a batch-based training procedure, make the development of\nrobust models particularly challenging. We hypothesize that the recent\nContinual Learning (CL) paradigm may represent an effective solution to enable\nincremental training, even through multiple sites. Indeed, a basic assumption\nof CL is that once a model has been trained, old data can no longer be used in\nsuccessive training iterations and in principle can be deleted. Therefore, in\nthis paper, we investigate the performance of different Continual Learning\nmethods in this scenario, simulating a learning model that is updated every\ntime a new chunk of data, even of variable size, is available. Experimental\nresults reveal that a particular CL method, namely Learning without Forgetting\n(LwF), is one of the best-performing algorithms. Then, we investigate its usage\nand parametrization in Morphing Attack Detection and Object Classification\ntasks, specifically with respect to the amount of new training data that became\navailable.\n", "date": "2023-07-27", "categories": []}
{"id": "2211.12292", "title": "Exemplar-free Continual Learning of Vision Transformers via Gated\n  Class-Attention and Cascaded Feature Drift Compensation", "abstract": "  We propose a new method for exemplar-free class incremental training of ViTs.\nThe main challenge of exemplar-free continual learning is maintaining\nplasticity of the learner without causing catastrophic forgetting of previously\nlearned tasks. This is often achieved via exemplar replay which can help\nrecalibrate previous task classifiers to the feature drift which occurs when\nlearning new tasks. Exemplar replay, however, comes at the cost of retaining\nsamples from previous tasks which for many applications may not be possible. To\naddress the problem of continual ViT training, we first propose gated\nclass-attention to minimize the drift in the final ViT transformer block. This\nmask-based gating is applied to class-attention mechanism of the last\ntransformer block and strongly regulates the weights crucial for previous\ntasks. Importantly, gated class-attention does not require the task-ID during\ninference, which distinguishes it from other parameter isolation methods.\nSecondly, we propose a new method of feature drift compensation that\naccommodates feature drift in the backbone when learning new tasks. The\ncombination of gated class-attention and cascaded feature drift compensation\nallows for plasticity towards new tasks while limiting forgetting of previous\nones. Extensive experiments performed on CIFAR-100, Tiny-ImageNet and\nImageNet100 demonstrate that our exemplar-free method obtains competitive\nresults when compared to rehearsal based ViT methods.\n", "date": "2022-11-22", "categories": []}
{"id": "2307.05747", "title": "Integrating Curricula with Replays: Its Effects on Continual Learning", "abstract": "  Humans engage in learning and reviewing processes with curricula when\nacquiring new skills or knowledge. This human learning behavior has inspired\nthe integration of curricula with replay methods in continual learning agents.\nThe goal is to emulate the human learning process, thereby improving knowledge\nretention and facilitating learning transfer. Existing replay methods in\ncontinual learning agents involve the random selection and ordering of data\nfrom previous tasks, which has shown to be effective. However, limited research\nhas explored the integration of different curricula with replay methods to\nenhance continual learning. Our study takes initial steps in examining the\nimpact of integrating curricula with replay methods on continual learning in\nthree specific aspects: the interleaved frequency of replayed exemplars with\ntraining data, the sequence in which exemplars are replayed, and the strategy\nfor selecting exemplars into the replay buffer. These aspects of curricula\ndesign align with cognitive psychology principles and leverage the benefits of\ninterleaved practice during replays, easy-to-hard rehearsal, and exemplar\nselection strategy involving exemplars from a uniform distribution of\ndifficulties. Based on our results, these three curricula effectively mitigated\ncatastrophic forgetting and enhanced positive knowledge transfer, demonstrating\nthe potential of curricula in advancing continual learning methodologies. Our\ncode and data are available:\nhttps://github.com/ZhangLab-DeepNeuroCogLab/Integrating-Curricula-with-Replays\n", "date": "2023-07-08", "categories": []}
{"id": "2307.12045", "title": "Revisiting Distillation for Continual Learning on Visual Question\n  Localized-Answering in Robotic Surgery", "abstract": "  The visual-question localized-answering (VQLA) system can serve as a\nknowledgeable assistant in surgical education. Except for providing text-based\nanswers, the VQLA system can highlight the interested region for better\nsurgical scene understanding. However, deep neural networks (DNNs) suffer from\ncatastrophic forgetting when learning new knowledge. Specifically, when DNNs\nlearn on incremental classes or tasks, their performance on old tasks drops\ndramatically. Furthermore, due to medical data privacy and licensing issues, it\nis often difficult to access old data when updating continual learning (CL)\nmodels. Therefore, we develop a non-exemplar continual surgical VQLA framework,\nto explore and balance the rigidity-plasticity trade-off of DNNs in a\nsequential learning paradigm. We revisit the distillation loss in CL tasks, and\npropose rigidity-plasticity-aware distillation (RP-Dist) and self-calibrated\nheterogeneous distillation (SH-Dist) to preserve the old knowledge. The weight\naligning (WA) technique is also integrated to adjust the weight bias between\nold and new tasks. We further establish a CL framework on three public surgical\ndatasets in the context of surgical settings that consist of overlapping\nclasses between old and new surgical VQLA tasks. With extensive experiments, we\ndemonstrate that our proposed method excellently reconciles learning and\nforgetting on the continual surgical VQLA over conventional CL methods. Our\ncode is publicly accessible.\n", "date": "2023-07-22", "categories": []}
{"id": "2307.12427", "title": "Augmented Box Replay: Overcoming Foreground Shift for Incremental Object\n  Detection", "abstract": "  In incremental learning, replaying stored samples from previous tasks\ntogether with current task samples is one of the most efficient approaches to\naddress catastrophic forgetting. However, unlike incremental classification,\nimage replay has not been successfully applied to incremental object detection\n(IOD). In this paper, we identify the overlooked problem of foreground shift as\nthe main reason for this. Foreground shift only occurs when replaying images of\nprevious tasks and refers to the fact that their background might contain\nforeground objects of the current task. To overcome this problem, a novel and\nefficient Augmented Box Replay (ABR) method is developed that only stores and\nreplays foreground objects and thereby circumvents the foreground shift\nproblem. In addition, we propose an innovative Attentive RoI Distillation loss\nthat uses spatial attention from region-of-interest (RoI) features to constrain\ncurrent model to focus on the most important information from old model. ABR\nsignificantly reduces forgetting of previous classes while maintaining high\nplasticity in current classes. Moreover, it considerably reduces the storage\nrequirements when compared to standard image replay. Comprehensive experiments\non Pascal-VOC and COCO datasets support the state-of-the-art performance of our\nmodel.\n", "date": "2023-07-23", "categories": []}
{"id": "2307.12660", "title": "Online Continual Learning in Keyword Spotting for Low-Resource Devices\n  via Pooling High-Order Temporal Statistics", "abstract": "  Keyword Spotting (KWS) models on embedded devices should adapt fast to new\nuser-defined words without forgetting previous ones. Embedded devices have\nlimited storage and computational resources, thus, they cannot save samples or\nupdate large models. We consider the setup of embedded online continual\nlearning (EOCL), where KWS models with frozen backbone are trained to\nincrementally recognize new words from a non-repeated stream of samples, seen\none at a time. To this end, we propose Temporal Aware Pooling (TAP) which\nconstructs an enriched feature space computing high-order moments of speech\nfeatures extracted by a pre-trained backbone. Our method, TAP-SLDA, updates a\nGaussian model for each class on the enriched feature space to effectively use\naudio representations. In experimental analyses, TAP-SLDA outperforms\ncompetitors on several setups, backbones, and baselines, bringing a relative\naverage gain of 11.3% on the GSC dataset.\n", "date": "2023-07-24", "categories": []}
{"id": "2307.11317", "title": "XLDA: Linear Discriminant Analysis for Scaling Continual Learning to\n  Extreme Classification at the Edge", "abstract": "  Streaming Linear Discriminant Analysis (LDA) while proven in\nClass-incremental Learning deployments at the edge with limited classes (upto\n1000), has not been proven for deployment in extreme classification scenarios.\nIn this paper, we present: (a) XLDA, a framework for Class-IL in edge\ndeployment where LDA classifier is proven to be equivalent to FC layer\nincluding in extreme classification scenarios, and (b) optimizations to enable\nXLDA-based training and inference for edge deployment where there is a\nconstraint on available compute resources. We show up to 42x speed up using a\nbatched training approach and up to 5x inference speedup with nearest neighbor\nsearch on extreme datasets like AliProducts (50k classes) and Google Landmarks\nV2 (81k classes)\n", "date": "2023-07-21", "categories": []}
{"id": "2307.11386", "title": "CLR: Channel-wise Lightweight Reprogramming for Continual Learning", "abstract": "  Continual learning aims to emulate the human ability to continually\naccumulate knowledge over sequential tasks. The main challenge is to maintain\nperformance on previously learned tasks after learning new tasks, i.e., to\navoid catastrophic forgetting. We propose a Channel-wise Lightweight\nReprogramming (CLR) approach that helps convolutional neural networks (CNNs)\novercome catastrophic forgetting during continual learning. We show that a CNN\nmodel trained on an old task (or self-supervised proxy task) could be\n``reprogrammed\" to solve a new task by using our proposed lightweight (very\ncheap) reprogramming parameter. With the help of CLR, we have a better\nstability-plasticity trade-off to solve continual learning problems: To\nmaintain stability and retain previous task ability, we use a common\ntask-agnostic immutable part as the shared ``anchor\" parameter set. We then add\ntask-specific lightweight reprogramming parameters to reinterpret the outputs\nof the immutable parts, to enable plasticity and integrate new knowledge. To\nlearn sequential tasks, we only train the lightweight reprogramming parameters\nto learn each new task. Reprogramming parameters are task-specific and\nexclusive to each task, which makes our method immune to catastrophic\nforgetting. To minimize the parameter requirement of reprogramming to learn new\ntasks, we make reprogramming lightweight by only adjusting essential kernels\nand learning channel-wise linear mappings from anchor parameters to\ntask-specific domain knowledge. We show that, for general CNNs, the CLR\nparameter increase is less than 0.6\\% for any new task. Our method outperforms\n13 state-of-the-art continual learning baselines on a new challenging sequence\nof 53 image classification datasets. Code and data are available at\nhttps://github.com/gyhandy/Channel-wise-Lightweight-Reprogramming\n", "date": "2023-07-21", "categories": []}
{"id": "2306.12619", "title": "Class-Incremental Learning based on Label Generation", "abstract": "  Despite the great success of pre-trained language models, it is still a\nchallenge to use these models for continual learning, especially for the\nclass-incremental learning (CIL) setting due to catastrophic forgetting (CF).\nThis paper reports our finding that if we formulate CIL as a continual label\ngeneration problem, CF is drastically reduced and the generalizable\nrepresentations of pre-trained models can be better retained. We thus propose a\nnew CIL method (VAG) that also leverages the sparsity of vocabulary to focus\nthe generation and creates pseudo-replay samples by using label semantics.\nExperimental results show that VAG outperforms baselines by a large margin.\n", "date": "2023-06-22", "categories": []}
{"id": "2307.10822", "title": "Gradient-Semantic Compensation for Incremental Semantic Segmentation", "abstract": "  Incremental semantic segmentation aims to continually learn the segmentation\nof new coming classes without accessing the training data of previously learned\nclasses. However, most current methods fail to address catastrophic forgetting\nand background shift since they 1) treat all previous classes equally without\nconsidering different forgetting paces caused by imbalanced gradient\nback-propagation; 2) lack strong semantic guidance between classes. To tackle\nthe above challenges, in this paper, we propose a Gradient-Semantic\nCompensation (GSC) model, which surmounts incremental semantic segmentation\nfrom both gradient and semantic perspectives. Specifically, to address\ncatastrophic forgetting from the gradient aspect, we develop a step-aware\ngradient compensation that can balance forgetting paces of previously seen\nclasses via re-weighting gradient backpropagation. Meanwhile, we propose a\nsoft-sharp semantic relation distillation to distill consistent inter-class\nsemantic relations via soft labels for alleviating catastrophic forgetting from\nthe semantic aspect. In addition, we develop a prototypical pseudo re-labeling\nthat provides strong semantic guidance to mitigate background shift. It\nproduces high-quality pseudo labels for old classes in the background by\nmeasuring distances between pixels and class-wise prototypes. Extensive\nexperiments on three public datasets, i.e., Pascal VOC 2012, ADE20K, and\nCityscapes, demonstrate the effectiveness of our proposed GSC model.\n", "date": "2023-07-20", "categories": []}
{"id": "2307.10845", "title": "Self-paced Weight Consolidation for Continual Learning", "abstract": "  Continual learning algorithms which keep the parameters of new tasks close to\nthat of previous tasks, are popular in preventing catastrophic forgetting in\nsequential task learning settings. However, 1) the performance for the new\ncontinual learner will be degraded without distinguishing the contributions of\npreviously learned tasks; 2) the computational cost will be greatly increased\nwith the number of tasks, since most existing algorithms need to regularize all\nprevious tasks when learning new tasks. To address the above challenges, we\npropose a self-paced Weight Consolidation (spWC) framework to attain robust\ncontinual learning via evaluating the discriminative contributions of previous\ntasks. To be specific, we develop a self-paced regularization to reflect the\npriorities of past tasks via measuring difficulty based on key performance\nindicator (i.e., accuracy). When encountering a new task, all previous tasks\nare sorted from \"difficult\" to \"easy\" based on the priorities. Then the\nparameters of the new continual learner will be learned via selectively\nmaintaining the knowledge amongst more difficult past tasks, which could well\novercome catastrophic forgetting with less computational cost. We adopt an\nalternative convex search to iteratively update the model parameters and\npriority weights in the bi-convex formulation. The proposed spWC framework is\nplug-and-play, which is applicable to most continual learning algorithms (e.g.,\nEWC, MAS and RCIL) in different directions (e.g., classification and\nsegmentation). Experimental results on several public benchmark datasets\ndemonstrate that our proposed framework can effectively improve performance\nwhen compared with other popular continual learning algorithms.\n", "date": "2023-07-20", "categories": []}
{"id": "2307.09827", "title": "Online Continual Learning for Robust Indoor Object Recognition", "abstract": "  Vision systems mounted on home robots need to interact with unseen classes in\nchanging environments. Robots have limited computational resources, labelled\ndata and storage capability. These requirements pose some unique challenges:\nmodels should adapt without forgetting past knowledge in a data- and\nparameter-efficient way. We characterize the problem as few-shot (FS) online\ncontinual learning (OCL), where robotic agents learn from a non-repeated stream\nof few-shot data updating only a few model parameters. Additionally, such\nmodels experience variable conditions at test time, where objects may appear in\ndifferent poses (e.g., horizontal or vertical) and environments (e.g., day or\nnight). To improve robustness of CL agents, we propose RobOCLe, which; 1)\nconstructs an enriched feature space computing high order statistical moments\nfrom the embedded features of samples; and 2) computes similarity between high\norder statistics of the samples on the enriched feature space, and predicts\ntheir class labels. We evaluate robustness of CL models to train/test\naugmentations in various cases. We show that different moments allow RobOCLe to\ncapture different properties of deformations, providing higher robustness with\nno decrease of inference speed.\n", "date": "2023-07-19", "categories": []}
{"id": "2307.00497", "title": "Don't Memorize; Mimic The Past: Federated Class Incremental Learning\n  Without Episodic Memory", "abstract": "  Deep learning models are prone to forgetting information learned in the past\nwhen trained on new data. This problem becomes even more pronounced in the\ncontext of federated learning (FL), where data is decentralized and subject to\nindependent changes for each user. Continual Learning (CL) studies this\nso-called \\textit{catastrophic forgetting} phenomenon primarily in centralized\nsettings, where the learner has direct access to the complete training dataset.\nHowever, applying CL techniques to FL is not straightforward due to privacy\nconcerns and resource limitations. This paper presents a framework for\nfederated class incremental learning that utilizes a generative model to\nsynthesize samples from past distributions instead of storing part of past\ndata. Then, clients can leverage the generative model to mitigate catastrophic\nforgetting locally. The generative model is trained on the server using\ndata-free methods at the end of each task without requesting data from clients.\nTherefore, it reduces the risk of data leakage as opposed to training it on the\nclient's private data. We demonstrate significant improvements for the\nCIFAR-100 dataset compared to existing baselines.\n", "date": "2023-07-02", "categories": []}
{"id": "2302.12200", "title": "A Neural Span-Based Continual Named Entity Recognition Model", "abstract": "  Named Entity Recognition (NER) models capable of Continual Learning (CL) are\nrealistically valuable in areas where entity types continuously increase (e.g.,\npersonal assistants). Meanwhile the learning paradigm of NER advances to new\npatterns such as the span-based methods. However, its potential to CL has not\nbeen fully explored. In this paper, we propose SpanKL, a simple yet effective\nSpan-based model with Knowledge distillation (KD) to preserve memories and\nmulti-Label prediction to prevent conflicts in CL-NER. Unlike prior sequence\nlabeling approaches, the inherently independent modeling in span and entity\nlevel with the designed coherent optimization on SpanKL promotes its learning\nat each incremental step and mitigates the forgetting. Experiments on synthetic\nCL datasets derived from OntoNotes and Few-NERD show that SpanKL significantly\noutperforms previous SoTA in many aspects, and obtains the smallest gap from CL\nto the upper bound revealing its high practiced value. The code is available at\nhttps://github.com/Qznan/SpanKL.\n", "date": "2023-02-23", "categories": []}
{"id": "2303.11165", "title": "Computationally Budgeted Continual Learning: What Does Matter?", "abstract": "  Continual Learning (CL) aims to sequentially train models on streams of\nincoming data that vary in distribution by preserving previous knowledge while\nadapting to new data. Current CL literature focuses on restricted access to\npreviously seen data, while imposing no constraints on the computational budget\nfor training. This is unreasonable for applications in-the-wild, where systems\nare primarily constrained by computational and time budgets, not storage. We\nrevisit this problem with a large-scale benchmark and analyze the performance\nof traditional CL approaches in a compute-constrained setting, where effective\nmemory samples used in training can be implicitly restricted as a consequence\nof limited computation. We conduct experiments evaluating various CL sampling\nstrategies, distillation losses, and partial fine-tuning on two large-scale\ndatasets, namely ImageNet2K and Continual Google Landmarks V2 in data\nincremental, class incremental, and time incremental settings. Through\nextensive experiments amounting to a total of over 1500 GPU-hours, we find\nthat, under compute-constrained setting, traditional CL approaches, with no\nexception, fail to outperform a simple minimal baseline that samples uniformly\nfrom memory. Our conclusions are consistent in a different number of stream\ntime steps, e.g., 20 to 200, and under several computational budgets. This\nsuggests that most existing CL methods are particularly too computationally\nexpensive for realistic budgeted deployment. Code for this project is available\nat: https://github.com/drimpossible/BudgetCL.\n", "date": "2023-03-20", "categories": []}
{"id": "2306.14369", "title": "Few-Shot Continual Learning via Flat-to-Wide Approaches", "abstract": "  Existing approaches on continual learning call for a lot of samples in their\ntraining processes. Such approaches are impractical for many real-world\nproblems having limited samples because of the overfitting problem. This paper\nproposes a few-shot continual learning approach, termed FLat-tO-WidE AppRoach\n(FLOWER), where a flat-to-wide learning process finding the flat-wide minima is\nproposed to address the catastrophic forgetting problem. The issue of data\nscarcity is overcome with a data augmentation approach making use of a ball\ngenerator concept to restrict the sampling space into the smallest enclosing\nball. Our numerical studies demonstrate the advantage of FLOWER achieving\nsignificantly improved performances over prior arts notably in the small base\ntasks. For further study, source codes of FLOWER, competitor algorithms and\nexperimental logs are shared publicly in\n\\url{https://github.com/anwarmaxsum/FLOWER}.\n", "date": "2023-06-26", "categories": []}
{"id": "2211.15944", "title": "The Effectiveness of World Models for Continual Reinforcement Learning", "abstract": "  World models power some of the most efficient reinforcement learning\nalgorithms. In this work, we showcase that they can be harnessed for continual\nlearning - a situation when the agent faces changing environments. World models\ntypically employ a replay buffer for training, which can be naturally extended\nto continual learning. We systematically study how different selective\nexperience replay methods affect performance, forgetting, and transfer. We also\nprovide recommendations regarding various modeling options for using world\nmodels. The best set of choices is called Continual-Dreamer, it is\ntask-agnostic and utilizes the world model for continual exploration.\nContinual-Dreamer is sample efficient and outperforms state-of-the-art\ntask-agnostic continual reinforcement learning methods on Minigrid and Minihack\nbenchmarks.\n", "date": "2022-11-29", "categories": []}
{"id": "2307.06877", "title": "The complexity of non-stationary reinforcement learning", "abstract": "  The problem of continual learning in the domain of reinforcement learning,\noften called non-stationary reinforcement learning, has been identified as an\nimportant challenge to the application of reinforcement learning. We prove a\nworst-case complexity result, which we believe captures this challenge:\nModifying the probabilities or the reward of a single state-action pair in a\nreinforcement learning problem requires an amount of time almost as large as\nthe number of states in order to keep the value function up to date, unless the\nstrong exponential time hypothesis (SETH) is false; SETH is a widely accepted\nstrengthening of the P $\\neq$ NP conjecture. Recall that the number of states\nin current applications of reinforcement learning is typically astronomical. In\ncontrast, we show that just $\\textit{adding}$ a new state-action pair is\nconsiderably easier to implement.\n", "date": "2023-07-13", "categories": []}
{"id": "2211.01413", "title": "Harnessing the Power of Explanations for Incremental Training: A\n  LIME-Based Approach", "abstract": "  Explainability of neural network prediction is essential to understand\nfeature importance and gain interpretable insight into neural network\nperformance. However, explanations of neural network outcomes are mostly\nlimited to visualization, and there is scarce work that looks to use these\nexplanations as feedback to improve model performance. In this work, model\nexplanations are fed back to the feed-forward training to help the model\ngeneralize better. To this extent, a custom weighted loss where the weights are\ngenerated by considering the Euclidean distances between true LIME (Local\nInterpretable Model-Agnostic Explanations) explanations and model-predicted\nLIME explanations is proposed. Also, in practical training scenarios,\ndeveloping a solution that can help the model learn sequentially without losing\ninformation on previous data distribution is imperative due to the\nunavailability of all the training data at once. Thus, the framework\nincorporates the custom weighted loss with Elastic Weight Consolidation (EWC)\nto maintain performance in sequential testing sets. The proposed custom\ntraining procedure results in a consistent enhancement of accuracy ranging from\n0.5% to 1.5% throughout all phases of the incremental learning setup compared\nto traditional loss-based training methods for the keyword spotting task using\nthe Google Speech Commands dataset.\n", "date": "2022-11-02", "categories": []}
{"id": "2307.05741", "title": "Towards Robust and Efficient Continual Language Learning", "abstract": "  As the application space of language models continues to evolve, a natural\nquestion to ask is how we can quickly adapt models to new tasks. We approach\nthis classic question from a continual learning perspective, in which we aim to\ncontinue fine-tuning models trained on past tasks on new tasks, with the goal\nof \"transferring\" relevant knowledge. However, this strategy also runs the risk\nof doing more harm than good, i.e., negative transfer. In this paper, we\nconstruct a new benchmark of task sequences that target different possible\ntransfer scenarios one might face, such as a sequence of tasks with high\npotential of positive transfer, high potential for negative transfer, no\nexpected effect, or a mixture of each. An ideal learner should be able to\nmaximally exploit information from all tasks that have any potential for\npositive transfer, while also avoiding the negative effects of any distracting\ntasks that may confuse it. We then propose a simple, yet effective, learner\nthat satisfies many of our desiderata simply by leveraging a selective strategy\nfor initializing new models from past task checkpoints. Still, limitations\nremain, and we hope this benchmark can help the community to further build and\nanalyze such learners.\n", "date": "2023-07-11", "categories": []}
{"id": "2304.02168", "title": "I2I: Initializing Adapters with Improvised Knowledge", "abstract": "  Adapters present a promising solution to the catastrophic forgetting problem\nin continual learning. However, training independent Adapter modules for every\nnew task misses an opportunity for cross-task knowledge transfer. We propose\nImprovise to Initialize (I2I), a continual learning algorithm that initializes\nAdapters for incoming tasks by distilling knowledge from previously-learned\ntasks' Adapters. We evaluate I2I on CLiMB, a multimodal continual learning\nbenchmark, by conducting experiments on sequences of visual question answering\ntasks. Adapters trained with I2I consistently achieve better task accuracy than\nindependently-trained Adapters, demonstrating that our algorithm facilitates\nknowledge transfer between task Adapters. I2I also results in better cross-task\nknowledge transfer than the state-of-the-art AdapterFusion without incurring\nthe associated parametric cost.\n", "date": "2023-04-04", "categories": []}
{"id": "2305.13825", "title": "Continual Learning on Dynamic Graphs via Parameter Isolation", "abstract": "  Many real-world graph learning tasks require handling dynamic graphs where\nnew nodes and edges emerge. Dynamic graph learning methods commonly suffer from\nthe catastrophic forgetting problem, where knowledge learned for previous\ngraphs is overwritten by updates for new graphs. To alleviate the problem,\ncontinual graph learning methods are proposed. However, existing continual\ngraph learning methods aim to learn new patterns and maintain old ones with the\nsame set of parameters of fixed size, and thus face a fundamental tradeoff\nbetween both goals. In this paper, we propose Parameter Isolation GNN (PI-GNN)\nfor continual learning on dynamic graphs that circumvents the tradeoff via\nparameter isolation and expansion. Our motivation lies in that different\nparameters contribute to learning different graph patterns. Based on the idea,\nwe expand model parameters to continually learn emerging graph patterns.\nMeanwhile, to effectively preserve knowledge for unaffected patterns, we find\nparameters that correspond to them via optimization and freeze them to prevent\nthem from being rewritten. Experiments on eight real-world datasets corroborate\nthe effectiveness of PI-GNN compared to state-of-the-art baselines.\n", "date": "2023-05-23", "categories": []}
{"id": "2307.05399", "title": "Domain-Agnostic Neural Architecture for Class Incremental Continual\n  Learning in Document Processing Platform", "abstract": "  Production deployments in complex systems require ML architectures to be\nhighly efficient and usable against multiple tasks. Particularly demanding are\nclassification problems in which data arrives in a streaming fashion and each\nclass is presented separately. Recent methods with stochastic gradient learning\nhave been shown to struggle in such setups or have limitations like memory\nbuffers, and being restricted to specific domains that disable its usage in\nreal-world scenarios. For this reason, we present a fully differentiable\narchitecture based on the Mixture of Experts model, that enables the training\nof high-performance classifiers when examples from each class are presented\nseparately. We conducted exhaustive experiments that proved its applicability\nin various domains and ability to learn online in production environments. The\nproposed technique achieves SOTA results without a memory buffer and clearly\noutperforms the reference methods.\n", "date": "2023-07-11", "categories": []}
{"id": "2209.02370", "title": "Continual Learning, Fast and Slow", "abstract": "  According to the Complementary Learning Systems (CLS)\ntheory~\\cite{mcclelland1995there} in neuroscience, humans do effective\n\\emph{continual learning} through two complementary systems: a fast learning\nsystem centered on the hippocampus for rapid learning of the specifics,\nindividual experiences; and a slow learning system located in the neocortex for\nthe gradual acquisition of structured knowledge about the environment.\nMotivated by this theory, we propose \\emph{DualNets} (for Dual Networks), a\ngeneral continual learning framework comprising a fast learning system for\nsupervised learning of pattern-separated representation from specific tasks and\na slow learning system for representation learning of task-agnostic general\nrepresentation via Self-Supervised Learning (SSL). DualNets can seamlessly\nincorporate both representation types into a holistic framework to facilitate\nbetter continual learning in deep neural networks. Via extensive experiments,\nwe demonstrate the promising results of DualNets on a wide range of continual\nlearning protocols, ranging from the standard offline, task-aware setting to\nthe challenging online, task-free scenario. Notably, on the\nCTrL~\\cite{veniat2020efficient} benchmark that has unrelated tasks with vastly\ndifferent visual images, DualNets can achieve competitive performance with\nexisting state-of-the-art dynamic architecture\nstrategies~\\cite{ostapenko2021continual}. Furthermore, we conduct comprehensive\nablation studies to validate DualNets efficacy, robustness, and scalability.\nCode will be made available at \\url{https://github.com/phquang/DualNet}.\n", "date": "2022-09-06", "categories": []}
{"id": "2307.04094", "title": "Class-Incremental Mixture of Gaussians for Deep Continual Learning", "abstract": "  Continual learning models for stationary data focus on learning and retaining\nconcepts coming to them in a sequential manner. In the most generic\nclass-incremental environment, we have to be ready to deal with classes coming\none by one, without any higher-level grouping. This requirement invalidates\nmany previously proposed methods and forces researchers to look for more\nflexible alternative approaches. In this work, we follow the idea of\ncentroid-driven methods and propose end-to-end incorporation of the mixture of\nGaussians model into the continual learning framework. By employing the\ngradient-based approach and designing losses capable of learning discriminative\nfeatures while avoiding degenerate solutions, we successfully combine the\nmixture model with a deep feature extractor allowing for joint optimization and\nadjustments in the latent space. Additionally, we show that our model can\neffectively learn in memory-free scenarios with fixed extractors. In the\nconducted experiments, we empirically demonstrate the effectiveness of the\nproposed solutions and exhibit the competitiveness of our model when compared\nwith state-of-the-art continual learning baselines evaluated in the context of\nimage classification problems.\n", "date": "2023-07-09", "categories": []}
{"id": "2303.13752", "title": "Leveraging Old Knowledge to Continually Learn New Classes in Medical\n  Images", "abstract": "  Class-incremental continual learning is a core step towards developing\nartificial intelligence systems that can continuously adapt to changes in the\nenvironment by learning new concepts without forgetting those previously\nlearned. This is especially needed in the medical domain where continually\nlearning from new incoming data is required to classify an expanded set of\ndiseases. In this work, we focus on how old knowledge can be leveraged to learn\nnew classes without catastrophic forgetting. We propose a framework that\ncomprises of two main components: (1) a dynamic architecture with expanding\nrepresentations to preserve previously learned features and accommodate new\nfeatures; and (2) a training procedure alternating between two objectives to\nbalance the learning of new features while maintaining the model's performance\non old classes. Experiment results on multiple medical datasets show that our\nsolution is able to achieve superior performance over state-of-the-art\nbaselines in terms of class accuracy and forgetting.\n", "date": "2023-03-24", "categories": []}
{"id": "2307.02641", "title": "Active Class Selection for Few-Shot Class-Incremental Learning", "abstract": "  For real-world applications, robots will need to continually learn in their\nenvironments through limited interactions with their users. Toward this,\nprevious works in few-shot class incremental learning (FSCIL) and active class\nselection (ACS) have achieved promising results but were tested in constrained\nsetups. Therefore, in this paper, we combine ideas from FSCIL and ACS to\ndevelop a novel framework that can allow an autonomous agent to continually\nlearn new objects by asking its users to label only a few of the most\ninformative objects in the environment. To this end, we build on a\nstate-of-the-art (SOTA) FSCIL model and extend it with techniques from ACS\nliterature. We term this model Few-shot Incremental Active class SeleCtiOn\n(FIASco). We further integrate a potential field-based navigation technique\nwith our model to develop a complete framework that can allow an agent to\nprocess and reason on its sensory data through the FIASco model, navigate\ntowards the most informative object in the environment, gather data about the\nobject through its sensors and incrementally update the FIASco model.\nExperimental results on a simulated agent and a real robot show the\nsignificance of our approach for long-term real-world robotics applications.\n", "date": "2023-07-05", "categories": []}
{"id": "2209.06606", "title": "PlaStIL: Plastic and Stable Memory-Free Class-Incremental Learning", "abstract": "  Plasticity and stability are needed in class-incremental learning in order to\nlearn from new data while preserving past knowledge. Due to catastrophic\nforgetting, finding a compromise between these two properties is particularly\nchallenging when no memory buffer is available. Mainstream methods need to\nstore two deep models since they integrate new classes using fine-tuning with\nknowledge distillation from the previous incremental state. We propose a method\nwhich has similar number of parameters but distributes them differently in\norder to find a better balance between plasticity and stability. Following an\napproach already deployed by transfer-based incremental methods, we freeze the\nfeature extractor after the initial state. Classes in the oldest incremental\nstates are trained with this frozen extractor to ensure stability. Recent\nclasses are predicted using partially fine-tuned models in order to introduce\nplasticity. Our proposed plasticity layer can be incorporated to any\ntransfer-based method designed for exemplar-free incremental learning, and we\napply it to two such methods. Evaluation is done with three large-scale\ndatasets. Results show that performance gains are obtained in all tested\nconfigurations compared to existing methods.\n", "date": "2022-09-14", "categories": []}
{"id": "2210.10209", "title": "Exclusive Supermask Subnetwork Training for Continual Learning", "abstract": "  Continual Learning (CL) methods focus on accumulating knowledge over time\nwhile avoiding catastrophic forgetting. Recently, Wortsman et al. (2020)\nproposed a CL method, SupSup, which uses a randomly initialized, fixed base\nnetwork (model) and finds a supermask for each new task that selectively keeps\nor removes each weight to produce a subnetwork. They prevent forgetting as the\nnetwork weights are not being updated. Although there is no forgetting, the\nperformance of SupSup is sub-optimal because fixed weights restrict its\nrepresentational power. Furthermore, there is no accumulation or transfer of\nknowledge inside the model when new tasks are learned. Hence, we propose\nExSSNeT (Exclusive Supermask SubNEtwork Training), that performs exclusive and\nnon-overlapping subnetwork weight training. This avoids conflicting updates to\nthe shared weights by subsequent tasks to improve performance while still\npreventing forgetting. Furthermore, we propose a novel KNN-based Knowledge\nTransfer (KKT) module that utilizes previously acquired knowledge to learn new\ntasks better and faster. We demonstrate that ExSSNeT outperforms strong\nprevious methods on both NLP and Vision domains while preventing forgetting.\nMoreover, ExSSNeT is particularly advantageous for sparse masks that activate\n2-10% of the model parameters, resulting in an average improvement of 8.3% over\nSupSup. Furthermore, ExSSNeT scales to a large number of tasks (100). Our code\nis available at https://github.com/prateeky2806/exessnet.\n", "date": "2022-10-18", "categories": []}
{"id": "2307.02246", "title": "S3C: Self-Supervised Stochastic Classifiers for Few-Shot\n  Class-Incremental Learning", "abstract": "  Few-shot class-incremental learning (FSCIL) aims to learn progressively about\nnew classes with very few labeled samples, without forgetting the knowledge of\nalready learnt classes. FSCIL suffers from two major challenges: (i)\nover-fitting on the new classes due to limited amount of data, (ii)\ncatastrophically forgetting about the old classes due to unavailability of data\nfrom these classes in the incremental stages. In this work, we propose a\nself-supervised stochastic classifier (S3C) to counter both these challenges in\nFSCIL. The stochasticity of the classifier weights (or class prototypes) not\nonly mitigates the adverse effect of absence of large number of samples of the\nnew classes, but also the absence of samples from previously learnt classes\nduring the incremental steps. This is complemented by the self-supervision\ncomponent, which helps to learn features from the base classes which generalize\nwell to unseen classes that are encountered in future, thus reducing\ncatastrophic forgetting. Extensive evaluation on three benchmark datasets using\nmultiple evaluation metrics show the effectiveness of the proposed framework.\nWe also experiment on two additional realistic scenarios of FSCIL, namely where\nthe number of annotated data available for each of the new classes can be\ndifferent, and also where the number of base classes is much lesser, and show\nthat the proposed S3C performs significantly better than the state-of-the-art\nfor all these challenging scenarios.\n", "date": "2023-07-05", "categories": []}
{"id": "2210.08710", "title": "Handling Label Uncertainty for Camera Incremental Person\n  Re-Identification", "abstract": "  Incremental learning for person re-identification (ReID) aims to develop\nmodels that can be trained with a continuous data stream, which is a more\npractical setting for real-world applications. However, the existing\nincremental ReID methods make two strong assumptions that the cameras are fixed\nand the new-emerging data is class-disjoint from previous classes. This is\nunrealistic as previously observed pedestrians may re-appear and be captured\nagain by new cameras. In this paper, we investigate person ReID in an\nunexplored scenario named Camera Incremental Person ReID (CIPR), which advances\nexisting lifelong person ReID by taking into account the class overlap issue.\nSpecifically, new data collected from new cameras may probably contain an\nunknown proportion of identities seen before. This subsequently leads to the\nlack of cross-camera annotations for new data due to privacy concerns. To\naddress these challenges, we propose a novel framework ExtendOVA. First, to\nhandle the class overlap issue, we introduce an instance-wise seen-class\nidentification module to discover previously seen identities at the instance\nlevel. Then, we propose a criterion for selecting confident ID-wise candidates\nand also devise an early learning regularization term to correct noise issues\nin pseudo labels. Furthermore, to compensate for the lack of previous data, we\nresort prototypical memory bank to create surrogate features, along with a\ncross-camera distillation loss to further retain the inter-camera relationship.\nThe comprehensive experimental results on multiple benchmarks show that\nExtendOVA significantly outperforms the state-of-the-arts with remarkable\nadvantages.\n", "date": "2022-10-17", "categories": []}
{"id": "2211.15875", "title": "Data Poisoning Attack Aiming the Vulnerability of Continual Learning", "abstract": "  Generally, regularization-based continual learning models limit access to the\nprevious task data to imitate the real-world constraints related to memory and\nprivacy. However, this introduces a problem in these models by not being able\nto track the performance on each task. In essence, current continual learning\nmethods are susceptible to attacks on previous tasks. We demonstrate the\nvulnerability of regularization-based continual learning methods by presenting\na simple task-specific data poisoning attack that can be used in the learning\nprocess of a new task. Training data generated by the proposed attack causes\nperformance degradation on a specific task targeted by the attacker. We\nexperiment with the attack on the two representative regularization-based\ncontinual learning methods, Elastic Weight Consolidation (EWC) and Synaptic\nIntelligence (SI), trained with variants of MNIST dataset. The experiment\nresults justify the vulnerability proposed in this paper and demonstrate the\nimportance of developing continual learning models that are robust to\nadversarial attacks.\n", "date": "2022-11-29", "categories": []}
{"id": "2306.16817", "title": "Improving Online Continual Learning Performance and Stability with\n  Temporal Ensembles", "abstract": "  Neural networks are very effective when trained on large datasets for a large\nnumber of iterations. However, when they are trained on non-stationary streams\nof data and in an online fashion, their performance is reduced (1) by the\nonline setup, which limits the availability of data, (2) due to catastrophic\nforgetting because of the non-stationary nature of the data. Furthermore,\nseveral recent works (Caccia et al., 2022; Lange et al., 2023) arXiv:2205.13452\nshowed that replay methods used in continual learning suffer from the stability\ngap, encountered when evaluating the model continually (rather than only on\ntask boundaries). In this article, we study the effect of model ensembling as a\nway to improve performance and stability in online continual learning. We\nnotice that naively ensembling models coming from a variety of training tasks\nincreases the performance in online continual learning considerably. Starting\nfrom this observation, and drawing inspirations from semi-supervised learning\nensembling methods, we use a lightweight temporal ensemble that computes the\nexponential moving average of the weights (EMA) at test time, and show that it\ncan drastically increase the performance and stability when used in combination\nwith several methods from the literature.\n", "date": "2023-06-29", "categories": []}
{"id": "2306.17091", "title": "The Importance of Robust Features in Mitigating Catastrophic Forgetting", "abstract": "  Continual learning (CL) is an approach to address catastrophic forgetting,\nwhich refers to forgetting previously learned knowledge by neural networks when\ntrained on new tasks or data distributions. The adversarial robustness has\ndecomposed features into robust and non-robust types and demonstrated that\nmodels trained on robust features significantly enhance adversarial robustness.\nHowever, no study has been conducted on the efficacy of robust features from\nthe lens of the CL model in mitigating catastrophic forgetting in CL. In this\npaper, we introduce the CL robust dataset and train four baseline models on\nboth the standard and CL robust datasets. Our results demonstrate that the CL\nmodels trained on the CL robust dataset experienced less catastrophic\nforgetting of the previously learned tasks than when trained on the standard\ndataset. Our observations highlight the significance of the features provided\nto the underlying CL models, showing that CL robust features can alleviate\ncatastrophic forgetting.\n", "date": "2023-06-29", "categories": []}
{"id": "2212.14193", "title": "A Unified Object Counting Network with Object Occupation Prior", "abstract": "  The counting task, which plays a fundamental role in numerous applications\n(e.g., crowd counting, traffic statistics), aims to predict the number of\nobjects with various densities. Existing object counting tasks are designed for\na single object class. However, it is inevitable to encounter newly coming data\nwith new classes in our real world. We name this scenario as \\textit{evolving\nobject counting}. In this paper, we build the first evolving object counting\ndataset and propose a unified object counting network as the first attempt to\naddress this task. The proposed model consists of two key components: a\nclass-agnostic mask module and a class-incremental module. The class-agnostic\nmask module learns generic object occupation prior via predicting a\nclass-agnostic binary mask (e.g., 1 denotes there exists an object at the\nconsidering position in an image and 0 otherwise). The class-incremental module\nis used to handle new coming classes and provides discriminative class guidance\nfor density map prediction. The combined outputs of class-agnostic mask module\nand image feature extractor are used to predict the final density map. When new\nclasses come, we first add new neural nodes into the last regression and\nclassification layers of class-incremental module. Then, instead of retraining\nthe model from scratch, we utilize knowledge distillation to help the model\nremember what have already learned about previous object classes. We also\nemploy a support sample bank to store a small number of typical training\nsamples of each class, which are used to prevent the model from forgetting key\ninformation of old data. With this design, our model can efficiently and\neffectively adapt to new coming classes while keeping good performance on\nalready seen data without large-scale retraining. Extensive experiments on the\ncollected dataset demonstrate the favorable performance.\n", "date": "2022-12-29", "categories": []}
{"id": "2301.05032", "title": "Online Hyperparameter Optimization for Class-Incremental Learning", "abstract": "  Class-incremental learning (CIL) aims to train a classification model while\nthe number of classes increases phase-by-phase. An inherent challenge of CIL is\nthe stability-plasticity tradeoff, i.e., CIL models should keep stable to\nretain old knowledge and keep plastic to absorb new knowledge. However, none of\nthe existing CIL models can achieve the optimal tradeoff in different\ndata-receiving settings--where typically the training-from-half (TFH) setting\nneeds more stability, but the training-from-scratch (TFS) needs more\nplasticity. To this end, we design an online learning method that can\nadaptively optimize the tradeoff without knowing the setting as a priori.\nSpecifically, we first introduce the key hyperparameters that influence the\ntrade-off, e.g., knowledge distillation (KD) loss weights, learning rates, and\nclassifier types. Then, we formulate the hyperparameter optimization process as\nan online Markov Decision Process (MDP) problem and propose a specific\nalgorithm to solve it. We apply local estimated rewards and a classic bandit\nalgorithm Exp3 to address the issues when applying online MDP methods to the\nCIL protocol. Our method consistently improves top-performing CIL methods in\nboth TFH and TFS settings, e.g., boosting the average accuracy of TFH and TFS\nby 2.2 percentage points on ImageNet-Full, compared to the state-of-the-art.\n", "date": "2023-01-11", "categories": []}
{"id": "2301.12467", "title": "Continual Learning for Predictive Maintenance: Overview and Challenges", "abstract": "  Deep learning techniques have become one of the main propellers for solving\nengineering problems effectively and efficiently. For instance, Predictive\nMaintenance methods have been used to improve predictions of when maintenance\nis needed on different machines and operative contexts. However, deep learning\nmethods are not without limitations, as these models are normally trained on a\nfixed distribution that only reflects the current state of the problem. Due to\ninternal or external factors, the state of the problem can change, and the\nperformance decreases due to the lack of generalization and adaptation.\nContrary to this stationary training set, real-world applications change their\nenvironments constantly, creating the need to constantly adapt the model to\nevolving scenarios. To aid in this endeavor, Continual Learning methods propose\nways to constantly adapt prediction models and incorporate new knowledge after\ndeployment. Despite the advantages of these techniques, there are still\nchallenges to applying them to real-world problems. In this work, we present a\nbrief introduction to predictive maintenance, non-stationary environments, and\ncontinual learning, together with an extensive review of the current state of\napplying continual learning in real-world applications and specifically in\npredictive maintenance. We then discuss the current challenges of both\npredictive maintenance and continual learning, proposing future directions at\nthe intersection of both areas. Finally, we propose a novel way to create\nbenchmarks that favor the application of continuous learning methods in more\nrealistic environments, giving specific examples of predictive maintenance.\n", "date": "2023-01-29", "categories": []}
{"id": "2303.14595", "title": "Preserving Linear Separability in Continual Learning by Backward Feature\n  Projection", "abstract": "  Catastrophic forgetting has been a major challenge in continual learning,\nwhere the model needs to learn new tasks with limited or no access to data from\npreviously seen tasks. To tackle this challenge, methods based on knowledge\ndistillation in feature space have been proposed and shown to reduce\nforgetting. However, most feature distillation methods directly constrain the\nnew features to match the old ones, overlooking the need for plasticity. To\nachieve a better stability-plasticity trade-off, we propose Backward Feature\nProjection (BFP), a method for continual learning that allows the new features\nto change up to a learnable linear transformation of the old features. BFP\npreserves the linear separability of the old classes while allowing the\nemergence of new feature directions to accommodate new classes. BFP can be\nintegrated with existing experience replay methods and boost performance by a\nsignificant margin. We also demonstrate that BFP helps learn a better\nrepresentation space, in which linear separability is well preserved during\ncontinual learning and linear probing achieves high classification accuracy.\nThe code can be found at https://github.com/rvl-lab-utoronto/BFP\n", "date": "2023-03-26", "categories": []}
{"id": "2305.18418", "title": "Just a Glimpse: Rethinking Temporal Information for Video Continual\n  Learning", "abstract": "  Class-incremental learning is one of the most important settings for the\nstudy of Continual Learning, as it closely resembles real-world application\nscenarios. With constrained memory sizes, catastrophic forgetting arises as the\nnumber of classes/tasks increases. Studying continual learning in the video\ndomain poses even more challenges, as video data contains a large number of\nframes, which places a higher burden on the replay memory. The current common\npractice is to sub-sample frames from the video stream and store them in the\nreplay memory. In this paper, we propose SMILE a novel replay mechanism for\neffective video continual learning based on individual/single frames. Through\nextensive experimentation, we show that under extreme memory constraints, video\ndiversity plays a more significant role than temporal information. Therefore,\nour method focuses on learning from a small number of frames that represent a\nlarge number of unique videos. On three representative video datasets,\nKinetics, UCF101, and ActivityNet, the proposed method achieves\nstate-of-the-art performance, outperforming the previous state-of-the-art by up\nto 21.49%.\n", "date": "2023-05-28", "categories": []}
{"id": "2306.15347", "title": "FedET: A Communication-Efficient Federated Class-Incremental Learning\n  Framework Based on Enhanced Transformer", "abstract": "  Federated Learning (FL) has been widely concerned for it enables\ndecentralized learning while ensuring data privacy. However, most existing\nmethods unrealistically assume that the classes encountered by local clients\nare fixed over time. After learning new classes, this assumption will make the\nmodel's catastrophic forgetting of old classes significantly severe. Moreover,\ndue to the limitation of communication cost, it is challenging to use\nlarge-scale models in FL, which will affect the prediction accuracy. To address\nthese challenges, we propose a novel framework, Federated Enhanced Transformer\n(FedET), which simultaneously achieves high accuracy and low communication\ncost. Specifically, FedET uses Enhancer, a tiny module, to absorb and\ncommunicate new knowledge, and applies pre-trained Transformers combined with\ndifferent Enhancers to ensure high precision on various tasks. To address local\nforgetting caused by new classes of new tasks and global forgetting brought by\nnon-i.i.d (non-independent and identically distributed) class imbalance across\ndifferent local clients, we proposed an Enhancer distillation method to modify\nthe imbalance between old and new knowledge and repair the non-i.i.d. problem.\nExperimental results demonstrate that FedET's average accuracy on\nrepresentative benchmark datasets is 14.1% higher than the state-of-the-art\nmethod, while FedET saves 90% of the communication cost compared to the\nprevious method.\n", "date": "2023-06-27", "categories": []}
{"id": "2306.14775", "title": "Parameter-Level Soft-Masking for Continual Learning", "abstract": "  Existing research on task incremental learning in continual learning has\nprimarily focused on preventing catastrophic forgetting (CF). Although several\ntechniques have achieved learning with no CF, they attain it by letting each\ntask monopolize a sub-network in a shared network, which seriously limits\nknowledge transfer (KT) and causes over-consumption of the network capacity,\ni.e., as more tasks are learned, the performance deteriorates. The goal of this\npaper is threefold: (1) overcoming CF, (2) encouraging KT, and (3) tackling the\ncapacity problem. A novel technique (called SPG) is proposed that soft-masks\n(partially blocks) parameter updating in training based on the importance of\neach parameter to old tasks. Each task still uses the full network, i.e., no\nmonopoly of any part of the network by any task, which enables maximum KT and\nreduction in capacity usage. To our knowledge, this is the first work that\nsoft-masks a model at the parameter-level for continual learning. Extensive\nexperiments demonstrate the effectiveness of SPG in achieving all three\nobjectives. More notably, it attains significant transfer of knowledge not only\namong similar tasks (with shared knowledge) but also among dissimilar tasks\n(with little shared knowledge) while mitigating CF.\n", "date": "2023-06-26", "categories": []}
{"id": "2306.13410", "title": "Explainable Lifelong Stream Learning Based on \"Glocal\" Pairwise Fusion", "abstract": "  Real-time on-device continual learning applications are used on mobile\nphones, consumer robots, and smart appliances. Such devices have limited\nprocessing and memory storage capabilities, whereas continual learning acquires\ndata over a long period of time. By necessity, lifelong learning algorithms\nhave to be able to operate under such constraints while delivering good\nperformance. This study presents the Explainable Lifelong Learning (ExLL)\nmodel, which incorporates several important traits: 1) learning to learn, in a\nsingle pass, from streaming data with scarce examples and resources; 2) a\nself-organizing prototype-based architecture that expands as needed and\nclusters streaming data into separable groups by similarity and preserves data\nagainst catastrophic forgetting; 3) an interpretable architecture to convert\nthe clusters into explainable IF-THEN rules as well as to justify model\npredictions in terms of what is similar and dissimilar to the inference; and 4)\ninferences at the global and local level using a pairwise decision fusion\nprocess to enhance the accuracy of the inference, hence ``Glocal Pairwise\nFusion.'' We compare ExLL against contemporary online learning algorithms for\nimage recognition, using OpenLoris, F-SIOL-310, and Places datasets to evaluate\nseveral continual learning scenarios for video streams, low-sample learning,\nability to scale, and imbalanced data streams. The algorithms are evaluated for\ntheir performance in accuracy, number of parameters, and experiment runtime\nrequirements. ExLL outperforms all algorithms for accuracy in the majority of\nthe tested scenarios.\n", "date": "2023-06-23", "categories": []}
{"id": "2306.12646", "title": "Learnability and Algorithm for Continual Learning", "abstract": "  This paper studies the challenging continual learning (CL) setting of Class\nIncremental Learning (CIL). CIL learns a sequence of tasks consisting of\ndisjoint sets of concepts or classes. At any time, a single model is built that\ncan be applied to predict/classify test instances of any classes learned thus\nfar without providing any task related information for each test instance.\nAlthough many techniques have been proposed for CIL, they are mostly empirical.\nIt has been shown recently that a strong CIL system needs a strong within-task\nprediction (WP) and a strong out-of-distribution (OOD) detection for each task.\nHowever, it is still not known whether CIL is actually learnable. This paper\nshows that CIL is learnable. Based on the theory, a new CIL algorithm is also\nproposed. Experimental results demonstrate its effectiveness.\n", "date": "2023-06-22", "categories": []}
{"id": "2306.11955", "title": "TADIL: Task-Agnostic Domain-Incremental Learning through Task-ID\n  Inference using Transformer Nearest-Centroid Embeddings", "abstract": "  Machine Learning (ML) models struggle with data that changes over time or\nacross domains due to factors such as noise, occlusion, illumination, or\nfrequency, unlike humans who can learn from such non independent and\nidentically distributed data. Consequently, a Continual Learning (CL) approach\nis indispensable, particularly, Domain-Incremental Learning. In this paper, we\npropose a novel pipeline for identifying tasks in domain-incremental learning\nscenarios without supervision. The pipeline comprises four steps. First, we\nobtain base embeddings from the raw data using an existing transformer-based\nmodel. Second, we group the embedding densities based on their similarity to\nobtain the nearest points to each cluster centroid. Third, we train an\nincremental task classifier using only these few points. Finally, we leverage\nthe lightweight computational requirements of the pipeline to devise an\nalgorithm that decides in an online fashion when to learn a new task using the\ntask classifier and a drift detector. We conduct experiments using the SODA10M\nreal-world driving dataset and several CL strategies. We demonstrate that the\nperformance of these CL strategies with our pipeline can match the ground-truth\napproach, both in classical experiments assuming task boundaries, and also in\nmore realistic task-agnostic scenarios that require detecting new tasks\non-the-fly\n", "date": "2023-06-21", "categories": []}
{"id": "2306.11967", "title": "Complementary Learning Subnetworks for Parameter-Efficient\n  Class-Incremental Learning", "abstract": "  In the scenario of class-incremental learning (CIL), deep neural networks\nhave to adapt their model parameters to non-stationary data distributions,\ne.g., the emergence of new classes over time. However, CIL models are\nchallenged by the well-known catastrophic forgetting phenomenon. Typical\nmethods such as rehearsal-based ones rely on storing exemplars of old classes\nto mitigate catastrophic forgetting, which limits real-world applications\nconsidering memory resources and privacy issues. In this paper, we propose a\nnovel rehearsal-free CIL approach that learns continually via the synergy\nbetween two Complementary Learning Subnetworks. Our approach involves jointly\noptimizing a plastic CNN feature extractor and an analytical feed-forward\nclassifier. The inaccessibility of historical data is tackled by holistically\ncontrolling the parameters of a well-trained model, ensuring that the decision\nboundary learned fits new classes while retaining recognition of previously\nlearned classes. Specifically, the trainable CNN feature extractor provides\ntask-dependent knowledge separately without interference; and the final\nclassifier integrates task-specific knowledge incrementally for decision-making\nwithout forgetting. In each CIL session, it accommodates new tasks by attaching\na tiny set of declarative parameters to its backbone, in which only one matrix\nper task or one vector per class is kept for knowledge retention. Extensive\nexperiments on a variety of task sequences show that our method achieves\ncompetitive results against state-of-the-art methods, especially in accuracy\ngain, memory cost, training efficiency, and task-order robustness. Furthermore,\nto make the non-growing backbone (i.e., a model with limited network capacity)\nsuffice to train on more incoming tasks, a graceful forgetting implementation\non previously learned trivial tasks is empirically investigated.\n", "date": "2023-06-21", "categories": []}
{"id": "2306.12026", "title": "Continual Learners are Incremental Model Generalizers", "abstract": "  Motivated by the efficiency and rapid convergence of pre-trained models for\nsolving downstream tasks, this paper extensively studies the impact of\nContinual Learning (CL) models as pre-trainers. In both supervised and\nunsupervised CL, we find that the transfer quality of the representation often\nincreases gradually without noticeable degradation in fine-tuning performance.\nThis is because CL models can learn improved task-general features when easily\nforgetting task-specific knowledge. Based on this observation, we suggest a new\nunsupervised CL framework with masked modeling, which aims to capture fluent\ntask-generic representation during training. Furthermore, we propose a new\nfine-tuning scheme, GLobal Attention Discretization (GLAD), that preserves rich\ntask-generic representation during solving downstream tasks. The model\nfine-tuned with GLAD achieves competitive performance and can also be used as a\ngood pre-trained model itself. We believe this paper breaks the barriers\nbetween pre-training and fine-tuning steps and leads to a sustainable learning\nframework in which the continual learner incrementally improves model\ngeneralization, yielding better transfer to unseen tasks.\n", "date": "2023-06-21", "categories": []}
{"id": "2301.11396", "title": "Class-Incremental Learning with Repetition", "abstract": "  Real-world data streams naturally include the repetition of previous\nconcepts. From a Continual Learning (CL) perspective, repetition is a property\nof the environment and, unlike replay, cannot be controlled by the agent.\nNowadays, the Class-Incremental (CI) scenario represents the leading test-bed\nfor assessing and comparing CL strategies. This scenario type is very easy to\nuse, but it never allows revisiting previously seen classes, thus completely\nneglecting the role of repetition. We focus on the family of Class-Incremental\nwith Repetition (CIR) scenario, where repetition is embedded in the definition\nof the stream. We propose two stochastic stream generators that produce a wide\nrange of CIR streams starting from a single dataset and a few interpretable\ncontrol parameters. We conduct the first comprehensive evaluation of repetition\nin CL by studying the behavior of existing CL strategies under different CIR\nstreams. We then present a novel replay strategy that exploits repetition and\ncounteracts the natural imbalance present in the stream. On both CIFAR100 and\nTinyImageNet, our strategy outperforms other replay approaches, which are not\ndesigned for environments with repetition.\n", "date": "2023-01-26", "categories": []}
{"id": "2306.09890", "title": "Studying Generalization on Memory-Based Methods in Continual Learning", "abstract": "  One of the objectives of Continual Learning is to learn new concepts\ncontinually over a stream of experiences and at the same time avoid\ncatastrophic forgetting. To mitigate complete knowledge overwriting,\nmemory-based methods store a percentage of previous data distributions to be\nused during training. Although these methods produce good results, few studies\nhave tested their out-of-distribution generalization properties, as well as\nwhether these methods overfit the replay memory. In this work, we show that\nalthough these methods can help in traditional in-distribution generalization,\nthey can strongly impair out-of-distribution generalization by learning\nspurious features and correlations. Using a controlled environment, the Synbol\nbenchmark generator (Lacoste et al., 2020), we demonstrate that this lack of\nout-of-distribution generalization mainly occurs in the linear classifier.\n", "date": "2023-06-16", "categories": []}
{"id": "2306.10181", "title": "Catastrophic Forgetting in the Context of Model Updates", "abstract": "  A large obstacle to deploying deep learning models in practice is the process\nof updating models post-deployment (ideally, frequently). Deep neural networks\ncan cost many thousands of dollars to train. When new data comes in the\npipeline, you can train a new model from scratch (randomly initialized weights)\non all existing data. Instead, you can take an existing model and fine-tune\n(continue to train) it on new data. The former is costly and slow. The latter\nis cheap and fast, but catastrophic forgetting generally causes the new model\nto 'forget' how to classify older data well. There are a plethora of\ncomplicated techniques to keep models from forgetting their past learnings.\nArguably the most basic is to mix in a small amount of past data into the new\ndata during fine-tuning: also known as 'data rehearsal'. In this paper, we\ncompare various methods of limiting catastrophic forgetting and conclude that\nif you can maintain access to a portion of your past data (or tasks), data\nrehearsal is ideal in terms of overall accuracy across all time periods, and\nperforms even better when combined with methods like Elastic Weight\nConsolidation (EWC). Especially when the amount of past data (past 'tasks') is\nlarge compared to new data, the cost of updating an existing model is far\ncheaper and faster than training a new model from scratch.\n", "date": "2023-06-16", "categories": []}
{"id": "2306.10480", "title": "IF2Net: Innately Forgetting-Free Networks for Continual Learning", "abstract": "  Continual learning can incrementally absorb new concepts without interfering\nwith previously learned knowledge. Motivated by the characteristics of neural\nnetworks, in which information is stored in weights on connections, we\ninvestigated how to design an Innately Forgetting-Free Network (IF2Net) for\ncontinual learning context. This study proposed a straightforward yet effective\nlearning paradigm by ingeniously keeping the weights relative to each seen task\nuntouched before and after learning a new task. We first presented the novel\nrepresentation-level learning on task sequences with random weights. This\ntechnique refers to tweaking the drifted representations caused by\nrandomization back to their separate task-optimal working states, but the\ninvolved weights are frozen and reused (opposite to well-known layer-wise\nupdates of weights). Then, sequential decision-making without forgetting can be\nachieved by projecting the output weight updates into the parsimonious\northogonal space, making the adaptations not disturb old knowledge while\nmaintaining model plasticity. IF2Net allows a single network to inherently\nlearn unlimited mapping rules without telling task identities at test time by\nintegrating the respective strengths of randomization and orthogonalization. We\nvalidated the effectiveness of our approach in the extensive theoretical\nanalysis and empirical study.\n", "date": "2023-06-18", "categories": []}
{"id": "2306.10724", "title": "Partial Hypernetworks for Continual Learning", "abstract": "  Hypernetworks mitigate forgetting in continual learning (CL) by generating\ntask-dependent weights and penalizing weight changes at a meta-model level.\nUnfortunately, generating all weights is not only computationally expensive for\nlarger architectures, but also, it is not well understood whether generating\nall model weights is necessary. Inspired by latent replay methods in CL, we\npropose partial weight generation for the final layers of a model using\nhypernetworks while freezing the initial layers. With this objective, we first\nanswer the question of how many layers can be frozen without compromising the\nfinal performance. Through several experiments, we empirically show that the\nnumber of layers that can be frozen is proportional to the distributional\nsimilarity in the CL stream. Then, to demonstrate the effectiveness of\nhypernetworks, we show that noisy streams can significantly impact the\nperformance of latent replay methods, leading to increased forgetting when\nfeatures from noisy experiences are replayed with old samples. In contrast,\npartial hypernetworks are more robust to noise by maintaining accuracy on\nprevious experiences. Finally, we conduct experiments on the split CIFAR-100\nand TinyImagenet benchmarks and compare different versions of partial\nhypernetworks to latent replay methods. We conclude that partial weight\ngeneration using hypernetworks is a promising solution to the problem of\nforgetting in neural networks. It can provide an effective balance between\ncomputation and final test accuracy in CL streams.\n", "date": "2023-06-19", "categories": []}
{"id": "2306.10942", "title": "Knowledge Transfer-Driven Few-Shot Class-Incremental Learning", "abstract": "  Few-shot class-incremental learning (FSCIL) aims to continually learn new\nclasses using a few samples while not forgetting the old classes. The key of\nthis task is effective knowledge transfer from the base session to the\nincremental sessions. Despite the advance of existing FSCIL methods, the\nproposed knowledge transfer learning schemes are sub-optimal due to the\ninsufficient optimization for the model's plasticity. To address this issue, we\npropose a Random Episode Sampling and Augmentation (RESA) strategy that relies\non diverse pseudo incremental tasks as agents to achieve the knowledge\ntransfer. Concretely, RESA mimics the real incremental setting and constructs\npseudo incremental tasks globally and locally, where the global pseudo\nincremental tasks are designed to coincide with the learning objective of FSCIL\nand the local pseudo incremental tasks are designed to improve the model's\nplasticity, respectively. Furthermore, to make convincing incremental\npredictions, we introduce a complementary model with a squared\nEuclidean-distance classifier as the auxiliary module, which couples with the\nwidely used cosine classifier to form our whole architecture. By such a way,\nequipped with model decoupling strategy, we can maintain the model's stability\nwhile enhancing the model's plasticity. Extensive quantitative and qualitative\nexperiments on three popular FSCIL benchmark datasets demonstrate that our\nproposed method, named Knowledge Transfer-driven Relation Complementation\nNetwork (KT-RCNet), outperforms almost all prior methods. More precisely, the\naverage accuracy of our proposed KT-RCNet outperforms the second-best method by\na margin of 5.26%, 3.49%, and 2.25% on miniImageNet, CIFAR100, and CUB200,\nrespectively. Our code is available at\nhttps://github.com/YeZiLaiXi/KT-RCNet.git.\n", "date": "2023-06-19", "categories": []}
{"id": "2210.12301", "title": "Continual Vision-based Reinforcement Learning with Group Symmetries", "abstract": "  Continual reinforcement learning aims to sequentially learn a variety of\ntasks, retaining the ability to perform previously encountered tasks while\nsimultaneously developing new policies for novel tasks. However, current\ncontinual RL approaches overlook the fact that certain tasks are identical\nunder basic group operations like rotations or translations, especially with\nvisual inputs. They may unnecessarily learn and maintain a new policy for each\nsimilar task, leading to poor sample efficiency and weak generalization\ncapability. To address this, we introduce a unique Continual Vision-based\nReinforcement Learning method that recognizes Group Symmetries, called COVERS,\ncultivating a policy for each group of equivalent tasks rather than individual\ntasks. COVERS employs a proximal policy optimization-based RL algorithm with an\nequivariant feature extractor and a novel task grouping mechanism that relies\non the extracted invariant features. We evaluate COVERS on sequences of\ntable-top manipulation tasks that incorporate image observations and robot\nproprioceptive information in both simulations and on real robot platforms. Our\nresults show that COVERS accurately assigns tasks to their respective groups\nand significantly outperforms existing methods in terms of generalization\ncapability.\n", "date": "2022-10-21", "categories": []}
{"id": "2303.13952", "title": "CCL: Continual Contrastive Learning for LiDAR Place Recognition", "abstract": "  Place recognition is an essential and challenging task in loop closing and\nglobal localization for robotics and autonomous driving applications.\nBenefiting from the recent advances in deep learning techniques, the\nperformance of LiDAR place recognition (LPR) has been greatly improved.\nHowever, current deep learning-based methods suffer from two major problems:\npoor generalization ability and catastrophic forgetting. In this paper, we\npropose a continual contrastive learning method, named CCL, to tackle the\ncatastrophic forgetting problem and generally improve the robustness of LPR\napproaches. Our CCL constructs a contrastive feature pool and utilizes\ncontrastive loss to train more transferable representations of places. When\ntransferred into new environments, our CCL continuously reviews the contrastive\nmemory bank and applies a distribution-based knowledge distillation to maintain\nthe retrieval ability of the past data while continually learning to recognize\nnew places from the new data. We thoroughly evaluate our approach on Oxford,\nMulRan, and PNV datasets using three different LPR methods. The experimental\nresults show that our CCL consistently improves the performance of different\nmethods in different environments outperforming the state-of-the-art continual\nlearning method. The implementation of our method has been released at\nhttps://github.com/cloudcjf/CCL.\n", "date": "2023-03-24", "categories": []}
{"id": "2306.08200", "title": "POP: Prompt Of Prompts for Continual Learning", "abstract": "  Continual learning (CL) has attracted increasing attention in the recent\npast. It aims to mimic the human ability to learn new concepts without\ncatastrophic forgetting. While existing CL methods accomplish this to some\nextent, they are still prone to semantic drift of the learned feature space.\nFoundation models, which are endowed with a robust feature representation,\nlearned from very large datasets, provide an interesting substrate for the\nsolution of the CL problem. Recent work has also shown that they can be adapted\nto specific tasks by prompt tuning techniques that leave the generality of the\nrepresentation mostly unscathed. An open question is, however, how to learn\nboth prompts that are task specific and prompts that are global, i.e. capture\ncross-task information. In this work, we propose the Prompt Of Prompts (POP)\nmodel, which addresses this goal by progressively learning a group of\ntask-specified prompts and a group of global prompts, denoted as POP, to\nintegrate information from the former. We show that a foundation model equipped\nwith POP learning is able to outperform classic CL methods by a significant\nmargin. Moreover, as prompt tuning only requires a small set of training\nsamples, POP is able to perform CL in the few-shot setting, while still\noutperforming competing methods trained on the entire dataset.\n", "date": "2023-06-14", "categories": []}
{"id": "2306.08593", "title": "Heterogeneous Continual Learning", "abstract": "  We propose a novel framework and a solution to tackle the continual learning\n(CL) problem with changing network architectures. Most CL methods focus on\nadapting a single architecture to a new task/class by modifying its weights.\nHowever, with rapid progress in architecture design, the problem of adapting\nexisting solutions to novel architectures becomes relevant. To address this\nlimitation, we propose Heterogeneous Continual Learning (HCL), where a wide\nrange of evolving network architectures emerge continually together with novel\ndata/tasks. As a solution, we build on top of the distillation family of\ntechniques and modify it to a new setting where a weaker model takes the role\nof a teacher; meanwhile, a new stronger architecture acts as a student.\nFurthermore, we consider a setup of limited access to previous data and propose\nQuick Deep Inversion (QDI) to recover prior task visual features to support\nknowledge transfer. QDI significantly reduces computational costs compared to\nprevious solutions and improves overall performance. In summary, we propose a\nnew setup for CL with a modified knowledge distillation paradigm and design a\nquick data inversion method to enhance distillation. Our evaluation of various\nbenchmarks shows a significant improvement on accuracy in comparison to\nstate-of-the-art methods over various networks architectures.\n", "date": "2023-06-14", "categories": []}
{"id": "2207.07730", "title": "How to Reuse and Compose Knowledge for a Lifetime of Tasks: A Survey on\n  Continual Learning and Functional Composition", "abstract": "  A major goal of artificial intelligence (AI) is to create an agent capable of\nacquiring a general understanding of the world. Such an agent would require the\nability to continually accumulate and build upon its knowledge as it encounters\nnew experiences. Lifelong or continual learning addresses this setting, whereby\nan agent faces a continual stream of problems and must strive to capture the\nknowledge necessary for solving each new task it encounters. If the agent is\ncapable of accumulating knowledge in some form of compositional representation,\nit could then selectively reuse and combine relevant pieces of knowledge to\nconstruct novel solutions. Despite the intuitive appeal of this simple idea,\nthe literatures on lifelong learning and compositional learning have proceeded\nlargely separately. In an effort to promote developments that bridge between\nthe two fields, this article surveys their respective research landscapes and\ndiscusses existing and future connections between them.\n", "date": "2022-07-15", "categories": []}
{"id": "2207.11240", "title": "Discrete Key-Value Bottleneck", "abstract": "  Deep neural networks perform well on classification tasks where data streams\nare i.i.d. and labeled data is abundant. Challenges emerge with non-stationary\ntraining data streams such as continual learning. One powerful approach that\nhas addressed this challenge involves pre-training of large encoders on volumes\nof readily available data, followed by task-specific tuning. Given a new task,\nhowever, updating the weights of these encoders is challenging as a large\nnumber of weights needs to be fine-tuned, and as a result, they forget\ninformation about the previous tasks. In the present work, we propose a model\narchitecture to address this issue, building upon a discrete bottleneck\ncontaining pairs of separate and learnable key-value codes. Our paradigm will\nbe to encode; process the representation via a discrete bottleneck; and decode.\nHere, the input is fed to the pre-trained encoder, the output of the encoder is\nused to select the nearest keys, and the corresponding values are fed to the\ndecoder to solve the current task. The model can only fetch and re-use a sparse\nnumber of these key-value pairs during inference, enabling localized and\ncontext-dependent model updates. We theoretically investigate the ability of\nthe discrete key-value bottleneck to minimize the effect of learning under\ndistribution shifts and show that it reduces the complexity of the hypothesis\nclass. We empirically verify the proposed method under challenging\nclass-incremental learning scenarios and show that the proposed model - without\nany task boundaries - reduces catastrophic forgetting across a wide variety of\npre-trained models, outperforming relevant baselines on this task.\n", "date": "2022-07-22", "categories": []}
{"id": "2303.14115", "title": "Principles of Forgetting in Domain-Incremental Semantic Segmentation in\n  Adverse Weather Conditions", "abstract": "  Deep neural networks for scene perception in automated vehicles achieve\nexcellent results for the domains they were trained on. However, in real-world\nconditions, the domain of operation and its underlying data distribution are\nsubject to change. Adverse weather conditions, in particular, can significantly\ndecrease model performance when such data are not available during\ntraining.Additionally, when a model is incrementally adapted to a new domain,\nit suffers from catastrophic forgetting, causing a significant drop in\nperformance on previously observed domains. Despite recent progress in reducing\ncatastrophic forgetting, its causes and effects remain obscure. Therefore, we\nstudy how the representations of semantic segmentation models are affected\nduring domain-incremental learning in adverse weather conditions. Our\nexperiments and representational analyses indicate that catastrophic forgetting\nis primarily caused by changes to low-level features in domain-incremental\nlearning and that learning more general features on the source domain using\npre-training and image augmentations leads to efficient feature reuse in\nsubsequent tasks, which drastically reduces catastrophic forgetting. These\nfindings highlight the importance of methods that facilitate generalized\nfeatures for effective continual learning algorithms.\n", "date": "2023-03-24", "categories": []}
{"id": "2305.00316", "title": "The Ideal Continual Learner: An Agent That Never Forgets", "abstract": "  The goal of continual learning is to find a model that solves multiple\nlearning tasks which are presented sequentially to the learner. A key challenge\nin this setting is that the learner may forget how to solve a previous task\nwhen learning a new task, a phenomenon known as catastrophic forgetting. To\naddress this challenge, many practical methods have been proposed, including\nmemory-based, regularization-based, and expansion-based methods. However, a\nrigorous theoretical understanding of these methods remains elusive. This paper\naims to bridge this gap between theory and practice by proposing a new\ncontinual learning framework called Ideal Continual Learner (ICL), which is\nguaranteed to avoid catastrophic forgetting by construction. We show that ICL\nunifies multiple well-established continual learning methods and gives new\ntheoretical insights into the strengths and weaknesses of these methods. We\nalso derive generalization bounds for ICL which allow us to theoretically\nquantify how rehearsal affects generalization. Finally, we connect ICL to\nseveral classic subjects and research topics of modern interest, which allows\nus to make historical remarks and inspire future directions.\n", "date": "2023-04-29", "categories": []}
{"id": "2306.02947", "title": "Continual Learning with Pretrained Backbones by Tuning in the Input\n  Space", "abstract": "  The intrinsic difficulty in adapting deep learning models to non-stationary\nenvironments limits the applicability of neural networks to real-world tasks.\nThis issue is critical in practical supervised learning settings, such as the\nones in which a pre-trained model computes projections toward a latent space\nwhere different task predictors are sequentially learned over time. As a matter\nof fact, incrementally fine-tuning the whole model to better adapt to new tasks\nusually results in catastrophic forgetting, with decreasing performance over\nthe past experiences and losing valuable knowledge from the pre-training stage.\nIn this paper, we propose a novel strategy to make the fine-tuning procedure\nmore effective, by avoiding to update the pre-trained part of the network and\nlearning not only the usual classification head, but also a set of\nnewly-introduced learnable parameters that are responsible for transforming the\ninput data. This process allows the network to effectively leverage the\npre-training knowledge and find a good trade-off between plasticity and\nstability with modest computational efforts, thus especially suitable for\non-the-edge settings. Our experiments on four image classification problems in\na continual learning setting confirm the quality of the proposed approach when\ncompared to several fine-tuning procedures and to popular continual learning\nmethods.\n", "date": "2023-06-05", "categories": []}
{"id": "2303.14771", "title": "Prototype-Sample Relation Distillation: Towards Replay-Free Continual\n  Learning", "abstract": "  In Continual learning (CL) balancing effective adaptation while combating\ncatastrophic forgetting is a central challenge. Many of the recent\nbest-performing methods utilize various forms of prior task data, e.g. a replay\nbuffer, to tackle the catastrophic forgetting problem. Having access to\nprevious task data can be restrictive in many real-world scenarios, for example\nwhen task data is sensitive or proprietary. To overcome the necessity of using\nprevious tasks' data, in this work, we start with strong representation\nlearning methods that have been shown to be less prone to forgetting. We\npropose a holistic approach to jointly learn the representation and class\nprototypes while maintaining the relevance of old class prototypes and their\nembedded similarities. Specifically, samples are mapped to an embedding space\nwhere the representations are learned using a supervised contrastive loss.\nClass prototypes are evolved continually in the same latent space, enabling\nlearning and prediction at any point. To continually adapt the prototypes\nwithout keeping any prior task data, we propose a novel distillation loss that\nconstrains class prototypes to maintain relative similarities as compared to\nnew task data. This method yields state-of-the-art performance in the\ntask-incremental setting, outperforming methods relying on large amounts of\ndata, and provides strong performance in the class-incremental setting without\nusing any stored data points.\n", "date": "2023-03-26", "categories": []}
{"id": "2306.03534", "title": "Continual Learning in Linear Classification on Separable Data", "abstract": "  We analyze continual learning on a sequence of separable linear\nclassification tasks with binary labels. We show theoretically that learning\nwith weak regularization reduces to solving a sequential max-margin problem,\ncorresponding to a special case of the Projection Onto Convex Sets (POCS)\nframework. We then develop upper bounds on the forgetting and other quantities\nof interest under various settings with recurring tasks, including cyclic and\nrandom orderings of tasks. We discuss several practical implications to popular\ntraining practices like regularization scheduling and weighting. We point out\nseveral theoretical differences between our continual classification setting\nand a recently studied continual regression setting.\n", "date": "2023-06-06", "categories": []}
{"id": "2306.03887", "title": "Fast Context Adaptation in Cost-Aware Continual Learning", "abstract": "  In the past few years, DRL has become a valuable solution to automatically\nlearn efficient resource management strategies in complex networks with\ntime-varying statistics. However, the increased complexity of 5G and Beyond\nnetworks requires correspondingly more complex learning agents and the learning\nprocess itself might end up competing with users for communication and\ncomputational resources. This creates friction: on the one hand, the learning\nprocess needs resources to quickly convergence to an effective strategy; on the\nother hand, the learning process needs to be efficient, i.e., take as few\nresources as possible from the user's data plane, so as not to throttle users'\nQoS. In this paper, we investigate this trade-off and propose a dynamic\nstrategy to balance the resources assigned to the data plane and those reserved\nfor learning. With the proposed approach, a learning agent can quickly converge\nto an efficient resource allocation strategy and adapt to changes in the\nenvironment as for the CL paradigm, while minimizing the impact on the users'\nQoS. Simulation results show that the proposed method outperforms static\nallocation methods with minimal learning overhead, almost reaching the\nperformance of an ideal out-of-band CL solution.\n", "date": "2023-06-06", "categories": []}
{"id": "2208.00777", "title": "D3Former: Debiased Dual Distilled Transformer for Incremental Learning", "abstract": "  In class incremental learning (CIL) setting, groups of classes are introduced\nto a model in each learning phase. The goal is to learn a unified model\nperformant on all the classes observed so far. Given the recent popularity of\nVision Transformers (ViTs) in conventional classification settings, an\ninteresting question is to study their continual learning behaviour. In this\nwork, we develop a Debiased Dual Distilled Transformer for CIL dubbed\n$\\textrm{D}^3\\textrm{Former}$. The proposed model leverages a hybrid nested ViT\ndesign to ensure data efficiency and scalability to small as well as large\ndatasets. In contrast to a recent ViT based CIL approach, our\n$\\textrm{D}^3\\textrm{Former}$ does not dynamically expand its architecture when\nnew tasks are learned and remains suitable for a large number of incremental\ntasks. The improved CIL behaviour of $\\textrm{D}^3\\textrm{Former}$ owes to two\nfundamental changes to the ViT design. First, we treat the incremental learning\nas a long-tail classification problem where the majority samples from new\nclasses vastly outnumber the limited exemplars available for old classes. To\navoid the bias against the minority old classes, we propose to dynamically\nadjust logits to emphasize on retaining the representations relevant to old\ntasks. Second, we propose to preserve the configuration of spatial attention\nmaps as the learning progresses across tasks. This helps in reducing\ncatastrophic forgetting by constraining the model to retain the attention on\nthe most discriminative regions. $\\textrm{D}^3\\textrm{Former}$ obtains\nfavorable results on incremental versions of CIFAR-100, MNIST, SVHN, and\nImageNet datasets. Code is available at https://tinyurl.com/d3former\n", "date": "2022-07-25", "categories": []}
{"id": "2305.18444", "title": "Continual Task Allocation in Meta-Policy Network via Sparse Prompting", "abstract": "  How to train a generalizable meta-policy by continually learning a sequence\nof tasks? It is a natural human skill yet challenging to achieve by current\nreinforcement learning: the agent is expected to quickly adapt to new tasks\n(plasticity) meanwhile retaining the common knowledge from previous tasks\n(stability). We address it by \"Continual Task Allocation via Sparse Prompting\n(CoTASP)\", which learns over-complete dictionaries to produce sparse masks as\nprompts extracting a sub-network for each task from a meta-policy network.\nCoTASP trains a policy for each task by optimizing the prompts and the\nsub-network weights alternatively. The dictionary is then updated to align the\noptimized prompts with tasks' embedding, thereby capturing tasks' semantic\ncorrelations. Hence, relevant tasks share more neurons in the meta-policy\nnetwork due to similar prompts while cross-task interference causing forgetting\nis effectively restrained. Given a meta-policy and dictionaries trained on\nprevious tasks, new task adaptation reduces to highly efficient sparse\nprompting and sub-network finetuning. In experiments, CoTASP achieves a\npromising plasticity-stability trade-off without storing or replaying any past\ntasks' experiences. It outperforms existing continual and multi-task RL methods\non all seen tasks, forgetting reduction, and generalization to unseen tasks.\n", "date": "2023-05-29", "categories": []}
{"id": "2306.02335", "title": "Towards Robust Feature Learning with t-vFM Similarity for Continual\n  Learning", "abstract": "  Continual learning has been developed using standard supervised contrastive\nloss from the perspective of feature learning. Due to the data imbalance during\nthe training, there are still challenges in learning better representations. In\nthis work, we suggest using a different similarity metric instead of cosine\nsimilarity in supervised contrastive loss in order to learn more robust\nrepresentations. We validate the our method on one of the image classification\ndatasets Seq-CIFAR-10 and the results outperform recent continual learning\nbaselines.\n", "date": "2023-06-04", "categories": []}
{"id": "2306.02389", "title": "Fast Continual Multi-View Clustering with Incomplete Views", "abstract": "  Multi-view clustering (MVC) has gained broad attention owing to its capacity\nto exploit consistent and complementary information across views. This paper\nfocuses on a challenging issue in MVC called the incomplete continual data\nproblem (ICDP). In specific, most existing algorithms assume that views are\navailable in advance and overlook the scenarios where data observations of\nviews are accumulated over time. Due to privacy considerations or memory\nlimitations, previous views cannot be stored in these situations. Some works\nare proposed to handle it, but all fail to address incomplete views. Such an\nincomplete continual data problem (ICDP) in MVC is tough to solve since\nincomplete information with continual data increases the difficulty of\nextracting consistent and complementary knowledge among views. We propose Fast\nContinual Multi-View Clustering with Incomplete Views (FCMVC-IV) to address it.\nSpecifically, it maintains a consensus coefficient matrix and updates knowledge\nwith the incoming incomplete view rather than storing and recomputing all the\ndata matrices. Considering that the views are incomplete, the newly collected\nview might contain samples that have yet to appear; two indicator matrices and\na rotation matrix are developed to match matrices with different dimensions.\nBesides, we design a three-step iterative algorithm to solve the resultant\nproblem in linear complexity with proven convergence. Comprehensive experiments\non various datasets show the superiority of FCMVC-IV.\n", "date": "2023-06-04", "categories": []}
{"id": "2306.02413", "title": "Evaluating Continual Learning on a Home Robot", "abstract": "  Robots in home environments need to be able to learn new skills continuously\nas data becomes available, becoming ever more capable over time while using as\nlittle real-world data as possible. However, traditional robot learning\napproaches typically assume large amounts of iid data, which is inconsistent\nwith this goal. In contrast, continual learning methods like CLEAR and SANE\nallow autonomous agents to learn off of a stream of non-iid samples; they,\nhowever, have not previously been demonstrated on real robotics platforms. In\nthis work, we show how continual learning methods can be adapted for use on a\nreal, low-cost home robot, and in particular look at the case where we have\nextremely small numbers of examples, in a task-id-free setting. Specifically,\nwe propose SANER, a method for continuously learning a library of skills, and\nABIP (Attention-Based Interaction Policies) as the backbone to support it. We\nlearn four sequential kitchen tasks on a low-cost home robot, using only a\nhandful of demonstrations per task.\n", "date": "2023-06-04", "categories": []}
{"id": "2211.17068", "title": "Self-Supervised Continual Graph Learning in Adaptive Riemannian Spaces", "abstract": "  Continual graph learning routinely finds its role in a variety of real-world\napplications where the graph data with different tasks come sequentially.\nDespite the success of prior works, it still faces great challenges. On the one\nhand, existing methods work with the zero-curvature Euclidean space, and\nlargely ignore the fact that curvature varies over the coming graph sequence.\nOn the other hand, continual learners in the literature rely on abundant\nlabels, but labeling graph in practice is particularly hard especially for the\ncontinuously emerging graphs on-the-fly. To address the aforementioned\nchallenges, we propose to explore a challenging yet practical problem, the\nself-supervised continual graph learning in adaptive Riemannian spaces. In this\npaper, we propose a novel self-supervised Riemannian Graph Continual Learner\n(RieGrace). In RieGrace, we first design an Adaptive Riemannian GCN (AdaRGCN),\na unified GCN coupled with a neural curvature adapter, so that Riemannian space\nis shaped by the learnt curvature adaptive to each graph. Then, we present a\nLabel-free Lorentz Distillation approach, in which we create teacher-student\nAdaRGCN for the graph sequence. The student successively performs\nintra-distillation from itself and inter-distillation from the teacher so as to\nconsolidate knowledge without catastrophic forgetting. In particular, we\npropose a theoretically grounded Generalized Lorentz Projection for the\ncontrastive distillation in Riemannian space. Extensive experiments on the\nbenchmark datasets show the superiority of RieGrace, and additionally, we\ninvestigate on how curvature changes over the graph sequence.\n", "date": "2022-11-30", "categories": []}
{"id": "2305.07437", "title": "Continual Vision-Language Representation Learning with Off-Diagonal\n  Information", "abstract": "  Large-scale multi-modal contrastive learning frameworks like CLIP typically\nrequire a large amount of image-text samples for training. However, these\nsamples are always collected continuously in real scenarios. This paper\ndiscusses the feasibility of continual CLIP training using streaming data.\nUnlike continual learning based on self-supervised learning methods for pure\nimages, which is empirically robust against catastrophic forgetting, CLIP's\nperformance degeneration in the continual setting is significant and\nnon-neglectable. By analyzing the changes in the model's representation space\nduring continual CLIP training from a spatial geometry perspective, we explore\nand summarize these spatial variations as Spatial Disorder (SD), which can be\ndivided into Intra-modal Rotation and Inter-modal Deviation. Moreover, we\nempirically and theoretically demonstrate how SD leads to a performance decline\nfor CLIP on cross-modal retrieval tasks. To alleviate SD, we propose a new\ncontinual vision-language representation learning framework Mod-X: Maintain\noff-diagonal information-matriX. By selectively aligning the off-diagonal\ninformation distribution of contrastive matrices, the Mod-X improves the\ncapability of the multi-modal model by maintaining the multi-modal\nrepresentation space alignment on the old data domain during continuously\nfitting the new training data domain. Experiments on commonly used datasets\nwith different scales and scopes have demonstrated the effectiveness of our\nmethod.\n", "date": "2023-05-11", "categories": []}
{"id": "2305.19879", "title": "RaSP: Relation-aware Semantic Prior for Weakly Supervised Incremental\n  Segmentation", "abstract": "  Class-incremental semantic image segmentation assumes multiple model updates,\neach enriching the model to segment new categories. This is typically carried\nout by providing expensive pixel-level annotations to the training algorithm\nfor all new objects, limiting the adoption of such methods in practical\napplications. Approaches that solely require image-level labels offer an\nattractive alternative, yet, such coarse annotations lack precise information\nabout the location and boundary of the new objects. In this paper we argue\nthat, since classes represent not just indices but semantic entities, the\nconceptual relationships between them can provide valuable information that\nshould be leveraged. We propose a weakly supervised approach that exploits such\nsemantic relations to transfer objectness prior from the previously learned\nclasses into the new ones, complementing the supervisory signal from\nimage-level labels. We validate our approach on a number of continual learning\ntasks, and show how even a simple pairwise interaction between classes can\nsignificantly improve the segmentation mask quality of both old and new\nclasses. We show these conclusions still hold for longer and, hence, more\nrealistic sequences of tasks and for a challenging few-shot scenario.\n", "date": "2023-05-31", "categories": []}
{"id": "2305.18563", "title": "SHARP: Sparsity and Hidden Activation RePlay for Neuro-Inspired\n  Continual Learning", "abstract": "  Deep neural networks (DNNs) struggle to learn in dynamic environments since\nthey rely on fixed datasets or stationary environments. Continual learning (CL)\naims to address this limitation and enable DNNs to accumulate knowledge\nincrementally, similar to human learning. Inspired by how our brain\nconsolidates memories, a powerful strategy in CL is replay, which involves\ntraining the DNN on a mixture of new and all seen classes. However, existing\nreplay methods overlook two crucial aspects of biological replay: 1) the brain\nreplays processed neural patterns instead of raw input, and 2) it prioritizes\nthe replay of recently learned information rather than revisiting all past\nexperiences. To address these differences, we propose SHARP, an efficient\nneuro-inspired CL method that leverages sparse dynamic connectivity and\nactivation replay. Unlike other activation replay methods, which assume layers\nnot subjected to replay have been pretrained and fixed, SHARP can continually\nupdate all layers. Also, SHARP is unique in that it only needs to replay few\nrecently seen classes instead of all past classes. Our experiments on five\ndatasets demonstrate that SHARP outperforms state-of-the-art replay methods in\nclass incremental learning. Furthermore, we showcase SHARP's flexibility in a\nnovel CL scenario where the boundaries between learning episodes are blurry.\nThe SHARP code is available at\n\\url{https://github.com/BurakGurbuz97/SHARP-Continual-Learning}.\n", "date": "2023-05-29", "categories": []}
{"id": "2305.18953", "title": "Sit Back and Relax: Learning to Drive Incrementally in All Weather\n  Conditions", "abstract": "  In autonomous driving scenarios, current object detection models show strong\nperformance when tested in clear weather. However, their performance\ndeteriorates significantly when tested in degrading weather conditions. In\naddition, even when adapted to perform robustly in a sequence of different\nweather conditions, they are often unable to perform well in all of them and\nsuffer from catastrophic forgetting. To efficiently mitigate forgetting, we\npropose Domain-Incremental Learning through Activation Matching (DILAM), which\nemploys unsupervised feature alignment to adapt only the affine parameters of a\nclear weather pre-trained network to different weather conditions. We propose\nto store these affine parameters as a memory bank for each weather condition\nand plug-in their weather-specific parameters during driving (i.e. test time)\nwhen the respective weather conditions are encountered. Our memory bank is\nextremely lightweight, since affine parameters account for less than 2% of a\ntypical object detector. Furthermore, contrary to previous domain-incremental\nlearning approaches, we do not require the weather label when testing and\npropose to automatically infer the weather condition by a majority voting\nlinear classifier.\n", "date": "2023-05-30", "categories": []}
{"id": "2305.17244", "title": "Mitigating Catastrophic Forgetting in Long Short-Term Memory Networks", "abstract": "  Continual learning on sequential data is critical for many machine learning\n(ML) deployments. Unfortunately, LSTM networks, which are commonly used to\nlearn on sequential data, suffer from catastrophic forgetting and are limited\nin their ability to learn multiple tasks continually. We discover that\ncatastrophic forgetting in LSTM networks can be overcome in two novel and\nreadily-implementable ways -- separating the LSTM memory either for each task\nor for each target label. Our approach eschews the need for explicit\nregularization, hypernetworks, and other complex methods. We quantify the\nbenefits of our approach on recently-proposed LSTM networks for computer memory\naccess prefetching, an important sequential learning problem in ML-based\ncomputer system optimization. Compared to state-of-the-art weight\nregularization methods to mitigate catastrophic forgetting, our approach is\nsimple, effective, and enables faster learning. We also show that our proposal\nenables the use of small, non-regularized LSTM networks for complex natural\nlanguage processing in the offline learning scenario, which was previously\nconsidered difficult.\n", "date": "2023-05-26", "categories": []}
{"id": "2305.16484", "title": "Batch Model Consolidation: A Multi-Task Model Consolidation Framework", "abstract": "  In Continual Learning (CL), a model is required to learn a stream of tasks\nsequentially without significant performance degradation on previously learned\ntasks. Current approaches fail for a long sequence of tasks from diverse\ndomains and difficulties. Many of the existing CL approaches are difficult to\napply in practice due to excessive memory cost or training time, or are tightly\ncoupled to a single device. With the intuition derived from the widely applied\nmini-batch training, we propose Batch Model Consolidation ($\\textbf{BMC}$) to\nsupport more realistic CL under conditions where multiple agents are exposed to\na range of tasks. During a $\\textit{regularization}$ phase, BMC trains multiple\n$\\textit{expert models}$ in parallel on a set of disjoint tasks. Each expert\nmaintains weight similarity to a $\\textit{base model}$ through a\n$\\textit{stability loss}$, and constructs a $\\textit{buffer}$ from a fraction\nof the task's data. During the $\\textit{consolidation}$ phase, we combine the\nlearned knowledge on 'batches' of $\\textit{expert models}$ using a\n$\\textit{batched consolidation loss}$ in $\\textit{memory}$ data that aggregates\nall buffers. We thoroughly evaluate each component of our method in an ablation\nstudy and demonstrate the effectiveness on standardized benchmark datasets\nSplit-CIFAR-100, Tiny-ImageNet, and the Stream dataset composed of 71 image\nclassification tasks from diverse domains and difficulties. Our method\noutperforms the next best CL approach by 70% and is the only approach that can\nmaintain performance at the end of 71 tasks; Our benchmark can be accessed at\nhttps://github.com/fostiropoulos/stream_benchmark\n", "date": "2023-05-25", "categories": []}
{"id": "2305.16559", "title": "Teamwork Is Not Always Good: An Empirical Study of Classifier Drift in\n  Class-incremental Information Extraction", "abstract": "  Class-incremental learning (CIL) aims to develop a learning system that can\ncontinually learn new classes from a data stream without forgetting previously\nlearned classes. When learning classes incrementally, the classifier must be\nconstantly updated to incorporate new classes, and the drift in decision\nboundary may lead to severe forgetting. This fundamental challenge, however,\nhas not yet been studied extensively, especially in the setting where no\nsamples from old classes are stored for rehearsal. In this paper, we take a\ncloser look at how the drift in the classifier leads to forgetting, and\naccordingly, design four simple yet (super-) effective solutions to alleviate\nthe classifier drift: an Individual Classifiers with Frozen Feature Extractor\n(ICE) framework where we individually train a classifier for each learning\nsession, and its three variants ICE-PL, ICE-O, and ICE-PL&O which further take\nthe logits of previously learned classes from old sessions or a constant logit\nof an Other class as a constraint to the learning of new classifiers. Extensive\nexperiments and analysis on 6 class-incremental information extraction tasks\ndemonstrate that our solutions, especially ICE-O, consistently show significant\nimprovement over the previous state-of-the-art approaches with up to 44.7%\nabsolute F-score gain, providing a strong baseline and insights for future\nresearch on class-incremental learning.\n", "date": "2023-05-26", "categories": []}
{"id": "2305.16687", "title": "Balanced Supervised Contrastive Learning for Few-Shot Class-Incremental\n  Learning", "abstract": "  Few-shot class-incremental learning (FSCIL) presents the primary challenge of\nbalancing underfitting to a new session's task and forgetting the tasks from\nprevious sessions. To address this challenge, we develop a simple yet powerful\nlearning scheme that integrates effective methods for each core component of\nthe FSCIL network, including the feature extractor, base session classifiers,\nand incremental session classifiers. In feature extractor training, our goal is\nto obtain balanced generic representations that benefit both current viewable\nand unseen or past classes. To achieve this, we propose a balanced supervised\ncontrastive loss that effectively balances these two objectives. In terms of\nclassifiers, we analyze and emphasize the importance of unifying initialization\nmethods for both the base and incremental session classifiers. Our method\ndemonstrates outstanding ability for new task learning and preventing\nforgetting on CUB200, CIFAR100, and miniImagenet datasets, with significant\nimprovements over previous state-of-the-art methods across diverse metrics. We\nconduct experiments to analyze the significance and rationale behind our\napproach and visualize the effectiveness of our representations on new tasks.\nFurthermore, we conduct diverse ablation studies to analyze the effects of each\nmodule.\n", "date": "2023-05-26", "categories": []}
{"id": "2212.01393", "title": "Continual Learning for On-Device Speech Recognition using Disentangled\n  Conformers", "abstract": "  Automatic speech recognition research focuses on training and evaluating on\nstatic datasets. Yet, as speech models are increasingly deployed on personal\ndevices, such models encounter user-specific distributional shifts. To simulate\nthis real-world scenario, we introduce LibriContinual, a continual learning\nbenchmark for speaker-specific domain adaptation derived from LibriVox\naudiobooks, with data corresponding to 118 individual speakers and 6 train\nsplits per speaker of different sizes. Additionally, current speech recognition\nmodels and continual learning algorithms are not optimized to be\ncompute-efficient. We adapt a general-purpose training algorithm NetAug for ASR\nand create a novel Conformer variant called the DisConformer (Disentangled\nConformer). This algorithm produces ASR models consisting of a frozen 'core'\nnetwork for general-purpose use and several tunable 'augment' networks for\nspeaker-specific tuning. Using such models, we propose a novel\ncompute-efficient continual learning algorithm called DisentangledCL. Our\nexperiments show that the DisConformer models significantly outperform\nbaselines on general ASR i.e. LibriSpeech (15.58% rel. WER on test-other). On\nspeaker-specific LibriContinual they significantly outperform\ntrainable-parameter-matched baselines (by 20.65% rel. WER on test) and even\nmatch fully finetuned baselines in some settings.\n", "date": "2022-12-02", "categories": []}
{"id": "2305.15591", "title": "Lightweight Learner for Shared Knowledge Lifelong Learning", "abstract": "  In Lifelong Learning (LL), agents continually learn as they encounter new\nconditions and tasks. Most current LL is limited to a single agent that learns\ntasks sequentially. Dedicated LL machinery is then deployed to mitigate the\nforgetting of old tasks as new tasks are learned. This is inherently slow. We\npropose a new Shared Knowledge Lifelong Learning (SKILL) challenge, which\ndeploys a decentralized population of LL agents that each sequentially learn\ndifferent tasks, with all agents operating independently and in parallel. After\nlearning their respective tasks, agents share and consolidate their knowledge\nover a decentralized communication network, so that, in the end, all agents can\nmaster all tasks. We present one solution to SKILL which uses Lightweight\nLifelong Learning (LLL) agents, where the goal is to facilitate efficient\nsharing by minimizing the fraction of the agent that is specialized for any\ngiven task. Each LLL agent thus consists of a common task-agnostic immutable\npart, where most parameters are, and individual task-specific modules that\ncontain fewer parameters but are adapted to each task. Agents share their\ntask-specific modules, plus summary information (\"task anchors\") representing\ntheir tasks in the common task-agnostic latent space of all agents. Receiving\nagents register each received task-specific module using the corresponding\nanchor. Thus, every agent improves its ability to solve new tasks each time new\ntask-specific modules and anchors are received. On a new, very challenging\nSKILL-102 dataset with 102 image classification tasks (5,033 classes in total,\n2,041,225 training, 243,464 validation, and 243,464 test images), we achieve\nmuch higher (and SOTA) accuracy over 8 LL baselines, while also achieving near\nperfect parallelization. Code and data can be found at\nhttps://github.com/gyhandy/Shared-Knowledge-Lifelong-Learning\n", "date": "2023-05-24", "categories": []}
{"id": "2305.15909", "title": "Camera-Incremental Object Re-Identification with Identity Knowledge\n  Evolution", "abstract": "  Object Re-identification (ReID) aims to retrieve the probe object from many\ngallery images with the ReID model inferred based on a stationary camera-free\ndataset by associating and collecting the identities across all camera views.\nWhen deploying the ReID algorithm in real-world scenarios, the aspect of\nstorage, privacy constraints, and dynamic changes of cameras would degrade its\ngeneralizability and applicability. Treating each camera's data independently,\nwe introduce a novel ReID task named Camera-Incremental Object\nRe-identification (CIOR) by continually optimizing the ReID mode from the\nincoming stream of the camera dataset. Since the identities under different\ncamera views might describe the same object, associating and distilling the\nknowledge of common identities would boost the discrimination and benefit from\nalleviating the catastrophic forgetting. In this paper, we propose a novel\nIdentity Knowledge Evolution (IKE) framework for CIOR, consisting of the\nIdentity Knowledge Association (IKA), Identity Knowledge Distillation (IKD),\nand Identity Knowledge Update (IKU). IKA is proposed to discover the common\nidentities between the current identity and historical identities. IKD has\napplied to distillate historical identity knowledge from common identities and\nquickly adapt the historical model to the current camera view. After each\ncamera has been trained, IKU is applied to continually expand the identity\nknowledge by combining the historical and current identity memories. The\nevaluation of Market-CL and Veri-CL shows the Identity Knowledge Evolution\n(IKE) effectiveness for CIOR.\ncode:https://github.com/htyao89/Camera-Incremental-Object-ReID\n", "date": "2023-05-25", "categories": []}
{"id": "2305.16143", "title": "Condensed Prototype Replay for Class Incremental Learning", "abstract": "  Incremental learning (IL) suffers from catastrophic forgetting of old tasks\nwhen learning new tasks. This can be addressed by replaying previous tasks'\ndata stored in a memory, which however is usually prone to size limits and\nprivacy leakage. Recent studies store only class centroids as prototypes and\naugment them with Gaussian noises to create synthetic data for replay. However,\nthey cannot effectively avoid class interference near their margins that leads\nto forgetting. Moreover, the injected noises distort the rich structure between\nreal data and prototypes, hence even detrimental to IL. In this paper, we\npropose YONO that You Only Need to replay One condensed prototype per class,\nwhich for the first time can even outperform memory-costly exemplar-replay\nmethods. To this end, we develop a novel prototype learning method that (1)\nsearches for more representative prototypes in high-density regions by an\nattentional mean-shift algorithm and (2) moves samples in each class to their\nprototype to form a compact cluster distant from other classes. Thereby, the\nclass margins are maximized, which effectively reduces interference causing\nfuture forgetting. In addition, we extend YONO to YONO+, which creates\nsynthetic replay data by random sampling in the neighborhood of each prototype\nin the representation space. We show that the synthetic data can further\nimprove YONO. Extensive experiments on IL benchmarks demonstrate the advantages\nof YONO/YONO+ over existing IL methods in terms of both accuracy and\nforgetting.\n", "date": "2023-05-25", "categories": []}
{"id": "2305.16252", "title": "Overcoming Catastrophic Forgetting in Massively Multilingual Continual\n  Learning", "abstract": "  Real-life multilingual systems should be able to efficiently incorporate new\nlanguages as data distributions fed to the system evolve and shift over time.\nTo do this, systems need to handle the issue of catastrophic forgetting, where\nthe model performance drops for languages or tasks seen further in its past. In\nthis paper, we study catastrophic forgetting, as well as methods to minimize\nthis, in a massively multilingual continual learning framework involving up to\n51 languages and covering both classification and sequence labeling tasks. We\npresent LR ADJUST, a learning rate scheduling method that is simple, yet\neffective in preserving new information without strongly overwriting past\nknowledge. Furthermore, we show that this method is effective across multiple\ncontinual learning approaches. Finally, we provide further insights into the\ndynamics of catastrophic forgetting in this massively multilingual setup.\n", "date": "2023-05-25", "categories": []}
{"id": "2209.00347", "title": "Dynamics-Adaptive Continual Reinforcement Learning via Progressive\n  Contextualization", "abstract": "  A key challenge of continual reinforcement learning (CRL) in dynamic\nenvironments is to promptly adapt the RL agent's behavior as the environment\nchanges over its lifetime, while minimizing the catastrophic forgetting of the\nlearned information. To address this challenge, in this article, we propose\nDaCoRL, i.e., dynamics-adaptive continual RL. DaCoRL learns a\ncontext-conditioned policy using progressive contextualization, which\nincrementally clusters a stream of stationary tasks in the dynamic environment\ninto a series of contexts and opts for an expandable multihead neural network\nto approximate the policy. Specifically, we define a set of tasks with similar\ndynamics as an environmental context and formalize context inference as a\nprocedure of online Bayesian infinite Gaussian mixture clustering on\nenvironment features, resorting to online Bayesian inference to infer the\nposterior distribution over contexts. Under the assumption of a Chinese\nrestaurant process prior, this technique can accurately classify the current\ntask as a previously seen context or instantiate a new context as needed\nwithout relying on any external indicator to signal environmental changes in\nadvance. Furthermore, we employ an expandable multihead neural network whose\noutput layer is synchronously expanded with the newly instantiated context, and\na knowledge distillation regularization term for retaining the performance on\nlearned tasks. As a general framework that can be coupled with various deep RL\nalgorithms, DaCoRL features consistent superiority over existing methods in\nterms of the stability, overall performance and generalization ability, as\nverified by extensive experiments on several robot navigation and MuJoCo\nlocomotion tasks.\n", "date": "2022-09-01", "categories": []}
{"id": "2305.14657", "title": "Dealing with Cross-Task Class Discrimination in Online Continual\n  Learning", "abstract": "  Existing continual learning (CL) research regards catastrophic forgetting\n(CF) as almost the only challenge. This paper argues for another challenge in\nclass-incremental learning (CIL), which we call cross-task class discrimination\n(CTCD),~i.e., how to establish decision boundaries between the classes of the\nnew task and old tasks with no (or limited) access to the old task data. CTCD\nis implicitly and partially dealt with by replay-based methods. A replay method\nsaves a small amount of data (replay data) from previous tasks. When a batch of\ncurrent task data arrives, the system jointly trains the new data and some\nsampled replay data. The replay data enables the system to partially learn the\ndecision boundaries between the new classes and the old classes as the amount\nof the saved data is small. However, this paper argues that the replay approach\nalso has a dynamic training bias issue which reduces the effectiveness of the\nreplay data in solving the CTCD problem. A novel optimization objective with a\ngradient-based adaptive method is proposed to dynamically deal with the problem\nin the online CL process. Experimental results show that the new method\nachieves much better results in online CL.\n", "date": "2023-05-24", "categories": []}
{"id": "2211.08161", "title": "An Investigation of the Combination of Rehearsal and Knowledge\n  Distillation in Continual Learning for Spoken Language Understanding", "abstract": "  Continual learning refers to a dynamical framework in which a model receives\na stream of non-stationary data over time and must adapt to new data while\npreserving previously acquired knowledge. Unluckily, neural networks fail to\nmeet these two desiderata, incurring the so-called catastrophic forgetting\nphenomenon. Whereas a vast array of strategies have been proposed to attenuate\nforgetting in the computer vision domain, for speech-related tasks, on the\nother hand, there is a dearth of works. In this paper, we consider the joint\nuse of rehearsal and knowledge distillation (KD) approaches for spoken language\nunderstanding under a class-incremental learning scenario. We report on\nmultiple KD combinations at different levels in the network, showing that\ncombining feature-level and predictions-level KDs leads to the best results.\nFinally, we provide an ablation study on the effect of the size of the\nrehearsal memory that corroborates the efficacy of our approach for\nlow-resource devices.\n", "date": "2022-11-15", "categories": []}
{"id": "2303.01904", "title": "EcoTTA: Memory-Efficient Continual Test-time Adaptation via\n  Self-distilled Regularization", "abstract": "  This paper presents a simple yet effective approach that improves continual\ntest-time adaptation (TTA) in a memory-efficient manner. TTA may primarily be\nconducted on edge devices with limited memory, so reducing memory is crucial\nbut has been overlooked in previous TTA studies. In addition, long-term\nadaptation often leads to catastrophic forgetting and error accumulation, which\nhinders applying TTA in real-world deployments. Our approach consists of two\ncomponents to address these issues. First, we present lightweight meta networks\nthat can adapt the frozen original networks to the target domain. This novel\narchitecture minimizes memory consumption by decreasing the size of\nintermediate activations required for backpropagation. Second, our novel\nself-distilled regularization controls the output of the meta networks not to\ndeviate significantly from the output of the frozen original networks, thereby\npreserving well-trained knowledge from the source domain. Without additional\nmemory, this regularization prevents error accumulation and catastrophic\nforgetting, resulting in stable performance even in long-term test-time\nadaptation. We demonstrate that our simple yet effective strategy outperforms\nother state-of-the-art methods on various benchmarks for image classification\nand semantic segmentation tasks. Notably, our proposed method with ResNet-50\nand WideResNet-40 takes 86% and 80% less memory than the recent\nstate-of-the-art method, CoTTA.\n", "date": "2023-03-03", "categories": []}
{"id": "2305.13332", "title": "Conditional Online Learning for Keyword Spotting", "abstract": "  Modern approaches for keyword spotting rely on training deep neural networks\non large static datasets with i.i.d. distributions. However, the resulting\nmodels tend to underperform when presented with changing data regimes in\nreal-life applications. This work investigates a simple but effective online\ncontinual learning method that updates a keyword spotter on-device via SGD as\nnew data becomes available. Contrary to previous research, this work focuses on\nlearning the same KWS task, which covers most commercial applications. During\nexperiments with dynamic audio streams in different scenarios, that method\nimproves the performance of a pre-trained small-footprint model by 34%.\nMoreover, experiments demonstrate that, compared to a naive online learning\nimplementation, conditional model updates based on its performance in a small\nhold-out set drawn from the training distribution mitigate catastrophic\nforgetting.\n", "date": "2023-05-19", "categories": []}
{"id": "2305.13678", "title": "Enhancing Accuracy and Robustness through Adversarial Training in Class\n  Incremental Continual Learning", "abstract": "  In real life, adversarial attack to deep learning models is a fatal security\nissue. However, the issue has been rarely discussed in a widely used\nclass-incremental continual learning (CICL). In this paper, we address problems\nof applying adversarial training to CICL, which is well-known defense method\nagainst adversarial attack. A well-known problem of CICL is class-imbalance\nthat biases a model to the current task by a few samples of previous tasks.\nMeeting with the adversarial training, the imbalance causes another imbalance\nof attack trials over tasks. Lacking clean data of a minority class by the\nclass-imbalance and increasing of attack trials from a majority class by the\nsecondary imbalance, adversarial training distorts optimal decision boundaries.\nThe distortion eventually decreases both accuracy and robustness than\nadversarial training. To exclude the effects, we propose a straightforward but\nsignificantly effective method, External Adversarial Training (EAT) which can\nbe applied to methods using experience replay. This method conduct adversarial\ntraining to an auxiliary external model for the current task data at each time\nstep, and applies generated adversarial examples to train the target model. We\nverify the effects on a toy problem and show significance on CICL benchmarks of\nimage classification. We expect that the results will be used as the first\nbaseline for robustness research of CICL.\n", "date": "2023-05-23", "categories": []}
{"id": "2305.12270", "title": "Mitigating Catastrophic Forgetting in Task-Incremental Continual\n  Learning with Adaptive Classification Criterion", "abstract": "  Task-incremental continual learning refers to continually training a model in\na sequence of tasks while overcoming the problem of catastrophic forgetting\n(CF). The issue arrives for the reason that the learned representations are\nforgotten for learning new tasks, and the decision boundary is destructed.\nPrevious studies mostly consider how to recover the representations of learned\ntasks. It is seldom considered to adapt the decision boundary for new\nrepresentations and in this paper we propose a Supervised Contrastive learning\nframework with adaptive classification criterion for Continual Learning (SCCL),\nIn our method, a contrastive loss is used to directly learn representations for\ndifferent tasks and a limited number of data samples are saved as the\nclassification criterion. During inference, the saved data samples are fed into\nthe current model to obtain updated representations, and a k Nearest Neighbour\nmodule is used for classification. In this way, the extensible model can solve\nthe learned tasks with adaptive criteria of saved samples. To mitigate CF, we\nfurther use an instance-wise relation distillation regularization term and a\nmemory replay module to maintain the information of previous tasks. Experiments\nshow that SCCL achieves state-of-the-art performance and has a stronger ability\nto overcome CF compared with the classification baselines.\n", "date": "2023-05-20", "categories": []}
{"id": "2305.10868", "title": "Advancing Incremental Few-shot Semantic Segmentation via Semantic-guided\n  Relation Alignment and Adaptation", "abstract": "  Incremental few-shot semantic segmentation (IFSS) aims to incrementally\nextend a semantic segmentation model to novel classes according to only a few\npixel-level annotated data, while preserving its segmentation capability on\npreviously learned base categories. This task faces a severe semantic-aliasing\nissue between base and novel classes due to data imbalance, which makes\nsegmentation results unsatisfactory. To alleviate this issue, we propose the\nSemantic-guided Relation Alignment and Adaptation (SRAA) method that fully\nconsiders the guidance of prior semantic information. Specifically, we first\nconduct Semantic Relation Alignment (SRA) in the base step, so as to\nsemantically align base class representations to their semantics. As a result,\nthe embeddings of base classes are constrained to have relatively low semantic\ncorrelations to categories that are different from them. Afterwards, based on\nthe semantically aligned base categories, Semantic-Guided Adaptation (SGA) is\nemployed during the incremental learning stage. It aims to ensure affinities\nbetween visual and semantic embeddings of encountered novel categories, thereby\nmaking the feature representations be consistent with their semantic\ninformation. In this way, the semantic-aliasing issue can be suppressed. We\nevaluate our model on the PASCAL VOC 2012 and the COCO dataset. The\nexperimental results on both these two datasets exhibit its competitive\nperformance, which demonstrates the superiority of our method.\n", "date": "2023-05-18", "categories": []}
{"id": "2211.11747", "title": "NEVIS'22: A Stream of 100 Tasks Sampled from 30 Years of Computer Vision\n  Research", "abstract": "  A shared goal of several machine learning communities like continual\nlearning, meta-learning and transfer learning, is to design algorithms and\nmodels that efficiently and robustly adapt to unseen tasks. An even more\nambitious goal is to build models that never stop adapting, and that become\nincreasingly more efficient through time by suitably transferring the accrued\nknowledge. Beyond the study of the actual learning algorithm and model\narchitecture, there are several hurdles towards our quest to build such models,\nsuch as the choice of learning protocol, metric of success and data needed to\nvalidate research hypotheses. In this work, we introduce the Never-Ending\nVIsual-classification Stream (NEVIS'22), a benchmark consisting of a stream of\nover 100 visual classification tasks, sorted chronologically and extracted from\npapers sampled uniformly from computer vision proceedings spanning the last\nthree decades. The resulting stream reflects what the research community\nthought was meaningful at any point in time, and it serves as an ideal test bed\nto assess how well models can adapt to new tasks, and do so better and more\nefficiently as time goes by. Despite being limited to classification, the\nresulting stream has a rich diversity of tasks from OCR, to texture analysis,\nscene recognition, and so forth. The diversity is also reflected in the wide\nrange of dataset sizes, spanning over four orders of magnitude. Overall,\nNEVIS'22 poses an unprecedented challenge for current sequential learning\napproaches due to the scale and diversity of tasks, yet with a low entry\nbarrier as it is limited to a single modality and well understood supervised\nlearning problems. Moreover, we provide a reference implementation including\nstrong baselines and an evaluation protocol to compare methods in terms of\ntheir trade-off between accuracy and compute.\n", "date": "2022-11-15", "categories": []}
{"id": "2207.04543", "title": "Challenging Common Assumptions about Catastrophic Forgetting", "abstract": "  Building learning agents that can progressively learn and accumulate\nknowledge is the core goal of the continual learning (CL) research field.\nUnfortunately, training a model on new data usually compromises the performance\non past data. In the CL literature, this effect is referred to as catastrophic\nforgetting (CF). CF has been largely studied, and a plethora of methods have\nbeen proposed to address it on short sequences of non-overlapping tasks. In\nsuch setups, CF always leads to a quick and significant drop in performance in\npast tasks. Nevertheless, despite CF, recent work showed that SGD training on\nlinear models accumulates knowledge in a CL regression setup. This phenomenon\nbecomes especially visible when tasks reoccur. We might then wonder if DNNs\ntrained with SGD or any standard gradient-based optimization accumulate\nknowledge in such a way. Such phenomena would have interesting consequences for\napplying DNNs to real continual scenarios. Indeed, standard gradient-based\noptimization methods are significantly less computationally expensive than\nexisting CL algorithms. In this paper, we study the progressive knowledge\naccumulation (KA) in DNNs trained with gradient-based algorithms in long\nsequences of tasks with data re-occurrence. We propose a new framework, SCoLe\n(Scaling Continual Learning), to investigate KA and discover that catastrophic\nforgetting has a limited effect on DNNs trained with SGD. When trained on long\nsequences with data sparsely re-occurring, the overall accuracy improves, which\nmight be counter-intuitive given the CF phenomenon. We empirically investigate\nKA in DNNs under various data occurrence frequencies and propose simple and\nscalable strategies to increase knowledge accumulation in DNNs.\n", "date": "2022-07-10", "categories": []}
{"id": "2301.07187", "title": "Artificial Neuronal Ensembles with Learned Context Dependent Gating", "abstract": "  Biological neural networks are capable of recruiting different sets of\nneurons to encode different memories. However, when training artificial neural\nnetworks on a set of tasks, typically, no mechanism is employed for selectively\nproducing anything analogous to these neuronal ensembles. Further, artificial\nneural networks suffer from catastrophic forgetting, where the network's\nperformance rapidly deteriorates as tasks are learned sequentially. By\ncontrast, sequential learning is possible for a range of biological organisms.\nWe introduce Learned Context Dependent Gating (LXDG), a method to flexibly\nallocate and recall `artificial neuronal ensembles', using a particular network\nstructure and a new set of regularization terms. Activities in the hidden\nlayers of the network are modulated by gates, which are dynamically produced\nduring training. The gates are outputs of networks themselves, trained with a\nsigmoid output activation. The regularization terms we have introduced\ncorrespond to properties exhibited by biological neuronal ensembles. The first\nterm penalizes low gate sparsity, ensuring that only a specified fraction of\nthe network is used. The second term ensures that previously learned gates are\nrecalled when the network is presented with input from previously learned\ntasks. Finally, there is a regularization term responsible for ensuring that\nnew tasks are encoded in gates that are as orthogonal as possible from\npreviously used ones. We demonstrate the ability of this method to alleviate\ncatastrophic forgetting on continual learning benchmarks. When the new\nregularization terms are included in the model along with Elastic Weight\nConsolidation (EWC) it achieves better performance on the benchmark `permuted\nMNIST' than with EWC alone. The benchmark `rotated MNIST' demonstrates how\nsimilar tasks recruit similar neurons to the artificial neuronal ensemble.\n", "date": "2023-01-17", "categories": []}
{"id": "2305.09275", "title": "Rapid Adaptation in Online Continual Learning: Are We Evaluating It\n  Right?", "abstract": "  We revisit the common practice of evaluating adaptation of Online Continual\nLearning (OCL) algorithms through the metric of online accuracy, which measures\nthe accuracy of the model on the immediate next few samples. However, we show\nthat this metric is unreliable, as even vacuous blind classifiers, which do not\nuse input images for prediction, can achieve unrealistically high online\naccuracy by exploiting spurious label correlations in the data stream. Our\nstudy reveals that existing OCL algorithms can also achieve high online\naccuracy, but perform poorly in retaining useful information, suggesting that\nthey unintentionally learn spurious label correlations. To address this issue,\nwe propose a novel metric for measuring adaptation based on the accuracy on the\nnear-future samples, where spurious correlations are removed. We benchmark\nexisting OCL approaches using our proposed metric on large-scale datasets under\nvarious computational budgets and find that better generalization can be\nachieved by retaining and reusing past seen information. We believe that our\nproposed metric can aid in the development of truly adaptive OCL methods. We\nprovide code to reproduce our results at\nhttps://github.com/drimpossible/EvalOCL.\n", "date": "2023-05-16", "categories": []}
{"id": "2305.08702", "title": "Recyclable Tuning for Continual Pre-training", "abstract": "  Continual pre-training is the paradigm where pre-trained language models\n(PLMs) continually acquire fresh knowledge from growing data and gradually get\nupgraded. Before an upgraded PLM is released, we may have tuned the original\nPLM for various tasks and stored the adapted weights. However, when tuning the\nupgraded PLM, these outdated adapted weights will typically be ignored and\ndiscarded, causing a potential waste of resources. We bring this issue to the\nforefront and contend that proper algorithms for recycling outdated adapted\nweights should be developed. To this end, we formulate the task of recyclable\ntuning for continual pre-training. In pilot studies, we find that after\ncontinual pre-training, the upgraded PLM remains compatible with the outdated\nadapted weights to some extent. Motivated by this finding, we analyze the\nconnection between continually pre-trained PLMs from two novel aspects, i.e.,\nmode connectivity, and functional similarity. Based on the corresponding\nfindings, we propose both an initialization-based method and a\ndistillation-based method for our task. We demonstrate their feasibility in\nimproving the convergence and performance for tuning the upgraded PLM. We also\nshow that both methods can be combined to achieve better performance. The\nsource codes are publicly available at\nhttps://github.com/thunlp/RecyclableTuning.\n", "date": "2023-05-15", "categories": []}
{"id": "2305.07085", "title": "Enhancing Contrastive Learning with Noise-Guided Attack: Towards\n  Continual Relation Extraction in the Wild", "abstract": "  The principle of continual relation extraction~(CRE) involves adapting to\nemerging novel relations while preserving od knowledge. While current endeavors\nin CRE succeed in preserving old knowledge, they tend to fail when exposed to\ncontaminated data streams. We assume this is attributed to their reliance on an\nartificial hypothesis that the data stream has no annotation errors, which\nhinders real-world applications for CRE. Considering the ubiquity of noisy\nlabels in real-world datasets, in this paper, we formalize a more practical\nlearning scenario, termed as \\textit{noisy-CRE}. Building upon this challenging\nsetting, we develop a noise-resistant contrastive framework named as\n\\textbf{N}oise-guided \\textbf{a}ttack in \\textbf{C}ontrative\n\\textbf{L}earning~(NaCL) to learn incremental corrupted relations. Compared to\ndirect noise discarding or inaccessible noise relabeling, we present modifying\nthe feature space to match the given noisy labels via attacking can better\nenrich contrastive representations. Extensive empirical validations highlight\nthat NaCL can achieve consistent performance improvements with increasing noise\nrates, outperforming state-of-the-art baselines.\n", "date": "2023-05-11", "categories": []}
{"id": "2305.07289", "title": "RepCL: Exploring Effective Representation for Continual Text\n  Classification", "abstract": "  Continual learning (CL) aims to constantly learn new knowledge over time\nwhile avoiding catastrophic forgetting on old tasks. In this work, we focus on\ncontinual text classification under the class-incremental setting. Recent CL\nstudies find that the representations learned in one task may not be effective\nfor other tasks, namely representation bias problem. For the first time we\nformally analyze representation bias from an information bottleneck perspective\nand suggest that exploiting representations with more class-relevant\ninformation could alleviate the bias. To this end, we propose a novel\nreplay-based continual text classification method, RepCL. Our approach utilizes\ncontrastive and generative representation learning objectives to capture more\nclass-relevant features. In addition, RepCL introduces an adversarial replay\nstrategy to alleviate the overfitting problem of replay. Experiments\ndemonstrate that RepCL effectively alleviates forgetting and achieves\nstate-of-the-art performance on three text classification tasks.\n", "date": "2023-05-12", "categories": []}
{"id": "2211.12701", "title": "Continual Learning of Natural Language Processing Tasks: A Survey", "abstract": "  Continual learning (CL) is a learning paradigm that emulates the human\ncapability of learning and accumulating knowledge continually without\nforgetting the previously learned knowledge and also transferring the learned\nknowledge to help learn new tasks better. This survey presents a comprehensive\nreview and analysis of the recent progress of CL in NLP, which has significant\ndifferences from CL in computer vision and machine learning. It covers (1) all\nCL settings with a taxonomy of existing techniques; (2) catastrophic forgetting\n(CF) prevention, (3) knowledge transfer (KT), which is particularly important\nfor NLP tasks; and (4) some theory and the hidden challenge of inter-task class\nseparation (ICS). (1), (3) and (4) have not been included in the existing\nsurvey. Finally, a list of future directions is discussed.\n", "date": "2022-11-23", "categories": []}
{"id": "2305.06555", "title": "Domain Incremental Lifelong Learning in an Open World", "abstract": "  Lifelong learning (LL) is an important ability for NLP models to learn new\ntasks continuously. Architecture-based approaches are reported to be effective\nimplementations for LL models. However, it is non-trivial to extend previous\napproaches to domain incremental LL scenarios since they either require access\nto task identities in the testing phase or cannot handle samples from unseen\ntasks. In this paper, we propose \\textbf{Diana}: a\n\\underline{d}ynam\\underline{i}c \\underline{a}rchitecture-based\nlifelo\\underline{n}g le\\underline{a}rning model that tries to learn a sequence\nof tasks with a prompt-enhanced language model. Four types of hierarchically\norganized prompts are used in Diana to capture knowledge from different\ngranularities. Specifically, we dedicate task-level prompts to capture\ntask-specific knowledge to retain high LL performances and maintain\ninstance-level prompts to learn knowledge shared across input samples to\nimprove the model's generalization performance. Moreover, we dedicate separate\nprompts to explicitly model unseen tasks and introduce a set of prompt key\nvectors to facilitate knowledge sharing between tasks. Extensive experiments\ndemonstrate that Diana outperforms state-of-the-art LL models, especially in\nhandling unseen tasks. We release the code and data at\n\\url{https://github.com/AlibabaResearch/DAMO-ConvAI/tree/main/diana}.\n", "date": "2023-05-11", "categories": []}
{"id": "2305.06616", "title": "Serial Contrastive Knowledge Distillation for Continual Few-shot\n  Relation Extraction", "abstract": "  Continual few-shot relation extraction (RE) aims to continuously train a\nmodel for new relations with few labeled training data, of which the major\nchallenges are the catastrophic forgetting of old relations and the overfitting\ncaused by data sparsity. In this paper, we propose a new model, namely SCKD, to\naccomplish the continual few-shot RE task. Specifically, we design serial\nknowledge distillation to preserve the prior knowledge from previous models and\nconduct contrastive learning with pseudo samples to keep the representations of\nsamples in different relations sufficiently distinguishable. Our experiments on\ntwo benchmark datasets validate the effectiveness of SCKD for continual\nfew-shot RE and its superiority in knowledge transfer and memory utilization\nover state-of-the-art models.\n", "date": "2023-05-11", "categories": []}
{"id": "2305.06620", "title": "Improving Continual Relation Extraction by Distinguishing Analogous\n  Semantics", "abstract": "  Continual relation extraction (RE) aims to learn constantly emerging\nrelations while avoiding forgetting the learned relations. Existing works store\na small number of typical samples to re-train the model for alleviating\nforgetting. However, repeatedly replaying these samples may cause the\noverfitting problem. We conduct an empirical study on existing works and\nobserve that their performance is severely affected by analogous relations. To\naddress this issue, we propose a novel continual extraction model for analogous\nrelations. Specifically, we design memory-insensitive relation prototypes and\nmemory augmentation to overcome the overfitting problem. We also introduce\nintegrated training and focal knowledge distillation to enhance the performance\non analogous relations. Experimental results show the superiority of our model\nand demonstrate its effectiveness in distinguishing analogous relations and\novercoming overfitting.\n", "date": "2023-05-11", "categories": []}
{"id": "2305.05968", "title": "Investigating Forgetting in Pre-Trained Representations Through\n  Continual Learning", "abstract": "  Representation forgetting refers to the drift of contextualized\nrepresentations during continual training. Intuitively, the representation\nforgetting can influence the general knowledge stored in pre-trained language\nmodels (LMs), but the concrete effect is still unclear. In this paper, we study\nthe effect of representation forgetting on the generality of pre-trained\nlanguage models, i.e. the potential capability for tackling future downstream\ntasks. Specifically, we design three metrics, including overall generality\ndestruction (GD), syntactic knowledge forgetting (SynF), and semantic knowledge\nforgetting (SemF), to measure the evolution of general knowledge in continual\nlearning. With extensive experiments, we find that the generality is destructed\nin various pre-trained LMs, and syntactic and semantic knowledge is forgotten\nthrough continual learning. Based on our experiments and analysis, we further\nget two insights into alleviating general knowledge forgetting: 1) training on\ngeneral linguistic tasks at first can mitigate general knowledge forgetting; 2)\nthe hybrid continual learning method can mitigate the generality destruction\nand maintain more general knowledge compared with those only considering\nrehearsal or regularization.\n", "date": "2023-05-10", "categories": []}
{"id": "2305.05175", "title": "SRIL: Selective Regularization for Class-Incremental Learning", "abstract": "  Human intelligence gradually accepts new information and accumulates\nknowledge throughout the lifespan. However, deep learning models suffer from a\ncatastrophic forgetting phenomenon, where they forget previous knowledge when\nacquiring new information. Class-Incremental Learning aims to create an\nintegrated model that balances plasticity and stability to overcome this\nchallenge. In this paper, we propose a selective regularization method that\naccepts new knowledge while maintaining previous knowledge. We first introduce\nan asymmetric feature distillation method for old and new classes inspired by\ncognitive science, using the gradient of classification and knowledge\ndistillation losses to determine whether to perform pattern completion or\npattern separation. We also propose a method to selectively interpolate the\nweight of the previous model for a balance between stability and plasticity,\nand we adjust whether to transfer through model confidence to ensure the\nperformance of the previous class and enable exploratory learning. We validate\nthe effectiveness of the proposed method, which surpasses the performance of\nexisting methods through extensive experimental protocols using CIFAR-100,\nImageNet-Subset, and ImageNet-Full.\n", "date": "2023-05-09", "categories": []}
{"id": "2305.04636", "title": "Enhancing Continual Relation Extraction via Classifier Decomposition", "abstract": "  Continual relation extraction (CRE) models aim at handling emerging new\nrelations while avoiding catastrophically forgetting old ones in the streaming\ndata. Though improvements have been shown by previous CRE studies, most of them\nonly adopt a vanilla strategy when models first learn representations of new\nrelations. In this work, we point out that there exist two typical biases after\ntraining of this vanilla strategy: classifier bias and representation bias,\nwhich causes the previous knowledge that the model learned to be shaded. To\nalleviate those biases, we propose a simple yet effective classifier\ndecomposition framework that splits the last FFN layer into separated previous\nand current classifiers, so as to maintain previous knowledge and encourage the\nmodel to learn more robust representations at this training stage. Experimental\nresults on two standard benchmarks show that our proposed framework\nconsistently outperforms the state-of-the-art CRE models, which indicates that\nthe importance of the first training stage to CRE models may be underestimated.\nOur code is available at https://github.com/hemingkx/CDec.\n", "date": "2023-05-08", "categories": []}
{"id": "2305.04769", "title": "BiRT: Bio-inspired Replay in Vision Transformers for Continual Learning", "abstract": "  The ability of deep neural networks to continually learn and adapt to a\nsequence of tasks has remained challenging due to catastrophic forgetting of\npreviously learned tasks. Humans, on the other hand, have a remarkable ability\nto acquire, assimilate, and transfer knowledge across tasks throughout their\nlifetime without catastrophic forgetting. The versatility of the brain can be\nattributed to the rehearsal of abstract experiences through a complementary\nlearning system. However, representation rehearsal in vision transformers lacks\ndiversity, resulting in overfitting and consequently, performance drops\nsignificantly compared to raw image rehearsal. Therefore, we propose BiRT, a\nnovel representation rehearsal-based continual learning approach using vision\ntransformers. Specifically, we introduce constructive noises at various stages\nof the vision transformer and enforce consistency in predictions with respect\nto an exponential moving average of the working model. Our method provides\nconsistent performance gain over raw image and vanilla representation rehearsal\non several challenging CL benchmarks, while being memory efficient and robust\nto natural and adversarial corruptions.\n", "date": "2023-05-08", "categories": []}
{"id": "2305.03648", "title": "On the Effectiveness of Equivariant Regularization for Robust Online\n  Continual Learning", "abstract": "  Humans can learn incrementally, whereas neural networks forget previously\nacquired information catastrophically. Continual Learning (CL) approaches seek\nto bridge this gap by facilitating the transfer of knowledge to both previous\ntasks (backward transfer) and future ones (forward transfer) during training.\n  Recent research has shown that self-supervision can produce versatile models\nthat can generalize well to diverse downstream tasks. However, contrastive\nself-supervised learning (CSSL), a popular self-supervision technique, has\nlimited effectiveness in online CL (OCL). OCL only permits one iteration of the\ninput dataset, and CSSL's low sample efficiency hinders its use on the input\ndata-stream.\n  In this work, we propose Continual Learning via Equivariant Regularization\n(CLER), an OCL approach that leverages equivariant tasks for self-supervision,\navoiding CSSL's limitations. Our method represents the first attempt at\ncombining equivariant knowledge with CL and can be easily integrated with\nexisting OCL methods. Extensive ablations shed light on how equivariant pretext\ntasks affect the network's information flow and its impact on CL dynamics.\n", "date": "2023-05-05", "categories": []}
{"id": "2304.14460", "title": "Gradient-based Maximally Interfered Retrieval for Domain Incremental 3D\n  Object Detection", "abstract": "  Accurate 3D object detection in all weather conditions remains a key\nchallenge to enable the widespread deployment of autonomous vehicles, as most\nwork to date has been performed on clear weather data. In order to generalize\nto adverse weather conditions, supervised methods perform best if trained from\nscratch on all weather data instead of finetuning a model pretrained on clear\nweather data. Training from scratch on all data will eventually become\ncomputationally infeasible and expensive as datasets continue to grow and\nencompass the full extent of possible weather conditions. On the other hand,\nnaive finetuning on data from a different weather domain can result in\ncatastrophic forgetting of the previously learned domain. Inspired by the\nsuccess of replay-based continual learning methods, we propose Gradient-based\nMaximally Interfered Retrieval (GMIR), a gradient based sampling strategy for\nreplay. During finetuning, GMIR periodically retrieves samples from the\nprevious domain dataset whose gradient vectors show maximal interference with\nthe gradient vector of the current update. Our 3D object detection experiments\non the SeeingThroughFog (STF) dataset show that GMIR not only overcomes\nforgetting but also offers competitive performance compared to scratch training\non all data with a 46.25% reduction in total training time.\n", "date": "2023-04-27", "categories": []}
{"id": "2305.01885", "title": "Evolving Dictionary Representation for Few-shot Class-incremental\n  Learning", "abstract": "  New objects are continuously emerging in the dynamically changing world and a\nreal-world artificial intelligence system should be capable of continual and\neffectual adaptation to new emerging classes without forgetting old ones. In\nview of this, in this paper we tackle a challenging and practical continual\nlearning scenario named few-shot class-incremental learning (FSCIL), in which\nlabeled data are given for classes in a base session but very limited labeled\ninstances are available for new incremental classes. To address this problem,\nwe propose a novel and succinct approach by introducing deep dictionary\nlearning which is a hybrid learning architecture that combines dictionary\nlearning and visual representation learning to provide a better space for\ncharacterizing different classes. We simultaneously optimize the dictionary and\nthe feature extraction backbone in the base session, while only finetune the\ndictionary in the incremental session for adaptation to novel classes, which\ncan alleviate the forgetting on base classes compared to finetuning the entire\nmodel. To further facilitate future adaptation, we also incorporate multiple\npseudo classes into the base session training so that certain space projected\nby dictionary can be reserved for future new concepts. The extensive\nexperimental results on CIFAR100, miniImageNet and CUB200 validate the\neffectiveness of our approach compared to other SOTA methods.\n", "date": "2023-05-03", "categories": []}
{"id": "2305.00380", "title": "DualHSIC: HSIC-Bottleneck and Alignment for Continual Learning", "abstract": "  Rehearsal-based approaches are a mainstay of continual learning (CL). They\nmitigate the catastrophic forgetting problem by maintaining a small fixed-size\nbuffer with a subset of data from past tasks. While most rehearsal-based\napproaches study how to effectively exploit the knowledge from the buffered\npast data, little attention is paid to the inter-task relationships with the\ncritical task-specific and task-invariant knowledge. By appropriately\nleveraging inter-task relationships, we propose a novel CL method named\nDualHSIC to boost the performance of existing rehearsal-based methods in a\nsimple yet effective way. DualHSIC consists of two complementary components\nthat stem from the so-called Hilbert Schmidt independence criterion (HSIC):\nHSIC-Bottleneck for Rehearsal (HBR) lessens the inter-task interference and\nHSIC Alignment (HA) promotes task-invariant knowledge sharing. Extensive\nexperiments show that DualHSIC can be seamlessly plugged into existing\nrehearsal-based methods for consistent performance improvements, and also\noutperforms recent state-of-the-art regularization-enhanced rehearsal methods.\nSource code will be released.\n", "date": "2023-04-30", "categories": []}
{"id": "2305.00620", "title": "Refined Response Distillation for Class-Incremental Player Detection", "abstract": "  Detecting players from sports broadcast videos is essential for intelligent\nevent analysis. However, existing methods assume fixed player categories,\nincapably accommodating the real-world scenarios where categories continue to\nevolve. Directly fine-tuning these methods on newly emerging categories also\nexist the catastrophic forgetting due to the non-stationary distribution.\nInspired by recent research on incremental object detection (IOD), we propose a\nRefined Response Distillation (R^2D) method to effectively mitigate\ncatastrophic forgetting for IOD tasks of the players. Firstly, we design a\nprogressive coarse-to-fine distillation region dividing scheme, separating\nhigh-value and low-value regions from classification and regression responses\nfor precise and fine-grained regional knowledge distillation. Subsequently, a\ntailored refined distillation strategy is developed on regions with varying\nsignificance to address the performance limitations posed by pronounced feature\nhomogeneity in the IOD tasks of the players. Furthermore, we present the\nNBA-IOD and Volleyball-IOD datasets as the benchmark and investigate the IOD\ntasks of the players systematically. Extensive experiments conducted on\nbenchmarks demonstrate that our method achieves state-of-the-art results.The\ncode and datasets are available at https://github.com/beiyan1911/Players-IOD.\n", "date": "2023-05-01", "categories": []}
{"id": "2304.14483", "title": "Adversary Aware Continual Learning", "abstract": "  Class incremental learning approaches are useful as they help the model to\nlearn new information (classes) sequentially, while also retaining the\npreviously acquired information (classes). However, it has been shown that such\napproaches are extremely vulnerable to the adversarial backdoor attacks, where\nan intelligent adversary can introduce small amount of misinformation to the\nmodel in the form of imperceptible backdoor pattern during training to cause\ndeliberate forgetting of a specific task or class at test time. In this work,\nwe propose a novel defensive framework to counter such an insidious attack\nwhere, we use the attacker's primary strength-hiding the backdoor pattern by\nmaking it imperceptible to humans-against it, and propose to learn a\nperceptible (stronger) pattern (also during the training) that can overpower\nthe attacker's imperceptible (weaker) pattern. We demonstrate the effectiveness\nof the proposed defensive mechanism through various commonly used Replay-based\n(both generative and exact replay-based) class incremental learning algorithms\nusing continual learning benchmark variants of CIFAR-10, CIFAR-100, and MNIST\ndatasets. Most noteworthy, our proposed defensive framework does not assume\nthat the attacker's target task and target class is known to the defender. The\ndefender is also unaware of the shape, size, and location of the attacker's\npattern. We show that our proposed defensive framework considerably improves\nthe performance of class incremental learning algorithms with no knowledge of\nthe attacker's target task, attacker's target class, and attacker's\nimperceptible pattern. We term our defensive framework as Adversary Aware\nContinual Learning (AACL).\n", "date": "2023-04-27", "categories": []}
{"id": "2302.03281", "title": "Utility-based Perturbed Gradient Descent: An Optimizer for Continual\n  Learning", "abstract": "  Modern representation learning methods often struggle to adapt quickly under\nnon-stationarity because they suffer from catastrophic forgetting and decaying\nplasticity. Such problems prevent learners from fast adaptation since they may\nforget useful features or have difficulty learning new ones. Hence, these\nmethods are rendered ineffective for continual learning. This paper proposes\nUtility-based Perturbed Gradient Descent (UPGD), an online learning algorithm\nwell-suited for continual learning agents. UPGD protects useful weights or\nfeatures from forgetting and perturbs less useful ones based on their\nutilities. Our empirical results show that UPGD helps reduce forgetting and\nmaintain plasticity, enabling modern representation learning methods to work\neffectively in continual learning.\n", "date": "2023-02-07", "categories": []}
{"id": "2304.10335", "title": "A baseline on continual learning methods for video action recognition", "abstract": "  Continual learning has recently attracted attention from the research\ncommunity, as it aims to solve long-standing limitations of classic\nsupervisedly-trained models. However, most research on this subject has tackled\ncontinual learning in simple image classification scenarios. In this paper, we\npresent a benchmark of state-of-the-art continual learning methods on video\naction recognition. Besides the increased complexity due to the temporal\ndimension, the video setting imposes stronger requirements on computing\nresources for top-performing rehearsal methods. To counteract the increased\nmemory requirements, we present two method-agnostic variants for rehearsal\nmethods, exploiting measures of either model confidence or data information to\nselect memorable samples. Our experiments show that, as expected from the\nliterature, rehearsal methods outperform other approaches; moreover, the\nproposed memory-efficient variants are shown to be effective at retaining a\ncertain level of performance with a smaller buffer size.\n", "date": "2023-04-20", "categories": []}
{"id": "2304.13327", "title": "Evaluation of Regularization-based Continual Learning Approaches:\n  Application to HAR", "abstract": "  Pervasive computing allows the provision of services in many important areas,\nincluding the relevant and dynamic field of health and well-being. In this\ndomain, Human Activity Recognition (HAR) has gained a lot of attention in\nrecent years. Current solutions rely on Machine Learning (ML) models and\nachieve impressive results. However, the evolution of these models remains\ndifficult, as long as a complete retraining is not performed. To overcome this\nproblem, the concept of Continual Learning is very promising today and, more\nparticularly, the techniques based on regularization. These techniques are\nparticularly interesting for their simplicity and their low cost. Initial\nstudies have been conducted and have shown promising outcomes. However, they\nremain very specific and difficult to compare. In this paper, we provide a\ncomprehensive comparison of three regularization-based methods that we adapted\nto the HAR domain, highlighting their strengths and limitations. Our\nexperiments were conducted on the UCI HAR dataset and the results showed that\nno single technique outperformed all others in all scenarios considered.\n", "date": "2023-04-26", "categories": []}
{"id": "2304.12067", "title": "Renate: A Library for Real-World Continual Learning", "abstract": "  Continual learning enables the incremental training of machine learning\nmodels on non-stationary data streams.While academic interest in the topic is\nhigh, there is little indication of the use of state-of-the-art continual\nlearning algorithms in practical machine learning deployment. This paper\npresents Renate, a continual learning library designed to build real-world\nupdating pipelines for PyTorch models. We discuss requirements for the use of\ncontinual learning algorithms in practice, from which we derive design\nprinciples for Renate. We give a high-level description of the library\ncomponents and interfaces. Finally, we showcase the strengths of the library by\npresenting experimental results. Renate may be found at\nhttps://github.com/awslabs/renate.\n", "date": "2023-04-24", "categories": []}
{"id": "2304.10857", "title": "SequeL: A Continual Learning Library in PyTorch and JAX", "abstract": "  Continual Learning is an important and challenging problem in machine\nlearning, where models must adapt to a continuous stream of new data without\nforgetting previously acquired knowledge. While existing frameworks are built\non PyTorch, the rising popularity of JAX might lead to divergent codebases,\nultimately hindering reproducibility and progress. To address this problem, we\nintroduce SequeL, a flexible and extensible library for Continual Learning that\nsupports both PyTorch and JAX frameworks. SequeL provides a unified interface\nfor a wide range of Continual Learning algorithms, including\nregularization-based approaches, replay-based approaches, and hybrid\napproaches. The library is designed towards modularity and simplicity, making\nthe API suitable for both researchers and practitioners. We release\nSequeL\\footnote{\\url{https://github.com/nik-dim/sequel}} as an open-source\nlibrary, enabling researchers and developers to easily experiment and extend\nthe library for their own purposes.\n", "date": "2023-04-21", "categories": []}
{"id": "2302.14250", "title": "Foundation Model Drives Weakly Incremental Learning for Semantic\n  Segmentation", "abstract": "  Modern incremental learning for semantic segmentation methods usually learn\nnew categories based on dense annotations. Although achieve promising results,\npixel-by-pixel labeling is costly and time-consuming. Weakly incremental\nlearning for semantic segmentation (WILSS) is a novel and attractive task,\nwhich aims at learning to segment new classes from cheap and widely available\nimage-level labels. Despite the comparable results, the image-level labels can\nnot provide details to locate each segment, which limits the performance of\nWILSS. This inspires us to think how to improve and effectively utilize the\nsupervision of new classes given image-level labels while avoiding forgetting\nold ones. In this work, we propose a novel and data-efficient framework for\nWILSS, named FMWISS. Specifically, we propose pre-training based\nco-segmentation to distill the knowledge of complementary foundation models for\ngenerating dense pseudo labels. We further optimize the noisy pseudo masks with\na teacher-student architecture, where a plug-in teacher is optimized with a\nproposed dense contrastive loss. Moreover, we introduce memory-based copy-paste\naugmentation to improve the catastrophic forgetting problem of old classes.\nExtensive experiments on Pascal VOC and COCO datasets demonstrate the superior\nperformance of our framework, e.g., FMWISS achieves 70.7% and 73.3% in the 15-5\nVOC setting, outperforming the state-of-the-art method by 3.4% and 6.1%,\nrespectively.\n", "date": "2023-02-28", "categories": []}
{"id": "2304.10103", "title": "eTag: Class-Incremental Learning with Embedding Distillation and\n  Task-Oriented Generation", "abstract": "  Class-Incremental Learning (CIL) aims to solve the neural networks'\ncatastrophic forgetting problem, which refers to the fact that once the network\nupdates on a new task, its performance on previously-learned tasks drops\ndramatically. Most successful CIL methods incrementally train a feature\nextractor with the aid of stored exemplars, or estimate the feature\ndistribution with the stored prototypes. However, the stored exemplars would\nviolate the data privacy concerns, while the stored prototypes might not\nreasonably be consistent with a proper feature distribution, hindering the\nexploration of real-world CIL applications. In this paper, we propose a method\nof \\textit{e}mbedding distillation and \\textit{Ta}sk-oriented\n\\textit{g}eneration (\\textit{eTag}) for CIL, which requires neither the\nexemplar nor the prototype. Instead, eTag achieves a data-free manner to train\nthe neural networks incrementally. To prevent the feature extractor from\nforgetting, eTag distills the embeddings of the network's intermediate blocks.\nAdditionally, eTag enables a generative network to produce suitable features,\nfitting the needs of the top incremental classifier. Experimental results\nconfirmed that our proposed eTag considerably outperforms the state-of-the-art\nmethods on CIFAR-100 and ImageNet-sub\\footnote{Our code is available in the\nSupplementary Materials.\n", "date": "2023-04-20", "categories": []}
{"id": "2304.10113", "title": "SATA: Source Anchoring and Target Alignment Network for Continual Test\n  Time Adaptation", "abstract": "  Adapting a trained model to perform satisfactorily on continually changing\ntesting domains/environments is an important and challenging task. In this\nwork, we propose a novel framework, SATA, which aims to satisfy the following\ncharacteristics required for online adaptation: 1) can work seamlessly with\ndifferent (preferably small) batch sizes to reduce latency; 2) should continue\nto work well for the source domain; 3) should have minimal tunable\nhyper-parameters and storage requirements. Given a pre-trained network trained\non source domain data, the proposed SATA framework modifies the batch-norm\naffine parameters using source anchoring based self-distillation. This ensures\nthat the model incorporates the knowledge of the newly encountered domains,\nwithout catastrophically forgetting about the previously seen ones. We also\npropose a source-prototype driven contrastive alignment to ensure natural\ngrouping of the target samples, while maintaining the already learnt semantic\ninformation. Extensive evaluation on three benchmark datasets under challenging\nsettings justify the effectiveness of SATA for real-world applications.\n", "date": "2023-04-20", "categories": []}
{"id": "2304.10177", "title": "Regularizing Second-Order Influences for Continual Learning", "abstract": "  Continual learning aims to learn on non-stationary data streams without\ncatastrophically forgetting previous knowledge. Prevalent replay-based methods\naddress this challenge by rehearsing on a small buffer holding the seen data,\nfor which a delicate sample selection strategy is required. However, existing\nselection schemes typically seek only to maximize the utility of the ongoing\nselection, overlooking the interference between successive rounds of selection.\nMotivated by this, we dissect the interaction of sequential selection steps\nwithin a framework built on influence functions. We manage to identify a new\nclass of second-order influences that will gradually amplify incidental bias in\nthe replay buffer and compromise the selection process. To regularize the\nsecond-order effects, a novel selection objective is proposed, which also has\nclear connections to two widely adopted criteria. Furthermore, we present an\nefficient implementation for optimizing the proposed criterion. Experiments on\nmultiple continual learning benchmarks demonstrate the advantage of our\napproach over state-of-the-art methods. Code is available at\nhttps://github.com/feifeiobama/InfluenceCL.\n", "date": "2023-04-20", "categories": []}
{"id": "2304.07374", "title": "Continual Source-Free Unsupervised Domain Adaptation", "abstract": "  Existing Source-free Unsupervised Domain Adaptation (SUDA) approaches\ninherently exhibit catastrophic forgetting. Typically, models trained on a\nlabeled source domain and adapted to unlabeled target data improve performance\non the target while dropping performance on the source, which is not available\nduring adaptation. In this study, our goal is to cope with the challenging\nproblem of SUDA in a continual learning setting, i.e., adapting to the\ntarget(s) with varying distributional shifts while maintaining performance on\nthe source. The proposed framework consists of two main stages: i) a SUDA model\nyielding cleaner target labels -- favoring good performance on target, and ii)\na novel method for synthesizing class-conditioned source-style images by\nleveraging only the source model and pseudo-labeled target data as a prior. An\nextensive pool of experiments on major benchmarks, e.g., PACS, Visda-C, and\nDomainNet demonstrates that the proposed Continual SUDA (C-SUDA) framework\nenables preserving satisfactory performance on the source domain without\nexploiting the source data at all.\n", "date": "2023-04-14", "categories": []}
{"id": "2304.07421", "title": "Peer-to-Peer Federated Continual Learning for Naturalistic Driving\n  Action Recognition", "abstract": "  Naturalistic driving action recognition (NDAR) has proven to be an effective\nmethod for detecting driver distraction and reducing the risk of traffic\naccidents. However, the intrusive design of in-cabin cameras raises concerns\nabout driver privacy. To address this issue, we propose a novel peer-to-peer\n(P2P) federated learning (FL) framework with continual learning, namely FedPC,\nwhich ensures privacy and enhances learning efficiency while reducing\ncommunication, computational, and storage overheads. Our framework focuses on\naddressing the clients' objectives within a serverless FL framework, with the\ngoal of delivering personalized and accurate NDAR models. We demonstrate and\nevaluate the performance of FedPC on two real-world NDAR datasets, including\nthe State Farm Distracted Driver Detection and Track 3 NDAR dataset in the 2023\nAICity Challenge. The results of our experiments highlight the strong\ncompetitiveness of FedPC compared to the conventional client-to-server (C2S)\nFLs in terms of performance, knowledge dissemination rate, and compatibility\nwith new clients.\n", "date": "2023-04-14", "categories": []}
{"id": "2304.07560", "title": "Continual Domain Adaptation through Pruning-aided Domain-specific Weight\n  Modulation", "abstract": "  In this paper, we propose to develop a method to address unsupervised domain\nadaptation (UDA) in a practical setting of continual learning (CL). The goal is\nto update the model on continually changing domains while preserving\ndomain-specific knowledge to prevent catastrophic forgetting of past-seen\ndomains. To this end, we build a framework for preserving domain-specific\nfeatures utilizing the inherent model capacity via pruning. We also perform\neffective inference using a novel batch-norm based metric to predict the final\nmodel parameters to be used accurately. Our approach achieves not only\nstate-of-the-art performance but also prevents catastrophic forgetting of past\ndomains significantly. Our code is made publicly available.\n", "date": "2023-04-15", "categories": []}
{"id": "2304.06738", "title": "A Study of Biologically Plausible Neural Network: The Role and\n  Interactions of Brain-Inspired Mechanisms in Continual Learning", "abstract": "  Humans excel at continually acquiring, consolidating, and retaining\ninformation from an ever-changing environment, whereas artificial neural\nnetworks (ANNs) exhibit catastrophic forgetting. There are considerable\ndifferences in the complexity of synapses, the processing of information, and\nthe learning mechanisms in biological neural networks and their artificial\ncounterparts, which may explain the mismatch in performance. We consider a\nbiologically plausible framework that constitutes separate populations of\nexclusively excitatory and inhibitory neurons that adhere to Dale's principle,\nand the excitatory pyramidal neurons are augmented with dendritic-like\nstructures for context-dependent processing of stimuli. We then conduct a\ncomprehensive study on the role and interactions of different mechanisms\ninspired by the brain, including sparse non-overlapping representations,\nHebbian learning, synaptic consolidation, and replay of past activations that\naccompanied the learning event. Our study suggests that the employing of\nmultiple complementary mechanisms in a biologically plausible architecture,\nsimilar to the brain, may be effective in enabling continual learning in ANNs.\n", "date": "2023-04-13", "categories": []}
{"id": "2304.06971", "title": "Preserving Locality in Vision Transformers for Class Incremental\n  Learning", "abstract": "  Learning new classes without forgetting is crucial for real-world\napplications for a classification model. Vision Transformers (ViT) recently\nachieve remarkable performance in Class Incremental Learning (CIL). Previous\nworks mainly focus on block design and model expansion for ViTs. However, in\nthis paper, we find that when the ViT is incrementally trained, the attention\nlayers gradually lose concentration on local features. We call this interesting\nphenomenon as \\emph{Locality Degradation} in ViTs for CIL. Since the low-level\nlocal information is crucial to the transferability of the representation, it\nis beneficial to preserve the locality in attention layers. In this paper, we\nencourage the model to preserve more local information as the training\nprocedure goes on and devise a Locality-Preserved Attention (LPA) layer to\nemphasize the importance of local features. Specifically, we incorporate the\nlocal information directly into the vanilla attention and control the initial\ngradients of the vanilla attention by weighting it with a small initial value.\nExtensive experiments show that the representations facilitated by LPA capture\nmore low-level general information which is easier to transfer to follow-up\ntasks. The improved model gets consistently better performance on CIFAR100 and\nImageNet100.\n", "date": "2023-04-14", "categories": []}
{"id": "2302.03241", "title": "Continual Pre-training of Language Models", "abstract": "  Language models (LMs) have been instrumental for the rapid advance of natural\nlanguage processing. This paper studies continual pre-training of LMs, in\nparticular, continual domain-adaptive pre-training (or continual DAP-training).\nExisting research has shown that further pre-training an LM using a domain\ncorpus to adapt the LM to the domain can improve the end-task performance in\nthe domain. This paper proposes a novel method to continually DAP-train an LM\nwith a sequence of unlabeled domain corpora to adapt the LM to these domains to\nimprove their end-task performances. The key novelty of our method is a\nsoft-masking mechanism that directly controls the update to the LM. A novel\nproxy is also proposed to preserve the general knowledge in the original LM.\nAdditionally, it contrasts the representations of the previously learned domain\nknowledge (including the general knowledge in the pre-trained LM) and the\nknowledge from the current full network to achieve knowledge integration. The\nmethod not only overcomes catastrophic forgetting, but also achieves knowledge\ntransfer to improve end-task performances. Empirical evaluation demonstrates\nthe effectiveness of the proposed method.\n", "date": "2023-02-07", "categories": []}
{"id": "2304.05547", "title": "Taxonomic Class Incremental Learning", "abstract": "  The problem of continual learning has attracted rising attention in recent\nyears. However, few works have questioned the commonly used learning setup,\nbased on a task curriculum of random class. This differs significantly from\nhuman continual learning, which is guided by taxonomic curricula. In this work,\nwe propose the Taxonomic Class Incremental Learning (TCIL) problem. In TCIL,\nthe task sequence is organized based on a taxonomic class tree. We unify\nexisting approaches to CIL and taxonomic learning as parameter inheritance\nschemes and introduce a new such scheme for the TCIL learning. This enables the\nincremental transfer of knowledge from ancestor to descendant class of a class\ntaxonomy through parameter inheritance. Experiments on CIFAR-100 and\nImageNet-100 show the effectiveness of the proposed TCIL method, which\noutperforms existing SOTA methods by 2% in terms of final accuracy on CIFAR-100\nand 3% on ImageNet-100.\n", "date": "2023-04-12", "categories": []}
{"id": "2304.05734", "title": "Few-shot Class-incremental Learning for Cross-domain Disease\n  Classification", "abstract": "  The ability to incrementally learn new classes from limited samples is\ncrucial to the development of artificial intelligence systems for real clinical\napplication. Although existing incremental learning techniques have attempted\nto address this issue, they still struggle with only few labeled data,\nparticularly when the samples are from varied domains. In this paper, we\nexplore the cross-domain few-shot incremental learning (CDFSCIL) problem.\nCDFSCIL requires models to learn new classes from very few labeled samples\nincrementally, and the new classes may be vastly different from the target\nspace. To counteract this difficulty, we propose a cross-domain enhancement\nconstraint and cross-domain data augmentation method. Experiments on MedMNIST\nshow that the classification performance of this method is better than other\nsimilar incremental learning methods.\n", "date": "2023-04-12", "categories": []}
{"id": "2208.11266", "title": "SCALE: Online Self-Supervised Lifelong Learning without Prior Knowledge", "abstract": "  Unsupervised lifelong learning refers to the ability to learn over time while\nmemorizing previous patterns without supervision. Although great progress has\nbeen made in this direction, existing work often assumes strong prior knowledge\nabout the incoming data (e.g., knowing the class boundaries), which can be\nimpossible to obtain in complex and unpredictable environments. In this paper,\nmotivated by real-world scenarios, we propose a more practical problem setting\ncalled online self-supervised lifelong learning without prior knowledge. The\nproposed setting is challenging due to the non-iid and single-pass data, the\nabsence of external supervision, and no prior knowledge. To address the\nchallenges, we propose Self-Supervised ContrAstive Lifelong LEarning without\nPrior Knowledge (SCALE) which can extract and memorize representations on the\nfly purely from the data continuum. SCALE is designed around three major\ncomponents: a pseudo-supervised contrastive loss, a self-supervised forgetting\nloss, and an online memory update for uniform subset selection. All three\ncomponents are designed to work collaboratively to maximize learning\nperformance. We perform comprehensive experiments of SCALE under iid and four\nnon-iid data streams. The results show that SCALE outperforms the\nstate-of-the-art algorithm in all settings with improvements up to 3.83%, 2.77%\nand 5.86% in terms of kNN accuracy on CIFAR-10, CIFAR-100, and TinyImageNet\ndatasets.\n", "date": "2022-08-24", "categories": []}
{"id": "2304.05015", "title": "Continual Semantic Segmentation with Automatic Memory Sample Selection", "abstract": "  Continual Semantic Segmentation (CSS) extends static semantic segmentation by\nincrementally introducing new classes for training. To alleviate the\ncatastrophic forgetting issue in CSS, a memory buffer that stores a small\nnumber of samples from the previous classes is constructed for replay. However,\nexisting methods select the memory samples either randomly or based on a\nsingle-factor-driven handcrafted strategy, which has no guarantee to be\noptimal. In this work, we propose a novel memory sample selection mechanism\nthat selects informative samples for effective replay in a fully automatic way\nby considering comprehensive factors including sample diversity and class\nperformance. Our mechanism regards the selection operation as a decision-making\nprocess and learns an optimal selection policy that directly maximizes the\nvalidation performance on a reward set. To facilitate the selection decision,\nwe design a novel state representation and a dual-stage action space. Our\nextensive experiments on Pascal-VOC 2012 and ADE 20K datasets demonstrate the\neffectiveness of our approach with state-of-the-art (SOTA) performance\nachieved, outperforming the second-place one by 12.54% for the 6stage setting\non Pascal-VOC 2012.\n", "date": "2023-04-11", "categories": []}
{"id": "2304.05288", "title": "Task Difficulty Aware Parameter Allocation & Regularization for Lifelong\n  Learning", "abstract": "  Parameter regularization or allocation methods are effective in overcoming\ncatastrophic forgetting in lifelong learning. However, they solve all tasks in\na sequence uniformly and ignore the differences in the learning difficulty of\ndifferent tasks. So parameter regularization methods face significant\nforgetting when learning a new task very different from learned tasks, and\nparameter allocation methods face unnecessary parameter overhead when learning\nsimple tasks. In this paper, we propose the Parameter Allocation &\nRegularization (PAR), which adaptively select an appropriate strategy for each\ntask from parameter allocation and regularization based on its learning\ndifficulty. A task is easy for a model that has learned tasks related to it and\nvice versa. We propose a divergence estimation method based on the\nNearest-Prototype distance to measure the task relatedness using only features\nof the new task. Moreover, we propose a time-efficient relatedness-aware\nsampling-based architecture search strategy to reduce the parameter overhead\nfor allocation. Experimental results on multiple benchmarks demonstrate that,\ncompared with SOTAs, our method is scalable and significantly reduces the\nmodel's redundancy while improving the model's performance. Further qualitative\nanalysis indicates that PAR obtains reasonable task-relatedness.\n", "date": "2023-04-11", "categories": []}
{"id": "2304.05362", "title": "MASIL: Towards Maximum Separable Class Representation for Few Shot Class\n  Incremental Learning", "abstract": "  Few Shot Class Incremental Learning (FSCIL) with few examples per class for\neach incremental session is the realistic setting of continual learning since\nobtaining large number of annotated samples is not feasible and cost effective.\nWe present the framework MASIL as a step towards learning the maximal separable\nclassifier. It addresses the common problem i.e forgetting of old classes and\nover-fitting to novel classes by learning the classifier weights to be\nmaximally separable between classes forming a simplex Equiangular Tight Frame.\nWe propose the idea of concept factorization explaining the collapsed features\nfor base session classes in terms of concept basis and use these to induce\nclassifier simplex for few shot classes. We further adds fine tuning to reduce\nany error occurred during factorization and train the classifier jointly on\nbase and novel classes without retaining any base class samples in memory.\nExperimental results on miniImageNet, CIFAR-100 and CUB-200 demonstrate that\nMASIL outperforms all the benchmarks.\n", "date": "2023-04-08", "categories": []}
{"id": "2303.14042", "title": "Class-Incremental Exemplar Compression for Class-Incremental Learning", "abstract": "  Exemplar-based class-incremental learning (CIL) finetunes the model with all\nsamples of new classes but few-shot exemplars of old classes in each\nincremental phase, where the \"few-shot\" abides by the limited memory budget. In\nthis paper, we break this \"few-shot\" limit based on a simple yet surprisingly\neffective idea: compressing exemplars by downsampling non-discriminative pixels\nand saving \"many-shot\" compressed exemplars in the memory. Without needing any\nmanual annotation, we achieve this compression by generating 0-1 masks on\ndiscriminative pixels from class activation maps (CAM). We propose an adaptive\nmask generation model called class-incremental masking (CIM) to explicitly\nresolve two difficulties of using CAM: 1) transforming the heatmaps of CAM to\n0-1 masks with an arbitrary threshold leads to a trade-off between the coverage\non discriminative pixels and the quantity of exemplars, as the total memory is\nfixed; and 2) optimal thresholds vary for different object classes, which is\nparticularly obvious in the dynamic environment of CIL. We optimize the CIM\nmodel alternatively with the conventional CIL model through a bilevel\noptimization problem. We conduct extensive experiments on high-resolution CIL\nbenchmarks including Food-101, ImageNet-100, and ImageNet-1000, and show that\nusing the compressed exemplars by CIM can achieve a new state-of-the-art CIL\naccuracy, e.g., 4.8 percentage points higher than FOSTER on 10-Phase\nImageNet-1000. Our code is available at https://github.com/xfflzl/CIM-CIL.\n", "date": "2023-03-24", "categories": []}
{"id": "2304.03931", "title": "Exploring Data Geometry for Continual Learning", "abstract": "  Continual learning aims to efficiently learn from a non-stationary stream of\ndata while avoiding forgetting the knowledge of old data. In many practical\napplications, data complies with non-Euclidean geometry. As such, the commonly\nused Euclidean space cannot gracefully capture non-Euclidean geometric\nstructures of data, leading to inferior results. In this paper, we study\ncontinual learning from a novel perspective by exploring data geometry for the\nnon-stationary stream of data. Our method dynamically expands the geometry of\nthe underlying space to match growing geometric structures induced by new data,\nand prevents forgetting by keeping geometric structures of old data into\naccount. In doing so, making use of the mixed curvature space, we propose an\nincremental search scheme, through which the growing geometric structures are\nencoded. Then, we introduce an angular-regularization loss and a\nneighbor-robustness loss to train the model, capable of penalizing the change\nof global geometric structures and local geometric structures. Experiments show\nthat our method achieves better performance than baseline methods designed in\nEuclidean space.\n", "date": "2023-04-08", "categories": []}
{"id": "2304.04152", "title": "Continual Graph Convolutional Network for Text Classification", "abstract": "  Graph convolutional network (GCN) has been successfully applied to capture\nglobal non-consecutive and long-distance semantic information for text\nclassification. However, while GCN-based methods have shown promising results\nin offline evaluations, they commonly follow a seen-token-seen-document\nparadigm by constructing a fixed document-token graph and cannot make\ninferences on new documents. It is a challenge to deploy them in online systems\nto infer steaming text data. In this work, we present a continual GCN model\n(ContGCN) to generalize inferences from observed documents to unobserved\ndocuments. Concretely, we propose a new all-token-any-document paradigm to\ndynamically update the document-token graph in every batch during both the\ntraining and testing phases of an online system. Moreover, we design an\noccurrence memory module and a self-supervised contrastive learning objective\nto update ContGCN in a label-free manner. A 3-month A/B test on Huawei public\nopinion analysis system shows ContGCN achieves 8.86% performance gain compared\nwith state-of-the-art methods. Offline experiments on five public datasets also\nshow ContGCN can improve inference quality. The source code will be released at\nhttps://github.com/Jyonn/ContGCN.\n", "date": "2023-04-09", "categories": []}
{"id": "2304.04158", "title": "Does Continual Learning Equally Forget All Parameters?", "abstract": "  Distribution shift (e.g., task or domain shift) in continual learning (CL)\nusually results in catastrophic forgetting of neural networks. Although it can\nbe alleviated by repeatedly replaying buffered data, the every-step replay is\ntime-consuming. In this paper, we study which modules in neural networks are\nmore prone to forgetting by investigating their training dynamics during CL.\nOur proposed metrics show that only a few modules are more task-specific and\nsensitively alter between tasks, while others can be shared across tasks as\ncommon knowledge. Hence, we attribute forgetting mainly to the former and find\nthat finetuning them only on a small buffer at the end of any CL method can\nbring non-trivial improvement. Due to the small number of finetuned parameters,\nsuch ``Forgetting Prioritized Finetuning (FPF)'' is efficient in computation.\nWe further propose a more efficient and simpler method that entirely removes\nthe every-step replay and replaces them by only $k$-times of FPF periodically\ntriggered during CL. Surprisingly, this ``$k$-FPF'' performs comparably to FPF\nand outperforms the SOTA CL methods but significantly reduces their\ncomputational overhead and cost. In experiments on several benchmarks of class-\nand domain-incremental CL, FPF consistently improves existing CL methods by a\nlarge margin, and $k$-FPF further excels in efficiency without degrading the\naccuracy. We also empirically studied the impact of buffer size, epochs per\ntask, and finetuning modules on the cost and accuracy of our methods.\n", "date": "2023-04-09", "categories": []}
{"id": "2304.04222", "title": "CILIATE: Towards Fairer Class-based Incremental Learning by Dataset and\n  Training Refinement", "abstract": "  Due to the model aging problem, Deep Neural Networks (DNNs) need updates to\nadjust them to new data distributions. The common practice leverages\nincremental learning (IL), e.g., Class-based Incremental Learning (CIL) that\nupdates output labels, to update the model with new data and a limited number\nof old data. This avoids heavyweight training (from scratch) using conventional\nmethods and saves storage space by reducing the number of old data to store.\nBut it also leads to poor performance in fairness. In this paper, we show that\nCIL suffers both dataset and algorithm bias problems, and existing solutions\ncan only partially solve the problem. We propose a novel framework, CILIATE,\nthat fixes both dataset and algorithm bias in CIL. It features a novel\ndifferential analysis guided dataset and training refinement process that\nidentifies unique and important samples overlooked by existing CIL and enforces\nthe model to learn from them. Through this process, CILIATE improves the\nfairness of CIL by 17.03%, 22.46%, and 31.79% compared to state-of-the-art\nmethods, iCaRL, BiC, and WA, respectively, based on our evaluation on three\npopular datasets and widely used ResNet models.\n", "date": "2023-04-09", "categories": []}
{"id": "2304.04408", "title": "PCR: Proxy-based Contrastive Replay for Online Class-Incremental\n  Continual Learning", "abstract": "  Online class-incremental continual learning is a specific task of continual\nlearning. It aims to continuously learn new classes from data stream and the\nsamples of data stream are seen only once, which suffers from the catastrophic\nforgetting issue, i.e., forgetting historical knowledge of old classes.\nExisting replay-based methods effectively alleviate this issue by saving and\nreplaying part of old data in a proxy-based or contrastive-based replay manner.\nAlthough these two replay manners are effective, the former would incline to\nnew classes due to class imbalance issues, and the latter is unstable and hard\nto converge because of the limited number of samples. In this paper, we conduct\na comprehensive analysis of these two replay manners and find that they can be\ncomplementary. Inspired by this finding, we propose a novel replay-based method\ncalled proxy-based contrastive replay (PCR). The key operation is to replace\nthe contrastive samples of anchors with corresponding proxies in the\ncontrastive-based way. It alleviates the phenomenon of catastrophic forgetting\nby effectively addressing the imbalance issue, as well as keeps a faster\nconvergence of the model. We conduct extensive experiments on three real-world\nbenchmark datasets, and empirical results consistently demonstrate the\nsuperiority of PCR over various state-of-the-art methods.\n", "date": "2023-04-10", "categories": []}
{"id": "2304.03626", "title": "Asynchronous Federated Continual Learning", "abstract": "  The standard class-incremental continual learning setting assumes a set of\ntasks seen one after the other in a fixed and predefined order. This is not\nvery realistic in federated learning environments where each client works\nindependently in an asynchronous manner getting data for the different tasks in\ntime-frames and orders totally uncorrelated with the other ones. We introduce a\nnovel federated learning setting (AFCL) where the continual learning of\nmultiple tasks happens at each client with different orderings and in\nasynchronous time slots. We tackle this novel task using prototype-based\nlearning, a representation loss, fractal pre-training, and a modified\naggregation policy. Our approach, called FedSpace, effectively tackles this\ntask as shown by the results on the CIFAR-100 dataset using 3 different\nfederated splits with 50, 100, and 500 clients, respectively. The code and\nfederated splits are available at https://github.com/LTTM/FedSpace.\n", "date": "2023-04-07", "categories": []}
{"id": "2301.11417", "title": "Are Labels Needed for Incremental Instance Learning?", "abstract": "  In this paper, we learn to classify visual object instances, incrementally\nand via self-supervision (self-incremental). Our learner observes a single\ninstance at a time, which is then discarded from the dataset. Incremental\ninstance learning is challenging, since longer learning sessions exacerbate\nforgetfulness, and labeling instances is cumbersome. We overcome these\nchallenges via three contributions: i. We propose VINIL, a self-incremental\nlearner that can learn object instances sequentially, ii. We equip VINIL with\nself-supervision to by-pass the need for instance labelling, iii. We compare\nVINIL to label-supervised variants on two large-scale benchmarks, and show that\nVINIL significantly improves accuracy while reducing forgetfulness.\n", "date": "2023-01-26", "categories": []}
{"id": "2212.04842", "title": "PIVOT: Prompting for Video Continual Learning", "abstract": "  Modern machine learning pipelines are limited due to data availability,\nstorage quotas, privacy regulations, and expensive annotation processes. These\nconstraints make it difficult or impossible to train and update large-scale\nmodels on such dynamic annotated sets. Continual learning directly approaches\nthis problem, with the ultimate goal of devising methods where a deep neural\nnetwork effectively learns relevant patterns for new (unseen) classes, without\nsignificantly altering its performance on previously learned ones. In this\npaper, we address the problem of continual learning for video data. We\nintroduce PIVOT, a novel method that leverages extensive knowledge in\npre-trained models from the image domain, thereby reducing the number of\ntrainable parameters and the associated forgetting. Unlike previous methods,\nours is the first approach that effectively uses prompting mechanisms for\ncontinual learning without any in-domain pre-training. Our experiments show\nthat PIVOT improves state-of-the-art methods by a significant 27% on the\n20-task ActivityNet setup.\n", "date": "2022-12-09", "categories": []}
{"id": "2304.01663", "title": "On the Stability-Plasticity Dilemma of Class-Incremental Learning", "abstract": "  A primary goal of class-incremental learning is to strike a balance between\nstability and plasticity, where models should be both stable enough to retain\nknowledge learned from previously seen classes, and plastic enough to learn\nconcepts from new classes. While previous works demonstrate strong performance\non class-incremental benchmarks, it is not clear whether their success comes\nfrom the models being stable, plastic, or a mixture of both. This paper aims to\nshed light on how effectively recent class-incremental learning algorithms\naddress the stability-plasticity trade-off. We establish analytical tools that\nmeasure the stability and plasticity of feature representations, and employ\nsuch tools to investigate models trained with various algorithms on large-scale\nclass-incremental benchmarks. Surprisingly, we find that the majority of\nclass-incremental learning algorithms heavily favor stability over plasticity,\nto the extent that the feature extractor of a model trained on the initial set\nof classes is no less effective than that of the final incremental model. Our\nobservations not only inspire two simple algorithms that highlight the\nimportance of feature representation analysis, but also suggest that\nclass-incremental learning approaches, in general, should strive for better\nfeature representation learning.\n", "date": "2023-04-04", "categories": []}
{"id": "2303.18171", "title": "How Efficient Are Today's Continual Learning Algorithms?", "abstract": "  Supervised Continual learning involves updating a deep neural network (DNN)\nfrom an ever-growing stream of labeled data. While most work has focused on\novercoming catastrophic forgetting, one of the major motivations behind\ncontinual learning is being able to efficiently update a network with new\ninformation, rather than retraining from scratch on the training dataset as it\ngrows over time. Despite recent continual learning methods largely solving the\ncatastrophic forgetting problem, there has been little attention paid to the\nefficiency of these algorithms. Here, we study recent methods for incremental\nclass learning and illustrate that many are highly inefficient in terms of\ncompute, memory, and storage. Some methods even require more compute than\ntraining from scratch! We argue that for continual learning to have real-world\napplicability, the research community cannot ignore the resources used by these\nalgorithms. There is more to continual learning than mitigating catastrophic\nforgetting.\n", "date": "2023-03-29", "categories": []}
{"id": "2304.00426", "title": "Learning with Fantasy: Semantic-Aware Virtual Contrastive Constraint for\n  Few-Shot Class-Incremental Learning", "abstract": "  Few-shot class-incremental learning (FSCIL) aims at learning to classify new\nclasses continually from limited samples without forgetting the old classes.\nThe mainstream framework tackling FSCIL is first to adopt the cross-entropy\n(CE) loss for training at the base session, then freeze the feature extractor\nto adapt to new classes. However, in this work, we find that the CE loss is not\nideal for the base session training as it suffers poor class separation in\nterms of representations, which further degrades generalization to novel\nclasses. One tempting method to mitigate this problem is to apply an additional\nnaive supervised contrastive learning (SCL) in the base session. Unfortunately,\nwe find that although SCL can create a slightly better representation\nseparation among different base classes, it still struggles to separate base\nclasses and new classes. Inspired by the observations made, we propose\nSemantic-Aware Virtual Contrastive model (SAVC), a novel method that\nfacilitates separation between new classes and base classes by introducing\nvirtual classes to SCL. These virtual classes, which are generated via\npre-defined transformations, not only act as placeholders for unseen classes in\nthe representation space, but also provide diverse semantic information. By\nlearning to recognize and contrast in the fantasy space fostered by virtual\nclasses, our SAVC significantly boosts base class separation and novel class\ngeneralization, achieving new state-of-the-art performance on the three\nwidely-used FSCIL benchmark datasets. Code is available at:\nhttps://github.com/zysong0113/SAVC.\n", "date": "2023-04-02", "categories": []}
{"id": "2303.09483", "title": "Achieving a Better Stability-Plasticity Trade-off via Auxiliary Networks\n  in Continual Learning", "abstract": "  In contrast to the natural capabilities of humans to learn new tasks in a\nsequential fashion, neural networks are known to suffer from catastrophic\nforgetting, where the model's performances on old tasks drop dramatically after\nbeing optimized for a new task. Since then, the continual learning (CL)\ncommunity has proposed several solutions aiming to equip the neural network\nwith the ability to learn the current task (plasticity) while still achieving\nhigh accuracy on the previous tasks (stability). Despite remarkable\nimprovements, the plasticity-stability trade-off is still far from being solved\nand its underlying mechanism is poorly understood. In this work, we propose\nAuxiliary Network Continual Learning (ANCL), a novel method that applies an\nadditional auxiliary network which promotes plasticity to the continually\nlearned model which mainly focuses on stability. More concretely, the proposed\nframework materializes in a regularizer that naturally interpolates between\nplasticity and stability, surpassing strong baselines on task incremental and\nclass incremental scenarios. Through extensive analyses on ANCL solutions, we\nidentify some essential principles beneath the stability-plasticity trade-off.\n", "date": "2023-03-16", "categories": []}
{"id": "2303.17764", "title": "Towards Adversarially Robust Continual Learning", "abstract": "  Recent studies show that models trained by continual learning can achieve the\ncomparable performances as the standard supervised learning and the learning\nflexibility of continual learning models enables their wide applications in the\nreal world. Deep learning models, however, are shown to be vulnerable to\nadversarial attacks. Though there are many studies on the model robustness in\nthe context of standard supervised learning, protecting continual learning from\nadversarial attacks has not yet been investigated. To fill in this research\ngap, we are the first to study adversarial robustness in continual learning and\npropose a novel method called \\textbf{T}ask-\\textbf{A}ware \\textbf{B}oundary\n\\textbf{A}ugmentation (TABA) to boost the robustness of continual learning\nmodels. With extensive experiments on CIFAR-10 and CIFAR-100, we show the\nefficacy of adversarial training and TABA in defending adversarial attacks.\n", "date": "2023-03-31", "categories": []}
{"id": "2211.09790", "title": "ConStruct-VL: Data-Free Continual Structured VL Concepts Learning", "abstract": "  Recently, large-scale pre-trained Vision-and-Language (VL) foundation models\nhave demonstrated remarkable capabilities in many zero-shot downstream tasks,\nachieving competitive results for recognizing objects defined by as little as\nshort text prompts. However, it has also been shown that VL models are still\nbrittle in Structured VL Concept (SVLC) reasoning, such as the ability to\nrecognize object attributes, states, and inter-object relations. This leads to\nreasoning mistakes, which need to be corrected as they occur by teaching VL\nmodels the missing SVLC skills; often this must be done using private data\nwhere the issue was found, which naturally leads to a data-free continual (no\ntask-id) VL learning setting. In this work, we introduce the first Continual\nData-Free Structured VL Concepts Learning (ConStruct-VL) benchmark and show it\nis challenging for many existing data-free CL strategies. We, therefore,\npropose a data-free method comprised of a new approach of Adversarial\nPseudo-Replay (APR) which generates adversarial reminders of past tasks from\npast task models. To use this method efficiently, we also propose a continual\nparameter-efficient Layered-LoRA (LaLo) neural architecture allowing\nno-memory-cost access to all past models at train time. We show this approach\noutperforms all data-free methods by as much as ~7% while even matching some\nlevels of experience-replay (prohibitive for applications where data-privacy\nmust be preserved). Our code is publicly available at\nhttps://github.com/jamessealesmith/ConStruct-VL\n", "date": "2022-11-17", "categories": []}
{"id": "2211.13218", "title": "CODA-Prompt: COntinual Decomposed Attention-based Prompting for\n  Rehearsal-Free Continual Learning", "abstract": "  Computer vision models suffer from a phenomenon known as catastrophic\nforgetting when learning novel concepts from continuously shifting training\ndata. Typical solutions for this continual learning problem require extensive\nrehearsal of previously seen data, which increases memory costs and may violate\ndata privacy. Recently, the emergence of large-scale pre-trained vision\ntransformer models has enabled prompting approaches as an alternative to\ndata-rehearsal. These approaches rely on a key-query mechanism to generate\nprompts and have been found to be highly resistant to catastrophic forgetting\nin the well-established rehearsal-free continual learning setting. However, the\nkey mechanism of these methods is not trained end-to-end with the task\nsequence. Our experiments show that this leads to a reduction in their\nplasticity, hence sacrificing new task accuracy, and inability to benefit from\nexpanded parameter capacity. We instead propose to learn a set of prompt\ncomponents which are assembled with input-conditioned weights to produce\ninput-conditioned prompts, resulting in a novel attention-based end-to-end\nkey-query scheme. Our experiments show that we outperform the current SOTA\nmethod DualPrompt on established benchmarks by as much as 4.5% in average final\naccuracy. We also outperform the state of art by as much as 4.4% accuracy on a\ncontinual learning benchmark which contains both class-incremental and\ndomain-incremental task shifts, corresponding to many practical settings. Our\ncode is available at https://github.com/GT-RIPL/CODA-Prompt\n", "date": "2022-11-23", "categories": []}
{"id": "2210.04428", "title": "A Simple Baseline that Questions the Use of Pretrained-Models in\n  Continual Learning", "abstract": "  With the success of pretraining techniques in representation learning, a\nnumber of continual learning methods based on pretrained models have been\nproposed. Some of these methods design continual learning mechanisms on the\npre-trained representations and only allow minimum updates or even no updates\nof the backbone models during the training of continual learning. In this\npaper, we question whether the complexity of these models is needed to achieve\ngood performance by comparing them to a simple baseline that we designed. We\nargue that the pretrained feature extractor itself can be strong enough to\nachieve a competitive or even better continual learning performance on\nSplit-CIFAR100 and CoRe 50 benchmarks. To validate this, we conduct a very\nsimple baseline that 1) use the frozen pretrained model to extract image\nfeatures for every class encountered during the continual learning stage and\ncompute their corresponding mean features on training data, and 2) predict the\nclass of the input based on the nearest neighbor distance between test samples\nand mean features of the classes; i.e., Nearest Mean Classifier (NMC). This\nbaseline is single-headed, exemplar-free, and can be task-free (by updating the\nmeans continually). This baseline achieved 88.53% on 10-Split-CIFAR-100,\nsurpassing most state-of-the-art continual learning methods that are all\ninitialized using the same pretrained transformer model. We hope our baseline\nmay encourage future progress in designing learning systems that can\ncontinually add quality to the learning representations even if they started\nfrom some pretrained weights.\n", "date": "2022-10-10", "categories": []}
{"id": "2303.15494", "title": "Semantic-visual Guided Transformer for Few-shot Class-incremental\n  Learning", "abstract": "  Few-shot class-incremental learning (FSCIL) has recently attracted extensive\nattention in various areas. Existing FSCIL methods highly depend on the\nrobustness of the feature backbone pre-trained on base classes. In recent\nyears, different Transformer variants have obtained significant processes in\nthe feature representation learning of massive fields. Nevertheless, the\nprogress of the Transformer in FSCIL scenarios has not achieved the potential\npromised in other fields so far. In this paper, we develop a semantic-visual\nguided Transformer (SV-T) to enhance the feature extracting capacity of the\npre-trained feature backbone on incremental classes. Specifically, we first\nutilize the visual (image) labels provided by the base classes to supervise the\noptimization of the Transformer. And then, a text encoder is introduced to\nautomatically generate the corresponding semantic (text) labels for each image\nfrom the base classes. Finally, the constructed semantic labels are further\napplied to the Transformer for guiding its hyperparameters updating. Our SV-T\ncan take full advantage of more supervision information from base classes and\nfurther enhance the training robustness of the feature backbone. More\nimportantly, our SV-T is an independent method, which can directly apply to the\nexisting FSCIL architectures for acquiring embeddings of various incremental\nclasses. Extensive experiments on three benchmarks, two FSCIL architectures,\nand two Transformer variants show that our proposed SV-T obtains a significant\nimprovement in comparison to the existing state-of-the-art FSCIL methods.\n", "date": "2023-03-27", "categories": []}
{"id": "2303.15888", "title": "Projected Latent Distillation for Data-Agnostic Consolidation in\n  Distributed Continual Learning", "abstract": "  Distributed learning on the edge often comprises self-centered devices (SCD)\nwhich learn local tasks independently and are unwilling to contribute to the\nperformance of other SDCs. How do we achieve forward transfer at zero cost for\nthe single SCDs? We formalize this problem as a Distributed Continual Learning\nscenario, where SCD adapt to local tasks and a CL model consolidates the\nknowledge from the resulting stream of models without looking at the SCD's\nprivate data. Unfortunately, current CL methods are not directly applicable to\nthis scenario. We propose Data-Agnostic Consolidation (DAC), a novel double\nknowledge distillation method that consolidates the stream of SC models without\nusing the original data. DAC performs distillation in the latent space via a\nnovel Projected Latent Distillation loss. Experimental results show that DAC\nenables forward transfer between SCDs and reaches state-of-the-art accuracy on\nSplit CIFAR100, CORe50 and Split TinyImageNet, both in reharsal-free and\ndistributed CL scenarios. Somewhat surprisingly, even a single\nout-of-distribution image is sufficient as the only source of data during\nconsolidation.\n", "date": "2023-03-28", "categories": []}
{"id": "2303.11934", "title": "Sparse Distributed Memory is a Continual Learner", "abstract": "  Continual learning is a problem for artificial neural networks that their\nbiological counterparts are adept at solving. Building on work using Sparse\nDistributed Memory (SDM) to connect a core neural circuit with the powerful\nTransformer model, we create a modified Multi-Layered Perceptron (MLP) that is\na strong continual learner. We find that every component of our MLP variant\ntranslated from biology is necessary for continual learning. Our solution is\nalso free from any memory replay or task information, and introduces novel\nmethods to train sparse networks that may be broadly applicable.\n", "date": "2023-03-20", "categories": []}
{"id": "2303.14423", "title": "Task-Attentive Transformer Architecture for Continual Learning of\n  Vision-and-Language Tasks Using Knowledge Distillation", "abstract": "  The size and the computational load of fine-tuning large-scale pre-trained\nneural network are becoming two major obstacles in adopting machine learning in\nmany applications. Continual learning (CL) can serve as a remedy through\nenabling knowledge-transfer across sequentially arriving tasks which relaxes\nthe need to fine-tune all network weights from scratch. However, existing CL\nalgorithms primarily consider learning unimodal vision-only or language-only\ntasks. We develop a transformer-based CL architecture for learning bimodal\nvision-and-language tasks based on increasing the number of the learnable\nparameters dynamically and using knowledge distillation. The new additional\nparameters are used to specialize the network for each task. Our approach\nenables sharing information between the tasks while addressing the challenge of\ncatastrophic forgetting. Our approach is scalable learning to a large number of\ntasks because it requires little memory and time overhead. Our model reaches\nstate-of-the-art performance on challenging vision-and-language tasks.\n", "date": "2023-03-25", "categories": []}
{"id": "2303.14962", "title": "Forget-free Continual Learning with Soft-Winning SubNetworks", "abstract": "  Inspired by Regularized Lottery Ticket Hypothesis (RLTH), which states that\ncompetitive smooth (non-binary) subnetworks exist within a dense network in\ncontinual learning tasks, we investigate two proposed architecture-based\ncontinual learning methods which sequentially learn and select adaptive binary-\n(WSN) and non-binary Soft-Subnetworks (SoftNet) for each task. WSN and SoftNet\njointly learn the regularized model weights and task-adaptive non-binary masks\nof subnetworks associated with each task whilst attempting to select a small\nset of weights to be activated (winning ticket) by reusing weights of the prior\nsubnetworks. Our proposed WSN and SoftNet are inherently immune to catastrophic\nforgetting as each selected subnetwork model does not infringe upon other\nsubnetworks in Task Incremental Learning (TIL). In TIL, binary masks spawned\nper winning ticket are encoded into one N-bit binary digit mask, then\ncompressed using Huffman coding for a sub-linear increase in network capacity\nto the number of tasks. Surprisingly, in the inference step, SoftNet generated\nby injecting small noises to the backgrounds of acquired WSN (holding the\nforegrounds of WSN) provides excellent forward transfer power for future tasks\nin TIL. SoftNet shows its effectiveness over WSN in regularizing parameters to\ntackle the overfitting, to a few examples in Few-shot Class Incremental\nLearning (FSCIL).\n", "date": "2023-03-27", "categories": []}
{"id": "2303.15342", "title": "Exploring Continual Learning of Diffusion Models", "abstract": "  Diffusion models have achieved remarkable success in generating high-quality\nimages thanks to their novel training procedures applied to unprecedented\namounts of data. However, training a diffusion model from scratch is\ncomputationally expensive. This highlights the need to investigate the\npossibility of training these models iteratively, reusing computation while the\ndata distribution changes. In this study, we take the first step in this\ndirection and evaluate the continual learning (CL) properties of diffusion\nmodels. We begin by benchmarking the most common CL methods applied to\nDenoising Diffusion Probabilistic Models (DDPMs), where we note the strong\nperformance of the experience replay with the reduced rehearsal coefficient.\nFurthermore, we provide insights into the dynamics of forgetting, which exhibit\ndiverse behavior across diffusion timesteps. We also uncover certain pitfalls\nof using the bits-per-dimension metric for evaluating CL.\n", "date": "2023-03-27", "categories": []}
{"id": "2303.15378", "title": "CoDeC: Communication-Efficient Decentralized Continual Learning", "abstract": "  Training at the edge utilizes continuously evolving data generated at\ndifferent locations. Privacy concerns prohibit the co-location of this\nspatially as well as temporally distributed data, deeming it crucial to design\ntraining algorithms that enable efficient continual learning over decentralized\nprivate data. Decentralized learning allows serverless training with spatially\ndistributed data. A fundamental barrier in such distributed learning is the\nhigh bandwidth cost of communicating model updates between agents. Moreover,\nexisting works under this training paradigm are not inherently suitable for\nlearning a temporal sequence of tasks while retaining the previously acquired\nknowledge. In this work, we propose CoDeC, a novel communication-efficient\ndecentralized continual learning algorithm which addresses these challenges. We\nmitigate catastrophic forgetting while learning a task sequence in a\ndecentralized learning setup by combining orthogonal gradient projection with\ngossip averaging across decentralized agents. Further, CoDeC includes a novel\nlossless communication compression scheme based on the gradient subspaces. We\nexpress layer-wise gradients as a linear combination of the basis vectors of\nthese gradient subspaces and communicate the associated coefficients. We\ntheoretically analyze the convergence rate for our algorithm and demonstrate\nthrough an extensive set of experiments that CoDeC successfully learns\ndistributed continual tasks with minimal forgetting. The proposed compression\nscheme results in up to 4.8x reduction in communication costs with\niso-performance as the full communication baseline.\n", "date": "2023-03-27", "categories": []}
{"id": "2302.01047", "title": "Real-Time Evaluation in Online Continual Learning: A New Hope", "abstract": "  Current evaluations of Continual Learning (CL) methods typically assume that\nthere is no constraint on training time and computation. This is an unrealistic\nassumption for any real-world setting, which motivates us to propose: a\npractical real-time evaluation of continual learning, in which the stream does\nnot wait for the model to complete training before revealing the next data for\npredictions. To do this, we evaluate current CL methods with respect to their\ncomputational costs. We conduct extensive experiments on CLOC, a large-scale\ndataset containing 39 million time-stamped images with geolocation labels. We\nshow that a simple baseline outperforms state-of-the-art CL methods under this\nevaluation, questioning the applicability of existing methods in realistic\nsettings. In addition, we explore various CL components commonly used in the\nliterature, including memory sampling strategies and regularization approaches.\nWe find that all considered methods fail to be competitive against our simple\nbaseline. This surprisingly suggests that the majority of existing CL\nliterature is tailored to a specific class of streams that is not practical. We\nhope that the evaluation we provide will be the first step towards a paradigm\nshift to consider the computational cost in the development of online continual\nlearning methods.\n", "date": "2023-02-02", "categories": []}
{"id": "2303.13862", "title": "Two-level Graph Network for Few-Shot Class-Incremental Learning", "abstract": "  Few-shot class-incremental learning (FSCIL) aims to design machine learning\nalgorithms that can continually learn new concepts from a few data points,\nwithout forgetting knowledge of old classes. The difficulty lies in that\nlimited data from new classes not only lead to significant overfitting issues\nbut also exacerbates the notorious catastrophic forgetting problems. However,\nexisting FSCIL methods ignore the semantic relationships between sample-level\nand class-level. % Using the advantage that graph neural network (GNN) can mine\nrich information among few samples, In this paper, we designed a two-level\ngraph network for FSCIL named Sample-level and Class-level Graph Neural Network\n(SCGN). Specifically, a pseudo incremental learning paradigm is designed in\nSCGN, which synthesizes virtual few-shot tasks as new tasks to optimize SCGN\nmodel parameters in advance. Sample-level graph network uses the relationship\nof a few samples to aggregate similar samples and obtains refined class-level\nfeatures. Class-level graph network aims to mitigate the semantic conflict\nbetween prototype features of new classes and old classes. SCGN builds\ntwo-level graph networks to guarantee the latent semantic of each few-shot\nclass can be effectively represented in FSCIL. Experiments on three popular\nbenchmark datasets show that our method significantly outperforms the baselines\nand sets new state-of-the-art results with remarkable advantages.\n", "date": "2023-03-24", "categories": []}
{"id": "2303.13898", "title": "Remind of the Past: Incremental Learning with Analogical Prompts", "abstract": "  Although data-free incremental learning methods are memory-friendly,\naccurately estimating and counteracting representation shifts is challenging in\nthe absence of historical data. This paper addresses this thorny problem by\nproposing a novel incremental learning method inspired by human analogy\ncapabilities. Specifically, we design an analogy-making mechanism to remap the\nnew data into the old class by prompt tuning. It mimics the feature\ndistribution of the target old class on the old model using only samples of new\nclasses. The learnt prompts are further used to estimate and counteract the\nrepresentation shift caused by fine-tuning for the historical prototypes. The\nproposed method sets up new state-of-the-art performance on four incremental\nlearning benchmarks under both the class and domain incremental learning\nsettings. It consistently outperforms data-replay methods by only saving\nfeature prototypes for each class. It has almost hit the empirical upper bound\nby joint training on the Core50 benchmark. The code will be released at\n\\url{https://github.com/ZhihengCV/A-Prompts}.\n", "date": "2023-03-24", "categories": []}
{"id": "2303.13157", "title": "Adiabatic replay for continual learning", "abstract": "  Conventional replay-based approaches to continual learning (CL) require, for\neach learning phase with new data, the replay of samples representing all of\nthe previously learned knowledge in order to avoid catastrophic forgetting.\nSince the amount of learned knowledge grows over time in CL problems,\ngenerative replay spends an increasing amount of time just re-learning what is\nalready known. In this proof-of-concept study, we propose a replay-based CL\nstrategy that we term adiabatic replay (AR), which derives its efficiency from\nthe (reasonable) assumption that each new learning phase is adiabatic, i.e.,\nrepresents only a small addition to existing knowledge. Each new learning phase\ntriggers a sampling process that selectively replays, from the body of existing\nknowledge, just such samples that are similar to the new data, in contrast to\nreplaying all of it. Complete replay is not required since AR represents the\ndata distribution by GMMs, which are capable of selectively updating their\ninternal representation only where data statistics have changed. As long as\nadditions are adiabatic, the amount of to-be-replayed samples need not to\ndepend on the amount of previously acquired knowledge at all. We verify\nexperimentally that AR is superior to state-of-the-art deep generative replay\nusing VAEs.\n", "date": "2023-03-23", "categories": []}
{"id": "2303.12696", "title": "Dense Network Expansion for Class Incremental Learning", "abstract": "  The problem of class incremental learning (CIL) is considered.\nState-of-the-art approaches use a dynamic architecture based on network\nexpansion (NE), in which a task expert is added per task. While effective from\na computational standpoint, these methods lead to models that grow quickly with\nthe number of tasks. A new NE method, dense network expansion (DNE), is\nproposed to achieve a better trade-off between accuracy and model complexity.\nThis is accomplished by the introduction of dense connections between the\nintermediate layers of the task expert networks, that enable the transfer of\nknowledge from old to new tasks via feature sharing and reusing. This sharing\nis implemented with a cross-task attention mechanism, based on a new task\nattention block (TAB), that fuses information across tasks. Unlike traditional\nattention mechanisms, TAB operates at the level of the feature mixing and is\ndecoupled with spatial attentions. This is shown more effective than a joint\nspatial-and-task attention for CIL. The proposed DNE approach can strictly\nmaintain the feature space of old classes while growing the network and feature\nscale at a much slower rate than previous methods. In result, it outperforms\nthe previous SOTA methods by a margin of 4\\% in terms of accuracy, with similar\nor even smaller model scale.\n", "date": "2023-03-22", "categories": []}
{"id": "2303.11624", "title": "Assessor-Guided Learning for Continual Environments", "abstract": "  This paper proposes an assessor-guided learning strategy for continual\nlearning where an assessor guides the learning process of a base learner by\ncontrolling the direction and pace of the learning process thus allowing an\nefficient learning of new environments while protecting against the\ncatastrophic interference problem. The assessor is trained in a meta-learning\nmanner with a meta-objective to boost the learning process of the base learner.\nIt performs a soft-weighting mechanism of every sample accepting positive\nsamples while rejecting negative samples. The training objective of a base\nlearner is to minimize a meta-weighted combination of the cross entropy loss\nfunction, the dark experience replay (DER) loss function and the knowledge\ndistillation loss function whose interactions are controlled in such a way to\nattain an improved performance. A compensated over-sampling (COS) strategy is\ndeveloped to overcome the class imbalanced problem of the episodic memory due\nto limited memory budgets. Our approach, Assessor-Guided Learning Approach\n(AGLA), has been evaluated in the class-incremental and task-incremental\nlearning problems. AGLA achieves improved performances compared to its\ncompetitors while the theoretical analysis of the COS strategy is offered.\nSource codes of AGLA, baseline algorithms and experimental logs are shared\npublicly in \\url{https://github.com/anwarmaxsum/AGLA} for further study.\n", "date": "2023-03-21", "categories": []}
{"id": "2303.11863", "title": "Continual Learning in the Presence of Spurious Correlation", "abstract": "  Most continual learning (CL) algorithms have focused on tackling the\nstability-plasticity dilemma, that is, the challenge of preventing the\nforgetting of previous tasks while learning new ones. However, they have\noverlooked the impact of the knowledge transfer when the dataset in a certain\ntask is biased - namely, when some unintended spurious correlations of the\ntasks are learned from the biased dataset. In that case, how would they affect\nlearning future tasks or the knowledge already learned from the past tasks? In\nthis work, we carefully design systematic experiments using one synthetic and\ntwo real-world datasets to answer the question from our empirical findings.\nSpecifically, we first show through two-task CL experiments that standard CL\nmethods, which are unaware of dataset bias, can transfer biases from one task\nto another, both forward and backward, and this transfer is exacerbated\ndepending on whether the CL methods focus on the stability or the plasticity.\nWe then present that the bias transfer also exists and even accumulate in\nlonger sequences of tasks. Finally, we propose a simple, yet strong plug-in\nmethod for debiasing-aware continual learning, dubbed as Group-class Balanced\nGreedy Sampling (BGS). As a result, we show that our BGS can always reduce the\nbias of a CL model, with a slight loss of CL performance at most.\n", "date": "2023-03-21", "categories": []}
{"id": "2303.11952", "title": "Hierarchical Memory Pool Based Edge Semi-Supervised Continual Learning\n  Method", "abstract": "  The continuous changes in the world have resulted in the performance\nregression of neural networks. Therefore, continual learning (CL) area\ngradually attracts the attention of more researchers. For edge intelligence,\nthe CL model not only needs to overcome catastrophic for-getting, but also\nneeds to face the huge challenge of severely limited resources: the lack of\nlabeled resources and powerful devices. However, the existing classic CL\nmethods usually rely on a large number of labeled samples to maintain the\nplasticity and stability, and the semi-supervised learning methods often need\nto pay a large computational and memory overhead for higher accuracy. In\nresponse to these prob-lems, a low-cost semi-supervised CL method named Edge\nHierarchical Memory Learner (EdgeHML) will be proposed. EdgeHML can\neffec-tively utilize a large number of unlabeled samples and a small number of\nlabeled samples. It is based on a hierarchical memory pool, lever-age\nmulti-level storage structure to store and replay samples. EdgeHML implements\nthe interaction between different levels through a combination of online and\noffline strategies. In addition, in order to further reduce the computational\noverhead for unlabeled samples, EdgeHML leverages a progressive learning\nmethod. It reduces the computation cycles of unlabeled samples by controlling\nthe learning process. The experimental results show that on three\nsemi-supervised CL tasks, EdgeHML can improve the model accuracy by up to\n16.35% compared with the classic CL method, and the training iterations time\ncan be reduced by more than 50% compared with semi-supervised methods. EdgeHML\nachieves a semi-supervised CL process with high performance and low overhead\nfor edge intelligence.\n", "date": "2023-01-17", "categories": []}
{"id": "2207.12819", "title": "S-Prompts Learning with Pre-trained Transformers: An Occam's Razor for\n  Domain Incremental Learning", "abstract": "  State-of-the-art deep neural networks are still struggling to address the\ncatastrophic forgetting problem in continual learning. In this paper, we\npropose one simple paradigm (named as S-Prompting) and two concrete approaches\nto highly reduce the forgetting degree in one of the most typical continual\nlearning scenarios, i.e., domain increment learning (DIL). The key idea of the\nparadigm is to learn prompts independently across domains with pre-trained\ntransformers, avoiding the use of exemplars that commonly appear in\nconventional methods. This results in a win-win game where the prompting can\nachieve the best for each domain. The independent prompting across domains only\nrequests one single cross-entropy loss for training and one simple K-NN\noperation as a domain identifier for inference. The learning paradigm derives\nan image prompt learning approach and a novel language-image prompt learning\napproach. Owning an excellent scalability (0.03% parameter increase per\ndomain), the best of our approaches achieves a remarkable relative improvement\n(an average of about 30%) over the best of the state-of-the-art exemplar-free\nmethods for three standard DIL tasks, and even surpasses the best of them\nrelatively by about 6% in average when they use exemplars. Source code is\navailable at \\url{https://github.com/iamwangyabin/S-Prompts}.\n", "date": "2022-07-26", "categories": []}
{"id": "2303.11076", "title": "From MNIST to ImageNet and Back: Benchmarking Continual Curriculum\n  Learning", "abstract": "  Continual learning (CL) is one of the most promising trends in recent machine\nlearning research. Its goal is to go beyond classical assumptions in machine\nlearning and develop models and learning strategies that present high\nrobustness in dynamic environments. The landscape of CL research is fragmented\ninto several learning evaluation protocols, comprising different learning\ntasks, datasets, and evaluation metrics. Additionally, the benchmarks adopted\nso far are still distant from the complexity of real-world scenarios, and are\nusually tailored to highlight capabilities specific to certain strategies. In\nsuch a landscape, it is hard to objectively assess strategies. In this work, we\nfill this gap for CL on image data by introducing two novel CL benchmarks that\ninvolve multiple heterogeneous tasks from six image datasets, with varying\nlevels of complexity and quality. Our aim is to fairly evaluate current\nstate-of-the-art CL strategies on a common ground that is closer to complex\nreal-world scenarios. We additionally structure our benchmarks so that tasks\nare presented in increasing and decreasing order of complexity -- according to\na curriculum -- in order to evaluate if current CL models are able to exploit\nstructure across tasks. We devote particular emphasis to providing the CL\ncommunity with a rigorous and reproducible evaluation protocol for measuring\nthe ability of a model to generalize and not to forget while learning.\nFurthermore, we provide an extensive experimental evaluation showing that\npopular CL strategies, when challenged with our benchmarks, yield sub-par\nperformance, high levels of forgetting, and present a limited ability to\neffectively leverage curriculum task ordering. We believe that these results\nhighlight the need for rigorous comparisons in future CL works as well as pave\nthe way to design new CL strategies that are able to deal with more complex\nscenarios.\n", "date": "2023-03-16", "categories": []}
{"id": "2211.05231", "title": "Biologically-Inspired Continual Learning of Human Motion Sequences", "abstract": "  This work proposes a model for continual learning on tasks involving temporal\nsequences, specifically, human motions. It improves on a recently proposed\nbrain-inspired replay model (BI-R) by building a biologically-inspired\nconditional temporal variational autoencoder (BI-CTVAE), which instantiates a\nlatent mixture-of-Gaussians for class representation. We investigate a novel\ncontinual-learning-to-generate (CL2Gen) scenario where the model generates\nmotion sequences of different classes. The generative accuracy of the model is\ntested over a set of tasks. The final classification accuracy of BI-CTVAE on a\nhuman motion dataset after sequentially learning all action classes is 78%,\nwhich is 63% higher than using no-replay, and only 5.4% lower than a\nstate-of-the-art offline trained GRU model.\n", "date": "2022-11-02", "categories": []}
{"id": "2303.08207", "title": "Is forgetting less a good inductive bias for forward transfer?", "abstract": "  One of the main motivations of studying continual learning is that the\nproblem setting allows a model to accrue knowledge from past tasks to learn new\ntasks more efficiently. However, recent studies suggest that the key metric\nthat continual learning algorithms optimize, reduction in catastrophic\nforgetting, does not correlate well with the forward transfer of knowledge. We\nbelieve that the conclusion previous works reached is due to the way they\nmeasure forward transfer. We argue that the measure of forward transfer to a\ntask should not be affected by the restrictions placed on the continual learner\nin order to preserve knowledge of previous tasks. Instead, forward transfer\nshould be measured by how easy it is to learn a new task given a set of\nrepresentations produced by continual learning on previous tasks. Under this\nnotion of forward transfer, we evaluate different continual learning algorithms\non a variety of image classification benchmarks. Our results indicate that less\nforgetful representations lead to a better forward transfer suggesting a strong\ncorrelation between retaining past information and learning efficiency on new\ntasks. Further, we found less forgetful representations to be more diverse and\ndiscriminative compared to their forgetful counterparts.\n", "date": "2023-03-14", "categories": []}
{"id": "2303.07477", "title": "Efficient Self-supervised Continual Learning with Progressive\n  Task-correlated Layer Freezing", "abstract": "  Inspired by the success of Self-supervised learning (SSL) in learning visual\nrepresentations from unlabeled data, a few recent works have studied SSL in the\ncontext of continual learning (CL), where multiple tasks are learned\nsequentially, giving rise to a new paradigm, namely self-supervised continual\nlearning (SSCL). It has been shown that the SSCL outperforms supervised\ncontinual learning (SCL) as the learned representations are more informative\nand robust to catastrophic forgetting. However, if not designed intelligently,\nthe training complexity of SSCL may be prohibitively high due to the inherent\ntraining cost of SSL. In this work, by investigating the task correlations in\nSSCL setup first, we discover an interesting phenomenon that, with the\nSSL-learned background model, the intermediate features are highly correlated\nbetween tasks. Based on this new finding, we propose a new SSCL method with\nlayer-wise freezing which progressively freezes partial layers with the highest\ncorrelation ratios for each task to improve training computation efficiency and\nmemory efficiency. Extensive experiments across multiple datasets are\nperformed, where our proposed method shows superior performance against the\nSoTA SSCL methods under various SSL frameworks. For example, compared to LUMP,\nour method achieves 12\\%/14\\%/12\\% GPU training time reduction, 23\\%/26\\%/24\\%\nmemory reduction, 35\\%/34\\%/33\\% backward FLOPs reduction, and\n1.31\\%/1.98\\%/1.21\\% forgetting reduction without accuracy degradation on three\ndatasets, respectively.\n", "date": "2023-03-13", "categories": []}
{"id": "2303.07507", "title": "Loss of Plasticity in Continual Deep Reinforcement Learning", "abstract": "  The ability to learn continually is essential in a complex and changing\nworld. In this paper, we characterize the behavior of canonical value-based\ndeep reinforcement learning (RL) approaches under varying degrees of\nnon-stationarity. In particular, we demonstrate that deep RL agents lose their\nability to learn good policies when they cycle through a sequence of Atari 2600\ngames. This phenomenon is alluded to in prior work under various guises --\ne.g., loss of plasticity, implicit under-parameterization, primacy bias, and\ncapacity loss. We investigate this phenomenon closely at scale and analyze how\nthe weights, gradients, and activations change over time in several experiments\nwith varying dimensions (e.g., similarity between games, number of games,\nnumber of frames per game), with some experiments spanning 50 days and 2\nbillion environment interactions. Our analysis shows that the activation\nfootprint of the network becomes sparser, contributing to the diminishing\ngradients. We investigate a remarkably simple mitigation strategy --\nConcatenated ReLUs (CReLUs) activation function -- and demonstrate its\neffectiveness in facilitating continual learning in a changing environment.\n", "date": "2023-03-13", "categories": []}
{"id": "2303.07864", "title": "DualMix: Unleashing the Potential of Data Augmentation for Online\n  Class-Incremental Learning", "abstract": "  Online Class-Incremental (OCI) learning has sparked new approaches to expand\nthe previously trained model knowledge from sequentially arriving data streams\nwith new classes. Unfortunately, OCI learning can suffer from catastrophic\nforgetting (CF) as the decision boundaries for old classes can become\ninaccurate when perturbated by new ones. Existing literature have applied the\ndata augmentation (DA) to alleviate the model forgetting, while the role of DA\nin OCI has not been well understood so far. In this paper, we theoretically\nshow that augmented samples with lower correlation to the original data are\nmore effective in preventing forgetting. However, aggressive augmentation may\nalso reduce the consistency between data and corresponding labels, which\nmotivates us to exploit proper DA to boost the OCI performance and prevent the\nCF problem. We propose the Enhanced Mixup (EnMix) method that mixes the\naugmented samples and their labels simultaneously, which is shown to enhance\nthe sample diversity while maintaining strong consistency with corresponding\nlabels. Further, to solve the class imbalance problem, we design an Adaptive\nMixup (AdpMix) method to calibrate the decision boundaries by mixing samples\nfrom both old and new classes and dynamically adjusting the label mixing ratio.\nOur approach is demonstrated to be effective on several benchmark datasets\nthrough extensive experiments, and it is shown to be compatible with other\nreplay-based techniques.\n", "date": "2023-03-14", "categories": []}
{"id": "2301.10418", "title": "DEJA VU: Continual Model Generalization For Unseen Domains", "abstract": "  In real-world applications, deep learning models often run in non-stationary\nenvironments where the target data distribution continually shifts over time.\nThere have been numerous domain adaptation (DA) methods in both online and\noffline modes to improve cross-domain adaptation ability. However, these DA\nmethods typically only provide good performance after a long period of\nadaptation, and perform poorly on new domains before and during adaptation - in\nwhat we call the \"Unfamiliar Period\", especially when domain shifts happen\nsuddenly and significantly. On the other hand, domain generalization (DG)\nmethods have been proposed to improve the model generalization ability on\nunadapted domains. However, existing DG works are ineffective for continually\nchanging domains due to severe catastrophic forgetting of learned knowledge. To\novercome these limitations of DA and DG in handling the Unfamiliar Period\nduring continual domain shift, we propose RaTP, a framework that focuses on\nimproving models' target domain generalization (TDG) capability, while also\nachieving effective target domain adaptation (TDA) capability right after\ntraining on certain domains and forgetting alleviation (FA) capability on past\ndomains. RaTP includes a training-free data augmentation module to prepare data\nfor TDG, a novel pseudo-labeling mechanism to provide reliable supervision for\nTDA, and a prototype contrastive alignment algorithm to align different domains\nfor achieving TDG, TDA and FA. Extensive experiments on Digits, PACS, and\nDomainNet demonstrate that RaTP significantly outperforms state-of-the-art\nworks from Continual DA, Source-Free DA, Test-Time/Online DA, Single DG,\nMultiple DG and Unified DA&DG in TDG, and achieves comparable TDA and FA\ncapabilities.\n", "date": "2023-01-25", "categories": []}
{"id": "2303.04183", "title": "Robustness-preserving Lifelong Learning via Dataset Condensation", "abstract": "  Lifelong learning (LL) aims to improve a predictive model as the data source\nevolves continuously. Most work in this learning paradigm has focused on\nresolving the problem of 'catastrophic forgetting,' which refers to a notorious\ndilemma between improving model accuracy over new data and retaining accuracy\nover previous data. Yet, it is also known that machine learning (ML) models can\nbe vulnerable in the sense that tiny, adversarial input perturbations can\ndeceive the models into producing erroneous predictions. This motivates the\nresearch objective of this paper - specification of a new LL framework that can\nsalvage model robustness (against adversarial attacks) from catastrophic\nforgetting. Specifically, we propose a new memory-replay LL strategy that\nleverages modern bi-level optimization techniques to determine the 'coreset' of\nthe current data (i.e., a small amount of data to be memorized) for ease of\npreserving adversarial robustness over time. We term the resulting LL framework\n'Data-Efficient Robustness-Preserving LL' (DERPLL). The effectiveness of DERPLL\nis evaluated for class-incremental image classification using ResNet-18 over\nthe CIFAR-10 dataset. Experimental results show that DERPLL outperforms the\nconventional coreset-guided LL baseline and achieves a substantial improvement\nin both standard accuracy and robust accuracy.\n", "date": "2023-03-07", "categories": []}
{"id": "2303.02954", "title": "Centroid Distance Distillation for Effective Rehearsal in Continual\n  Learning", "abstract": "  Rehearsal, retraining on a stored small data subset of old tasks, has been\nproven effective in solving catastrophic forgetting in continual learning.\nHowever, due to the sampled data may have a large bias towards the original\ndataset, retraining them is susceptible to driving continual domain drift of\nold tasks in feature space, resulting in forgetting. In this paper, we focus on\ntackling the continual domain drift problem with centroid distance\ndistillation. First, we propose a centroid caching mechanism for sampling data\npoints based on constructed centroids to reduce the sample bias in rehearsal.\nThen, we present a centroid distance distillation that only stores the centroid\ndistance to reduce the continual domain drift. The experiments on four\ncontinual learning datasets show the superiority of the proposed method, and\nthe continual domain drift can be reduced.\n", "date": "2023-03-06", "categories": []}
{"id": "2211.10445", "title": "Building a Subspace of Policies for Scalable Continual Learning", "abstract": "  The ability to continuously acquire new knowledge and skills is crucial for\nautonomous agents. Existing methods are typically based on either fixed-size\nmodels that struggle to learn a large number of diverse behaviors, or\ngrowing-size models that scale poorly with the number of tasks. In this work,\nwe aim to strike a better balance between an agent's size and performance by\ndesigning a method that grows adaptively depending on the task sequence. We\nintroduce Continual Subspace of Policies (CSP), a new approach that\nincrementally builds a subspace of policies for training a reinforcement\nlearning agent on a sequence of tasks. The subspace's high expressivity allows\nCSP to perform well for many different tasks while growing sublinearly with the\nnumber of tasks. Our method does not suffer from forgetting and displays\npositive transfer to new tasks. CSP outperforms a number of popular baselines\non a wide range of scenarios from two challenging domains, Brax (locomotion)\nand Continual World (manipulation).\n", "date": "2022-11-18", "categories": []}
{"id": "2303.01081", "title": "Can BERT Refrain from Forgetting on Sequential Tasks? A Probing Study", "abstract": "  Large pre-trained language models help to achieve state of the art on a\nvariety of natural language processing (NLP) tasks, nevertheless, they still\nsuffer from forgetting when incrementally learning a sequence of tasks. To\nalleviate this problem, recent works enhance existing models by sparse\nexperience replay and local adaption, which yield satisfactory performance.\nHowever, in this paper we find that pre-trained language models like BERT have\na potential ability to learn sequentially, even without any sparse memory\nreplay. To verify the ability of BERT to maintain old knowledge, we adopt and\nre-finetune single-layer probe networks with the parameters of BERT fixed. We\ninvestigate the models on two types of NLP tasks, text classification and\nextractive question answering. Our experiments reveal that BERT can actually\ngenerate high quality representations for previously learned tasks in a long\nterm, under extremely sparse replay or even no replay. We further introduce a\nseries of novel methods to interpret the mechanism of forgetting and how memory\nrehearsal plays a significant role in task incremental learning, which bridges\nthe gap between our new discovery and previous studies about catastrophic\nforgetting.\n", "date": "2023-03-02", "categories": []}
{"id": "2303.01421", "title": "Semiparametric Language Models Are Scalable Continual Learners", "abstract": "  Semiparametric language models (LMs) have shown promise in continuously\nlearning from new text data by combining a parameterized neural LM with a\ngrowable non-parametric memory for memorizing new content. However,\nconventional semiparametric LMs will finally become prohibitive for computing\nand storing if they are applied to continual learning over streaming data,\nbecause the non-parametric memory grows linearly with the amount of data they\nlearn from over time. To address the issue of scalability, we present a simple\nand intuitive approach called Selective Memorization (SeMem), which only\nmemorizes difficult samples that the model is likely to struggle with. We\ndemonstrate that SeMem improves the scalability of semiparametric LMs for\ncontinual learning over streaming data in two ways: (1) data-wise scalability:\nas the model becomes stronger through continual learning, it will encounter\nfewer difficult cases that need to be memorized, causing the growth of the\nnon-parametric memory to slow down over time rather than growing at a linear\nrate with the size of training data; (2) model-wise scalability: SeMem allows a\nlarger model to memorize fewer samples than its smaller counterpart because it\nis rarer for a larger model to encounter incomprehensible cases, resulting in a\nnon-parametric memory that does not scale linearly with model size. We conduct\nextensive experiments in language modeling and downstream tasks to test SeMem's\nresults, showing SeMem enables a semiparametric LM to be a scalable continual\nlearner with little forgetting.\n", "date": "2023-03-02", "categories": []}
{"id": "2209.07529", "title": "On the Soft-Subnetwork for Few-shot Class Incremental Learning", "abstract": "  Inspired by Regularized Lottery Ticket Hypothesis (RLTH), which hypothesizes\nthat there exist smooth (non-binary) subnetworks within a dense network that\nachieve the competitive performance of the dense network, we propose a few-shot\nclass incremental learning (FSCIL) method referred to as \\emph{Soft-SubNetworks\n(SoftNet)}. Our objective is to learn a sequence of sessions incrementally,\nwhere each session only includes a few training instances per class while\npreserving the knowledge of the previously learned ones. SoftNet jointly learns\nthe model weights and adaptive non-binary soft masks at a base training session\nin which each mask consists of the major and minor subnetwork; the former aims\nto minimize catastrophic forgetting during training, and the latter aims to\navoid overfitting to a few samples in each new training session. We provide\ncomprehensive empirical validations demonstrating that our SoftNet effectively\ntackles the few-shot incremental learning problem by surpassing the performance\nof state-of-the-art baselines over benchmark datasets.\n", "date": "2022-09-15", "categories": []}
{"id": "2302.14712", "title": "Generating Accurate Virtual Examples For Lifelong Machine Learning", "abstract": "  Lifelong machine learning (LML) is an area of machine learning research\nconcerned with human-like persistent and cumulative nature of learning. LML\nsystem's objective is consolidating new information into an existing machine\nlearning model without catastrophically disrupting the prior information. Our\nresearch addresses this LML retention problem for creating a knowledge\nconsolidation network through task rehearsal without retaining the prior task's\ntraining examples. We discovered that the training data reconstruction error\nfrom a trained Restricted Boltzmann Machine can be successfully used to\ngenerate accurate virtual examples from the reconstructed set of a uniform\nrandom set of examples given to the trained model. We also defined a measure\nfor comparing the probability distributions of two datasets given to a trained\nnetwork model based on their reconstruction mean square errors.\n", "date": "2023-02-28", "categories": []}
{"id": "2302.13289", "title": "Improving Representational Continuity via Continued Pretraining", "abstract": "  We consider the continual representation learning setting: sequentially\npretrain a model $M'$ on tasks $T_1, \\ldots, T_T$, and then adapt $M'$ on a\nsmall amount of data from each task $T_i$ to check if it has forgotten\ninformation from old tasks. Under a kNN adaptation protocol, prior work shows\nthat continual learning methods improve forgetting over naive training (SGD).\nIn reality, practitioners do not use kNN classifiers -- they use the adaptation\nmethod that works best (e.g., fine-tuning) -- here, we find that strong\ncontinual learning baselines do worse than naive training. Interestingly, we\nfind that a method from the transfer learning community (LP-FT) outperforms\nnaive training and the other continual learning methods. Even with standard kNN\nevaluation protocols, LP-FT performs comparably with strong continual learning\nmethods (while being simpler and requiring less memory) on three standard\nbenchmarks: sequential CIFAR-10, CIFAR-100, and TinyImageNet. LP-FT also\nreduces forgetting in a real world satellite remote sensing dataset (FMoW), and\na variant of LP-FT gets state-of-the-art accuracies on an NLP continual\nlearning benchmark.\n", "date": "2023-02-26", "categories": []}
{"id": "2302.11730", "title": "Detachedly Learn a Classifier for Class-Incremental Learning", "abstract": "  In continual learning, model needs to continually learn a feature extractor\nand classifier on a sequence of tasks. This paper focuses on how to learn a\nclassifier based on a pretrained feature extractor under continual learning\nsetting. We present an probabilistic analysis that the failure of vanilla\nexperience replay (ER) comes from unnecessary re-learning of previous tasks and\nincompetence to distinguish current task from the previous ones, which is the\ncause of knowledge degradation and prediction bias. To overcome these\nweaknesses, we propose a novel replay strategy task-aware experience replay. It\nrebalances the replay loss and detaches classifier weight for the old tasks\nfrom the update process, by which the previous knowledge is kept intact and the\noverfitting on episodic memory is alleviated. Experimental results show our\nmethod outperforms current state-of-the-art methods.\n", "date": "2023-02-23", "categories": []}
{"id": "2302.11074", "title": "Preventing Catastrophic Forgetting in Continual Learning of New Natural\n  Language Tasks", "abstract": "  Multi-Task Learning (MTL) is widely-accepted in Natural Language Processing\nas a standard technique for learning multiple related tasks in one model.\nTraining an MTL model requires having the training data for all tasks available\nat the same time. As systems usually evolve over time, (e.g., to support new\nfunctionalities), adding a new task to an existing MTL model usually requires\nretraining the model from scratch on all the tasks and this can be\ntime-consuming and computationally expensive. Moreover, in some scenarios, the\ndata used to train the original training may be no longer available, for\nexample, due to storage or privacy concerns. In this paper, we approach the\nproblem of incrementally expanding MTL models' capability to solve new tasks\nover time by distilling the knowledge of an already trained model on n tasks\ninto a new one for solving n+1 tasks. To avoid catastrophic forgetting, we\npropose to exploit unlabeled data from the same distributions of the old tasks.\nOur experiments on publicly available benchmarks show that such a technique\ndramatically benefits the distillation by preserving the already acquired\nknowledge (i.e., preventing up to 20% performance drops on old tasks) while\nobtaining good performance on the incrementally added tasks. Further, we also\nshow that our approach is beneficial in practical settings by using data from a\nleading voice assistant.\n", "date": "2023-02-22", "categories": []}
{"id": "2302.11344", "title": "Error Sensitivity Modulation based Experience Replay: Mitigating Abrupt\n  Representation Drift in Continual Learning", "abstract": "  Humans excel at lifelong learning, as the brain has evolved to be robust to\ndistribution shifts and noise in our ever-changing environment. Deep neural\nnetworks (DNNs), however, exhibit catastrophic forgetting and the learned\nrepresentations drift drastically as they encounter a new task. This alludes to\na different error-based learning mechanism in the brain. Unlike DNNs, where\nlearning scales linearly with the magnitude of the error, the sensitivity to\nerrors in the brain decreases as a function of their magnitude. To this end, we\npropose \\textit{ESMER} which employs a principled mechanism to modulate error\nsensitivity in a dual-memory rehearsal-based system. Concretely, it maintains a\nmemory of past errors and uses it to modify the learning dynamics so that the\nmodel learns more from small consistent errors compared to large sudden errors.\nWe also propose \\textit{Error-Sensitive Reservoir Sampling} to maintain\nepisodic memory, which leverages the error history to pre-select low-loss\nsamples as candidates for the buffer, which are better suited for retaining\ninformation. Empirical results show that ESMER effectively reduces forgetting\nand abrupt drift in representations at the task boundary by gradually adapting\nto the new task while consolidating knowledge. Remarkably, it also enables the\nmodel to learn under high levels of label noise, which is ubiquitous in\nreal-world data streams.\n", "date": "2023-02-14", "categories": []}
{"id": "2302.11346", "title": "Task-Aware Information Routing from Common Representation Space in\n  Lifelong Learning", "abstract": "  Intelligent systems deployed in the real world suffer from catastrophic\nforgetting when exposed to a sequence of tasks. Humans, on the other hand,\nacquire, consolidate, and transfer knowledge between tasks that rarely\ninterfere with the consolidated knowledge. Accompanied by self-regulated\nneurogenesis, continual learning in the brain is governed by a rich set of\nneurophysiological processes that harbor different types of knowledge, which\nare then integrated by conscious processing. Thus, inspired by the Global\nWorkspace Theory of conscious information access in the brain, we propose\nTAMiL, a continual learning method that entails task-attention modules to\ncapture task-specific information from the common representation space. We\nemploy simple, undercomplete autoencoders to create a communication bottleneck\nbetween the common representation space and the global workspace, allowing only\nthe task-relevant information to the global workspace, thus greatly reducing\ntask interference. Experimental results show that our method outperforms\nstate-of-the-art rehearsal-based and dynamic sparse approaches and bridges the\ngap between fixed capacity and parameter isolation approaches while being\nscalable. We also show that our method effectively mitigates catastrophic\nforgetting while being well-calibrated with reduced task-recency bias.\n", "date": "2023-02-14", "categories": []}
{"id": "2302.10718", "title": "Effects of Architectures on Continual Semantic Segmentation", "abstract": "  Research in the field of Continual Semantic Segmentation is mainly\ninvestigating novel learning algorithms to overcome catastrophic forgetting of\nneural networks. Most recent publications have focused on improving learning\nalgorithms without distinguishing effects caused by the choice of neural\narchitecture.Therefore, we study how the choice of neural network architecture\naffects catastrophic forgetting in class- and domain-incremental semantic\nsegmentation. Specifically, we compare the well-researched CNNs to recently\nproposed Transformers and Hybrid architectures, as well as the impact of the\nchoice of novel normalization layers and different decoder heads. We find that\ntraditional CNNs like ResNet have high plasticity but low stability, while\ntransformer architectures are much more stable. When the inductive biases of\nCNN architectures are combined with transformers in hybrid architectures, it\nleads to higher plasticity and stability. The stability of these models can be\nexplained by their ability to learn general features that are robust against\ndistribution shifts. Experiments with different normalization layers show that\nContinual Normalization achieves the best trade-off in terms of adaptability\nand stability of the model. In the class-incremental setting, the choice of the\nnormalization layer has much less impact. Our experiments suggest that the\nright choice of architecture can significantly reduce forgetting even with\nnaive fine-tuning and confirm that for real-world applications, the\narchitecture is an important factor in designing a continual learning model.\n", "date": "2023-02-21", "categories": []}
{"id": "2302.09310", "title": "On Handling Catastrophic Forgetting for Incremental Learning of Human\n  Physical Activity on the Edge", "abstract": "  Human activity recognition (HAR) has been a classic research problem. In\nparticular, with recent machine learning (ML) techniques, the recognition task\nhas been largely investigated by companies and integrated into their products\nfor customers. However, most of them apply a predefined activity set and\nconduct the learning process on the cloud, hindering specific personalizations\nfrom end users (i.e., edge devices). Even though recent progress in Incremental\nLearning allows learning new-class data on the fly, the learning process is\ngenerally conducted on the cloud, requiring constant data exchange between\ncloud and edge devices, thus leading to data privacy issues. In this paper, we\npropose PILOTE, which pushes the incremental learning process to the extreme\nedge, while providing reliable data privacy and practical utility, e.g., low\nprocessing latency, personalization, etc. In particular, we consider the\npractical challenge of extremely limited data during the incremental learning\nprocess on edge, where catastrophic forgetting is required to be handled in a\npractical way. We validate PILOTE with extensive experiments on human activity\ndata collected from mobile sensors. The results show PILOTE can work on edge\ndevices with extremely limited resources while providing reliable performance.\n", "date": "2023-02-18", "categories": []}
{"id": "2302.09886", "title": "InOR-Net: Incremental 3D Object Recognition Network for Point Cloud\n  Representation", "abstract": "  3D object recognition has successfully become an appealing research topic in\nthe real-world. However, most existing recognition models unreasonably assume\nthat the categories of 3D objects cannot change over time in the real-world.\nThis unrealistic assumption may result in significant performance degradation\nfor them to learn new classes of 3D objects consecutively, due to the\ncatastrophic forgetting on old learned classes. Moreover, they cannot explore\nwhich 3D geometric characteristics are essential to alleviate the catastrophic\nforgetting on old classes of 3D objects. To tackle the above challenges, we\ndevelop a novel Incremental 3D Object Recognition Network (i.e., InOR-Net),\nwhich could recognize new classes of 3D objects continuously via overcoming the\ncatastrophic forgetting on old classes. Specifically, a category-guided\ngeometric reasoning is proposed to reason local geometric structures with\ndistinctive 3D characteristics of each class by leveraging intrinsic category\ninformation. We then propose a novel critic-induced geometric attention\nmechanism to distinguish which 3D geometric characteristics within each class\nare beneficial to overcome the catastrophic forgetting on old classes of 3D\nobjects, while preventing the negative influence of useless 3D characteristics.\nIn addition, a dual adaptive fairness compensations strategy is designed to\novercome the forgetting brought by class imbalance, by compensating biased\nweights and predictions of the classifier. Comparison experiments verify the\nstate-of-the-art performance of the proposed InOR-Net model on several public\npoint cloud datasets.\n", "date": "2023-02-20", "categories": []}
{"id": "2302.08741", "title": "New Insights for the Stability-Plasticity Dilemma in Online Continual\n  Learning", "abstract": "  The aim of continual learning is to learn new tasks continuously (i.e.,\nplasticity) without forgetting previously learned knowledge from old tasks\n(i.e., stability). In the scenario of online continual learning, wherein data\ncomes strictly in a streaming manner, the plasticity of online continual\nlearning is more vulnerable than offline continual learning because the\ntraining signal that can be obtained from a single data point is limited. To\novercome the stability-plasticity dilemma in online continual learning, we\npropose an online continual learning framework named multi-scale feature\nadaptation network (MuFAN) that utilizes a richer context encoding extracted\nfrom different levels of a pre-trained network. Additionally, we introduce a\nnovel structure-wise distillation loss and replace the commonly used batch\nnormalization layer with a newly proposed stability-plasticity normalization\nmodule to train MuFAN that simultaneously maintains high plasticity and\nstability. MuFAN outperforms other state-of-the-art continual learning methods\non the SVHN, CIFAR100, miniImageNet, and CORe50 datasets. Extensive experiments\nand ablation studies validate the significance and scalability of each proposed\ncomponent: 1) multi-scale feature maps from a pre-trained encoder, 2) the\nstructure-wise distillation loss, and 3) the stability-plasticity normalization\nmodule in MuFAN. Code is publicly available at\nhttps://github.com/whitesnowdrop/MuFAN.\n", "date": "2023-02-17", "categories": []}
{"id": "2210.06101", "title": "Federated Continual Learning for Text Classification via Selective\n  Inter-client Transfer", "abstract": "  In this work, we combine the two paradigms: Federated Learning (FL) and\nContinual Learning (CL) for text classification task in cloud-edge continuum.\nThe objective of Federated Continual Learning (FCL) is to improve deep learning\nmodels over life time at each client by (relevant and efficient) knowledge\ntransfer without sharing data. Here, we address challenges in minimizing\ninter-client interference while knowledge sharing due to heterogeneous tasks\nacross clients in FCL setup. In doing so, we propose a novel framework,\nFederated Selective Inter-client Transfer (FedSeIT) which selectively combines\nmodel parameters of foreign clients. To further maximize knowledge transfer, we\nassess domain overlap and select informative tasks from the sequence of\nhistorical tasks at each foreign client while preserving privacy. Evaluating\nagainst the baselines, we show improved performance, a gain of (average) 12.4\\%\nin text classification over a sequence of tasks using five datasets from\ndiverse domains. To the best of our knowledge, this is the first work that\napplies FCL to NLP.\n", "date": "2022-10-12", "categories": []}
{"id": "2212.04145", "title": "Decorate the Newcomers: Visual Domain Prompt for Continual Test Time\n  Adaptation", "abstract": "  Continual Test-Time Adaptation (CTTA) aims to adapt the source model to\ncontinually changing unlabeled target domains without access to the source\ndata. Existing methods mainly focus on model-based adaptation in a\nself-training manner, such as predicting pseudo labels for new domain datasets.\nSince pseudo labels are noisy and unreliable, these methods suffer from\ncatastrophic forgetting and error accumulation when dealing with dynamic data\ndistributions. Motivated by the prompt learning in NLP, in this paper, we\npropose to learn an image-level visual domain prompt for target domains while\nhaving the source model parameters frozen. During testing, the changing target\ndatasets can be adapted to the source model by reformulating the input data\nwith the learned visual prompts. Specifically, we devise two types of prompts,\ni.e., domains-specific prompts and domains-agnostic prompts, to extract current\ndomain knowledge and maintain the domain-shared knowledge in the continual\nadaptation. Furthermore, we design a homeostasis-based prompt adaptation\nstrategy to suppress domain-sensitive parameters in domain-invariant prompts to\nlearn domain-shared knowledge more effectively. This transition from the\nmodel-dependent paradigm to the model-free one enables us to bypass the\ncatastrophic forgetting and error accumulation problems. Experiments show that\nour proposed method achieves significant performance gains over\nstate-of-the-art methods on four widely-used benchmarks, including CIFAR-10C,\nCIFAR-100C, ImageNet-C, and VLCS datasets.\n", "date": "2022-12-08", "categories": []}
{"id": "2302.00353", "title": "Towards Label-Efficient Incremental Learning: A Survey", "abstract": "  The current dominant paradigm when building a machine learning model is to\niterate over a dataset over and over until convergence. Such an approach is\nnon-incremental, as it assumes access to all images of all categories at once.\nHowever, for many applications, non-incremental learning is unrealistic. To\nthat end, researchers study incremental learning, where a learner is required\nto adapt to an incoming stream of data with a varying distribution while\npreventing forgetting of past knowledge. Significant progress has been made,\nhowever, the vast majority of works focus on the fully supervised setting,\nmaking these algorithms label-hungry thus limiting their real-life deployment.\nTo that end, in this paper, we make the first attempt to survey recently\ngrowing interest in label-efficient incremental learning. We identify three\nsubdivisions, namely semi-, few-shot- and self-supervised learning to reduce\nlabeling efforts. Finally, we identify novel directions that can further\nenhance label-efficiency and improve incremental learning scalability. Project\nwebsite: https://github.com/kilickaya/label-efficient-il.\n", "date": "2023-02-01", "categories": []}
{"id": "2302.05836", "title": "Theory on Forgetting and Generalization of Continual Learning", "abstract": "  Continual learning (CL), which aims to learn a sequence of tasks, has\nattracted significant recent attention. However, most work has focused on the\nexperimental performance of CL, and theoretical studies of CL are still\nlimited. In particular, there is a lack of understanding on what factors are\nimportant and how they affect \"catastrophic forgetting\" and generalization\nperformance. To fill this gap, our theoretical analysis, under\noverparameterized linear models, provides the first-known explicit form of the\nexpected forgetting and generalization error. Further analysis of such a key\nresult yields a number of theoretical explanations about how\noverparameterization, task similarity, and task ordering affect both forgetting\nand generalization error of CL. More interestingly, by conducting experiments\non real datasets using deep neural networks (DNNs), we show that some of these\ninsights even go beyond the linear models and can be carried over to practical\nsetups. In particular, we use concrete examples to show that our results not\nonly explain some interesting empirical observations in recent studies, but\nalso motivate better practical algorithm designs of CL.\n", "date": "2023-02-12", "categories": []}
{"id": "2302.05936", "title": "Generalized Few-Shot Continual Learning with Contrastive Mixture of\n  Adapters", "abstract": "  The goal of Few-Shot Continual Learning (FSCL) is to incrementally learn\nnovel tasks with limited labeled samples and preserve previous capabilities\nsimultaneously, while current FSCL methods are all for the class-incremental\npurpose. Moreover, the evaluation of FSCL solutions is only the cumulative\nperformance of all encountered tasks, but there is no work on exploring the\ndomain generalization ability. Domain generalization is a challenging yet\npractical task that aims to generalize beyond training domains. In this paper,\nwe set up a Generalized FSCL (GFSCL) protocol involving both class- and\ndomain-incremental situations together with the domain generalization\nassessment. Firstly, two benchmark datasets and protocols are newly arranged,\nand detailed baselines are provided for this unexplored configuration. We find\nthat common continual learning methods have poor generalization ability on\nunseen domains and cannot better cope with the catastrophic forgetting issue in\ncross-incremental tasks. In this way, we further propose a rehearsal-free\nframework based on Vision Transformer (ViT) named Contrastive Mixture of\nAdapters (CMoA). Due to different optimization targets of class increment and\ndomain increment, the CMoA contains two parts: (1) For the class-incremental\nissue, the Mixture of Adapters (MoA) module is incorporated into ViT, then\ncosine similarity regularization and the dynamic weighting are designed to make\neach adapter learn specific knowledge and concentrate on particular classes.\n(2) For the domain-related issues and domain-invariant representation learning,\nwe alleviate the inner-class variation by prototype-calibrated contrastive\nlearning. The codes and protocols are available at\nhttps://github.com/yawencui/CMoA.\n", "date": "2023-02-12", "categories": []}
{"id": "2208.05610", "title": "Memorizing Complementation Network for Few-Shot Class-Incremental\n  Learning", "abstract": "  Few-shot Class-Incremental Learning (FSCIL) aims at learning new concepts\ncontinually with only a few samples, which is prone to suffer the catastrophic\nforgetting and overfitting problems. The inaccessibility of old classes and the\nscarcity of the novel samples make it formidable to realize the trade-off\nbetween retaining old knowledge and learning novel concepts. Inspired by that\ndifferent models memorize different knowledge when learning novel concepts, we\npropose a Memorizing Complementation Network (MCNet) to ensemble multiple\nmodels that complements the different memorized knowledge with each other in\nnovel tasks. Additionally, to update the model with few novel samples, we\ndevelop a Prototype Smoothing Hard-mining Triplet (PSHT) loss to push the novel\nsamples away from not only each other in current task but also the old\ndistribution. Extensive experiments on three benchmark datasets, e.g.,\nCIFAR100, miniImageNet and CUB200, have demonstrated the superiority of our\nproposed method.\n", "date": "2022-08-11", "categories": []}
{"id": "2302.03004", "title": "Neural Collapse Inspired Feature-Classifier Alignment for Few-Shot Class\n  Incremental Learning", "abstract": "  Few-shot class-incremental learning (FSCIL) has been a challenging problem as\nonly a few training samples are accessible for each novel class in the new\nsessions. Finetuning the backbone or adjusting the classifier prototypes\ntrained in the prior sessions would inevitably cause a misalignment between the\nfeature and classifier of old classes, which explains the well-known\ncatastrophic forgetting problem. In this paper, we deal with this misalignment\ndilemma in FSCIL inspired by the recently discovered phenomenon named neural\ncollapse, which reveals that the last-layer features of the same class will\ncollapse into a vertex, and the vertices of all classes are aligned with the\nclassifier prototypes, which are formed as a simplex equiangular tight frame\n(ETF). It corresponds to an optimal geometric structure for classification due\nto the maximized Fisher Discriminant Ratio. We propose a neural collapse\ninspired framework for FSCIL. A group of classifier prototypes are pre-assigned\nas a simplex ETF for the whole label space, including the base session and all\nthe incremental sessions. During training, the classifier prototypes are not\nlearnable, and we adopt a novel loss function that drives the features into\ntheir corresponding prototypes. Theoretical analysis shows that our method\nholds the neural collapse optimality and does not break the feature-classifier\nalignment in an incremental fashion. Experiments on the miniImageNet, CUB-200,\nand CIFAR-100 datasets demonstrate that our proposed framework outperforms the\nstate-of-the-art performances. Code address:\nhttps://github.com/NeuralCollapseApplications/FSCIL\n", "date": "2023-02-06", "categories": []}
{"id": "2302.01386", "title": "Continual Learning with Scaled Gradient Projection", "abstract": "  In neural networks, continual learning results in gradient interference among\nsequential tasks, leading to catastrophic forgetting of old tasks while\nlearning new ones. This issue is addressed in recent methods by storing the\nimportant gradient spaces for old tasks and updating the model orthogonally\nduring new tasks. However, such restrictive orthogonal gradient updates hamper\nthe learning capability of the new tasks resulting in sub-optimal performance.\nTo improve new learning while minimizing forgetting, in this paper we propose a\nScaled Gradient Projection (SGP) method, where we combine the orthogonal\ngradient projections with scaled gradient steps along the important gradient\nspaces for the past tasks. The degree of gradient scaling along these spaces\ndepends on the importance of the bases spanning them. We propose an efficient\nmethod for computing and accumulating importance of these bases using the\nsingular value decomposition of the input representations for each task. We\nconduct extensive experiments ranging from continual image classification to\nreinforcement learning tasks and report better performance with less training\noverhead than the state-of-the-art approaches.\n", "date": "2023-02-02", "categories": []}
{"id": "2302.01766", "title": "Avalanche: A PyTorch Library for Deep Continual Learning", "abstract": "  Continual learning is the problem of learning from a nonstationary stream of\ndata, a fundamental issue for sustainable and efficient training of deep neural\nnetworks over time. Unfortunately, deep learning libraries only provide\nprimitives for offline training, assuming that model's architecture and data\nare fixed. Avalanche is an open source library maintained by the ContinualAI\nnon-profit organization that extends PyTorch by providing first-class support\nfor dynamic architectures, streams of datasets, and incremental training and\nevaluation methods. Avalanche provides a large set of predefined benchmarks and\ntraining algorithms and it is easy to extend and modular while supporting a\nwide range of continual learning scenarios. Documentation is available at\n\\url{https://avalanche.continualai.org}.\n", "date": "2023-02-02", "categories": []}
{"id": "2302.00858", "title": "Online Continual Learning via the Knowledge Invariant and Spread-out\n  Properties", "abstract": "  The goal of continual learning is to provide intelligent agents that are\ncapable of learning continually a sequence of tasks using the knowledge\nobtained from previous tasks while performing well on prior tasks. However, a\nkey challenge in this continual learning paradigm is catastrophic forgetting,\nnamely adapting a model to new tasks often leads to severe performance\ndegradation on prior tasks. Current memory-based approaches show their success\nin alleviating the catastrophic forgetting problem by replaying examples from\npast tasks when new tasks are learned. However, these methods are infeasible to\ntransfer the structural knowledge from previous tasks i.e., similarities or\ndissimilarities between different instances. Furthermore, the learning bias\nbetween the current and prior tasks is also an urgent problem that should be\nsolved. In this work, we propose a new method, named Online Continual Learning\nvia the Knowledge Invariant and Spread-out Properties (OCLKISP), in which we\nconstrain the evolution of the embedding features via Knowledge Invariant and\nSpread-out Properties (KISP). Thus, we can further transfer the inter-instance\nstructural knowledge of previous tasks while alleviating the forgetting due to\nthe learning bias. We empirically evaluate our proposed method on four popular\nbenchmarks for continual learning: Split CIFAR 100, Split SVHN, Split CUB200\nand Split Tiny-Image-Net. The experimental results show the efficacy of our\nproposed method compared to the state-of-the-art continual learning algorithms.\n", "date": "2023-02-02", "categories": []}
{"id": "2210.01987", "title": "ImpressLearn: Continual Learning via Combined Task Impressions", "abstract": "  This work proposes a new method to sequentially train deep neural networks on\nmultiple tasks without suffering catastrophic forgetting, while endowing it\nwith the capability to quickly adapt to unseen tasks. Starting from existing\nwork on network masking (Wortsman et al., 2020), we show that simply learning a\nlinear combination of a small number of task-specific supermasks (impressions)\non a randomly initialized backbone network is sufficient to both retain\naccuracy on previously learned tasks, as well as achieve high accuracy on\nunseen tasks. In contrast to previous methods, we do not require to generate\ndedicated masks or contexts for each new task, instead leveraging transfer\nlearning to keep per-task parameter overhead small. Our work illustrates the\npower of linearly combining individual impressions, each of which fares poorly\nin isolation, to achieve performance comparable to a dedicated mask. Moreover,\neven repeated impressions from the same task (homogeneous masks), when\ncombined, can approach the performance of heterogeneous combinations if\nsufficiently many impressions are used. Our approach scales more efficiently\nthan existing methods, often requiring orders of magnitude fewer parameters and\ncan function without modification even when task identity is missing. In\naddition, in the setting where task labels are not given at inference, our\nalgorithm gives an often favorable alternative to the one-shot procedure used\nby Wortsman et al., 2020. We evaluate our method on a number of well-known\nimage classification datasets and network architectures.\n", "date": "2022-10-05", "categories": []}
{"id": "2210.05398", "title": "Continual Learning by Modeling Intra-Class Variation", "abstract": "  It has been observed that neural networks perform poorly when the data or\ntasks are presented sequentially. Unlike humans, neural networks suffer greatly\nfrom catastrophic forgetting, making it impossible to perform life-long\nlearning. To address this issue, memory-based continual learning has been\nactively studied and stands out as one of the best-performing methods. We\nexamine memory-based continual learning and identify that large variation in\nthe representation space is crucial for avoiding catastrophic forgetting.\nMotivated by this, we propose to diversify representations by using two types\nof perturbations: model-agnostic variation (i.e., the variation is generated\nwithout the knowledge of the learned neural network) and model-based variation\n(i.e., the variation is conditioned on the learned neural network). We\ndemonstrate that enlarging representational variation serves as a general\nprinciple to improve continual learning. Finally, we perform empirical studies\nwhich demonstrate that our method, as a simple plug-and-play component, can\nconsistently improve a number of memory-based continual learning methods by a\nlarge margin.\n", "date": "2022-10-11", "categories": []}
{"id": "2301.12055", "title": "TIDo: Source-free Task Incremental Learning in Non-stationary\n  Environments", "abstract": "  This work presents an incremental learning approach for autonomous agents to\nlearn new tasks in a non-stationary environment. Updating a DNN model-based\nagent to learn new target tasks requires us to store past training data and\nneeds a large labeled target task dataset. Few-shot task incremental learning\nmethods overcome the limitation of labeled target datasets by adapting trained\nmodels to learn private target classes using a few labeled representatives and\na large unlabeled target dataset. However, the methods assume that the source\nand target tasks are stationary. We propose a one-shot task incremental\nlearning approach that can adapt to non-stationary source and target tasks. Our\napproach minimizes adversarial discrepancy between the model's feature space\nand incoming incremental data to learn an updated hypothesis. We also use\ndistillation loss to reduce catastrophic forgetting of previously learned\ntasks. Finally, we use Gaussian prototypes to generate exemplar instances\neliminating the need to store past training data. Unlike current work in task\nincremental learning, our model can learn both source and target task updates\nincrementally. We evaluate our method on various problem settings for\nincremental object detection and disease prediction model update. We evaluate\nour approach by measuring the performance of shared class and target private\nclass prediction. Our results show that our approach achieved improved\nperformance compared to existing state-of-the-art task incremental learning\nmethods.\n", "date": "2023-01-28", "categories": []}
{"id": "2301.12104", "title": "Unbiased and Efficient Self-Supervised Incremental Contrastive Learning", "abstract": "  Contrastive Learning (CL) has been proved to be a powerful self-supervised\napproach for a wide range of domains, including computer vision and graph\nrepresentation learning. However, the incremental learning issue of CL has\nrarely been studied, which brings the limitation in applying it to real-world\napplications. Contrastive learning identifies the samples with the negative\nones from the noise distribution that changes in the incremental scenarios.\nTherefore, only fitting the change of data without noise distribution causes\nbias, and directly retraining results in low efficiency. To bridge this\nresearch gap, we propose a self-supervised Incremental Contrastive Learning\n(ICL) framework consisting of (i) a novel Incremental InfoNCE (NCE-II) loss\nfunction by estimating the change of noise distribution for old data to\nguarantee no bias with respect to the retraining, (ii) a meta-optimization with\ndeep reinforced Learning Rate Learning (LRL) mechanism which can adaptively\nlearn the learning rate according to the status of the training processes and\nachieve fast convergence which is critical for incremental learning.\nTheoretically, the proposed ICL is equivalent to retraining, which is based on\nsolid mathematical derivation. In practice, extensive experiments in different\ndomains demonstrate that, without retraining a new model, ICL achieves up to\n16.7x training speedup and 16.8x faster convergence with competitive results.\n", "date": "2023-01-28", "categories": []}
{"id": "2301.12131", "title": "Restricted Orthogonal Gradient Projection for Continual Learning", "abstract": "  Continual learning aims to avoid catastrophic forgetting and effectively\nleverage learned experiences to master new knowledge. Existing gradient\nprojection approaches impose hard constraints on the optimization space for new\ntasks to minimize interference, which simultaneously hinders forward knowledge\ntransfer. To address this issue, recent methods reuse frozen parameters with a\ngrowing network, resulting in high computational costs. Thus, it remains a\nchallenge whether we can improve forward knowledge transfer for gradient\nprojection approaches using a fixed network architecture. In this work, we\npropose the Restricted Orthogonal Gradient prOjection (ROGO) framework. The\nbasic idea is to adopt a restricted orthogonal constraint allowing parameters\noptimized in the direction oblique to the whole frozen space to facilitate\nforward knowledge transfer while consolidating previous knowledge. Our\nframework requires neither data buffers nor extra parameters. Extensive\nexperiments have demonstrated the superiority of our framework over several\nstrong baselines. We also provide theoretical guarantees for our relaxing\nstrategy.\n", "date": "2023-01-28", "categories": []}
{"id": "2301.12314", "title": "Progressive Prompts: Continual Learning for Language Models", "abstract": "  We introduce Progressive Prompts - a simple and efficient approach for\ncontinual learning in language models. Our method allows forward transfer and\nresists catastrophic forgetting, without relying on data replay or a large\nnumber of task-specific parameters. Progressive Prompts learns a new soft\nprompt for each task and sequentially concatenates it with the previously\nlearned prompts, while keeping the base model frozen. Experiments on standard\ncontinual learning benchmarks show that our approach outperforms\nstate-of-the-art methods, with an improvement >20% in average test accuracy\nover the previous best-preforming method on T5 model. We also explore a more\nchallenging continual learning setup with longer sequences of tasks and show\nthat Progressive Prompts significantly outperforms prior methods.\n", "date": "2023-01-29", "categories": []}
{"id": "2301.11487", "title": "Projected Subnetworks Scale Adaptation", "abstract": "  Large models support great zero-shot and few-shot capabilities. However,\nupdating these models on new tasks can break performance on previous seen tasks\nand their zero/few-shot unseen tasks. Our work explores how to update\nzero/few-shot learners such that they can maintain performance on seen/unseen\ntasks of previous tasks as well as new tasks. By manipulating the parameter\nupdates of a gradient-based meta learner as the projected task-specific\nsubnetworks, we show improvements for large models to retain seen and zero/few\nshot task performance in online settings.\n", "date": "2023-01-27", "categories": []}
{"id": "2301.11892", "title": "Streaming LifeLong Learning With Any-Time Inference", "abstract": "  Despite rapid advancements in lifelong learning (LLL) research, a large body\nof research mainly focuses on improving the performance in the existing\n\\textit{static} continual learning (CL) setups. These methods lack the ability\nto succeed in a rapidly changing \\textit{dynamic} environment, where an AI\nagent needs to quickly learn new instances in a `single pass' from the\nnon-i.i.d (also possibly temporally contiguous/coherent) data streams without\nsuffering from catastrophic forgetting. For practical applicability, we propose\na novel lifelong learning approach, which is streaming, i.e., a single input\nsample arrives in each time step, single pass, class-incremental, and subject\nto be evaluated at any moment. To address this challenging setup and various\nevaluation protocols, we propose a Bayesian framework, that enables fast\nparameter update, given a single training example, and enables any-time\ninference. We additionally propose an implicit regularizer in the form of\nsnap-shot self-distillation, which effectively minimizes the forgetting\nfurther. We further propose an effective method that efficiently selects a\nsubset of samples for online memory rehearsal and employs a new replay buffer\nmanagement scheme that significantly boosts the overall performance. Our\nempirical evaluations and ablations demonstrate that the proposed method\noutperforms the prior works by large margins.\n", "date": "2023-01-27", "categories": []}
{"id": "2301.09964", "title": "Uncertainty-Aware Distillation for Semi-Supervised Few-Shot\n  Class-Incremental Learning", "abstract": "  Given a model well-trained with a large-scale base dataset, Few-Shot\nClass-Incremental Learning (FSCIL) aims at incrementally learning novel classes\nfrom a few labeled samples by avoiding overfitting, without catastrophically\nforgetting all encountered classes previously. Currently, semi-supervised\nlearning technique that harnesses freely-available unlabeled data to compensate\nfor limited labeled data can boost the performance in numerous vision tasks,\nwhich heuristically can be applied to tackle issues in FSCIL, i.e., the\nSemi-supervised FSCIL (Semi-FSCIL). So far, very limited work focuses on the\nSemi-FSCIL task, leaving the adaptability issue of semi-supervised learning to\nthe FSCIL task unresolved. In this paper, we focus on this adaptability issue\nand present a simple yet efficient Semi-FSCIL framework named Uncertainty-aware\nDistillation with Class-Equilibrium (UaD-CE), encompassing two modules UaD and\nCE. Specifically, when incorporating unlabeled data into each incremental\nsession, we introduce the CE module that employs a class-balanced self-training\nto avoid the gradual dominance of easy-to-classified classes on pseudo-label\ngeneration. To distill reliable knowledge from the reference model, we further\nimplement the UaD module that combines uncertainty-guided knowledge refinement\nwith adaptive distillation. Comprehensive experiments on three benchmark\ndatasets demonstrate that our method can boost the adaptability of unlabeled\ndata with the semi-supervised learning technique in FSCIL tasks.\n", "date": "2023-01-24", "categories": []}
{"id": "2301.07799", "title": "A Domain-Agnostic Approach for Characterization of Lifelong Learning\n  Systems", "abstract": "  Despite the advancement of machine learning techniques in recent years,\nstate-of-the-art systems lack robustness to \"real world\" events, where the\ninput distributions and tasks encountered by the deployed systems will not be\nlimited to the original training context, and systems will instead need to\nadapt to novel distributions and tasks while deployed. This critical gap may be\naddressed through the development of \"Lifelong Learning\" systems that are\ncapable of 1) Continuous Learning, 2) Transfer and Adaptation, and 3)\nScalability. Unfortunately, efforts to improve these capabilities are typically\ntreated as distinct areas of research that are assessed independently, without\nregard to the impact of each separate capability on other aspects of the\nsystem. We instead propose a holistic approach, using a suite of metrics and an\nevaluation framework to assess Lifelong Learning in a principled way that is\nagnostic to specific domains or system techniques. Through five case studies,\nwe show that this suite of metrics can inform the development of varied and\ncomplex Lifelong Learning systems. We highlight how the proposed suite of\nmetrics quantifies performance trade-offs present during Lifelong Learning\nsystem development - both the widely discussed Stability-Plasticity dilemma and\nthe newly proposed relationship between Sample Efficient and Robust Learning.\nFurther, we make recommendations for the formulation and use of metrics to\nguide the continuing development of Lifelong Learning systems and assess their\nprogress in the future.\n", "date": "2023-01-18", "categories": []}
{"id": "2301.07316", "title": "Adaptively Integrated Knowledge Distillation and Prediction Uncertainty\n  for Continual Learning", "abstract": "  Current deep learning models often suffer from catastrophic forgetting of old\nknowledge when continually learning new knowledge. Existing strategies to\nalleviate this issue often fix the trade-off between keeping old knowledge\n(stability) and learning new knowledge (plasticity). However, the\nstability-plasticity trade-off during continual learning may need to be\ndynamically changed for better model performance. In this paper, we propose two\nnovel ways to adaptively balance model stability and plasticity. The first one\nis to adaptively integrate multiple levels of old knowledge and transfer it to\neach block level in the new model. The second one uses prediction uncertainty\nof old knowledge to naturally tune the importance of learning new knowledge\nduring model training. To our best knowledge, this is the first time to connect\nmodel prediction uncertainty and knowledge distillation for continual learning.\nIn addition, this paper applies a modified CutMix particularly to augment the\ndata for old knowledge, further alleviating the catastrophic forgetting issue.\nExtensive evaluations on the CIFAR100 and the ImageNet datasets confirmed the\neffectiveness of the proposed method for continual learning.\n", "date": "2023-01-18", "categories": []}
{"id": "2208.04952", "title": "Continual Prune-and-Select: Class-incremental learning with specialized\n  subnetworks", "abstract": "  The human brain is capable of learning tasks sequentially mostly without\nforgetting. However, deep neural networks (DNNs) suffer from catastrophic\nforgetting when learning one task after another. We address this challenge\nconsidering a class-incremental learning scenario where the DNN sees test data\nwithout knowing the task from which this data originates. During training,\nContinual-Prune-and-Select (CP&S) finds a subnetwork within the DNN that is\nresponsible for solving a given task. Then, during inference, CP&S selects the\ncorrect subnetwork to make predictions for that task. A new task is learned by\ntraining available neuronal connections of the DNN (previously untrained) to\ncreate a new subnetwork by pruning, which can include previously trained\nconnections belonging to other subnetwork(s) because it does not update shared\nconnections. This enables to eliminate catastrophic forgetting by creating\nspecialized regions in the DNN that do not conflict with each other while still\nallowing knowledge transfer across them. The CP&S strategy is implemented with\ndifferent subnetwork selection strategies, revealing superior performance to\nstate-of-the-art continual learning methods tested on various datasets\n(CIFAR-100, CUB-200-2011, ImageNet-100 and ImageNet-1000). In particular, CP&S\nis capable of sequentially learning 10 tasks from ImageNet-1000 keeping an\naccuracy around 94% with negligible forgetting, a first-of-its-kind result in\nclass-incremental learning. To the best of the authors' knowledge, this\nrepresents an improvement in accuracy above 10% when compared to the best\nalternative method.\n", "date": "2022-08-09", "categories": []}
{"id": "2301.05792", "title": "RMM: Reinforced Memory Management for Class-Incremental Learning", "abstract": "  Class-Incremental Learning (CIL) [40] trains classifiers under a strict\nmemory budget: in each incremental phase, learning is done for new data, most\nof which is abandoned to free space for the next phase. The preserved data are\nexemplars used for replaying. However, existing methods use a static and ad hoc\nstrategy for memory allocation, which is often sub-optimal. In this work, we\npropose a dynamic memory management strategy that is optimized for the\nincremental phases and different object classes. We call our method reinforced\nmemory management (RMM), leveraging reinforcement learning. RMM training is not\nnaturally compatible with CIL as the past, and future data are strictly\nnon-accessible during the incremental phases. We solve this by training the\npolicy function of RMM on pseudo CIL tasks, e.g., the tasks built on the data\nof the 0-th phase, and then applying it to target tasks. RMM propagates two\nlevels of actions: Level-1 determines how to split the memory between old and\nnew classes, and Level-2 allocates memory for each specific class. In essence,\nit is an optimizable and general method for memory management that can be used\nin any replaying-based CIL method. For evaluation, we plug RMM into two\ntop-performing baselines (LUCIR+AANets and POD+AANets [30]) and conduct\nexperiments on three benchmarks (CIFAR-100, ImageNet-Subset, and\nImageNet-Full). Our results show clear improvements, e.g., boosting POD+AANets\nby 3.6%, 4.4%, and 1.9% in the 25-Phase settings of the above benchmarks,\nrespectively.\n", "date": "2023-01-14", "categories": []}
{"id": "2301.06030", "title": "Self-recovery of memory via generative replay", "abstract": "  A remarkable capacity of the brain is its ability to autonomously reorganize\nmemories during offline periods. Memory replay, a mechanism hypothesized to\nunderlie biological offline learning, has inspired offline methods for reducing\nforgetting in artificial neural networks in continual learning settings. A\nmemory-efficient and neurally-plausible method is generative replay, which\nachieves state of the art performance on continual learning benchmarks.\nHowever, unlike the brain, standard generative replay does not self-reorganize\nmemories when trained offline on its own replay samples. We propose a novel\narchitecture that augments generative replay with an adaptive, brain-like\ncapacity to autonomously recover memories. We demonstrate this capacity of the\narchitecture across several continual learning tasks and environments.\n", "date": "2023-01-15", "categories": []}
{"id": "2211.05347", "title": "Mitigating Forgetting in Online Continual Learning via Contrasting\n  Semantically Distinct Augmentations", "abstract": "  Online continual learning (OCL) aims to enable model learning from a\nnon-stationary data stream to continuously acquire new knowledge as well as\nretain the learnt one, under the constraints of having limited system size and\ncomputational cost, in which the main challenge comes from the \"catastrophic\nforgetting\" issue -- the inability to well remember the learnt knowledge while\nlearning the new ones. With the specific focus on the class-incremental OCL\nscenario, i.e. OCL for classification, the recent advance incorporates the\ncontrastive learning technique for learning more generalised feature\nrepresentation to achieve the state-of-the-art performance but is still unable\nto fully resolve the catastrophic forgetting. In this paper, we follow the\nstrategy of adopting contrastive learning but further introduce the\nsemantically distinct augmentation technique, in which it leverages strong\naugmentation to generate more data samples, and we show that considering these\nsamples semantically different from their original classes (thus being related\nto the out-of-distribution samples) in the contrastive learning mechanism\ncontributes to alleviate forgetting and facilitate model stability. Moreover,\nin addition to contrastive learning, the typical classification mechanism and\nobjective (i.e. softmax classifier and cross-entropy loss) are included in our\nmodel design for faster convergence and utilising the label information, but\nparticularly equipped with a sampling strategy to tackle the tendency of\nfavouring the new classes (i.e. model bias towards the recently learnt\nclasses). Upon conducting extensive experiments on CIFAR-10, CIFAR-100, and\nMini-Imagenet datasets, our proposed method is shown to achieve superior\nperformance against various baselines.\n", "date": "2022-11-10", "categories": []}
{"id": "2301.05058", "title": "Sparse Coding in a Dual Memory System for Lifelong Learning", "abstract": "  Efficient continual learning in humans is enabled by a rich set of\nneurophysiological mechanisms and interactions between multiple memory systems.\nThe brain efficiently encodes information in non-overlapping sparse codes,\nwhich facilitates the learning of new associations faster with controlled\ninterference with previous associations. To mimic sparse coding in DNNs, we\nenforce activation sparsity along with a dropout mechanism which encourages the\nmodel to activate similar units for semantically similar inputs and have less\noverlap with activation patterns of semantically dissimilar inputs. This\nprovides us with an efficient mechanism for balancing the reusability and\ninterference of features, depending on the similarity of classes across tasks.\nFurthermore, we employ sparse coding in a multiple-memory replay mechanism. Our\nmethod maintains an additional long-term semantic memory that aggregates and\nconsolidates information encoded in the synaptic weights of the working model.\nOur extensive evaluation and characteristics analysis show that equipped with\nthese biologically inspired mechanisms, the model can further mitigate\nforgetting.\n", "date": "2022-12-28", "categories": []}
{"id": "2301.04506", "title": "A Distinct Unsupervised Reference Model From The Environment Helps\n  Continual Learning", "abstract": "  The existing continual learning methods are mainly focused on\nfully-supervised scenarios and are still not able to take advantage of\nunlabeled data available in the environment. Some recent works tried to\ninvestigate semi-supervised continual learning (SSCL) settings in which the\nunlabeled data are available, but it is only from the same distribution as the\nlabeled data. This assumption is still not general enough for real-world\napplications and restricts the utilization of unsupervised data. In this work,\nwe introduce Open-Set Semi-Supervised Continual Learning (OSSCL), a more\nrealistic semi-supervised continual learning setting in which\nout-of-distribution (OoD) unlabeled samples in the environment are assumed to\ncoexist with the in-distribution ones. Under this configuration, we present a\nmodel with two distinct parts: (i) the reference network captures\ngeneral-purpose and task-agnostic knowledge in the environment by using a broad\nspectrum of unlabeled samples, (ii) the learner network is designed to learn\ntask-specific representations by exploiting supervised samples. The reference\nmodel both provides a pivotal representation space and also segregates\nunlabeled data to exploit them more efficiently. By performing a diverse range\nof experiments, we show the superior performance of our model compared with\nother competitors and prove the effectiveness of each component of the proposed\nmodel.\n", "date": "2023-01-11", "categories": []}
{"id": "2211.06866", "title": "Mining Unseen Classes via Regional Objectness: A Simple Baseline for\n  Incremental Segmentation", "abstract": "  Incremental or continual learning has been extensively studied for image\nclassification tasks to alleviate catastrophic forgetting, a phenomenon that\nearlier learned knowledge is forgotten when learning new concepts. For class\nincremental semantic segmentation, such a phenomenon often becomes much worse\ndue to the background shift, i.e., some concepts learned at previous stages are\nassigned to the background class at the current training stage, therefore,\nsignificantly reducing the performance of these old concepts. To address this\nissue, we propose a simple yet effective method in this paper, named Mining\nunseen Classes via Regional Objectness for Segmentation (MicroSeg). Our\nMicroSeg is based on the assumption that background regions with strong\nobjectness possibly belong to those concepts in the historical or future\nstages. Therefore, to avoid forgetting old knowledge at the current training\nstage, our MicroSeg first splits the given image into hundreds of segment\nproposals with a proposal generator. Those segment proposals with strong\nobjectness from the background are then clustered and assigned newly-defined\nlabels during the optimization. In this way, the distribution characterizes of\nold concepts in the feature space could be better perceived, relieving the\ncatastrophic forgetting caused by the background shift accordingly. Extensive\nexperiments on Pascal VOC and ADE20K datasets show competitive results with\nstate-of-the-art, well validating the effectiveness of the proposed MicroSeg.\n", "date": "2022-11-13", "categories": []}
{"id": "2301.02464", "title": "Architect, Regularize and Replay (ARR): a Flexible Hybrid Approach for\n  Continual Learning", "abstract": "  In recent years we have witnessed a renewed interest in machine learning\nmethodologies, especially for deep representation learning, that could overcome\nbasic i.i.d. assumptions and tackle non-stationary environments subject to\nvarious distributional shifts or sample selection biases. Within this context,\nseveral computational approaches based on architectural priors, regularizers\nand replay policies have been proposed with different degrees of success\ndepending on the specific scenario in which they were developed and assessed.\nHowever, designing comprehensive hybrid solutions that can flexibly and\ngenerally be applied with tunable efficiency-effectiveness trade-offs still\nseems a distant goal. In this paper, we propose \"Architect, Regularize and\nReplay\" (ARR), an hybrid generalization of the renowned AR1 algorithm and its\nvariants, that can achieve state-of-the-art results in classic scenarios (e.g.\nclass-incremental learning) but also generalize to arbitrary data streams\ngenerated from real-world datasets such as CIFAR-100, CORe50 and ImageNet-1000.\n", "date": "2023-01-06", "categories": []}
{"id": "2212.14284", "title": "Resolving Task Confusion in Dynamic Expansion Architectures for Class\n  Incremental Learning", "abstract": "  The dynamic expansion architecture is becoming popular in class incremental\nlearning, mainly due to its advantages in alleviating catastrophic forgetting.\nHowever, task confusion is not well assessed within this framework, e.g., the\ndiscrepancy between classes of different tasks is not well learned (i.e.,\ninter-task confusion, ITC), and certain priority is still given to the latest\nclass batch (i.e., old-new confusion, ONC). We empirically validate the side\neffects of the two types of confusion. Meanwhile, a novel solution called Task\nCorrelated Incremental Learning (TCIL) is proposed to encourage discriminative\nand fair feature utilization across tasks. TCIL performs a multi-level\nknowledge distillation to propagate knowledge learned from old tasks to the new\none. It establishes information flow paths at both feature and logit levels,\nenabling the learning to be aware of old classes. Besides, attention mechanism\nand classifier re-scoring are applied to generate more fair classification\nscores. We conduct extensive experiments on CIFAR100 and ImageNet100 datasets.\nThe results demonstrate that TCIL consistently achieves state-of-the-art\naccuracy. It mitigates both ITC and ONC, while showing advantages in battle\nwith catastrophic forgetting even no rehearsal memory is reserved.\n", "date": "2022-12-29", "categories": []}
{"id": "2301.01033", "title": "Dissecting Continual Learning a Structural and Data Analysis", "abstract": "  Continual Learning (CL) is a field dedicated to devise algorithms able to\nachieve lifelong learning. Overcoming the knowledge disruption of previously\nacquired concepts, a drawback affecting deep learning models and that goes by\nthe name of catastrophic forgetting, is a hard challenge. Currently, deep\nlearning methods can attain impressive results when the data modeled does not\nundergo a considerable distributional shift in subsequent learning sessions,\nbut whenever we expose such systems to this incremental setting, performance\ndrop very quickly. Overcoming this limitation is fundamental as it would allow\nus to build truly intelligent systems showing stability and plasticity.\nSecondly, it would allow us to overcome the onerous limitation of retraining\nthese architectures from scratch with the new updated data. In this thesis, we\ntackle the problem from multiple directions. In a first study, we show that in\nrehearsal-based techniques (systems that use memory buffer), the quantity of\ndata stored in the rehearsal buffer is a more important factor over the quality\nof the data. Secondly, we propose one of the early works of incremental\nlearning on ViTs architectures, comparing functional, weight and attention\nregularization approaches and propose effective novel a novel asymmetric loss.\nAt the end we conclude with a study on pretraining and how it affects the\nperformance in Continual Learning, raising some questions about the effective\nprogression of the field. We then conclude with some future directions and\nclosing remarks.\n", "date": "2023-01-03", "categories": []}
{"id": "2301.00141", "title": "Self-Activating Neural Ensembles for Continual Reinforcement Learning", "abstract": "  The ability for an agent to continuously learn new skills without\ncatastrophically forgetting existing knowledge is of critical importance for\nthe development of generally intelligent agents. Most methods devised to\naddress this problem depend heavily on well-defined task boundaries, and thus\ndepend on human supervision. Our task-agnostic method, Self-Activating Neural\nEnsembles (SANE), uses a modular architecture designed to avoid catastrophic\nforgetting without making any such assumptions. At the beginning of each\ntrajectory, a module in the SANE ensemble is activated to determine the agent's\nnext policy. During training, new modules are created as needed and only\nactivated modules are updated to ensure that unused modules remain unchanged.\nThis system enables our method to retain and leverage old skills, while growing\nand learning new ones. We demonstrate our approach on visually rich\nprocedurally generated environments.\n", "date": "2022-12-31", "categories": []}
{"id": "2301.00620", "title": "Dynamically Modular and Sparse General Continual Learning", "abstract": "  Real-world applications often require learning continuously from a stream of\ndata under ever-changing conditions. When trying to learn from such\nnon-stationary data, deep neural networks (DNNs) undergo catastrophic\nforgetting of previously learned information. Among the common approaches to\navoid catastrophic forgetting, rehearsal-based methods have proven effective.\nHowever, they are still prone to forgetting due to task-interference as all\nparameters respond to all tasks. To counter this, we take inspiration from\nsparse coding in the brain and introduce dynamic modularity and sparsity\n(Dynamos) for rehearsal-based general continual learning. In this setup, the\nDNN learns to respond to stimuli by activating relevant subsets of neurons. We\ndemonstrate the effectiveness of Dynamos on multiple datasets under challenging\ncontinual learning evaluation protocols. Finally, we show that our method\nlearns representations that are modular and specialized, while maintaining\nreusability by activating subsets of neurons with overlaps corresponding to the\nsimilarity of stimuli.\n", "date": "2023-01-02", "categories": []}
{"id": "2208.08367", "title": "Ask Question First for Enhancing Lifelong Language Learning", "abstract": "  Lifelong language learning aims to stream learning NLP tasks while retaining\nknowledge of previous tasks. Previous works based on the language model and\nfollowing data-free constraint approaches have explored formatting all data as\n\"begin token (\\textit{B}) + context (\\textit{C}) + question (\\textit{Q}) +\nanswer (\\textit{A})\" for different tasks. However, they still suffer from\ncatastrophic forgetting and are exacerbated when the previous task's pseudo\ndata is insufficient for the following reasons: (1) The model has difficulty\ngenerating task-corresponding pseudo data, and (2) \\textit{A} is prone to error\nwhen \\textit{A} and \\textit{C} are separated by \\textit{Q} because the\ninformation of the \\textit{C} is diminished before generating \\textit{A}.\nTherefore, we propose the Ask Question First and Replay Question (AQF-RQ),\nincluding a novel data format \"\\textit{BQCA}\" and a new training task to train\npseudo questions of previous tasks. Experimental results demonstrate that\nAQF-RQ makes it easier for the model to generate more pseudo data that match\ncorresponding tasks, and is more robust to both sufficient and insufficient\npseudo-data when the task boundary is both clear and unclear. AQF-RQ can\nachieve only 0.36\\% lower performance than multi-task learning.\n", "date": "2022-08-17", "categories": []}
{"id": "2212.12643", "title": "Utilizing Priming to Identify Optimal Class Ordering to Alleviate\n  Catastrophic Forgetting", "abstract": "  In order for artificial neural networks to begin accurately mimicking\nbiological ones, they must be able to adapt to new exigencies without\nforgetting what they have learned from previous training. Lifelong learning\napproaches to artificial neural networks attempt to strive towards this goal,\nyet have not progressed far enough to be realistically deployed for natural\nlanguage processing tasks. The proverbial roadblock of catastrophic forgetting\nstill gate-keeps researchers from an adequate lifelong learning model. While\nefforts are being made to quell catastrophic forgetting, there is a lack of\nresearch that looks into the importance of class ordering when training on new\nclasses for incremental learning. This is surprising as the ordering of\n\"classes\" that humans learn is heavily monitored and incredibly important.\nWhile heuristics to develop an ideal class order have been researched, this\npaper examines class ordering as it relates to priming as a scheme for\nincremental class learning. By examining the connections between various\nmethods of priming found in humans and how those are mimicked yet remain\nunexplained in life-long machine learning, this paper provides a better\nunderstanding of the similarities between our biological systems and the\nsynthetic systems while simultaneously improving current practices to combat\ncatastrophic forgetting. Through the merging of psychological priming practices\nwith class ordering, this paper is able to identify a generalizable method for\nclass ordering in NLP incremental learning tasks that consistently outperforms\nrandom class ordering.\n", "date": "2022-12-24", "categories": []}
{"id": "2212.13242", "title": "Saliency-Augmented Memory Completion for Continual Learning", "abstract": "  Continual Learning is considered a key step toward next-generation Artificial\nIntelligence. Among various methods, replay-based approaches that maintain and\nreplay a small episodic memory of previous samples are one of the most\nsuccessful strategies against catastrophic forgetting. However, since\nforgetting is inevitable given bounded memory and unbounded tasks, how to\nforget is a problem continual learning must address. Therefore, beyond simply\navoiding catastrophic forgetting, an under-explored issue is how to reasonably\nforget while ensuring the merits of human memory, including 1. storage\nefficiency, 2. generalizability, and 3. some interpretability. To achieve these\nsimultaneously, our paper proposes a new saliency-augmented memory completion\nframework for continual learning, inspired by recent discoveries in memory\ncompletion separation in cognitive neuroscience. Specifically, we innovatively\npropose to store the part of the image most important to the tasks in episodic\nmemory by saliency map extraction and memory encoding. When learning new tasks,\nprevious data from memory are inpainted by an adaptive data generation module,\nwhich is inspired by how humans complete episodic memory. The module's\nparameters are shared across all tasks and it can be jointly trained with a\ncontinual learning classifier as bilevel optimization. Extensive experiments on\nseveral continual learning and image classification benchmarks demonstrate the\nproposed method's effectiveness and efficiency.\n", "date": "2022-12-26", "categories": []}
{"id": "2210.04993", "title": "Continual Learning with Evolving Class Ontologies", "abstract": "  Lifelong learners must recognize concept vocabularies that evolve over time.\nA common yet underexplored scenario is learning with class labels that\ncontinually refine/expand old classes. For example, humans learn to recognize\n${\\tt dog}$ before dog breeds. In practical settings, dataset\n$\\textit{versioning}$ often introduces refinement to ontologies, such as\nautonomous vehicle benchmarks that refine a previous ${\\tt vehicle}$ class into\n${\\tt school-bus}$ as autonomous operations expand to new cities. This paper\nformalizes a protocol for studying the problem of $\\textit{Learning with\nEvolving Class Ontology}$ (LECO). LECO requires learning classifiers in\ndistinct time periods (TPs); each TP introduces a new ontology of \"fine\" labels\nthat refines old ontologies of \"coarse\" labels (e.g., dog breeds that refine\nthe previous ${\\tt dog}$). LECO explores such questions as whether to annotate\nnew data or relabel the old, how to leverage coarse labels, and whether to\nfinetune the previous TP's model or train from scratch. To answer these\nquestions, we leverage insights from related problems such as class-incremental\nlearning. We validate them under the LECO protocol through the lens of image\nclassification (CIFAR and iNaturalist) and semantic segmentation (Mapillary).\nOur experiments lead to surprising conclusions; while the current status quo is\nto relabel existing datasets with new ontologies (such as COCO-to-LVIS or\nMapillary1.2-to-2.0), LECO demonstrates that a far better strategy is to\nannotate $\\textit{new}$ data with the new ontology. However, this produces an\naggregate dataset with inconsistent old-vs-new labels, complicating learning.\nTo address this challenge, we adopt methods from semi-supervised and\npartial-label learning. Such strategies can surprisingly be made near-optimal,\napproaching an \"oracle\" that learns on the aggregate dataset exhaustively\nlabeled with the newest ontology.\n", "date": "2022-10-10", "categories": []}
{"id": "2212.06833", "title": "3rd Continual Learning Workshop Challenge on Egocentric Category and\n  Instance Level Object Understanding", "abstract": "  Continual Learning, also known as Lifelong or Incremental Learning, has\nrecently gained renewed interest among the Artificial Intelligence research\ncommunity. Recent research efforts have quickly led to the design of novel\nalgorithms able to reduce the impact of the catastrophic forgetting phenomenon\nin deep neural networks. Due to this surge of interest in the field, many\ncompetitions have been held in recent years, as they are an excellent\nopportunity to stimulate research in promising directions. This paper\nsummarizes the ideas, design choices, rules, and results of the challenge held\nat the 3rd Continual Learning in Computer Vision (CLVision) Workshop at CVPR\n2022. The focus of this competition is the complex continual object detection\ntask, which is still underexplored in literature compared to classification\ntasks. The challenge is based on the challenge version of the novel EgoObjects\ndataset, a large-scale egocentric object dataset explicitly designed to\nbenchmark continual learning algorithms for egocentric category-/instance-level\nobject understanding, which covers more than 1k unique main objects and 250+\ncategories in around 100k video frames.\n", "date": "2022-12-13", "categories": []}
{"id": "2212.04316", "title": "Bio-Inspired, Task-Free Continual Learning through Activity\n  Regularization", "abstract": "  The ability to sequentially learn multiple tasks without forgetting is a key\nskill of biological brains, whereas it represents a major challenge to the\nfield of deep learning. To avoid catastrophic forgetting, various continual\nlearning (CL) approaches have been devised. However, these usually require\ndiscrete task boundaries. This requirement seems biologically implausible and\noften limits the application of CL methods in the real world where tasks are\nnot always well defined. Here, we take inspiration from neuroscience, where\nsparse, non-overlapping neuronal representations have been suggested to prevent\ncatastrophic forgetting. As in the brain, we argue that these sparse\nrepresentations should be chosen on the basis of feed forward\n(stimulus-specific) as well as top-down (context-specific) information. To\nimplement such selective sparsity, we use a bio-plausible form of hierarchical\ncredit assignment known as Deep Feedback Control (DFC) and combine it with a\nwinner-take-all sparsity mechanism. In addition to sparsity, we introduce\nlateral recurrent connections within each layer to further protect previously\nlearned representations. We evaluate the new sparse-recurrent version of DFC on\nthe split-MNIST computer vision benchmark and show that only the combination of\nsparsity and intra-layer recurrent connections improves CL performance with\nrespect to standard backpropagation. Our method achieves similar performance to\nwell-known CL methods, such as Elastic Weight Consolidation and Synaptic\nIntelligence, without requiring information about task boundaries. Overall, we\nshowcase the idea of adopting computational principles from the brain to derive\nnew, task-free learning algorithms for CL.\n", "date": "2022-12-08", "categories": []}
{"id": "2212.01738", "title": "FedKNOW: Federated Continual Learning with Signature Task Knowledge\n  Integration at Edge", "abstract": "  Deep Neural Networks (DNNs) have been ubiquitously adopted in internet of\nthings and are becoming an integral of our daily life. When tackling the\nevolving learning tasks in real world, such as classifying different types of\nobjects, DNNs face the challenge to continually retrain themselves according to\nthe tasks on different edge devices. Federated continual learning is a\npromising technique that offers partial solutions but yet to overcome the\nfollowing difficulties: the significant accuracy loss due to the limited\non-device processing, the negative knowledge transfer caused by the limited\ncommunication of non-IID data, and the limited scalability on the tasks and\nedge devices. In this paper, we propose FedKNOW, an accurate and scalable\nfederated continual learning framework, via a novel concept of signature task\nknowledge. FedKNOW is a client side solution that continuously extracts and\nintegrates the knowledge of signature tasks which are highly influenced by the\ncurrent task. Each client of FedKNOW is composed of a knowledge extractor, a\ngradient restorer and, most importantly, a gradient integrator. Upon training\nfor a new task, the gradient integrator ensures the prevention of catastrophic\nforgetting and mitigation of negative knowledge transfer by effectively\ncombining signature tasks identified from the past local tasks and other\nclients' current tasks through the global model. We implement FedKNOW in\nPyTorch and extensively evaluate it against state-of-the-art techniques using\npopular federated continual learning benchmarks. Extensive evaluation results\non heterogeneous edge devices show that FedKNOW improves model accuracy by\n63.24% without increasing model training time, reduces communication cost by\n34.28%, and achieves more improvements under difficult scenarios such as large\nnumbers of tasks or clients, and training different complex networks.\n", "date": "2022-12-04", "categories": []}
{"id": "2212.02057", "title": "DA-CIL: Towards Domain Adaptive Class-Incremental 3D Object Detection", "abstract": "  Deep learning has achieved notable success in 3D object detection with the\nadvent of large-scale point cloud datasets. However, severe performance\ndegradation in the past trained classes, i.e., catastrophic forgetting, still\nremains a critical issue for real-world deployment when the number of classes\nis unknown or may vary. Moreover, existing 3D class-incremental detection\nmethods are developed for the single-domain scenario, which fail when\nencountering domain shift caused by different datasets, varying environments,\netc. In this paper, we identify the unexplored yet valuable scenario, i.e.,\nclass-incremental learning under domain shift, and propose a novel 3D domain\nadaptive class-incremental object detection framework, DA-CIL, in which we\ndesign a novel dual-domain copy-paste augmentation method to construct multiple\naugmented domains for diversifying training distributions, thereby facilitating\ngradual domain adaptation. Then, multi-level consistency is explored to\nfacilitate dual-teacher knowledge distillation from different domains for\ndomain adaptive class-incremental learning. Extensive experiments on various\ndatasets demonstrate the effectiveness of the proposed method over baselines in\nthe domain adaptive class-incremental learning scenario.\n", "date": "2022-12-05", "categories": []}
{"id": "2212.00972", "title": "Cloud-Device Collaborative Adaptation to Continual Changing Environments\n  in the Real-world", "abstract": "  When facing changing environments in the real world, the lightweight model on\nclient devices suffers from severe performance drops under distribution shifts.\nThe main limitations of the existing device model lie in (1) unable to update\ndue to the computation limit of the device, (2) the limited generalization\nability of the lightweight model. Meanwhile, recent large models have shown\nstrong generalization capability on the cloud while they can not be deployed on\nclient devices due to poor computation constraints. To enable the device model\nto deal with changing environments, we propose a new learning paradigm of\nCloud-Device Collaborative Continual Adaptation, which encourages collaboration\nbetween cloud and device and improves the generalization of the device model.\nBased on this paradigm, we further propose an Uncertainty-based Visual Prompt\nAdapted (U-VPA) teacher-student model to transfer the generalization capability\nof the large model on the cloud to the device model. Specifically, we first\ndesign the Uncertainty Guided Sampling (UGS) to screen out challenging data\ncontinuously and transmit the most out-of-distribution samples from the device\nto the cloud. Then we propose a Visual Prompt Learning Strategy with\nUncertainty guided updating (VPLU) to specifically deal with the selected\nsamples with more distribution shifts. We transmit the visual prompts to the\ndevice and concatenate them with the incoming data to pull the device testing\ndistribution closer to the cloud training distribution. We conduct extensive\nexperiments on two object detection datasets with continually changing\nenvironments. Our proposed U-VPA teacher-student framework outperforms previous\nstate-of-the-art test time adaptation and device-cloud collaboration methods.\nThe code and datasets will be released.\n", "date": "2022-12-02", "categories": []}
{"id": "2212.00572", "title": "GMM-IL: Image Classification using Incrementally Learnt, Independent\n  Probabilistic Models for Small Sample Sizes", "abstract": "  Current deep learning classifiers, carry out supervised learning and store\nclass discriminatory information in a set of shared network weights. These\nweights cannot be easily altered to incrementally learn additional classes,\nsince the classification weights all require retraining to prevent old class\ninformation from being lost and also require the previous training data to be\npresent. We present a novel two stage architecture which couples visual feature\nlearning with probabilistic models to represent each class in the form of a\nGaussian Mixture Model. By using these independent class representations within\nour classifier, we outperform a benchmark of an equivalent network with a\nSoftmax head, obtaining increased accuracy for sample sizes smaller than 12 and\nincreased weighted F1 score for 3 imbalanced class profiles in that sample\nrange. When learning new classes our classifier exhibits no catastrophic\nforgetting issues and only requires the new classes' training images to be\npresent. This enables a database of growing classes over time which can be\nvisually indexed and reasoned over.\n", "date": "2022-12-01", "categories": []}
{"id": "2211.16759", "title": "General policy mapping: online continual reinforcement learning inspired\n  on the insect brain", "abstract": "  We have developed a model for online continual or lifelong reinforcement\nlearning (RL) inspired on the insect brain. Our model leverages the offline\ntraining of a feature extraction and a common general policy layer to enable\nthe convergence of RL algorithms in online settings. Sharing a common policy\nlayer across tasks leads to positive backward transfer, where the agent\ncontinuously improved in older tasks sharing the same underlying general\npolicy. Biologically inspired restrictions to the agent's network are key for\nthe convergence of RL algorithms. This provides a pathway towards efficient\nonline RL in resource-constrained scenarios.\n", "date": "2022-11-30", "categories": []}
{"id": "2211.17012", "title": "Correlation of the importances of neural network weights calculated by\n  modern methods of overcoming catastrophic forgetting", "abstract": "  Following the invention in 2017 of the EWC method, several methods have been\nproposed to calculate the importance of neural network weights for use in the\nEWC method. Despite the significant difference in calculating the importance of\nweights, they all proved to be effective. Accordingly, a reasonable question\narises as to how similar the importances of the weights calculated by different\nmethods. To answer this question, we calculated layer-by-layer correlations of\nthe importance of weights calculated by all those methods. As a result, it\nturned out that the importances of several of the methods correlated with each\nother quite strongly and we were able to present an explanation for such a\ncorrelation. At the same time, for other methods, the correlation can vary from\nstrong on some layers of the network to negative on other layers. Which raises\na reasonable question: why, despite the very different calculation methods, all\nthose importances allow EWC method to overcome the catastrophic forgetting of\nneural networks perfectly?\n", "date": "2022-10-24", "categories": []}
{"id": "2211.15969", "title": "Isolation and Impartial Aggregation: A Paradigm of Incremental Learning\n  without Interference", "abstract": "  This paper focuses on the prevalent performance imbalance in the stages of\nincremental learning. To avoid obvious stage learning bottlenecks, we propose a\nbrand-new stage-isolation based incremental learning framework, which leverages\na series of stage-isolated classifiers to perform the learning task of each\nstage without the interference of others. To be concrete, to aggregate multiple\nstage classifiers as a uniform one impartially, we first introduce a\ntemperature-controlled energy metric for indicating the confidence score levels\nof the stage classifiers. We then propose an anchor-based energy\nself-normalization strategy to ensure the stage classifiers work at the same\nenergy level. Finally, we design a voting-based inference augmentation strategy\nfor robust inference. The proposed method is rehearsal free and can work for\nalmost all continual learning scenarios. We evaluate the proposed method on\nfour large benchmarks. Extensive results demonstrate the superiority of the\nproposed method in setting up new state-of-the-art overall performance.\n\\emph{Code is available at} \\url{https://github.com/iamwangyabin/ESN}.\n", "date": "2022-11-29", "categories": []}
{"id": "2211.16201", "title": "Lifelong Person Re-Identification via Knowledge Refreshing and\n  Consolidation", "abstract": "  Lifelong person re-identification (LReID) is in significant demand for\nreal-world development as a large amount of ReID data is captured from diverse\nlocations over time and cannot be accessed at once inherently. However, a key\nchallenge for LReID is how to incrementally preserve old knowledge and\ngradually add new capabilities to the system. Unlike most existing LReID\nmethods, which mainly focus on dealing with catastrophic forgetting, our focus\nis on a more challenging problem, which is, not only trying to reduce the\nforgetting on old tasks but also aiming to improve the model performance on\nboth new and old tasks during the lifelong learning process. Inspired by the\nbiological process of human cognition where the somatosensory neocortex and the\nhippocampus work together in memory consolidation, we formulated a model called\nKnowledge Refreshing and Consolidation (KRC) that achieves both positive\nforward and backward transfer. More specifically, a knowledge refreshing scheme\nis incorporated with the knowledge rehearsal mechanism to enable bi-directional\nknowledge transfer by introducing a dynamic memory model and an adaptive\nworking model. Moreover, a knowledge consolidation scheme operating on the dual\nspace further improves model stability over the long term. Extensive\nevaluations show KRC's superiority over the state-of-the-art LReID methods on\nchallenging pedestrian benchmarks.\n", "date": "2022-11-29", "categories": []}
{"id": "2211.14763", "title": "Multi-Label Continual Learning using Augmented Graph Convolutional\n  Network", "abstract": "  Multi-Label Continual Learning (MLCL) builds a class-incremental framework in\na sequential multi-label image recognition data stream. The critical challenges\nof MLCL are the construction of label relationships on past-missing and\nfuture-missing partial labels of training data and the catastrophic forgetting\non old classes, resulting in poor generalization. To solve the problems, the\nstudy proposes an Augmented Graph Convolutional Network (AGCN++) that can\nconstruct the cross-task label relationships in MLCL and sustain catastrophic\nforgetting. First, we build an Augmented Correlation Matrix (ACM) across all\nseen classes, where the intra-task relationships derive from the hard label\nstatistics. In contrast, the inter-task relationships leverage hard and soft\nlabels from data and a constructed expert network. Then, we propose a novel\npartial label encoder (PLE) for MLCL, which can extract dynamic class\nrepresentation for each partial label image as graph nodes and help generate\nsoft labels to create a more convincing ACM and suppress forgetting. Last, to\nsuppress the forgetting of label dependencies across old tasks, we propose a\nrelationship-preserving constrainter to construct label relationships. The\ninter-class topology can be augmented automatically, which also yields\neffective class representations. The proposed method is evaluated using two\nmulti-label image benchmarks. The experimental results show that the proposed\nway is effective for MLCL image recognition and can build convincing\ncorrelations across tasks even if the labels of previous tasks are missing.\n", "date": "2022-11-27", "categories": []}
{"id": "2211.15215", "title": "Progressive Learning without Forgetting", "abstract": "  Learning from changing tasks and sequential experience without forgetting the\nobtained knowledge is a challenging problem for artificial neural networks. In\nthis work, we focus on two challenging problems in the paradigm of Continual\nLearning (CL) without involving any old data: (i) the accumulation of\ncatastrophic forgetting caused by the gradually fading knowledge space from\nwhich the model learns the previous knowledge; (ii) the uncontrolled tug-of-war\ndynamics to balance the stability and plasticity during the learning of new\ntasks. In order to tackle these problems, we present Progressive Learning\nwithout Forgetting (PLwF) and a credit assignment regime in the optimizer. PLwF\ndensely introduces model functions from previous tasks to construct a knowledge\nspace such that it contains the most reliable knowledge on each task and the\ndistribution information of different tasks, while credit assignment controls\nthe tug-of-war dynamics by removing gradient conflict through projection.\nExtensive ablative experiments demonstrate the effectiveness of PLwF and credit\nassignment. In comparison with other CL methods, we report notably better\nresults even without relying on any raw data.\n", "date": "2022-11-28", "categories": []}
{"id": "2206.09059", "title": "CLiMB: A Continual Learning Benchmark for Vision-and-Language Tasks", "abstract": "  Current state-of-the-art vision-and-language models are evaluated on tasks\neither individually or in a multi-task setting, overlooking the challenges of\ncontinually learning (CL) tasks as they arrive. Existing CL benchmarks have\nfacilitated research on task adaptation and mitigating \"catastrophic\nforgetting\", but are limited to vision-only and language-only tasks. We present\nCLiMB, a benchmark to study the challenge of learning multimodal tasks in a CL\nsetting, and to systematically evaluate how upstream continual learning can\nrapidly generalize to new multimodal and unimodal tasks. CLiMB includes\nimplementations of several CL algorithms and a modified Vision-Language\nTransformer (ViLT) model that can be deployed on both multimodal and unimodal\ntasks. We find that common CL methods can help mitigate forgetting during\nmultimodal task learning, but do not enable cross-task knowledge transfer. We\nenvision that CLiMB will facilitate research on a new class of CL algorithms\nfor this challenging multimodal setting.\n", "date": "2022-06-18", "categories": []}
{"id": "2210.07783", "title": "Prompt Conditioned VAE: Enhancing Generative Replay for Lifelong\n  Learning in Task-Oriented Dialogue", "abstract": "  Lifelong learning (LL) is vital for advanced task-oriented dialogue (ToD)\nsystems. To address the catastrophic forgetting issue of LL, generative replay\nmethods are widely employed to consolidate past knowledge with generated pseudo\nsamples. However, most existing generative replay methods use only a single\ntask-specific token to control their models. This scheme is usually not strong\nenough to constrain the generative model due to insufficient information\ninvolved. In this paper, we propose a novel method, prompt conditioned VAE for\nlifelong learning (PCLL), to enhance generative replay by incorporating tasks'\nstatistics. PCLL captures task-specific distributions with a conditional\nvariational autoencoder, conditioned on natural language prompts to guide the\npseudo-sample generation. Moreover, it leverages a distillation process to\nfurther consolidate past knowledge by alleviating the noise in pseudo samples.\nExperiments on natural language understanding tasks of ToD systems demonstrate\nthat PCLL significantly outperforms competitive baselines in building LL\nmodels.\n", "date": "2022-10-14", "categories": []}
{"id": "2211.13823", "title": "Neural Weight Search for Scalable Task Incremental Learning", "abstract": "  Task incremental learning aims to enable a system to maintain its performance\non previously learned tasks while learning new tasks, solving the problem of\ncatastrophic forgetting. One promising approach is to build an individual\nnetwork or sub-network for future tasks. However, this leads to an ever-growing\nmemory due to saving extra weights for new tasks and how to address this issue\nhas remained an open problem in task incremental learning. In this paper, we\nintroduce a novel Neural Weight Search technique that designs a fixed search\nspace where the optimal combinations of frozen weights can be searched to build\nnew models for novel tasks in an end-to-end manner, resulting in scalable and\ncontrollable memory growth. Extensive experiments on two benchmarks, i.e.,\nSplit-CIFAR-100 and CUB-to-Sketches, show our method achieves state-of-the-art\nperformance with respect to both average inference accuracy and total memory\ncost.\n", "date": "2022-11-24", "categories": []}
{"id": "2211.13999", "title": "CoMFormer: Continual Learning in Semantic and Panoptic Segmentation", "abstract": "  Continual learning for segmentation has recently seen increasing interest.\nHowever, all previous works focus on narrow semantic segmentation and disregard\npanoptic segmentation, an important task with real-world impacts. %a In this\npaper, we present the first continual learning model capable of operating on\nboth semantic and panoptic segmentation. Inspired by recent transformer\napproaches that consider segmentation as a mask-classification problem, we\ndesign CoMFormer. Our method carefully exploits the properties of transformer\narchitectures to learn new classes over time. Specifically, we propose a novel\nadaptive distillation loss along with a mask-based pseudo-labeling technique to\neffectively prevent forgetting. To evaluate our approach, we introduce a novel\ncontinual panoptic segmentation benchmark on the challenging ADE20K dataset.\nOur CoMFormer outperforms all the existing baselines by forgetting less old\nclasses but also learning more effectively new classes. In addition, we also\nreport an extensive evaluation in the large-scale continual semantic\nsegmentation scenario showing that CoMFormer also significantly outperforms\nstate-of-the-art methods.\n", "date": "2022-11-25", "categories": []}
{"id": "2211.14177", "title": "Overcoming Catastrophic Forgetting by XAI", "abstract": "  Explaining the behaviors of deep neural networks, usually considered as black\nboxes, is critical especially when they are now being adopted over diverse\naspects of human life. Taking the advantages of interpretable machine learning\n(interpretable ML), this work proposes a novel tool called Catastrophic\nForgetting Dissector (or CFD) to explain catastrophic forgetting in continual\nlearning settings. We also introduce a new method called Critical Freezing\nbased on the observations of our tool. Experiments on ResNet articulate how\ncatastrophic forgetting happens, particularly showing which components of this\nfamous network are forgetting. Our new continual learning algorithm defeats\nvarious recent techniques by a significant margin, proving the capability of\nthe investigation. Critical freezing not only attacks catastrophic forgetting\nbut also exposes explainability.\n", "date": "2022-11-25", "categories": []}
{"id": "2211.13050", "title": "Semi-Supervised Lifelong Language Learning", "abstract": "  Lifelong learning aims to accumulate knowledge and alleviate catastrophic\nforgetting when learning tasks sequentially. However, existing lifelong\nlanguage learning methods only focus on the supervised learning setting.\nUnlabeled data, which can be easily accessed in real-world scenarios, are\nunderexplored. In this paper, we explore a novel setting, semi-supervised\nlifelong language learning (SSLL), where a model learns sequentially arriving\nlanguage tasks with both labeled and unlabeled data. We propose an unlabeled\ndata enhanced lifelong learner to explore SSLL. Specially, we dedicate\ntask-specific modules to alleviate catastrophic forgetting and design two\nmodules to exploit unlabeled data: (1) a virtual supervision enhanced task\nsolver is constructed on a teacher-student framework to mine the underlying\nknowledge from unlabeled data; and (2) a backward augmented learner is built to\nencourage knowledge transfer from newly arrived unlabeled data to previous\ntasks. Experimental results on various language tasks demonstrate our model's\neffectiveness and superiority over competitive baselines under the new setting\nSSLL.\n", "date": "2022-11-23", "categories": []}
{"id": "2207.05615", "title": "Contrastive Learning for Online Semi-Supervised General Continual\n  Learning", "abstract": "  We study Online Continual Learning with missing labels and propose SemiCon, a\nnew contrastive loss designed for partly labeled data. We demonstrate its\nefficiency by devising a memory-based method trained on an unlabeled data\nstream, where every data added to memory is labeled using an oracle. Our\napproach outperforms existing semi-supervised methods when few labels are\navailable, and obtain similar results to state-of-the-art supervised methods\nwhile using only 2.6% of labels on Split-CIFAR10 and 10% of labels on\nSplit-CIFAR100.\n", "date": "2022-07-12", "categories": []}
{"id": "2211.10567", "title": "CL-CrossVQA: A Continual Learning Benchmark for Cross-Domain Visual\n  Question Answering", "abstract": "  Visual Question Answering (VQA) is a multi-discipline research task. To\nproduce the right answer, it requires an understanding of the visual content of\nimages, the natural language questions, as well as commonsense reasoning over\nthe information contained in the image and world knowledge. Recently,\nlarge-scale Vision-and-Language Pre-trained Models (VLPMs) have been the\nmainstream approach to VQA tasks due to their superior performance. The\nstandard practice is to fine-tune large-scale VLPMs pre-trained on huge\ngeneral-domain datasets using the domain-specific VQA datasets. However, in\nreality, the application domain can change over time, necessitating VLPMs to\ncontinually learn and adapt to new domains without forgetting previously\nacquired knowledge. Most existing continual learning (CL) research concentrates\non unimodal tasks, whereas a more practical application scenario, i.e, CL on\ncross-domain VQA, has not been studied. Motivated by this, we introduce\nCL-CrossVQA, a rigorous Continual Learning benchmark for Cross-domain Visual\nQuestion Answering, through which we conduct extensive experiments on 4 VLPMs,\n4 CL approaches, and 5 VQA datasets from different domains. In addition, by\nprobing the forgetting phenomenon of the intermediate layers, we provide\ninsights into how model architecture affects CL performance, why CL approaches\ncan help mitigate forgetting in VLPMs to some extent, and how to design CL\napproaches suitable for VLPMs in this challenging continual learning\nenvironment. To facilitate future work on CL for cross-domain VQA, we will\nrelease our datasets and code.\n", "date": "2022-11-19", "categories": []}
{"id": "2211.10253", "title": "Delving into Transformer for Incremental Semantic Segmentation", "abstract": "  Incremental semantic segmentation(ISS) is an emerging task where old model is\nupdated by incrementally adding new classes. At present, methods based on\nconvolutional neural networks are dominant in ISS. However, studies have shown\nthat such methods have difficulty in learning new tasks while maintaining good\nperformance on old ones (catastrophic forgetting). In contrast, a Transformer\nbased method has a natural advantage in curbing catastrophic forgetting due to\nits ability to model both long-term and short-term tasks. In this work, we\nexplore the reasons why Transformer based architecture are more suitable for\nISS, and accordingly propose propose TISS, a Transformer based method for\nIncremental Semantic Segmentation. In addition, to better alleviate\ncatastrophic forgetting while preserving transferability on ISS, we introduce\ntwo patch-wise contrastive losses to imitate similar features and enhance\nfeature diversity respectively, which can further improve the performance of\nTISS. Under extensive experimental settings with Pascal-VOC 2012 and ADE20K\ndatasets, our method significantly outperforms state-of-the-art incremental\nsemantic segmentation methods.\n", "date": "2022-11-18", "categories": []}
{"id": "2211.09032", "title": "CL2R: Compatible Lifelong Learning Representations", "abstract": "  In this paper, we propose a method to partially mimic natural intelligence\nfor the problem of lifelong learning representations that are compatible. We\ntake the perspective of a learning agent that is interested in recognizing\nobject instances in an open dynamic universe in a way in which any update to\nits internal feature representation does not render the features in the gallery\nunusable for visual search. We refer to this learning problem as Compatible\nLifelong Learning Representations (CL2R) as it considers compatible\nrepresentation learning within the lifelong learning paradigm. We identify\nstationarity as the property that the feature representation is required to\nhold to achieve compatibility and propose a novel training procedure that\nencourages local and global stationarity on the learned representation. Due to\nstationarity, the statistical properties of the learned features do not change\nover time, making them interoperable with previously learned features.\nExtensive experiments on standard benchmark datasets show that our CL2R\ntraining procedure outperforms alternative baselines and state-of-the-art\nmethods. We also provide novel metrics to specifically evaluate compatible\nrepresentation learning under catastrophic forgetting in various sequential\nlearning tasks. Code at\nhttps://github.com/NiccoBiondi/CompatibleLifelongRepresentation.\n", "date": "2022-11-16", "categories": []}
{"id": "2211.07725", "title": "Hierarchically Structured Task-Agnostic Continual Learning", "abstract": "  One notable weakness of current machine learning algorithms is the poor\nability of models to solve new problems without forgetting previously acquired\nknowledge. The Continual Learning paradigm has emerged as a protocol to\nsystematically investigate settings where the model sequentially observes\nsamples generated by a series of tasks. In this work, we take a task-agnostic\nview of continual learning and develop a hierarchical information-theoretic\noptimality principle that facilitates a trade-off between learning and\nforgetting. We derive this principle from a Bayesian perspective and show its\nconnections to previous approaches to continual learning. Based on this\nprinciple, we propose a neural network layer, called the\nMixture-of-Variational-Experts layer, that alleviates forgetting by creating a\nset of information processing paths through the network which is governed by a\ngating policy. Equipped with a diverse and specialized set of parameters, each\npath can be regarded as a distinct sub-network that learns to solve tasks. To\nimprove expert allocation, we introduce diversity objectives, which we evaluate\nin additional ablation studies. Importantly, our approach can operate in a\ntask-agnostic way, i.e., it does not require task-specific knowledge, as is the\ncase with many existing continual learning algorithms. Due to the general\nformulation based on generic utility functions, we can apply this optimality\nprinciple to a large variety of learning problems, including supervised\nlearning, reinforcement learning, and generative modeling. We demonstrate the\ncompetitive performance of our method on continual reinforcement learning and\nvariants of the MNIST, CIFAR-10, and CIFAR-100 datasets.\n", "date": "2022-11-14", "categories": []}
{"id": "2209.13917", "title": "A simple but strong baseline for online continual learning: Repeated\n  Augmented Rehearsal", "abstract": "  Online continual learning (OCL) aims to train neural networks incrementally\nfrom a non-stationary data stream with a single pass through data.\nRehearsal-based methods attempt to approximate the observed input distributions\nover time with a small memory and revisit them later to avoid forgetting.\nDespite its strong empirical performance, rehearsal methods still suffer from a\npoor approximation of the loss landscape of past data with memory samples. This\npaper revisits the rehearsal dynamics in online settings. We provide\ntheoretical insights on the inherent memory overfitting risk from the viewpoint\nof biased and dynamic empirical risk minimization, and examine the merits and\nlimits of repeated rehearsal. Inspired by our analysis, a simple and intuitive\nbaseline, Repeated Augmented Rehearsal (RAR), is designed to address the\nunderfitting-overfitting dilemma of online rehearsal. Surprisingly, across four\nrather different OCL benchmarks, this simple baseline outperforms vanilla\nrehearsal by 9%-17% and also significantly improves state-of-the-art\nrehearsal-based methods MIR, ASER, and SCR. We also demonstrate that RAR\nsuccessfully achieves an accurate approximation of the loss landscape of past\ndata and high-loss ridge aversion in its learning trajectory. Extensive\nablation studies are conducted to study the interplay between repeated and\naugmented rehearsal and reinforcement learning (RL) is applied to dynamically\nadjust the hyperparameters of RAR to balance the stability-plasticity trade-off\nonline. Code is available at\nhttps://github.com/YaqianZhang/RepeatedAugmentedRehearsal\n", "date": "2022-09-28", "categories": []}
{"id": "2209.10320", "title": "Continual VQA for Disaster Response Systems", "abstract": "  Visual Question Answering (VQA) is a multi-modal task that involves answering\nquestions from an input image, semantically understanding the contents of the\nimage and answering it in natural language. Using VQA for disaster management\nis an important line of research due to the scope of problems that are answered\nby the VQA system. However, the main challenge is the delay caused by the\ngeneration of labels in the assessment of the affected areas. To tackle this,\nwe deployed pre-trained CLIP model, which is trained on visual-image pairs.\nhowever, we empirically see that the model has poor zero-shot performance.\nThus, we instead use pre-trained embeddings of text and image from this model\nfor our supervised training and surpass previous state-of-the-art results on\nthe FloodNet dataset. We expand this to a continual setting, which is a more\nreal-life scenario. We tackle the problem of catastrophic forgetting using\nvarious experience replay methods. Our training runs are available at:\nhttps://wandb.ai/compyle/continual_vqa_final. Our code is available at\nhttps://github.com/AdityaKane2001/continual_vqa.\n", "date": "2022-09-21", "categories": []}
{"id": "2211.04624", "title": "Cold Start Streaming Learning for Deep Networks", "abstract": "  The ability to dynamically adapt neural networks to newly-available data\nwithout performance deterioration would revolutionize deep learning\napplications. Streaming learning (i.e., learning from one data example at a\ntime) has the potential to enable such real-time adaptation, but current\napproaches i) freeze a majority of network parameters during streaming and ii)\nare dependent upon offline, base initialization procedures over large subsets\nof data, which damages performance and limits applicability. To mitigate these\nshortcomings, we propose Cold Start Streaming Learning (CSSL), a simple,\nend-to-end approach for streaming learning with deep networks that uses a\ncombination of replay and data augmentation to avoid catastrophic forgetting.\n  Because CSSL updates all model parameters during streaming, the algorithm is\ncapable of beginning streaming from a random initialization, making base\ninitialization optional. Going further, the algorithm's simplicity allows\ntheoretical convergence guarantees to be derived using analysis of the Neural\nTangent Random Feature (NTRF). In experiments, we find that CSSL outperforms\nexisting baselines for streaming learning in experiments on CIFAR100, ImageNet,\nand Core50 datasets. Additionally, we propose a novel multi-task streaming\nlearning setting and show that CSSL performs favorably in this domain. Put\nsimply, CSSL performs well and demonstrates that the complicated, multi-step\ntraining pipelines adopted by most streaming methodologies can be replaced with\na simple, end-to-end learning approach without sacrificing performance.\n", "date": "2022-11-09", "categories": []}
{"id": "2211.04770", "title": "Continual learning autoencoder training for a particle-in-cell\n  simulation via streaming", "abstract": "  The upcoming exascale era will provide a new generation of physics\nsimulations. These simulations will have a high spatiotemporal resolution,\nwhich will impact the training of machine learning models since storing a high\namount of simulation data on disk is nearly impossible. Therefore, we need to\nrethink the training of machine learning models for simulations for the\nupcoming exascale era. This work presents an approach that trains a neural\nnetwork concurrently to a running simulation without storing data on a disk.\nThe training pipeline accesses the training data by in-memory streaming.\nFurthermore, we apply methods from the domain of continual learning to enhance\nthe generalization of the model. We tested our pipeline on the training of a 3d\nautoencoder trained concurrently to laser wakefield acceleration\nparticle-in-cell simulation. Furthermore, we experimented with various\ncontinual learning methods and their effect on the generalization.\n", "date": "2022-11-09", "categories": []}
{"id": "2211.03966", "title": "Parameter and Data Efficient Continual Pre-training for Robustness to\n  Dialectal Variance in Arabic", "abstract": "  The use of multilingual language models for tasks in low and high-resource\nlanguages has been a success story in deep learning. In recent times, Arabic\nhas been receiving widespread attention on account of its dialectal variance.\nWhile prior research studies have tried to adapt these multilingual models for\ndialectal variants of Arabic, it still remains a challenging problem owing to\nthe lack of sufficient monolingual dialectal data and parallel translation data\nof such dialectal variants. It remains an open problem on whether the limited\ndialectical data can be used to improve the models trained in Arabic on its\ndialectal variants. First, we show that multilingual-BERT (mBERT) incrementally\npretrained on Arabic monolingual data takes less training time and yields\ncomparable accuracy when compared to our custom monolingual Arabic model and\nbeat existing models (by an avg metric of +$6.41$). We then explore two\ncontinual pre-training methods -- (1) using small amounts of dialectical data\nfor continual finetuning and (2) parallel Arabic to English data and a\nTranslation Language Modeling loss function. We show that both approaches help\nimprove performance on dialectal classification tasks ($+4.64$ avg. gain) when\nused on monolingual models.\n", "date": "2022-11-08", "categories": []}
{"id": "2209.07326", "title": "A Continual Development Methodology for Large-scale Multitask Dynamic ML\n  Systems", "abstract": "  The traditional Machine Learning (ML) methodology requires to fragment the\ndevelopment and experimental process into disconnected iterations whose\nfeedback is used to guide design or tuning choices. This methodology has\nmultiple efficiency and scalability disadvantages, such as leading to spend\nsignificant resources into the creation of multiple trial models that do not\ncontribute to the final solution.The presented work is based on the intuition\nthat defining ML models as modular and extensible artefacts allows to introduce\na novel ML development methodology enabling the integration of multiple design\nand evaluation iterations into the continuous enrichment of a single unbounded\nintelligent system. We define a novel method for the generation of dynamic\nmultitask ML models as a sequence of extensions and generalizations. We first\nanalyze the capabilities of the proposed method by using the standard ML\nempirical evaluation methodology. Finally, we propose a novel continuous\ndevelopment methodology that allows to dynamically extend a pre-existing\nmultitask large-scale ML system while analyzing the properties of the proposed\nmethod extensions. This results in the generation of an ML model capable of\njointly solving 124 image classification tasks achieving state of the art\nquality with improved size and compute cost.\n", "date": "2022-09-15", "categories": []}
{"id": "2211.03186", "title": "Momentum-based Weight Interpolation of Strong Zero-Shot Models for\n  Continual Learning", "abstract": "  Large pre-trained, zero-shot capable models have shown considerable success\nboth for standard transfer and adaptation tasks, with particular robustness\ntowards distribution shifts. In addition, subsequent fine-tuning can\nconsiderably improve performance on a selected downstream task. However,\nthrough naive fine-tuning, these zero-shot models lose their generalizability\nand robustness towards distribution shifts. This is a particular problem for\ntasks such as Continual Learning (CL), where continuous adaptation has to be\nperformed as new task distributions are introduced sequentially. In this work,\nwe showcase that where fine-tuning falls short to adapt such zero-shot capable\nmodels, simple momentum-based weight interpolation can provide consistent\nimprovements for CL tasks in both memory-free and memory-based settings. In\nparticular, we find improvements of over $+4\\%$ on standard CL benchmarks,\nwhile reducing the error to the upper limit of jointly training on all tasks at\nonce in parts by more than half, allowing the continual learner to inch closer\nto the joint training limits.\n", "date": "2022-11-06", "categories": []}
{"id": "2211.01542", "title": "Continual Learning of Neural Machine Translation within Low Forgetting\n  Risk Regions", "abstract": "  This paper considers continual learning of large-scale pretrained neural\nmachine translation model without accessing the previous training data or\nintroducing model separation. We argue that the widely used\nregularization-based methods, which perform multi-objective learning with an\nauxiliary loss, suffer from the misestimate problem and cannot always achieve a\ngood balance between the previous and new tasks. To solve the problem, we\npropose a two-stage training method based on the local features of the real\nloss. We first search low forgetting risk regions, where the model can retain\nthe performance on the previous task as the parameters are updated, to avoid\nthe catastrophic forgetting problem. Then we can continually train the model\nwithin this region only with the new training data to fit the new task.\nSpecifically, we propose two methods to search the low forgetting risk regions,\nwhich are based on the curvature of loss and the impacts of the parameters on\nthe model output, respectively. We conduct experiments on domain adaptation and\nmore challenging language adaptation tasks, and the experimental results show\nthat our method can achieve significant improvements compared with several\nstrong baselines.\n", "date": "2022-11-03", "categories": []}
{"id": "2211.02633", "title": "A Theoretical Study on Solving Continual Learning", "abstract": "  Continual learning (CL) learns a sequence of tasks incrementally. There are\ntwo popular CL settings, class incremental learning (CIL) and task incremental\nlearning (TIL). A major challenge of CL is catastrophic forgetting (CF). While\na number of techniques are already available to effectively overcome CF for\nTIL, CIL remains to be highly challenging. So far, little theoretical study has\nbeen done to provide a principled guidance on how to solve the CIL problem.\nThis paper performs such a study. It first shows that probabilistically, the\nCIL problem can be decomposed into two sub-problems: Within-task Prediction\n(WP) and Task-id Prediction (TP). It further proves that TP is correlated with\nout-of-distribution (OOD) detection, which connects CIL and OOD detection. The\nkey conclusion of this study is that regardless of whether WP and TP or OOD\ndetection are defined explicitly or implicitly by a CIL algorithm, good WP and\ngood TP or OOD detection are necessary and sufficient for good CIL\nperformances. Additionally, TIL is simply WP. Based on the theoretical result,\nnew CIL methods are also designed, which outperform strong baselines in both\nCIL and TIL settings by a large margin.\n", "date": "2022-11-04", "categories": []}
{"id": "2211.00789", "title": "Beyond Not-Forgetting: Continual Learning with Backward Knowledge\n  Transfer", "abstract": "  By learning a sequence of tasks continually, an agent in continual learning\n(CL) can improve the learning performance of both a new task and `old' tasks by\nleveraging the forward knowledge transfer and the backward knowledge transfer,\nrespectively. However, most existing CL methods focus on addressing\ncatastrophic forgetting in neural networks by minimizing the modification of\nthe learnt model for old tasks. This inevitably limits the backward knowledge\ntransfer from the new task to the old tasks, because judicious model updates\ncould possibly improve the learning performance of the old tasks as well. To\ntackle this problem, we first theoretically analyze the conditions under which\nupdating the learnt model of old tasks could be beneficial for CL and also lead\nto backward knowledge transfer, based on the gradient projection onto the input\nsubspaces of old tasks. Building on the theoretical analysis, we next develop a\nContinUal learning method with Backward knowlEdge tRansfer (CUBER), for a fixed\ncapacity neural network without data replay. In particular, CUBER first\ncharacterizes the task correlation to identify the positively correlated old\ntasks in a layer-wise manner, and then selectively modifies the learnt model of\nthe old tasks when learning the new task. Experimental studies show that CUBER\ncan even achieve positive backward knowledge transfer on several existing CL\nbenchmarks for the first time without data replay, where the related baselines\nstill suffer from catastrophic forgetting (negative backward knowledge\ntransfer). The superior performance of CUBER on the backward knowledge transfer\nalso leads to higher accuracy accordingly.\n", "date": "2022-11-01", "categories": []}
{"id": "2211.00833", "title": "Learning a Condensed Frame for Memory-Efficient Video Class-Incremental\n  Learning", "abstract": "  Recent incremental learning for action recognition usually stores\nrepresentative videos to mitigate catastrophic forgetting. However, only a few\nbulky videos can be stored due to the limited memory. To address this problem,\nwe propose FrameMaker, a memory-efficient video class-incremental learning\napproach that learns to produce a condensed frame for each selected video.\nSpecifically, FrameMaker is mainly composed of two crucial components: Frame\nCondensing and Instance-Specific Prompt. The former is to reduce the memory\ncost by preserving only one condensed frame instead of the whole video, while\nthe latter aims to compensate the lost spatio-temporal details in the Frame\nCondensing stage. By this means, FrameMaker enables a remarkable reduction in\nmemory but keep enough information that can be applied to following incremental\ntasks. Experimental results on multiple challenging benchmarks, i.e., HMDB51,\nUCF101 and Something-Something V2, demonstrate that FrameMaker can achieve\nbetter performance to recent advanced methods while consuming only 20% memory.\nAdditionally, under the same memory consumption conditions, FrameMaker\nsignificantly outperforms existing state-of-the-arts by a convincing margin.\n", "date": "2022-11-02", "categories": []}
{"id": "2208.12967", "title": "Anti-Retroactive Interference for Lifelong Learning", "abstract": "  Humans can continuously learn new knowledge. However, machine learning models\nsuffer from drastic dropping in performance on previous tasks after learning\nnew tasks. Cognitive science points out that the competition of similar\nknowledge is an important cause of forgetting. In this paper, we design a\nparadigm for lifelong learning based on meta-learning and associative mechanism\nof the brain. It tackles the problem from two aspects: extracting knowledge and\nmemorizing knowledge. First, we disrupt the sample's background distribution\nthrough a background attack, which strengthens the model to extract the key\nfeatures of each task. Second, according to the similarity between incremental\nknowledge and base knowledge, we design an adaptive fusion of incremental\nknowledge, which helps the model allocate capacity to the knowledge of\ndifferent difficulties. It is theoretically analyzed that the proposed learning\nparadigm can make the models of different tasks converge to the same optimum.\nThe proposed method is validated on the MNIST, CIFAR100, CUB200 and ImageNet100\ndatasets.\n", "date": "2022-08-27", "categories": []}
{"id": "2210.17322", "title": "Generative Negative Text Replay for Continual Vision-Language\n  Pretraining", "abstract": "  Vision-language pre-training (VLP) has attracted increasing attention\nrecently. With a large amount of image-text pairs, VLP models trained with\ncontrastive loss have achieved impressive performance in various tasks,\nespecially the zero-shot generalization on downstream datasets. In practical\napplications, however, massive data are usually collected in a streaming\nfashion, requiring VLP models to continuously integrate novel knowledge from\nincoming data and retain learned knowledge. In this work, we focus on learning\na VLP model with sequential chunks of image-text pair data. To tackle the\ncatastrophic forgetting issue in this multi-modal continual learning setting,\nwe first introduce pseudo text replay that generates hard negative texts\nconditioned on the training images in memory, which not only better preserves\nlearned knowledge but also improves the diversity of negative samples in the\ncontrastive loss. Moreover, we propose multi-modal knowledge distillation\nbetween images and texts to align the instance-wise prediction between old and\nnew models. We incrementally pre-train our model on both the instance and class\nincremental splits of the Conceptual Caption dataset, and evaluate the model on\nzero-shot image classification and image-text retrieval tasks. Our method\nconsistently outperforms the existing baselines with a large margin, which\ndemonstrates its superiority. Notably, we realize an average performance boost\nof $4.60\\%$ on image-classification downstream datasets for the class\nincremental split.\n", "date": "2022-10-31", "categories": []}
{"id": "2210.15051", "title": "Federated Continual Learning to Detect Accounting Anomalies in Financial\n  Auditing", "abstract": "  The International Standards on Auditing require auditors to collect\nreasonable assurance that financial statements are free of material\nmisstatement. At the same time, a central objective of Continuous Assurance is\nthe real-time assessment of digital accounting journal entries. Recently,\ndriven by the advances in artificial intelligence, Deep Learning techniques\nhave emerged in financial auditing to examine vast quantities of accounting\ndata. However, learning highly adaptive audit models in decentralised and\ndynamic settings remains challenging. It requires the study of data\ndistribution shifts over multiple clients and time periods. In this work, we\npropose a Federated Continual Learning framework enabling auditors to learn\naudit models from decentral clients continuously. We evaluate the framework's\nability to detect accounting anomalies in common scenarios of organizational\nactivity. Our empirical results, using real-world datasets and combined\nfederated continual learning strategies, demonstrate the learned model's\nability to detect anomalies in audit settings of data distribution shifts.\n", "date": "2022-10-26", "categories": []}
{"id": "2210.14307", "title": "On Robust Incremental Learning over Many Multilingual Steps", "abstract": "  Recent work in incremental learning has introduced diverse approaches to\ntackle catastrophic forgetting from data augmentation to optimized training\nregimes. However, most of them focus on very few training steps. We propose a\nmethod for robust incremental learning over dozens of fine-tuning steps using\ndata from a variety of languages. We show that a combination of\ndata-augmentation and an optimized training regime allows us to continue\nimproving the model even for as many as fifty training steps. Crucially, our\naugmentation strategy does not require retaining access to previous training\ndata and is suitable in scenarios with privacy constraints.\n", "date": "2022-10-25", "categories": []}
{"id": "2210.14797", "title": "Is Multi-Task Learning an Upper Bound for Continual Learning?", "abstract": "  Continual and multi-task learning are common machine learning approaches to\nlearning from multiple tasks. The existing works in the literature often assume\nmulti-task learning as a sensible performance upper bound for various continual\nlearning algorithms. While this assumption is empirically verified for\ndifferent continual learning benchmarks, it is not rigorously justified.\nMoreover, it is imaginable that when learning from multiple tasks, a small\nsubset of these tasks could behave as adversarial tasks reducing the overall\nlearning performance in a multi-task setting. In contrast, continual learning\napproaches can avoid the performance drop caused by such adversarial tasks to\npreserve their performance on the rest of the tasks, leading to better\nperformance than a multi-task learner. This paper proposes a novel continual\nself-supervised learning setting, where each task corresponds to learning an\ninvariant representation for a specific class of data augmentations. In this\nsetting, we show that continual learning often beats multi-task learning on\nvarious benchmark datasets, including MNIST, CIFAR-10, and CIFAR-100.\n", "date": "2022-10-26", "categories": []}
{"id": "2210.06443", "title": "On the Effectiveness of Lipschitz-Driven Rehearsal in Continual Learning", "abstract": "  Rehearsal approaches enjoy immense popularity with Continual Learning (CL)\npractitioners. These methods collect samples from previously encountered data\ndistributions in a small memory buffer; subsequently, they repeatedly optimize\non the latter to prevent catastrophic forgetting. This work draws attention to\na hidden pitfall of this widespread practice: repeated optimization on a small\npool of data inevitably leads to tight and unstable decision boundaries, which\nare a major hindrance to generalization. To address this issue, we propose\nLipschitz-DrivEn Rehearsal (LiDER), a surrogate objective that induces\nsmoothness in the backbone network by constraining its layer-wise Lipschitz\nconstants w.r.t. replay examples. By means of extensive experiments, we show\nthat applying LiDER delivers a stable performance gain to several\nstate-of-the-art rehearsal CL methods across multiple datasets, both in the\npresence and absence of pre-training. Through additional ablative experiments,\nwe highlight peculiar aspects of buffer overfitting in CL and better\ncharacterize the effect produced by LiDER. Code is available at\nhttps://github.com/aimagelab/LiDER\n", "date": "2022-10-12", "categories": []}
{"id": "2210.08442", "title": "Navigating Memory Construction by Global Pseudo-Task Simulation for\n  Continual Learning", "abstract": "  Continual learning faces a crucial challenge of catastrophic forgetting. To\naddress this challenge, experience replay (ER) that maintains a tiny subset of\nsamples from previous tasks has been commonly used. Existing ER works usually\nfocus on refining the learning objective for each task with a static memory\nconstruction policy. In this paper, we formulate the dynamic memory\nconstruction in ER as a combinatorial optimization problem, which aims at\ndirectly minimizing the global loss across all experienced tasks. We first\napply three tactics to solve the problem in the offline setting as a starting\npoint. To provide an approximate solution to this problem in the online\ncontinual learning setting, we further propose the Global Pseudo-task\nSimulation (GPS), which mimics future catastrophic forgetting of the current\ntask by permutation. Our empirical results and analyses suggest that the GPS\nconsistently improves accuracy across four commonly used vision benchmarks. We\nhave also shown that our GPS can serve as the unified framework for integrating\nvarious memory construction policies in existing ER works.\n", "date": "2022-10-16", "categories": []}
{"id": "2210.05561", "title": "Schedule-Robust Online Continual Learning", "abstract": "  A continual learning (CL) algorithm learns from a non-stationary data stream.\nThe non-stationarity is modeled by some schedule that determines how data is\npresented over time. Most current methods make strong assumptions on the\nschedule and have unpredictable performance when such requirements are not met.\nA key challenge in CL is thus to design methods robust against arbitrary\nschedules over the same underlying data, since in real-world scenarios\nschedules are often unknown and dynamic. In this work, we introduce the notion\nof schedule-robustness for CL and a novel approach satisfying this desirable\nproperty in the challenging online class-incremental setting. We also present a\nnew perspective on CL, as the process of learning a schedule-robust predictor,\nfollowed by adapting the predictor using only replay data. Empirically, we\ndemonstrate that our approach outperforms existing methods on CL benchmarks for\nimage classification by a large margin.\n", "date": "2022-10-11", "categories": []}
{"id": "2210.04137", "title": "Few-Shot Continual Active Learning by a Robot", "abstract": "  In this paper, we consider a challenging but realistic continual learning\n(CL) problem, Few-Shot Continual Active Learning (FoCAL), where a CL agent is\nprovided with unlabeled data for a new or a previously learned task in each\nincrement and the agent only has limited labeling budget available. Towards\nthis, we build on the continual learning and active learning literature and\ndevelop a framework that can allow a CL agent to continually learn new object\nclasses from a few labeled training examples. Our framework represents each\nobject class using a uniform Gaussian mixture model (GMM) and uses\npseudo-rehearsal to mitigate catastrophic forgetting. The framework also uses\nuncertainty measures on the Gaussian representations of the previously learned\nclasses to find the most informative samples to be labeled in an increment. We\nevaluate our approach on the CORe-50 dataset and on a real humanoid robot for\nthe object classification task. The results show that our approach not only\nproduces state-of-the-art results on the dataset but also allows a real robot\nto continually learn unseen objects in a real environment with limited labeling\nsupervision provided by its user.\n", "date": "2022-10-09", "categories": []}
{"id": "2210.06579", "title": "Task-Free Continual Learning via Online Discrepancy Distance Learning", "abstract": "  Learning from non-stationary data streams, also called Task-Free Continual\nLearning (TFCL) remains challenging due to the absence of explicit task\ninformation. Although recently some methods have been proposed for TFCL, they\nlack theoretical guarantees. Moreover, forgetting analysis during TFCL was not\nstudied theoretically before. This paper develops a new theoretical analysis\nframework which provides generalization bounds based on the discrepancy\ndistance between the visited samples and the entire information made available\nfor training the model. This analysis gives new insights into the forgetting\nbehaviour in classification tasks. Inspired by this theoretical model, we\npropose a new approach enabled by the dynamic component expansion mechanism for\na mixture model, namely the Online Discrepancy Distance Learning (ODDL). ODDL\nestimates the discrepancy between the probabilistic representation of the\ncurrent memory buffer and the already accumulated knowledge and uses it as the\nexpansion signal to ensure a compact network architecture with optimal\nperformance. We then propose a new sample selection approach that selectively\nstores the most relevant samples into the memory buffer through the\ndiscrepancy-based measure, further improving the performance. We perform\nseveral TFCL experiments with the proposed methodology, which demonstrate that\nthe proposed approach achieves the state of the art performance.\n", "date": "2022-10-12", "categories": []}
{"id": "2210.06816", "title": "ALIFE: Adaptive Logit Regularizer and Feature Replay for Incremental\n  Semantic Segmentation", "abstract": "  We address the problem of incremental semantic segmentation (ISS) recognizing\nnovel object/stuff categories continually without forgetting previous ones that\nhave been learned. The catastrophic forgetting problem is particularly severe\nin ISS, since pixel-level ground-truth labels are available only for the novel\ncategories at training time. To address the problem, regularization-based\nmethods exploit probability calibration techniques to learn semantic\ninformation from unlabeled pixels. While such techniques are effective, there\nis still a lack of theoretical understanding of them. Replay-based methods\npropose to memorize a small set of images for previous categories. They achieve\nstate-of-the-art performance at the cost of large memory footprint. We propose\nin this paper a novel ISS method, dubbed ALIFE, that provides a better\ncompromise between accuracy and efficiency. To this end, we first show an\nin-depth analysis on the calibration techniques to better understand the\neffects on ISS. Based on this, we then introduce an adaptive logit regularizer\n(ALI) that enables our model to better learn new categories, while retaining\nknowledge for previous ones. We also present a feature replay scheme that\nmemorizes features, instead of images directly, in order to reduce memory\nrequirements significantly. Since a feature extractor is changed continually,\nmemorized features should also be updated at every incremental stage. To handle\nthis, we introduce category-specific rotation matrices updating the features\nfor each category separately. We demonstrate the effectiveness of our approach\nwith extensive experiments on standard ISS benchmarks, and show that our method\nachieves a better trade-off in terms of accuracy and efficiency.\n", "date": "2022-10-13", "categories": []}
{"id": "2210.07016", "title": "Learning with Style: Continual Semantic Segmentation Across Tasks and\n  Domains", "abstract": "  Deep learning models dealing with image understanding in real-world settings\nmust be able to adapt to a wide variety of tasks across different domains.\nDomain adaptation and class incremental learning deal with domain and task\nvariability separately, whereas their unified solution is still an open\nproblem. We tackle both facets of the problem together, taking into account the\nsemantic shift within both input and label spaces. We start by formally\nintroducing continual learning under task and domain shift. Then, we address\nthe proposed setup by using style transfer techniques to extend knowledge\nacross domains when learning incremental tasks and a robust distillation\nframework to effectively recollect task knowledge under incremental domain\nshift. The devised framework (LwS, Learning with Style) is able to generalize\nincrementally acquired task knowledge across all the domains encountered,\nproving to be robust against catastrophic forgetting. Extensive experimental\nevaluation on multiple autonomous driving datasets shows how the proposed\nmethod outperforms existing approaches, which prove to be ill-equipped to deal\nwith continual semantic segmentation under both task and domain shift.\n", "date": "2022-10-13", "categories": []}
{"id": "2210.07207", "title": "Attribution-aware Weight Transfer: A Warm-Start Initialization for\n  Class-Incremental Semantic Segmentation", "abstract": "  In class-incremental semantic segmentation (CISS), deep learning\narchitectures suffer from the critical problems of catastrophic forgetting and\nsemantic background shift. Although recent works focused on these issues,\nexisting classifier initialization methods do not address the background shift\nproblem and assign the same initialization weights to both background and new\nforeground class classifiers. We propose to address the background shift with a\nnovel classifier initialization method which employs gradient-based attribution\nto identify the most relevant weights for new classes from the classifier's\nweights for the previous background and transfers these weights to the new\nclassifier. This warm-start weight initialization provides a general solution\napplicable to several CISS methods. Furthermore, it accelerates learning of new\nclasses while mitigating forgetting. Our experiments demonstrate significant\nimprovement in mIoU compared to the state-of-the-art CISS methods on the\nPascal-VOC 2012, ADE20K and Cityscapes datasets.\n", "date": "2022-10-13", "categories": []}
{"id": "2207.12964", "title": "Incremental Few-Shot Semantic Segmentation via Embedding Adaptive-Update\n  and Hyper-class Representation", "abstract": "  Incremental few-shot semantic segmentation (IFSS) targets at incrementally\nexpanding model's capacity to segment new class of images supervised by only a\nfew samples. However, features learned on old classes could significantly\ndrift, causing catastrophic forgetting. Moreover, few samples for pixel-level\nsegmentation on new classes lead to notorious overfitting issues in each\nlearning session. In this paper, we explicitly represent class-based knowledge\nfor semantic segmentation as a category embedding and a hyper-class embedding,\nwhere the former describes exclusive semantical properties, and the latter\nexpresses hyper-class knowledge as class-shared semantic properties. Aiming to\nsolve IFSS problems, we present EHNet, i.e., Embedding adaptive-update and\nHyper-class representation Network from two aspects. First, we propose an\nembedding adaptive-update strategy to avoid feature drift, which maintains old\nknowledge by hyper-class representation, and adaptively update category\nembeddings with a class-attention scheme to involve new classes learned in\nindividual sessions. Second, to resist overfitting issues caused by few\ntraining samples, a hyper-class embedding is learned by clustering all category\nembeddings for initialization and aligned with category embedding of the new\nclass for enhancement, where learned knowledge assists to learn new knowledge,\nthus alleviating performance dependence on training data scale. Significantly,\nthese two designs provide representation capability for classes with sufficient\nsemantics and limited biases, enabling to perform segmentation tasks requiring\nhigh semantic dependence. Experiments on PASCAL-5i and COCO datasets show that\nEHNet achieves new state-of-the-art performance with remarkable advantages.\n", "date": "2022-07-26", "categories": []}
{"id": "2210.05751", "title": "Toward Sustainable Continual Learning: Detection and Knowledge\n  Repurposing of Similar Tasks", "abstract": "  Most existing works on continual learning (CL) focus on overcoming the\ncatastrophic forgetting (CF) problem, with dynamic models and replay methods\nperforming exceptionally well. However, since current works tend to assume\nexclusivity or dissimilarity among learning tasks, these methods require\nconstantly accumulating task-specific knowledge in memory for each task. This\nresults in the eventual prohibitive expansion of the knowledge repository if we\nconsider learning from a long sequence of tasks. In this work, we introduce a\nparadigm where the continual learner gets a sequence of mixed similar and\ndissimilar tasks. We propose a new continual learning framework that uses a\ntask similarity detection function that does not require additional learning,\nwith which we analyze whether there is a specific task in the past that is\nsimilar to the current task. We can then reuse previous task knowledge to slow\ndown parameter expansion, ensuring that the CL system expands the knowledge\nrepository sublinearly to the number of learned tasks. Our experiments show\nthat the proposed framework performs competitively on widely used computer\nvision benchmarks such as CIFAR10, CIFAR100, and EMNIST.\n", "date": "2022-10-11", "categories": []}
{"id": "2210.06401", "title": "Improving information retention in large scale online continual learning", "abstract": "  Given a stream of data sampled from non-stationary distributions, online\ncontinual learning (OCL) aims to adapt efficiently to new data while retaining\nexisting knowledge. The typical approach to address information retention (the\nability to retain previous knowledge) is keeping a replay buffer of a fixed\nsize and computing gradients using a mixture of new data and the replay buffer.\nSurprisingly, the recent work (Cai et al., 2021) suggests that information\nretention remains a problem in large scale OCL even when the replay buffer is\nunlimited, i.e., the gradients are computed using all past data. This paper\nfocuses on this peculiarity to understand and address information retention. To\npinpoint the source of this problem, we theoretically show that, given limited\ncomputation budgets at each time step, even without strict storage limit,\nnaively applying SGD with constant or constantly decreasing learning rates\nfails to optimize information retention in the long term. We propose using a\nmoving average family of methods to improve optimization for non-stationary\nobjectives. Specifically, we design an adaptive moving average (AMA) optimizer\nand a moving-average-based learning rate schedule (MALR). We demonstrate the\neffectiveness of AMA+MALR on large-scale benchmarks, including Continual\nLocalization (CLOC), Google Landmarks, and ImageNet. Code will be released upon\npublication.\n", "date": "2022-10-12", "categories": []}
{"id": "2210.05549", "title": "Continual Training of Language Models for Few-Shot Learning", "abstract": "  Recent work on applying large language models (LMs) achieves impressive\nperformance in many NLP applications. Adapting or posttraining an LM using an\nunlabeled domain corpus can produce even better performance for end-tasks in\nthe domain. This paper proposes the problem of continually extending an LM by\nincrementally post-train the LM with a sequence of unlabeled domain corpora to\nexpand its knowledge without forgetting its previous skills. The goal is to\nimprove the few-shot end-task learning in these domains. The resulting system\nis called CPT (Continual PostTraining), which to our knowledge, is the first\ncontinual post-training system. Experimental results verify its effectiveness.\n", "date": "2022-10-11", "categories": []}
{"id": "2210.03980", "title": "Distilling Causal Effect from Miscellaneous Other-Class for Continual\n  Named Entity Recognition", "abstract": "  Continual Learning for Named Entity Recognition (CL-NER) aims to learn a\ngrowing number of entity types over time from a stream of data. However, simply\nlearning Other-Class in the same way as new entity types amplifies the\ncatastrophic forgetting and leads to a substantial performance drop. The main\ncause behind this is that Other-Class samples usually contain old entity types,\nand the old knowledge in these Other-Class samples is not preserved properly.\nThanks to the causal inference, we identify that the forgetting is caused by\nthe missing causal effect from the old data. To this end, we propose a unified\ncausal framework to retrieve the causality from both new entity types and\nOther-Class. Furthermore, we apply curriculum learning to mitigate the impact\nof label noise and introduce a self-adaptive weight for balancing the causal\neffects between new entity types and Other-Class. Experimental results on three\nbenchmark datasets show that our method outperforms the state-of-the-art method\nby a large margin. Moreover, our method can be combined with the existing\nstate-of-the-art methods to improve the performance in CL-NER\n", "date": "2022-10-08", "categories": []}
{"id": "2210.04485", "title": "A Memory Transformer Network for Incremental Learning", "abstract": "  We study class-incremental learning, a training setup in which new classes of\ndata are observed over time for the model to learn from. Despite the\nstraightforward problem formulation, the naive application of classification\nmodels to class-incremental learning results in the \"catastrophic forgetting\"\nof previously seen classes. One of the most successful existing methods has\nbeen the use of a memory of exemplars, which overcomes the issue of\ncatastrophic forgetting by saving a subset of past data into a memory bank and\nutilizing it to prevent forgetting when training future tasks. In our paper, we\npropose to enhance the utilization of this memory bank: we not only use it as a\nsource of additional training data like existing works but also integrate it in\nthe prediction process explicitly.Our method, the Memory Transformer Network\n(MTN), learns how to combine and aggregate the information from the nearest\nneighbors in the memory with a transformer to make more accurate predictions.\nWe conduct extensive experiments and ablations to evaluate our approach. We\nshow that MTN achieves state-of-the-art performance on the challenging\nImageNet-1k and Google-Landmarks-1k incremental learning benchmarks.\n", "date": "2022-10-10", "categories": []}
{"id": "2210.04497", "title": "Learning Robust Representations for Continual Relation Extraction via\n  Adversarial Class Augmentation", "abstract": "  Continual relation extraction (CRE) aims to continually learn new relations\nfrom a class-incremental data stream. CRE model usually suffers from\ncatastrophic forgetting problem, i.e., the performance of old relations\nseriously degrades when the model learns new relations. Most previous work\nattributes catastrophic forgetting to the corruption of the learned\nrepresentations as new relations come, with an implicit assumption that the CRE\nmodels have adequately learned the old relations. In this paper, through\nempirical studies we argue that this assumption may not hold, and an important\nreason for catastrophic forgetting is that the learned representations do not\nhave good robustness against the appearance of analogous relations in the\nsubsequent learning process. To address this issue, we encourage the model to\nlearn more precise and robust representations through a simple yet effective\nadversarial class augmentation mechanism (ACA), which is easy to implement and\nmodel-agnostic. Experimental results show that ACA can consistently improve the\nperformance of state-of-the-art CRE models on two popular benchmarks.\n", "date": "2022-10-10", "categories": []}
{"id": "2210.04524", "title": "Margin-Based Few-Shot Class-Incremental Learning with Class-Level\n  Overfitting Mitigation", "abstract": "  Few-shot class-incremental learning (FSCIL) is designed to incrementally\nrecognize novel classes with only few training samples after the (pre-)training\non base classes with sufficient samples, which focuses on both base-class\nperformance and novel-class generalization. A well known modification to the\nbase-class training is to apply a margin to the base-class classification.\nHowever, a dilemma exists that we can hardly achieve both good base-class\nperformance and novel-class generalization simultaneously by applying the\nmargin during the base-class training, which is still under explored. In this\npaper, we study the cause of such dilemma for FSCIL. We first interpret this\ndilemma as a class-level overfitting (CO) problem from the aspect of pattern\nlearning, and then find its cause lies in the easily-satisfied constraint of\nlearning margin-based patterns. Based on the analysis, we propose a novel\nmargin-based FSCIL method to mitigate the CO problem by providing the pattern\nlearning process with extra constraint from the margin-based patterns\nthemselves. Extensive experiments on CIFAR100, Caltech-USCD Birds-200-2011\n(CUB200), and miniImageNet demonstrate that the proposed method effectively\nmitigates the CO problem and achieves state-of-the-art performance.\n", "date": "2022-10-10", "categories": []}
{"id": "2210.04865", "title": "Tracking changes using Kullback-Leibler divergence for the continual\n  learning", "abstract": "  Recently, continual learning has received a lot of attention. One of the\nsignificant problems is the occurrence of \\emph{concept drift}, which consists\nof changing probabilistic characteristics of the incoming data. In the case of\nthe classification task, this phenomenon destabilizes the model's performance\nand negatively affects the achieved prediction quality. Most current methods\napply statistical learning and similarity analysis over the raw data. However,\nsimilarity analysis in streaming data remains a complex problem due to time\nlimitation, non-precise values, fast decision speed, scalability, etc. This\narticle introduces a novel method for monitoring changes in the probabilistic\ndistribution of multi-dimensional data streams. As a measure of the rapidity of\nchanges, we analyze the popular Kullback-Leibler divergence. During the\nexperimental study, we show how to use this metric to predict the concept drift\noccurrence and understand its nature. The obtained results encourage further\nwork on the proposed methods and its application in the real tasks where the\nprediction of the future appearance of concept drift plays a crucial role, such\nas predictive maintenance.\n", "date": "2022-10-10", "categories": []}
{"id": "2210.03482", "title": "CLAD: A realistic Continual Learning benchmark for Autonomous Driving", "abstract": "  In this paper we describe the design and the ideas motivating a new Continual\nLearning benchmark for Autonomous Driving (CLAD), that focuses on the problems\nof object classification and object detection. The benchmark utilises SODA10M,\na recently released large-scale dataset that concerns autonomous driving\nrelated problems. First, we review and discuss existing continual learning\nbenchmarks, how they are related, and show that most are extreme cases of\ncontinual learning. To this end, we survey the benchmarks used in continual\nlearning papers at three highly ranked computer vision conferences. Next, we\nintroduce CLAD-C, an online classification benchmark realised through a\nchronological data stream that poses both class and domain incremental\nchallenges; and CLAD-D, a domain incremental continual object detection\nbenchmark. We examine the inherent difficulties and challenges posed by the\nbenchmark, through a survey of the techniques and methods used by the top-3\nparticipants in a CLAD-challenge workshop at ICCV 2021. We conclude with\npossible pathways to improve the current continual learning state of the art,\nand which directions we deem promising for future research.\n", "date": "2022-10-07", "categories": []}
{"id": "2210.02661", "title": "Topological Continual Learning with Wasserstein Distance and Barycenter", "abstract": "  Continual learning in neural networks suffers from a phenomenon called\ncatastrophic forgetting, in which a network quickly forgets what was learned in\na previous task. The human brain, however, is able to continually learn new\ntasks and accumulate knowledge throughout life. Neuroscience findings suggest\nthat continual learning success in the human brain is potentially associated\nwith its modular structure and memory consolidation mechanisms. In this paper\nwe propose a novel topological regularization that penalizes cycle structure in\na neural network during training using principled theory from persistent\nhomology and optimal transport. The penalty encourages the network to learn\nmodular structure during training. The penalization is based on the closed-form\nexpressions of the Wasserstein distance and barycenter for the topological\nfeatures of a 1-skeleton representation for the network. Our topological\ncontinual learning method combines the proposed regularization with a tiny\nepisodic memory to mitigate forgetting. We demonstrate that our method is\neffective in both shallow and deep network architectures for multiple image\nclassification datasets.\n", "date": "2022-10-06", "categories": []}
{"id": "2210.03114", "title": "CLIP model is an Efficient Continual Learner", "abstract": "  The continual learning setting aims to learn new tasks over time without\nforgetting the previous ones. The literature reports several significant\nefforts to tackle this problem with limited or no access to previous task data.\nAmong such efforts, typical solutions offer sophisticated techniques involving\nmemory replay, knowledge distillation, model regularization, and dynamic\nnetwork expansion. The resulting methods have a retraining cost at each\nlearning task, dedicated memory requirements, and setting-specific design\nchoices. In this work, we show that a frozen CLIP (Contrastive Language-Image\nPretraining) model offers astounding continual learning performance without any\nfine-tuning (zero-shot evaluation). We evaluate CLIP under a variety of\nsettings including class-incremental, domain-incremental and task-agnostic\nincremental learning on five popular benchmarks (ImageNet-100 & 1K, CORe50,\nCIFAR-100, and TinyImageNet). Without any bells and whistles, the CLIP model\noutperforms the state-of-the-art continual learning approaches in the majority\nof the settings. We show the effect on the CLIP model's performance by varying\ntext inputs with simple prompt templates. To the best of our knowledge, this is\nthe first work to report the CLIP zero-shot performance in a continual setting.\nWe advocate the use of this strong yet embarrassingly simple baseline for\nfuture comparisons in the continual learning tasks.\n", "date": "2022-10-06", "categories": []}
{"id": "2210.02313", "title": "Multi-stream Fusion for Class Incremental Learning in Pill Image\n  Classification", "abstract": "  Classifying pill categories from real-world images is crucial for various\nsmart healthcare applications. Although existing approaches in image\nclassification might achieve a good performance on fixed pill categories, they\nfail to handle novel instances of pill categories that are frequently presented\nto the learning algorithm. To this end, a trivial solution is to train the\nmodel with novel classes. However, this may result in a phenomenon known as\ncatastrophic forgetting, in which the system forgets what it learned in\nprevious classes. In this paper, we address this challenge by introducing the\nclass incremental learning (CIL) ability to traditional pill image\nclassification systems. Specifically, we propose a novel incremental\nmulti-stream intermediate fusion framework enabling incorporation of an\nadditional guidance information stream that best matches the domain of the\nproblem into various state-of-the-art CIL methods. From this framework, we\nconsider color-specific information of pill images as a guidance stream and\ndevise an approach, namely \"Color Guidance with Multi-stream intermediate\nfusion\"(CG-IMIF) for solving CIL pill image classification task. We conduct\ncomprehensive experiments on real-world incremental pill image classification\ndataset, namely VAIPE-PCIL, and find that the CG-IMIF consistently outperforms\nseveral state-of-the-art methods by a large margin in different task settings.\nOur code, data, and trained model are available at\nhttps://github.com/vinuni-vishc/CG-IMIF.\n", "date": "2022-10-05", "categories": []}
{"id": "2210.01600", "title": "Positive Pair Distillation Considered Harmful: Continual Meta Metric\n  Learning for Lifelong Object Re-Identification", "abstract": "  Lifelong object re-identification incrementally learns from a stream of\nre-identification tasks. The objective is to learn a representation that can be\napplied to all tasks and that generalizes to previously unseen\nre-identification tasks. The main challenge is that at inference time the\nrepresentation must generalize to previously unseen identities. To address this\nproblem, we apply continual meta metric learning to lifelong object\nre-identification. To prevent forgetting of previous tasks, we use knowledge\ndistillation and explore the roles of positive and negative pairs. Based on our\nobservation that the distillation and metric losses are antagonistic, we\npropose to remove positive pairs from distillation to robustify model updates.\nOur method, called Distillation without Positive Pairs (DwoPP), is evaluated on\nextensive intra-domain experiments on person and vehicle re-identification\ndatasets, as well as inter-domain experiments on the LReID benchmark. Our\nexperiments demonstrate that DwoPP significantly outperforms the\nstate-of-the-art. The code is here: https://github.com/wangkai930418/DwoPP_code\n", "date": "2022-10-04", "categories": []}
{"id": "2210.00232", "title": "Learnable Distribution Calibration for Few-Shot Class-Incremental\n  Learning", "abstract": "  Few-shot class-incremental learning (FSCIL) faces challenges of memorizing\nold class distributions and estimating new class distributions given few\ntraining samples. In this study, we propose a learnable distribution\ncalibration (LDC) approach, with the aim to systematically solve these two\nchallenges using a unified framework. LDC is built upon a parameterized\ncalibration unit (PCU), which initializes biased distributions for all classes\nbased on classifier vectors (memory-free) and a single covariance matrix. The\ncovariance matrix is shared by all classes, so that the memory costs are fixed.\nDuring base training, PCU is endowed with the ability to calibrate biased\ndistributions by recurrently updating sampled features under the supervision of\nreal distributions. During incremental learning, PCU recovers distributions for\nold classes to avoid `forgetting', as well as estimating distributions and\naugmenting samples for new classes to alleviate `over-fitting' caused by the\nbiased distributions of few-shot samples. LDC is theoretically plausible by\nformatting a variational inference procedure. It improves FSCIL's flexibility\nas the training procedure requires no class similarity priori. Experiments on\nCUB200, CIFAR100, and mini-ImageNet datasets show that LDC outperforms the\nstate-of-the-arts by 4.64%, 1.98%, and 3.97%, respectively. LDC's effectiveness\nis also validated on few-shot learning scenarios.\n", "date": "2022-10-01", "categories": []}
{"id": "2210.00266", "title": "Long-Tailed Class Incremental Learning", "abstract": "  In class incremental learning (CIL) a model must learn new classes in a\nsequential manner without forgetting old ones. However, conventional CIL\nmethods consider a balanced distribution for each new task, which ignores the\nprevalence of long-tailed distributions in the real world. In this work we\npropose two long-tailed CIL scenarios, which we term ordered and shuffled\nLT-CIL. Ordered LT-CIL considers the scenario where we learn from head classes\ncollected with more samples than tail classes which have few. Shuffled LT-CIL,\non the other hand, assumes a completely random long-tailed distribution for\neach task. We systematically evaluate existing methods in both LT-CIL scenarios\nand demonstrate very different behaviors compared to conventional CIL\nscenarios. Additionally, we propose a two-stage learning baseline with a\nlearnable weight scaling layer for reducing the bias caused by long-tailed\ndistribution in LT-CIL and which in turn also improves the performance of\nconventional CIL due to the limited exemplars. Our results demonstrate the\nsuperior performance (up to 6.44 points in average incremental accuracy) of our\napproach on CIFAR-100 and ImageNet-Subset. The code is available at\nhttps://github.com/xialeiliu/Long-Tailed-CIL\n", "date": "2022-10-01", "categories": []}
{"id": "2210.00713", "title": "Efficient Meta-Learning for Continual Learning with Taylor Expansion\n  Approximation", "abstract": "  Continual learning aims to alleviate catastrophic forgetting when handling\nconsecutive tasks under non-stationary distributions. Gradient-based\nmeta-learning algorithms have shown the capability to implicitly solve the\ntransfer-interference trade-off problem between different examples. However,\nthey still suffer from the catastrophic forgetting problem in the setting of\ncontinual learning, since the past data of previous tasks are no longer\navailable. In this work, we propose a novel efficient meta-learning algorithm\nfor solving the online continual learning problem, where the regularization\nterms and learning rates are adapted to the Taylor approximation of the\nparameter's importance to mitigate forgetting. The proposed method expresses\nthe gradient of the meta-loss in closed-form and thus avoid computing\nsecond-order derivative which is computationally inhibitable. We also use\nProximal Gradient Descent to further improve computational efficiency and\naccuracy. Experiments on diverse benchmarks show that our method achieves\nbetter or on-par performance and much higher efficiency compared to the\nstate-of-the-art approaches.\n", "date": "2022-10-03", "categories": []}
{"id": "2210.00940", "title": "How Relevant is Selective Memory Population in Lifelong Language\n  Learning?", "abstract": "  Lifelong language learning seeks to have models continuously learn multiple\ntasks in a sequential order without suffering from catastrophic forgetting.\nState-of-the-art approaches rely on sparse experience replay as the primary\napproach to prevent forgetting. Experience replay usually adopts sampling\nmethods for the memory population; however, the effect of the chosen sampling\nstrategy on model performance has not yet been studied. In this paper, we\ninvestigate how relevant the selective memory population is in the lifelong\nlearning process of text classification and question-answering tasks. We found\nthat methods that randomly store a uniform number of samples from the entire\ndata stream lead to high performances, especially for low memory size, which is\nconsistent with computer vision studies.\n", "date": "2022-10-03", "categories": []}
{"id": "2209.14996", "title": "Multiple Modes for Continual Learning", "abstract": "  Adapting model parameters to incoming streams of data is a crucial factor to\ndeep learning scalability. Interestingly, prior continual learning strategies\nin online settings inadvertently anchor their updated parameters to a local\nparameter subspace to remember old tasks, else drift away from the subspace and\nforget. From this observation, we formulate a trade-off between constructing\nmultiple parameter modes and allocating tasks per mode. Mode-Optimized Task\nAllocation (MOTA), our contributed adaptation strategy, trains multiple modes\nin parallel, then optimizes task allocation per mode. We empirically\ndemonstrate improvements over baseline continual learning strategies and across\nvarying distribution shifts, namely sub-population, domain, and task shift.\n", "date": "2022-09-29", "categories": []}
{"id": "2209.13900", "title": "Disentangling Transfer in Continual Reinforcement Learning", "abstract": "  The ability of continual learning systems to transfer knowledge from\npreviously seen tasks in order to maximize performance on new tasks is a\nsignificant challenge for the field, limiting the applicability of continual\nlearning solutions to realistic scenarios. Consequently, this study aims to\nbroaden our understanding of transfer and its driving forces in the specific\ncase of continual reinforcement learning. We adopt SAC as the underlying RL\nalgorithm and Continual World as a suite of continuous control tasks. We\nsystematically study how different components of SAC (the actor and the critic,\nexploration, and data) affect transfer efficacy, and we provide recommendations\nregarding various modeling options. The best set of choices, dubbed ClonEx-SAC,\nis evaluated on the recent Continual World benchmark. ClonEx-SAC achieves 87%\nfinal success rate compared to 80% of PackNet, the best method in the\nbenchmark. Moreover, the transfer grows from 0.18 to 0.54 according to the\nmetric provided by Continual World.\n", "date": "2022-09-28", "categories": []}
{"id": "2209.12241", "title": "Exploring Example Influence in Continual Learning", "abstract": "  Continual Learning (CL) sequentially learns new tasks like human beings, with\nthe goal to achieve better Stability (S, remembering past tasks) and Plasticity\n(P, adapting to new tasks). Due to the fact that past training data is not\navailable, it is valuable to explore the influence difference on S and P among\ntraining examples, which may improve the learning pattern towards better SP.\nInspired by Influence Function (IF), we first study example influence via\nadding perturbation to example weight and computing the influence derivation.\nTo avoid the storage and calculation burden of Hessian inverse in neural\nnetworks, we propose a simple yet effective MetaSP algorithm to simulate the\ntwo key steps in the computation of IF and obtain the S- and P-aware example\ninfluence. Moreover, we propose to fuse two kinds of example influence by\nsolving a dual-objective optimization problem, and obtain a fused influence\ntowards SP Pareto optimality. The fused influence can be used to control the\nupdate of model and optimize the storage of rehearsal. Empirical results show\nthat our algorithm significantly outperforms state-of-the-art methods on both\ntask- and class-incremental benchmark CL datasets.\n", "date": "2022-09-25", "categories": []}
{"id": "2209.11469", "title": "Optimizing Class Distribution in Memory for Multi-Label Online Continual\n  Learning", "abstract": "  Online continual learning, especially when task identities and task\nboundaries are unavailable, is a challenging continual learning setting. One\nrepresentative kind of methods for online continual learning is replay-based\nmethods, in which a replay buffer called memory is maintained to keep a small\npart of past samples for overcoming catastrophic forgetting. When tackling with\nonline continual learning, most existing replay-based methods focus on\nsingle-label problems in which each sample in the data stream has only one\nlabel. But multi-label problems may also happen in the online continual\nlearning setting in which each sample may have more than one label. In the\nonline setting with multi-label samples, the class distribution in data stream\nis typically highly imbalanced, and it is challenging to control class\ndistribution in memory since changing the number of samples belonging to one\nclass may affect the number of samples belonging to other classes. But class\ndistribution in memory is critical for replay-based memory to get good\nperformance, especially when the class distribution in data stream is highly\nimbalanced. In this paper, we propose a simple but effective method, called\noptimizing class distribution in memory (OCDM), for multi-label online\ncontinual learning. OCDM formulates the memory update mechanism as an\noptimization problem and updates the memory by solving this problem.\nExperiments on two widely used multi-label datasets show that OCDM can control\nthe class distribution in memory well and can outperform other state-of-the-art\nmethods.\n", "date": "2022-09-23", "categories": []}
{"id": "2209.09476", "title": "SparCL: Sparse Continual Learning on the Edge", "abstract": "  Existing work in continual learning (CL) focuses on mitigating catastrophic\nforgetting, i.e., model performance deterioration on past tasks when learning a\nnew task. However, the training efficiency of a CL system is\nunder-investigated, which limits the real-world application of CL systems under\nresource-limited scenarios. In this work, we propose a novel framework called\nSparse Continual Learning(SparCL), which is the first study that leverages\nsparsity to enable cost-effective continual learning on edge devices. SparCL\nachieves both training acceleration and accuracy preservation through the\nsynergy of three aspects: weight sparsity, data efficiency, and gradient\nsparsity. Specifically, we propose task-aware dynamic masking (TDM) to learn a\nsparse network throughout the entire CL process, dynamic data removal (DDR) to\nremove less informative training data, and dynamic gradient masking (DGM) to\nsparsify the gradient updates. Each of them not only improves efficiency, but\nalso further mitigates catastrophic forgetting. SparCL consistently improves\nthe training efficiency of existing state-of-the-art (SOTA) CL methods by at\nmost 23X less training FLOPs, and, surprisingly, further improves the SOTA\naccuracy by at most 1.7%. SparCL also outperforms competitive baselines\nobtained from adapting SOTA sparse training methods to the CL setting in both\nefficiency and accuracy. We also evaluate the effectiveness of SparCL on a real\nmobile phone, further indicating the practical potential of our method.\n", "date": "2022-09-20", "categories": []}
{"id": "2209.09839", "title": "Improving Replay-Based Continual Semantic Segmentation with Smart Data\n  Selection", "abstract": "  Continual learning for Semantic Segmentation (CSS) is a rapidly emerging\nfield, in which the capabilities of the segmentation model are incrementally\nimproved by learning new classes or new domains. A central challenge in\nContinual Learning is overcoming the effects of catastrophic forgetting, which\nrefers to the sudden drop in accuracy on previously learned tasks after the\nmodel is trained on new classes or domains. In continual classification this\nchallenge is often overcome by replaying a small selection of samples from\nprevious tasks, however replay is rarely considered in CSS. Therefore, we\ninvestigate the influences of various replay strategies for semantic\nsegmentation and evaluate them in class- and domain-incremental settings. Our\nfindings suggest that in a class-incremental setting, it is critical to achieve\na uniform distribution for the different classes in the buffer to avoid a bias\ntowards newly learned classes. In the domain-incremental setting, it is most\neffective to select buffer samples by uniformly sampling from the distribution\nof learned feature representations or by choosing samples with median entropy.\nFinally, we observe that the effective sampling methods help to decrease the\nrepresentation shift significantly in early layers, which is a major cause of\nforgetting in domain-incremental learning.\n", "date": "2022-09-20", "categories": []}
{"id": "2209.07712", "title": "Continual Learning with Dependency Preserving Hypernetworks", "abstract": "  Humans learn continually throughout their lifespan by accumulating diverse\nknowledge and fine-tuning it for future tasks. When presented with a similar\ngoal, neural networks suffer from catastrophic forgetting if data distributions\nacross sequential tasks are not stationary over the course of learning. An\neffective approach to address such continual learning (CL) problems is to use\nhypernetworks which generate task dependent weights for a target network.\nHowever, the continual learning performance of existing hypernetwork based\napproaches are affected by the assumption of independence of the weights across\nthe layers in order to maintain parameter efficiency. To address this\nlimitation, we propose a novel approach that uses a dependency preserving\nhypernetwork to generate weights for the target network while also maintaining\nthe parameter efficiency. We propose to use recurrent neural network (RNN)\nbased hypernetwork that can generate layer weights efficiently while allowing\nfor dependencies across them. In addition, we propose novel regularisation and\nnetwork growth techniques for the RNN based hypernetwork to further improve the\ncontinual learning performance. To demonstrate the effectiveness of the\nproposed methods, we conducted experiments on several image classification\ncontinual learning tasks and settings. We found that the proposed methods based\non the RNN hypernetworks outperformed the baselines in all these CL settings\nand tasks.\n", "date": "2022-09-16", "categories": []}
{"id": "2209.08023", "title": "Continual Learning for Class- and Domain-Incremental Semantic\n  Segmentation", "abstract": "  The field of continual deep learning is an emerging field and a lot of\nprogress has been made. However, concurrently most of the approaches are only\ntested on the task of image classification, which is not relevant in the field\nof intelligent vehicles. Only recently approaches for class-incremental\nsemantic segmentation were proposed. However, all of those approaches are based\non some form of knowledge distillation. At the moment there are no\ninvestigations on replay-based approaches that are commonly used for object\nrecognition in a continual setting. At the same time while unsupervised domain\nadaption for semantic segmentation gained a lot of traction, investigations\nregarding domain-incremental learning in an continual setting is not\nwell-studied. Therefore, the goal of our work is to evaluate and adapt\nestablished solutions for continual object recognition to the task of semantic\nsegmentation and to provide baseline methods and evaluation protocols for the\ntask of continual semantic segmentation. We firstly introduce evaluation\nprotocols for the class- and domain-incremental segmentation and analyze\nselected approaches. We show that the nature of the task of semantic\nsegmentation changes which methods are most effective in mitigating forgetting\ncompared to image classification. Especially, in class-incremental learning\nknowledge distillation proves to be a vital tool, whereas in domain-incremental\nlearning replay methods are the most effective method.\n", "date": "2022-09-16", "categories": []}
{"id": "2209.06107", "title": "Online Continual Learning via the Meta-learning Update with Multi-scale\n  Knowledge Distillation and Data Augmentation", "abstract": "  Continual learning aims to rapidly and continually learn the current task\nfrom a sequence of tasks.\n  Compared to other kinds of methods, the methods based on experience replay\nhave shown great advantages to overcome catastrophic forgetting. One common\nlimitation of this method is the data imbalance between the previous and\ncurrent tasks, which would further aggravate forgetting. Moreover, how to\neffectively address the stability-plasticity dilemma in this setting is also an\nurgent problem to be solved. In this paper, we overcome these challenges by\nproposing a novel framework called Meta-learning update via Multi-scale\nKnowledge Distillation and Data Augmentation (MMKDDA). Specifically, we apply\nmultiscale knowledge distillation to grasp the evolution of long-range and\nshort-range spatial relationships at different feature levels to alleviate the\nproblem of data imbalance. Besides, our method mixes the samples from the\nepisodic memory and current task in the online continual training procedure,\nthus alleviating the side influence due to the change of probability\ndistribution. Moreover, we optimize our model via the meta-learning update\nresorting to the number of tasks seen previously, which is helpful to keep a\nbetter balance between stability and plasticity. Finally, our experimental\nevaluation on four benchmark datasets shows the effectiveness of the proposed\nMMKDDA framework against other popular baselines, and ablation studies are also\nconducted to further analyze the role of each component in our framework.\n", "date": "2022-09-12", "categories": []}
{"id": "2208.02048", "title": "Centroids Matching: an efficient Continual Learning approach operating\n  in the embedding space", "abstract": "  Catastrophic forgetting (CF) occurs when a neural network loses the\ninformation previously learned while training on a set of samples from a\ndifferent distribution, i.e., a new task. Existing approaches have achieved\nremarkable results in mitigating CF, especially in a scenario called task\nincremental learning. However, this scenario is not realistic, and limited work\nhas been done to achieve good results on more realistic scenarios. In this\npaper, we propose a novel regularization method called Centroids Matching,\nthat, inspired by meta-learning approaches, fights CF by operating in the\nfeature space produced by the neural network, achieving good results while\nrequiring a small memory footprint. Specifically, the approach classifies the\nsamples directly using the feature vectors produced by the neural network, by\nmatching those vectors with the centroids representing the classes from the\ncurrent task, or all the tasks up to that point. Centroids Matching is faster\nthan competing baselines, and it can be exploited to efficiently mitigate CF,\nby preserving the distances between the embedding space produced by the model\nwhen past tasks were over, and the one currently produced, leading to a method\nthat achieves high accuracy on all the tasks, without using an external memory\nwhen operating on easy scenarios, or using a small one for more realistic ones.\nExtensive experiments demonstrate that Centroids Matching achieves accuracy\ngains on multiple datasets and scenarios.\n", "date": "2022-08-03", "categories": []}
{"id": "2209.05245", "title": "Continual learning benefits from multiple sleep mechanisms: NREM, REM,\n  and Synaptic Downscaling", "abstract": "  Learning new tasks and skills in succession without losing prior learning\n(i.e., catastrophic forgetting) is a computational challenge for both\nartificial and biological neural networks, yet artificial systems struggle to\nachieve parity with their biological analogues. Mammalian brains employ\nnumerous neural operations in support of continual learning during sleep. These\nare ripe for artificial adaptation. Here, we investigate how modeling three\ndistinct components of mammalian sleep together affects continual learning in\nartificial neural networks: (1) a veridical memory replay process observed\nduring non-rapid eye movement (NREM) sleep; (2) a generative memory replay\nprocess linked to REM sleep; and (3) a synaptic downscaling process which has\nbeen proposed to tune signal-to-noise ratios and support neural upkeep. We find\nbenefits from the inclusion of all three sleep components when evaluating\nperformance on a continual learning CIFAR-100 image classification benchmark.\nMaximum accuracy improved during training and catastrophic forgetting was\nreduced during later tasks. While some catastrophic forgetting persisted over\nthe course of network training, higher levels of synaptic downscaling lead to\nbetter retention of early tasks and further facilitated the recovery of early\ntask accuracy during subsequent training. One key takeaway is that there is a\ntrade-off at hand when considering the level of synaptic downscaling to use -\nmore aggressive downscaling better protects early tasks, but less downscaling\nenhances the ability to learn new tasks. Intermediate levels can strike a\nbalance with the highest overall accuracies during training. Overall, our\nresults both provide insight into how to adapt sleep components to enhance\nartificial continual learning systems and highlight areas for future\nneuroscientific sleep research to further such systems.\n", "date": "2022-09-09", "categories": []}
{"id": "2209.04212", "title": "Selecting Related Knowledge via Efficient Channel Attention for Online\n  Continual Learning", "abstract": "  Continual learning aims to learn a sequence of tasks by leveraging the\nknowledge acquired in the past in an online-learning manner while being able to\nperform well on all previous tasks, this ability is crucial to the artificial\nintelligence (AI) system, hence continual learning is more suitable for most\nreal-word and complex applicative scenarios compared to the traditional\nlearning pattern. However, the current models usually learn a generic\nrepresentation base on the class label on each task and an effective strategy\nis selected to avoid catastrophic forgetting. We postulate that selecting the\nrelated and useful parts only from the knowledge obtained to perform each task\nis more effective than utilizing the whole knowledge. Based on this fact, in\nthis paper we propose a new framework, named Selecting Related Knowledge for\nOnline Continual Learning (SRKOCL), which incorporates an additional efficient\nchannel attention mechanism to pick the particular related knowledge for every\ntask. Our model also combines experience replay and knowledge distillation to\ncircumvent the catastrophic forgetting. Finally, extensive experiments are\nconducted on different benchmarks and the competitive experimental results\ndemonstrate that our proposed SRKOCL is a promised approach against the\nstate-of-the-art.\n", "date": "2022-09-09", "categories": []}
{"id": "2209.03603", "title": "nVFNet-RDC: Replay and Non-Local Distillation Collaboration for\n  Continual Object Detection", "abstract": "  Continual Learning (CL) focuses on developing algorithms with the ability to\nadapt to new environments and learn new skills. This very challenging task has\ngenerated a lot of interest in recent years, with new solutions appearing\nrapidly. In this paper, we propose a nVFNet-RDC approach for continual object\ndetection. Our nVFNet-RDC consists of teacher-student models, and adopts replay\nand feature distillation strategies. As the 1st place solutions, we achieve\n55.94% and 54.65% average mAP on the 3rd CLVision Challenge Track 2 and Track\n3, respectively.\n", "date": "2022-09-08", "categories": []}
{"id": "2209.01326", "title": "Continual Learning for Steganalysis", "abstract": "  To detect the existing steganographic algorithms, recent steganalysis methods\nusually train a Convolutional Neural Network (CNN) model on the dataset\nconsisting of corresponding paired cover/stego-images. However, it is\ninefficient and impractical for those steganalysis tools to completely retrain\nthe CNN model to make it effective against both the existing steganographic\nalgorithms and a new emerging steganographic algorithm. Thus, existing\nsteganalysis models usually lack dynamic extensibility for new steganographic\nalgorithms, which limits their application in real-world scenarios. To address\nthis issue, we propose an accurate parameter importance estimation (APIE)\nbased-continual learning scheme for steganalysis. In this scheme, when a\nsteganalysis model is trained on the new image dataset generated by the new\nsteganographic algorithm, its network parameters are effectively and\nefficiently updated with sufficient consideration of their importance evaluated\nin the previous training process. This approach can guide the steganalysis\nmodel to learn the patterns of the new steganographic algorithm without\nsignificantly degrading the detectability against the previous steganographic\nalgorithms. Experimental results demonstrate the proposed scheme has promising\nextensibility for new emerging steganographic algorithms.\n", "date": "2022-09-03", "categories": []}
{"id": "2209.01556", "title": "Reinforced Continual Learning for Graphs", "abstract": "  Graph Neural Networks (GNNs) have become the backbone for a myriad of tasks\npertaining to graphs and similar topological data structures. While many works\nhave been established in domains related to node and graph\nclassification/regression tasks, they mostly deal with a single task. Continual\nlearning on graphs is largely unexplored and existing graph continual learning\napproaches are limited to the task-incremental learning scenarios. This paper\nproposes a graph continual learning strategy that combines the\narchitecture-based and memory-based approaches. The structural learning\nstrategy is driven by reinforcement learning, where a controller network is\ntrained in such a way to determine an optimal number of nodes to be\nadded/pruned from the base network when new tasks are observed, thus assuring\nsufficient network capacities. The parameter learning strategy is underpinned\nby the concept of Dark Experience replay method to cope with the catastrophic\nforgetting problem. Our approach is numerically validated with several graph\ncontinual learning benchmark problems in both task-incremental learning and\nclass-incremental learning settings. Compared to recently published works, our\napproach demonstrates improved performance in both the settings. The\nimplementation code can be found at \\url{https://github.com/codexhammer/gcl}.\n", "date": "2022-09-04", "categories": []}
{"id": "2209.01558", "title": "Scalable Adversarial Online Continual Learning", "abstract": "  Adversarial continual learning is effective for continual learning problems\nbecause of the presence of feature alignment process generating task-invariant\nfeatures having low susceptibility to the catastrophic forgetting problem.\nNevertheless, the ACL method imposes considerable complexities because it\nrelies on task-specific networks and discriminators. It also goes through an\niterative training process which does not fit for online (one-epoch) continual\nlearning problems. This paper proposes a scalable adversarial continual\nlearning (SCALE) method putting forward a parameter generator transforming\ncommon features into task-specific features and a single discriminator in the\nadversarial game to induce common features. The training process is carried out\nin meta-learning fashions using a new combination of three loss functions.\nSCALE outperforms prominent baselines with noticeable margins in both accuracy\nand execution time.\n", "date": "2022-09-04", "categories": []}
{"id": "2209.02112", "title": "Class-Incremental Learning via Knowledge Amalgamation", "abstract": "  Catastrophic forgetting has been a significant problem hindering the\ndeployment of deep learning algorithms in the continual learning setting.\nNumerous methods have been proposed to address the catastrophic forgetting\nproblem where an agent loses its generalization power of old tasks while\nlearning new tasks. We put forward an alternative strategy to handle the\ncatastrophic forgetting with knowledge amalgamation (CFA), which learns a\nstudent network from multiple heterogeneous teacher models specializing in\nprevious tasks and can be applied to current offline methods. The knowledge\namalgamation process is carried out in a single-head manner with only a\nselected number of memorized samples and no annotations. The teachers and\nstudents do not need to share the same network structure, allowing\nheterogeneous tasks to be adapted to a compact or sparse data representation.\nWe compare our method with competitive baselines from different strategies,\ndemonstrating our approach's advantages.\n", "date": "2022-09-05", "categories": []}
{"id": "2209.00243", "title": "Less is More: Rethinking State-of-the-art Continual Relation Extraction\n  Models with a Frustratingly Easy but Effective Approach", "abstract": "  Continual relation extraction (CRE) requires the model to continually learn\nnew relations from class-incremental data streams. In this paper, we propose a\nFrustratingly easy but Effective Approach (FEA) method with two learning stages\nfor CRE: 1) Fast Adaption (FA) warms up the model with only new data. 2)\nBalanced Tuning (BT) finetunes the model on the balanced memory data. Despite\nits simplicity, FEA achieves comparable (on TACRED or superior (on FewRel)\nperformance compared with the state-of-the-art baselines. With careful\nexaminations, we find that the data imbalance between new and old relations\nleads to a skewed decision boundary in the head classifiers over the pretrained\nencoders, thus hurting the overall performance. In FEA, the FA stage unleashes\nthe potential of memory data for the subsequent finetuning, while the BT stage\nhelps establish a more balanced decision boundary. With a unified view, we find\nthat two strong CRE baselines can be subsumed into the proposed training\npipeline. The success of FEA also provides actionable insights and suggestions\nfor future model designing in CRE.\n", "date": "2022-09-01", "categories": []}
{"id": "2209.00591", "title": "Incremental Online Learning Algorithms Comparison for Gesture and Visual\n  Smart Sensors", "abstract": "  Tiny machine learning (TinyML) in IoT systems exploits MCUs as edge devices\nfor data processing. However, traditional TinyML methods can only perform\ninference, limited to static environments or classes. Real case scenarios\nusually work in dynamic environments, thus drifting the context where the\noriginal neural model is no more suitable. For this reason, pre-trained models\nreduce accuracy and reliability during their lifetime because the data recorded\nslowly becomes obsolete or new patterns appear. Continual learning strategies\nmaintain the model up to date, with runtime fine-tuning of the parameters. This\npaper compares four state-of-the-art algorithms in two real applications: i)\ngesture recognition based on accelerometer data and ii) image classification.\nOur results confirm these systems' reliability and the feasibility of deploying\nthem in tiny-memory MCUs, with a drop in the accuracy of a few percentage\npoints with respect to the original models for unconstrained computing\nplatforms.\n", "date": "2022-09-01", "categories": []}
{"id": "2208.12037", "title": "Symbolic Replay: Scene Graph as Prompt for Continual Learning on VQA\n  Task", "abstract": "  VQA is an ambitious task aiming to answer any image-related question.\nHowever, in reality, it is hard to build such a system once for all since the\nneeds of users are continuously updated, and the system has to implement new\nfunctions. Thus, Continual Learning (CL) ability is a must in developing\nadvanced VQA systems. Recently, a pioneer work split a VQA dataset into\ndisjoint answer sets to study this topic. However, CL on VQA involves not only\nthe expansion of label sets (new Answer sets). It is crucial to study how to\nanswer questions when deploying VQA systems to new environments (new Visual\nscenes) and how to answer questions requiring new functions (new Question\ntypes). Thus, we propose CLOVE, a benchmark for Continual Learning On Visual\nquEstion answering, which contains scene- and function-incremental settings for\nthe two aforementioned CL scenarios. In terms of methodology, the main\ndifference between CL on VQA and classification is that the former additionally\ninvolves expanding and preventing forgetting of reasoning mechanisms, while the\nlatter focusing on class representation. Thus, we propose a real-data-free\nreplay-based method tailored for CL on VQA, named Scene Graph as Prompt for\nSymbolic Replay. Using a piece of scene graph as a prompt, it replays pseudo\nscene graphs to represent the past images, along with correlated QA pairs. A\nunified VQA model is also proposed to utilize the current and replayed data to\nenhance its QA ability. Finally, experimental results reveal challenges in\nCLOVE and demonstrate the effectiveness of our method. The dataset and code\nwill be available at https://github.com/showlab/CLVQA.\n", "date": "2022-08-24", "categories": []}
{"id": "2208.14307", "title": "Beyond Supervised Continual Learning: a Review", "abstract": "  Continual Learning (CL, sometimes also termed incremental learning) is a\nflavor of machine learning where the usual assumption of stationary data\ndistribution is relaxed or omitted. When naively applying, e.g., DNNs in CL\nproblems, changes in the data distribution can cause the so-called catastrophic\nforgetting (CF) effect: an abrupt loss of previous knowledge. Although many\nsignificant contributions to enabling CL have been made in recent years, most\nworks address supervised (classification) problems. This article reviews\nliterature that study CL in other settings, such as learning with reduced\nsupervision, fully unsupervised learning, and reinforcement learning. Besides\nproposing a simple schema for classifying CL approaches w.r.t. their level of\nautonomy and supervision, we discuss the specific challenges associated with\neach setting and the potential contributions to the field of CL in general.\n", "date": "2022-08-30", "categories": []}
{"id": "2207.07256", "title": "Improving Task-free Continual Learning by Distributionally Robust Memory\n  Evolution", "abstract": "  Task-free continual learning (CL) aims to learn a non-stationary data stream\nwithout explicit task definitions and not forget previous knowledge. The widely\nadopted memory replay approach could gradually become less effective for long\ndata streams, as the model may memorize the stored examples and overfit the\nmemory buffer. Second, existing methods overlook the high uncertainty in the\nmemory data distribution since there is a big gap between the memory data\ndistribution and the distribution of all the previous data examples. To address\nthese problems, for the first time, we propose a principled memory evolution\nframework to dynamically evolve the memory data distribution by making the\nmemory buffer gradually harder to be memorized with distributionally robust\noptimization (DRO). We then derive a family of methods to evolve the memory\nbuffer data in the continuous probability measure space with Wasserstein\ngradient flow (WGF). The proposed DRO is w.r.t the worst-case evolved memory\ndata distribution, thus guarantees the model performance and learns\nsignificantly more robust features than existing memory-replay-based methods.\nExtensive experiments on existing benchmarks demonstrate the effectiveness of\nthe proposed methods for alleviating forgetting. As a by-product of the\nproposed framework, our method is more robust to adversarial examples than\nexisting task-free CL methods. Code is available on GitHub\n\\url{https://github.com/joey-wang123/DRO-Task-free}\n", "date": "2022-07-15", "categories": []}
{"id": "2208.09734", "title": "A Multi-Head Model for Continual Learning via Out-of-Distribution Replay", "abstract": "  This paper studies class incremental learning (CIL) of continual learning\n(CL). Many approaches have been proposed to deal with catastrophic forgetting\n(CF) in CIL. Most methods incrementally construct a single classifier for all\nclasses of all tasks in a single head network. To prevent CF, a popular\napproach is to memorize a small number of samples from previous tasks and\nreplay them during training of the new task. However, this approach still\nsuffers from serious CF as the parameters learned for previous tasks are\nupdated or adjusted with only the limited number of saved samples in the\nmemory. This paper proposes an entirely different approach that builds a\nseparate classifier (head) for each task (called a multi-head model) using a\ntransformer network, called MORE. Instead of using the saved samples in memory\nto update the network for previous tasks/classes in the existing approach, MORE\nleverages the saved samples to build a task specific classifier (adding a new\nclassification head) without updating the network learned for previous\ntasks/classes. The model for the new task in MORE is trained to learn the\nclasses of the task and also to detect samples that are not from the same data\ndistribution (i.e., out-of-distribution (OOD)) of the task. This enables the\nclassifier for the task to which the test instance belongs to produce a high\nscore for the correct class and the classifiers of other tasks to produce low\nscores because the test instance is not from the data distributions of these\nclassifiers. Experimental results show that MORE outperforms state-of-the-art\nbaselines and is also naturally capable of performing OOD detection in the\ncontinual learning setting.\n", "date": "2022-08-20", "categories": []}
{"id": "2208.08112", "title": "DLCFT: Deep Linear Continual Fine-Tuning for General Incremental\n  Learning", "abstract": "  Pre-trained representation is one of the key elements in the success of\nmodern deep learning. However, existing works on continual learning methods\nhave mostly focused on learning models incrementally from scratch. In this\npaper, we explore an alternative framework to incremental learning where we\ncontinually fine-tune the model from a pre-trained representation. Our method\ntakes advantage of linearization technique of a pre-trained neural network for\nsimple and effective continual learning. We show that this allows us to design\na linear model where quadratic parameter regularization method is placed as the\noptimal continual learning policy, and at the same time enjoying the high\nperformance of neural networks. We also show that the proposed algorithm\nenables parameter regularization methods to be applied to class-incremental\nproblems. Additionally, we provide a theoretical reason why the existing\nparameter-space regularization algorithms such as EWC underperform on neural\nnetworks trained with cross-entropy loss. We show that the proposed method can\nprevent forgetting while achieving high continual fine-tuning performance on\nimage classification tasks. To show that our method can be applied to general\ncontinual learning settings, we evaluate our method in data-incremental,\ntask-incremental, and class-incremental learning problems.\n", "date": "2022-08-17", "categories": []}
{"id": "2208.03767", "title": "Class-Incremental Learning with Cross-Space Clustering and Controlled\n  Transfer", "abstract": "  In class-incremental learning, the model is expected to learn new classes\ncontinually while maintaining knowledge on previous classes. The challenge here\nlies in preserving the model's ability to effectively represent prior classes\nin the feature space, while adapting it to represent incoming new classes. We\npropose two distillation-based objectives for class incremental learning that\nleverage the structure of the feature space to maintain accuracy on previous\nclasses, as well as enable learning the new classes. In our first objective,\ntermed cross-space clustering (CSC), we propose to use the feature space\nstructure of the previous model to characterize directions of optimization that\nmaximally preserve the class: directions that all instances of a specific class\nshould collectively optimize towards, and those that they should collectively\noptimize away from. Apart from minimizing forgetting, this indirectly\nencourages the model to cluster all instances of a class in the current feature\nspace, and gives rise to a sense of herd-immunity, allowing all samples of a\nclass to jointly combat the model from forgetting the class. Our second\nobjective termed controlled transfer (CT) tackles incremental learning from an\nunderstudied perspective of inter-class transfer. CT explicitly approximates\nand conditions the current model on the semantic similarities between\nincrementally arriving classes and prior classes. This allows the model to\nlearn classes in such a way that it maximizes positive forward transfer from\nsimilar prior classes, thus increasing plasticity, and minimizes negative\nbackward transfer on dissimilar prior classes, whereby strengthening stability.\nWe perform extensive experiments on two benchmark datasets, adding our method\n(CSCCT) on top of three prominent class-incremental learning methods. We\nobserve consistent performance improvement on a variety of experimental\nsettings.\n", "date": "2022-08-07", "categories": []}
{"id": "2208.05056", "title": "Model-Free Generative Replay for Lifelong Reinforcement Learning:\n  Application to Starcraft-2", "abstract": "  One approach to meet the challenges of deep lifelong reinforcement learning\n(LRL) is careful management of the agent's learning experiences, to learn\n(without forgetting) and build internal meta-models (of the tasks,\nenvironments, agents, and world). Generative replay (GR) is a biologically\ninspired replay mechanism that augments learning experiences with self-labelled\nexamples drawn from an internal generative model that is updated over time. We\npresent a version of GR for LRL that satisfies two desiderata: (a)\nIntrospective density modelling of the latent representations of policies\nlearned using deep RL, and (b) Model-free end-to-end learning. In this paper,\nwe study three deep learning architectures for model-free GR, starting from a\nna\\\"ive GR and adding ingredients to achieve (a) and (b). We evaluate our\nproposed algorithms on three different scenarios comprising tasks from the\nStarcraft-2 and Minigrid domains. We report several key findings showing the\nimpact of the design choices on quantitative metrics that include transfer\nlearning, generalization to unseen tasks, fast adaptation after task change,\nperformance wrt task expert, and catastrophic forgetting. We observe that our\nGR prevents drift in the features-to-action mapping from the latent vector\nspace of a deep RL agent. We also show improvements in established lifelong\nlearning metrics. We find that a small random replay buffer significantly\nincreases the stability of training. Overall, we find that \"hidden replay\" (a\nwell-known architecture for class-incremental classification) is the most\npromising approach that pushes the state-of-the-art in GR for LRL and observe\nthat the architecture of the sleep model might be more important for improving\nperformance than the types of replay used. Our experiments required only 6% of\ntraining samples to achieve 80-90% of expert performance in most Starcraft-2\nscenarios.\n", "date": "2022-08-09", "categories": []}
{"id": "2208.05791", "title": "Empirical investigations on WVA structural issues", "abstract": "  In this paper we want to present the results of empirical verification of\nsome issues concerning the methods for overcoming catastrophic forgetting in\nneural networks. First, in the introduction, we will try to describe in detail\nthe problem of catastrophic forgetting and methods for overcoming it for those\nwho are not yet familiar with this topic. Then we will discuss the essence and\nlimitations of the WVA method which we presented in previous papers. Further,\nwe will touch upon the issues of applying the WVA method to gradients or\noptimization steps of weights, choosing the optimal attenuation function in\nthis method, as well as choosing the optimal hyper-parameters of the method\ndepending on the number of tasks in sequential training of neural networks.\n", "date": "2022-08-11", "categories": []}
{"id": "2208.06931", "title": "A Theory for Knowledge Transfer in Continual Learning", "abstract": "  Continual learning of a stream of tasks is an active area in deep neural\nnetworks. The main challenge investigated has been the phenomenon of\ncatastrophic forgetting or interference of newly acquired knowledge with\nknowledge from previous tasks. Recent work has investigated forward knowledge\ntransfer to new tasks. Backward transfer for improving knowledge gained during\nprevious tasks has received much less attention. There is in general limited\nunderstanding of how knowledge transfer could aid tasks learned continually. We\npresent a theory for knowledge transfer in continual supervised learning, which\nconsiders both forward and backward transfer. We aim at understanding their\nimpact for increasingly knowledgeable learners. We derive error bounds for each\nof these transfer mechanisms. These bounds are agnostic to specific\nimplementations (e.g. deep neural networks). We demonstrate that, for a\ncontinual learner that observes related tasks, both forward and backward\ntransfer can contribute to an increasing performance as more tasks are\nobserved.\n", "date": "2022-08-14", "categories": []}
{"id": "2208.05217", "title": "Continual Machine Reading Comprehension via Uncertainty-aware Fixed\n  Memory and Adversarial Domain Adaptation", "abstract": "  Continual Machine Reading Comprehension aims to incrementally learn from a\ncontinuous data stream across time without access the previous seen data, which\nis crucial for the development of real-world MRC systems. However, it is a\ngreat challenge to learn a new domain incrementally without catastrophically\nforgetting previous knowledge. In this paper, MA-MRC, a continual MRC model\nwith uncertainty-aware fixed Memory and Adversarial domain adaptation, is\nproposed. In MA-MRC, a fixed size memory stores a small number of samples in\nprevious domain data along with an uncertainty-aware updating strategy when new\ndomain data arrives. For incremental learning, MA-MRC not only keeps a stable\nunderstanding by learning both memory and new domain data, but also makes full\nuse of the domain adaptation relationship between them by adversarial learning\nstrategy. The experimental results show that MA-MRC is superior to strong\nbaselines and has a substantial incremental learning ability without\ncatastrophically forgetting under two different continual MRC settings.\n", "date": "2022-08-10", "categories": []}
{"id": "2208.05388", "title": "ATLAS: Universal Function Approximator for Memory Retention", "abstract": "  Artificial neural networks (ANNs), despite their universal function\napproximation capability and practical success, are subject to catastrophic\nforgetting. Catastrophic forgetting refers to the abrupt unlearning of a\nprevious task when a new task is learned. It is an emergent phenomenon that\nhinders continual learning. Existing universal function approximation theorems\nfor ANNs guarantee function approximation ability, but do not predict\ncatastrophic forgetting. This paper presents a novel universal approximation\ntheorem for multi-variable functions using only single-variable functions and\nexponential functions. Furthermore, we present ATLAS: a novel ANN architecture\nbased on the new theorem. It is shown that ATLAS is a universal function\napproximator capable of some memory retention, and continual learning. The\nmemory of ATLAS is imperfect, with some off-target effects during continual\nlearning, but it is well-behaved and predictable. An efficient implementation\nof ATLAS is provided. Experiments are conducted to evaluate both the function\napproximation and memory retention capabilities of ATLAS.\n", "date": "2022-08-10", "categories": []}
{"id": "2208.04227", "title": "A Multi-label Continual Learning Framework to Scale Deep Learning\n  Approaches for Packaging Equipment Monitoring", "abstract": "  Continual Learning aims to learn from a stream of tasks, being able to\nremember at the same time both new and old tasks. While many approaches were\nproposed for single-class classification, multi-label classification in the\ncontinual scenario remains a challenging problem. For the first time, we study\nmulti-label classification in the Domain Incremental Learning scenario.\nMoreover, we propose an efficient approach that has a logarithmic complexity\nwith regard to the number of tasks, and can be applied also in the Class\nIncremental Learning scenario. We validate our approach on a real-world\nmulti-label Alarm Forecasting problem from the packaging industry. For the sake\nof reproducibility, the dataset and the code used for the experiments are\npublicly available.\n", "date": "2022-08-08", "categories": []}
{"id": "2208.04287", "title": "Continual Reinforcement Learning with TELLA", "abstract": "  Training reinforcement learning agents that continually learn across multiple\nenvironments is a challenging problem. This is made more difficult by a lack of\nreproducible experiments and standard metrics for comparing different continual\nlearning approaches. To address this, we present TELLA, a tool for the Test and\nEvaluation of Lifelong Learning Agents. TELLA provides specified, reproducible\ncurricula to lifelong learning agents while logging detailed data for\nevaluation and standardized analysis. Researchers can define and share their\nown curricula over various learning environments or run against a curriculum\ncreated under the DARPA Lifelong Learning Machines (L2M) Program.\n", "date": "2022-08-08", "categories": []}
{"id": "2208.03206", "title": "Task-agnostic Continual Hippocampus Segmentation for Smooth Population\n  Shifts", "abstract": "  Most continual learning methods are validated in settings where task\nboundaries are clearly defined and task identity information is available\nduring training and testing. We explore how such methods perform in a\ntask-agnostic setting that more closely resembles dynamic clinical environments\nwith gradual population shifts. We propose ODEx, a holistic solution that\ncombines out-of-distribution detection with continual learning techniques.\nValidation on two scenarios of hippocampus segmentation shows that our proposed\nmethod reliably maintains performance on earlier tasks without losing\nplasticity.\n", "date": "2022-08-05", "categories": []}
{"id": "2208.02660", "title": "A Benchmark and Empirical Analysis for Replay Strategies in Continual\n  Learning", "abstract": "  With the capacity of continual learning, humans can continuously acquire\nknowledge throughout their lifespan. However, computational systems are not, in\ngeneral, capable of learning tasks sequentially. This long-standing challenge\nfor deep neural networks (DNNs) is called catastrophic forgetting. Multiple\nsolutions have been proposed to overcome this limitation. This paper makes an\nin-depth evaluation of the memory replay methods, exploring the efficiency,\nperformance, and scalability of various sampling strategies when selecting\nreplay data. All experiments are conducted on multiple datasets under various\ndomains. Finally, a practical solution for selecting replay methods for various\ndata distributions is provided.\n", "date": "2022-08-04", "categories": []}
{"id": "2208.00147", "title": "Few-Shot Class-Incremental Learning from an Open-Set Perspective", "abstract": "  The continual appearance of new objects in the visual world poses\nconsiderable challenges for current deep learning methods in real-world\ndeployments. The challenge of new task learning is often exacerbated by the\nscarcity of data for the new categories due to rarity or cost. Here we explore\nthe important task of Few-Shot Class-Incremental Learning (FSCIL) and its\nextreme data scarcity condition of one-shot. An ideal FSCIL model needs to\nperform well on all classes, regardless of their presentation order or paucity\nof data. It also needs to be robust to open-set real-world conditions and be\neasily adapted to the new tasks that always arise in the field. In this paper,\nwe first reevaluate the current task setting and propose a more comprehensive\nand practical setting for the FSCIL task. Then, inspired by the similarity of\nthe goals for FSCIL and modern face recognition systems, we propose our method\n-- Augmented Angular Loss Incremental Classification or ALICE. In ALICE,\ninstead of the commonly used cross-entropy loss, we propose to use the angular\npenalty loss to obtain well-clustered features. As the obtained features not\nonly need to be compactly clustered but also diverse enough to maintain\ngeneralization for future incremental classes, we further discuss how class\naugmentation, data augmentation, and data balancing affect classification\nperformance. Experiments on benchmark datasets, including CIFAR100,\nminiImageNet, and CUB200, demonstrate the improved performance of ALICE over\nthe state-of-the-art FSCIL methods.\n", "date": "2022-07-30", "categories": []}
{"id": "2207.10856", "title": "Prototype-Guided Continual Adaptation for Class-Incremental Unsupervised\n  Domain Adaptation", "abstract": "  This paper studies a new, practical but challenging problem, called\nClass-Incremental Unsupervised Domain Adaptation (CI-UDA), where the labeled\nsource domain contains all classes, but the classes in the unlabeled target\ndomain increase sequentially. This problem is challenging due to two\ndifficulties. First, source and target label sets are inconsistent at each time\nstep, which makes it difficult to conduct accurate domain alignment. Second,\nprevious target classes are unavailable in the current step, resulting in the\nforgetting of previous knowledge. To address this problem, we propose a novel\nPrototype-guided Continual Adaptation (ProCA) method, consisting of two\nsolution strategies. 1) Label prototype identification: we identify target\nlabel prototypes by detecting shared classes with cumulative prediction\nprobabilities of target samples. 2) Prototype-based alignment and replay: based\non the identified label prototypes, we align both domains and enforce the model\nto retain previous knowledge. With these two strategies, ProCA is able to adapt\nthe source model to a class-incremental unlabeled target domain effectively.\nExtensive experiments demonstrate the effectiveness and superiority of ProCA in\nresolving CI-UDA. The source code is available at\nhttps://github.com/Hongbin98/ProCA.git\n", "date": "2022-07-22", "categories": []}
{"id": "2207.14202", "title": "Progressive Voronoi Diagram Subdivision: Towards A Holistic Geometric\n  Framework for Exemplar-free Class-Incremental Learning", "abstract": "  Exemplar-free Class-incremental Learning (CIL) is a challenging problem\nbecause rehearsing data from previous phases is strictly prohibited, causing\ncatastrophic forgetting of Deep Neural Networks (DNNs). In this paper, we\npresent iVoro, a holistic framework for CIL, derived from computational\ngeometry. We found Voronoi Diagram (VD), a classical model for space\nsubdivision, is especially powerful for solving the CIL problem, because VD\nitself can be constructed favorably in an incremental manner -- the newly added\nsites (classes) will only affect the proximate classes, making the\nnon-contiguous classes hardly forgettable. Further, in order to find a better\nset of centers for VD construction, we colligate DNN with VD using Power\nDiagram and show that the VD structure can be optimized by integrating local\nDNN models using a divide-and-conquer algorithm. Moreover, our VD construction\nis not restricted to the deep feature space, but is also applicable to multiple\nintermediate feature spaces, promoting VD to be multi-centered VD (CIVD) that\nefficiently captures multi-grained features from DNN. Importantly, iVoro is\nalso capable of handling uncertainty-aware test-time Voronoi cell assignment\nand has exhibited high correlations between geometric uncertainty and\npredictive accuracy (up to ~0.9). Putting everything together, iVoro achieves\nup to 25.26%, 37.09%, and 33.21% improvements on CIFAR-100, TinyImageNet, and\nImageNet-Subset, respectively, compared to the state-of-the-art non-exemplar\nCIL approaches. In conclusion, iVoro enables highly accurate,\nprivacy-preserving, and geometrically interpretable CIL that is particularly\nuseful when cross-phase data sharing is forbidden, e.g. in medical\napplications. Our code is available at https://machunwei.github.io/ivoro.\n", "date": "2022-07-28", "categories": []}
{"id": "2207.13516", "title": "Online Continual Learning with Contrastive Vision Transformer", "abstract": "  Online continual learning (online CL) studies the problem of learning\nsequential tasks from an online data stream without task boundaries, aiming to\nadapt to new data while alleviating catastrophic forgetting on the past tasks.\nThis paper proposes a framework Contrastive Vision Transformer (CVT), which\ndesigns a focal contrastive learning strategy based on a transformer\narchitecture, to achieve a better stability-plasticity trade-off for online CL.\nSpecifically, we design a new external attention mechanism for online CL that\nimplicitly captures previous tasks' information. Besides, CVT contains\nlearnable focuses for each class, which could accumulate the knowledge of\nprevious classes to alleviate forgetting. Based on the learnable focuses, we\ndesign a focal contrastive loss to rebalance contrastive learning between new\nand past classes and consolidate previously learned representations. Moreover,\nCVT contains a dual-classifier structure for decoupling learning current\nclasses and balancing all observed classes. The extensive experimental results\nshow that our approach achieves state-of-the-art performance with even fewer\nparameters on online CL benchmarks and effectively alleviates the catastrophic\nforgetting.\n", "date": "2022-07-24", "categories": []}
{"id": "2206.00388", "title": "Transfer without Forgetting", "abstract": "  This work investigates the entanglement between Continual Learning (CL) and\nTransfer Learning (TL). In particular, we shed light on the widespread\napplication of network pretraining, highlighting that it is itself subject to\ncatastrophic forgetting. Unfortunately, this issue leads to the\nunder-exploitation of knowledge transfer during later tasks. On this ground, we\npropose Transfer without Forgetting (TwF), a hybrid approach building upon a\nfixed pretrained sibling network, which continuously propagates the knowledge\ninherent in the source domain through a layer-wise loss term. Our experiments\nindicate that TwF steadily outperforms other CL methods across a variety of\nsettings, averaging a 4.81% gain in Class-Incremental accuracy over a variety\nof datasets and different buffer sizes.\n", "date": "2022-06-01", "categories": []}
{"id": "2207.11436", "title": "Facing Changes: Continual Entity Alignment for Growing Knowledge Graphs", "abstract": "  Entity alignment is a basic and vital technique in knowledge graph (KG)\nintegration. Over the years, research on entity alignment has resided on the\nassumption that KGs are static, which neglects the nature of growth of\nreal-world KGs. As KGs grow, previous alignment results face the need to be\nrevisited while new entity alignment waits to be discovered. In this paper, we\npropose and dive into a realistic yet unexplored setting, referred to as\ncontinual entity alignment. To avoid retraining an entire model on the whole\nKGs whenever new entities and triples come, we present a continual alignment\nmethod for this task. It reconstructs an entity's representation based on\nentity adjacency, enabling it to generate embeddings for new entities quickly\nand inductively using their existing neighbors. It selects and replays partial\npre-aligned entity pairs to train only parts of KGs while extracting\ntrustworthy alignment for knowledge augmentation. As growing KGs inevitably\ncontain non-matchable entities, different from previous works, the proposed\nmethod employs bidirectional nearest neighbor matching to find new entity\nalignment and update old alignment. Furthermore, we also construct new datasets\nby simulating the growth of multilingual DBpedia. Extensive experiments\ndemonstrate that our continual alignment method is more effective than\nbaselines based on retraining or inductive learning.\n", "date": "2022-07-23", "categories": []}
{"id": "2207.11577", "title": "Augmented Bilinear Network for Incremental Multi-Stock Time-Series\n  Classification", "abstract": "  Deep Learning models have become dominant in tackling financial time-series\nanalysis problems, overturning conventional machine learning and statistical\nmethods. Most often, a model trained for one market or security cannot be\ndirectly applied to another market or security due to differences inherent in\nthe market conditions. In addition, as the market evolves through time, it is\nnecessary to update the existing models or train new ones when new data is made\navailable. This scenario, which is inherent in most financial forecasting\napplications, naturally raises the following research question: How to\nefficiently adapt a pre-trained model to a new set of data while retaining\nperformance on the old data, especially when the old data is not accessible? In\nthis paper, we propose a method to efficiently retain the knowledge available\nin a neural network pre-trained on a set of securities and adapt it to achieve\nhigh performance in new ones. In our method, the prior knowledge encoded in a\npre-trained neural network is maintained by keeping existing connections fixed,\nand this knowledge is adjusted for the new securities by a set of augmented\nconnections, which are optimized using the new data. The auxiliary connections\nare constrained to be of low rank. This not only allows us to rapidly optimize\nfor the new task but also reduces the storage and run-time complexity during\nthe deployment phase. The efficiency of our approach is empirically validated\nin the stock mid-price movement prediction problem using a large-scale limit\norder book dataset. Experimental results show that our approach enhances\nprediction performance as well as reduces the overall number of network\nparameters.\n", "date": "2022-07-23", "categories": []}
{"id": "2207.12061", "title": "Balancing Stability and Plasticity through Advanced Null Space in\n  Continual Learning", "abstract": "  Continual learning is a learning paradigm that learns tasks sequentially with\nresources constraints, in which the key challenge is stability-plasticity\ndilemma, i.e., it is uneasy to simultaneously have the stability to prevent\ncatastrophic forgetting of old tasks and the plasticity to learn new tasks\nwell. In this paper, we propose a new continual learning approach, Advanced\nNull Space (AdNS), to balance the stability and plasticity without storing any\nold data of previous tasks. Specifically, to obtain better stability, AdNS\nmakes use of low-rank approximation to obtain a novel null space and projects\nthe gradient onto the null space to prevent the interference on the past tasks.\nTo control the generation of the null space, we introduce a non-uniform\nconstraint strength to further reduce forgetting. Furthermore, we present a\nsimple but effective method, intra-task distillation, to improve the\nperformance of the current task. Finally, we theoretically find that null space\nplays a key role in plasticity and stability, respectively. Experimental\nresults show that the proposed method can achieve better performance compared\nto state-of-the-art continual learning approaches.\n", "date": "2022-07-25", "categories": []}
{"id": "2207.12256", "title": "Lifelong Machine Learning of Functionally Compositional Structures", "abstract": "  A hallmark of human intelligence is the ability to construct self-contained\nchunks of knowledge and reuse them in novel combinations for solving different\nproblems. Learning such compositional structures has been a challenge for\nartificial systems, due to the underlying combinatorial search. To date,\nresearch into compositional learning has largely proceeded separately from work\non lifelong or continual learning. This dissertation integrated these two lines\nof work to present a general-purpose framework for lifelong learning of\nfunctionally compositional structures. The framework separates the learning\ninto two stages: learning how to combine existing components to assimilate a\nnovel problem, and learning how to adapt the existing components to accommodate\nthe new problem. This separation explicitly handles the trade-off between\nstability and flexibility. This dissertation instantiated the framework into\nvarious supervised and reinforcement learning (RL) algorithms. Supervised\nlearning evaluations found that 1) compositional models improve lifelong\nlearning of diverse tasks, 2) the multi-stage process permits lifelong learning\nof compositional knowledge, and 3) the components learned by the framework\nrepresent self-contained and reusable functions. Similar RL evaluations\ndemonstrated that 1) algorithms under the framework accelerate the discovery of\nhigh-performing policies, and 2) these algorithms retain or improve performance\non previously learned tasks. The dissertation extended one lifelong\ncompositional RL algorithm to the nonstationary setting, where the task\ndistribution varies over time, and found that modularity permits individually\ntracking changes to different elements in the environment. The final\ncontribution of this dissertation was a new benchmark for compositional RL,\nwhich exposed that existing methods struggle to discover the compositional\nproperties of the environment.\n", "date": "2022-07-25", "categories": []}
{"id": "2207.12303", "title": "Continual Few-Shot Learning with Adversarial Class Storage", "abstract": "  Humans have a remarkable ability to quickly and effectively learn new\nconcepts in a continuous manner without forgetting old knowledge. Though deep\nlearning has made tremendous successes on various computer vision tasks, it\nfaces challenges for achieving such human-level intelligence. In this paper, we\ndefine a new problem called continual few-shot learning, in which tasks arrive\nsequentially and each task is associated with a few training samples. We\npropose Continual Meta-Learner (CML) to solve this problem. CML integrates\nmetric-based classification and a memory-based mechanism along with adversarial\nlearning into a meta-learning framework, which leads to the desirable\nproperties: 1) it can quickly and effectively learn to handle a new task; 2) it\novercomes catastrophic forgetting; 3) it is model-agnostic. We conduct\nextensive experiments on two image datasets, MiniImageNet and CIFAR100.\nExperimental results show that CML delivers state-of-the-art performance in\nterms of classification accuracy on few-shot learning tasks without\ncatastrophic forgetting.\n", "date": "2022-07-10", "categories": []}
{"id": "2207.11213", "title": "Few-Shot Class-Incremental Learning via Entropy-Regularized Data-Free\n  Replay", "abstract": "  Few-shot class-incremental learning (FSCIL) has been proposed aiming to\nenable a deep learning system to incrementally learn new classes with limited\ndata. Recently, a pioneer claims that the commonly used replay-based method in\nclass-incremental learning (CIL) is ineffective and thus not preferred for\nFSCIL. This has, if truth, a significant influence on the fields of FSCIL. In\nthis paper, we show through empirical results that adopting the data replay is\nsurprisingly favorable. However, storing and replaying old data can lead to a\nprivacy concern. To address this issue, we alternatively propose using\ndata-free replay that can synthesize data by a generator without accessing real\ndata. In observing the the effectiveness of uncertain data for knowledge\ndistillation, we impose entropy regularization in the generator training to\nencourage more uncertain examples. Moreover, we propose to relabel the\ngenerated data with one-hot-like labels. This modification allows the network\nto learn by solely minimizing the cross-entropy loss, which mitigates the\nproblem of balancing different objectives in the conventional knowledge\ndistillation approach. Finally, we show extensive experimental results and\nanalysis on CIFAR-100, miniImageNet and CUB-200 to demonstrate the\neffectiveness of our proposed one.\n", "date": "2022-07-22", "categories": []}
{"id": "2206.06957", "title": "Continual-Learning-as-a-Service (CLaaS): On-Demand Efficient Adaptation\n  of Predictive Models", "abstract": "  Predictive machine learning models nowadays are often updated in a stateless\nand expensive way. The two main future trends for companies that want to build\nmachine learning-based applications and systems are real-time inference and\ncontinual updating. Unfortunately, both trends require a mature infrastructure\nthat is hard and costly to realize on-premise. This paper defines a novel\nsoftware service and model delivery infrastructure termed Continual\nLearning-as-a-Service (CLaaS) to address these issues. Specifically, it\nembraces continual machine learning and continuous integration techniques. It\nprovides support for model updating and validation tools for data scientists\nwithout an on-premise solution and in an efficient, stateful and easy-to-use\nmanner. Finally, this CL model service is easy to encapsulate in any machine\nlearning infrastructure or cloud system. This paper presents the design and\nimplementation of a CLaaS instantiation, called LiquidBrain, evaluated in two\nreal-world scenarios. The former is a robotic object recognition setting using\nthe CORe50 dataset while the latter is a named category and attribute\nprediction using the DeepFashion-C dataset in the fashion domain. Our\npreliminary results suggest the usability and efficiency of the Continual\nLearning model services and the effectiveness of the solution in addressing\nreal-world use-cases regardless of where the computation happens in the\ncontinuum Edge-Cloud.\n", "date": "2022-06-14", "categories": []}
{"id": "2207.10131", "title": "Continual Variational Autoencoder Learning via Online Cooperative\n  Memorization", "abstract": "  Due to their inference, data representation and reconstruction properties,\nVariational Autoencoders (VAE) have been successfully used in continual\nlearning classification tasks. However, their ability to generate images with\nspecifications corresponding to the classes and databases learned during\nContinual Learning (CL) is not well understood and catastrophic forgetting\nremains a significant challenge. In this paper, we firstly analyze the\nforgetting behaviour of VAEs by developing a new theoretical framework that\nformulates CL as a dynamic optimal transport problem. This framework proves\napproximate bounds to the data likelihood without requiring the task\ninformation and explains how the prior knowledge is lost during the training\nprocess. We then propose a novel memory buffering approach, namely the Online\nCooperative Memorization (OCM) framework, which consists of a Short-Term Memory\n(STM) that continually stores recent samples to provide future information for\nthe model, and a Long-Term Memory (LTM) aiming to preserve a wide diversity of\nsamples. The proposed OCM transfers certain samples from STM to LTM according\nto the information diversity selection criterion without requiring any\nsupervised signals. The OCM framework is then combined with a dynamic VAE\nexpansion mixture network for further enhancing its performance.\n", "date": "2022-07-20", "categories": []}
{"id": "2207.10659", "title": "Novel Class Discovery without Forgetting", "abstract": "  Humans possess an innate ability to identify and differentiate instances that\nthey are not familiar with, by leveraging and adapting the knowledge that they\nhave acquired so far. Importantly, they achieve this without deteriorating the\nperformance on their earlier learning. Inspired by this, we identify and\nformulate a new, pragmatic problem setting of NCDwF: Novel Class Discovery\nwithout Forgetting, which tasks a machine learning model to incrementally\ndiscover novel categories of instances from unlabeled data, while maintaining\nits performance on the previously seen categories. We propose 1) a method to\ngenerate pseudo-latent representations which act as a proxy for (no longer\navailable) labeled data, thereby alleviating forgetting, 2) a\nmutual-information based regularizer which enhances unsupervised discovery of\nnovel classes, and 3) a simple Known Class Identifier which aids generalized\ninference when the testing data contains instances form both seen and unseen\ncategories. We introduce experimental protocols based on CIFAR-10, CIFAR-100\nand ImageNet-1000 to measure the trade-off between knowledge retention and\nnovel class discovery. Our extensive evaluations reveal that existing models\ncatastrophically forget previously seen categories while identifying novel\ncategories, while our method is able to effectively balance between the\ncompeting objectives. We hope our work will attract further research into this\nnewly identified pragmatic problem setting.\n", "date": "2022-07-21", "categories": []}
{"id": "2207.09248", "title": "Don't Stop Learning: Towards Continual Learning for the CLIP Model", "abstract": "  The Contrastive Language-Image Pre-training (CLIP) Model is a recently\nproposed large-scale pre-train model which attracts increasing attention in the\ncomputer vision community. Benefiting from its gigantic image-text training\nset, the CLIP model has learned outstanding capabilities in zero-shot learning\nand image-text matching. To boost the recognition performance of CLIP on some\ntarget visual concepts, it is often desirable to further update the CLIP model\nby fine-tuning some classes-of-interest on extra training data. This operation,\nhowever, raises an important concern: will the update hurt the zero-shot\nlearning or image-text matching capability of the CLIP, i.e., the catastrophic\nforgetting issue? If yes, could existing continual learning algorithms be\nadapted to alleviate the risk of catastrophic forgetting? To answer these\nquestions, this work conducts a systemic study on the continual learning issue\nof the CLIP model. We construct evaluation protocols to measure the impact of\nfine-tuning updates and explore different ways to upgrade existing continual\nlearning methods to mitigate the forgetting issue of the CLIP model. Our study\nreveals the particular challenges of CLIP continual learning problem and lays a\nfoundation for further researches. Moreover, we propose a new algorithm, dubbed\nLearning without Forgetting via Replayed Vocabulary (VR-LwF), which shows exact\neffectiveness for alleviating the forgetting issue of the CLIP model.\n", "date": "2022-07-19", "categories": []}
{"id": "2207.09963", "title": "Rethinking Few-Shot Class-Incremental Learning with Open-Set Hypothesis\n  in Hyperbolic Geometry", "abstract": "  Few-Shot Class-Incremental Learning (FSCIL) aims at incrementally learning\nnovel classes from a few labeled samples by avoiding the overfitting and\ncatastrophic forgetting simultaneously. The current protocol of FSCIL is built\nby mimicking the general class-incremental learning setting, while it is not\ntotally appropriate due to the different data configuration, i.e., novel\nclasses are all in the limited data regime. In this paper, we rethink the\nconfiguration of FSCIL with the open-set hypothesis by reserving the\npossibility in the first session for incoming categories. To assign better\nperformances on both close-set and open-set recognition to the model,\nHyperbolic Reciprocal Point Learning module (Hyper-RPL) is built on Reciprocal\nPoint Learning (RPL) with hyperbolic neural networks. Besides, for learning\nnovel categories from limited labeled data, we incorporate a hyperbolic metric\nlearning (Hyper-Metric) module into the distillation-based framework to\nalleviate the overfitting issue and better handle the trade-off issue between\nthe preservation of old knowledge and the acquisition of new knowledge. The\ncomprehensive assessments of the proposed configuration and modules on three\nbenchmark datasets are executed to validate the effectiveness concerning three\nevaluation indicators.\n", "date": "2022-07-20", "categories": []}
{"id": "2207.09074", "title": "Incremental Task Learning with Incremental Rank Updates", "abstract": "  Incremental Task learning (ITL) is a category of continual learning that\nseeks to train a single network for multiple tasks (one after another), where\ntraining data for each task is only available during the training of that task.\nNeural networks tend to forget older tasks when they are trained for the newer\ntasks; this property is often known as catastrophic forgetting. To address this\nissue, ITL methods use episodic memory, parameter regularization, masking and\npruning, or extensible network structures. In this paper, we propose a new\nincremental task learning framework based on low-rank factorization. In\nparticular, we represent the network weights for each layer as a linear\ncombination of several rank-1 matrices. To update the network for a new task,\nwe learn a rank-1 (or low-rank) matrix and add that to the weights of every\nlayer. We also introduce an additional selector vector that assigns different\nweights to the low-rank matrices learned for the previous tasks. We show that\nour approach performs better than the current state-of-the-art methods in terms\nof accuracy and forgetting. Our method also offers better memory efficiency\ncompared to episodic memory- and mask-based approaches. Our code will be\navailable at https://github.com/CSIPlab/task-increment-rank-update.git\n", "date": "2022-07-19", "categories": []}
{"id": "2207.07429", "title": "Continual Learning For On-Device Environmental Sound Classification", "abstract": "  Continuously learning new classes without catastrophic forgetting is a\nchallenging problem for on-device environmental sound classification given the\nrestrictions on computation resources (e.g., model size, running memory). To\naddress this issue, we propose a simple and efficient continual learning\nmethod. Our method selects the historical data for the training by measuring\nthe per-sample classification uncertainty. Specifically, we measure the\nuncertainty by observing how the classification probability of data fluctuates\nagainst the parallel perturbations added to the classifier embedding. In this\nway, the computation cost can be significantly reduced compared with adding\nperturbation to the raw data. Experimental results on the DCASE 2019 Task 1 and\nESC-50 dataset show that our proposed method outperforms baseline continual\nlearning methods on classification accuracy and computational efficiency,\nindicating our method can efficiently and incrementally learn new classes\nwithout the catastrophic forgetting problem for on-device environmental sound\nclassification.\n", "date": "2022-07-15", "categories": []}
{"id": "2207.07840", "title": "Class-Incremental Lifelong Learning in Multi-Label Classification", "abstract": "  Existing class-incremental lifelong learning studies only the data is with\nsingle-label, which limits its adaptation to multi-label data. This paper\nstudies Lifelong Multi-Label (LML) classification, which builds an online\nclass-incremental classifier in a sequential multi-label classification data\nstream. Training on the data with Partial Labels in LML classification may\nresult in more serious Catastrophic Forgetting in old classes. To solve the\nproblem, the study proposes an Augmented Graph Convolutional Network (AGCN)\nwith a built Augmented Correlation Matrix (ACM) across sequential partial-label\ntasks. The results of two benchmarks show that the method is effective for LML\nclassification and reducing forgetting.\n", "date": "2022-07-16", "categories": []}
{"id": "2207.08181", "title": "Federated Continual Learning through distillation in pervasive computing", "abstract": "  Federated Learning has been introduced as a new machine learning paradigm\nenhancing the use of local devices. At a server level, FL regularly aggregates\nmodels learned locally on distributed clients to obtain a more general model.\nCurrent solutions rely on the availability of large amounts of stored data at\nthe client side in order to fine-tune the models sent by the server. Such\nsetting is not realistic in mobile pervasive computing where data storage must\nbe kept low and data characteristic can change dramatically. To account for\nthis variability, a solution is to use the data regularly collected by the\nclient to progressively adapt the received model. But such naive approach\nexposes clients to the well-known problem of catastrophic forgetting. To\naddress this problem, we have defined a Federated Continual Learning approach\nwhich is mainly based on distillation. Our approach allows a better use of\nresources, eliminating the need to retrain from scratch at the arrival of new\ndata and reducing memory usage by limiting the amount of data to be stored.\nThis proposal has been evaluated in the Human Activity Recognition (HAR) domain\nand has shown to effectively reduce the catastrophic forgetting effect.\n", "date": "2022-07-17", "categories": []}
{"id": "2207.08224", "title": "Learning with Recoverable Forgetting", "abstract": "  Life-long learning aims at learning a sequence of tasks without forgetting\nthe previously acquired knowledge. However, the involved training data may not\nbe life-long legitimate due to privacy or copyright reasons. In practical\nscenarios, for instance, the model owner may wish to enable or disable the\nknowledge of specific tasks or specific samples from time to time. Such\nflexible control over knowledge transfer, unfortunately, has been largely\noverlooked in previous incremental or decremental learning methods, even at a\nproblem-setup level. In this paper, we explore a novel learning scheme, termed\nas Learning wIth Recoverable Forgetting (LIRF), that explicitly handles the\ntask- or sample-specific knowledge removal and recovery. Specifically, LIRF\nbrings in two innovative schemes, namely knowledge deposit and withdrawal,\nwhich allow for isolating user-designated knowledge from a pre-trained network\nand injecting it back when necessary. During the knowledge deposit process, the\nspecified knowledge is extracted from the target network and stored in a\ndeposit module, while the insensitive or general knowledge of the target\nnetwork is preserved and further augmented. During knowledge withdrawal, the\ntaken-off knowledge is added back to the target network. The deposit and\nwithdraw processes only demand for a few epochs of finetuning on the removal\ndata, ensuring both data and time efficiency. We conduct experiments on several\ndatasets, and demonstrate that the proposed LIRF strategy yields encouraging\nresults with gratifying generalization capability.\n", "date": "2022-07-17", "categories": []}
{"id": "2207.08309", "title": "CULT: Continual Unsupervised Learning with Typicality-Based Environment\n  Detection", "abstract": "  We introduce CULT (Continual Unsupervised Representation Learning with\nTypicality-Based Environment Detection), a new algorithm for continual\nunsupervised learning with variational auto-encoders. CULT uses a simple\ntypicality metric in the latent space of a VAE to detect distributional shifts\nin the environment, which is used in conjunction with generative replay and an\nauxiliary environmental classifier to limit catastrophic forgetting in\nunsupervised representation learning. In our experiments, CULT significantly\noutperforms baseline continual unsupervised learning approaches. Code for this\npaper can be found here: https://github.com/oliveradk/cult\n", "date": "2022-07-17", "categories": []}
{"id": "2207.08605", "title": "Class-incremental Novel Class Discovery", "abstract": "  We study the new task of class-incremental Novel Class Discovery\n(class-iNCD), which refers to the problem of discovering novel categories in an\nunlabelled data set by leveraging a pre-trained model that has been trained on\na labelled data set containing disjoint yet related categories. Apart from\ndiscovering novel classes, we also aim at preserving the ability of the model\nto recognize previously seen base categories. Inspired by rehearsal-based\nincremental learning methods, in this paper we propose a novel approach for\nclass-iNCD which prevents forgetting of past information about the base classes\nby jointly exploiting base class feature prototypes and feature-level knowledge\ndistillation. We also propose a self-training clustering strategy that\nsimultaneously clusters novel categories and trains a joint classifier for both\nthe base and novel classes. This makes our method able to operate in a\nclass-incremental setting. Our experiments, conducted on three common\nbenchmarks, demonstrate that our method significantly outperforms\nstate-of-the-art approaches. Code is available at\nhttps://github.com/OatmealLiu/class-iNCD\n", "date": "2022-07-18", "categories": []}
{"id": "2207.06475", "title": "Analysis of Catastrophic Forgetting for Random Orthogonal Transformation\n  Tasks in the Overparameterized Regime", "abstract": "  Overparameterization is known to permit strong generalization performance in\nneural networks. In this work, we provide an initial theoretical analysis of\nits effect on catastrophic forgetting in a continual learning setup. We show\nexperimentally that in permuted MNIST image classification tasks, the\ngeneralization performance of multilayer perceptrons trained by vanilla\nstochastic gradient descent can be improved by overparameterization, and the\nextent of the performance increase achieved by overparameterization is\ncomparable to that of state-of-the-art continual learning algorithms. We\nprovide a theoretical explanation of this effect by studying a qualitatively\nsimilar two-task linear regression problem, where each task is related by a\nrandom orthogonal transformation. We show that when a model is trained on the\ntwo tasks in sequence without any additional regularization, the risk gain on\nthe first task is small if the model is sufficiently overparameterized.\n", "date": "2022-06-01", "categories": []}
{"id": "2207.06543", "title": "CoSCL: Cooperation of Small Continual Learners is Stronger than a Big\n  One", "abstract": "  Continual learning requires incremental compatibility with a sequence of\ntasks. However, the design of model architecture remains an open question: In\ngeneral, learning all tasks with a shared set of parameters suffers from severe\ninterference between tasks; while learning each task with a dedicated parameter\nsubspace is limited by scalability. In this work, we theoretically analyze the\ngeneralization errors for learning plasticity and memory stability in continual\nlearning, which can be uniformly upper-bounded by (1) discrepancy between task\ndistributions, (2) flatness of loss landscape and (3) cover of parameter space.\nThen, inspired by the robust biological learning system that processes\nsequential experiences with multiple parallel compartments, we propose\nCooperation of Small Continual Learners (CoSCL) as a general strategy for\ncontinual learning. Specifically, we present an architecture with a fixed\nnumber of narrower sub-networks to learn all incremental tasks in parallel,\nwhich can naturally reduce the two errors through improving the three\ncomponents of the upper bound. To strengthen this advantage, we encourage to\ncooperate these sub-networks by penalizing the difference of predictions made\nby their feature representations. With a fixed parameter budget, CoSCL can\nimprove a variety of representative continual learning approaches by a large\nmargin (e.g., up to 10.64% on CIFAR-100-SC, 9.33% on CIFAR-100-RS, 11.45% on\nCUB-200-2011 and 6.72% on Tiny-ImageNet) and achieve the new state-of-the-art\nperformance.\n", "date": "2022-07-13", "categories": []}
{"id": "2207.06754", "title": "E2-AEN: End-to-End Incremental Learning with Adaptively Expandable\n  Network", "abstract": "  Expandable networks have demonstrated their advantages in dealing with\ncatastrophic forgetting problem in incremental learning. Considering that\ndifferent tasks may need different structures, recent methods design dynamic\nstructures adapted to different tasks via sophisticated skills. Their routine\nis to search expandable structures first and then train on the new tasks,\nwhich, however, breaks tasks into multiple training stages, leading to\nsuboptimal or overmuch computational cost. In this paper, we propose an\nend-to-end trainable adaptively expandable network named E2-AEN, which\ndynamically generates lightweight structures for new tasks without any accuracy\ndrop in previous tasks. Specifically, the network contains a serial of powerful\nfeature adapters for augmenting the previously learned representations to new\ntasks, and avoiding task interference. These adapters are controlled via an\nadaptive gate-based pruning strategy which decides whether the expanded\nstructures can be pruned, making the network structure dynamically changeable\naccording to the complexity of the new tasks. Moreover, we introduce a novel\nsparsity-activation regularization to encourage the model to learn\ndiscriminative features with limited parameters. E2-AEN reduces cost and can be\nbuilt upon any feed-forward architectures in an end-to-end manner. Extensive\nexperiments on both classification (i.e., CIFAR and VDD) and detection (i.e.,\nCOCO, VOC and ICCV2021 SSLAD challenge) benchmarks demonstrate the\neffectiveness of the proposed method, which achieves the new remarkable\nresults.\n", "date": "2022-07-14", "categories": []}
{"id": "2206.15189", "title": "Multi-Granularity Regularized Re-Balancing for Class Incremental\n  Learning", "abstract": "  Deep learning models suffer from catastrophic forgetting when learning new\ntasks incrementally. Incremental learning has been proposed to retain the\nknowledge of old classes while learning to identify new classes. A typical\napproach is to use a few exemplars to avoid forgetting old knowledge. In such a\nscenario, data imbalance between old and new classes is a key issue that leads\nto performance degradation of the model. Several strategies have been designed\nto rectify the bias towards the new classes due to data imbalance. However,\nthey heavily rely on the assumptions of the bias relation between old and new\nclasses. Therefore, they are not suitable for complex real-world applications.\nIn this study, we propose an assumption-agnostic method, Multi-Granularity\nRegularized re-Balancing (MGRB), to address this problem. Re-balancing methods\nare used to alleviate the influence of data imbalance; however, we empirically\ndiscover that they would under-fit new classes. To this end, we further design\na novel multi-granularity regularization term that enables the model to\nconsider the correlations of classes in addition to re-balancing the data. A\nclass hierarchy is first constructed by grouping the semantically or visually\nsimilar classes. The multi-granularity regularization then transforms the\none-hot label vector into a continuous label distribution, which reflects the\nrelations between the target class and other classes based on the constructed\nclass hierarchy. Thus, the model can learn the inter-class relational\ninformation, which helps enhance the learning of both old and new classes.\nExperimental results on both public datasets and a real-world fault diagnosis\ndataset verify the effectiveness of the proposed method.\n", "date": "2022-06-30", "categories": []}
{"id": "2207.05897", "title": "D-CBRS: Accounting For Intra-Class Diversity in Continual Learning", "abstract": "  Continual learning -- accumulating knowledge from a sequence of learning\nexperiences -- is an important yet challenging problem. In this paradigm, the\nmodel's performance for previously encountered instances may substantially drop\nas additional data are seen. When dealing with class-imbalanced data,\nforgetting is further exacerbated. Prior work has proposed replay-based\napproaches which aim at reducing forgetting by intelligently storing instances\nfor future replay. Although Class-Balancing Reservoir Sampling (CBRS) has been\nsuccessful in dealing with imbalanced data, the intra-class diversity has not\nbeen accounted for, implicitly assuming that each instance of a class is\nequally informative. We present Diverse-CBRS (D-CBRS), an algorithm that allows\nus to consider within class diversity when storing instances in the memory. Our\nresults show that D-CBRS outperforms state-of-the-art memory management\ncontinual learning algorithms on data sets with considerable intra-class\ndiversity.\n", "date": "2022-07-13", "categories": []}
{"id": "2207.06233", "title": "Continual Learning with Deep Learning Methods in an Application-Oriented\n  Context", "abstract": "  Abstract knowledge is deeply grounded in many computer-based applications. An\nimportant research area of Artificial Intelligence (AI) deals with the\nautomatic derivation of knowledge from data. Machine learning offers the\naccording algorithms. One area of research focuses on the development of\nbiologically inspired learning algorithms. The respective machine learning\nmethods are based on neurological concepts so that they can systematically\nderive knowledge from data and store it. One type of machine learning\nalgorithms that can be categorized as \"deep learning\" model is referred to as\nDeep Neural Networks (DNNs). DNNs consist of multiple artificial neurons\narranged in layers that are trained by using the backpropagation algorithm.\nThese deep learning methods exhibit amazing capabilities for inferring and\nstoring complex knowledge from high-dimensional data. However, DNNs are\naffected by a problem that prevents new knowledge from being added to an\nexisting base. The ability to continuously accumulate knowledge is an important\nfactor that contributed to evolution and is therefore a prerequisite for the\ndevelopment of strong AIs. The so-called \"catastrophic forgetting\" (CF) effect\ncauses DNNs to immediately loose already derived knowledge after a few training\niterations on a new data distribution. Only an energetically expensive\nretraining with the joint data distribution of past and new data enables the\nabstraction of the entire new set of knowledge. In order to counteract the\neffect, various techniques have been and are still being developed with the\ngoal to mitigate or even solve the CF problem. These published CF avoidance\nstudies usually imply the effectiveness of their approaches for various\ncontinual learning tasks. This dissertation is set in the context of continual\nmachine learning with deep learning methods. The first part deals with the\ndevelopment of an ...\n", "date": "2022-07-12", "categories": []}
{"id": "2207.06267", "title": "Task Agnostic Representation Consolidation: a Self-supervised based\n  Continual Learning Approach", "abstract": "  Continual learning (CL) over non-stationary data streams remains one of the\nlong-standing challenges in deep neural networks (DNNs) as they are prone to\ncatastrophic forgetting. CL models can benefit from self-supervised\npre-training as it enables learning more generalizable task-agnostic features.\nHowever, the effect of self-supervised pre-training diminishes as the length of\ntask sequences increases. Furthermore, the domain shift between pre-training\ndata distribution and the task distribution reduces the generalizability of the\nlearned representations. To address these limitations, we propose Task Agnostic\nRepresentation Consolidation (TARC), a two-stage training paradigm for CL that\nintertwines task-agnostic and task-specific learning whereby self-supervised\ntraining is followed by supervised learning for each task. To further restrict\nthe deviation from the learned representations in the self-supervised stage, we\nemploy a task-agnostic auxiliary loss during the supervised stage. We show that\nour training paradigm can be easily added to memory- or regularization-based\napproaches and provides consistent performance gain across more challenging CL\nsettings. We further show that it leads to more robust and well-calibrated\nmodels.\n", "date": "2022-07-13", "categories": []}
{"id": "2207.04354", "title": "An Introduction to Lifelong Supervised Learning", "abstract": "  This primer is an attempt to provide a detailed summary of the different\nfacets of lifelong learning. We start with Chapter 2 which provides a\nhigh-level overview of lifelong learning systems. In this chapter, we discuss\nprominent scenarios in lifelong learning (Section 2.4), provide 8 Introduction\na high-level organization of different lifelong learning approaches (Section\n2.5), enumerate the desiderata for an ideal lifelong learning system (Section\n2.6), discuss how lifelong learning is related to other learning paradigms\n(Section 2.7), describe common metrics used to evaluate lifelong learning\nsystems (Section 2.8). This chapter is more useful for readers who are new to\nlifelong learning and want to get introduced to the field without focusing on\nspecific approaches or benchmarks. The remaining chapters focus on specific\naspects (either learning algorithms or benchmarks) and are more useful for\nreaders who are looking for specific approaches or benchmarks. Chapter 3\nfocuses on regularization-based approaches that do not assume access to any\ndata from previous tasks. Chapter 4 discusses memory-based approaches that\ntypically use a replay buffer or an episodic memory to save subset of data\nacross different tasks. Chapter 5 focuses on different architecture families\n(and their instantiations) that have been proposed for training lifelong\nlearning systems. Following these different classes of learning algorithms, we\ndiscuss the commonly used evaluation benchmarks and metrics for lifelong\nlearning (Chapter 6) and wrap up with a discussion of future challenges and\nimportant research directions in Chapter 7.\n", "date": "2022-07-10", "categories": []}
{"id": "2207.05071", "title": "Online Continual Learning of End-to-End Speech Recognition Models", "abstract": "  Continual Learning, also known as Lifelong Learning, aims to continually\nlearn from new data as it becomes available. While prior research on continual\nlearning in automatic speech recognition has focused on the adaptation of\nmodels across multiple different speech recognition tasks, in this paper we\npropose an experimental setting for \\textit{online continual learning} for\nautomatic speech recognition of a single task. Specifically focusing on the\ncase where additional training data for the same task becomes available\nincrementally over time, we demonstrate the effectiveness of performing\nincremental model updates to end-to-end speech recognition models with an\nonline Gradient Episodic Memory (GEM) method. Moreover, we show that with\nonline continual learning and a selective sampling strategy, we can maintain an\naccuracy that is similar to retraining a model from scratch while requiring\nsignificantly lower computation costs. We have also verified our method with\nself-supervised learning (SSL) features.\n", "date": "2022-07-11", "categories": []}
{"id": "2207.05080", "title": "Learning an evolved mixture model for task-free continual learning", "abstract": "  Recently, continual learning (CL) has gained significant interest because it\nenables deep learning models to acquire new knowledge without forgetting\npreviously learnt information. However, most existing works require knowing the\ntask identities and boundaries, which is not realistic in a real context. In\nthis paper, we address a more challenging and realistic setting in CL, namely\nthe Task-Free Continual Learning (TFCL) in which a model is trained on\nnon-stationary data streams with no explicit task information. To address TFCL,\nwe introduce an evolved mixture model whose network architecture is dynamically\nexpanded to adapt to the data distribution shift. We implement this expansion\nmechanism by evaluating the probability distance between the knowledge stored\nin each mixture model component and the current memory buffer using the Hilbert\nSchmidt Independence Criterion (HSIC). We further introduce two simple dropout\nmechanisms to selectively remove stored examples in order to avoid memory\noverload while preserving memory diversity. Empirical results demonstrate that\nthe proposed approach achieves excellent performance.\n", "date": "2022-07-11", "categories": []}
{"id": "2207.00010", "title": "Continual Learning for Human State Monitoring", "abstract": "  Continual Learning (CL) on time series data represents a promising but\nunder-studied avenue for real-world applications. We propose two new CL\nbenchmarks for Human State Monitoring. We carefully designed the benchmarks to\nmirror real-world environments in which new subjects are continuously added. We\nconducted an empirical evaluation to assess the ability of popular CL\nstrategies to mitigate forgetting in our benchmarks. Our results show that,\npossibly due to the domain-incremental properties of our benchmarks, forgetting\ncan be easily tackled even with a simple finetuning and that existing\nstrategies struggle in accumulating knowledge over a fixed, held-out, test\nsubject.\n", "date": "2022-06-29", "categories": []}
{"id": "2207.04874", "title": "Hebbian Continual Representation Learning", "abstract": "  Continual Learning aims to bring machine learning into a more realistic\nscenario, where tasks are learned sequentially and the i.i.d. assumption is not\npreserved. Although this setting is natural for biological systems, it proves\nvery difficult for machine learning models such as artificial neural networks.\nTo reduce this performance gap, we investigate the question whether\nbiologically inspired Hebbian learning is useful for tackling continual\nchallenges. In particular, we highlight a realistic and often overlooked\nunsupervised setting, where the learner has to build representations without\nany supervision. By combining sparse neural networks with Hebbian learning\nprinciple, we build a simple yet effective alternative (HebbCL) to typical\nneural network models trained via the gradient descent. Due to Hebbian\nlearning, the network have easily interpretable weights, which might be\nessential in critical application such as security or healthcare. We\ndemonstrate the efficacy of HebbCL in an unsupervised learning setting applied\nto MNIST and Omniglot datasets. We also adapt the algorithm to the supervised\nscenario and obtain promising results in the class-incremental learning.\n", "date": "2022-06-28", "categories": []}
{"id": "2207.04998", "title": "Consistency is the key to further mitigating catastrophic forgetting in\n  continual learning", "abstract": "  Deep neural networks struggle to continually learn multiple sequential tasks\ndue to catastrophic forgetting of previously learned tasks. Rehearsal-based\nmethods which explicitly store previous task samples in the buffer and\ninterleave them with the current task samples have proven to be the most\neffective in mitigating forgetting. However, Experience Replay (ER) does not\nperform well under low-buffer regimes and longer task sequences as its\nperformance is commensurate with the buffer size. Consistency in predictions of\nsoft-targets can assist ER in preserving information pertaining to previous\ntasks better as soft-targets capture the rich similarity structure of the data.\nTherefore, we examine the role of consistency regularization in ER framework\nunder various continual learning scenarios. We also propose to cast consistency\nregularization as a self-supervised pretext task thereby enabling the use of a\nwide variety of self-supervised learning methods as regularizers. While\nsimultaneously enhancing model calibration and robustness to natural\ncorruptions, regularizing consistency in predictions results in lesser\nforgetting across all continual learning scenarios. Among the different\nfamilies of regularizers, we find that stricter consistency constraints\npreserve previous task information in ER better.\n", "date": "2022-07-11", "categories": []}
{"id": "2207.01562", "title": "Progressive Latent Replay for efficient Generative Rehearsal", "abstract": "  We introduce a new method for internal replay that modulates the frequency of\nrehearsal based on the depth of the network. While replay strategies mitigate\nthe effects of catastrophic forgetting in neural networks, recent works on\ngenerative replay show that performing the rehearsal only on the deeper layers\nof the network improves the performance in continual learning. However, the\ngenerative approach introduces additional computational overhead, limiting its\napplications. Motivated by the observation that earlier layers of neural\nnetworks forget less abruptly, we propose to update network layers with varying\nfrequency using intermediate-level features during replay. This reduces the\ncomputational burden by omitting computations for both deeper layers of the\ngenerator and earlier layers of the main model. We name our method Progressive\nLatent Replay and show that it outperforms Internal Replay while using\nsignificantly fewer resources.\n", "date": "2022-07-04", "categories": []}
{"id": "2207.00005", "title": "Class Impression for Data-free Incremental Learning", "abstract": "  Standard deep learning-based classification approaches require collecting all\nsamples from all classes in advance and are trained offline. This paradigm may\nnot be practical in real-world clinical applications, where new classes are\nincrementally introduced through the addition of new data. Class incremental\nlearning is a strategy allowing learning from such data. However, a major\nchallenge is catastrophic forgetting, i.e., performance degradation on previous\nclasses when adapting a trained model to new data. Prior methodologies to\nalleviate this challenge save a portion of training data require perpetual\nstorage of such data that may introduce privacy issues. Here, we propose a\nnovel data-free class incremental learning framework that first synthesizes\ndata from the model trained on previous classes to generate a \\ours.\nSubsequently, it updates the model by combining the synthesized data with new\nclass data. Furthermore, we incorporate a cosine normalized Cross-entropy loss\nto mitigate the adverse effects of the imbalance, a margin loss to increase\nseparation among previous classes and new ones, and an intra-domain contrastive\nloss to generalize the model trained on the synthesized data to real data. We\ncompare our proposed framework with state-of-the-art methods in class\nincremental learning, where we demonstrate improvement in accuracy for the\nclassification of 11,062 echocardiography cine series of patients.\n", "date": "2022-06-26", "categories": []}
{"id": "2207.00693", "title": "Few-shot incremental learning in the context of solar cell quality\n  inspection", "abstract": "  In industry, Deep Neural Networks have shown high defect detection rates\nsurpassing other more traditional manual feature engineering based proposals.\nThis has been achieved mainly through supervised training where a great amount\nof data is required in order to learn good classification models. However, such\namount of data is sometimes hard to obtain in industrial scenarios, as few\ndefective pieces are produced normally. In addition, certain kinds of defects\nare very rare and usually just appear from time to time, which makes the\ngeneration of a proper dataset for training a classification model even harder.\nMoreover, the lack of available data limits the adaptation of inspection models\nto new defect types that appear in production as it might require a model\nretraining in order to incorporate the detects and detect them. In this work,\nwe have explored the technique of weight imprinting in the context of solar\ncell quality inspection where we have trained a network on three base defect\nclasses, and then we have incorporated new defect classes using few samples.\nThe results have shown that this technique allows the network to extend its\nknowledge with regard to defect classes with few samples, which can be\ninteresting for industrial practitioners.\n", "date": "2022-07-01", "categories": []}
{"id": "2207.01036", "title": "Memory-Based Label-Text Tuning for Few-Shot Class-Incremental Learning", "abstract": "  Few-shot class-incremental learning(FSCIL) focuses on designing learning\nalgorithms that can continually learn a sequence of new tasks from a few\nsamples without forgetting old ones. The difficulties are that training on a\nsequence of limited data from new tasks leads to severe overfitting issues and\ncauses the well-known catastrophic forgetting problem. Existing researches\nmainly utilize the image information, such as storing the image knowledge of\nprevious tasks or limiting classifiers updating. However, they ignore analyzing\nthe informative and less noisy text information of class labels. In this work,\nwe propose leveraging the label-text information by adopting the memory prompt.\nThe memory prompt can learn new data sequentially, and meanwhile store the\nprevious knowledge. Furthermore, to optimize the memory prompt without\nundermining the stored knowledge, we propose a stimulation-based training\nstrategy. It optimizes the memory prompt depending on the image embedding\nstimulation, which is the distribution of the image embedding elements.\nExperiments show that our proposed method outperforms all prior\nstate-of-the-art approaches, significantly mitigating the catastrophic\nforgetting and overfitting problems.\n", "date": "2022-07-03", "categories": []}
{"id": "2207.00429", "title": "Modular Lifelong Reinforcement Learning via Neural Composition", "abstract": "  Humans commonly solve complex problems by decomposing them into easier\nsubproblems and then combining the subproblem solutions. This type of\ncompositional reasoning permits reuse of the subproblem solutions when tackling\nfuture tasks that share part of the underlying compositional structure. In a\ncontinual or lifelong reinforcement learning (RL) setting, this ability to\ndecompose knowledge into reusable components would enable agents to quickly\nlearn new RL tasks by leveraging accumulated compositional structures. We\nexplore a particular form of composition based on neural modules and present a\nset of RL problems that intuitively admit compositional solutions. Empirically,\nwe demonstrate that neural composition indeed captures the underlying structure\nof this space of problems. We further propose a compositional lifelong RL\nmethod that leverages accumulated neural components to accelerate the learning\nof future tasks while retaining performance on previous tasks via off-line RL\nover replayed experiences.\n", "date": "2022-07-01", "categories": []}
{"id": "2206.13720", "title": "SHELS: Exclusive Feature Sets for Novelty Detection and Continual\n  Learning Without Class Boundaries", "abstract": "  While deep neural networks (DNNs) have achieved impressive classification\nperformance in closed-world learning scenarios, they typically fail to\ngeneralize to unseen categories in dynamic open-world environments, in which\nthe number of concepts is unbounded. In contrast, human and animal learners\nhave the ability to incrementally update their knowledge by recognizing and\nadapting to novel observations. In particular, humans characterize concepts via\nexclusive (unique) sets of essential features, which are used for both\nrecognizing known classes and identifying novelty. Inspired by natural\nlearners, we introduce a Sparse High-level-Exclusive, Low-level-Shared feature\nrepresentation (SHELS) that simultaneously encourages learning exclusive sets\nof high-level features and essential, shared low-level features. The\nexclusivity of the high-level features enables the DNN to automatically detect\nout-of-distribution (OOD) data, while the efficient use of capacity via sparse\nlow-level features permits accommodating new knowledge. The resulting approach\nuses OOD detection to perform class-incremental continual learning without\nknown class boundaries. We show that using SHELS for novelty detection results\nin statistically significant improvements over state-of-the-art OOD detection\napproaches over a variety of benchmark datasets. Further, we demonstrate that\nthe SHELS model mitigates catastrophic forgetting in a class-incremental\nlearning setting,enabling a combined novelty detection and accommodation\nframework that supports learning in open-world settings\n", "date": "2022-06-28", "categories": []}
{"id": "2206.14085", "title": "Continual Learning with Transformers for Image Classification", "abstract": "  In many real-world scenarios, data to train machine learning models become\navailable over time. However, neural network models struggle to continually\nlearn new concepts without forgetting what has been learnt in the past. This\nphenomenon is known as catastrophic forgetting and it is often difficult to\nprevent due to practical constraints, such as the amount of data that can be\nstored or the limited computation sources that can be used. Moreover, training\nlarge neural networks, such as Transformers, from scratch is very costly and\nrequires a vast amount of training data, which might not be available in the\napplication domain of interest. A recent trend indicates that dynamic\narchitectures based on an expansion of the parameters can reduce catastrophic\nforgetting efficiently in continual learning, but this needs complex tuning to\nbalance the growing number of parameters and barely share any information\nacross tasks. As a result, they struggle to scale to a large number of tasks\nwithout significant overhead. In this paper, we validate in the computer vision\ndomain a recent solution called Adaptive Distillation of Adapters (ADA), which\nis developed to perform continual learning using pre-trained Transformers and\nAdapters on text classification tasks. We empirically demonstrate on different\nclassification tasks that this method maintains a good predictive performance\nwithout retraining the model or increasing the number of model parameters over\nthe time. Besides it is significantly faster at inference time compared to the\nstate-of-the-art methods.\n", "date": "2022-06-28", "categories": []}
{"id": "2206.13336", "title": "Continual Learning of Dynamical Systems with Competitive Federated\n  Reservoir Computing", "abstract": "  Machine learning recently proved efficient in learning differential equations\nand dynamical systems from data. However, the data is commonly assumed to\noriginate from a single never-changing system. In contrast, when modeling\nreal-world dynamical processes, the data distribution often shifts due to\nchanges in the underlying system dynamics. Continual learning of these\nprocesses aims to rapidly adapt to abrupt system changes without forgetting\nprevious dynamical regimes. This work proposes an approach to continual\nlearning based on reservoir computing, a state-of-the-art method for training\nrecurrent neural networks on complex spatiotemporal dynamical systems.\nReservoir computing fixes the recurrent network weights - hence these cannot be\nforgotten - and only updates linear projection heads to the output. We propose\nto train multiple competitive prediction heads concurrently. Inspired by\nneuroscience's predictive coding, only the most predictive heads activate,\nlaterally inhibiting and thus protecting the inactive heads from forgetting\ninduced by interfering parameter updates. We show that this multi-head\nreservoir minimizes interference and catastrophic forgetting on several\ndynamical systems, including the Van-der-Pol oscillator, the chaotic Lorenz\nattractor, and the high-dimensional Lorenz-96 weather model. Our results\nsuggest that reservoir computing is a promising candidate framework for the\ncontinual learning of dynamical systems. We provide our code for data\ngeneration, method, and comparisons at\n\\url{https://github.com/leonardbereska/multiheadreservoir}.\n", "date": "2022-06-27", "categories": []}
{"id": "2206.11849", "title": "Sample Condensation in Online Continual Learning", "abstract": "  Online Continual learning is a challenging learning scenario where the model\nmust learn from a non-stationary stream of data where each sample is seen only\nonce. The main challenge is to incrementally learn while avoiding catastrophic\nforgetting, namely the problem of forgetting previously acquired knowledge\nwhile learning from new data. A popular solution in these scenario is to use a\nsmall memory to retain old data and rehearse them over time. Unfortunately, due\nto the limited memory size, the quality of the memory will deteriorate over\ntime. In this paper we propose OLCGM, a novel replay-based continual learning\nstrategy that uses knowledge condensation techniques to continuously compress\nthe memory and achieve a better use of its limited size. The sample\ncondensation step compresses old samples, instead of removing them like other\nreplay strategies. As a result, the experiments show that, whenever the memory\nbudget is limited compared to the complexity of the data, OLCGM improves the\nfinal accuracy compared to state-of-the-art replay strategies.\n", "date": "2022-06-23", "categories": []}
{"id": "2206.10596", "title": "Demystifying the Base and Novel Performances for Few-shot\n  Class-incremental Learning", "abstract": "  Few-shot class-incremental learning (FSCIL) has addressed challenging\nreal-world scenarios where unseen novel classes continually arrive with few\nsamples. In these scenarios, it is required to develop a model that recognizes\nthe novel classes without forgetting prior knowledge. In other words, FSCIL\naims to maintain the base performance and improve the novel performance\nsimultaneously. However, there is little study to investigate the two\nperformances separately. In this paper, we first decompose the entire model\ninto four types of parameters and demonstrate that the tendency of the two\nperformances varies greatly with the updated parameters when the novel classes\nappear. Based on the analysis, we propose a simple method for FSCIL, coined as\nNoNPC, which uses normalized prototype classifiers without further training for\nincremental novel classes. It is shown that our straightforward method has\ncomparable performance with the sophisticated state-of-the-art algorithms.\n", "date": "2022-06-18", "categories": []}
{"id": "2206.09117", "title": "NISPA: Neuro-Inspired Stability-Plasticity Adaptation for Continual\n  Learning in Sparse Networks", "abstract": "  The goal of continual learning (CL) is to learn different tasks over time.\nThe main desiderata associated with CL are to maintain performance on older\ntasks, leverage the latter to improve learning of future tasks, and to\nintroduce minimal overhead in the training process (for instance, to not\nrequire a growing model or retraining). We propose the Neuro-Inspired\nStability-Plasticity Adaptation (NISPA) architecture that addresses these\ndesiderata through a sparse neural network with fixed density. NISPA forms\nstable paths to preserve learned knowledge from older tasks. Also, NISPA uses\nconnection rewiring to create new plastic paths that reuse existing knowledge\non novel tasks. Our extensive evaluation on EMNIST, FashionMNIST, CIFAR10, and\nCIFAR100 datasets shows that NISPA significantly outperforms representative\nstate-of-the-art continual learning baselines, and it uses up to ten times\nfewer learnable parameters compared to baselines. We also make the case that\nsparsity is an essential ingredient for continual learning. The NISPA code is\navailable at https://github.com/BurakGurbuz97/NISPA.\n", "date": "2022-06-18", "categories": []}
{"id": "2206.07842", "title": "Queried Unlabeled Data Improves and Robustifies Class-Incremental\n  Learning", "abstract": "  Class-incremental learning (CIL) suffers from the notorious dilemma between\nlearning newly added classes and preserving previously learned class knowledge.\nThat catastrophic forgetting issue could be mitigated by storing historical\ndata for replay, which yet would cause memory overheads as well as imbalanced\nprediction updates. To address this dilemma, we propose to leverage \"free\"\nexternal unlabeled data querying in continual learning. We first present a CIL\nwith Queried Unlabeled Data (CIL-QUD) scheme, where we only store a handful of\npast training samples as anchors and use them to query relevant unlabeled\nexamples each time. Along with new and past stored data, the queried unlabeled\nare effectively utilized, through learning-without-forgetting (LwF)\nregularizers and class-balance training. Besides preserving model\ngeneralization over past and current tasks, we next study the problem of\nadversarial robustness for CIL-QUD. Inspired by the recent success of learning\nrobust models with unlabeled data, we explore a new robustness-aware CIL\nsetting, where the learned adversarial robustness has to resist forgetting and\nbe transferred as new tasks come in continually. While existing options easily\nfail, we show queried unlabeled data can continue to benefit, and seamlessly\nextend CIL-QUD into its robustified versions, RCIL-QUD. Extensive experiments\ndemonstrate that CIL-QUD achieves substantial accuracy gains on CIFAR-10 and\nCIFAR-100, compared to previous state-of-the-art CIL approaches. Moreover,\nRCIL-QUD establishes the first strong milestone for robustness-aware CIL. Codes\nare available in https://github.com/VITA-Group/CIL-QUD.\n", "date": "2022-06-15", "categories": []}
{"id": "2206.08489", "title": "Debugging using Orthogonal Gradient Descent", "abstract": "  In this report we consider the following problem: Given a trained model that\nis partially faulty, can we correct its behaviour without having to train the\nmodel from scratch? In other words, can we ``debug\" neural networks similar to\nhow we address bugs in our mathematical models and standard computer code. We\nbase our approach on the hypothesis that debugging can be treated as a two-task\ncontinual learning problem. In particular, we employ a modified version of a\ncontinual learning algorithm called Orthogonal Gradient Descent (OGD) to\ndemonstrate, via two simple experiments on the MNIST dataset, that we can\nin-fact \\textit{unlearn} the undesirable behaviour while retaining the general\nperformance of the model, and we can additionally \\textit{relearn} the\nappropriate behaviour, both without having to train the model from scratch.\n", "date": "2022-06-17", "categories": []}
{"id": "2206.08492", "title": "TKIL: Tangent Kernel Approach for Class Balanced Incremental Learning", "abstract": "  When learning new tasks in a sequential manner, deep neural networks tend to\nforget tasks that they previously learned, a phenomenon called catastrophic\nforgetting. Class incremental learning methods aim to address this problem by\nkeeping a memory of a few exemplars from previously learned tasks, and\ndistilling knowledge from them. However, existing methods struggle to balance\nthe performance across classes since they typically overfit the model to the\nlatest task. In our work, we propose to address these challenges with the\nintroduction of a novel methodology of Tangent Kernel for Incremental Learning\n(TKIL) that achieves class-balanced performance. The approach preserves the\nrepresentations across classes and balances the accuracy for each class, and as\nsuch achieves better overall accuracy and variance. TKIL approach is based on\nNeural Tangent Kernel (NTK), which describes the convergence behavior of neural\nnetworks as a kernel function in the limit of infinite width. In TKIL, the\ngradients between feature layers are treated as the distance between the\nrepresentations of these layers and can be defined as Gradients Tangent Kernel\nloss (GTK loss) such that it is minimized along with averaging weights. This\nallows TKIL to automatically identify the task and to quickly adapt to it\nduring inference. Experiments on CIFAR-100 and ImageNet datasets with various\nincremental learning settings show that these strategies allow TKIL to\noutperform existing state-of-the-art methods.\n", "date": "2022-06-17", "categories": []}
{"id": "2206.07932", "title": "Lifelong Wandering: A realistic few-shot online continual learning\n  setting", "abstract": "  Online few-shot learning describes a setting where models are trained and\nevaluated on a stream of data while learning emerging classes. While prior work\nin this setting has achieved very promising performance on instance\nclassification when learning from data-streams composed of a single indoor\nenvironment, we propose to extend this setting to consider object\nclassification on a series of several indoor environments, which is likely to\noccur in applications such as robotics. Importantly, our setting, which we\nrefer to as online few-shot continual learning, injects the well-studied issue\nof catastrophic forgetting into the few-shot online learning paradigm. In this\nwork, we benchmark several existing methods and adapted baselines within our\nsetting, and show there exists a trade-off between catastrophic forgetting and\nonline performance. Our findings motivate the need for future work in this\nsetting, which can achieve better online performance without catastrophic\nforgetting.\n", "date": "2022-06-16", "categories": []}
{"id": "2206.07996", "title": "Continual Learning with Guarantees via Weight Interval Constraints", "abstract": "  We introduce a new training paradigm that enforces interval constraints on\nneural network parameter space to control forgetting. Contemporary Continual\nLearning (CL) methods focus on training neural networks efficiently from a\nstream of data, while reducing the negative impact of catastrophic forgetting,\nyet they do not provide any firm guarantees that network performance will not\ndeteriorate uncontrollably over time. In this work, we show how to put bounds\non forgetting by reformulating continual learning of a model as a continual\ncontraction of its parameter space. To that end, we propose Hyperrectangle\nTraining, a new training methodology where each task is represented by a\nhyperrectangle in the parameter space, fully contained in the hyperrectangles\nof the previous tasks. This formulation reduces the NP-hard CL problem back to\npolynomial time while providing full resilience against forgetting. We validate\nour claim by developing InterContiNet (Interval Continual Learning) algorithm\nwhich leverages interval arithmetic to effectively model parameter regions as\nhyperrectangles. Through experimental results, we show that our approach\nperforms well in a continual learning setup without storing data from previous\ntasks.\n", "date": "2022-06-16", "categories": []}
{"id": "2206.04016", "title": "SYNERgy between SYNaptic consolidation and Experience Replay for general\n  continual learning", "abstract": "  Continual learning (CL) in the brain is facilitated by a complex set of\nmechanisms. This includes the interplay of multiple memory systems for\nconsolidating information as posited by the complementary learning systems\n(CLS) theory and synaptic consolidation for protecting the acquired knowledge\nfrom erasure. Thus, we propose a general CL method that creates a synergy\nbetween SYNaptic consolidation and dual memory Experience Replay (SYNERgy). Our\nmethod maintains a semantic memory that accumulates and consolidates\ninformation across the tasks and interacts with the episodic memory for\neffective replay. It further employs synaptic consolidation by tracking the\nimportance of parameters during the training trajectory and anchoring them to\nthe consolidated parameters in the semantic memory. To the best of our\nknowledge, our study is the first to employ dual memory experience replay in\nconjunction with synaptic consolidation that is suitable for general CL whereby\nthe network does not utilize task boundaries or task labels during training or\ninference. Our evaluation on various challenging CL scenarios and\ncharacteristics analyses demonstrate the efficacy of incorporating both\nsynaptic consolidation and CLS theory in enabling effective CL in DNNs.\n", "date": "2022-06-08", "categories": []}
{"id": "2206.02577", "title": "Effects of Auxiliary Knowledge on Continual Learning", "abstract": "  In Continual Learning (CL), a neural network is trained on a stream of data\nwhose distribution changes over time. In this context, the main problem is how\nto learn new information without forgetting old knowledge (i.e., Catastrophic\nForgetting). Most existing CL approaches focus on finding solutions to preserve\nacquired knowledge, so working on the past of the model. However, we argue that\nas the model has to continually learn new tasks, it is also important to put\nfocus on the present knowledge that could improve following tasks learning. In\nthis paper we propose a new, simple, CL algorithm that focuses on solving the\ncurrent task in a way that might facilitate the learning of the next ones. More\nspecifically, our approach combines the main data stream with a secondary,\ndiverse and uncorrelated stream, from which the network can draw auxiliary\nknowledge. This helps the model from different perspectives, since auxiliary\ndata may contain useful features for the current and the next tasks and\nincoming task classes can be mapped onto auxiliary classes. Furthermore, the\naddition of data to the current task is implicitly making the classifier more\nrobust as we are forcing the extraction of more discriminative features. Our\nmethod can outperform existing state-of-the-art models on the most common CL\nImage Classification benchmarks.\n", "date": "2022-06-03", "categories": []}
{"id": "2606.21253", "title": "Gradient-Free Warm-Start Library Recovery: an Amortized-Regret Separation", "abstract": "Continual learning that is gradient-free, local, online, and append-only is attractive for edge and streaming deployment, but its value is usually argued informally. We give a provable account on recurring-regime streams. Given segmentation, a warm-start library learner attains amortized recovery cost $O\\!\\big(KD/\\varepsilon^2+(R-K)\\logK/\\Delta^2\\big)$ versus a memoryless re-estimator's $\\Theta(RD/\\varepsilon^2)$, an advantage $(R-K)\\,\\Theta(D/\\varepsilon^2)$ growing with dimension $D$ and recurrence density. The mechanism is a decoupling: recognizing which of $K$ seen regimes is active costs $O(\\log K/\\Delta^2)$, independent of $D$, whereas estimating a regime costs $\\Theta(D/\\varepsilon^2)$. We prove this is tight: matching lower bounds give recognition $\\Theta(\\log K/\\Delta^2)$ and a memoryless-class bound $\\Omega(RD/\\varepsilon^2)$, so each term is individually minimax-tight (the joint statement is conditional). The separation is born-immune (a memoryless learner's advantage is identically zero) and paradigm-level: it matches, and does not beat, a fair spawn-capable Bayesian baseline; the contribution is attaining this cost structure without end-to-end backprop and with zero forgetting by construction. A count-calibrated variant ties the baseline's leading constant up to a bounded, never-negative per-recurrence overshoot, hyperparameter-free and with no per-step transcendentals. We bound the scope: recognizable regimes are capped by simplex packing (walls $e^{\\Theta(D)}$); autonomous segmentation is impossible at the packing wall (no detector escapes the false-alarm/delay frontier as regimes overlap); the advantage vanishes under overlap. The dimension-dependent separation is corroborated on synthetic streams and real $k$-mer genome distributions (memoryless cost $\\propto D^{1.04}$, recognition $D$-independent); the one real sequential stream sits in the $D{=}1$ near-null corner.", "date": "2026-06-19", "categories": []}
{"id": "2606.21307", "title": "Task-Differentiated Atomic Skill Expansion and Routing for Continual Learning Across Highly Heterogeneous Tasks", "abstract": "Continual learning (CL) is commonly studied under the assumption that sequential tasks are semantically related or structurally similar. However, in highly heterogeneous settings, where tasks differ substantially in reasoning patterns and input-output formats, existing methods often suffer from catastrophic forgetting and inefficient capacity allocation. To address this challenge, we propose Task-differentiated Atomic Skill Expansion and Routing (\\texttt{TASER}), a CL framework that jointly determines how many new atomic skills to introduce for each task and which skills to activate. The framework first uses atomic skill incremental learning to dynamically expand capacity based on task divergence and model uncertainty. It then applies orthogonality-enhanced skill detection to ensure these skills remain semantically distinct and independently reusable. Finally, a skill dynamic routing mechanism composes task-relevant skills through lightweight task-conditioned gating. We further introduce \\texttt{HeteroCLBench}, a highly heterogeneous benchmark for CL, comprising 19 diverse tasks across 9 cognitive dimensions under a standardized sequential protocol. Experiments on \\texttt{HeteroCLBench} show that \\texttt{TASER} consistently outperforms strong baselines by improving plasticity and reducing catastrophic forgetting.", "date": "2026-06-19", "categories": []}
{"id": "2606.21315", "title": "Social World Model for Lifelong Social Intelligence", "abstract": "Social intelligence is a core competency for language agents, yet current research primarily focuses on static capability evaluation rather than how these skills are continuously shaped and accumulated. This gap calls for a shift toward sustainable learning paradigms. Currently, two methodological pain points exist: social interaction trajectories lack unified structured representations to form iterable learning signals, and capability improvement and retention are typically studied in isolation, hindering the assessment of continuous evolution.\n  To bridge this gap, we propose the Social World Model. We decompose social interaction into five dimensions (scene setting, observation, mental state, action, and dialogue) to build a closed-loop learning framework. In this setup, agents collect interaction experiences, convert them into preference signals for model updating, and redeploy the updated policy for continued learning. Additionally, we provide a reusable data synthesis mechanism and a lifelong learning benchmark, transforming social capabilities from an \"object of evaluation\" into an \"object of sustainable training\".\n  Validating our framework on the ASCENT-Bench, the interactively trained Qwen2.5-7B model outperforms its baseline across all five core metrics. Notably, it matches the closed-source Gemini 3 Flash in completion rate, exceeds it in pass rate, and achieves zero forgetting across three difficulty levels. Unlike prior works that merely report static comparisons or capability decay, this end-to-end approach provides a trainable, verifiable, and retainable pathway, demonstrating that small open-source models can sustainably acquire competitive social coordination capabilities.", "date": "2026-06-19", "categories": []}
{"id": "2606.22248", "title": "SamatNext v0.2-B: An Exploratory Study of RMS-Normalized Hybrid Decoders for Curriculum Retention in Small Code Models", "abstract": "Standard autoregressive Transformer decoders can often exhibit substantial forgetting under sequential fine-tuning on shifting curriculum distributions. This technical report evaluates SamatNext v0.2-B, an experimental 356M-parameter hybrid sequence decoder that alternates Differential-Attention-style layers with DeltaNet-inspired simplified linear-state mixer layers using RMS normalization and output scale calibration. We study the model under a controlled staged Python code curriculum and compare it with a parameter-matched Transformer baseline. In this setting, SamatNext v0.2-B achieves a 100.0% pass rate on the controlled Stage 5 holdout while retaining 98.8% of adjacent Stage 3 semantic behavior and reaching 12.0% on the Stage 2E early syntax holdout. The strongest Transformer baseline reaches 97.6% on Stage 5 but retains only 6.0% of Stage 3 behavior. Both architectures remain weak on long-horizon early-stage retention, so the result should be interpreted as evidence of an altered retention/plasticity tradeoff in this controlled setting, not as a general solution to catastrophic forgetting. Code, model specifications, evaluation scripts, and result tables are provided for independent verification.", "date": "2026-06-20", "categories": []}
{"id": "2606.22999", "title": "Black-Box Continual Learning for Vision-Language Models", "abstract": "The rapid deployment of Vision-Language Models (VLMs) in dynamic environments necessitates the ability to learn continuously without forgetting. However, traditional continual learning (CL) settings often rely on white-box paradigms, which is increasingly invalidated by the shift toward cloud-hosted models. In this paper, we introduce Black-CL, a more realistic benchmark for VLMs that enforces three primary real-world challenges: weight and architecture inaccessibility, constrained computation, and task-agnostic inference. The learner can query only output embeddings or logits, with no gradient flow through or structural modification of the backbone. Current CL methodologies, which rely on backbone backpropagation or complex parameter expansion, are fundamentally incompatible with these constraints. Under this setting, we propose BETA, a simple yet effective baseline built on the key insight that solely optimizing textual prototypes can navigate the complexities of CL. BETA integrates three core components: Semantic Projection Accumulation (SPA) for incremental knowledge acquisition, Latent Distribution Replay (LDR) for anchoring the embedding space against catastrophic forgetting, and Test-Time Prototype Adaptation (TTPA) for dynamic, instance-aware boundary refinement. Extensive experiments across ten diverse datasets and various backbones demonstrate that BETA significantly outperforms existing black-box tuners. Remarkably, with only 0.05 M trainable parameters, a 180--3000$\\times$ reduction compared to competitive methods, BETA achieves performance on par with or even exceeding white-box CL methods. We believe Black-CL and BETA provide a foundational framework for future advancements in continual learning and accelerates the transition of continual learning from academia to real-world systems.", "date": "2026-06-22", "categories": []}
{"id": "2606.23063", "title": "Attention-Spectrum Regularization for Replay-Free Continual Multimodal LLMs", "abstract": "Multimodal large language models (MLLMs) are increasingly required to adapt to non-stationary streams of visual domains, question types, and user instructions, yet continual fine-tuning often causes severe forgetting of previously acquired multimodal skills. Existing continual vision-language methods mainly preserve outputs, replay data or pseudo-data, regularize embedding geometry, or allocate task-specific parameters, but they provide limited control over how internal cross-modal attention patterns supporting old skills drift during adaptation. We propose Attention-Spectrum Regularization (ASR), a replay-free continual learning framework that preserves skill-conditioned structures of cross-modal attention. ASR treats cross-attention maps as two-dimensional signals, summarizes their scale and directional properties into compact spectral statistics, and stores only skill-wise prototype distributions instead of replaying past image-question pairs, generated pseudo-examples, or old-stage teacher snapshots. In later stages, a phase-invariant spectral regularizer constrains harmful drift of these prototypes while allowing instance-level attention to adapt to new tasks. We provide theoretical analysis showing that skill-conditioned spectral drift controls forgetting under a spectral sufficiency assumption, and that Fourier power spectra are stable to spatial translations and bounded perturbations. Experiments on continual VQA and multimodal instruction-tuning benchmarks, including VQA v2, VQACL, CLT-VQA, CoIN, and UCIT, show that ASR consistently improves final performance and reduces forgetting over strong replay-, regularization-, and adapter-based baselines. Preserving skill-level attention structure is an effective and lightweight mechanism for continual MLLMs. Code is available at https://github.com/Creative-zcx/attention-spectrum-replay", "date": "2026-06-22", "categories": []}
{"id": "2606.23487", "title": "CADRE: Stable, Parameter Efficient Adaptation of Medical Vision Language Models with Bounded Forgetting and Prior Drift", "abstract": "Medical vision-language models (VLMs) such as BiomedCLIP generalize broadly, but adapting them to a clinical service is as much a safety problem as an accuracy one. Updating a deployed model for a new imaging modality can fail silently in two ways that harm patients: it can forget modalities it already handled (catastrophic forgetting), and it can drift from its trustworthy pretrained prior toward modality-specific shortcuts. We study parameter-efficient continual adaptation through these two properties rather than leaderboard accuracy, presenting CADRE: a frozen-backbone framework combining low-rank adaptation (LoRA) with an online, self-scaling, similarity-aware elastic weight consolidation term that bounds retained-competence loss, and an anchor-to-prior penalty bounding embedding drift from the frozen prior. Two short guarantees, a bound on total consolidation mass and a scale-invariance property, remove the scale-related sources of vanilla EWC's order fragility. Using breast cancer across three maximally dissimilar modalities (histopathology, ultrasound, chest radiography) as a controlled cross-modality stress test, under a multi-seed, multi-order protocol with paired significance testing and training approximately 0.23% of parameters, CADRE attains the highest accuracy, SPQ, and backward transfer and the lowest forgetting among adapting methods, reducing forgetting roughly sevenfold versus the strongest regularized baseline (0.075 to 0.011; paired p=0.023) and achieving positive backward transfer where every baseline is negative. We frame these as stability properties aligned with clinical-safety desiderata, not a deployment guarantee; robustness to distribution shift and adversarial inputs is out of scope.", "date": "2026-06-22", "categories": []}
{"id": "2606.24007", "title": "Fast and Slow Variational Continual Learning", "abstract": "Continual learning remains a major challenge for modern deep networks, partly because commonly used optimizers lack inherent mechanisms for continual adaptation. One such natural mechanism is fast and slow adaptation to balance stability and plasticity. This mechanism has deep roots in neuroscience and biology, but there is no consensus on how to best incorporate it in commonly used optimizers. Here, we show that this can be easily done via the VCL framework, where past posteriors are used as priors in the future. Our key idea is to incorporate slow adaptation via merging of past posteriors to slow down the drift in the knowledge as learning progresses. The merged posterior is then used as the prior in the VCL update to implement the fast-weight updates. These steps can be seamlessly implemented in the IVON optimizer, whose form and costs are nearly identical to that of Adam. We call this new optimizer the Continual IVON (CoVON) optimizer and show that it not only consistently improves over existing VCL optimizers, but also performs better than other weight-regularization strategies across domain-incremental learning, continual pre-training, and fine-tuning of large language models.", "date": "2026-06-22", "categories": []}
{"id": "2606.24752", "title": "Can Scale Save Us From Plasticity Loss in Large Language Models?", "abstract": "The loss of plasticity - the ability of a network to learn new information after having already learned older information - is a fundamental challenge in creating artificial neural networks capable of continual learning. Although this phenomenon has been known for decades, it has mostly been studied in older, relatively small architectures and rarely in natural-language domains. To determine whether loss of plasticity remains a problem in the modern transformer-based LLM paradigm, we study plasticity loss in GPT-style Transformer models trained on a multilingual continual learning problem. Consistent with prior work, we find evidence of plasticity loss across models ranging from 5M to 314M non-embedding parameters, as measured by deterioration on a held-out Vietnamese probing task. We further find that the onset of plasticity loss follows a predictable scaling law, growing sublinearly with model size. These results suggest that larger models may delay the measurable effects of plasticity loss, but that increasing parameter count alone is likely to be insufficient to completely prevent it. We also find evidence of plasticity loss under stationary multilingual training, challenging the view that the phenomenon is exclusive to continual learning with abrupt task changes. Overall, our results suggest that even large Transformer language models trained on natural-language will eventually lose the ability to efficiently adapt to new data after sufficiently long training, in both continual and stationary settings.", "date": "2026-06-23", "categories": []}
{"id": "2606.25115", "title": "Forget to Improve: On-Device LLM-Agent Continual Learning via Budget-Curated Memory", "abstract": "On-device language-model agents improve by accumulating experience in retrieved memory rather than by updating weights. This memory is hard-bounded and exposed: it consumes RAM and energy, reaches peers through a thin uplink, and becomes an attack surface because it is writable by what the agent reads. Existing systems each cover one part of this problem: agentic memories grow without a budget, on-device methods keep entries by success alone, and poisoning is studied mainly as an attack rather than as a memory-governance problem. We propose \\sys{}, a single net-value-per-byte score that governs an agent's experience-memory lifecycle. The main idea is to let the budget act as the curator: each entry is scored as value minus harm, per byte, so one ruler decides what to keep, share, and trust. \\sys{} makes three decisions: (1) \\textbf{KEEP} evicts low-value bytes under the RAM and energy budget; (2) \\textbf{SHARE} sends an insight only when its value exceeds its uplink cost; and (3) \\textbf{TRUST} gates a peer entry by provenance. On language-model-agent task-drift benchmarks and a real heterogeneous Jetson testbed with two robot-arm nodes and a hub, \\sys{} reduces memory by $2.7\\times$ and uplink by $2.4\\times$, drives injection success from 0.75 to zero, and raises accuracy on cases corrupted by poison or stale memory. Curating by net value reduces footprint, energy, uplink, and injection success together without reducing accuracy. In this setting, forgetting by net value improves the agent rather than weakening it.", "date": "2026-06-23", "categories": []}
{"id": "2606.25165", "title": "The Gentle Collapse: Distributional Metrics for Continual Learning", "abstract": "Accuracy degradation is the standard metric for Catastrophic Forgetting (CF), however, it records only whether forgetting occurred or not. It saturates at the extremes and collapses discretely at task boundaries, hiding the internal structure of what is being forgotten. We introduce six softmax-derived metrics spanning true-label rank (TLR), predictive confidence, and distributional divergence that characterize forgetting continuously, each normalized to [0, 1] with no modification to training. On CIFAR-100, these metrics carry information where accuracy does not: at 0% accuracy, the Confusion Margin spans an IQR of [0.32, 0.50] across classes that accuracy treats identically. We demonstrate that this richer signal is actionable in mitigating catastrophic forgetting. Per-sample metric scores used as loss weights reduce forgetting by 1.3 percentage points over uniform experience replay (ER) on CIFAR-100. Furthermore, the slope of a metric over a small window provides a stable sampling criterion: at a small-window size (e.g. 3 epochs), accuracy-trend degrades to 34.79% (std. = 2.32) while log-TLR achieves 41.07% (std. = 0.57). This gap is structural since reliable small-window trend estimation requires a continuous signal. On TinyImageNet, log-TLR trend sampling reduces forgetting by 7.7 percentage points over the ER baseline.", "date": "2026-06-23", "categories": []}
{"id": "2606.25342", "title": "Lifelong In-Context Learning with Transformers Requires Parametric Forms of Attention", "abstract": "Lifelong continual learning remains an obstacle on the path to human-like intelligence. Modern transformers show sparks of intelligence with in-context learning. The quadratic nature of attention, however, prohibits transformers from performing this process on arbitrarily long sequences. In this work, we argue that extending in-context learning to lifelong settings is a practical solution for continual learning in AI agents. In particular, we argue that \\emph{parametric forms of attention} are needed to understand a lifetime of context with transformers on a fixed hardware budget. These attention mechanisms learn the relationship between keys and their associated values at test-time with parametric regression. Our generalization of parametric approaches (linear attention, state-space models, fast weight programmers, and test-time training layers) contrasts with nonparametric counterparts like softmax attention. They replace the ever-growing key-value cache with an online-trainable neural network, maintaining a constant memory footprint. We highlight how parametric attention currently fall short of lifelong learning due to limited memory capacity or costly online updates. To address these issues, we pose a set of open questions with novel insights to guide the field toward long-horizon agents.", "date": "2026-06-24", "categories": []}
{"id": "2606.25347", "title": "Geometry-Anchored Transport Framework for Exemplar-Free Class-Incremental Learning", "abstract": "Exemplar-free class-incremental learning (EFCIL) requires stable decision boundaries within a shifting feature space. While maintaining class-conditional Gaussian statistics provides a principled classification strategy, these parametric summaries remain sensitive to anisotropic representation drift. Existing methods often transport these statistics across tasks using a decoupled, post-hoc paradigm: optimizing a backbone without explicit geometric constraints can distort the legacy manifold, limiting the precision of retroactive alignment. In this paper, we formulate feature transport as an endogenous training constraint rather than a separate post-task step, presenting the Geometry-Anchored Transport Framework. First, we derive an Analytic Geometric Anchor via Mahalanobis-aligned regression to mitigate macroscopic anisotropic drift. Second, we introduce a Topology-Aware Evolution objective that regularizes localized manifold degradation while calibrating a residual network against the analytic prior. By coupling manifold evolution with transport constraints during the primary training phase, our framework mitigates evaluation errors without requiring decoupled fine-tuning. Experiments across CIFAR-100, TinyImageNet, and ImageNet-100 demonstrate that the proposed framework consistently improves upon existing post-hoc alternatives under strict exemplar-free constraints.", "date": "2026-06-24", "categories": []}
{"id": "2606.26629", "title": "From Weights to Features: SAE-Guided Activation Regularization for LLM Continual Learning", "abstract": "Weight-space regularization methods such as Elastic Weight Consolidation (EWC) are the standard approach to catastrophic forgetting in continual learning. However, those methods tend to underperform when applied to large language models. We argue that such underperformance can be partly explained by the ``polysemantic'' nature of large language models: per-weight importance estimates utilized by EWC-style regularization are too coarse and cannot isolate the knowledge that needs protection. In this paper, we propose regularizing instead in the model's activation space, using pretrained Sparse Autoencoders (SAEs) as a monosemantic feature dictionary. From the perspective of constrained optimization, we derive a new loss function that uses the SAE feature dictionary to explicitly balance stability and plasticity, and show that EWC is a special case in the one-sided weight-space penalty setting. Unlike replay-based methods that store or revisit examples from earlier tasks, our method requires no previous-task data after mask construction: current-task data is used to compute a compact SAE feature mask, and only this mask is retained for later training. Further, since the feature space has significantly lower dimensionality than the parameter space, the proposed method is more memory efficient. On the TRACE and MedCL continual learning benchmarks, the method achieves the strongest result among approaches without introducing task-specific architectural components, also surpassing traditional weight-space regularization methods like EWC. Beyond performance comparisons, we provide empirical evidence for the polysemanticity thesis: task-relevant representations are linearly separable in the SAE feature basis but indistinguishable from chance in the weight basis, and weight-space protection is nearly non-selective at the concept level.", "date": "2026-06-25", "categories": []}
{"id": "2606.27095", "title": "Data-Free Reservoir Features for Efficient Long-Horizon Cold-Start Continual Learning", "abstract": "Cold-start exemplar-free class-incremental learning requires learning a growing set of classes without replay, external pretraining, or a large initial task. Existing cold-start methods typically either train the backbone throughout the stream and compensate for semantic drift, or freeze a backbone after the first task, producing features biased toward the initial classes. These choices also create a computational tension: drift-compensation methods require repeated backbone training and increasingly expensive updates as the task horizon grows, while frozen-backbone methods are cheap but weak under cold start. We study a third option: a feature extractor that is never fit to image data at all. We propose CIRCLE, a class-incremental classifier built from fixed bidirectional two-dimensional reservoir features, adapted from BiRC2D for image classification, and streaming linear discriminant analysis heads. CIRCLE groups multiple random reservoir instantiations into feature ensembles and averages the softmax outputs of independent SLDA heads, yielding a tunable bias-variance tradeoff between richer random features and prediction-level ensembling. Because the feature extractor is fixed and the head admits streaming closed-form updates, CIRCLE performs sample-wise training without replay, task-boundary information, or backbone backpropagation. On CIFAR-100, TinyImageNet, ImageNet-Subset, and ImageNet-1k, CIRCLE is competitive at 10-20 task splits and substantially outperforms strong CS-EFCIL baselines at 50, 100, and 500 task splits, while training much faster than trained-backbone drift-compensation methods. Ablations show that the BiRC2D-style extractor, SLDA head, and balanced feature/prediction ensembling each contribute to the final performance.", "date": "2026-06-25", "categories": []}
{"id": "2606.27374", "title": "World Action Models Enable Continual Imitation Learning with Recurrent Generative Replays", "abstract": "Going beyond predicting robot actions, World Action Models (WAMs) can also generate future visual observations. We build on this generative capability to propose Recurrent Generative Replay (REGEN), a continual imitation learning framework that synthesizes pseudo-replay trajectories, enabling a robot policy to rehearse previously learned tasks without storing their original human demonstrations. During continual adaptation, REGEN recursively queries the WAM to synthesize pseudo-replay trajectories conditioned only on prior task instructions and current-task observations. Experiments in both simulation and real-world manipulation settings show that REGEN reduces catastrophic forgetting by up to $50\\%$ relative to sequential fine-tuning, while approaching the performance of privileged experience replay methods that require access to real replay data. Finally, we analyze the factors limiting generated replay, identifying long-horizon visual degradation and action-observation inconsistency as the primary bottlenecks. Our results establish WAMs as a promising foundation for continual robot learning without stored demonstrations.", "date": "2026-06-25", "categories": []}
{"id": "2606.27634", "title": "Continual Learning for Sequential Personalization of Small Language Models: A Stability Monitoring Analysis", "abstract": "Small Language Models (SLMs) are increasingly being considered for deployment on edge devices such as laptops, enabling private, low-latency, and locally personalized applications. However, personalization requires models to adapt over time to evolving user- or task-specific data, placing them in a continual learning setting. This creates the risk of catastrophic forgetting, where learning new information degrades performance on previously learned tasks or broader model capabilities. Recent benchmarks such as TRACE have shown that continual fine-tuning can significantly degrade the general abilities of aligned large language models. In this work, we present a study for sequential LoRA personalization of SLMs. We save model checkpoints after each adaptation stage and evaluate them on current tasks, previously seen tasks, and a fixed reference set. This checkpoint-level protocol enables us to monitor task performance, forgetting, and reference set drift over time. We show that lightweight reference set distributional diagnostics can reveal model-specific instability patterns during sequential LoRA personalization of SLMs, including cases where task-level metrics alone hide harmful adaptation. We hope this can highlight new research avenues for monitoring stability of SLMs in a continual learning setting.", "date": "2026-06-26", "categories": []}
{"id": "2606.28117", "title": "When One Adapter Speaks for Many: Discovering Low-Rank Redundancy in Continual Fine-Tuning", "abstract": "Low-Rank Adaptation (LoRA) has become the standard tool for parameter-efficient fine-tuning of large pretrained models. When applied sequentially across tasks in Continual Learning (CL), the standard assumption is that each new task requires a dedicated low-rank adapter. In this work, we challenge this assumption empirically and structurally. We show that task-specific LoRA adapters in CL exhibit significant low-rank redundancy: the subspaces spanned by adapters trained on different tasks substantially overlap, and in many cases earlier adapters can faithfully represent later tasks. Building on this observation, we propose LiteLoRA, a plug-and-play gating mechanism that learns at train time whether to recruit a new adapter or reuse existing low-rank representations. Our method reduces the number of active adapters by 20-70% while matching or exceeding state-of-the-art performance on standard CL benchmarks, revealing that structural redundancy is pervasive and that selective learning is sufficient to achieve stability without sacrificing plasticity.", "date": "2026-06-26", "categories": []}
{"id": "2606.28835", "title": "Fisher-Routed Mixture of Experts for Federated Class-Incremental Learning", "abstract": "Federated Learning (FL) emerged as a promising distributed machine learning paradigm. However, extending FL to the class incremental learning scenarios introduces unique challenges: 1) Capacity conflict and catastrophic forgetting from the shared model overloading, 2) Heterogeneity from Non-Independent and Identically Distributed (Non-IID) data, and 3) Synchronized class misalignment. In this paper, we propose \\textbf{F}isher-Routed \\textbf{M}i\\textbf{X}ture of Experts for \\textbf{Fed}erated Class-Incremental Learning (\\textsc{FedFMX}), a novel framework to address these challenges via adaptive expert specialization across clients. The crucial insight is to route each sample to an expert subset that jointly optimizes knowledge acquisition and retention. Specifically, we introduce a Fisher-Routed Expert Scoring (FRES) module to estimate expert importance via Fisher-based stability cost and gradient-based plasticity gain. Then, we design an Adaptive Expert Selection (AES) module by quantifying marginal contributions for adaptive expert subset determination. Finally, by the routing-aware regularization (RAR), we achieve load balance and efficient FL training. We theoretically prove the $\\mathcal{O}(T^{-1})$ convergence rate. Extensive experiments on multiple benchmarks compared with state-of-the-art methods demonstrate the superiority of \\textsc{FedFMX}.", "date": "2026-06-27", "categories": []}
{"id": "2606.29465", "title": "Prototype Latent World Model Replay for Class-Incremental Learning", "abstract": "Class-incremental learning requires a model to learn new classes while preserving decision regions for old ones. This is difficult when raw old samples are no longer available. We propose Prototype Latent World Model Replay, a memory-free framework that stores old classes as distributions over stable hidden states rather than as images. A frozen ImageNet-pretrained encoder maps each image into a latent state space. In this space, each class is summarized by several prototype-centered distributions with class-specific variances. When new classes arrive, the model samples old latent states from this prototype world model. It then trains a lightweight adapter and classifier using both sampled old states and real new-class features. We also add a supervised contrastive term in the adapter space to promote intra-class compactness and old-new class separation. On Split CIFAR-100, our method improves over fine-tuning under Inc5, Inc10, and Inc20 without storing raw exemplars. The full Ours-LWM+Con model raises LastAcc from 4.55% to 31.64%, from 9.06% to 37.06%, and from 16.96% to 43.10% in Inc5, Inc10, and Inc20, respectively. It also achieves AvgAcc of 45.86%, 52.19%, and 56.18%. Ablation and retention analyses show that stable latent-state replay is the main source of the gain. Contrastive separation further refines the old-new geometry. These results suggest that prototype latent memory preserves reusable class-state distributions, rather than only fitting the current classifier.", "date": "2026-06-28", "categories": []}
{"id": "2606.29832", "title": "The Forgetting-Retention Dilemma: Certified Unlearning Theory in Continual Learning", "abstract": "Machine unlearning aims to eliminate the influence of specific data from trained models to safeguard privacy. However, this presents a significant challenge in the context of continual learning (CL), where models update sequentially on dynamic datasets. A major limitation is that current certified unlearning algorithms fail to account for the complex, cumulative model evolution inherent to CL framework. In this work, we establish the first theoretical foundation bridging CL and machine unlearning. We formulate the CL's unlearning objective as the minimization of post-unlearning excess risk, which decomposes into CL excess risk and unlearning loss, characterizing the fundamental trade-off between preserving historical knowledge and targeted forgetting. Under mild assumptions, we first establish an upper bound for the CL excess risk in non-convex models. We then adapt two certified unlearning approaches, gradient-based and Hessian-based, to the CL framework. Our analysis reveals that while the gradient-based approach is less effective than the Hessian-based method in minimizing unlearning loss, it offers the distinct advantage of nearly zero storage overhead for enabling unlearning. This insight motivates a hybrid strategy that reduces storage costs while maintaining post-unlearning performance. Experimental results further validate our theoretical findings.", "date": "2026-06-29", "categories": []}
{"id": "2606.29841", "title": "Theory of Continual Learning Against Data Poisoning Attacks", "abstract": "Continual learning (CL), where a model is trained on a sequence of data tasks, is increasingly being adopted across key fields such as large language models and image recognition, yet it remains highly vulnerable to data poisoning that triggers learning divergence or severe excess risk. Despite these threats, a principled theoretical foundation in CL for understanding attack and defense remains lacking. In this paper, we develop a theoretical framework to analyze strategic attacks and defenses in regularization-based CL, a cornerstone of recent CL theory. By framing the adversary-defender interaction as an online zero-sum game, we first establish a fundamental performance limit: no defense succeeds when an adversary poisons a linear proportion of tasks by injecting unbounded noise or pattern shifts in regularization-based CL. We then analyze two possibly defensible scenarios: infrequent attacks and bounded noise per attack. For the former regime, we propose a task-to-task verification mechanism to detect data poisoning and reduce cumulative bias for learning convergence. For the latter regime, we derive a robust defense that minimizes the model's sensitivity to poisoned features, provably accelerating the convergence rate. Extensive experiments on realistic tasks further validate our theoretical results.", "date": "2026-06-29", "categories": []}
{"id": "2606.29916", "title": "EVAF: A Test-Retest Protocol for Selective Parametric Consolidation", "abstract": "Long-running language agents need mechanisms for deciding which experiences should persist after the working context is gone. Retrieval systems can reinsert past text, but they do not by themselves show that an experience has been selectively consolidated into the model's own behavior. We introduce EVAF, an Echo-Valence Attractor Field mechanism for gated LoRA consolidation, and a test-retest protocol for measuring selective parametric consolidation under controlled interference. Across GPT-2 and TinyLlama, EVAF preferentially consolidates high-valence, high-surprise experiences while preserving retrieval-accessible factual memory through a complementary routed memory path. Test-retest measurements show stronger post-interference behavioral persistence than frozen, retrieval-only, and ungated continual-update baselines, while keeping parameter drift and cross-persona contamination low. The results support a separation between memory access and memory depth: retrieving a fact and internalizing an experience are distinct computational operations.", "date": "2026-06-29", "categories": []}
{"id": "2606.30067", "title": "Neural Subspace Reallocation: Continual Learning as Retrieval-Based Subspace Memory Management", "abstract": "We introduce Neural Subspace Reallocation (NSR), which reframes continual learning as memory management over parameter subspaces. Instead of treating Low-Rank Adaptation (LoRA) modules as disposable per-task adapters, NSR manages them as compressible, retrievable memory units on a frozen backbone through a recurring cycle: (1) compress learned LoRAs via SVD, (2) reserve them in a TaskKnowledgeBank, (3) recall related past LoRAs by embedding similarity to warm-start new or returning tasks, and (4) reallocate the active subspace accordingly, with distillation protecting prior tasks. We prove that in cyclic environments any memoryless allocation policy incurs cumulative regret Omega(T(M-1)Delta_switch) relative to a history-aware policy backed by the Bank (Theorem 1). Empirically, on Split-CIFAR-100 the Bank reduces cyclic recovery time by 10x, exactly as predicted, and on the heterogeneous 5-Datasets benchmark NSR achieves the highest accuracy and the least forgetting, about 9x closer to zero backward transfer than the memoryless heuristics. Crucially, we run a controlled study that isolates which component matters: holding the Bank fixed and varying only the allocation rule, we find that a simple similarity-based retrieval rule matches or beats a learned reinforcement-learning controller (recovering recurring tasks in 0 vs 1.8 steps and reaching equal accuracy). Our central, honest finding is therefore that the memory mechanism -- compression and similarity retrieval -- rather than a learned allocation policy, drives continual-learning performance under fixed capacity. A memory-budget analysis confirms the compressed Bank stays small -- 0.29 MB of parameter memory per task -- so a top-K retention cap bounds the total footprint while preserving fast recovery for retained tasks.", "date": "2026-06-29", "categories": []}
{"id": "2606.30190", "title": "Few-Shot Domain Incremental Learning via Continual Vision-Language Consolidation", "abstract": "Existing domain-incremental learning (DIL) strategies call for massive amounts of data to adapt to new domains and suffer from the overfitting problem in the case of data scarcity. This paper puts forward a relatively uncharted problem, namely, few-shot domain incremental learning (FSDIL), taking into account the problem of extreme data shortages in the realm of DIL. A novel algorithm, namely Continual Vision-Language Consolidation (CVLC), is proposed to address the FSDIL problem, where the key idea lies in the concept of latent space reservation in the base domain coupled with dual coalescent projection (DCP) as a parameter-efficient fine-tuning method. First, the vision prototype is calibrated while multiple templates and synonyms are generated via LLMs to induce the language prototype. The vision and language prototypes are fused. Adaptation to never-ending arrivals of new domains is done by the DCP technique, fine-tuned in such a way to prepare the model to unseen domains via latent-space reservations committed in the base domain. CVLC is structured under shared and domain-specific components to combine general knowledge and domain-specific details. The advantage of our approach is demonstrated through a range of benchmark problems and comparisons with prior arts, in which CVLC outperforms them by up to a 16% gap. Our codes are shared publicly in https://github.com/Naeem-Paeedeh/CVLC .", "date": "2026-06-29", "categories": []}
{"id": "2606.30266", "title": "Towards Continual Motion-Language Agents: LoRA Variants for Incremental Motion Understanding and Generation", "abstract": "Motion-language agents must possess the bidirectional capability to both understand human movement (motion-to-text, M2T) and generate it from natural language (text-to-motion, T2M). While foundational models have achieved strong performance in static settings, autonomous agents operating in dynamic environments must continuously incorporate new motion concepts -- such as novel athletic styles or specialized gestures -- without catastrophic forgetting of previously acquired skills. We investigate the stability-plasticity trade-off in bidirectional motion-language learning under sequential task exposure. Building on a frozen large language model backbone, we introduce low-rank adaptation (LoRA) variants designed to mitigate inter-task interference. We specifically propose mixture-of-experts architectures that utilize an autoencoder-based router to select task-specific experts at inference time, so that no task-label is needed. To evaluate these methods, we establish a reproducible five-task benchmark derived from HumanML3D through semantic clustering of motion descriptions. Our experimental results demonstrate near-zero forgetting across both M2T and T2M directions while maintaining high generation and captioning quality. Furthermore, we show that hard expert selection via routing significantly outperforms soft expert blending in quality metrics, indicating that preserving expert isolation is critical for maintaining performance in our continual learning setting. Finally, we observe that a divergence between token-level accuracy and downstream generation quality may occur, highlighting the need for more comprehensive evaluation protocols in future research on lifelong motion-language agents.", "date": "2026-06-29", "categories": []}
{"id": "2606.30537", "title": "Learning from Mistakes: Rollout-Retrieval Lifelong Policy Learning for Autonomous Driving", "abstract": "Autonomous driving policies should be able to improve continually as deployment exposes them to increasingly diverse and long-tail traffic situations. However, most learning-based policies are trained or fine-tuned on expert demonstrations and then rely largely on generalization to handle challenging closed-loop scenarios, lacking an explicit mechanism to correct and retain the mistakes exposed in these scenarios. This paper studies autonomous driving policy improvement from a lifelong learning perspective: Can a pretrained policy improve continually by accumulating corrective knowledge derived from its own mistakes, while retaining previously acquired driving competence? To answer this question, we propose Rollout-Retrieval Lifelong Policy Learning (R$^2$LPL), a policy learning framework that retrieves corrective targets from recoverable policy-induced mistakes and retains the resulting knowledge through lifelong policy learning. R^2LPL addresses a key bottleneck in continual policy improvement: closed-loop mistakes reveal where the policy is weak, but do not directly specify what the policy should learn. By filtering recoverable mistake-related states and retrieving feasible corrective targets, R$^2$LPL turns sparse failure evidence into compact supervised knowledge for stable and sample-efficient policy improvement. We evaluate R$^2$LPL on large-scale closed-loop nuPlan benchmarks. With only a few rollout and continual-learning cycles, R$^2$LPL elevates a learning-based planner with moderate initial performance to state-of-the-art performance across the evaluated benchmarks, especially on the challenging and long-tail Test14-hard split. These results demonstrate the effectiveness of R$^2$LPL in converting recoverable closed-loop mistakes into corrective knowledge for sustained policy improvement.", "date": "2026-06-29", "categories": []}
{"id": "2606.30559", "title": "Convergence of Continual Learning in Homogeneous Deep Networks", "abstract": "We characterize weakly regularized continual classification in homogeneous models as sequential projections onto task margin sets. This result generalizes prior analyses restricted to either stationary (single-task) deep models or continual linear models. We show that global convergence generally fails, even for simple models linear in data but nonlinear in parameters. Nevertheless, by leveraging results from nonconvex projection theory, we identify regularity properties of homogeneous deep networks that guarantee local linear convergence under random and cyclic task sequences. Finally, we extend our analysis to continual regression, unifying the framework for homogeneous models.", "date": "2026-06-29", "categories": []}
{"id": "2606.31191", "title": "ISM:Self-Improving Strategy Memory for Continual Mathematical Reasoning", "abstract": "We propose Intelligent Schema Memory (ISM), a self-evolving memory-augmented system that improves mathematical reasoning for a frozen LLM under continual learning with hard episodic resets. ISM maintains a compact, self-refined bank of strategy schemas learned from both successful and failed episodes, with symbolic tools that check intermediate steps and certify answers. Without updating model parameters, ISM outperforms passive, retrieval, and reflection baselines on MATH-Hard and OlympiadBench, using 64% and 86% fewer schemas respectively than the strongest passive baseline. These results show that small, actively maintained, and verified strategy memories can support reliable continual mathematical reasoning under strict episodic isolation. The codebase is available at https://github.com/pdx97/ISM .", "date": "2026-06-30", "categories": []}
{"id": "2606.31275", "title": "CLIMB: Centroid-Based Hierarchical Memory for Online Continual Self-Supervised Learning", "abstract": "Online Continual Self-Supervised Learning (OCSSL) aims to learn representations from a continuous stream of unlabeled data, without knowledge of task boundaries and under memory constraints. Existing methods rely either on replay buffers that exploit latent space structure, or on regularization alone. We present CLIMB (Continual Learning with Intelligent Memory Bank), which combines both simultaneously. Our method introduces a hierarchical centroid-based memory, bounded in total number of stored images, combined with knowledge distillation on replayed examples to limit representation drift. The memory groups similar images into centroids, providing hard-to-discriminate examples for contrastive learning while covering the diversity of observed distributions. Experiments on Split CIFAR-100 and Split ImageNet-100, on standard benchmarks from the state-of-the-art as well as a new protocol with irregular task distributions show that CLIMB outperforms state-of-the-art OCSSL methods.", "date": "2026-06-30", "categories": []}
{"id": "2606.31495", "title": "Surprise as a Signal for Plasticity and Metacognition", "abstract": "We study a single idea across two settings: that a prediction-error signal, computed by a small predictor over the latent space of a frozen encoder, can serve both as a gate on plasticity and as a substrate for metacognition. In the first system, a non-parametric episodic memory writes a new concept only when this surprise is high, and a periodic offline replay phase consolidates recent traces into a slow linear readout. On a continual stream of 1000 ImageNet classes with a frozen DINOv2 or I-JEPA backbone, the consolidation phase recovers 17.7 points of retention on the oldest classes for DINOv2 and 51.3 points for I-JEPA (single-seed runs), and an ablation shows that replaying only a recent window is worse than no replay at all. In few-shot evaluation the same memory reaches 91.6% on 5-way 1-shot mini-ImageNet, above a task-specific baseline, while a harder 500-way regime exposes the true difficulty. In the second system, the same surprise signal, computed in a shared text-image space, modulates the behaviour of a vision-language model: it answers assertively when a concept is known, hedges when it is partially familiar, and refuses to identify the object and asks for an explanation when it is novel, learning the concept from a single user utterance. The external detector separates known from novel concepts at an AUROC of 0.966 (95% CI +/-0.024), far above the model's own verbalised confidence (0.618), while its token-level confidence sits below chance under greedy decoding; after a sleep phase that empties the fast store, the system recalls 99.2% of fifty taught facts from the consolidated store while a base model recovers none. We report both systems as proof-of-concept, with explicit limitations, and position the second against recent episodic-memory and personalised-VLM work.", "date": "2026-06-30", "categories": []}