{"id": "2510.06048", "title": "BLISS: A Lightweight Bilevel Influence Scoring Method for Data Selection in Language Model Pretraining", "abstract": "Effective data selection is essential for pretraining large language models (LLMs), enhancing efficiency and improving generalization to downstream tasks. However, existing approaches often require leveraging external pretrained models, making it difficult to disentangle the effects of data selection from those of the external pretrained models. In addition, they often overlook the long-term impact of selected data if the model is trained to convergence, primarily due to the prohibitive cost of full-scale LLM pretraining. In this paper, we introduce BLISS (\\textbf{B}ileve\\textbf{L} \\textbf{I}nfluence \\textbf{S}coring method for data \\textbf{S}election): a lightweight data selection method that operates entirely \\emph{from scratch}, without relying on any external pretrained oracle models, while explicitly accounting for the long-term impact of selected data. BLISS leverages a small proxy model as a surrogate for the LLM and employs a score model to estimate the long-term influence of training samples if the proxy model is trained to convergence. We formulate data selection as a bilevel optimization problem, where the upper-level objective optimizes the score model to assign importance weights to training samples, ensuring that minimizing the lower-level objective (i.e., training the proxy model over the weighted training loss until convergence) leads to best validation performance. Once optimized, the trained score model predicts influence scores for the dataset, enabling efficient selection of high-quality samples for LLM pretraining. We validate BLISS by pretraining 410M/1B/2.8B Pythia and LLaMA-0.5B models on selected subsets of the C4 dataset. Notably, under the 1B model setting, BLISS achieves $1.7\\times$ speedup in reaching the same performance as the state-of-the-art method, demonstrating superior performance across multiple downstream tasks.", "date": "2025-10-07", "categories": []}
{"id": "2602.14696", "title": "A Critical Look at Targeted Instruction Selection: Disentangling What Matters (and What Doesn't)", "abstract": "Instruction fine-tuning of large language models (LLMs) often involves selecting a subset of instruction training data from a large candidate pool, using a small query set from the target task. Despite growing interest, the literature on targeted instruction selection remains fragmented and opaque: methods vary widely in selection budgets, often omit zero-shot baselines, and frequently entangle the contributions of key components. As a result, practitioners lack actionable guidance on selecting instructions for their target tasks. In this work, we aim to bring clarity to this landscape by disentangling and systematically analyzing the two core ingredients: data representation and selection algorithms. Our framework enables controlled comparisons across models, tasks, and budgets. We find that only gradient-based data representations choose subsets whose similarity to the query consistently predicts performance across datasets, models, and candidate pools. While no single method dominates, gradient-based representations paired with greedy round-robin selection often perform best on average at low budgets, but these gains diminish at larger budgets. Finally, we unify several existing selection algorithms as forms of approximate distance minimization between the selected subset and the query set, and support this view with new generalization bounds. More broadly, our findings provide critical insights and a foundation for more principled data selection in LLM fine-tuning. The code is available at https://github.com/dcml-lab/targeted-instruction-selection.", "date": "2026-02-16", "categories": []}
{"id": "2606.19411", "title": "Spectral DPPs via NEPv: A Scalable Continuous Relaxation of Determinantal MAP for Diversity-Aware Data Selection", "abstract": "Selecting a small, diverse, high-quality subset from a massive pool of candidates is a recurring primitive in modern machine learning -- data curation and coreset selection for training and fine-tuning large models, active-learning batch acquisition, prompt and exemplar selection for in-context learning, retrieval diversification, and experimental design. Determinantal Point Processes (\\DPP s) give a principled, well-calibrated notion of diversity for this task, but their \\emph{MAP} objective -- pick a size-$k$ subset $S$ maximizing $\\logdet(L_S)$ -- is NP-hard, and the standard greedy and sampling algorithms scale superlinearly in the ground-set size $n$. This cost is prohibitive precisely in the data-centric regime where diversity matters most, where $n$ ranges over millions to billions of candidate examples, features, or embeddings. We recast \\DPP-MAP as a continuous optimization problem over the Stiefel manifold, and show that its first-order optimality conditions form a \\emph{Nonlinear Eigenvalue Problem with eigenvector dependency} (\\NEPv) of a previously unstudied form. This \\NEPv\\ admits a self-consistent field (\\SCF) iteration with a spectral-gap-based local contraction guarantee, giving a principled iterative solver where the diversity objective drives an eigenvector-dependent operator. The resulting algorithm, \\OurMethod, requires only matrix-vector products with the kernel and runs in time $O\\!\\big((ndk+nk^2)\\,t\\big)$ for a small number of iterations $t$, scaling near-linearly in $n$ and integrating directly with low-rank and feature-map kernels common in ML. This paper focuses on the relaxation, solver, and scaling analysis; full real-data benchmarking is left to a planned empirical study.", "date": "2026-06-17", "categories": []}
{"id": "2606.20196", "title": "Distill Once, Adapt Life-Long: Exploring Dataset Distillation for Continual Test-Time Adaptation", "abstract": "Continual Test-Time Adaptation (CTTA) aims to maintain model performance under evolving target domains by adapting online without labeled data. However, practical deployments often cannot retain the source dataset due to privacy or licensing constraints, and purely source-free CTTA methods tend to become unstable under long-term distribution shift, suffering from compounding self-training errors and catastrophic forgetting. We introduce DO-ALL (Distill Once, Adapt Life-Long), a plug-and-play framework that revisits source information in a compact and privacy-conscious form via Dataset Distillation (DD). Before deployment, DO-ALL performs DD to produce a small set of synthetic distilled anchors that summarize the source distribution. During adaptation, each target sample is matched with its most semantically aligned anchor, which provides a stable reference for various CTTA via source replay, representation alignment, and manifold-smoothing regularization. DO-ALL can be seamlessly integrated into existing CTTA algorithms, consistently improving long-term robustness across CIFAR100-C, ImageNet-C, and the CCC benchmark. This demonstrates the potential of leveraging DD to enable stable and continuous adaptation without retaining raw source data. The code is available at https://github.com/blue-531/DOALL.", "date": "2026-06-18", "categories": []}
{"id": "2606.05883", "title": "Geometry-Aware Dataset Condensation for Diffusion Model Training", "abstract": "Dataset condensation aims to construct compact datasets from real data via synthesis or selection. However, existing approaches are ill-suited for diffusion model training: synthetic data generation often yields low-fidelity samples unsuitable for authentic modeling, while real subset selection typically fails to preserve the distributional geometry required by diffusion likelihood objectives. To address this, we propose to reformulate real subset selection as a geometry-aware distribution alignment problem. By incorporating one-sided partial optimal transport, our method selectively aligns a compact subset with the full data distribution while allowing unmatched mass in low-density regions, ensuring the preserved geometric structure necessary for effective diffusion model training. To further ensure distributional fidelity, we complement geometric alignment with lightweight feature-statistics and semantic consistency regularization. An efficient two-stage discrete optimization strategy is proposed to achieve this alignment objective. Extensive experiments across diffusion variants, subset sizes, image resolutions, and training rounds show that our method achieves superior fidelity and distributional coverage in diffusion model training. Codes are available at https://github.com/2018cx/GADC.", "date": "2026-06-04", "categories": []}
{"id": "2606.18663", "title": "RegMix-D: Dynamic Data Mixing via Proxy Training Trajectories", "abstract": "Data mixture selection is critical for Large Language Model pretraining. Existing methods such as RegMix select a single static mixture by fitting a regression model on small-scale proxy runs. We propose RegMix-D, a simple extension of RegMix to dynamic mixing. Our key observation is that proxy runs produce not only endpoint losses, but also full loss trajectories, which can be used to further improve data mixture. By training regression model on these trajectories, we can predict optimal mixtures at multiple training stages. RegMix-D supports two deployment modes: an offline variant that generates a complete mixture schedule before target training, and an online variant that adapts the mixture during training using observed loss. Experiments on 25B tokens of the Pile dataset with a 1B parameter target model show that RegMix-D consistently improves over RegMix and DoReMi across 13 downstream tasks while remaining proxy-efficient: it surpasses RegMix even with only 128 proxy models (25% of RegMix's proxy compute budget).", "date": "2026-06-17", "categories": []}
{"id": "2606.18729", "title": "TimeLAVA: Learning-Agnostic Data Valuation for Time Series", "abstract": "Data valuation quantifies the intrinsic quality of individual samples to enable principled data curation, quality control, and robust learning. For time series in critical domains such as healthcare, finance, and industrial monitoring, effective valuation methods are essential yet fundamentally lacking. Existing approaches are either model-dependent, limiting their generalizability, or designed for i.i.d. data and thus fail to capture temporal dependencies, multi-scale patterns, and non-stationary dynamics inherent to sequential data. We introduce TimeLAVA, a learning-agnostic framework that values temporal segments by their marginal contribution to minimizing distributional discrepancy between evaluated and reference data. At its core is a novel Selective Wavelet-based Wasserstein discrepancy combining multi-scale wavelet transforms for temporal localization with unbalanced optimal transport for robustness to distributional shifts. Segment values are efficiently computed via sensitivity analysis without requiring model training and aggregated into point-wise scores. We provide theoretical guarantees linking valuation to model-agnostic generalization and prove bounded sensitivity to outlier contamination. Extensive experiments across anomaly detection, data pruning, and label noise detection demonstrate that TimeLAVA produces significantly more informative value scores than existing methods on diverse real-world datasets.", "date": "2026-06-17", "categories": []}
{"id": "2602.03300", "title": "R1-SyntheticVL: Is Synthetic Data from Generative Models Ready for Multimodal Large Language Model?", "abstract": "In this work, we aim to develop effective data synthesis techniques that autonomously synthesize multimodal training data for enhancing MLLMs in solving complex real-world tasks. To this end, we propose Collective Adversarial Data Synthesis (CADS), a novel and general approach to synthesize high-quality, diverse and challenging multimodal data for MLLMs. The core idea of CADS is to leverage collective intelligence to ensure high-quality and diverse generation, while exploring adversarial learning to synthesize challenging samples for effectively driving model improvement. Specifically, CADS operates with two cyclic phases, i.e., Collective Adversarial Data Generation (CAD-Generate) and Collective Adversarial Data Judgment (CAD-Judge). CAD-Generate leverages collective knowledge to jointly generate new and diverse multimodal data, while CAD-Judge collaboratively assesses the quality of synthesized data. In addition, CADS introduces an Adversarial Context Optimization mechanism to optimize the generation context to encourage challenging and high-value data generation. With CADS, we construct MMSynthetic-20K and train our model R1-SyntheticVL, which demonstrates superior performance on various benchmarks.", "date": "2026-02-03", "categories": []}
{"id": "2606.11616", "title": "DeMix: Debugging Training Data with Mixed Data Error Types by Investigating Influence Vectors", "abstract": "High-quality training data is essential for the success of machine learning models. However, real-world datasets often contain mixed types of errors arising from systematic flaws in data preparation pipelines, including label errors, feature errors, and spurious correlations. Effective debugging of training data requires both detecting erroneous samples and identifying their specific error types to enable targeted repair, yet existing data cleaning and attribution methods fail to adequately address this dual requirement. In this paper, we propose DeMix, a novel framework that simultaneously diagnoses erroneous samples and their error types. Our key insight is that different error types produce distinct patterns on model behavior. DeMix captures such error-specific patterns by influence vectors that characterize how each training sample affects model predictions across all validation samples. We formulate training data debugging as a multi-label classification problem where a classifier is developed to predict error types directly from influence vectors. We further introduce an intervention-based learning strategy that guides the classifier to capture invariant rationales specific to each error type, ensuring the learned classifier generalizes effectively. Empirical evaluations on 11 tasks across tabular data prediction, recommendation systems, and LLM alignment demonstrate that DeMix significantly outperforms state-of-the-art approaches, achieving a 22.61% improvement in data debugging F1-score and a 9.32% gain in task model performance after data repair. Code is available at: https://github.com/SJTU-DMTai/DeMix.", "date": "2026-06-10", "categories": []}
{"id": "2606.18209", "title": "Rethinking Dataset Distillation for Classification: Do Distilled Sets Outperform Coresets?", "abstract": "Dataset distillation (DD) has emerged as a prominent approach in data centric machine learning, aiming to synthesize compact training sets for efficient training by compressing the information in large datasets into a small number of synthetic samples. However, DD methods are often evaluated under inconsistent evaluation protocols, ranging from standard ERM to single/multi-teacher supervision, making it difficult to isolate the effectiveness of distilled data from evaluation. Moreover, many prior methods claim that DD outperforms data pruning approaches such as coreset selection (CS), based on the assumption that restricting condensed datasets to subsets of real samples fundamentally limits their expressiveness. In this work, we critically evaluate DD methods through large-scale experiments using standardized datasets and evaluation protocols to assess their intrinsic effectiveness. We benchmark seven state-of-the-art (SOTA) DD methods on ImageNet-1K, ImageNet100, and ImageNette, using three widely adopted training protocols against three CS strategies. Our results show that while some DD methods fail to outperform even simple random subsets, the SOTA DD approaches are comparable to or worse than coresets on large-scale datasets and incur a substantially higher cost for construction. Beyond accuracy, we also evaluate the representativeness, diversity, and quality of condensed sets, and find that coresets consistently achieve better coverage of the original data distribution. These findings highlight the limited practical advantages of current DD methods and show that coresets remain competitive and are often a more computationally efficient alternative for data-centric learning.", "date": "2026-06-16", "categories": []}
{"id": "2502.17614", "title": "Scalable Graph Condensation with Evolving Capabilities", "abstract": "The rapid growth of graph data creates significant scalability challenges as most graph algorithms scale quadratically with size. To mitigate these issues, Graph Condensation (GC) methods have been proposed to learn a small graph from a larger one, accelerating downstream tasks. However, existing approaches critically assume a static training set, which conflicts with the inherently dynamic and evolving nature of real-world graph data. This work introduces a novel framework for continual graph condensation, enabling efficient updates to the distilled graph that handle data streams without requiring costly retraining. This limitation leads to inefficiencies when condensing growing training sets. In this paper, we introduce GECC (\\underline{G}raph \\underline{E}volving \\underline{C}lustering \\underline{C}ondensation), a scalable graph condensation method designed to handle large-scale and evolving graph data. GECC employs a traceable and efficient approach by performing class-wise clustering on aggregated features. Furthermore, it can inherit previous condensation results as clustering centroids when the condensed graph expands, thereby attaining an evolving capability. This methodology is supported by robust theoretical foundations and demonstrates superior empirical performance. Comprehensive experiments including real world scenario show that GECC achieves better performance than most state-of-the-art graph condensation methods while delivering an around 1000$\\times$ speedup on large datasets.", "date": "2025-02-24", "categories": []}
{"id": "2510.16882", "title": "Utility-Diversity Aware Online Batch Selection for LLM Supervised Fine-tuning", "abstract": "Supervised fine-tuning (SFT) is a commonly used technique to adapt large language models (LLMs) to downstream tasks. In practice, SFT on a full dataset is computationally expensive and sometimes suffers from overfitting or bias amplification. This facilitates the rise of data curation in SFT, which prioritizes the most valuable data to optimze. This work studies the online batch selection family that dynamically scores and filters samples during the training process. However, existing popular methods often (i) rely merely on the utility of data to select a subset while neglecting other crucial factors like diversity, (ii) rely on external resources such as reference models or validation sets, and (iii) incur extra training time over full-dataset training. To address these limitations, this work develops UDS (Utility-Diversity Sampling), a framework for efficient online batch selection in SFT. UDS leverages the nuclear norm of the logits matrix to capture both data utility and intra-sample diversity, while estimating inter-sample diversity through efficient low-dimensional embedding comparisons with a lightweight memory buffer of historical samples. Such a design eliminates the need for external resources and unnecessary backpropagation, securing computational efficiency. Experiments on multiple benchmarks demonstrate that UDS consistently outperforms state-of-the-art online batch selection methods under varying data budgets, and significantly reduces training time compared to full-dataset fine-tuning. Code is available at https://github.com/gfyddha/UDS.", "date": "2025-10-19", "categories": []}
{"id": "2602.09326", "title": "Priority-Aware Shapley Value", "abstract": "Shapley values are widely used for model-agnostic data valuation and feature attribution, yet they implicitly assume contributors are interchangeable. This can be problematic when contributors are dependent (e.g., reused/augmented data or causal feature orderings) or when contributions should be adjusted by factors such as trust or risk. We propose Priority-Aware Shapley Value (PASV), which incorporates both hard precedence constraints and soft, contributor-specific priority weights. PASV is applicable to general precedence structures, recovers precedence-only and weight-only Shapley variants as special cases, and is uniquely characterized by natural axioms. We develop an efficient adjacent-swap Metropolis-Hastings sampler for scalable Monte Carlo estimation and analyze limiting regimes induced by extreme priority weights. Experiments on data valuation (MNIST/CIFAR10) and feature attribution (Census Income) demonstrate more structure-faithful allocations and a practical sensitivity analysis via our proposed \"priority sweeping\".", "date": "2026-02-10", "categories": []}
{"id": "2606.07678", "title": "DOG-DPO:Dynamic Optimization in Geometry for Safety Alignment", "abstract": "Safety alignment for large language models relies on preference data, but current pipelines often train on large, redundant datasets. Existing data selection methods typically score each preference pair independently, collapsing directional preference information into scalar quality or diversity scores. This sample-centric view is especially limiting in multi-dataset settings, where shared safety directions coexist with dataset-specific residual risks. We propose DOG-DPO, a training-free data selection framework that treats preference pairs as structured geometric signals. DOG-DPO first represents each preference pair as a direction in model representation space. It then decomposes multi-dataset preference geometry into a global anchor subspace and dataset-specific residual subspaces. Finally, it selects subsets by maximizing diversity-based coverage, encouraging broad, non-redundant coverage of alignment directions before DPO training. Across six safety benchmarks and two model backbones, DOG-DPO achieves a strong utility-robustness trade-off using only 11% of the preference pairs. It recovers most of the safety gains of full-data training while remaining entirely teacher-free, training-free, and substantially faster than representative selection baselines.", "date": "2026-06-04", "categories": []}
{"id": "2606.14971", "title": "FastMix: Fast Data Mixture Optimization via Gradient Descent", "abstract": "While large and diverse datasets have driven recent advances in large models, identifying the optimal data mixture for pre-training and post-training remains a significant open problem. We address this challenge with FASTMIX, a novel framework that automates data mixture discovery while training only a single proxy model. Instead of relying on predefined heuristics or resource-intensive simulations, FASTMIX jointly optimizes mixture coefficients and model parameters, substantially improving efficiency and scalability over prior approaches. At the core of FASTMIX is a reformulation of mixture selection as a bilevel optimization problem. Under this reformulation, we show that optimizing mixture ratios is mathematically equivalent to assigning per-source loss weights under uniform source sampling. This embeds the mixture coefficients directly into the differentiable iterative optimization objective, enabling efficient, gradient-based optimization of both mixture and model. To solve the optimization problem, FASTMIX implements an approximate iterative optimization procedure, alternating between (i) updating model parameters on data sampled according to current mixture ratios (inner loop) and (ii) updating mixture ratios based on validation feedback (outer loop). Across pre- and post-training, FASTMIX outperforms baselines while drastically reducing search cost. Code (https://github.com/hrtan/fastmix)", "date": "2026-06-12", "categories": []}
{"id": "2606.15216", "title": "Spokes: Optimizing for Diverse Pretraining Data Selection", "abstract": "Diversity plays a critical role in data selection, improving performance under fixed data budgets by reducing redundancy and repetition. However, optimizing for diversity is inherently challenging, as it is a set-level property that depends on interactions between data points rather than individual examples. As a result, existing approaches typically rely on proxies or approximations, which often fail to ensure sufficiently diverse subsets. In this work, we directly optimize diversity by introducing a probabilistic diversification framework based on the G-Vendi score, optimized via exponentiated gradient descent. Our method produces subsets that are substantially more diverse than those obtained via random sampling, achieving a +489 increase in G-Vendi score on a 500k-sample subset. We evaluate our approach on FineWeb and DCLM, where it consistently outperforms existing methods. Notably, SPOKES (diversity-only) improves average downstream performance by +0.4 and +0.5 points over random sampling on DCLM and FineWeb, respectively. More importantly, jointly optimizing for both quality and diversity yields the strongest results: SPOKES achieves gains of +1.5 and +1.4 points on DCLM and FineWeb, outperforming all baselines, including semantic deduplication and quality filtering.", "date": "2026-06-13", "categories": []}
{"id": "2606.13668", "title": "Influcoder: Distilling Decoders' Gradient Influence Rankings into an Encoder for Data Attribution", "abstract": "With the growth of LLMs' (Large Language Models) capabilities, there has been an increasing push to curate high quality datasets by filtering samples in the training data. In general, Data Attribution (DA) methods aim to estimate how individual samples in a training dataset can precondition a model to generate certain outputs. As an example, one might be interested in which samples in the data could be the source of toxic behavior after training the LLM. Many methods quantify this conditioning through the paradigm of influence functions. While methods of this family are effective in its function, they lack the necessary processing speed and storage compactness to be practically implemented on large datasets. We propose a method, Influcoder, as a quick and cost-effective approach to influence-based Data Attribution at scale.", "date": "2026-06-11", "categories": []}
{"id": "2510.06596", "title": "SDQM: Synthetic Data Quality Metric for Object Detection Dataset Evaluation", "abstract": "The performance of machine learning models depends heavily on training data. The scarcity of large-scale, well-annotated datasets poses significant challenges in creating robust models. To address this, synthetic data generated through simulations and generative models has emerged as a promising solution, enhancing dataset diversity and improving the performance, reliability, and resilience of models. However, evaluating the quality of this generated data requires an effective metric. We introduce the Synthetic Dataset Quality Metric (SDQM) to assess data quality for object detection tasks without requiring model training to converge. This metric enables more efficient generation and selection of synthetic datasets, addressing a key challenge in resource-constrained object detection tasks. In our experiments, SDQM demonstrated a strong correlation with the mean average precision (mAP) scores of YOLO11, a leading object detection model, whereas previous metrics only exhibited moderate or weak correlations. In addition, it provides actionable insights into improving dataset quality, minimizing the need for costly iterative training. This scalable and efficient metric sets a new standard for evaluating synthetic data. The code for SDQM is available at https://github.com/ayushzenith/SDQM", "date": "2025-10-08", "categories": []}
{"id": "2606.07001", "title": "DataEvolver: Automatic Data Preparation for Large Language Models through Multi-Level Self-Evolving", "abstract": "High-quality training data is essential to large language models (LLMs) and typically requires extensive and costly manual curation. Existing automatic data preparation methods rely on predefined pipelines or customized human instructions, which limits their adaptability to diverse data distributions and lacks principled guidance from high-quality examples. In this paper, we introduce DataEvolver, the first self-evolving data preparation system that automatically constructs pipelines to transform raw data into high-quality data. DataEvolver employs a multi-level mechanism to ensure both pipeline executability and effectiveness. At the operator level, it incrementally expands the operator set to construct a logical plan while resolving dependency conflicts. At the pipeline level, it instantiates logical plans into executable code and iteratively refines pipeline orchestration through a feedback loop that reduces the distribution gap between prepared data and high-quality examples. Experiments on seven benchmarks show that DataEvolver substantially improves data quality and achieves an average 10\\% gain in downstream LLM performance compared with training on original data, highlighting new opportunities for the iterative co-evolution of LLMs and data.", "date": "2026-06-05", "categories": []}
{"id": "2606.11499", "title": "Hubs or Fringes: Pretraining Data Selection via Web Graph Centrality", "abstract": "The performance of modern language models depends critically on pretraining data composition. Yet existing data selection methods rely on auxiliary classifiers for document scoring or mixture optimization, adding computational overhead and dependence on labeled data. We propose WebGraphMix, a lightweight data selection framework that computes structural centrality scores over the Common Crawl host-level web graph and uses them to vary the proportion of central versus peripheral documents in the pretraining mixture. We hypothesize that central hosts expose models to reusable abstractions, while peripheral hosts encode specialized, long-tail knowledge. WebGraphMix computes centrality scores efficiently at web scale, requiring no model training, labeled data, or downstream supervision. We integrate WebGraphMix into the DataComp-LM pipeline and train models at 400M and 1B parameter scales with 8B and 28B tokens respectively, evaluating on 23 tasks ranging from factual knowledge to symbolic reasoning. Our experiments show that central and peripheral web regions encode complementary capabilities. Mixture combining both at a ratio of 1:1 achieves 41.4% on average, compared to 39.8% for uniform sampling. Combining structural scores with document-level quality classifier scores further improves performance to 43.8%. These findings demonstrate that web graph topology is a meaningful axis for pretraining data curation, capturing information that is largely orthogonal to existing content-based approaches.", "date": "2026-06-09", "categories": []}
{"id": "2606.11761", "title": "RCAP: Robust, Class-Aware, Probabilistic Dynamic Dataset Pruning", "abstract": "Dynamic data pruning techniques aim to reduce computational cost while minimizing information loss by periodically selecting representative subsets of input data during model training. However, existing methods often struggle to maintain strong worst-group accuracy, particularly at high pruning rates, across balanced and imbalanced datasets. To address this challenge, we propose RCAP, a Robust, Class-Aware, Probabilistic dynamic dataset pruning algorithm for classification tasks. RCAP applies a closed-form solution to estimate the fraction of samples to be included in the training subset for each individual class. This fraction is adaptively adjusted in every epoch using class-wise aggregated loss. Thereafter, it employs an adaptive sampling strategy that prioritizes samples having high loss for populating the class-wise subsets. We evaluate RCAP on six diverse datasets ranging from class-balanced to highly imbalanced using five distinct models across three training paradigms: training from scratch, transfer learning, and fine-tuning. Our approach consistently outperforms state-of-the-art dataset pruning methods, achieving superior worst-group accuracy at all pruning rates. Remarkably, with only $10\\%$ data, RCAP delivers $>1\\%$ improvement in performance on class-imbalanced datasets compared to full data training while providing an average $8.69\\times$ speedup. The code can be accessed at https://github.com/atif-hassan/RCAP-dynamic-dataset-pruning", "date": "2026-06-10", "categories": []}
{"id": "2505.11034", "title": "CleanPatrick: A Benchmark for Image Data Cleaning", "abstract": "Robust machine learning depends on clean data, yet current image data cleaning benchmarks rely on synthetic noise or narrow human studies, limiting comparison and real-world relevance. We introduce CleanPatrick, the first large-scale benchmark for data cleaning in the image domain, built upon the publicly available Fitzpatrick17k dermatology dataset. We collect 496,377 binary annotations from 933 medical crowd workers, identify off-topic samples (4%), near-duplicates (21%), and label errors (32%), and employ an aggregation model inspired by item-response theory followed by expert review to derive high-quality ground truth. CleanPatrick formalizes issue detection as a ranking task and employs standard ranking metrics that mirror real audit workflows. We benchmark classical anomaly detectors, perceptual hashing, SSIM, Confident Learning, NoiseRank, FINE, BHN, and SelfClean. On CleanPatrick, self-supervised representations excel at near-duplicate detection, classical methods achieve competitive off-topic detection under constrained review budgets, and detecting implausible labels under conservative human judgment remains challenging for fine-grained medical classification. By releasing both the dataset and the evaluation framework, CleanPatrick enables a systematic comparison of image-cleaning strategies.", "date": "2025-05-16", "categories": []}
{"id": "2510.12071", "title": "Influence Dynamics and Stagewise Data Attribution", "abstract": "Current training data attribution (TDA) methods treat the influence one sample has on another as static, but neural networks learn in distinct stages that exhibit changing patterns of influence. In this work, we introduce a framework for stagewise data attribution grounded in singular learning theory. We predict that influence can change non-monotonically, including sign flips and sharp peaks at developmental transitions. We first validate these predictions analytically and empirically in a toy model, showing that dynamic shifts in influence directly map to the model's progressive learning of a semantic hierarchy. Finally, we demonstrate these phenomena at scale in language models, where token-level influence changes align with known developmental stages.", "date": "2025-10-14", "categories": []}
{"id": "2506.20573", "title": "LARP: Learner-Agnostic Robust Data Prefiltering", "abstract": "Public datasets, crucial for modern machine learning and statistical inference, often contain low-quality or contaminated samples that can harm model performance. This creates a need for principled prefiltering procedures that a data provider can apply to protect the accuracy of a range of potential downstream statistical and learning procedures simultaneously. In this work, we formalize and analyze Learner-Agnostic Robust data Prefiltering (LARP), the problem of designing prefiltering procedures with guarantees on the worst-case loss over a pre-specified set of learners. We establish the feasibility of LARP in two theoretical settings, by providing upper-bound guarantees on the worst-case loss. Our theoretical results indicate that protecting heterogeneous learner sets via LARP comes at the price of some performance loss compared to individual, learner-specific prefiltering; we call this gap the price of LARP. To assess this gap in performance, we empirically measure the price of LARP across image and tabular tasks. We further explore potential benefits of LARP from the perspective of saving on repeated data curation efforts, in a game-theoretic model where the downstream learners can split the cost of the single prefiltering.", "date": "2025-06-25", "categories": []}
{"id": "2605.30407", "title": "Exploring Autonomous Agentic Data Engineering for Model Specialization", "abstract": "Large Language Models (LLMs) have demonstrated strong performance on general tasks, while often struggling to adapt to specialized domains without high-quality domain-specific data. Existing LLM-based data curation methods primarily rely on human-designed workflows, leaving it unexamined whether LLMs can autonomously execute an end-to-end data engineering pipeline for model specialization. We formalize Autonomous Agentic Data Engineering, a novel task designed to evaluate LLMs as autonomous data engineers that drive model specialization through end-to-end data curation. We frame data as an optimizable component and study agents that plan, generate, and iteratively optimize training data across multiple domains, guided by post-training performance improvement. Experiments show that autonomous LLM data engineers yield substantial gains, as GPT-5.2 constructs a training curriculum that improves a student model by 57.29%, entirely through iterative, agent-driven data adaptation. By illuminating both potential and bottlenecks, our study establishes autonomous data engineering as a measurable capability and charts a path toward agent-driven model specialization (Code will be released at https://github.com/zjunlp/DataAgent).", "date": "2026-05-28", "categories": []}
{"id": "2606.07597", "title": "Repetition Mismatch: Why Data Mixture Experiments Don't Scale and How to Fix Them", "abstract": "Pre-training data mixtures are commonly tuned by running small-scale experiments and extrapolating to the target training budget. When high-quality data is scarce and must be repeated, this extrapolation frequently fails, but the source of the failure has not been isolated. We show that a primary culprit is a repetition mismatch: because high-quality datasets are small, their repetition rate changes as the training budget grows, shifting the optimal mixture in ways that small-scale proxy experiments do not anticipate. A subsampling procedure that matches the target repetition rate controls for this effect. In a two-source setting combining limited high-quality data with web crawl, a single repetition-controlled experiment using only 1/16 of the target tokens recovers a mixture within 0.05 of the optimum for a 757M parameter model, compared to an error of 0.75 without repetition control. Achieving comparable accuracy without repetition control requires three to four horizons, consuming 44 to 94% of the target token budget. With three data sources, the larger mixture space requires more than a single experiment to constrain, but the approach remains effective: at the 757M scale, just two repetition-controlled horizons recover the optimal mixture, outperforming baselines that instead require the full two-source experiments to construct. Our results reveal that repetition dynamics, not scale alone, shape whether small-scale mixture experiments generalize. More broadly, they suggest that data repetition deserves treatment as a first-class variable in mixture optimization, rather than an inconvenient side effect of limited data.", "date": "2026-05-29", "categories": []}
{"id": "2606.07778", "title": "Unlocking Latent Value: Taxonomy-Guided Recovery of High-Performing Data from Low-Tier Web Corpora", "abstract": "Dominant web data curation pipelines for pretraining collapse document quality into a single composite score, systematically missing high-value content along dimensions the scorer underweights. We present a taxonomy-driven framework that recovers this value by filtering along semantically meaningful dimensions that composite scores fail to capture. First, building on the ESSENTIAL-WEB taxonomy, we introduce two novel dimensions: timeliness and cultural specificity, both of which show low pairwise NMI with existing ones. We annotate 14M documents using Qwen2.5 32B and distill into a lightweight 0.5B model. To enable rapid corpus-wide annotation, we additionally train a 73M multi-task MLP on E5 embeddings, achieving 50x inference throughput. Second, to navigate the combinatorial explosion of filter configurations, we introduce a compute-efficient two-pass framework: Pass 1 identifies the strongest dimension signals at small scale; Pass 2 constructs and evaluates conjunctive and disjunctive compound filters from the top performers - identifying high-performing configurations at a fraction of full scaling-law cost. Applying the selected filters to deprioritized web data, taxonomy-filtered subsets outperform their unfiltered baselines and even surpass the highest-quality tier. On mid-tier data, our best filter improves over its unfiltered baseline by 12.1% on reasoning, 9.5% on coding, and 2.0% on knowledge benchmarks, exceeding unfiltered top-tier data by 6.7% on reasoning and 13.7% on coding. Furthermore, filtered data from two tiers below the typical production threshold improves by 22.3% on reasoning and 19.5% on coding over its unfiltered baseline, surpassing top-tier data on coding benchmarks. These results establish that vast latent value remains locked in deprioritized web data, and that multi-dimensional taxonomy filtering is a principled, compute-efficient key to unlocking it.", "date": "2026-06-05", "categories": []}
{"id": "2502.06434", "title": "Unifying Dataset Pruning and Distillation for Efficient Large-scale Compression", "abstract": "Dataset pruning (DP) and dataset distillation (DD) fundamentally differ in their outputs: DP selects original image subsets, while DD generates synthetic images. Recently, DD's increasing reliance on original images suggests a convergence of the two directions. To investigate this convergence trend, we propose a unified dataset compression (DC) benchmark. This benchmark reveals an interesting trade-off for soft-label-DD: while soft labels provide valuable information, they can make the distillation process less essential, as distilled images may not always outperform random subsets. In addition, the benchmark reveals that in current stages, dataset pruning outperforms dataset distillation at small dataset sizes. Given these observations, we explore hard-label-DC as a complementary approach that emphasizes image quality while offering substantial storage efficiency. Our PCA (Prune, Combine, and Augment) is the first framework that does not rely on soft labels but instead focuses on image quality. (1) \"P'' means selecting easy samples based on dataset pruning metrics, (2) \"C'' indicates combining these samples effectively, and (3) \"A'' is to apply constrained image augmentation during training. Our code is available at https://github.com/ArmandXiao/Unifying-Dataset-Pruning-and-Distillation", "date": "2025-02-10", "categories": []}
{"id": "2606.04261", "title": "Can Generalist Agents Automate Data Curation?", "abstract": "Curating training data is among the most consequential yet labor-intensive parts of modern AI development: practitioners iteratively propose, implement, evaluate, and revise data policies against noisy benchmark feedback. We ask whether generalist coding agents can automate this data-curation loop. We introduce *Curation-Bench*, an agent-centric benchmark that fixes the model, training recipe, and evaluation suite while giving agents command-line access to inspect data, implement policies, submit them to a fixed training/evaluation pipeline, and revise. In a vision-language instruction-tuning instantiation, out-of-the-box agents reach strong published data-selection baselines within ten iterations. However, trajectory analysis reveals a persistent *execution-research gap*: agents mainly tune local policy variants rather than explore new policy families, even when given strategy guides and paper references. Scaffolds requiring each iteration to cite, instantiate, and adapt a prior method shift agents toward method-guided exploration. The scaffolded agent autonomously composes -- without human design input -- a data-selection policy that outperforms strong published baselines at one-tenth their data budget. Overall, current agents can run the curation loop, but reliable data research requires scaffolded method adaptation, not open-ended prompting alone. Code and benchmark are open-sourced.", "date": "2026-06-02", "categories": []}
{"id": "2606.04820", "title": "OA-CutMix: Correcting the Label Bias of CutMix", "abstract": "CutMix has become the de facto standard mixing augmentation, yet its label assignment rests on a flawed assumption: The area of the pasted patch faithfully reflects its semantic contribution to the mixed image. In practice, however, patches frequently land on background regions, assigning label credit to classes whose objects are not visible. The mean discrepancy of the CutMix label and the semantic object area is $21.5\\%$. In $17\\%$ of samples an image contributes zero visible object pixels yet receives nonzero label weight. We propose Object-Aware CutMix (OA-CutMix), which corrects this bias by replacing the area-based CutMix weight with one derived from precomputed segmentation masks, assigning labels in proportion to the visible object area each image contributes to the mix. The image mixing procedure is left entirely unchanged. We evaluate OA-CutMix against 10+ static and dynamic mixing methods across 4 architectures and 6 datasets. OA-CutMix consistently achieves the highest accuracy over all tasks, outperforming even dynamic mixing methods, but at a fraction of the training-time cost. Improvements are largest for small objects, where the label bias from CutMix is greatest. Thus, correcting the label is sufficient to match or exceed the performance of methods modifying the image mixing algorithm.", "date": "2026-06-03", "categories": []}
{"id": "2606.03839", "title": "Text-attributed Graph Condensation via Text Selection and Attribute Matching", "abstract": "Text-Attributed Graph (TAG) is an important type of graph structured data, where each node has a text description. TAG models usually train a Graph Neural Network (GNN) and language model jointly, which leads to high space and time consumption, especially on large datasets. To mitigate this, we propose TAGSAM, a condensation method that compresses TAGs while preserving training accuracy. TAGSAM comes with two key designs, i.e., subgraph text Selection and Attribute similarity Matching, which compress the text description and graph topology of TAG, respectively. For the texts, subgraph text selection selects and merges representative text chunks from multiple related text descriptions by maximizing mutual information. For the graph topology, popular condensation methods based on Matching Training Trajectories (MTT) suffer from high variance, which hinders accuracy. Our attribute similarity matching mitigates this issue by aligning stable similarity matrices. We evaluate TAGSAM against six state-of-the-art baselines, where it showcases superior performance. For the same compressed size, TAGSAM improves upon the best-performing baseline by an average of 4.9% in accuracy. Furthermore, it maintains competitive training accuracy even when the TAG is condensed to just 1% size. Our code is available at https://github.com/SundayVHan/TAGSAM", "date": "2026-06-02", "categories": []}
{"id": "2512.15647", "title": "Hard Labels In! Rethinking the Role of Hard Labels in Mitigating Local Semantic Drift", "abstract": "Soft labels from teacher models are a de facto practice for knowledge transfer and large-scale dataset distillation (e.g., SRe2L, LPLD). However, when we limit the number of crops per image to reduce the substantial cost of storing precomputed soft labels, these methods suffer severely from local semantic drift: visually ambiguous crops can cause soft supervision to deviate from the image-level ground-truth semantics, leading to persistent errors and a train-test distribution mismatch. We revisit the overlooked role of hard labels and show that, when properly integrated, they can act as a content-invariant semantic anchor that calibrates such drift. We theoretically analyze the emergence of drift under sparse soft-label supervision and demonstrate that hybridizing hard and soft labels restores alignment between visual content and semantic supervision. Building on this insight, we propose a new training paradigm, Hard Label for Alleviating Local Semantic Drift (HALD), which uses hard labels as intermediate corrective signals while preserving the fine-grained benefits of soft labels. Extensive experiments on dataset distillation and large-scale classification benchmarks show consistent generalization improvements. On ImageNet-1K, our method achieves 42.7% accuracy with only 285M soft-label storage (reduces by 100X), outperforming prior state-of-the-art LPLD 9.0%.", "date": "2025-12-17", "categories": []}
{"id": "2601.22276", "title": "SurrogateSHAP: Training-Free Contributor Attribution for Text-to-Image (T2I) Models", "abstract": "As Text-to-Image (T2I) diffusion models are increasingly used in real-world creative workflows, a principled framework for valuing contributors who provide a collection of data is essential for fair compensation and sustainable data marketplaces. While the Shapley value offers a theoretically grounded approach to attribution, it faces a dual computational bottleneck: (i) the prohibitive cost of exhaustive model retraining for each sampled subset of players (i.e., data contributors) and (ii) the combinatorial number of subsets needed to estimate marginal contributions due to contributor interactions. To this end, we propose SurrogateSHAP, a retraining-free framework that approximates the expensive retraining game through inference from a pretrained model. To further improve efficiency, we employ a gradient-boosted tree to approximate the utility function and derive Shapley values analytically from the tree-based model. We evaluate SurrogateSHAP across three diverse attribution tasks: (i) image quality for DDPM-CFG on CIFAR-20, (ii) aesthetics for Stable Diffusion on Post-Impressionist artworks, and (iii) product diversity for FLUX.1 on Fashion-Product data. Across settings, SurrogateSHAP outperforms prior methods while substantially reducing computational overhead, consistently identifying influential contributors across multiple utility metrics. Finally, we demonstrate that SurrogateSHAP effectively localizes data sources responsible for spurious correlations in clinical images, providing a scalable path toward auditing safety-critical generative models.", "date": "2026-01-29", "categories": []}
{"id": "2602.19789", "title": "Position: Stop Preaching and Start Practising Data Frugality for Responsible Development of AI", "abstract": "This position paper argues that the machine learning community must move from preaching to practising data frugality for responsible artificial intelligence (AI) development. For too long, progress has been equated with ever-larger datasets, driving remarkable advances but now yielding increasingly diminishing performance gains alongside rising energy use and carbon emissions. While awareness of data frugal approaches has grown, their adoption has remained rhetorical, and data scaling continues to dominate development practice. We argue that this gap between preach and practice must be closed, as continued data scaling entails substantial and under-accounted environmental impacts. To ground our position, we provide indicative estimates of the energy use and carbon emissions associated with the downstream use of ImageNet-1K. We then present empirical evidence that data frugality is both practical and beneficial, demonstrating that subset selection methods can substantially reduce training energy consumption with little loss in accuracy, while also mitigating dataset bias. Finally, we outline actionable recommendations for moving data frugality from rhetorical preaching to concrete practice for responsible development of AI.", "date": "2026-02-23", "categories": []}
{"id": "2606.01920", "title": "Pool-Select-Refine: Allocation-Aware Generative Dataset Distillation with Soft-Label-Guided Latent Refinement", "abstract": "Diffusion-based dataset distillation has recently emerged as a promising paradigm for condensing large-scale datasets into compact synthetic sets. By leveraging pretrained generative priors, these methods can produce realistic class-conditional samples more efficiently than traditional matching-based approaches. However, most existing diffusion-based methods still adopt a rigid ``Generate-and-Use'' strategy, where the generated samples are directly treated as the final distilled set under a fixed images-per-class budget. Such a design tightly couples candidate generation with final budget allocation, which may result in redundant waste of the limited budget or insufficiently informative samples. In this paper, we propose ``Pool-Select-Refine'', a two-stage framework for allocation-aware generative dataset distillation. First, instead of directly using a fixed number of generated samples, we construct an over-complete candidate pool and select a compact subset under the target budget. Second, we refine the selected samples in latent space using soft-label supervision derived from the teacher model, improving semantic alignment while preserving the generative prior. This design explicitly decouples generation, selection, and refinement, enabling more effective use of the distillation budget. Experiments on large-scale and fine-grained image classification benchmarks show that the proposed framework delivers consistent gains over diffusion-based baselines. The results suggest that introducing a curation stage before refinement is a simple yet effective way to improve diffusion-based dataset distillation.", "date": "2026-06-01", "categories": []}
{"id": "2502.04554", "title": "Unifying and Optimizing Data Values for Selection via Sequential Decision-Making", "abstract": "Data selection has emerged as a crucial downstream application of data valuation, yet the theoretical foundations for using data values in selection remain underexplored. We reformulate data selection as a sequential decision-making problem where the optimal selection sequence arises from dynamic programming, and data values can be understood as encodings of this optimal sequence. This framework unifies and reinterprets existing methods like Data Shapley through the lens of approximate dynamic programming, revealing them as myopic linear approximations to the sequential problem. We further analyze how selection optimality degrades with utility curvature under submodularity, explaining when and why these approximations fail. To bridge theory and practice, we propose an efficient bipartite graph-based surrogate that preserves submodular structure while enabling scalable greedy selection with provable guarantees. Experiments on classical ML benchmarks and large-scale LLM fine-tuning data selection demonstrate substantial improvements over existing methods. Code is publicly available at https://github.com/frankhlchi/SeqDataVal", "date": "2025-02-06", "categories": []}
{"id": "2602.10388", "title": "Less is Enough: Synthesizing Diverse Data in LLM Feature Space with Sparse Autoencoders", "abstract": "The diversity of post-training data is critical for effective downstream performance in large language models (LLMs). Many existing approaches to constructing post-training data quantify diversity using text-based metrics that capture linguistic variation, but such metrics provide only weak signals for the task-relevant features that determine downstream performance. In this work, we introduce Feature Activation Coverage (FAC) which measures data diversity in an interpretable feature space. Building upon this metric, we further propose a diversity-driven data synthesis framework, named FAC Synthesis, that first uses a sparse autoencoder to identify missing features from a seed dataset, and then generates synthetic samples that explicitly reflect these features. Experiments show that our approach consistently improves both data diversity and downstream performance on various tasks, including instruction following, toxicity detection, reward modeling, and behavior steering. Interestingly, we identify a shared, interpretable feature space across model families (i.e., LLaMA, Mistral, and Qwen), enabling cross-model knowledge transfer. Our work provides a solid and practical methodology for exploring data-centric optimization of LLMs.", "date": "2026-02-11", "categories": []}
{"id": "2605.26121", "title": "GEM: Geometric Entropy Mixing for Optimal LLM Data Curation", "abstract": "LLM pre-training efficacy increasingly depends on data composition rather than sheer volume. Yet, optimal mixing is hindered by categorization flaws: human taxonomies suffer from ontological misalignment, and Euclidean clustering fails to address embedding anisotropy. We introduce GEM (Geometric Entropy Mixing), a framework reformulating data curation as a variational problem on the hypersphere augmented with a mixing-balance regularizer. By decoupling the generative prior and optimizing the objective via a provable MM (Minorize-Maximize) algorithm, GEM effectively counteracts the cluster collapse to discover balanced semantic structures invisible to Euclidean heuristics. We employ teacher-student distillation to scale this geometric fidelity to web-scale corpora and introduce the Geometric Influence Score (GIS) for interpretable taxonomy generation. Experiments with 1.1B-parameter models demonstrate that GEM establishes a new state-of-the-art when integrated into mixing strategies like DoReMi and RegMix, improving average downstream accuracy by up to 1.2% and offering a robust coordinate system for predictable data mixing.", "date": "2026-04-27", "categories": []}
{"id": "2605.30288", "title": "MIRA: Mid-training Rubric Anchoring for Source-Aware Data Selection", "abstract": "Mid-training has become an important stage in modern LLM development, using large-scale curated mixtures to strengthen capabilities before final post-training. Its data selection problem is distinct: the data are optimized under a pretraining-style objective at near-pretraining scale, but are curated toward downstream capabilities and drawn from heterogeneous sources with different formats and training roles. As a result, effective selection requires both scalability and source-adaptive semantic criteria. Existing model-based methods scale well, but provide only implicit quality signals. Semantic selection methods offer stronger judgments, but usually assume fixed rubrics or standardized data formats. To address this mismatch, we propose MIRA, a source-aware filtering framework based on self-anchored rubric discovery. The key idea is to make rubric construction part of data selection: MIRA first discovers what should be evaluated for each source group, then distills those judgments into scalable student scorers for full-corpus filtering. On code-oriented mid-training with 21 sources and 5 source groups, MIRA outperforms selection baselines across nine code benchmarks and matches the full-corpus run while using only half the tokens.", "date": "2026-05-28", "categories": []}
{"id": "2605.30857", "title": "MADS: Model-Aware Diverse Core Set Selection for Instruction Tuning", "abstract": "Instruction fine-tuning is employed to enhance the instruction-following ability of large language models (LLMs). As the amount of instruction fine-tuning data increases, selecting the optimal core set becomes particularly important. However, ensuring the diversity of the core set remains a significant challenge. Existing methods predominantly distinguish different training data based on the text features themselves, decoupled from LLMs' own understanding and representation of the data. To address this issue, we propose a Model-Aware Diverse Core Set Selection method, which distinguishes data features based on the neural activation states during LLM inference. This approach serves as an efficient instantiation of coverage-based selection using model-intrinsic activation features to ensure the diversity in the core set. We extensively evaluate our method on six benchmarks that cover five distinct tasks. In our method, the core set selected by the 3B-parameter LLM performs effectively when utilized to fine-tune larger models with 7B, 8B, and 13B parameters. Experimental results on the Alpaca-GPT4 dataset, which comprises 52K instruction-response pairs, show that the core set, sized at 15\\% of the original dataset and selected by Llama-3.2-3B-Instruct, achieves an average improvement of 2.5\\% when fine-tuning four larger base models compared with training on the full dataset. The experimental results demonstrate that our method enhances model performance on multiple downstream tasks while reducing data requirements.", "date": "2026-05-29", "categories": []}
{"id": "2605.29448", "title": "How Much Is a Dataset Worth? Scaling Laws, the Vendi Score, and Matrix Spectral Functions", "abstract": "Neural scaling laws appraise data through dataset size, while the Vendi Score uses quantum entropy to measure dataset value. We show both that common neural-scaling-law objectives and the Vendi Score are submodular. We further show that the Vendi Score is a special case of a broader class of submodular objectives that we call matrix spectral functions. This also includes determinantal (DPP) objectives, as well as many others. We also introduce weakly matrix monotone functions and show how they lead to weakly submodular matrix spectral functions, yielding a broad family of practical objectives for data appraisal. We develop secular-equation-based updates that avoid repeated eigendecompositions during greedy optimization, reducing marginal-gain evaluation for $m$-dimensional embeddings by an $O(m)$ factor relative to oracle queries. This yields an average empirical speedup of about 35,000x, making direct optimization of the Vendi Score feasible on ImageNet-1K-scale datasets. Thus enabled, we compare how well several objectives predict the value of training subsets for held-out test performance under fixed-size, class-balanced, and fixed training-budget regimes, including the Vendi Score, DPPs, facility location, and three new matrix spectral variants. Across multiple datasets, facility location performs the best. Direct optimization also reveals that, while the Vendi Score is predictive over moderate score ranges, pushing the objective to higher values can make it a poor downstream performance proxy. We also find that uniformly at random fixed-size subsets, both unconstrained and class-balanced, are remarkably concentrated in both appraisal scores and held-out performance. Finally, we show that size, class balance, and training budget do not alone determine data value: even when controlling for these factors, performance ranges smoothly from good to bad.", "date": "2026-05-28", "categories": []}
{"id": "2605.29807", "title": "Data filtering methods for training language models", "abstract": "Data quality is a critical factor in the effectiveness of machine learning models. Label errors, present even in widely used benchmarks, introduce noise into training data and reduce model generalization. In this work, we conduct a comparative analysis of two automatic label error detection methods - Confident Learning and Dataset Cartography - on three Russian text classification corpora of varying size, number of classes, and domain: ru_emotion_e-culture (49,123 examples, emotion classification), RuCoLA (8,524 examples, linguistic acceptability), and TERRa (2,337 examples, textual entailment recognition). We use the pre-trained rubert-base-cased model fine-tuned on each corpus. To verify the meaningfulness of filtering, we conduct control experiments with random removal of an equivalent number of examples. Results show that the effectiveness of both methods depends strongly on dataset characteristics: on large corpora with low noise levels, filtering does not improve performance, while on small datasets with high noise, Confident Learning achieves a significant F1-macro improvement. Dataset Cartography demonstrates more conservative behavior, removing fewer examples. Across all corpora, targeted removal by both methods outperforms random removal, confirming the meaningfulness of the approaches.", "date": "2026-05-28", "categories": []}
{"id": "2605.30334", "title": "Demystifying Data Organization for Enhanced LLM Training", "abstract": "Large Language Models (LLMs) have revolutionized various fields, yet their training efficiency is heavily reliant on effective data curation. While data selection has been widely studied, the strategic data organization for enhanced training remains an underexplored area, particularly since current LLMs are often trained for only one or a few epochs. This paper systematically explores the influence of data organization on LLM training by reusing pre-computed sample-level scores originally generated for data efficiency, thereby incurring minimal additional computational overhead. We identify and formalize four key guidelines for optimizing data organization: Boundary Sharpening, Cyclic Scheduling, Curriculum Continuity, and Local Diversity. Guided by them, we introduce two novel data ordering methods termed STR and SAW. Extensive experiments across different model scales and data sizes, encompassing both pre-training and SFT stages, validate the effectiveness of our summarized guidelines. They also demonstrate the robustness of our proposed data ordering methods in enhancing the stability and performance of LLM training. Github Link: https://github.com/microsoft/data-efficacy/", "date": "2026-05-28", "categories": []}
{"id": "2601.21309", "title": "Transferable Graph Condensation from the Causal Perspective", "abstract": "The increasing scale of graph datasets has significantly improved the performance of graph representation learning methods, but it has also introduced substantial training challenges. Graph dataset condensation techniques have emerged to compress large datasets into smaller yet information-rich datasets, while maintaining similar test performance. However, these methods strictly require downstream applications to match the original dataset and task, which often fails in cross-task and cross-domain scenarios. To address these challenges, we propose a novel causal-invariance-based and transferable graph dataset condensation method, named TGCC, providing effective and transferable condensed datasets. Specifically, to preserve domain-invariant knowledge, we first extract domain causal-invariant features from the spatial domain of the graph using causal interventions. Then, to fully capture the structural and feature information of the original graph, we perform enhanced condensation operations. Finally, through spectral-domain enhanced contrastive learning, we inject the causal-invariant features into the condensed graph, ensuring that the compressed graph retains the causal information of the original graph. Experimental results on five public datasets and our novel FinReport dataset demonstrate that TGCC achieves up to a 13.41% improvement in cross-task and cross-domain complex scenarios compared to existing methods, and achieves state-of-the-art performance on 5 out of 6 datasets in the single dataset and task scenario.", "date": "2026-01-29", "categories": []}
{"id": "2605.27495", "title": "Representation-Conditioned Diffusion Models for Guided Training Data Generation", "abstract": "Data availability remains a critical bottleneck in many deep learning applications. Large-scale datasets are often expensive to collect, curate and annotate, which can limit the scalability and applicability of supervised learning methods. In this work, we evaluate the classification performance of models trained on synthetic image datasets produced by generative deep learning. In particular, we use latent diffusion models conditioned on learned representations from DINOv2, DINOv3, and CLIP. Our results demonstrates that this representation-conditioned formulation significantly outperforms class-conditioned generation by a large margin (+10.76 p.p. top-1 accuracy on ImageNet100), by improving sample quality and mode coverage. Furthermore, by scaling the size of the synthetic dataset, we are able to outperform a classifier trained on the real data (+2.0 p.p top-1 accuracy).\n  We also demonstrate how generated images can be used for augmentation purposes, outperforming classical augmentation methods, and how the conditioning space can be used for sample filtering to further improve training value. Collectively, these findings highlight that representation-conditioned diffusion models provide a promising approach for augmenting, complementing, or potentially replacing real-world datasets in large-scale visual learning tasks.", "date": "2026-05-26", "categories": []}
{"id": "2605.28247", "title": "IRDS: Interpretable RLVR Data Selection via Verifier-Coupled Sparse Autoencoder Coverage", "abstract": "Reinforcement learning with verifiable rewards (RLVR) has become a key technique for en- hancing LLM reasoning, yet its data ineffi- ciency remains a major bottleneck. Existing methods address this problem only partially, each missing at least one of subset-level cov- erage, verifier signal use, or interpretability. To address this gap, we present IRDS (Inter- pretable RLVR Data Selection), which selects RLVR training instances on a sparse autoen- coder (SAE) cluster basis so the selection itself is auditable on recognizable problem motifs. To select instances the model both fails on and can still learn from, we introduce a verifier- coupled coverage objective on the SAE basis and solve it by greedy log-determinant max- imization. Experiments on three instruction- tuned models and six math reasoning bench- marks show that IRDS achieves the highest overall accuracy, exceeding the strongest base- line by +3.9/+4.0 pp on the two Qwen models and by +0.5 pp on Llama-3.1-8B, while run- ning an order of magnitude cheaper than the trajectory-based baseline.", "date": "2026-05-27", "categories": []}
{"id": "2605.28631", "title": "Single-Rollout Hidden-State Dynamics for Training-Free RLVR Data Selection", "abstract": "Reinforcement learning with verifiable rewards (RLVR) can yield large reasoning gains from very few training instances, yet its strong sensitivity to which instances are used makes data selection a central bottleneck. Most existing selection pipelines rely on training-time optimization signals and/or require access to verifiable rewards or ground-truth answers over large candidate pools, which is costly and often infeasible in specialized domains. We study RLVR data selection in a setting where selection must be performed before any RL training and without labels or reward evaluation on the full pool. We propose SHIFT, a one-shot, training-free selector based solely on inference-time hidden-state dynamics. For each candidate instance, SHIFT runs a single deterministic reasoning rollout and computes a reasoning-induced representation shift (RIRS) as the start-to-end hidden-state delta. SHIFT uses the RIRS magnitude as a lightweight proxy for instance utility and enforces coverage via a quality-weighted farthest-first CoreSet procedure in an RIRS-augmented feature space, producing compact subsets that scale to large unlabeled pools. Across mathematical reasoning and medical QA benchmarks under ultra-low budgets, SHIFT consistently outperforms training-free diversity and difficulty/uncertainty baselines, improving both in-domain accuracy and transfer to harder evaluation settings. Ablations show that RIRS-based coverage and quality-weighting contribute complementary gains, and analyses indicate that RIRS is not explained by simple input/output length statistics. Code is available at github.com/JianghaoWu/SHIFT.", "date": "2026-05-27", "categories": []}
{"id": "2501.06708", "title": "Evaluating Sample Utility for Efficient Data Selection by Mimicking Model Weights", "abstract": "Large-scale web-crawled datasets contain noise, bias, and irrelevant information, necessitating data selection techniques. Existing methods depend on hand-crafted heuristics, downstream datasets, or require expensive influence-based computations -- all of which limit scalability and introduce unwanted data dependencies. To address this, we introduce the Mimic Score, a simple and geometry-based data-quality metric that evaluates utility by measuring alignment between a sample's gradients and a target direction induced by a pre-trained reference model. This leverages readily available model weights, avoids needing validation datasets, and incurs minimal computational overheads. Building on this metric, we propose Grad-Mimic, a two-stage framework that re-weights samples online to accelerate training and aggregates sample utilities offline to construct effective data filters. Empirically, we show that using mimic scores to guide training improves data efficiency, accelerates convergence, yields consistent performance gains across six image datasets, and enhances CLIP models with 20.7% fewer training steps. Additionally, mimic score-based filters augment existing filtering techniques, enabling improved CLIP models trained with 4.7 million fewer samples.", "date": "2025-01-12", "categories": []}
{"id": "2605.20606", "title": "Mind Your Margin and Boundary: Are Your Distilled Datasets Truly Robust?", "abstract": "Dataset distillation (DD) compresses a large training set into a small synthetic set for efficient training, but most DD methods optimize only clean accuracy and leave robustness uncontrolled. Recent robust DD methods improve robustness, yet they often suffer from a poor accuracy-robustness trade-off because they (i) treat all adversarially perturbed examples uniformly, despite robust risk being dominated by near-zero robust margins, and (ii) do not explicitly increase inter-class separation in the decision boundary where attacks concentrate. We present Contrastive Curriculum for Robust Dataset Distillation (C$^2$R), a framework that couples an attack-aware curriculum with a contrastive robustness objective. From a robust-margin perspective, we derive a perturbation score that approximates each sample's robust hinge, enabling a curriculum that prioritizes the smallest-margin adversaries that most directly drive robust error. In parallel, a class-balanced contrastive robustness loss enforces adversarial invariance while explicitly widening boundary separation across classes. Experiments on CIFAR-10/100, Tiny-ImageNet, and multiple ImageNet-1K subsets under six attacks show that C$^2$R achieves the best robust accuracy, outperforming prior robust DD by $2.8$% on average.", "date": "2026-05-20", "categories": []}
{"id": "2605.26971", "title": "RLVR Datasets and Where to Find Them: Tracing Data Lineage for Better Training Data", "abstract": "The proliferation of Reinforcement Learning from Verifiable Rewards (RLVR) datasets has exacerbated provenance collapse due to unclear lineage among existing datasets. To bridge this fragmented RLVR data landscape, we propose Atomic-source Tracing via Lineage-Aware Search (ATLAS), a systematic framework for tracing RLVR datasets back to their atomic sources, attributing over 99.7% of 1.45M instances to 20 atomic sources. Our analysis reveals that most RLVR datasets are variants of a small set of shared upstream sources, with few introducing genuinely new data, and many facing data contamination risks. These findings naturally motivate us to curate a new RLVR dataset, DAPO++, and to benchmark existing datasets from a lineage-aware perspective. To this end, we propose Source-level Counterfactual Attribution (SCA) as a guiding principle to curate a decontaminated training dataset with concentrated learning signals. Essentially, SCA measures a sample's marginal utility by comparing per-atomic-source RL checkpoints against a shared base model. Building upon these attribution signals, we further design a composite dataset quality score Q that strongly correlates with downstream RLVR performance. Experiments on Qwen3 series models verify that DAPO++ consistently improves performance on held-out benchmarks, while Q reliably predicts downstream RLVR training effectiveness. Our code and data is available at https://github.com/Celine-hxy/ATLAS.", "date": "2026-05-26", "categories": []}
{"id": "2505.22322", "title": "A Closer Look on Memorization in Tabular Diffusion Model: A Data-Centric Perspective", "abstract": "Diffusion models have shown strong performance in generating high-quality tabular data, but they carry privacy risks by reproducing exact training samples. While prior work focuses on dataset-level augmentation to reduce memorization, little is known about which individual samples contribute most. We present the first data-centric study of memorization dynamics in tabular diffusion models. We quantify memorization for each real sample based on how many generated samples are flagged as replicas, using a relative distance ratio. Our empirical analysis reveals a heavy-tailed distribution of memorization counts: a small subset of samples contributes disproportionately to leakage, confirmed via sample-removal experiments. To understand this, we divide real samples into top- and non-top-memorized groups and analyze their training-time behaviors. We track when each sample is first memorized and monitor per-epoch memorization intensity (AUC). Memorized samples are memorized slightly earlier and show stronger signals in early training. Based on these insights, we propose DynamicCut, a two-stage, model-agnostic mitigation method: (a) rank samples by epoch-wise intensity, (b) prune a tunable top fraction, and (c) retrain on the filtered dataset. Across multiple tabular datasets and models, DynamicCut reduces memorization with minimal impact on data diversity and downstream performance. It also complements augmentation-based defenses. Furthermore, DynamicCut enables cross-model transferability: high-ranked samples identified from one model (e.g., a diffusion model) are also effective for reducing memorization when removed from others, such as GANs and VAEs.", "date": "2025-05-28", "categories": []}
{"id": "2605.12649", "title": "DIVER:Diving Deeper into Distilled Data via Expressive Semantic Recovery", "abstract": "Dataset distillation aims to synthesize a compact proxy dataset that is unreadable or non-raw from the original dataset for privacy protection and highly efficient learning. However, previous approaches typically adopt a single-stage distillation paradigm, which suffers from learning specific patterns that overfit on a prior architecture, consequently suppressing the expression of semantics and leading to performance degradation across heterogeneous architectures. To address this issue, we propose a novel dual-stage distillation framework called ${\\textbf{DIVER}}$, which leverages the pre-trained diffusion model to dive deeper into $\\textbf{DI}$stilled data $\\textbf{V}$ia $\\textbf{E}$xpressive semantic $\\textbf{R}$ecovery, an entire process of semantic inheritance, guidance, and fusion. Semantic inheritance distills high-level semantics of abstract distilled images into the latent space to filter out architecture-specific ``noise\" and retain the intrinsic semantics. Furthermore, semantic guidance improves the preservation of the original semantics by directing the reverse procedure. Finally, semantic fusion is designed to provide semantic guidance only during the concrete phase of the reverse process, preventing semantic ambiguity and artifacts while maintaining the guidance information. Extensive experiments validate the effectiveness and efficiency of DIVER in improving classical distillation techniques and significantly improving cross-architecture generalization, requiring processing time comparable to raw DiT on ImageNet (256$\\times$256) with only 4 GB of GPU memory usage.", "date": "2026-05-12", "categories": []}
{"id": "2605.25022", "title": "D3S2: Diffusion-Guided Dataset Distillation for Semantic Segmentation", "abstract": "Dataset distillation (DD) aims to compress large-scale datasets into compact synthetic sets while preserving training efficacy. However, existing studies mainly focus on image classification, leaving dense prediction tasks such as semantic segmentation largely underexplored. In this work, we identify three key challenges for segmentation DD: (i) long-tailed class imbalance, (ii) the need for strict pixel-wise alignment between images and dense labels, and (iii) the high computational cost of optimizing high-resolution data with complex models. To address these challenges, we propose D3S2, a Diffusion-guided Dataset Distillation framework for Semantic Segmentation. Our method adopts a two-stage design. In Class-Balanced Mask Selection, we construct a representative mask set via a greedy strategy that prioritizes underrepresented classes. In Diffusion-Guided Image Synthesis, we employ a pretrained layout-to-image diffusion model to generate images conditioned on the selected masks, naturally ensuring spatial alignment. To further enhance the training utility of synthesized data, we introduce guided diffusion sampling with two complementary objectives: a segmentation-consistency loss for pixel-level alignment, and a class-wise feature matching loss for aligning per-class feature statistics across layers. Extensive experiments demonstrate the superiority of D3S2. Notably, at an extremely compression rate of 1%, our method achieves 24.99% and 35.49% mIoU on ADE20K and COCO-Stuff with Mask2Former (Swin-S), outperforming random selection by 9.34% and 5.70%, respectively.", "date": "2026-05-24", "categories": []}
{"id": "2605.25162", "title": "STREAM: A Data-Centric Framework for Mining High-Value Task-Oriented Dialogues from Streaming Media", "abstract": "Large language models for vertical domains are bottlenecked by the scarcity of complex, domain-specific task-oriented dialogues. Existing data acquisition pipelines face a persistent trilemma: expert annotation is expensive, real-world service conversations are constrained by privacy and commercial restrictions, and static corpora quickly become temporally stale. We propose Stream, a data-centric framework that leverages publicly available streaming media (live streams and short videos) to synthesize high-value service dialogues at scale. Stream mines authentic interaction signals from noisy streams and synthesizes conversations by integrating role-grounded persona construction with Conversational Blueprint construction; it further adopts retrieval-augmented generation (RAG) to support knowledge-aware responses. Based on Stream, we release StreamDial, a large-scale multi-domain dataset covering Automotive, Restaurant, and Hotel. StreamDial contains 87,498 dialogue sessions and 1,497,320 turns in total, with an average of 17.11 turns per session and a comparable scale across domains. Each session is organized as a structured quadruplet $\\langle P_u, P_a, B, H \\rangle$ that pairs dialogue history with explicit user/agent personas and a Conversational Blueprint, capturing realistic service behaviors such as requirement mining, constraint conflicts, negotiation, and recovery. Evaluations with automatic judges and downstream tasks show that StreamDial improves intrinsic dialogue quality over strong baselines, and models trained with StreamDial improve Dialogue State Tracking across backbones; we further report a completed human-evaluation set and encouraging multilingual transfer on Qwen3-8B under a controlled training budget. The data is released in https://github.com/hitxueliang/DialogDataSetBySTREAM.", "date": "2026-05-24", "categories": []}
{"id": "2605.23482", "title": "Multimodal Distribution Matching for Vision-Language Dataset Distillation", "abstract": "Dataset distillation compresses large training sets into compact synthetic datasets while preserving downstream performance. As modern systems increasingly operate on paired vision-language inputs, multimodal distillation must preserve representation quality and cross-modal alignment under tight compute and memory budgets, yet prior methods often require heavy computes and overlook their correlations. To address this, we present Multimodal Distribution Matching (MDM), a geometry-aware framework for efficient and generalizable multimodal distillation. Specifically, MDM integrates complementary components at the data, model, and loss levels. At the data level, it initializes synthetic image-text pairs by sampling from clusters in the joint embedding space. At the model level, it forms a mixed teacher by interpolating independently fine-tuned models in weight space according to their angular deviation from the pretrained anchor. At the loss level, it matches joint distributions on the unit hypersphere using a geometry-aware matching objective that exploits the joint features in the cross-modal agreement and discrepancy directions along with symmetric contrastive learning. Across image-text retrieval benchmarks with cross-architecture evaluation, MDM yields compact synthetic sets that preserve multimodal semantics, substantially reduce distillation cost, and remain robust across architectures.", "date": "2026-05-22", "categories": []}
{"id": "2605.00185", "title": "Fair Dataset Distillation via Cross-Group Barycenter Alignment", "abstract": "Dataset Distillation aims to compress a large dataset into a small synthetic one while maintaining predictive performance. We show that as different demographic groups exhibit distinct predictive patterns, the distillation process struggles to simultaneously preserve informative signals for all subgroups, regardless of whether group sizes are mildly or severely imbalanced. Consequently, models trained on distilled data can experience substantial performance drops for certain subgroups, leading to fairness gaps. Crucially, these gaps do not disappear by merely correcting group imbalance, since they stem from fundamental mismatches in subgroup predictive patterns rather than from sample-size disparities alone. We therefore formally analyze the interaction between these two sources of bias and cast the solution as identifying a group-imbalance-agnostic barycenter of the predictive information that induces similar representations across all subgroups. By distilling toward this shared aggregate representation, we show that group fairness concerns can be reduced. Our approach is compatible with existing distillation methods, and empirical results show that it substantially reduces bias introduced by dataset distillation. Code is available at https://github.com/mhmoslemi/COBRA.", "date": "2026-04-30", "categories": []}
{"id": "2605.18893", "title": "Position: Graph Condensation Needs a Reset -- Move Beyond Full-dataset Training and Model-Dependence", "abstract": "Graph Neural Networks (GNNs) are powerful tools for learning from graph-structured data, but their scalability is increasingly strained by the size of real-world graphs in domains like recommender systems, fraud detection, and molecular biology. Graph condensation -- the task of generating a smaller synthetic graph that retains the performance of models trained on the original -- has emerged as a promising solution. However, the dominant approach of gradient matching introduces a fundamental contradiction: it requires training on the full dataset to create the compressed version, thereby undermining the goal of efficiency. Worse still, these methods suffer from high computational overhead, poor generalization across GNN architectures, and brittle reliance on specific model configurations. Equally concerning is the community's reliance on misleading evaluation protocols such as node compression ratios, which fail to reflect true resource savings, condensation overhead, and illusory application to neural architecture search. These shortcomings are not incidental -- they are systemic, and they obstruct meaningful progress.\n  In this position paper, we argue that graph condensation, in its current form, needs a reset. We call for moving beyond full-dataset training and model-dependent design, and instead advocate for methods that are lightweight, architecture-agnostic, and practically deployable. By identifying key methodological flaws and outlining concrete research directions, we aim to reorient the field toward approaches that deliver on the true promise of condensation: efficient, generalizable, and usable GNN training at scale.", "date": "2026-05-17", "categories": []}
{"id": "2605.15691", "title": "SEED: Targeted Data Selection by Weighted Independent Set", "abstract": "Data selection seeks to identify a compact yet informative subset from large-scale training corpora, balancing sample quality against collection diversity. We formulate this problem as a Weighted Independent Set (WIS) on a similarity graph, where nodes represent data samples weighted by influence, and edges connect semantically redundant pairs. This formulation naturally yields subsets that are simultaneously high-quality and diverse. However, two challenges arise in practice: naive node weights fail to distinguish informative signals from gradient noise, and edge construction under heterogeneous domain distributions produces structurally imbalanced graphs that bias selection toward sparse regions. To address these issues, we introduce two principled refinements from a unified graph perspective: (1) \\textit{node value calibration} that restricts influence estimation to the bilateral salient subspace to ground node importance in task-relevant signals rather than surface-level statistics; (2) \\textit{local scale normalization} that adapts edge thresholds to local neighborhood density, mitigating graph imbalance induced by cross-domain distribution shifts. Together, these components yield a robust and scalable data selection pipeline dubbed SEED. We further construct \\texttt{Honeybee-Remake-SEED-200K}, a compact multimodal dataset curated by SEED. Extensive experiments show that SEED consistently outperforms state-of-the-art methods on instruction tuning, visual instruction tuning, and semantic segmentation across diverse model families.", "date": "2026-05-15", "categories": []}
{"id": "2605.20620", "title": "Dynamic Shapley Computation", "abstract": "Shapley-based data valuation provides a principled way to quantify the contribution of training data, but its high computational cost makes it impractical in dynamic settings where tasks and training players evolve. Existing methods treat Shapley computation as a one-shot process and collapse contributions into aggregated scores, preventing reuse and requiring recomputation under any change. We introduce a new perspective that represents Shapley values as a player-by-task matrix and formulates dynamic valuation as a structured matrix maintenance problem. We exploit the fact that each task depends on a small subset of training players and that similar tasks yield similar valuations, leading to utility locality and coalition locality. Based on these insights, we propose D-Shap, a dynamic valuation framework that enables efficient updates by modifying only a small portion of the matrix: new task valuations are inferred via structure-aware interpolation, while updates induced by new players are confined to affected local matrix blocks. To eliminate the need for pre-specified evaluation tasks, we introduce self-valuation, which constructs the initial matrix directly from training data, supported by scalable subset reuse and coverage-aware anchor selection. Experiments across diverse models show that D-Shap performs task updates in milliseconds and reduces the cost of player updates by up to three orders of magnitude, while achieving valuation quality competitive with full recomputation.", "date": "2026-05-20", "categories": []}
{"id": "2605.21033", "title": "Efficient Banzhaf-Based Data Valuation for $k$-Nearest Neighbors Classification", "abstract": "Data valuation, the task of quantifying the contribution of individual data points to model performance, has emerged as a fundamental challenge in machine learning. Game-theoretic approaches, such as the Banzhaf value, offer principled frameworks for fair data valuation; however, they suffer from exponential computational complexity. We address this challenge by developing efficient algorithms specifically tailored for computing Banzhaf values in $k$-nearest neighbor ($k$NN) classifiers. We first establish the theoretical hardness of the problem by proving that it is \\#P-hard. Despite this intractability, we exploit the locality properties of $k$NN classifiers to develop practical exact algorithms. Our main contribution is a dynamic programming framework that achieves significant computational improvements: we present a pseudo-polynomial algorithm with $O(Wkn^2)$ time complexity for weighted $k$NN classifiers, where $W$ is the maximum sum of top-$k$ weights, and a specialized algorithm for unweighted $k$NN that achieves $O(nk^2)$ time complexity, that is, linear in the number of data points. We also offer efficient Monte Carlo estimation methods. Extensive experiments on real-world datasets demonstrate the practical efficiency of our approach and its effectiveness in data valuation applications.", "date": "2026-05-20", "categories": []}
{"id": "2605.21372", "title": "Closed Loop Dynamic Driving Data Mixture for Real-Synthetic Co-Training", "abstract": "Data scaling is fundamental to modern deep learning, and grows increasingly critical as autonomous driving shifts to end-to-end learning. Real-world driving data is expensive to annotate and scene-biased, making real-synthetic co-training with near-infinite synthetic data a promising direction. However, naively incorporating all available synthetic data is inefficient and leads to distribution shifts, and optimizing data mixture under practical training budgets remains a critical yet under-explored problem. In this sense, we claim that the mixture of training data requires clear guidance in terms of scene types and quantities. Particularly in this work, we conceptualize the data mixture approximately as a dynamic optimization process that iteratively adjusts the training data mixture to maximize model performance, guided by closed-loop evaluation feedback, and propose AutoScale, a fully automated closed-loop data engine unifying scene representation, data mixture optimization and retrieval, as well as model training and evaluation. Specifically, we propose Graph Regularized AutoEncoder (Graph-RAE) for driving scene representations, introduce Cluster-aware Gradient Ascent (Cluster-GA) for cluster-wise importance estimation and reweighting, and perform cluster-guided vector retrieval to select high-value samples. Experiments on NavSim demonstrate that AutoScale outperforms vanilla co-training and cross-domain baselines, achieving better performance with fewer synthetic samples under constrained budgets.", "date": "2026-05-20", "categories": []}
{"id": "2605.18836", "title": "Spectral Gradient Surgery for Domain-Generalizable Dataset Distillation", "abstract": "Dataset Distillation (DD) synthesizes a compact synthetic dataset that preserves the training utility of a full dataset. However, its standard formulation assumes that test data follow the same distribution as training data, an assumption that rarely holds in practice. A straightforward extension-applying post-hoc Domain Generalization (DG) techniques to distilled data-is ill-suited because existing DG methods rely on the natural diversity of real datasets, which compact synthetic sets inherently lack, while also incurring substantial augmentation overhead that conflicts with the efficiency objective of dataset distillation. To address this limitation, we introduce Domain Generalizable Dataset Distillation (DGDD), a new problem setting that explicitly targets out-of-distribution (OOD) generalization of distilled datasets. We study this problem through a widely adopted DD baseline of Distribution Matching (DM). We attribute the OOD vulnerability of DM to the entanglement of class-discriminative and domain-specific information within the compressed synthetic set, and propose Spectral Gradient Surgery (SGS) to disentangle the two. The key insight of SGS is that cross-domain agreement among domain-wise gradients in the spectral domain reveals which gradient components are shared across source domains-and are therefore class-discriminative-and which are domain-specific. Based on this observation, SGS augments the standard DM update with two complementary gradients: one that reinforces cross-domain shared components and another that explicitly promotes diversity within the distilled dataset. Extensive experiments on diverse-scale benchmarks demonstrate that SGS substantially improves OOD generalization while remaining plug-and-play compatible with existing DM methods.", "date": "2026-05-13", "categories": []}
{"id": "2605.19407", "title": "A Bitter Lesson for Data Filtering", "abstract": "We investigate data filtering for large model pretraining via new scaling studies that target the high compute, data-scarce regime. In spite of an apparently common belief that filtering data to include only high-quality information is essential, our experiments suggest that with enough compute, the best data filter is no data filter. We find that sufficiently trained large parameter models not only tolerate low-quality and distractor data, but in fact benefit from nominally ``poor'' data.", "date": "2026-05-19", "categories": []}
{"id": "2602.18584", "title": "GIST: Targeted Data Selection for Instruction Tuning via Coupled Optimization Geometry", "abstract": "Targeted data selection has emerged as a crucial paradigm for efficient instruction tuning, aiming to identify a small yet influential subset of training examples for a specific target task. In practice, influence is often measured through the effect of an example on parameter updates. To make selection scalable, many approaches leverage optimizer statistics (e.g., Adam states) as an axis-aligned surrogate for update geometry (i.e., diagonal precondition), implicitly treating parameters as coordinate-wise independent. We show that this assumption breaks down in parameter-efficient fine-tuning (PEFT) methods such as LoRA. In this setting, the induced optimization geometry exhibits strong cross-parameter coupling with non-trivial off-diagonal interactions, while the task-relevant update directions are confined to a low-dimensional subspace. Motivated by this mismatch, we propose GIST (Gradient Isometric Subspace Transformation), a simple yet principled alternative that replaces axis-aligned scaling with robust subspace alignment. GIST recovers a task-specific subspace from validation gradients via singular value decomposition (SVD), projects training gradients into this coupled subspace, and scores examples by their alignment with target directions. Extensive experiments have demonstrated that GIST matches or outperforms the state-of-the-art baseline with only 0.29% of the storage and 25% of the computational time under the same selection budget.", "date": "2026-02-20", "categories": []}
{"id": "2605.16704", "title": "Convex Dataset Valuation for Post-Training", "abstract": "Improving LLM performance on downstream tasks sometimes requires leveraging auxiliary datasets during post-training. In practice, however, developers face constraints on compute, labeling, and licensing costs that preclude using all available data, necessitating principled dataset-level selection. These constraints are increasingly shaped by dataset marketplaces, where data acquisition is governed by budgets and negotiation. We study dataset valuation as a subset selection problem during LLM post-training. Our goal is to identify and weight auxiliary datasets so as to maximize target task performance given constrained budgets. We first show that commonly used gradient alignment scores provide a reasonable yet incomplete valuation signal, as they ignore redundancy among datasets. To address this, we propose a scalable convex dataset-level valuation method based on kernel mean matching (KMM) in gradient space, which jointly accounts for alignment with the target task and redundancy across auxiliary datasets. Through extensive experiments across diverse post-training settings and tasks, we show that our approach consistently outperforms existing valuation baselines, achieving stronger performance with low computational overhead. Our results position dataset valuation as a practical decision tool for post-training data selection in market-constrained large language model settings. The code is available at https://github.com/uiuctml/convex_data_valuation.", "date": "2026-05-15", "categories": []}
{"id": "2605.18012", "title": "SAS: Semantic-aware Sampling for Generative Dataset Distillation", "abstract": "Deep neural networks have achieved impressive performance across a wide range of tasks, but this success often comes with substantial computational and storage costs due to large-scale training data. Dataset distillation addresses this challenge by constructing compact yet informative datasets that enable efficient model training while maintaining downstream performance. However, most existing approaches primarily emphasize matching data distributions or downstream training statistics, with limited attention to preserving high-level semantic information in the distilled data. In this work, we introduce a semantic-aware perspective for dataset distillation by leveraging Contrastive Language-Image Pretraining (CLIP) as a semantic prior for post-sampling. Our goal is to obtain distilled datasets that are not only compact but also semantically class-discriminative and diverse. To this end, we design three semantic scoring functions that quantify class relevance, inter-class separability, and intra-set diversity in a pretrained semantic space. Based on image pools generated by existing distillation methods, we further develop a two-stage strategy for effective sampling: the first stage filters semantically discriminative samples to form a reliable candidate set, and the second stage performs a dynamic diversity-aware selection to reduce redundancy while preserving semantic coverage. Extensive experiments across multiple datasets, image pools, and downstream models demonstrate consistent performance gains, highlighting the effectiveness of incorporating semantic information into dataset distillation.", "date": "2026-05-18", "categories": []}
{"id": "2605.14773", "title": "Beyond What to Select: A Plug-and-play Oscillatory Data-Volume Scheduling for Efficient Model Training", "abstract": "Data selection accelerates training by identifying representative training data while preserving model performance. However, existing methods mainly focus on designing sample-importance criteria, i.e., deciding what to select, while typically fixing the selected data volume as the target ratio throughout training. Thus, they are often dynamic in sample identity but static in data volume. In this work, we revisit data selection from an optimization perspective and show that selected-data training induces an implicit regularization effect modulated by the instantaneous selection ratio. This reveals a key trade-off: lower ratios amplify selection-induced regularization, whereas higher ratios preserve data coverage and optimization fidelity. Motivated by this insight, we propose PODS, a Plug-and-play Oscillatory Data-volume Scheduling framework. Rather than introducing another sample-scoring metric, PODS serves as a lightweight module that dynamically schedules how much data to select over training. Under the target selection ratio, PODS alternates between low-ratio regularization phases and high-ratio recovery phases to exploit selection-induced regularization without sacrificing optimization stability. With its lightweight, ratio-level, and task-agnostic design, PODS is compatible with existing static and dynamic selection methods and broadly applicable across training paradigms. Experiments across various datasets, architectures, and tasks show that PODS consistently improves the efficiency-generalization trade-off, e.g., reducing ImageNet-1k training cost by 50% with improved accuracy and accelerating LLM instruction tuning by over 2x without performance degradation.", "date": "2026-05-14", "categories": []}
{"id": "2509.13858", "title": "EDITS: Enhancing Dataset Distillation with Implicit Textual Semantics", "abstract": "Dataset distillation aims to synthesize a compact dataset from the original large-scale one, enabling highly efficient learning while preserving competitive model performance. However, traditional techniques primarily capture low-level visual features, neglecting the high-level semantic and structural information inherent in images. In this paper, we propose EDITS, a novel framework that exploits the implicit textual semantics within the image data to achieve enhanced distillation. First, external texts generated by a Vision Language Model (VLM) are fused with image features through a Global Semantic Query module, forming the prior clustered buffer. Local Semantic Awareness then selects representative samples from the buffer to construct image and text prototypes, with the latter produced by guiding a Large Language Model (LLM) with meticulously crafted prompt. Ultimately, Dual Prototype Guidance strategy generates the final synthetic dataset through a diffusion model. Extensive experiments confirm the effectiveness of our method.Source code is available in: https://github.com/einsteinxia/EDITS.", "date": "2025-09-17", "categories": []}
{"id": "2604.00001", "title": "Filter-then-Weight: Online Data Selection and Reweighting for LLM Fine-Tuning", "abstract": "Gradient-based data selection offers a principled framework for estimating sample utility in large language model (LLM) fine-tuning, but existing methods are mostly designed for offline settings. They are therefore less suited to online fine-tuning, where data arrives sequentially, sample utility is step-dependent, and the effective update geometry is shaped by adaptive optimizers. We propose an optimizer-aware framework for gradient-based online data selection and reweighting in LLM fine-tuning. Our key idea is to view online selection not as static sample ranking, but as shaping the next target-oriented update under the current optimizer state. We formulate this as an optimizer-aware update-matching problem, establish its connection to second-order target utility, and show why subset-level construction must account for interactions and redundancy among selected samples. Based on this view, we develop a two-stage Filter-then-Weight algorithm that first filters geometrically useful candidates and then optimizes their coefficients. To make the framework practical for LLMs, we introduce a factorized outer-product gradient representation and optimized matrix computations for long-context data. Experiments show that our method consistently improves convergence and downstream performance over existing online data selection baselines under the same data budget.", "date": "2026-03-08", "categories": []}
{"id": "2605.10906", "title": "DataMaster: Data-Centric Autonomous AI Research", "abstract": "As model families, training recipes, and compute budgets become increasingly standardized, further gains in machine learning systems depend increasingly on data. Yet data engineering remains largely manual and ad hoc: practitioners repeatedly search for external datasets, adapt them to existing pipelines, validate candidate data through downstream training, and carry forward lessons from prior attempts. We study task-conditioned autonomous data engineering, where an autonomous agent improves a fixed learning algorithm by optimizing only the data side, including external data discovery, data selection and composition, cleaning and transformation. The goal is to obtain a stronger downstream solution while leaving the learning algorithm unchanged. To address the open-ended search space, branch-dependent refinement, and delayed validation inherent in autonomous data engineering, we propose DataMaster, a data-agent framework that integrates tree-structured search, shared candidate data, and cumulative memory. DataMaster consists of three key components: a DataTree that organizes alternative data-engineering branches, a shared Data Pool that stores discovered external data sources for reuse, and a Global Memory that records node outcomes, artifacts, and reusable findings. Together, these components allow the agent to discover candidate data, construct executable training inputs, evaluate them through downstream feedback, and carry useful evidence across branches. We evaluate DataMaster on two types of benchmarks, MLE-Bench Lite and PostTrainBench. On MLE-Bench Lite, it improves medal rate by 32.27% over the initial score; on PostTrainBench, it surpasses the instruct model on GPQA (31.02% vs 30.35%).", "date": "2026-05-11", "categories": []}
{"id": "2605.11231", "title": "LiBaGS: Lightweight Boundary Gap Synthesis for Targeted Synthetic Data Selection", "abstract": "Synthetic data is useful only when the added samples fill missing parts of the training distribution that matter for the downstream task. We introduce LiBaGS, a lightweight, generator-agnostic method for targeted synthetic training data selection. LiBaGS scores candidate synthetic samples by combining decision-boundary proximity, predictive uncertainty, real-data density, and support validity, so that selected samples are both informative and likely to remain on the real data manifold. We then use a boundary-gap allocation rule that targets sparse but realistic decision-boundary neighborhoods, rather than simply adding more data or selecting only the most uncertain candidates. LiBaGS also learns when enough synthetic samples have been added through a marginal-value stopping rule, assigns softer labels near ambiguous boundaries, and uses a diversity objective to avoid redundant near-duplicate selections. Experiments show that LiBaGS improves accuracy over classical oversampling, hard augmentation, uncertainty and density ablations, and targeted-generation selection criteria.", "date": "2026-05-11", "categories": []}
{"id": "2605.12944", "title": "From Instance Selection to Fixed-Pool Data Recipe Search for Supervised Fine-Tuning", "abstract": "Supervised fine-tuning (SFT) data selection is commonly formulated as instance ranking: score each example and retain a top-$k$ subset. However, effective SFT training subsets are often produced through ordered curation recipes, where filtering, mixing, and deduplication operators jointly shape the final data distribution. We formulate this problem as fixed-pool data recipe search: given a raw instruction pool and a library of grounded operators, the goal is to discover an executable recipe that constructs a high-quality selected subset under a limited budget of full SFT evaluations, without generating, rewriting, or augmenting training samples. We introduce AutoSelection, a two-layer solver that decouples fixed-pool materialization based on cached task-, data-, and model-side signals from expensive full evaluation, using warmup probes, realized subset states, local recipe edits, Gaussian-process-assisted ranking, and stagnation-triggered reseeding. Experiments on a 90K instruction pool show that AutoSelection achieves the strongest in-distribution reasoning average across three base models, outperforming full-data training, random recipe search, random top-$k$, and single-operator selectors. Additional Out-of-distribution graph-reasoning results, search-stability analyses, structural ablations, and 1.5B-to-7B transfer checks further show that recipe structure matters beyond individual selection operators. Code is available at https://github.com/w253/AutoSelection.", "date": "2026-05-13", "categories": []}
{"id": "2605.13130", "title": "GRACE: Gradient-aligned Reasoning Data Curation for Efficient Post-training", "abstract": "Existing reasoning data curation pipelines score whole samples, treating every intermediate step as equally valuable. In reality, steps within a trace contribute very unevenly, and selecting reasoning data well requires assessing them individually. We present GRACE, a gradient-aligned curation method that views each reasoning trace as a sequence of optimization events and scores every step by two complementary signals: its alignment with the answer-oriented gradient direction, and its consistency with the preceding reasoning trajectory. Step-level scores are aggregated into a sample-level value for subset selection, using only the model's internal optimization signals and no external reward models or step annotations. To make this scalable, GRACE introduces a representation-level gradient proxy that estimates step-level alignment from token-level upstream signals in a single forward pass. Post-training Qwen3-VL-2B-Instruct on MMathCoT-1M, GRACE reaches 108.8% of the full-data performance with 20% of the data and retains 100.2% with only 5%, with subsets that transfer effectively across model backbones.", "date": "2026-05-13", "categories": []}
{"id": "2605.13149", "title": "AcquisitionSynthesis: Targeted Data Generation using Acquisition Functions", "abstract": "Data quality remains a critical bottleneck in developing capable, competitive models. Researchers have explored many ways to generate top quality samples. Some works rely on rejection sampling: generating lots of synthetic samples and filtering out low-quality samples. Other works rely on larger or closed-source models to extract model weaknesses, necessary skills, or a curriculum off of which to base data generation. These works have one common limitation: there is no quantitative approach to measure the impact of the generated samples on the downstream learner. Active learning literature provides exactly this, in the form of acquisition functions. Acquisition functions measure the informativeness and/or influence of data, providing interpretable, model-centric signals. Inspired by this, we propose AcquisitionSynthesis: using acquisition functions as reward models to train language models to generate higher-quality synthetic data. We conduct experiments on classic verifiable tasks of math, medical question-answering, and coding. Our experimental results indicate that (1) student models trained with AcquisitionSynthesis data achieve good performance on in-distribution tasks (2-7% gain) and is more robust to catastrophic forgetting, and (2) AcquisitionSynthesis models can generate data for other models and for low-to-high resource training paradigms. By leveraging acquisition rewards, we seek to demonstrate a principled path toward model-aware self-improvement that surpasses static datasets.", "date": "2026-05-13", "categories": []}
{"id": "2605.10684", "title": "Is Data Shapley Not Better than Random in Data Selection? Ask NASH", "abstract": "Data selection studies the problem of identifying high-quality subsets of training data. While some existing works have considered selecting the subset of data with top-$m$ Data Shapley or other semivalues as they account for the interaction among every subset of data, other works argue that Data Shapley can sometimes perform ineffectively in practice and select subsets that are no better than random. This raises the questions: (I) Are there certain \"Shapley-informative\" settings where Data Shapley consistently works well? (II) Can we strategically utilize these settings to select high-quality subsets consistently and efficiently? In this paper, we propose a novel data selection framework, NASH (Non-linear Aggregation of SHapley-informative components), which (I) decomposes the target utility function (e.g., validation accuracy) into simpler, Shapley-informative component functions, and selects data by optimizing an objective that (II) aggregates these components non-linearly. We demonstrate that NASH substantially boosts the effectiveness of Shapley/semivalue-based data selection with minimal additional runtime cost.", "date": "2026-05-11", "categories": []}
{"id": "2605.11312", "title": "Constraint-Data-Value-Maximization: Utilizing Data Attribution for Effective Data Pruning in Low-Data Environments", "abstract": "Attributing model behavior to training data is an evolving research field. A common benchmark is data removal, which involves eliminating data instances with either low or high values, then assessing a model's performance trained on the modified dataset. Many existing studies leverage Shapley-based data values for this task. In this paper, we demonstrate that these data values are not optimally suited for pruning low-value data when only a limited amount of data remains. To address this limitation, we introduce the Constraint-Data-Value-Maximization (CDVM) approach, which effectively utilizes data attributions for pruning in low-data scenarios. By casting pruning as a constrained optimization that both maximizes total influence and penalizes excessive per-test contributions, CDVM delivers robust performance when only a small fraction of the data is retained. On the OpenDataVal benchmark, CDVM shows strong performance and competitive runtime.", "date": "2026-05-11", "categories": []}
{"id": "2605.11705", "title": "CAST: Collapse-Aware multi-Scale Topology Fusion for Multimodal Coreset Selection", "abstract": "The training of large multimodal models fundamentally relies on massive image-text datasets, which inevitably incur prohibitive computational overhead. Dataset selection offers a promising paradigm by identifying a highly informative coreset. However, existing approaches suffer from two critical limitations: (i) single-modality-dominated sampling methods, which ignore the fine-grained cross-modal information imbalance inherent in multimodal datasets and thus lead to semantic loss in the other modality; and (ii) coarse-grained sample-scoring-based sampling methods, where the selected coreset tends to be biased toward the scoring model, making it difficult to guarantee distributional equivalence between the coreset and the original dataset. Meanwhile, existing distribution matching and discrete sampling strategies often fail to jointly account for global semantic structure, local fine-grained details, and redundancy-aware coverage in dense regions. To this end, we propose CAST, a Collapse-Aware multi-Scale Topology fusion framework for multimodal coreset selection. We first construct image- and text-modality topologies, and derive a unified topology via local-collapse-aware refinement and cross-modal fusion. We then introduce a multi-scale distribution matching criterion in the diffusion wavelet domain, encouraging the coreset to approximate the original dataset at multiple scales. Finally, we introduce a local soft relational coverage mechanism that extends pure geometric coverage to relation-aware indirect coverage, penalizing redundant selections in dense clusters. Extensive experiments on Flickr30K and MS-COCO show that CAST outperforms existing dataset selection baselines, showcasing great superiority in cross-architecture generalization and energy efficiency over state-of-the-art multimodal synthesis methods.", "date": "2026-05-12", "categories": []}
{"id": "2602.05391", "title": "Efficient Dataset Distillation for Pre-Trained Self-Supervised Models via Statistical Flow Matching", "abstract": "Dataset distillation seeks to synthesize a highly compact dataset that achieves performance comparable to the original dataset on downstream tasks. For the classification task that use pre-trained self-supervised models as backbones, previous linear gradient matching optimizes synthetic images by encouraging them to mimic the gradient updates induced by real images on the linear classifier. However, this batch-level formulation requires loading thousands of real images and applying multiple rounds of differentiable augmentations to synthetic images at each distillation step, leading to substantial computational and memory overhead. In this paper, we introduce statistical flow matching , a stable and efficient supervised learning framework that optimizes synthetic images by aligning constant statistical flows from target class centers to non-target class centers in the original data. Our approach loads raw statistics only once and performs a single augmentation pass on the synthetic data, achieving performance comparable to or better than the state-of-the-art methods with 10x lower GPU memory usage and 4x shorter runtime. Furthermore, we propose a classifier inheritance strategy that reuses the classifier trained on the original dataset for inference, requiring only an extremely lightweight linear projector and marginal storage while achieving substantial performance gains.", "date": "2026-02-05", "categories": []}
{"id": "2605.08153", "title": "Temporal-Decay Shapley: A Time-Aware Data Valuation Framework for Time-Series Data", "abstract": "With the rapid development of machine learning applications on time-series data, accurately assessing the value of training samples has become essential for data selection, noise detection, and model optimization. However, traditional data valuation methods usually assume that samples are independent and identically distributed, and thus ignore the time-varying nature of sample value in time-series data. This paper proposes an improved temporal Shapley data valuation method that enables accurate sample valuation for time-series data through a temporal decay mechanism and a multi-scale fusion strategy. Specifically, we propose three progressively enhanced temporal Shapley methods. Temporal-Decay Shapley (TDS) incorporates temporal information into Shapley value computation through exponential decay weights; the improved TDS adopts power exponential decay to better adapt to nonlinear temporal drift; and Multi-Scale Temporal-Decay Shapley (MS-TDS) constructs a multi-scale fusion mechanism that balances the value of short-term hotspot samples and long-term foundational samples through parallel multi-scale valuation and sample-level adaptive fusion. Experimental results show that the proposed methods generally outperform traditional methods in noise detection and high-value data identification tasks, with more evident advantages under most strongly temporal settings, thereby effectively improving the accuracy and robustness of data valuation.", "date": "2026-05-04", "categories": []}
{"id": "2605.09404", "title": "Let the Target Select for Itself: Data Selection via Target-Aligned Paths", "abstract": "Targeted data selection aims to identify training samples from a large candidate pool that improve performance on a specific downstream task. Many recent methods estimate candidate utility by aggregating local attribution scores along a trajectory induced by the candidate pool. When the pool is heterogeneous, however, this reference trajectory may be misaligned with the dynamics of a target-aligned selected subset, creating what we call reference path bias. We propose an alternative reference path: a validation-induced flow obtained from a short, capacity-limited warmup on the available target validation proxy. Along this path, candidates are scored by a normalized endpoint loss drop, yielding a simple zero-order selection rule that requires no candidate gradients or Hessian approximations. Across controlled logistic, vision, and instruction-tuning experiments, this score is competitive with strong dynamic attribution baselines while substantially reducing warmup and storage cost. Moreover, since the reference trajectory is decoupled from any specific candidate pool, the same compact warmup can be reused across additional pools without recomputing the trajectory.", "date": "2026-05-10", "categories": []}
{"id": "2605.09665", "title": "Learning Multi-Indicator Weights for Data Selection: A Joint Task-Model Adaptation Framework with Efficient Proxies", "abstract": "Data selection is a key component of efficient instruction tuning for large language models, as recent work has shown that data quality often matters more than data quantity. Accordingly, prior studies have introduced various multi-dimensional heuristics to evaluate and filter instruction data. However, most existing methods rely on static task-agnostic and model-agnostic weighting schemes, which overlook the varying requirements of specific downstream tasks and the differing pre-existing capabilities of models. In this paper, we propose a framework for learning multi-indicator weights that jointly adapts data selection to both the downstream task and the specific model. Our method identifies optimal weight configurations without full-scale fine-tuning by utilizing in-context learning (ICL) signals on compact tiny-validation sets. These signals serve as efficient performance proxies that ensure high-fidelity evaluation at minimal computational cost. Experiments across multiple benchmarks and model families, including Mistral, Qwen, and Llama, show that the approach achieves performance comparable to or exceeding full-dataset tuning while using only 30\\% of the training samples on GSM8K. Furthermore, our analysis reveals a trade-off between semantic diversity and logical complexity in reasoning tasks, highlighting the necessity of joint task-model adaptation.", "date": "2026-05-10", "categories": []}
{"id": "2605.09699", "title": "A Real-Calibrated Synthetic-First Data Engine", "abstract": "Modern computer vision systems increasingly encounter performance limitations in data-scarce domains, where collecting large-scale, high-quality labeled data is costly or impractical. While controllable diffusion models enable scalable synthetic image generation, directly applying synthetic augmentation often leads to unstable performance gains due to dataset-level quality issues and insufficient feedback mechanisms.\n  In this work, we present a Real-Calibrated Synthetic-First Data Engine, a modular data engineering framework that combines controllable diffusion generation and multi-stage curation/filtering within a unified pipeline, with optional support for uncertainty-driven selection and human verification. Instead of introducing new generative algorithms, our approach focuses on systematic dataset construction for improving the practical reliability of synthetic augmentation in low-data regimes.\n  The framework is implemented as a modular CLI-based pipeline, where generation, filtering, selection, and validation components can be independently configured and replaced. This design emphasizes reproducibility, flexibility, and practical deployment in real-world data workflows.\n  Through empirical evaluation centered on human pose estimation, we show that synthetic data improves a real-data baseline when used as near-zero-human-annotation-cost augmentation alongside real anchors, while synthetic-only training remains substantially below real-only performance. Supplementary segmentation diagnostics show the same domain-gap pattern. These results highlight the practical value of data-centric orchestration for low-data augmentation.", "date": "2026-05-10", "categories": []}
{"id": "2605.10315", "title": "Active Tabular Augmentation via Policy-Guided Diffusion Inpainting", "abstract": "Generative tabular augmentation is appealing in data-scarce domains, yet the prevailing focus on distributional fidelity does not reliably translate into better downstream models. We formalize a fidelity-utility gap: common generative objectives prioritize distributional plausibility, whereas augmentation succeeds only when injected samples reduce the current learner's held-out evaluation loss. This gap motivates learning not just how to generate, but what to generate and when to inject as training evolves. We propose TAP (Tabular Augmentation Policy), which couples diffusion inpainting with a lightweight, learner-conditioned policy to steer generation toward high-utility regions and controls safe injection via explicit gating and conservative windowed commitment. Under severe data scarcity, TAP consistently outperforms strong generative baselines on seven real-world datasets, improving classification accuracy by up to 15.6 percentage points and reducing regression RMSE by up to 32%.", "date": "2026-05-11", "categories": []}
{"id": "2605.07194", "title": "Closed-Form Linear-Probe Dataset Distillation for Pre-trained Vision Models", "abstract": "Dataset distillation compresses a large training set into a small synthetic set that preserves downstream training utility. While most existing methods target training networks from scratch, modern visual transfer learning often uses frozen pre-trained encoders followed by lightweight linear probing. Existing distillation methods for this setting either unroll iterative linear-probe updates with trajectory-based gradient matching, or rely on closed-form formulations originally designed for from-scratch training with neural-tangent-kernel (NTK) approximations. Neither route exploits the fact that frozen-feature linear probing admits a closed-form solution determined directly by the pre-trained features themselves, with no infinite-width approximation and no inner-loop trajectory. We propose Closed-Form Linear-Probe Dataset Distillation (CLP-DD), a bilevel formulation that computes the linear probe induced by the synthetic set with a sample-space kernel ridge solver. The synthetic images are then updated by evaluating this induced classifier on real features through a temperature-scaled softmax cross-entropy, where the classifier columns act as learned class anchors in feature space. We further show that the choice of outer objective is decisive: pairing the closed-form inner solver with a standard MSE outer loss substantially underperforms trajectory-based methods, while the discriminative outer loss closes most of the gap. On ImageNet-100 with four pre-trained backbones, CLP-DD substantially improves over LGM without DSA and approaches LGM with DSA at a fraction of the computational cost. On ImageNet-1K, CLP-DD matches or surpasses LGM with DSA on three of four backbones while running roughly $14\\times$ faster and using less than one-eighth of the GPU memory.", "date": "2026-05-08", "categories": []}
{"id": "2605.07663", "title": "Quotient Semivalues for False-Name-Resistant Data Attribution", "abstract": "Data valuation methods allocate payments and audit training data's contribution to machine-learning pipelines; however, they often assume passive contributors. In reality, contributors can split datasets across pseudonymous identities, duplicate high-value examples, create near-duplicates, or launder synthetic variants to inflate their share. We formalize this as false-name manipulation in ML data attribution. Our main construction is the quotient semivalue mechanism: compute Shapley-, Banzhaf-, or Beta-style values over evidence-backed attribution clusters instead of raw identities, using a canonical-representative operator to absorb within-cluster duplication. We prove an impossibility: on a fixed monotone data-value game, exact Shapley-fair attribution over reported identities is incompatible with unrestricted false-name-proofness, even on binary-valued instances, and characterize the split-gain of a general semivalue on a unanimity counter-example. The mechanism is exactly false-name-proof under two structural conditions: false-name-neutral within-cluster allocation and quotient-stable manipulations. Under imperfect provenance, when these conditions hold approximately, manipulation gain and fairness loss are bounded by three measurable quantities: escaped-cluster mass, value-estimation error, and clustering distance. We instantiate the mechanisms in DataMarket-Gym, a benchmark for attribution under strategic provider attacks. On synthetic classification tasks, quotient semivalues with example-level evidence reduce manipulation gain on duplicate and near-duplicate Sybil attacks from $1.74$ under baseline Shapley to $0.96$, near the honest level. The cosine-threshold and (false-merge, false-split) rate sweeps trace the corresponding fairness--Sybil frontier.", "date": "2026-05-08", "categories": []}
{"id": "2605.03877", "title": "DMGD: Train-Free Dataset Distillation with Semantic-Distribution Matching in Diffusion Models", "abstract": "Dataset distillation enables efficient training by distilling the information of large-scale datasets into significantly smaller synthetic datasets. Diffusion based paradigms have emerged in recent years, offering novel perspectives for dataset distillation. However, they typically necessitate additional fine-tuning stages, and effective guidance mechanisms remain underexplored. To address these limitations, we rethink diffusion based dataset distillation and propose a Dual Matching Guided Diffusion (DMGD) framework, centered on efficient training-free guidance. We first establish Semantic Matching via conditional likelihood optimization, eliminating the need for auxiliary classifiers. Furthermore, we propose a dynamic guidance mechanism that enhances the diversity of synthetic data while maintaining semantic alignment. Simultaneously, we introduce an optimal transport (OT) based Distribution Matching approach to further align with the target distribution structure. To ensure efficiency, we develop two enhanced strategies for diffusion based framework: Distribution Approximate Matching and Greedy Progressive Matching. These strategies enable effective distribution matching guidance with minimal computational overhead. Experimental results on ImageNet-Woof, ImageNet-Nette, and ImageNet-1K demonstrate that our training-free approach achieves significant improvements, outperforming state-of-the-art (SOTA) methods requiring additional fine-tuning by average accuracy gains of 2.1%, 5.4%, and 2.4%, respectively.", "date": "2026-05-05", "categories": []}
{"id": "2605.00944", "title": "SCARV: Structure-Constrained Aggregation for Stable Sample Ranking in Redundant NLP Datasets", "abstract": "Sample-level rankings are increasingly used in data-centric NLP for analysis, filtering, debugging, and curation, yet existing pipelines typically score training examples pointwise and rank them as if they were independent. This assumption is fragile in the presence of exact duplicates, near-duplicates, paraphrases, and other redundant structure common in NLP corpora, where stochastic training can make highly similar examples receive unstable relative orderings across random seeds. We study stable sample-level ranking under redundancy and propose \\textsc{SCARV}, a modular aggregation framework that operates on top of an existing scoring proxy. \\textsc{SCARV} combines robust multi-seed aggregation with a structure-aware aggregation/allocation step over redundancy clusters. Across synthetic redundancy, naturally mined QQP redundancy, multiple proxy families, several NLP tasks, and end-to-end DistilBERT fine-tuning, \\textsc{SCARV} substantially improves over bare proxy rankings in global and local stability and yields more reproducible ranking-based decisions such as subset selection and suspicious-example retrieval. Our decomposition and compute-aware frontier sharpen the mechanism: robust multi-seed aggregation is the dominant generic stabilizer, while the structure-aware component adds value mainly under low aggregation budgets or when redundancy clusters are informative, naturally occurring, or sufficiently covered. These results position \\textsc{SCARV} not as a universal data selector or a universally dominant replacement for seed-only aggregation, but as a stability-oriented aggregation layer for proxy-induced rankings in redundant NLP datasets.", "date": "2026-05-01", "categories": []}
{"id": "2511.03928", "title": "SynQuE: Estimating Synthetic Dataset Quality Without Annotations", "abstract": "We introduce and formalize the Synthetic Dataset Quality Estimation (SynQuE) problem: ranking synthetic datasets by their expected real-world task performance using only limited unannotated real data. This addresses a critical and open challenge where data is scarce due to collection costs or privacy constraints. We establish the first comprehensive benchmarks for this problem by introducing and evaluating proxy metrics that choose synthetic data for training to maximize task performance on real data. We introduce the first proxy metrics for SynQuE by adapting distribution and diversity-based distance measures to our context via embedding models. To address the shortcomings of these metrics on complex planning tasks, we propose LENS, a novel proxy that leverages large language model reasoning. Our results show that SynQuE proxies correlate with real task performance across diverse tasks, including sentiment analysis, Text2SQL, web navigation, and image classification, with LENS consistently outperforming others on complex tasks by capturing nuanced characteristics. For instance, on text-to-SQL parsing, training on the top-3 synthetic datasets selected via SynQuE proxies can raise accuracy from 30.4% to 38.4 (+8.1)% on average compared to selecting data indiscriminately. This work establishes SynQuE as a practical framework for synthetic data selection under real-data scarcity and motivates future research on foundation model-based data characterization and fine-grained data selection. We release our code.", "date": "2025-11-06", "categories": []}
{"id": "2604.28075", "title": "Repetition over Diversity: High-Signal Data Filtering for Sample-Efficient German Language Modeling", "abstract": "Recent research has shown that filtering massive English web corpora into high-quality subsets significantly improves training efficiency. However, for high-resource non-English languages like German, French, or Japanese, aggressive filtering creates a strategic dilemma: should practitioners prioritize diversity by training once on large amounts of lightly filtered web data, or prioritize quality by strictly filtering for a high-quality core and repeating it over multiple epochs? We investigate this trade-off for German by constructing hierarchical quality filters applied to 500M web documents, comparing multi-epoch training on the filtered subsets against single-pass training on a diverse corpus. Our experiments across multiple model scales and token budgets show that repeating high-quality data consistently outperforms single-pass training on larger, less filtered sets. Notably, the performance gap persists even after 7 epochs. Our findings suggest that for non-English LLMs, semantic concentration through quality filtering offers a more viable path to efficient language modeling than simply maximizing unique data volume. We release our German language models (called Boldt), as well as our cleaned evaluation benchmarks to the research community. Our experiments indicate that they achieve state-of-the-art results despite training on 10-360x fewer tokens than comparable models.", "date": "2026-04-30", "categories": []}
{"id": "2604.25132", "title": "What Makes Good Instruction-Tuning Data? An In-Context Learning Perspective", "abstract": "Instruction-tuning datasets often contain substantial redundancy and low-quality samples, necessitating effective data selection methods. We propose an instruction data selection framework based on weighted in-context influence (wICI), which measures how effectively each candidate example reduces instruction-following difficulty for semantically related peers. Through systematic experiments, we address three key questions: what constitutes effective instruction tuning data from an in-context perspective, whether sample difficulty correlates with in-context influence, and how in-context influence translates to instruction tuning effectiveness. Experiments across multiple models and benchmarks demonstrate that our method consistently outperforms existing baselines under constrained data budgets, while empirically showing that sample difficulty negatively correlates with in-context influence.", "date": "2026-04-28", "categories": []}
{"id": "2508.10180", "title": "For-Value: Efficient Forward-Only Data Valuation for finetuning LLMs and VLMs", "abstract": "Data valuation is essential for enhancing the transparency and accountability of large language models (LLMs) and vision-language models (VLMs). However, existing methods typically rely on gradient computations, making them computationally prohibitive for billion-parameter models and precluding batch parallelization. In this work, we introduce For-Value, a forward-only data valuation framework that enables efficient batch-scalable value estimation while maintaining effectiveness. Leveraging the expressive power of pretrained LLMs/VLMs, we theoretically demonstrate that data valuation can be captured by the alignment between the final hidden representations and prediction errors at the last layer. In light of this insight, For-Value computes data value using a simple closed-form expression with a single forward pass, eliminating the need for costly backpropagation and enabling efficient batch calculating at scale. Extensive experiments show that For-Value matches or outperforms gradient-based baselines in detecting influential data and mislabeled data, while achieving significant efficiency improvements.", "date": "2025-08-13", "categories": []}
{"id": "2509.26158", "title": "Towards Continual Expansion of Data Coverage: Automatic Text-guided Edge-case Synthesis", "abstract": "The performance of deep neural networks is strongly influenced by the quality of their training data. However, mitigating dataset bias by manually curating challenging edge cases remains a major bottleneck. To address this, we propose an automated pipeline for text-guided edge-case synthesis. Our approach employs a Large Language Model, fine-tuned via preference learning, to rephrase image captions into diverse textual prompts that steer a Text-to-Image model toward generating difficult visual scenarios. Evaluated on the FishEye8K object detection benchmark, our method achieves superior robustness, surpassing both naive augmentation and manually engineered prompts. This work establishes a scalable framework that shifts data curation from manual effort to automated, targeted synthesis, offering a promising direction for developing more reliable and continuously improving AI systems. Code is available at https://github.com/gokyeongryeol/ATES.", "date": "2025-09-30", "categories": []}
{"id": "2511.10161", "title": "DenoGrad: A Gradient-Based Framework for Data Refinement in Tabular and Time-Series Learning", "abstract": "In the Data-Centric Artificial Intelligence (AI) paradigm, improving data quality is essential for robust machine learning. However, many denoising methods rely on rigid statistical assumptions or require clean reference data, which limits their applicability in real-world scenarios. In this work, we propose DenoGrad, a gradient-based framework for data refinement that leverages a pretrained neural network to iteratively correct noisy observations by optimizing the input space while keeping the model fixed. DenoGrad is applicable to both tabular regression and time-series forecasting, and incorporates a consensus-based strategy to ensure temporally coherent updates in sequential settings. Experiments on ten real-world datasets show that the proposed approach yields consistent improvements in downstream predictive performance while preserving the statistical structure of the data, as measured by distributional and correlation-based metrics. In addition, DenoGrad can improve generalization in nominally clean datasets, acting as a form of dataset-level regularization. These results support model-guided data refinement as a practical component of data-centric machine learning workflows. Code is available at: https://github.com/ari-dasci/S-DenoGrad.", "date": "2025-11-13", "categories": []}
{"id": "2602.11318", "title": "The Consensus Trap: Dissecting Subjectivity and the \"Ground Truth\" Illusion in Data Annotation", "abstract": "In machine learning, \"ground truth\" refers to the assumed correct labels used to train and evaluate models. However, the foundational \"ground truth\" paradigm rests on a positivistic fallacy that treats human disagreement as technical noise rather than a vital sociotechnical signal. This systematic literature review analyzes research published between 2020 and 2025 across seven premier venues: ACL, AIES, CHI, CSCW, EAAMO, FAccT, and NeurIPS, investigating the mechanisms in data annotation practices that facilitate this \"consensus trap\". Our reflexive thematic analysis of 346 papers reveals that systemic failures in positional legibility, combined with the recent architectural shift toward human-as-verifier models, specifically the reliance on model-mediated annotations, introduce deep-seated anchoring bias and effectively remove human voices from the loop. We further demonstrate how geographic hegemony imposes Western norms as universal benchmarks, often enforced by the performative alignment of precarious data workers who prioritize requester compliance over honest subjectivity to avoid economic penalties. Critiquing the \"noisy sensor\" fallacy, where statistical models misdiagnose pluralism as error, we argue for reclaiming disagreement as a high-fidelity signal essential for building culturally competent models. To address these systemic tensions, we propose a roadmap for pluralistic annotation infrastructures that shift the objective from discovering a singular \"right\" answer to mapping the diversity of human experience.", "date": "2026-02-11", "categories": []}
{"id": "2604.22893", "title": "Utility-Aware Data Pricing: Token-Level Quality and Empirical Training Gain for LLMs", "abstract": "Traditional data valuation methods based on ``row-count $\\times$ quality coefficient'' paradigms fail to capture the nuanced, nonlinear contributions that data makes to Large Language Model (LLM) capabilities. This paper presents a dynamic data valuation framework that transitions from static accounting to utility-based pricing. Our approach operates on three layers: (1) token-level information density metrics using Shannon entropy and Data Quality Scores; (2) empirical training gain measurement through influence functions, proxy model strategies, and Data Shapley values; and (3) cryptographic verifiability through hash-based commitments, Merkle trees, and a tamper-evident training ledger. We provide comprehensive experimental validation on three real domains (instruction following, mathematical reasoning, and code summarization), demonstrating that proxy-based empirical gain achieves near-perfect ranking alignment with realized utility, substantially outperforming row-count and token-count baselines. This framework enables a fair Data-as-a-Service economy where high-reasoning data is priced according to its actual contribution to model intelligence, while providing the transparency and auditability necessary for trustworthy data markets.", "date": "2026-04-24", "categories": []}
{"id": "2604.22693", "title": "CRAFT: Clustered Regression for Adaptive Filtering of Training data", "abstract": "Selecting a small, high-quality subset from a large corpus for fine-tuning is increasingly important as corpora grow to tens of millions of datapoints, making full fine-tuning expensive and often unnecessary. We propose CRAFT (Clustered Regression for Adaptive Filtering of Training data), a vectorization-agnostic selection method for training sequence-to-sequence models. CRAFT decomposes the joint source-target distribution and performs a two-stage selection: (i) match the validation source distribution through proportional budget allocation across k-means clusters, and (ii) within each source cluster, select training pairs whose target embeddings minimize a conditional expected distance derived from the validation target distribution. We prove that proportional cluster allocation bounds the continuous KL divergence between selected and validation distributions, with the residual controlled by cluster diameters. We evaluate CRAFT on English-Hindi translation by selecting training data from 33 million NLLB sentence pairs and fine-tuning mBART via LoRA. CRAFT achieves 43.34 BLEU, outperforming TSDS (41.21) by 2.13 points on the same candidate pool and encoder while completing selection over 40 times faster. With TF-IDF vectorization, the entire pipeline completes in under one minute on CPU. TAROT achieves 45.61 BLEU, but CRAFT completes selection in 26.86 seconds versus TAROT's 75.6 seconds, a 2.8 time speedup.", "date": "2026-04-24", "categories": []}
{"id": "2604.21638", "title": "Geometric Characterisation and Structured Trajectory Surrogates for Clinical Dataset Condensation", "abstract": "Dataset condensation constructs compact synthetic datasets that retain the training utility of large real-world datasets, enabling efficient model development and potentially supporting downstream research in governed domains such as healthcare. Trajectory matching (TM) is a widely used condensation approach that supervises synthetic data using changes in model parameters observed during training on real data, yet the structure of this supervision signal remains poorly understood. In this paper, we provide a geometric characterisation of trajectory matching, showing that a fixed synthetic dataset can only reproduce a limited span of such training-induced parameter changes. When the resulting supervision signal is spectrally broad, this creates a conditional representability bottleneck. Motivated by this mismatch, we propose Bezier Trajectory Matching (BTM), which replaces SGD trajectories with quadratic Bezier trajectory surrogates between initial and final model states. These surrogates are optimised to reduce average loss along the path while replacing broad SGD-derived supervision with a more structured, lower-rank signal that is better aligned with the optimisation constraints of a fixed synthetic dataset, and they substantially reduce trajectory storage. Experiments on five clinical datasets demonstrate that BTM consistently matches or improves upon standard trajectory matching, with the largest gains in low-prevalence and low-synthetic-budget settings. These results indicate that effective trajectory matching depends on structuring the supervision signal rather than reproducing stochastic optimisation paths.", "date": "2026-04-23", "categories": []}
{"id": "2604.20544", "title": "Evian: Towards Explainable Visual Instruction-tuning Data Auditing", "abstract": "The efficacy of Large Vision-Language Models (LVLMs) is critically dependent on the quality of their training data, requiring a precise balance between visual fidelity and instruction-following capability. Existing datasets, however, are plagued by inconsistent quality, and current data filtering methods rely on coarse-grained scores that lack the granularity to identify nuanced semantic flaws like logical fallacies or factual errors. This creates a fundamental bottleneck in developing more reliable models. To address this, we make three core contributions. First, we construct a large-scale, 300K-sample benchmark by systematically injecting diverse, subtle defects to provide a challenging testbed for data auditing. Second, we introduce a novel \"Decomposition-then-Evaluation\" paradigm that breaks model responses into constituent cognitive components: visual description, subjective inference, and factual claim, enabling targeted analysis. Third, we instantiate this paradigm via EVIAN (Explainable Visual Instruction-tuning Data AuditiNg), an automated framework that evaluates these components along the orthogonal axes of Image-Text Consistency, Logical Coherence, and Factual Accuracy. Our empirical findings challenge the prevailing scale-centric paradigm: a model fine-tuned on a compact, high-quality subset curated by EVIAN consistently surpassed models trained on orders-of-magnitude larger datasets. We also reveal that dividing complex auditing into verifiable subtasks enables robust curation, and that Logical Coherence is the most critical factor in data quality evaluation.", "date": "2026-04-22", "categories": []}
{"id": "2604.20549", "title": "Toward Cross-Lingual Quality Classifiers for Multilingual Pretraining Data Selection", "abstract": "As Large Language Models (LLMs) scale, data curation has shifted from maximizing volume to optimizing the signal-to-noise ratio by performing quality filtering. However, for many languages, native high quality data is insufficient to train robust quality classifiers. This work investigates the idea that quality markers in embedding space may show cross-lingual consistency, which would allow high-resource languages to subsidize the filtering of low-resource ones. We evaluate various filtering strategies, including cross-lingual transfer, third quartile sampling (Q3), and retention rate tuning. Our results demonstrate that massive multilingual pooling frequently outperforms monolingual baselines in both rank stability and aggregate accuracy for a 1B model trained on 103B tokens, delivering gains for high resource languages (1.2% increase in aggregate normalized accuracy for French) and matching or exceeding monolingual baselines for low-resource languages. However, we find that scale alone does not guarantee stability. Furthermore, for high-resource languages like French, we show that refining the decision boundary through third quartile sampling (Q3) or tuning the retention rate is necessary to fully leverage the multilingual signal.", "date": "2026-04-22", "categories": []}
{"id": "2604.18811", "title": "Rethinking Dataset Distillation: Hard Truths about Soft Labels", "abstract": "Despite the perceived success of large-scale dataset distillation (DD) methods, recent evidence finds that simple random image baselines perform on-par with state-of-theart DD methods like SRe2L due to the use of soft labels during downstream model training. This is in contrast with the findings in coreset literature, where high-quality coresets consistently outperform random subsets in the hardlabel (HL) setting. To understand this discrepancy, we perform a detailed scalability analysis to examine the role of data quality under different label regimes, ranging from abundant soft labels (termed as SL+KD regime) to fixed soft labels (SL) and hard labels (HL). Our analysis reveals that high-quality coresets fail to convincingly outperform the random baseline in both SL and SL+KD regimes. In the SL+KD setting, performance further approaches nearoptimal levels relative to the full dataset, regardless of subset size or quality, for a given compute budget. This performance saturation calls into question the widespread practice of using soft labels for model evaluation, where unlike the HL setting, subset quality has negligible influence. A subsequent systematic evaluation of five large-scale and four small-scale DD methods in the HL setting reveals that only RDED reliably outperforms random baselines on ImageNet-1K, but can still lag behind strong coreset methods due to its over-reliance on easy sample patches. Based on this, we introduce CAD-Prune, a compute-aware pruning metric that efficiently identifies samples of optimal difficulty for a given compute budget, and use it to develop CA2D, a compute-aligned DD method, outperforming current DD methods on ImageNet-1K at various IPC settings. Together, our findings uncover many insights into current DD research and establish useful tools to advance dataefficient learning for both coresets and DD.", "date": "2026-04-20", "categories": []}
{"id": "2408.11338", "title": "Automatic Dataset Construction (ADC): Sample Collection, Data Curation, and Beyond", "abstract": "Large-scale data collection is essential for developing personalized training data, mitigating the shortage of training data, and fine-tuning specialized models. However, creating high-quality datasets quickly and accurately remains a challenge due to annotation errors, the substantial time and costs associated with human labor. To address these issues, we propose Automatic Dataset Construction (ADC), an innovative methodology that automates dataset creation with negligible cost and high efficiency. Taking the image classification task as a starting point, ADC leverages LLMs for the detailed class design and code generation to collect relevant samples via search engines, significantly reducing the need for manual annotation and speeding up the data generation process. To demonstrate ADC at scale, we construct Clothing-ADC: a dataset of over 1 million images spanning 12 main classes and 12,000 fine-grained subclasses. Our automated curation achieves 79\\% agreement with human annotators and reduces label noise from 22.2\\% to 10.7\\%. Despite these advantages, ADC also encounters real-world challenges such as label errors (label noise) and imbalanced data distributions (label bias). We provide open-source software that incorporates existing methods for label error detection, robust learning under noisy and biased data, ensuring a higher-quality training data and more robust model training procedure. Furthermore, we design three benchmark datasets focused on label noise detection, label noise learning, and class-imbalanced learning. These datasets are vital because there are few existing datasets specifically for label noise detection, despite its importance. Finally, we evaluate the performance of existing popular methods on these datasets, thereby facilitating further research in the field.", "date": "2024-08-21", "categories": []}
{"id": "2505.23114", "title": "Alignment Data Map for Efficient Preference Data Selection and Diagnosis", "abstract": "Human preference data is essential for aligning large language models (LLMs) with human values, but collecting such data is often costly and inefficient-motivating the need for efficient data selection methods that reduce annotation costs while preserving alignment effectiveness. To address this issue, we propose Alignment Data Map, a data analysis tool for identifying and selecting effective preference data. We first evaluate alignment scores of the preference data by LLM-as-a-judge, explicit reward model, and reference-based approaches. The Alignment Data Map considers both response quality and inter-response variability based on the alignment scores. From our experimental findings, training on only 33% of samples that exhibit high-quality and low-variability, achieves comparable or superior alignment performance on MT-Bench, Evol-Instruct, and AlpacaEval, compared to training with the full dataset. In addition, Alignment Data Map detects potential label misannotations by analyzing correlations between annotated labels and alignment scores, improving annotation accuracy. The implementation is available at https://github.com/01choco/Alignment-Data-Map.", "date": "2025-05-29", "categories": []}
{"id": "2506.01942", "title": "OD3: Optimization-free Dataset Distillation for Object Detection", "abstract": "Training large neural networks on large-scale datasets requires substantial computational resources, particularly for dense prediction tasks such as object detection. Although dataset distillation (DD) has been proposed to alleviate these demands by synthesizing compact datasets from larger ones, most existing work focuses solely on image classification, leaving the more complex detection setting largely unexplored. In this paper, we introduce OD3, a novel optimization-free data distillation framework specifically designed for object detection. Our approach involves two stages: first, a candidate selection process in which object instances are iteratively placed in synthesized images based on their suitable locations, and second, a candidate screening process using a pre-trained observer model to remove low-confidence objects. We perform our data synthesis framework on MS COCO and PASCAL VOC, two popular detection datasets, with compression ratios ranging from 0.25% to 5%. Compared to the prior solely existing dataset distillation method on detection and conventional core set selection methods, OD3 delivers superior accuracy, establishes new state-of-the-art results, surpassing prior best method by more than 14% on COCO mAP50 at a compression ratio of 1.0%. Code is available at: https://github.com/VILA-Lab/OD3.", "date": "2025-06-02", "categories": []}
{"id": "2604.10448", "title": "Instruction Data Selection via Answer Divergence", "abstract": "Instruction tuning relies on large instruction-response corpora whose quality and composition strongly affect downstream performance. We propose Answer Divergence-Guided Selection (ADG), which selects instruction data based on the geometric structure of multi-sample outputs. ADG draws several high-temperature generations per instruction, maps responses into an embedding space, and computes an output divergence score that jointly encodes dispersion magnitude and shape anisotropy. High scores correspond to instructions whose answers are both far apart and multi-modal, rather than clustered paraphrases along a single direction. Across two backbones and three public instruction pools, fine-tuning on only 10K ADG-selected examples consistently outperforms strong selectors on six benchmarks spanning reasoning, knowledge, and coding. Analyses further show that both dispersion magnitude and shape anisotropy are necessary, supporting answer divergence as a practical signal for instruction data selection. Code and appendix are included in the supplementary materials.", "date": "2026-04-12", "categories": []}
{"id": "2604.16979", "title": "DOSE: Data Selection for Multi-Modal LLMs via Off-the-Shelf Models", "abstract": "High-quality and diverse multimodal data are essential for improving vision-language models (VLMs), yet existing datasets often contain noisy, redundant, and poorly aligned samples. To address these problems, data filtering is commonly used to enhance the efficiency and performance of multimodal learning, but it introduces extra computational cost because filtering models are usually trained on the same data they are meant to screen. To reduce this cost, we study DOSE, which explores whether off-the-shelf pretrained models that have never seen the target data can be used to select training samples for larger and stronger multimodal models without any task-specific training. Even without fine-tuning, these models can effectively assess text quality and image-text alignment to guide data selection. Based on this, we build a joint quality-alignment distribution and apply adaptive weighted sampling to select informative samples while maintaining long-tail diversity. This approach enhances data diversity, enabling models trained on DOSE-filtered data to match or surpass those trained on the full dataset on standard VQA and math benchmarks. Extensive experiments demonstrate its effectiveness, efficiency, and scalability.", "date": "2026-04-18", "categories": []}
{"id": "2604.18135", "title": "Soft Label Pruning and Quantization for Large-Scale Dataset Distillation", "abstract": "Large-scale dataset distillation requires storing auxiliary soft labels that can be 30-40x larger on ImageNet-1K and 200x larger on ImageNet-21K than the condensed images, undermining the goal of dataset compression. We identify two fundamental issues necessitating such extensive labels: (1) insufficient image diversity, where high within-class similarity in synthetic images requires extensive augmentation, and (2) insufficient supervision diversity, where limited variety in supervisory signals during training leads to performance degradation at high compression rates. To address these challenges, we propose Label Pruning and Quantization for Large-scale Distillation (LPQLD). We enhance image diversity via class-wise batching and batch-normalization supervision during synthesis. For supervision diversity, we introduce Label Pruning with Dynamic Knowledge Reuse to improve label-per-augmentation diversity, and Label Quantization with Calibrated Student-Teacher Alignment to improve augmentation-per-image diversity. Our approach reduces soft label storage by 78x on ImageNet-1K and 500x on ImageNet-21K while improving accuracy by up to 7.2% and 2.8%, respectively. Extensive experiments validate the superiority of LPQLD across different network architectures and dataset distillation methods. Code is available at https://github.com/he-y/soft-label-pruning-quantization-for-dataset-distillation.", "date": "2026-04-20", "categories": []}
{"id": "2604.15706", "title": "Target-Oriented Pretraining Data Selection via Neuron-Activated Graph", "abstract": "Everyday tasks come with a target, and pretraining models around this target is what turns them into experts. In this paper, we study target-oriented language model (LM) pretraining by introducing Neuron-Activated Graph Ranking (NAG-based Ranking), a training-free and interpretable framework for target pretraining data selection. Rather than using black-box representations, our approach directly characterizes each target input by a sparse set of high-impact neurons in any off-the-shelf LLMs. Concretely, we quantify neuron impact and select the most influential neurons across layers into a compact Neuron-Activated Graph (NAG), and rank candidate data by NAG similarity to target examples. We conduct experiments across six benchmarks, where our NAG-based Ranking improves target-oriented pretraining by 4.9% on average over random sampling, and also outperforms state-of-the-art baselines by 5.3% accuracy on HellaSwag. It also remains effective under a more applicable multi-target setting, where our best setup surpasses two baselines by 1.1% and 4.1%, respectively. Furthermore, we provide a comprehensive analysis on why and how our NAG works, e.g., deactivating NAG-selected neurons (only 0.12% of all) causes a 23.5% performance collapse, and restricting NAG to the final layer incurs a 4.1% average drop, indicating that NAG captures a sparse \"functional backbone\" for learning target features. We release the code at https://github.com/asillycat/NAG.", "date": "2026-04-17", "categories": []}
{"id": "2510.26491", "title": "Data-Efficient RLVR via Off-Policy Influence Guidance", "abstract": "Data selection is a critical aspect of Reinforcement Learning with Verifiable Rewards (RLVR) for enhancing the reasoning capabilities of large language models (LLMs). Current data selection methods are largely heuristic-based, lacking theoretical guarantees and generalizability. This work proposes a theoretically-grounded approach using influence functions to estimate the contribution of each data point to the learning objective. To overcome the prohibitive computational cost of policy rollouts required for online influence estimation, we introduce an off-policy influence estimation method that efficiently approximates data influence using pre-collected offline trajectories. Furthermore, to manage the high-dimensional gradients of LLMs, we employ sparse random projection to reduce dimensionality and improve storage and computation efficiency. Leveraging these techniques, we develop \\textbf{C}urriculum \\textbf{R}L with \\textbf{O}ff-\\textbf{P}olicy \\text{I}nfluence guidance (\\textbf{CROPI}), a multi-stage RL framework that iteratively selects the most influential data for the current policy. Experiments on models up to 7B parameters demonstrate that CROPI significantly accelerates training. On a 1.5B model, it achieves a 2.66x step-level acceleration while using only 10\\% of the data per stage compared to full-dataset training. Our results highlight the substantial potential of influence-based data selection for efficient RLVR.", "date": "2025-10-30", "categories": []}
{"id": "2604.13610", "title": "What Are We Really Measuring? Rethinking Dataset Bias in Web-Scale Natural Image Collections via Unsupervised Semantic Clustering", "abstract": "In computer vision, a prevailing method for quantifying dataset bias is to train a model to distinguish between datasets. High classification accuracy is then interpreted as evidence of meaningful semantic differences. This approach assumes that standard image augmentations successfully suppress low-level, non-semantic cues, and that any remaining performance must therefore reflect true semantic divergence. We demonstrate that this fundamental assumption is flawed within the domain of large-scale natural image collections. High classification accuracy is often driven by resolution-based artifacts, which are structural fingerprints arising from native image resolution distributions and interpolation effects during resizing. These artifacts form robust, dataset-specific signatures that persist despite conventional image corruptions. Through controlled experiments, we show that models achieve strong dataset classification even on non-semantic, procedurally generated images, proving their reliance on superficial cues. To address this issue, we revisit this decades-old idea of dataset separability, but not with supervised classification. Instead, we introduce an unsupervised approach that measures true semantic separability. Our framework directly assesses semantic similarity by clustering semantically-rich features from foundational vision models, deliberately bypassing supervised classification on dataset labels. When applied to major web-scale datasets, the primary focus of this work, the high separability reported by supervised methods largely vanishes, with clustering accuracy dropping to near-chance levels. This reveals that conventional classification-based evaluation systematically overstates semantic bias by an overwhelming margin.", "date": "2026-04-15", "categories": []}
{"id": "2512.24503", "title": "Can Small Training Runs Reliably Guide Data Curation? Rethinking Proxy-Model Practice", "abstract": "Data teams at frontier AI companies routinely train small proxy models to make critical decisions about pretraining data recipes for full-scale training runs. However, the community has a limited understanding of whether and when conclusions drawn from small-scale experiments reliably transfer to full-scale model training. In this work, we uncover a subtle yet critical issue in the standard experimental protocol for data recipe assessment: the use of identical small-scale model training configurations across all data recipes in the name of \"fair\" comparison. We show that the experiment conclusions about data quality can flip with even minor adjustments to training hyperparameters, as the optimal training configuration is inherently data-dependent. Moreover, this fixed-configuration protocol diverges from full-scale model development pipelines, where hyperparameter optimization is a standard step. Consequently, we posit that the objective of data recipe assessment should be to identify the recipe that yields the best performance under data-specific tuning. To mitigate the high cost of hyperparameter tuning, we introduce a simple patch to the evaluation protocol: using reduced learning rates for proxy model training. We show that this approach yields relative performance that strongly correlates with that of fully tuned large-scale LLM pretraining runs. Theoretically, we prove that for random-feature models, this approach preserves the ordering of datasets according to their optimal achievable loss. Empirically, we validate this approach across 23 data recipes covering four critical dimensions of data curation, demonstrating dramatic improvements in the reliability of small-scale experiments.", "date": "2025-12-30", "categories": []}
{"id": "2604.10480", "title": "Tracing the Roots: A Multi-Agent Framework for Uncovering Data Lineage in Post-Training LLMs", "abstract": "Post-training data plays a pivotal role in shaping the capabilities of Large Language Models (LLMs), yet datasets are often treated as isolated artifacts, overlooking the systemic connections that underlie their evolution. To disentangle these complex relationships, we introduce the concept of \\textbf{data lineage} to the LLM ecosystem and propose an automated multi-agent framework to reconstruct the evolutionary graph of dataset development. Through large-scale lineage analysis, we characterize domain-specific structural patterns, such as vertical refinement in math-oriented datasets and horizontal aggregation in general-domain corpora. Moreover, we uncover pervasive systemic issues, including \\textit{structural redundancy} induced by implicit dataset intersections and the \\textit{propagation of benchmark contamination} along lineage paths. To demonstrate the practical value of lineage analysis for data construction, we leverage the reconstructed lineage graph to create a \\textit{lineage-aware diversity-oriented dataset}. By anchoring instruction sampling at upstream root sources, this approach mitigates downstream homogenization and hidden redundancy, yielding a more diverse post-training corpus. We further highlight lineage-centric analysis as an efficient and robust topological alternative to sample-level dataset comparison for large-scale data ecosystems. By grounding data construction in explicit lineage structures, our work advances post-training data curation toward a more systematic and controllable paradigm.", "date": "2026-04-12", "categories": []}
{"id": "2604.10666", "title": "Omnimodal Dataset Distillation via High-order Proxy Alignment", "abstract": "Dataset distillation compresses large-scale datasets into compact synthetic sets while preserving training performance, but existing methods are largely restricted to single-modal or bimodal settings. Extending dataset distillation to scenarios involving more than two modalities, i.e., Omnimodal Dataset Distillation, remains underexplored and challenging due to increased heterogeneity and complex cross-modal interactions. In this work, we identify the key determinant that bounds the endpoint discrepancy in the omnimodal setting, which is exacerbated with an increasing number of modalities. To this end, we propose HoPA, a unified method that captures high-order cross-modal alignments via a compact proxy, which is compatible with trajectory matching as well. By abstracting omnimodal alignment with a shared similarity structure, our method avoids the combinatorial complexity of pairwise modality modeling and enables scalable joint distillation across heterogeneous modalities. Theoretical analysis from the spectral perspective reveals the rationality of our proposed method against bimodal dataset distillation techniques. Extensive experiments on various benchmarks demonstrate that the proposed method achieves superior compression-performance trade-offs compared to existing competitors. The source code will be publicly released.", "date": "2026-04-12", "categories": []}
{"id": "2604.10952", "title": "UniPROT: Uniform Prototype Selection via Partial Optimal Transport with Submodular Guarantees", "abstract": "Selecting prototypical examples from a source distribution to represent a target data distribution is a fundamental problem in machine learning. Existing subset selection methods often rely on implicit importance scores, which can be skewed towards majority classes and lead to low-quality prototypes for minority classes. We present $\\methodprop$, a novel subset selection framework that minimizes the optimal transport (OT) distance between a uniformly weighted prototypical distribution and the target distribution. While intuitive, this formulation leads to a cardinality-constrained maximization of a \\emph{super-additive} objective, which is generally intractable to approximate efficiently. To address this, we propose a principled reformulation of the OT marginal constraints, yielding a partial optimal transport-based submodular objective. We prove that this reformulation enables a greedy algorithm with a $(1-1/e)$ approximation guarantee relative to the original super-additive maximization problem. Empirically, we showcase that enforcing uniform prototype weights in UniPROT consistently improves minority-class representation in imbalanced classification benchmarks without compromising majority-class accuracy. In both finetuning and pretraining regimes for large language models under domain imbalance, UniPROT enforces uniform source contributions, yielding robust performance gains. Our results establish UniPROT as a scalable, theoretically grounded solution for uniform-weighted prototype selection. Our code is publicly available at GitHub\\footnote{Code: https://github.com/efficiency-learning/UniPROT}", "date": "2026-04-13", "categories": []}
{"id": "2604.04771", "title": "MinerU2.5-Pro: Pushing the Limits of Data-Centric Document Parsing at Scale", "abstract": "Current document parsing methods advance primarily through model architecture innovation, while systematic engineering of training data remains underexplored. Yet state-of-the-art models spanning diverse architectures and parameter scales exhibit highly consistent failure patterns on the same set of hard samples, suggesting that the performance bottleneck stems from shared deficiencies in training data rather than from architectural differences. Building on this finding, we present MinerU2.5-Pro, which advances the state of the art purely through data engineering and training strategy design while retaining the 1.2B-parameter architecture of MinerU2.5 unchanged. At its core is a Data Engine co-designed around coverage, informativeness, and annotation accuracy: Diversity-and-Difficulty-Aware Sampling expands training data from under 10M to 65.5M samples while mitigating distribution shift; Cross-Model Consistency Verification leverages output consensus among heterogeneous models to assess sample difficulty and generate reliable annotations; the Judge-and-Refine pipeline improves annotation quality for hard samples through render-then-verify iterative correction. A three-stage progressive training strategy--large-scale pre-training, hard sample fine-tuning, and GRPO alignment--sequentially exploits these data at different quality tiers. On the evaluation front, we rectify element-matching biases in OmniDocBench v1.5 and introduce a Hard subset, establishing the more discriminative OmniDocBench v1.6 protocol. Without any architectural modification, MinerU2.5-Pro achieves 95.69 on OmniDocBench v1.6, improving over the same-architecture baseline by 2.71 points and surpassing all existing methods, including those based on models with over 200x more parameters.", "date": "2026-04-06", "categories": []}
{"id": "2604.07591", "title": "From Ground Truth to Measurement: A Statistical Framework for Human Labeling", "abstract": "Supervised machine learning assumes that labeled data provide accurate measurements of the concepts models are meant to learn. Yet in practice, human labeling introduces systematic variation arising from ambiguous items, divergent interpretations, and simple mistakes. Machine learning research commonly treats all disagreement as noise, which obscures these distinctions and limits our understanding of what models actually learn. This paper reframes annotation as a measurement process and introduces a statistical framework for decomposing labeling outcomes into interpretable sources of variation: instance difficulty, annotator bias, situational noise, and relational alignment. The framework extends classical measurement-error models to accommodate both shared and individualized notions of truth, reflecting traditional and human label variation interpretations of error, and provides a diagnostic for assessing which regime better characterizes a given task. Applying the proposed model to a multi-annotator natural language inference dataset, we find empirical evidence for all four theorized components and demonstrate the effectiveness of our approach. We conclude with implications for data-centric machine learning and outline how this approach can guide the development of a more systematic science of labeling.", "date": "2026-04-08", "categories": []}
{"id": "2604.08366", "title": "Scaling-Aware Data Selection for End-to-End Autonomous Driving Systems", "abstract": "Large-scale deep learning models for physical AI applications depend on diverse training data collection efforts. These models and correspondingly, the training data, must address different evaluation criteria necessary for the models to be deployable in real-world environments. Data selection policies can guide the development of the training set, but current frameworks do not account for the ambiguity in how data points affect different metrics. In this work, we propose Mixture Optimization via Scaling-Aware Iterative Collection (MOSAIC), a general data selection framework that operates by: (i) partitioning the dataset into domains; (ii) fitting neural scaling laws from each data domain to the evaluation metrics; and (iii) optimizing a data mixture by iteratively adding data from domains that maximize the change in metrics. We apply MOSAIC to autonomous driving (AD), where an End-to-End (E2E) planner model is evaluated on the Extended Predictive Driver Model Score (EPDMS), an aggregate of driving rule compliance metrics. Here, MOSAIC outperforms a diverse set of baselines on EPDMS with up to 80\\% less data.", "date": "2026-04-09", "categories": []}
{"id": "2601.23155", "title": "SPICE: Submodular Penalized Information-Conflict Selection for Efficient Large Language Model Training", "abstract": "Information-based data selection for instruction tuning is compelling: maximizing the log-determinant of the Fisher information yields a monotone submodular objective, enabling greedy algorithms to achieve a $(1-1/e)$ approximation under a cardinality budget. In practice, however, we identify alleviating gradient conflicts, misalignment between per-sample gradients, is a key factor that slows down the decay of marginal log-determinant information gains, thereby preventing significant loss of information. We formalize this via an $\\varepsilon$-decomposition that quantifies the deviation from ideal submodularity as a function of conflict statistics, yielding data-dependent approximation factors that tighten as conflicts diminish. Guided by this analysis, we propose SPICE, a conflict-aware selector that maximizes information while penalizing misalignment, and that supports early stopping and proxy models for efficiency. Empirically, SPICE selects subsets with higher log-determinant information than original criteria, and these informational gains translate into performance improvements: across 8 benchmarks with LLaMA2-7B and Qwen2-7B, SPICE uses only 10% of the data, yet matches or exceeds 6 methods including full-data tuning. This achieves performance improvements with substantially lower training cost.", "date": "2026-01-30", "categories": []}
{"id": "2604.06834", "title": "On the Step Length Confounding in LLM Reasoning Data Selection", "abstract": "Large reasoning models have recently demonstrated strong performance on complex tasks that require long chain-of-thought reasoning, through supervised fine-tuning on large-scale and high-quality datasets. To construct such datasets, existing pipelines generate long reasoning data from more capable Large Language Models (LLMs) and apply manually heuristic or naturalness-based selection methods to filter high-quality samples. Despite the proven effectiveness of naturalness-based data selection, which ranks data by the average log probability assigned by LLMs, our analysis shows that, when applied to LLM reasoning datasets, it systematically prefers samples with longer reasoning steps (i.e., more tokens per step) rather than higher-quality ones, a phenomenon we term step length confounding. Through quantitative analysis, we attribute this phenomenon to low-probability first tokens in reasoning steps; longer steps dilute their influence, thereby inflating the average log probabilities. To address this issue, we propose two variant methods: ASLEC-DROP, which drops first-token probabilities when computing average log probability, and ASLEC-CASL, which applies a causal debiasing regression to remove the first tokens' confounding effect. Experiments across four LLMs and five evaluation benchmarks demonstrate the effectiveness of our approach in mitigating the step length confounding problem.", "date": "2026-04-08", "categories": []}
{"id": "2510.23409", "title": "Eigen-Value: Efficient Domain-Robust Data Valuation via Eigenvalue-Based Approach", "abstract": "Data valuation has become central in the era of data-centric AI. It drives efficient training pipelines and enables objective pricing in data markets by assigning a numeric value to each data point. Most existing data valuation methods estimate the effect of removing individual data points by evaluating changes in model validation performance under in-distribution (ID) settings, as opposed to out-of-distribution (OOD) scenarios where data follow different patterns. Since ID and OOD data behave differently, data valuation methods based on ID loss often fail to generalize to OOD settings, particularly when the validation set contains no OOD data. Furthermore, although OOD-aware methods exist, they involve heavy computational costs, which hinder practical deployment. To address these challenges, we introduce \\emph{Eigen-Value} (EV), a plug-and-play data valuation framework for OOD robustness that uses only an ID data subset, including during validation. EV provides a new spectral approximation of domain discrepancy, which is the gap of loss between ID and OOD using ratios of eigenvalues of ID data's covariance matrix. EV then estimates the marginal contribution of each data point to this discrepancy via perturbation theory, alleviating the computational burden. Subsequently, EV plugs into ID loss-based methods by adding an EV term without any additional training loop. We demonstrate that EV achieves improved OOD robustness and stable value rankings across real-world datasets, while remaining computationally lightweight. These results indicate that EV is practical for large-scale settings with domain shift, offering an efficient path to OOD-robust data valuation.", "date": "2025-10-27", "categories": []}
{"id": "2604.05993", "title": "Data Distribution Valuation Using Generalized Bayesian Inference", "abstract": "We investigate the data distribution valuation problem, which aims to quantify the values of data distributions from their samples. This is a recently proposed problem that is related to but different from classical data valuation and can be applied to various applications. For this problem, we develop a novel framework called Generalized Bayes Valuation that utilizes generalized Bayesian inference with a loss constructed from transferability measures. This framework allows us to solve, in a unified way, seemingly unrelated practical problems, such as annotator evaluation and data augmentation. Using the Bayesian principles, we further improve and enhance the applicability of our framework by extending it to the continuous data stream setting. Our experiment results confirm the effectiveness and efficiency of our framework in different real-world scenarios.", "date": "2026-04-07", "categories": []}
{"id": "2601.02081", "title": "ASSS: A Differentiable Adversarial Framework for Task-Aware Data Reduction", "abstract": "Massive datasets often contain redundancy that inflates computational costs without improving generalization. Existing data reduction methods are typically task-agnostic, discarding informative boundary samples and yielding suboptimal performance. We propose Adversarial Soft-Selection Subsampling (ASSS), a differentiable framework that casts data reduction as a minimax game between a learnable selector and a task network. Using Gumbel-Softmax relaxation, ASSS enables end-to-end gradient flow and is theoretically grounded in the information bottleneck principle. Experiments on multiple benchmarks show that ASSS achieves a performance retention rate (PRR) of 98.9% while using only 30% of the data, significantly outperforming random sampling, K-means, and gradient-based methods. Visualizations confirm that ASSS preferentially retains samples near decision boundaries. The framework is scalable, fully differentiable, and easily integrated into existing training pipelines. This work introduces a new paradigm for task-aware data reduction that directly optimizes subset selection for the downstream objective, offering a principled and practical solution to the scalability challenges in modern deep learning.", "date": "2026-01-05", "categories": []}
{"id": "2604.04681", "title": "Batch Loss Score for Dynamic Data Pruning", "abstract": "Dynamic data pruning accelerates deep learning by selectively omitting less informative samples during training. While per-sample loss is a common importance metric, obtaining it can be challenging or infeasible for complex models or loss functions, often requiring significant implementation effort. This work proposes the Batch Loss Score (BLS), a computationally efficient alternative using an Exponential Moving Average (EMA) of readily available batch losses to assign scores to individual samples. We frame the batch loss, from the perspective of a single sample, as a noisy measurement of its scaled individual loss, with noise originating from stochastic batch composition. It is formally shown that the EMA mechanism functions as a first-order low-pass filter, attenuating high-frequency batch composition noise. This yields a score approximating the smoothed and persistent contribution of the individual sample to the loss, providing a theoretical grounding for BLS as a proxy for sample importance. BLS demonstrates remarkable code integration simplicity (\\textbf{three-line injection}) and readily adapts existing per-sample loss-based methods (\\textbf{one-line proxy}). Its effectiveness is demonstrated by enhancing two such methods to losslessly prune \\textbf{20\\%-50\\%} of samples across \\textit{14 datasets}, \\textit{11 tasks} and \\textit{18 models}, highlighting its utility and broad applicability, especially for complex scenarios where per-sample loss is difficult to access. Code is available at https://github.com/mrazhou/BLS.", "date": "2026-04-06", "categories": []}
{"id": "2510.17421", "title": "Diffusion Models as Dataset Distillation Priors", "abstract": "Dataset distillation aims to synthesize compact yet informative datasets from large ones. A significant challenge in this field is achieving a trifecta of diversity, generalization, and representativeness in a single distilled dataset. Although recent generative dataset distillation methods adopt powerful diffusion models as their foundation models, the inherent representativeness prior in diffusion models is overlooked. Consequently, these approaches often necessitate the integration of external constraints to enhance data quality. To address this, we propose Diffusion As Priors (DAP), which formalizes representativeness by quantifying the similarity between synthetic and real data in feature space using a Mercer kernel. We then introduce this prior as guidance to steer the reverse diffusion process, enhancing the representativeness of distilled samples without any retraining. Extensive experiments on large-scale datasets, such as ImageNet-1K and its subsets, demonstrate that DAP outperforms state-of-the-art methods in generating high-fidelity datasets while achieving superior cross-architecture generalization. Our work not only establishes a theoretical connection between diffusion priors and the objectives of dataset distillation but also provides a practical, training-free framework for improving the quality of the distilled dataset.", "date": "2025-10-20", "categories": []}
{"id": "2604.00519", "title": "Learnability-Guided Diffusion for Dataset Distillation", "abstract": "Training machine learning models on massive datasets is expensive and time-consuming. Dataset distillation addresses this by creating a small synthetic dataset that achieves the same performance as the full dataset. Recent methods use diffusion models to generate distilled data, either by promoting diversity or matching training gradients. However, existing approaches produce redundant training signals, where samples convey overlapping information. Empirically, disjoint subsets of distilled datasets capture 80-90% overlapping signals. This redundancy stems from optimizing visual diversity or average training dynamics without accounting for similarity across samples, leading to datasets where multiple samples share similar information rather than complementary knowledge. We propose learnability-driven dataset distillation, which constructs synthetic datasets incrementally through successive stages. Starting from a small set, we train a model and generate new samples guided by learnability scores that identify what the current model can learn from, creating an adaptive curriculum. We introduce Learnability-Guided Diffusion (LGD), which balances training utility for the current model with validity under a reference model to generate curriculum-aligned samples. Our approach reduces redundancy by 39.1%, promotes specialization across training stages, and achieves state-of-the-art results on ImageNet-1K (60.1%), ImageNette (87.2%), and ImageWoof (72.9%). Our code is available on our project page https://jachansantiago.github.io/learnability-guided-distillation/.", "date": "2026-04-01", "categories": []}
{"id": "2603.29791", "title": "Reasoning-Driven Synthetic Data Generation and Evaluation", "abstract": "Although many AI applications of interest require specialized multi-modal models, relevant data to train such models is inherently scarce or inaccessible. Filling these gaps with human annotators is prohibitively expensive, error-prone, and time-consuming, leading model builders to increasingly consider synthetic data as a scalable alternative. However, existing synthetic data generation methods often rely on manual prompts, evolutionary algorithms, or extensive seed data from the target distribution - limiting their scalability, explainability, and control. In this paper, we introduce Simula: a novel reasoning-driven framework for data generation and evaluation. It employs a seedless, agentic approach to generate synthetic datasets at scale, allowing users to define desired dataset characteristics through an explainable and controllable process that enables fine-grained resource allocation. We show the efficacy of our approach on a variety of datasets, rigorously testing both intrinsic and downstream properties. Our work (1) offers guidelines for synthetic data mechanism design, (2) provides insights into generating and evaluating synthetic data at scale, and (3) unlocks new opportunities for developing and deploying AI in domains where data scarcity or privacy concerns are paramount.", "date": "2026-03-31", "categories": []}
{"id": "2501.10677", "title": "Class-Imbalanced-Aware Adaptive Dataset Distillation for Scalable Pretrained Model on Credit Scoring", "abstract": "The advent of artificial intelligence has significantly enhanced credit scoring technologies. Despite the remarkable efficacy of advanced deep learning models, mainstream adoption continues to favor tree-structured models due to their robust predictive performance on tabular data. Although pretrained models have seen considerable development, their application within the financial realm predominantly revolves around question-answering tasks and the use of such models for tabular-structured credit scoring datasets remains largely unexplored. Tabular-oriented large models, such as TabPFN, has made the application of large models in credit scoring feasible, albeit can only processing with limited sample sizes. This paper provides a novel framework to combine tabular-tailored dataset distillation technique with the pretrained model, empowers the scalability for TabPFN. Furthermore, though class imbalance distribution is the common nature in financial datasets, its influence during dataset distillation has not been explored. We thus integrate the imbalance-aware techniques during dataset distillation, resulting in improved performance in financial datasets (e.g., a 2.5% enhancement in AUC). This study presents a novel framework for scaling up the application of large pretrained models on financial tabular datasets and offers a comparative analysis of the influence of class imbalance on the dataset distillation process. We believe this approach can broaden the applications and downstream tasks of large models in the financial domain.", "date": "2025-01-18", "categories": []}
{"id": "2603.14830", "title": "Dataset Distillation Efficiently Encodes Low-Dimensional Representations from Gradient-Based Learning of Non-Linear Tasks", "abstract": "Dataset distillation, a training-aware data compression technique, has recently attracted increasing attention as an effective tool for mitigating costs of optimization and data storage. However, progress remains largely empirical. Mechanisms underlying the extraction of task-relevant information from the training process and the efficient encoding of such information into synthetic data points remain elusive. In this paper, we theoretically analyze practical algorithms of dataset distillation applied to the gradient-based training of two-layer neural networks with width $L$. By focusing on a non-linear task structure called multi-index model, we prove that the low-dimensional structure of the problem is efficiently encoded into the resulting distilled data. This dataset reproduces a model with high generalization ability for a required memory complexity of $\\tilde{\\Theta}$$(r^2d+L)$, where $d$ and $r$ are the input and intrinsic dimensions of the task. To the best of our knowledge, this is one of the first theoretical works that include a specific task structure, leverage its intrinsic dimensionality to quantify the compression rate and study dataset distillation implemented solely via gradient-based algorithms.", "date": "2026-03-16", "categories": []}
{"id": "2603.27987", "title": "Beyond Dataset Distillation: Lossless Dataset Concentration via Diffusion-Assisted Distribution Alignment", "abstract": "The high cost and accessibility problem associated with large datasets hinder the development of large-scale visual recognition systems. Dataset Distillation addresses these problems by synthesizing compact surrogate datasets for efficient training, storage, transfer, and privacy preservation. The existing state-of-the-art diffusion-based dataset distillation methods face three issues: lack of theoretical justification, poor efficiency in scaling to high data volumes, and failure in data-free scenarios. To address these issues, we establish a theoretical framework that justifies the use of diffusion models by proving the equivalence between dataset distillation and distribution matching, and reveals an inherent efficiency limit in the dataset distillation paradigm. We then propose a Dataset Concentration (DsCo) framework that uses a diffusion-based Noise-Optimization (NOpt) method to synthesize a small yet representative set of samples, and optionally augments the synthetic data via \"Doping\", which mixes selected samples from the original dataset with the synthetic samples to overcome the efficiency limit of dataset distillation. DsCo is applicable in both data-accessible and data-free scenarios, achieving SOTA performances for low data volumes, and it extends well to high data volumes, where it nearly reduces the dataset size by half with no performance degradation.", "date": "2026-03-30", "categories": []}
{"id": "2603.28092", "title": "InkDrop: Invisible Backdoor Attacks Against Dataset Condensation", "abstract": "Dataset Condensation (DC) is a data-efficient learning paradigm that synthesizes small yet informative datasets, enabling models to match the performance of full-data training. However, recent work exposes a critical vulnerability of DC to backdoor attacks, where malicious patterns (\\textit{e.g.}, triggers) are implanted into the condensation dataset, inducing targeted misclassification on specific inputs. Existing attacks always prioritize attack effectiveness and model utility, overlooking the crucial dimension of stealthiness. To bridge this gap, we propose InkDrop, which enhances the imperceptibility of malicious manipulation without degrading attack effectiveness and model utility. InkDrop leverages the inherent uncertainty near model decision boundaries, where minor input perturbations can induce semantic shifts, to construct a stealthy and effective backdoor attack. Specifically, InkDrop first selects candidate samples near the target decision boundary that exhibit latent semantic affinity to the target class. It then learns instance-dependent perturbations constrained by perceptual and spatial consistency, embedding targeted malicious behavior into the condensed dataset. Extensive experiments across diverse datasets validate the overall effectiveness of InkDrop, demonstrating its ability to integrate adversarial intent into condensed datasets while preserving model utility and minimizing detectability. Our code is available at https://github.com/lvdongyi/InkDrop.", "date": "2026-03-30", "categories": []}
{"id": "2603.25144", "title": "FD$^2$: A Dedicated Framework for Fine-Grained Dataset Distillation", "abstract": "Dataset distillation (DD) compresses a large training set into a small synthetic set, reducing storage and training cost, and has shown strong results on general benchmarks. Decoupled DD further improves efficiency by splitting the pipeline into pretraining, sample distillation, and soft-label generation. However, existing decoupled methods largely rely on coarse class-label supervision and optimize samples within each class in a nearly identical manner. On fine-grained datasets, this often yields distilled samples that (i) retain large intra-class variation with subtle inter-class differences and (ii) become overly similar within the same class, limiting localized discriminative cues and hurting recognition. To solve the above-mentioned problems, we propose FD$^{2}$, a dedicated framework for Fine-grained Dataset Distillation. FD$^{2}$ localizes discriminative regions and constructs fine-grained representations for distillation. During pretraining, counterfactual attention learning aggregates discriminative representations to update class prototypes. During distillation, a fine-grained characteristic constraint aligns each sample with its class prototype while repelling others, and a similarity constraint diversifies attention across same-class samples. Experiments on multiple fine-grained and general datasets show that FD$^{2}$ integrates seamlessly with decoupled DD and improves performance in most settings, indicating strong transferability.", "date": "2026-03-26", "categories": []}
{"id": "2603.25388", "title": "Multimodal Dataset Distillation via Phased Teacher Models", "abstract": "Multimodal dataset distillation aims to construct compact synthetic datasets that enable efficient compression and knowledge transfer from large-scale image-text data. However, existing approaches often fail to capture the complex, dynamically evolving knowledge embedded in the later training stages of teacher models. This limitation leads to degraded student performance and compromises the quality of the distilled data. To address critical challenges such as pronounced cross-stage performance gaps and unstable teacher trajectories, we propose Phased Teacher Model with Shortcut Trajectory (PTM-ST) -- a novel phased distillation framework. PTM-ST leverages stage-aware teacher modeling and a shortcut-based trajectory construction strategy to accurately fit the teacher's learning dynamics across distinct training phases. This enhances both the stability and expressiveness of the distillation process. Through theoretical analysis and comprehensive experiments, we show that PTM-ST significantly mitigates optimization oscillations and inter-phase knowledge gaps, while also reducing storage overhead. Our method consistently surpasses state-of-the-art baselines on Flickr30k and COCO, achieving up to 13.5% absolute improvement and an average gain of 9.53% on Flickr30k. Code: https://github.com/Previsior/PTM-ST.", "date": "2026-03-26", "categories": []}
{"id": "2505.22564", "title": "PRISM: Video Dataset Condensation with Progressive Refinement and Insertion for Sparse Motion", "abstract": "Video dataset condensation aims to reduce the immense computational cost of video processing. However, it faces a fundamental challenge regarding the inseparable interdependence between spatial appearance and temporal dynamics. Prior work follows a static/dynamic disentanglement paradigm where videos are decomposed into static content and auxiliary motion signals. This multi-stage approach often misrepresents the intrinsic coupling of real-world actions. We introduce Progressive Refinement and Insertion for Sparse Motion (PRISM), a holistic approach that treats the video as a unified and fully coupled spatiotemporal structure from the outset. To maximize representational efficiency, PRISM addresses the inherent temporal redundancy of video by avoiding fixed-frame optimization. It begins with minimal temporal anchors and progressively inserts key-frames only where linear interpolation fails to capture non-linear dynamics. These critical moments are identified through gradient misalignments. Such an adaptive process ensures that representational capacity is allocated precisely where needed, minimizing storage requirements while preserving complex motion. Extensive experiments demonstrate that PRISM achieves competitive performance across standard benchmarks while providing state-of-the-art storage efficiency through its sparse and holistically learned representation.", "date": "2025-05-28", "categories": []}
{"id": "2603.20644", "title": "ScaleEdit-12M: Scaling Open-Source Image Editing Data Generation via Multi-Agent Framework", "abstract": "Instruction-based image editing has emerged as a key capability for unified multimodal models (UMMs), yet constructing large-scale, diverse, and high-quality editing datasets without costly proprietary APIs remains challenging. Previous image editing datasets either rely on closed-source models for annotation, which prevents cost-effective scaling, or employ fixed synthetic editing pipelines, which suffer from limited quality and generalizability. To address these challenges, we propose ScaleEditor, a fully open-source hierarchical multi-agent framework for end-to-end construction of large-scale, high-quality image editing datasets. Our pipeline consists of three key components: source image expansion with world-knowledge infusion, adaptive multi-agent editing instruction-image synthesis, and a task-aware data quality verification mechanism. Using ScaleEditor, we curate ScaleEdit-12M, the largest open-source image editing dataset to date, spanning 23 task families across diverse real and synthetic domains. Fine-tuning UniWorld-V1 and Bagel on ScaleEdit yields consistent gains, improving performance by up to 10.4% on ImgEdit and 35.1% on GEdit for general editing benchmarks and by up to 150.0% on RISE and 26.5% on KRIS-Bench for knowledge-infused benchmarks. These results demonstrate that open-source, agentic pipelines can approach commercial-grade data quality while retaining cost-effectiveness and scalability. Both the framework and dataset will be open-sourced.", "date": "2026-03-21", "categories": []}
{"id": "2603.21819", "title": "Ctrl-A: Control-Driven Online Data Augmentation", "abstract": "We introduce ControlAugment (Ctrl-A), an automated data augmentation algorithm for image-vision tasks, which incorporates principles from control theory for online adjustment of augmentation strength distributions during model training. Ctrl-A eliminates the need for initialization of individual augmentation strengths. Instead, augmentation strength distributions are dynamically, and individually, adapted during training based on a control-loop architecture and what we define as relative operation response curves. Using an operation-dependent update procedure provides Ctrl-A with the potential to suppress augmentation styles that negatively impact model performance, alleviating the need for manually engineering augmentation policies for new image-vision tasks. Experiments on the CIFAR-10, CIFAR-100, and SVHN-core benchmark datasets using the common WideResNet-28-10 architecture demonstrate that Ctrl-A is highly competitive with existing state-of-the-art data augmentation strategies.", "date": "2026-03-23", "categories": []}
{"id": "2511.09833", "title": "ACT as Human: Multimodal Large Language Model Data Annotation with Critical Thinking", "abstract": "Supervised learning relies on high-quality labeled data, but obtaining such data through human annotation is both expensive and time-consuming. Recent work explores using large language models (LLMs) for annotation, but LLM-generated labels still fall short of human-level quality. To address this problem, we propose the Annotation with Critical Thinking (ACT) data pipeline, where LLMs serve not only as annotators but also as judges to critically identify potential errors. Human effort is then directed towards reviewing only the most \"suspicious\" cases, significantly improving the human annotation efficiency. Our major contributions are as follows: (1) ACT is applicable to a wide range of domains, including natural language processing (NLP), computer vision (CV), and multimodal understanding, by leveraging multimodal-LLMs (MLLMs). (2) Through empirical studies, we derive 7 insights on how to enhance annotation quality while efficiently reducing the human cost, and then translate these findings into user-friendly guidelines. (3) We theoretically analyze how to modify the loss function so that models trained on ACT data achieve similar performance to those trained on fully human-annotated data. Our experiments show that the performance gap can be reduced to less than 2% on most benchmark datasets while saving up to 90% of human costs.", "date": "2025-11-13", "categories": []}
{"id": "2603.18534", "title": "Data-efficient pre-training by scaling synthetic megadocs", "abstract": "Synthetic data augmentation has emerged as a promising solution when pre-training is constrained by data rather than compute. We study how to design synthetic data algorithms that achieve better loss scaling: not only lowering loss at finite compute but especially as compute approaches infinity. We first show that pre-training on web data mixed with synthetically generated rephrases improves i.i.d. validation loss on the web data, despite the synthetic data coming from an entirely different distribution. With optimal mixing and epoching, loss and benchmark accuracy improve without overfitting as the number of synthetic generations grows, plateauing near $1.48\\times$ data efficiency at 32 rephrases per document. We find even better loss scaling under a new perspective: synthetic generations from the same document can form a single substantially longer megadocument instead of many short documents. We show two ways to construct megadocs: stitching synthetic rephrases from the same web document or stretching a document by inserting rationales. Both methods improve i.i.d. loss, downstream benchmarks, and especially long-context loss relative to simple rephrasing, increasing data efficiency from $1.48\\times$ to $1.80\\times$ at $32$ generations per document. Importantly, the improvement of megadocs over simple rephrasing widens as more synthetic data is generated. Our results show how to design synthetic data algorithms that benefit more from increasing compute when data-constrained.", "date": "2026-03-19", "categories": []}
{"id": "2509.26219", "title": "Parameterizing Dataset Distillation via Gaussian Splatting", "abstract": "Dataset distillation aims to compress training data while preserving training-aware knowledge, alleviating the reliance on large-scale datasets in modern model training. Dataset parameterization provides a more efficient storage structure for dataset distillation, reducing redundancy and accommodating richer information. However, existing methods either rely on complex auxiliary modules or fail to balance representational capacity and efficiency. In this paper, we propose GSDD, a simple, novel, and effective dataset parameterization technique for Dataset Distillation based on Gaussian Splatting. We adapt CUDA-based splatting operators for parallel training in batch, enabling high-quality rendering with minimal computational and memory overhead. Gaussian primitives can effectively capture meaningful training features, allowing a sparse yet expressive representation of individual images. Leveraging both high representational capacity and efficiency, GSDD substantially increases the diversity of distilled datasets under a given storage budget, thereby improving distillation performance. Beyond achieving competitive results on multiple standard benchmarks, GSDD also delivers significant performance gains on large-scale datasets such as ImageNet-1K and on video distillation tasks. In addition, we conduct comprehensive benchmarks to evaluate the computational efficiency, memory footprint, and cross-GPU architectural stability of GSDD. Code is available on https://github.com/j-cyoung/GSDatasetDistillation", "date": "2025-09-30", "categories": []}
{"id": "2602.24144", "title": "Fixed Anchors Are Not Enough: Dynamic Retrieval and Persistent Homology for Dataset Distillation", "abstract": "Decoupled dataset distillation (DD) compresses large corpora into a few synthetic images by matching a frozen teacher's statistics. However, current residual-matching pipelines rely on static real patches, creating a fit-complexity gap and a pull-to-anchor effect that reduce intra-class diversity and hurt generalization. To address these issues, we introduce RETA -- a Retrieval and Topology Alignment framework for decoupled DD. First, Dynamic Retrieval Connection (DRC) selects a real patch from a prebuilt pool by minimizing a fit-complexity score in teacher feature space; the chosen patch is injected via a residual connection to tighten feature fit while controlling injected complexity. Second, Persistent Topology Alignment (PTA) regularizes synthesis with persistent homology: we build a mutual k-NN feature graph, compute persistence images of components and loops, and penalize topology discrepancies between real and synthetic sets, mitigating pull-to-anchor effect. Across CIFAR-100, Tiny-ImageNet, ImageNet-1K, and multiple ImageNet subsets, RETA consistently outperforms various baselines under comparable time and memory, especially reaching 64.3% top-1 accuracy on ImageNet-1K with ResNet-18 at 50 images per class, +3.1% over the best prior.", "date": "2026-02-27", "categories": []}
{"id": "2603.13346", "title": "Post Training Quantization for Efficient Dataset Condensation", "abstract": "Dataset Condensation (DC) distills knowledge from large datasets into smaller ones, accelerating training and reducing storage requirements. However, despite notable progress, prior methods have largely overlooked the potential of quantization for further reducing storage costs. In this paper, we take the first step to explore post-training quantization in dataset condensation, demonstrating its effectiveness in reducing storage size while maintaining representation quality without requiring expensive training cost. However, we find that at extremely low bit-widths (e.g., 2-bit), conventional quantization leads to substantial degradation in representation quality, negatively impacting the networks trained on these data. To address this, we propose a novel \\emph{patch-based post-training quantization} approach that ensures localized quantization with minimal loss of information. To reduce the overhead of quantization parameters, especially for small patch sizes, we employ quantization-aware clustering to identify similar patches and subsequently aggregate them for efficient quantization. Furthermore, we introduce a refinement module to align the distribution between original images and their dequantized counterparts, compensating for quantization errors. Our method is a plug-and-play framework that can be applied to synthetic images generated by various DC methods. Extensive experiments across diverse benchmarks including CIFAR-10/100, Tiny ImageNet, and ImageNet subsets demonstrate that our method consistently outperforms prior works under the same storage constraints. Notably, our method nearly \\textbf{doubles the test accuracy} of existing methods at extreme compression regimes (e.g., 26.0\\% $\\rightarrow$ 54.1\\% for DM at IPC=1), while operating directly on 2-bit images without additional distillation.", "date": "2026-03-07", "categories": []}
{"id": "2603.13790", "title": "Greedy Information Projection for LLM Data Selection", "abstract": "We present \\emph{Greedy Information Projection} (\\textsc{GIP}), a principled framework for choosing training examples for large language model fine-tuning. \\textsc{GIP} casts selection as maximizing mutual information between a subset of examples and task-specific query signals, which may originate from LLM quality judgments, metadata, or other sources. The framework involves optimizing a closed-form mutual information objective defined using both data and query embeddings, naturally balancing {\\it quality} and {\\it diversity}. Optimizing this score is equivalent to maximizing the projection of the query embedding matrix onto the span of the selected data, which provides a geometric explanation for the co-emergence of quality and diversity. Building on this view, we employ a fast greedy matching-pursuit procedure with efficient projection-based updates. On instruction-following and mathematical reasoning datasets, \\textsc{GIP} selects small subsets that match full-data fine-tuning while using only a fraction of examples and compute, unifying quality-aware and diversity-aware selection for efficient fine-tuning.", "date": "2026-03-14", "categories": []}
{"id": "2603.13960", "title": "IMS3: Breaking Distributional Aggregation in Diffusion-Based Dataset Distillation", "abstract": "Dataset Distillation aims to synthesize compact datasets that can approximate the training efficacy of large-scale real datasets, offering an efficient solution to the increasing computational demands of modern deep learning. Recently, diffusion-based dataset distillation methods have shown great promise by leveraging the strong generative capacity of diffusion models to produce diverse and structurally consistent samples. However, a fundamental goal misalignment persists: diffusion models are optimized for generative likelihood rather than discriminative utility, resulting in over-concentration in high-density regions and inadequate coverage of boundary samples crucial for classification. To address this issue, we propose two complementary strategies. Inversion-Matching (IM) introduces an inversion-guided fine-tuning process that aligns denoising trajectories with their inversion counterparts, broadening distributional coverage and enhancing diversity. Selective Subgroup Sampling(S^3) is a training-free sampling mechanism that improves inter-class separability by selecting synthetic subsets that are both representative and distinctive. Extensive experiments demonstrate that our approach significantly enhances the discriminative quality and generalization of distilled datasets, achieving state-of-the-art performance among diffusion-based methods.", "date": "2026-03-14", "categories": []}
{"id": "2603.15276", "title": "Dataset Diversity Metrics and Impact on Classification Models", "abstract": "The diversity of training datasets is usually perceived as an important aspect to obtain a robust model. However, the definition of diversity is often not defined or differs across papers, and while some metrics exist, the quantification of this diversity is often overlooked when developing new algorithms. In this work, we study the behaviour of multiple dataset diversity metrics for image, text and metadata using MorphoMNIST, a toy dataset with controlled perturbations, and PadChest, a publicly available chest X-ray dataset. We evaluate whether these metrics correlate with each other but also with the intuition of a clinical expert. We also assess whether they correlate with downstream-task performance and how they impact the training dynamic of the models. We find limited correlations between the AUC and image or metadata reference-free diversity metrics, but higher correlations with the FID and the semantic diversity metrics. Finally, the clinical expert indicates that scanners are the main source of diversity in practice. However, we find that the addition of another scanner to the training set leads to shortcut learning. The code used in this study is available at https://github.com/TheoSourget/dataset_diversity_evaluation", "date": "2026-03-16", "categories": []}
{"id": "2503.01212", "title": "Understanding Dataset Distillation via Spectral Filtering", "abstract": "Dataset distillation (DD) has emerged as a promising approach to compress datasets and speed up model training. However, the underlying connections among various DD methods remain largely unexplored. In this paper, we introduce UniDD, a spectral filtering framework that unifies diverse DD objectives. UniDD interprets each DD objective as a specific filter function that affects the eigenvalues of the feature-feature correlation (FFC) matrix and modulates the frequency components of the feature-label correlation (FLC) matrix. In this way, UniDD reveals that the essence of DD fundamentally lies in matching frequency-specific features. Moreover, according to the filter behaviors, we classify existing methods into low-frequency matching and high-frequency matching, encoding global texture and local details, respectively. However, existing methods rely on fixed filter functions throughout distillation, which cannot capture the low- and high-frequency information simultaneously. To address this limitation, we further propose Curriculum Frequency Matching (CFM), which gradually adjusts the filter parameter to cover both low- and high-frequency information of the FFC and FLC matrices. Extensive experiments on small-scale datasets, such as CIFAR-10/100, and large-scale datasets, including ImageNet-1K, demonstrate the superior performance of CFM over existing baselines and validate the practicality of UniDD.", "date": "2025-03-03", "categories": []}
{"id": "2507.05914", "title": "Accelerating Diffusion Model Training under Minimal Budgets: A Condensation-Based Perspective", "abstract": "Diffusion models have achieved remarkable performance on a wide range of generative tasks, yet training them from scratch is notoriously resource-intensive, typically requiring millions of training images and many GPU days. Motivated by a data-centric view of this bottleneck, we adopt a condensation-based perspective: given a large training set, the goal is to construct a much smaller condensed dataset that still supports training strong diffusion models under minimal data and compute budgets. To operationalize this perspective, we introduce Diffusion Dataset Condensation (D2C), a two-phase framework comprising Select and Attach. In the Select phase, a diffusion difficulty score combined with interval sampling is used to identify a compact, informative training subset from the original data. Building on this subset, the Attach phase further strengthens the conditional signals by augmenting each selected image with rich semantic and visual representations. To our knowledge, D2C is the first framework that systematically investigates dataset condensation for diffusion models, whereas prior condensation methods have mainly targeted discriminative architectures. Extensive experiments across data budgets (0.8%-8% of ImageNet), model architectures, and image resolutions demonstrate that D2C dramatically accelerates diffusion model training while preserving high generative quality. On ImageNet 256x256 with SiT-XL/2, D2C attains an FID of 4.3 in just 40k steps using only 0.8% of the training images, corresponding to about 233x and 100x faster training than vanilla SiT-XL/2 and SiT-XL/2 + REPA, respectively.", "date": "2025-07-08", "categories": []}
{"id": "2603.12478", "title": "Less Data, Faster Convergence: Goal-Driven Data Optimization for Multimodal Instruction Tuning", "abstract": "Multimodal instruction tuning is often compute-inefficient because training budgets are spread across large mixed image-video pools whose utility is highly uneven. We present Goal-Driven Data Optimization (GDO), a framework that computes six sample descriptors for each candidate and constructs optimized 1$\\times$ training subsets for different goals. Under a fixed one-epoch Qwen3-VL-8B-Instruct training and evaluation recipe on 8 H20 GPUs, GDO uses far fewer training samples than the Uni-10x baseline while converging faster and achieving higher accuracy. Relative to the fixed 512k-sample Uni-10x baseline, GDO reaches the Uni-10x reference after 35.4k samples on MVBench, 26.6k on VideoMME, 27.3k on MLVU, and 34.7k on LVBench, while improving Accuracy by +1.38, +1.67, +3.08, and +0.84 percentage points, respectively. The gains are largest on MVBench and MLVU, while LVBench improves more modestly, consistent with its ultra-long-video setting and the mismatch between that benchmark and the short-video/image-dominant training pool. Across MinLoss, Diverse, Temp, and Temp+, stronger temporal emphasis yields steadily better long-video understanding behavior. Overall, GDO provides a goal-driven data optimization framework that enables faster convergence with fewer training samples under a fixed training protocol. Code is available at https://github.com/rujiewu/GDO.", "date": "2026-03-12", "categories": []}
{"id": "2603.09987", "title": "Evolving Demonstration Optimization for Chain-of-Thought Feature Transformation", "abstract": "Feature Transformation (FT) is a core data-centric AI task that improves feature space quality to advance downstream predictive performance. However, discovering effective transformations remains challenging due to the large space of feature-operator combinations. Existing solutions rely on discrete search or latent generation, but they are frequently limited by sample inefficiency, invalid candidates, and redundant generations with limited coverage. Large Language Models (LLMs) offer strong priors for producing valid transformations, but current LLM-based FT methods typically rely on static demonstrations, resulting in limited diversity, redundant outputs, and weak alignment with downstream objectives. We propose a framework that optimizes context data for LLM-driven FT by evolving trajectory-level experiences in a closed loop. Starting from high-performing feature transportation sequences explored by reinforcement learning, we construct and continuously update an experience library of downstream task-verified transformation trajectories, and use a diversity-aware selector to form contexts along with a chain-of-thought and guide transformed feature generation toward higher performance. Experiments on diverse tabular benchmarks show that our method outperforms classical and LLM-based baselines and is more stable than one-shot generation. The framework generalizes across API-based and open-source LLMs and remains robust across downstream evaluators.", "date": "2026-02-13", "categories": []}
{"id": "2603.10410", "title": "Effective Dataset Distillation for Spatio-Temporal Forecasting with Bi-dimensional Compression", "abstract": "Spatio-temporal time series are widely used in real-world applications, including traffic prediction and weather forecasting. They are sequences of observations over extensive periods and multiple locations, naturally represented as multidimensional data. Forecasting is a central task in spatio-temporal analysis, and numerous deep learning methods have been developed to address it. However, as dataset sizes and model complexities continue to grow in practice, training deep learning models has become increasingly time- and resource-intensive. A promising solution to this challenge is dataset distillation, which synthesizes compact datasets that can effectively replace the original data for model training. Although successful in various domains, including time series analysis, existing dataset distillation methods compress only one dimension, making them less suitable for spatio-temporal datasets, where both spatial and temporal dimensions jointly contribute to the large data volume. To address this limitation, we propose STemDist, the first dataset distillation method specialized for spatio-temporal time series forecasting. A key idea of our solution is to compress both temporal and spatial dimensions in a balanced manner, reducing training time and memory. We further reduce the distillation cost by performing distillation at the cluster level rather than the individual location level, and we complement this coarse-grained approach with a subset-based granular distillation technique that enhances forecasting performance. On five real-world datasets, we show empirically that, compared to both general and time-series dataset distillation methods, datasets distilled by our STemDist method enable model training (1) faster (up to 6X) (2) more memory-efficient (up to 8X), and (3) more effective (with up to 12% lower prediction error).", "date": "2026-03-11", "categories": []}
{"id": "2410.01611", "title": "DRUPI: Dataset Reduction Using Privileged Information", "abstract": "Dataset Condensation (DC) seeks to select or distill samples from large datasets into smaller subsets while preserving performance on target tasks. Existing methods primarily focus on pruning or synthesizing data in the same format as the original dataset, typically being the input data and corresponding labels. However, in DC settings, we find it is possible to synthesize more information beyond the data-label pair as an additional learning target to facilitate model training. In this paper, we introduce Dataset Condensation using Privileged Information (DCPI), which enriches DC by synthesizing privileged information alongside the reduced dataset. This privileged information can take the form of feature labels or attention labels, providing auxiliary supervision to improve model learning. Our findings reveal that effective feature labels must balance between being overly discriminative and excessively diverse, with a moderate level proves optimal for improving the reduced dataset's efficacy. Extensive experiments on ImageNet-1K, CIFAR-10/100 and Tiny ImageNet demonstrate that DCPI integrates seamlessly with existing dataset condensation methods, offering significant performance gains.", "date": "2024-10-02", "categories": []}
{"id": "2502.06574", "title": "On the Impact of the Utility in Semivalue-based Data Valuation", "abstract": "Semivalue-based data valuation uses cooperative-game theory intuitions to assign each data point a value reflecting its contribution to a downstream task. Still, those values depend on the practitioner's choice of utility, raising the question: How robust is semivalue-based data valuation to changes in the utility? This issue is critical when the utility is set as a trade-off between several criteria and when practitioners must select among multiple equally valid utilities. We address this by introducing the notion of a dataset's spatial signature: given a semivalue, we embed each data point into a lower-dimensional space in which any utility becomes a linear functional, making the data valuation framework amenable to a simpler geometric picture. Building on this, we propose a practical methodology centered on an explicit robustness metric that informs practitioners whether and by how much their data valuation results will shift as the utility changes. We validate this approach across diverse datasets and semivalues, demonstrating strong agreement with rank-correlation analyses and offering analytical insight into how choosing a semivalue can amplify or diminish robustness.", "date": "2025-02-10", "categories": []}
{"id": "2506.01290", "title": "Rating Quality of Diverse Time Series Data by Meta-learning from LLM Judgment", "abstract": "High-quality time series (TS) data are essential for ensuring TS model performance, rendering research on rating TS data quality indispensable. Existing methods have shown promising rating accuracy within individual domains, primarily by extending data quality rating techniques such as influence functions and Shapley values to account for temporal characteristics. However, they neglect the fact that real-world TS data can span vastly different domains and exhibit distinct properties, hampering the accurate and efficient rating of diverse TS data. In this paper, we propose TSRating, a novel and unified framework for rating the quality of time series data crawled from diverse domains. TSRating leverages LLMs' inherent ample knowledge, acquired during their extensive pretraining, to comprehend and discern quality differences in diverse TS data. We verify this by devising a series of prompts to elicit quality comparisons from LLMs for pairs of TS samples. We then fit a dedicated rating model, termed TSRater, to convert the LLMs' judgments into efficient quality predictions by inferring future TS samples through TSRater's inference. To ensure cross-domain adaptability, we develop a meta-learning scheme to train TSRater on quality comparisons collected from nine distinct domains. To improve training efficiency, we employ signSGD for inner-loop updates, thus circumventing the demanding computation of hypergradients. Extensive experimental results on eleven benchmark datasets across three time series tasks, each using both conventional TS models and TS foundation models, demonstrate that TSRating outperforms baselines in terms of estimation accuracy, efficiency, and domain adaptability.", "date": "2025-06-02", "categories": []}
{"id": "2511.11648", "title": "Lightweight Time Series Data Valuation on Time Series Foundation Models via In-Context Finetuning", "abstract": "Time series foundation models (TSFMs) have demonstrated increasing capabilities due to their extensive pretraining on large volumes of diverse time series data. Consequently, the quality of time series data is crucial to TSFM performance, rendering an accurate and efficient data valuation of time series for TSFMs indispensable. However, traditional data valuation methods, such as influence functions, face severe computational bottlenecks due to their poor scalability with growing TSFM model sizes and often fail to preserve temporal dependencies. In this paper, we propose LTSV, a Lightweight Time Series Valuation on TSFMS via in-context finetuning. Grounded in the theoretical evidence that in-context finetuning approximates the influence function, LTSV estimates a sample's contribution by measuring the change in context loss after in-context finetuning, leveraging the strong generalization capabilities of TSFMs to produce robust and transferable data valuations. To capture temporal dependencies, we introduce temporal block aggregation, which integrates per-block influence scores across overlapping time windows. Experiments across multiple time series datasets and models demonstrate that LTSV consistently provides reliable and strong valuation performance, while maintaining manageable computational requirements. Our results suggest that in-context finetuning on time series foundation models provides a practical and effective bridge between data attribution and model generalization in time series learning.", "date": "2025-11-10", "categories": []}
{"id": "2603.09356", "title": "Democratising Clinical AI through Dataset Condensation for Classical Clinical Models", "abstract": "Dataset condensation (DC) learns a compact synthetic dataset that enables models to match the performance of full-data training, prioritising utility over distributional fidelity. While typically explored for computational efficiency, DC also holds promise for healthcare data democratisation, especially when paired with differential privacy, allowing synthetic data to serve as a safe alternative to real records. However, existing DC methods rely on differentiable neural networks, limiting their compatibility with widely used clinical models such as decision trees and Cox regression. We address this gap using a differentially private, zero-order optimisation framework that extends DC to non-differentiable models using only function evaluations. Empirical results across six datasets, including both classification and survival tasks, show that the proposed method produces condensed datasets that preserve model utility while providing effective differential privacy guarantees - enabling model-agnostic data sharing for clinical prediction tasks without exposing sensitive patient information.", "date": "2026-03-10", "categories": []}
{"id": "2602.00329", "title": "In-Run Data Shapley for Adam Optimizer", "abstract": "Reliable data attribution is essential for mitigating bias and reducing computational waste in modern machine learning, with the Shapley value serving as the theoretical gold standard. While recent \"In-Run\" methods bypass the prohibitive cost of retraining by estimating contributions dynamically, they heavily rely on the linear structure of Stochastic Gradient Descent (SGD) and fail to capture the complex dynamics of adaptive optimizers like Adam. In this work, we demonstrate that data attribution is inherently optimizer-dependent: we show that SGD-based proxies diverge significantly from true contributions under Adam (Pearson $R \\approx 0.11$), rendering them ineffective for modern training pipelines. To bridge this gap, we propose Adam-Aware In-Run Data Shapley. We derive a closed-form approximation that restores additivity by redefining utility under a fixed-state assumption and enable scalable computation via a novel Linearized Ghost Approximation. This technique linearizes the variance-dependent scaling term, allowing us to compute pairwise gradient dot-products without materializing per-sample gradients. Extensive experiments show that our method achieves near-perfect fidelity to ground-truth marginal contributions ($R > 0.99$) while retaining $\\sim$95\\% of standard training throughput. Furthermore, our Adam-aware attribution significantly outperforms SGD-based baselines in data attribution downstream tasks.", "date": "2026-01-30", "categories": []}
{"id": "2603.06932", "title": "HIERAMP: Coarse-to-Fine Autoregressive Amplification for Generative Dataset Distillation", "abstract": "Dataset distillation often prioritizes global semantic proximity when creating small surrogate datasets for original large-scale ones. However, object semantics are inherently hierarchical. For example, the position and appearance of a bird's eyes are constrained by the outline of its head. Global proximity alone fails to capture how object-relevant structures at different levels support recognition. In this work, we investigate the contributions of hierarchical semantics to effective distilled data. We leverage the vision autoregressive (VAR) model whose coarse-to-fine generation mirrors this hierarchy and propose HIERAMP to amplify semantics at different levels. At each VAR scale, we inject class tokens that dynamically identify salient regions and use their induced maps to guide amplification at that scale. This adds only marginal inference cost while steering synthesis toward discriminative parts and structures. Empirically, we find that semantic amplification leads to more diverse token choices in constructing coarse-scale object layouts. Conversely, at fine scales, the amplification concentrates token usage, increasing focus on object-related details. Across popular dataset distillation benchmarks, HIERAMP consistently improves validation performance without explicitly optimizing global proximity, demonstrating the importance of semantic amplification for effective dataset distillation.", "date": "2026-03-06", "categories": []}
{"id": "2603.07476", "title": "EVLF: Early Vision-Language Fusion for Generative Dataset Distillation", "abstract": "Dataset distillation (DD) aims to synthesize compact training sets that enable models to achieve high accuracy with significantly fewer samples. Recent diffusion-based DD methods commonly introduce semantic guidance through late-stage cross-attention, where textual prompts tend to dominate the generative process. Although this strategy enforces label relevance, it diminishes the contribution of visual latents, resulting in over-corrected samples that mirror prompt patterns rather than reflecting intrinsic visual features. To solve this problem, we introduce an Early Vision-Language Fusion (EVLF) method that aligns textual and visual embeddings at the transition between the encoder and the generative backbone. By incorporating a lightweight cross-attention module at this transition, the early representations simultaneously encode local textures and global semantic directions across the denoising process. Importantly, EVLF is plug-and-play and can be easily integrated into any diffusion-based dataset distillation pipeline with an encoder. It works across different denoiser architectures and sampling schedules without any task-specific modifications. Extensive experiments demonstrate that EVLF generates semantically faithful and visually coherent synthetic data, yielding consistent improvements in downstream classification accuracy across varied settings. Source code is available at https://github.com/wenqi-cai297/earlyfusion-for-dd/.", "date": "2026-03-08", "categories": []}
{"id": "2603.08364", "title": "Diffusion-Based Data Augmentation for Image Recognition: A Systematic Analysis and Evaluation", "abstract": "Diffusion-based data augmentation (DiffDA) has emerged as a promising approach to improving classification performance under data scarcity. However, existing works vary significantly in task configurations, model choices, and experimental pipelines, making it difficult to fairly compare methods or assess their effectiveness across different scenarios. Moreover, there remains a lack of systematic understanding of the full DiffDA workflow. In this work, we introduce UniDiffDA, a unified analytical framework that decomposes DiffDA methods into three core components: model fine-tuning, sample generation, and sample utilization. This perspective enables us to identify key differences among existing methods and clarify the overall design space. Building on this framework, we develop a comprehensive and fair evaluation protocol, benchmarking representative DiffDA methods across diverse low-data classification tasks. Extensive experiments reveal the relative strengths and limitations of different DiffDA strategies and offer practical insights into method design and deployment. All methods are re-implemented within a unified codebase, with full release of code and configurations to ensure reproducibility and to facilitate future research.", "date": "2026-03-09", "categories": []}
{"id": "2603.05729", "title": "Unlocking ImageNet's Multi-Object Nature: Automated Large-Scale Multilabel Annotation", "abstract": "The original ImageNet benchmark enforces a single-label assumption, despite many images depicting multiple objects. This leads to label noise and limits the richness of the learning signal. Multi-label annotations more accurately reflect real-world visual scenes, where multiple objects co-occur and contribute to semantic understanding, enabling models to learn richer and more robust representations. While prior efforts (e.g., ReaL, ImageNetv2) have improved the validation set, there has not yet been a scalable, high-quality multi-label annotation for the training set. To this end, we present an automated pipeline to convert the ImageNet training set into a multi-label dataset, without human annotations. Using self-supervised Vision Transformers, we perform unsupervised object discovery, select regions aligned with original labels to train a lightweight classifier, and apply it to all regions to generate coherent multi-label annotations across the dataset. Our labels show strong alignment with human judgment in qualitative evaluations and consistently improve performance across quantitative benchmarks. Compared to traditional single-label scheme, models trained with our multi-label supervision achieve consistently better in-domain accuracy across architectures (up to +2.0 top-1 accuracy on ReaL and +1.5 on ImageNet-V2) and exhibit stronger transferability to downstream tasks (up to +4.2 and +2.3 mAP on COCO and VOC, respectively). These results underscore the importance of accurate multi-label annotations for enhancing both classification performance and representation learning. Project code and the generated multi-label annotations are available at https://github.com/jchen175/MultiLabel-ImageNet.", "date": "2026-03-05", "categories": []}
{"id": "2603.04981", "title": "Rethinking Representativeness and Diversity in Dynamic Data Selection", "abstract": "Dynamic data selection accelerates training by sampling a changing subset of the dataset while preserving accuracy. We rethink two core notions underlying sample evaluation: representativeness and diversity. Instead of local geometric centrality, we define representativeness as coverage of dataset-level common or high-frequency feature factors. Instead of within-subset dispersion, we define diversity at the process level, requiring the selection trajectory to gradually include complementary rare factors over training. Based on this view, we propose a dynamic selection framework with three components. First, we score representativeness in a plug-in feature space to prioritize samples covering frequent factors. We instantiate this with a sparse autoencoder trained on the target dataset, using sparse unit activations to summarize both individual samples and dataset-wide factor statistics. Second, we realize process-level diversity by combining rare-factor sampling with a Usage-Frequency Penalty that promotes sample rotation, provably discourages monopoly, and reduces gradient bias. Third, we couple the two-dimensional scoring with a smooth scheduler that transitions selection from core-pattern consolidation to rare-factor exploration, without extra gradients, influence estimates, or second-order computations on the training model. Extensive experiments on five benchmarks across vision and text tasks demonstrate improved accuracy-efficiency trade-offs across models. Our method matches or exceeds full-data accuracy with over 2x training acceleration. Code will be released.", "date": "2026-03-05", "categories": []}
{"id": "2505.21574", "title": "Do We Need All the Synthetic Data? Targeted Image Augmentation via Diffusion Models", "abstract": "Synthetically augmenting training datasets with diffusion models has become an effective strategy for improving the generalization of image classifiers. However, existing approaches typically increase dataset size by 10-30x and struggle to ensure generation diversity, leading to substantial computational overhead. In this work, we introduce TADA (TArgeted Diffusion Augmentation), a principled framework that selectively augments examples that are not learned early in training using faithful synthetic images that preserve semantic features while varying noise. We show that augmenting only this targeted subset consistently outperforms augmenting the entire dataset. Through theoretical analysis on a two-layer CNN, we prove that TADA improves generalization by promoting homogeneity in feature learning speed without amplifying noise. Extensive experiments demonstrate that by augmenting only 30-40% of the training data, TADA improves generalization by up to 2.8% across diverse architectures including ResNet, ViT, ConvNeXt, and Swin Transformer on CIFAR-10/100, TinyImageNet, and ImageNet, using optimizers such as SGD and SAM. Notably, TADA combined with SGD outperforms the state-of-the-art optimizer SAM on CIFAR-100 and TinyImageNet. Furthermore, TADA shows promising improvements on object detection benchmarks, demonstrating its applicability beyond image classification. Our code is available at https://github.com/BigML-CS-UCLA/TADA.", "date": "2025-05-27", "categories": []}
{"id": "2603.03672", "title": "Local Shapley: Model-Induced Locality and Optimal Reuse in Data Valuation", "abstract": "The Shapley value provides a principled foundation for data valuation, but exact computation is #P-hard due to the exponential coalition space. Existing accelerations remain global and ignore a structural property of modern predictors: for a given test instance, only a small subset of training points influences the prediction. We formalize this model-induced locality through support sets defined by the model's computational pathway (e.g., neighbors in KNN, leaves in trees, receptive fields in GNNs), showing that Shapley computation can be projected onto these supports without loss when locality is exact. This reframes Shapley evaluation as a structured data processing problem over overlapping support-induced subset families rather than exhaustive coalition enumeration. We prove that the intrinsic complexity of Local Shapley is governed by the number of distinct influential subsets, establishing an information-theoretic lower bound on retraining operations. Guided by this result, we propose LSMR (Local Shapley via Model Reuse), an optimal subset-centric algorithm that trains each influential subset exactly once via support mapping and pivot scheduling. For larger supports, we develop LSMR-A, a reuse-aware Monte Carlo estimator that remains unbiased with exponential concentration, with runtime determined by the number of distinct sampled subsets rather than total draws. Experiments across multiple model families demonstrate substantial retraining reductions and speedups while preserving high valuation fidelity.", "date": "2026-03-04", "categories": []}
{"id": "2603.03760", "title": "Harmonic Dataset Distillation for Time Series Forecasting", "abstract": "Time Series forecasting (TSF) in the modern era faces significant computational and storage cost challenges due to the massive scale of real-world data. Dataset Distillation (DD), a paradigm that synthesizes a small, compact dataset to achieve training performance comparable to that of the original dataset, has emerged as a promising solution. However, conventional DD methods are not tailored for time series and suffer from architectural overfitting and limited scalability. To address these issues, we propose Harmonic Dataset Distillation for Time Series Forecasting (HDT). HDT decomposes the time series into its sinusoidal basis through the FFT and aligns the core periodic structure by Harmonic Matching. Since this process operates in the frequency domain, all updates during distillation are applied globally without disrupting temporal dependencies of time series. Extensive experiments demonstrate that HDT achieves strong cross-architecture generalization and scalability, validating its practicality for large-scale, real-world applications.", "date": "2026-03-04", "categories": []}
{"id": "2603.03808", "title": "Vector-Quantized Soft Label Compression for Dataset Distillation", "abstract": "Dataset distillation is an emerging technique for reducing the computational and storage costs of training machine learning models by synthesizing a small, informative subset of data that captures the essential characteristics of a much larger dataset. Recent methods pair synthetic samples and their augmentations with soft labels from a teacher model, enabling student models to generalize effectively despite the small size of the distilled dataset. While soft labels are critical for effective distillation, the storage and communication overhead they incur, especially when accounting for augmentations, is often overlooked. In practice, each distilled sample is associated with multiple soft labels, making them the dominant contributor to storage costs, particularly in large-class settings such as ImageNet-1K. In this paper, we present a rigorous analysis of bit requirements across dataset distillation frameworks, quantifying the storage demands of both distilled samples and their soft labels. To address the overhead, we introduce a vector-quantized autoencoder (VQAE) for compressing soft labels, achieving substantial compression while preserving the effectiveness of the distilled data. We validate our method on both vision and language distillation benchmarks. On ImageNet-1K, our proposed VQAE achieves 30--40x additional compression over RDED, LPLD, SRE2L, and CDA baselines while retaining over $90\\%$ of their original performance.", "date": "2026-03-04", "categories": []}
{"id": "2505.21813", "title": "Optimizing Data Augmentation through Bayesian Model Selection", "abstract": "Data Augmentation (DA) has become an essential tool to improve robustness and generalization of modern machine learning. However, when deciding on DA strategies it is critical to choose parameters carefully, and this can be a daunting task which is traditionally left to trial-and-error or expensive optimization based on validation performance. In this paper, we counter these limitations by proposing a novel framework for optimizing DA. In particular, we take a probabilistic view of DA, which leads to the interpretation of augmentation parameters as model (hyper)-parameters, and the optimization of the marginal likelihood with respect to these parameters as a Bayesian model selection problem. Due to its intractability, we derive a tractable ELBO, which allows us to optimize augmentation parameters jointly with model parameters. We provide extensive theoretical results on variational approximation quality, generalization guarantees, invariance properties, and connections to empirical Bayes. Through experiments on computer vision and NLP tasks, we show that our approach improves calibration and yields robust performance over fixed or no augmentation. Our work provides a rigorous foundation for optimizing DA through Bayesian principles with significant potential for robust machine learning.", "date": "2025-05-27", "categories": []}
{"id": "2509.18577", "title": "Prior-based Noisy Text Data Filtering: Fast and Strong Alternative For Perplexity", "abstract": "As large language models (LLMs) are pretrained on massive web corpora, careful selection of data becomes essential to ensure effective and efficient learning. While perplexity (PPL)-based filtering has shown strong performance, it suffers from drawbacks: substantial time costs and inherent unreliability of the model when handling noisy or out-of-distribution samples. In this work, we propose a simple yet powerful alternative: a prior-based data filtering method that estimates token priors using corpus-level term frequency statistics, inspired by linguistic insights on word roles and lexical density. Our approach filters documents based on the mean and standard deviation of token priors, serving as a fast proxy to PPL while requiring no model inference. Despite its simplicity, the prior-based filter achieves the highest average performance across 20 downstream benchmarks, while reducing time cost by over 1000x compared to PPL-based filtering. We further demonstrate its applicability to symbolic languages such as code and math, and its dynamic adaptability to multilingual corpora without supervision", "date": "2025-09-23", "categories": []}
{"id": "2603.02411", "title": "From Fewer Samples to Fewer Bits: Reframing Dataset Distillation as Joint Optimization of Precision and Compactness", "abstract": "Dataset Distillation (DD) compresses large datasets into compact synthetic ones that maintain training performance. However, current methods mainly target sample reduction, with limited consideration of data precision and its impact on efficiency. We propose Quantization-aware Dataset Distillation (QuADD), a unified framework that jointly optimizes dataset compactness and precision under fixed bit budgets. QuADD integrates a differentiable quantization module within the distillation loop, enabling end-to-end co-optimization of synthetic samples and quantization parameters. Guided by the rate-distortion perspective, we empirically analyze how bit allocation between sample count and precision influences learning performance. Our framework supports both uniform and adaptive non-uniform quantization, where the latter learns quantization levels from data to represent information-dense regions better. Experiments on image classification and 3GPP beam management tasks show that QuADD surpasses existing DD and post-quantized baselines in accuracy per bit, establishing a new standard for information-efficient dataset distillation.", "date": "2026-03-02", "categories": []}
{"id": "2505.02881", "title": "Rewriting Pre-Training Data Boosts LLM Performance in Math and Code", "abstract": "The performance of large language models (LLMs) in program synthesis and mathematical reasoning is fundamentally limited by the quality of their pre-training corpora. We introduce two openly licensed pre-training datasets, released under the Llama 3.3 Community License, that significantly enhance LLM performance by systematically rewriting public data. SwallowCode ($\\approx$16.1 billion tokens) refines Python snippets from The-Stack-v2 through a novel four-stage pipeline: syntax validation, pylint-based style filtering, and a two-stage LLM rewriting process that enforces style conformity and transforms snippets into self-contained, algorithmically efficient examples. Unlike prior methods that rely on exclusionary filtering or limited transformations, our transform-and-retain approach refines low-quality code, maximizing data utility. SwallowMath ($\\approx$2.3 billion tokens) enhances Finemath-4+ by removing boilerplate, restoring context, and reformatting solutions into concise, step-by-step explanations. Within a fixed 50 billion token training budget, continual pre-training of Llama-3.1-8B with SwallowCode boosts pass@1 by +17.0 on HumanEval and +16.1 on HumanEval+ compared to Stack-Edu, surpassing the baseline model's code generation capabilities. Similarly, substituting SwallowMath yields +12.4 accuracy on GSM8K and +7.6 on MATH. Ablation studies confirm that each pipeline stage contributes incrementally, with rewriting yielding the largest gains. By releasing datasets, prompts, checkpoints, and pipeline code, we ensure reproducibility and provide a transferable transform-and-retain methodology that can be adapted to other base models and LLM rewriting setups.", "date": "2025-05-05", "categories": []}
{"id": "2602.20650", "title": "Dataset Color Quantization: A Training-Oriented Framework for Dataset-Level Compression", "abstract": "Large-scale image datasets are fundamental to deep learning, but their high storage demands pose challenges for deployment in resource-constrained environments. While existing approaches reduce dataset size by discarding samples, they often ignore the significant redundancy within each image -- particularly in the color space. To address this, we propose Dataset Color Quantization (DCQ), a unified framework that compresses visual datasets by reducing color-space redundancy while preserving information crucial for model training. DCQ achieves this by enforcing consistent palette representations across similar images, selectively retaining semantically important colors guided by model perception, and maintaining structural details necessary for effective feature learning. Extensive experiments across CIFAR-10, CIFAR-100, Tiny-ImageNet, and ImageNet-1K show that DCQ significantly improves training performance under aggressive compression, offering a scalable and robust solution for dataset-level storage reduction.", "date": "2026-02-24", "categories": []}
{"id": "2603.00031", "title": "GRIP: Geometric Refinement and Adaptive Information Potential for Data Efficiency", "abstract": "The performance of Large Language Models (LLMs) is increasingly governed by data efficiency rather than raw scaling volume. However, existing selection methods often decouple global distribution balancing from local instance selection, compromising the hierarchical integrity of the training set. We introduce \\textbf{GRIP} (Geometric Refinement and Adaptive Information Potential), a framework that unifies these dimensions by modeling the corpus as an information-dense geometric space. GRIP employs a \\textbf{Rapid Adaptation Probe (RAP)} to quantify the information potential of semantic clusters, dynamically re-allocating the sampling budget to regions with the highest representation deficits. Subsequently, we perform Intra-Cluster Selection using a \\textbf{length-rectified geometric prior} to counteract embedding density artifacts and preserve long-tail logical sequences. Extensive evaluations on Mixture-of-Experts (MoE) models up to 300B tokens demonstrate that GRIP consistently outperforms state-of-the-art baselines, \\textbf{surpassing the performance of models trained on $3\\times$ larger uncurated datasets}. Our work establishes a robust geometric foundation for adaptive data curation in large-scale pre-training.", "date": "2026-02-04", "categories": []}
{"id": "2603.00604", "title": "Data-Centric Benchmark for Label Noise Estimation and Ranking in Remote Sensing Image Segmentation", "abstract": "High-quality pixel-level annotations are essential for the semantic segmentation of remote sensing imagery. However, such labels are expensive to obtain and often affected by noise due to the labor-intensive and time-consuming nature of pixel-wise annotation, which makes it challenging for human annotators to label every pixel accurately. Annotation errors can significantly degrade the performance and robustness of modern segmentation models, motivating the need for reliable mechanisms to identify and quantify noisy training samples. This paper introduces a novel Data-Centric benchmark, together with a novel, publicly available dataset and two techniques for identifying, quantifying, and ranking training samples according to their level of label noise in remote sensing semantic segmentation. Such proposed methods leverage complementary strategies based on model uncertainty, prediction consistency, and representation analysis, and consistently outperform established baselines across a range of experimental settings. The outcomes of this work are publicly available at https://github.com/keillernogueira/label_noise_segmentation.", "date": "2026-02-28", "categories": []}
{"id": "2603.00651", "title": "Exploring 3D Dataset Pruning", "abstract": "Dataset pruning has been widely studied for 2D images to remove redundancy and accelerate training, while particular pruning methods for 3D data remain largely unexplored. In this work, we study dataset pruning for 3D data, where its observed common long-tail class distribution nature make optimization under conventional evaluation metrics Overall Accuracy (OA) and Mean Accuracy (mAcc) inherently conflicting, and further make pruning particularly challenging. To address this, we formulate pruning as approximating the full-data expected risk with a weighted subset, which reveals two key errors: coverage error from insufficient representativeness and prior-mismatch bias from inconsistency between subset-induced class weights and target metrics. We propose representation-aware subset selection with per-class retention quotas for long-tail coverage, and prior-invariant teacher supervision using calibrated soft labels and embedding-geometry distillation. The retention quota also serves as a switch to control the OA-mAcc trade-off. Extensive experiments on 3D datasets show that our method can improve both metrics across multiple settings while adapting to different downstream preferences. Our code is available at https://github.com/XiaohanZhao123/3D-Dataset-Pruning.", "date": "2026-02-28", "categories": []}
{"id": "2603.01053", "title": "Turning Black Box into White Box: Dataset Distillation Leaks", "abstract": "Dataset distillation compresses a large real dataset into a small synthetic one, enabling models trained on the synthetic data to achieve performance comparable to those trained on the real data. Although synthetic datasets are assumed to be privacy-preserving, we show that existing distillation methods can cause severe privacy leakage because synthetic datasets implicitly encode the weight trajectories of the distilled model, they become over-informative and exploitable by adversaries. To expose this risk, we introduce the Information Revelation Attack (IRA) against state-of-the-art distillation techniques. Experiments show that IRA accurately predicts both the distillation algorithm and model architecture, and can successfully infer membership and recover sensitive samples from the real dataset.", "date": "2026-03-01", "categories": []}
{"id": "2603.01698", "title": "Towards Principled Dataset Distillation: A Spectral Distribution Perspective", "abstract": "Dataset distillation (DD) aims to compress large-scale datasets into compact synthetic counterparts for efficient model training. However, existing DD methods exhibit substantial performance degradation on long-tailed datasets. We identify two fundamental challenges: heuristic design choices for distribution discrepancy measure and uniform treatment of imbalanced classes. To address these limitations, we propose Class-Aware Spectral Distribution Matching (CSDM), which reformulates distribution alignment via the spectrum of a well-behaved kernel function. This technique maps the original samples into frequency space, resulting in the Spectral Distribution Distance (SDD). To mitigate class imbalance, we exploit the unified form of SDD to perform amplitude-phase decomposition, which adaptively prioritizes the realism in tail classes. On CIFAR-10-LT, with 10 images per class, CSDM achieves a 14.0% improvement over state-of-the-art DD methods, with only a 5.7% performance drop when the number of images in tail classes decreases from 500 to 25, demonstrating strong stability on long-tailed data.", "date": "2026-03-02", "categories": []}
{"id": "2602.19756", "title": "Multimodal Dataset Distillation Made Simple by Prototype-Guided Data Synthesis", "abstract": "Recent advances in multimodal learning have achieved remarkable success across diverse vision-language tasks. However, such progress heavily relies on large-scale image-text datasets, making training costly and inefficient. Prior efforts in dataset filtering and pruning attempt to mitigate this issue, but still require relatively large subsets to maintain performance and fail under very small subsets. Dataset distillation offers a promising alternative, yet existing multimodal dataset distillation methods require full-dataset training and joint optimization of image pixels and text features, making them architecture-dependent and limiting cross-architecture generalization. To overcome this, we propose a learning-free dataset distillation framework that eliminates the need for large-scale training and optimization while enhancing generalization across architectures. Our method uses CLIP to extract aligned image-text embeddings, obtains prototypes, and employs an unCLIP decoder to synthesize images, enabling efficient and scalable multimodal dataset distillation. Extensive experiments demonstrate that our approach consistently outperforms optimization-based dataset distillation and subset selection methods, achieving state-of-the-art cross-architecture generalization.", "date": "2026-02-23", "categories": []}
{"id": "2602.23295", "title": "ManifoldGD: Training-Free Hierarchical Manifold Guidance for Diffusion-Based Dataset Distillation", "abstract": "In recent times, large datasets hinder efficient model training while also containing redundant concepts. Dataset distillation aims to synthesize compact datasets that preserve the knowledge of large-scale training sets while drastically reducing storage and computation. Recent advances in diffusion models have enabled training-free distillation by leveraging pre-trained generative priors; however, existing guidance strategies remain limited. Current score-based methods either perform unguided denoising or rely on simple mode-based guidance toward instance prototype centroids (IPC centroids), which often are rudimentary and suboptimal. We propose Manifold-Guided Distillation (ManifoldGD), a training-free diffusion-based framework that integrates manifold consistent guidance at every denoising timestep. Our method employs IPCs computed via a hierarchical, divisive clustering of VAE latent features, yielding a multi-scale coreset of IPCs that captures both coarse semantic modes and fine intra-class variability. Using a local neighborhood of the extracted IPC centroids, we create the latent manifold for each diffusion denoising timestep. At each denoising step, we project the mode-alignment vector onto the local tangent space of the estimated latent manifold, thus constraining the generation trajectory to remain manifold-faithful while preserving semantic consistency. This formulation improves representativeness, diversity, and image fidelity without requiring any model retraining. Empirical results demonstrate consistent gains over existing training-free and training-based baselines in terms of FID, l2 distance among real and synthetic dataset embeddings, and classification accuracy, establishing ManifoldGD as the first geometry-aware training-free data distillation framework.", "date": "2026-02-26", "categories": []}
{"id": "2602.23358", "title": "A Dataset is Worth 1 MB", "abstract": "A dataset server must often distribute the same large payload to many clients, incurring massive communication costs. Since clients frequently operate on diverse hardware and software frameworks, transmitting a pre-trained model is often infeasible; instead, agents require raw data to train their own task-specific models locally. While dataset distillation attempts to compress training signals, current methods struggle to scale to high-resolution data and rarely achieve sufficiently small files. In this paper, we propose Pseudo-Labels as Data (PLADA), a method that completely eliminates pixel transmission. We assume agents are preloaded with a large, generic, unlabeled reference dataset (e.g., ImageNet-1K, ImageNet-21K) and communicate a new task by transmitting only the class labels for specific images. To address the distribution mismatch between the reference and target datasets, we introduce a pruning mechanism that filters the reference dataset to retain only the labels of the most semantically relevant images for the target task. This selection process simultaneously maximizes training efficiency and minimizes transmission payload. Experiments on 10 diverse datasets demonstrate that our approach can transfer task knowledge with a payload of less than 1 MB while retaining high classification accuracy, offering a promising solution for efficient dataset serving.", "date": "2026-02-26", "categories": []}
{"id": "2602.21717", "title": "C$^{2}$TC: A Training-Free Framework for Efficient Tabular Data Condensation", "abstract": "Tabular data is the primary data format in industrial relational databases, underpinning modern data analytics and decision-making. However, the increasing scale of tabular data poses significant computational and storage challenges to learning-based analytical systems. This highlights the need for data-efficient learning, which enables effective model training and generalization using substantially fewer samples. Dataset condensation (DC) has emerged as a promising data-centric paradigm that synthesizes small yet informative datasets to preserve data utility while reducing storage and training costs. However, existing DC methods are computationally intensive due to reliance on complex gradient-based optimization. Moreover, they often overlook key characteristics of tabular data, such as heterogeneous features and class imbalance. To address these limitations, we introduce C$^{2}$TC (Class-Adaptive Clustering for Tabular Condensation), the first training-free tabular dataset condensation framework that jointly optimizes class allocation and feature representation, enabling efficient and scalable condensation. Specifically, we reformulate the dataset condensation objective into a novel class-adaptive cluster allocation problem (CCAP), which eliminates costly training and integrates adaptive label allocation to handle class imbalance. To solve the NP-hard CCAP, we develop HFILS, a heuristic local search that alternates between soft allocation and class-wise clustering to efficiently obtain high-quality solutions. Moreover, a hybrid categorical feature encoding (HCFE) is proposed for semantics-preserving clustering of heterogeneous discrete attributes. Extensive experiments on 10 real-world datasets demonstrate that C$^{2}$TC improves efficiency by at least 2 orders of magnitude over state-of-the-art baselines, while achieving superior downstream performance.", "date": "2026-02-25", "categories": []}
{"id": "2602.22014", "title": "A Diversity Diet for a Healthier Model: A Case Study of French ModernBERT", "abstract": "Diversity has been gaining interest in the NLP community in recent years. At the same time, state-of-the-art transformer models such as ModernBERT use very large pre-training datasets, which are driven by size rather than by diversity. This summons for an investigation of the impact of diversity on the ModernBERT pre-training. We do so in this study, with the express intent of reducing pre-training dataset size, while retaining at least comparable performance. We compare diversity-driven sampling algorithms, so as to pick the best one. We find that diversity-driven sampling allows in some tasks to gain 10 points relative to randomly-sampled pre-training data of commensurate size. We also see that a model pre-trained for 483h on a diversity-driven dataset of 150M tokens can yield a commensurate performance to a model pre-trained for 1,775h on a randomly-driven dataset of 2.4B tokens.", "date": "2026-02-25", "categories": []}
{"id": "2511.09905", "title": "PRISM: Diversifying Dataset Distillation by Decoupling Architectural Priors", "abstract": "Dataset distillation (DD) promises compact yet faithful synthetic data, but existing approaches often inherit the inductive bias of a single teacher model. As dataset size increases, this bias drives generation toward overly smooth, homogeneous samples, reducing intra-class diversity and limiting generalization. We present PRISM (PRIors from diverse Source Models), a framework that disentangles architectural priors during synthesis. PRISM decouples the logit-matching and regularization objectives, supervising them with different teacher architectures: a primary model for logits and a stochastic subset for batch-normalization (BN) alignment. On ImageNet-1K, PRISM consistently and reproducibly outperforms single-teacher methods (e.g., SRe2L) and recent multi-teacher variants (e.g., G-VBSM) at low- and mid-IPC regimes. The generated data also show significantly richer intra-class diversity, as reflected by a notable drop in cosine similarity between features. We further analyze teacher selection strategies (pre- vs. intra-distillation) and introduce a scalable cross-class batch formation scheme for fast parallel synthesis. Code will be released after the review period.", "date": "2025-11-13", "categories": []}
{"id": "2506.14457", "title": "Dataset distillation for memorized data: Soft labels can leak held-out teacher knowledge", "abstract": "Dataset distillation aims to compress training data into fewer examples via a teacher, from which a student can learn effectively. While its success is often attributed to structure in the data, modern neural networks also memorize specific facts, but if and how such memorized information is can transferred in distillation settings remains less understood. In this work, we show that students trained on soft labels from teachers can achieve non-trivial accuracy on held-out memorized data they never directly observed. This effect persists on structured data when the teacher has not generalized.To analyze it in isolation, we consider finite random i.i.d. datasets where generalization is a priori impossible and a successful teacher fit implies pure memorization. Still, students can learn non-trivial information about the held-out data, in some cases up to perfect accuracy. In those settings, enough soft labels are available to recover the teacher functionally - the student matches the teacher's predictions on all possible inputs, including the held-out memorized data. We show that these phenomena strongly depend on the temperature with which the logits are smoothed, but persist across varying network capacities, architectures and dataset compositions.", "date": "2025-06-17", "categories": []}
{"id": "2602.17835", "title": "Influence-Preserving Proxies for Gradient-Based Data Selection in LLM Fine-tuning", "abstract": "Supervised fine-tuning (SFT) relies critically on selecting training data that most benefits a model's downstream performance. Gradient-based data selection methods such as TracIn and Influence Functions leverage influence to identify useful samples, but their computational cost scales poorly, making them impractical for multi-billion-parameter large language models (LLMs). A common alternative is to use off-the-shelf smaller models as proxies, but they remain suboptimal since their learning dynamics are unclear, their sizes cannot be flexibly adjusted, and they cannot be further aligned with the target model in terms of gradient-based influence estimation. To address these challenges, we introduce Iprox, a two-stage framework that derives influence-preserving proxies directly from the target model. It first applies a low-rank compression stage to preserve influence information of the target model, and then an aligning stage to align both model gradients and logits, thereby constructing proxies that flexibly control computational cost while retaining the target model's influence. Experimental results across diverse LLM families and evaluation tasks show that Iprox consistently outperforms off-the-shelf proxies and baseline methods. On Qwen3-4B, a 1.5B proxy constructed with Iprox achieves stronger performance than the larger 1.7B off-the-shelf proxy. Notably, on Llama3.2, Iprox achieves better performance than baselines while reducing computational cost by more than half relative to the full 3B model. These results show that Iprox provides effective influence-preserving proxies, making gradient-based data selection more scalable for LLMs.", "date": "2026-02-19", "categories": []}
{"id": "2602.12414", "title": "propella-1: Multi-Property Document Annotation for LLM Data Curation at Scale", "abstract": "Since FineWeb-Edu, data curation for LLM pretraining has predominantly relied on single scalar quality scores produced by small classifiers. A single score conflates multiple quality dimensions, prevents flexible filtering, and offers no interpretability. We introduce propella-1, a family of small multilingual LLMs (0.6B, 1.7B, 4B parameters) that annotate text documents across 18 properties organized into six categories: core content, classification, quality and value, audience and purpose, safety and compliance, and geographic relevance. The models support 57 languages and produce structured JSON annotations conforming to a predefined schema. Evaluated against a frontier commercial LLM as a reference annotator, the 4B model achieves higher agreement than much larger general-purpose models. We release propella-annotations, a dataset of over three billion document annotations covering major pretraining corpora including data from FineWeb-2, FinePDFs, HPLT 3.0, and Nemotron-CC. Using these annotations, we present a multi-dimensional compositional analysis of widely used pretraining datasets, revealing substantial differences in quality, reasoning depth, and content composition that single-score approaches cannot capture. All model weights and annotations are released under permissive, commercial-use licenses.", "date": "2026-02-12", "categories": []}
{"id": "2602.15277", "title": "Accelerating Large-Scale Dataset Distillation via Exploration-Exploitation Optimization", "abstract": "Dataset distillation compresses the original data into compact synthetic datasets, reducing training time and storage while retaining model performance, enabling deployment under limited resources. Although recent decoupling-based distillation methods enable dataset distillation at large scale, they continue to face an efficiency gap: optimization-based decoupling methods achieve higher accuracy but demand intensive computation, whereas optimization-free decoupling methods are efficient but sacrifice accuracy. To overcome this trade-off, we propose Exploration--Exploitation Distillation (E$^2$D), a simple, practical method that minimizes redundant computation through an efficient pipeline that begins with full-image initialization to preserve semantic integrity and feature diversity. It then uses a two-phase optimization strategy: an exploration phase that performs uniform updates and identifies high-loss regions, and an exploitation phase that focuses updates on these regions to accelerate convergence. We evaluate E$^2$D on large-scale benchmarks, surpassing the state-of-the-art on ImageNet-1K while being $18\\times$ faster, and on ImageNet-21K, our method substantially improves accuracy while remaining $4.3\\times$ faster. These results demonstrate that targeted, redundancy-reducing updates, rather than brute-force optimization, bridge the gap between accuracy and efficiency in large-scale dataset distillation. Code is available at https://github.com/ncsu-dk-lab/E2D.", "date": "2026-02-17", "categories": []}
{"id": "2602.17465", "title": "Entropy-Based Data Selection for Language Models", "abstract": "Modern language models (LMs) increasingly require two critical resources: computational resources and data resources. Data selection techniques can effectively reduce the amount of training data required for fine-tuning LMs. However, their effectiveness is closely related to computational resources, which always require a high compute budget. Owing to the resource limitations in practical fine-tuning scenario, we systematically reveal the relationship between data selection and uncertainty estimation of selected data. Although large language models (LLMs) exhibit exceptional capabilities in language understanding and generation, which provide new ways to alleviate data scarcity, evaluating data usability remains a challenging task. This makes efficient data selection indispensable. To mitigate these issues, we propose Entropy-Based Unsupervised Data Selection (EUDS) framework. Empirical experiments on sentiment analysis (SA), topic classification (Topic-CLS), and question answering (Q&A) tasks validate its effectiveness. EUDS establishes a computationally efficient data-filtering mechanism. Theoretical analysis and experimental results confirm the effectiveness of our approach. EUDS significantly reduces computational costs and improves training time efficiency with less data requirement. This provides an innovative solution for the efficient fine-tuning of LMs in the compute-constrained scenarios.", "date": "2026-02-19", "categories": []}
{"id": "2507.13074", "title": "Label-Consistent Dataset Distillation with Detector-Guided Refinement", "abstract": "Dataset distillation (DD) aims to generate a compact yet informative dataset that achieves performance comparable to the original dataset, thereby reducing demands on storage and computational resources. Although diffusion models have made significant progress in dataset distillation, the generated surrogate datasets often contain samples with label inconsistencies or insufficient structural detail, leading to suboptimal downstream performance. To address these issues, we propose a detector-guided dataset distillation framework that explicitly leverages a pre-trained detector to identify and refine anomalous synthetic samples, thereby ensuring label consistency and improving image quality. Specifically, a detector model trained on the original dataset is employed to identify anomalous images exhibiting label mismatches or low classification confidence. For each defective image, multiple candidates are generated using a pre-trained diffusion model conditioned on the corresponding image prototype and label. The optimal candidate is then selected by jointly considering the detector's confidence score and dissimilarity to existing qualified synthetic samples, thereby ensuring both label accuracy and intra-class diversity. Experimental results demonstrate that our method can synthesize high-quality representative images with richer details, achieving state-of-the-art performance on the validation set.", "date": "2025-07-17", "categories": []}
{"id": "2602.11320", "title": "Efficient Analysis of the Distilled Neural Tangent Kernel", "abstract": "Neural tangent kernel (NTK) methods are computationally limited by the need to evaluate large Jacobians across many data points. Existing approaches reduce this cost primarily through projecting and sketching the Jacobian. We show that NTK computation can also be reduced by compressing the data dimension itself using NTK-tuned dataset distillation. We demonstrate that the neural tangent space spanned by the input data can be induced by dataset distillation, yielding a 20-100$\\times$ reduction in required Jacobian calculations. We further show that per-class NTK matrices have low effective rank that is preserved by this reduction. Building on these insights, we propose the distilled neural tangent kernel (DNTK), which combines NTK-tuned dataset distillation with state-of-the-art projection methods to reduce up NTK computational complexity by up to five orders of magnitude while preserving kernel structure and predictive performance.", "date": "2026-02-11", "categories": []}
{"id": "2501.07575", "title": "Dataset Distillation via Committee Voting", "abstract": "Dataset distillation aims to synthesize a compact yet representative dataset that preserves the essential characteristics of the original data for efficient model training. Existing methods mainly focus on improving data-synthetic alignment or scaling distillation to large datasets. In this work, we propose $\\textbf{C}$ommittee $\\textbf{V}$oting for $\\textbf{D}$ataset $\\textbf{D}$istillation ($\\textbf{CV-DD}$), an orthogonal approach that leverages the collective knowledge of multiple models to produce higher-quality distilled data. We first establish a strong baseline that achieves state-of-the-art performance through modern architectural and optimization choices. By integrating distributions and predictions from multiple models and generating high-quality soft labels, our method captures a broader range of data characteristics, reduces model-specific bias and the impact of distribution shifts, and significantly improves generalization. This voting-based strategy enhances diversity and robustness, alleviates overfitting, and improves post-evaluation performance. Extensive experiments across multiple datasets and IPC settings demonstrate that CV-DD consistently outperforms single- and multi-model distillation methods and generalizes well to non-training-based frameworks and challenging synthetic-to-real transfer tasks. Code is available at: https://github.com/Jiacheng8/CV-DD.", "date": "2025-01-13", "categories": []}
{"id": "2602.13773", "title": "On Representation Redundancy in Large-Scale Instruction Tuning Data Selection", "abstract": "Data quality is a crucial factor in large language models training. While prior work has shown that models trained on smaller, high-quality datasets can outperform those trained on much larger but noisy or low-quality corpora, systematic methods for industrial-scale data selection in instruction tuning remain underexplored. In this work, we study instruction-tuning data selection through the lens of semantic representation similarity and identify a key limitation of state-of-the-art LLM encoders: they produce highly redundant semantic embeddings. To mitigate this redundancy, we propose Compressed Representation Data Selection (CRDS), a novel framework with two variants. CRDS-R applies Rademacher random projection followed by concatenation of transformer hidden-layer representations, while CRDS-W employs whitening-based dimensionality reduction to improve representational quality. Experimental results demonstrate that both variants substantially enhance data quality and consistently outperform state-of-the-art representation-based selection methods. Notably, CRDS-W achieves strong performance using only 3.5% of the data, surpassing the full-data baseline by an average of 0.71% across four datasets. Our code is available at https://github.com/tdano1/CRDS.", "date": "2026-02-14", "categories": []}
{"id": "2506.04755", "title": "Truth in the Few: High-Value Data Selection for Efficient Multi-Modal Reasoning", "abstract": "While multi-modal large language models (MLLMs) have made significant progress in complex reasoning tasks via reinforcement learning, it is commonly believed that extensive training data is necessary for improving multi-modal reasoning ability, inevitably leading to data redundancy and substantial computational costs. However, can smaller high-value datasets match or outperform full corpora for multi-modal reasoning in MLLMs? In this work, we challenge this assumption through a key observation: meaningful multi-modal reasoning is triggered by only a sparse subset of training samples, termed cognitive samples, whereas the majority contribute marginally. Building on this insight, we propose a novel data selection paradigm termed Reasoning Activation Potential (RAP)}, which identifies cognitive samples by estimating each sample's potential to stimulate genuine multi-modal reasoning by two complementary estimators: 1) Causal Discrepancy Estimator (CDE) based on the potential outcome model principle, eliminates samples that overly rely on language priors by comparing outputs between multi-modal and text-only inputs; 2) Attention Confidence Estimator (ACE), which exploits token-level self-attention to discard samples dominated by irrelevant but over-emphasized tokens in intermediate reasoning stages. Moreover, we introduce a Difficulty-aware Replacement Module (DRM) to substitute trivial instances with cognitively challenging ones, thereby ensuring complexity for robust multi-modal reasoning. Experiments on six datasets show that our RAP method consistently achieves superior performance using only 9.3% of the training data, while reducing computational costs by over 43%.", "date": "2025-06-05", "categories": []}
{"id": "2602.11615", "title": "SkillRater: Untangling Capabilities in Multimodal Data", "abstract": "Data curation methods typically assign samples a single quality score. We argue this scalar framing is fundamentally limited: when training requires multiple distinct capabilities, a monolithic scorer cannot maximize useful signals for all of them simultaneously. Quality is better understood as multidimensional, with each dimension corresponding to a capability the model must acquire. We introduce SkillRater, a framework that decomposes data filtering into specialized raters - one per capability, each trained via meta-learning on a disjoint validation objective - and composes their scores through a progressive selection rule: at each training stage, a sample is retained if any rater ranks it above a threshold that tightens over time, preserving diversity early while concentrating on high-value samples late. We validate this approach on vision language models, decomposing quality into three capability dimensions: visual understanding, OCR, and STEM reasoning. At 2B parameters, SkillRater improves over unfiltered baselines by 5.63% on visual understanding, 2.00% on OCR, and 3.53% on STEM on held out benchmarks. The learned rater signals are near orthogonal, confirming that the decomposition captures genuinely independent quality dimensions and explaining why it outperforms both unfiltered training and monolithic learned filtering.", "date": "2026-02-12", "categories": []}
{"id": "2602.12237", "title": "Olmix: A Framework for Data Mixing Throughout LM Development", "abstract": "Data mixing -- determining the ratios of data from different domains -- is a first-order concern for training language models (LMs). While existing mixing methods show promise, they fall short when applied during real-world LM development. We present Olmix, a framework that addresses two such challenges. First, the configuration space for developing a mixing method is not well understood -- design choices across existing methods lack justification or consensus and overlook practical issues like data constraints. We conduct a comprehensive empirical study of this space, identifying which design choices lead to a strong mixing method. Second, in practice, the domain set evolves throughout LM development as datasets are added, removed, partitioned, and revised -- a problem setting largely unaddressed by existing works, which assume fixed domains. We study how to efficiently recompute the mixture after the domain set is updated, leveraging information from past mixtures. We introduce mixture reuse, a mechanism that reuses existing ratios and recomputes ratios only for domains affected by the update. Over a sequence of five domain-set updates mirroring real-world LM development, mixture reuse matches the performance of fully recomputing the mix after each update with 74% less compute and improves over training without mixing by 11.6% on downstream tasks.", "date": "2026-02-12", "categories": []}
{"id": "2505.19013", "title": "Faithful Group Shapley Value", "abstract": "Data Shapley is an important tool for data valuation, which quantifies the contribution of individual data points to machine learning models. In practice, group-level data valuation is desirable when data providers contribute data in batch. However, we identify that existing group-level extensions of Data Shapley are vulnerable to shell company attacks, where strategic group splitting can unfairly inflate valuations. We propose Faithful Group Shapley Value (FGSV) that uniquely defends against such attacks. Building on original mathematical insights, we develop a provably fast and accurate approximation algorithm for computing FGSV. Empirical experiments demonstrate that our algorithm significantly outperforms state-of-the-art methods in computational efficiency and approximation accuracy, while ensuring faithful group-level valuation.", "date": "2025-05-25", "categories": []}
{"id": "2511.10192", "title": "Text2SQL-Flow: A Robust SQL-Aware Data Augmentation Framework for Text-to-SQL", "abstract": "The data-centric paradigm has emerged as a pivotal direction in artificial intelligence (AI), emphasizing the role of high-quality training data. This shift is especially critical in the Text-to-SQL task, where the scarcity, limited diversity, and structural simplicity of existing datasets constrain model performance. To address these challenges, we propose Text2SQL-Flow, a SQL-aware data augmentation framework that systematically generates large-scale, semantically valid, and structurally diverse Text-to-SQL pairs from limited seed data. Our framework spans six augmentation dimensions and integrates an end-to-end pipeline with auxiliary database selection, SQL executability verification, natural language (NL) question generation, NL-SQL correspondence verification, and chain-of-thought (CoT) reasoning trace generation. Leveraging this framework, we construct SQLFlow, a high-quality dataset comprising 75,386 annotated examples. We demonstrate the utility of SQLFlow in both fine-tuning and prompt-based settings. (1) For open-source large language models (LLMs), fine-tuning with SQLFlow improves problem-solving ability, delivering competitive gains across multiple benchmarks under the same data budget. (2) For closed-source LLMs, we propose a masked alignment retrieval method that uses SQLFlow as both a knowledge base and training data for the retrieval model, enabling structure-aware example matching via fine-grained NL-SQL alignments. Experiments show that our retrieval strategy outperforms existing example retrieval methods, highlighting the combined value of SQLFlow's data quality and our retrieval technique. Overall, our work provides a scalable, data-centric foundation for advancing Text-to-SQL systems and underscores the importance of structured, high-fidelity data in modern AI development. Our code is available at https://github.com/TechNomad-ds/Text2SQL-Flow.", "date": "2025-11-13", "categories": []}
{"id": "2602.09372", "title": "AgentSkiller: Scaling Generalist Agent Intelligence through Semantically Integrated Cross-Domain Data Synthesis", "abstract": "Large Language Model agents demonstrate potential in solving real-world problems via tools, yet generalist intelligence is bottlenecked by scarce high-quality, long-horizon data. Existing methods collect privacy-constrained API logs or generate scripted interactions lacking diversity, which struggle to produce data requisite for scaling capabilities. We propose AgentSkiller, a fully automated framework synthesizing multi-turn interaction data across realistic, semantically linked domains. It employs a DAG-based architecture with explicit state transitions to ensure determinism and recoverability. The pipeline builds a domain ontology and Person-Centric Entity Graph, defines tool interfaces via Service Blueprints for Model Context Protocol servers, and populates environments with consistent databases and strict Domain Policies. A cross-domain fusion mechanism links services to simulate complex tasks. Finally, the pipeline creates user tasks by verifying solution paths, filtering via execution-based validation, and generating queries using a Persona-based Simulator for automated rollout. This produces reliable environments with clear state changes. To demonstrate effectiveness, we synthesized $\\approx$ 11K interaction samples; experimental results indicate that models trained on this dataset achieve significant improvements on function calling over baselines, particularly in larger parameter regimes.", "date": "2026-02-10", "categories": []}
{"id": "2506.21910", "title": "AutoMixer: Checkpoint Artifacts as Automatic Data Mixers", "abstract": "In language model training, it is desirable to equip models with capabilities from various tasks. However, it is not clear how to directly obtain the right data mixtures for these capabilities as the relationship between data and tasks is difficult to be modeled. In this work, we observe that checkpoint models exhibit emerging capabilities at different points in the training trajectory. Often, the training process saves checkpoints as artifacts that are under-utilized as a source of in-training data signals. We identify these artifact models based on their respective capabilities on the benchmarks and leverage them as data mixers by using their aggregated first-order influence approximation over source data. We demonstrated on eight reasoning benchmarks that the proposed framework shows significant improvements in the pretraining setting, with performance improvements of up to 1.93%. Overall, this shows the potential of checkpoint models to enhance data quality and optimize data mixtures.", "date": "2025-06-27", "categories": []}
{"id": "2511.20222", "title": "Decoupling and Damping: Structurally-Regularized Gradient Matching for Multimodal Graph Condensation", "abstract": "In multimodal graph learning, graph structures that integrate information from multiple sources, such as vision and text, can more comprehensively model complex entity relationships. However, the continuous growth of their data scale poses a significant computational bottleneck for training. Graph condensation methods provide a feasible path forward by synthesizing compact and representative datasets. Nevertheless, existing condensation approaches generally suffer from performance limitations in multimodal scenarios, mainly due to two reasons: (1) semantic misalignment between different modalities leads to gradient conflicts; (2) the message-passing mechanism of graph neural networks further structurally amplifies such gradient noise. Based on this, we propose Structural Regularized Gradient Matching (SR-GM), a condensation framework for multimodal graphs. This method alleviates gradient conflicts between modalities through a gradient decoupling mechanism and introduces a structural damping regularizer to suppress the propagation of gradient noise in the topology, thereby transforming the graph structure from a noise amplifier into a training stabilizer. Extensive experiments on four multimodal graph datasets demonstrate the effectiveness of SR-GM, highlighting its state-of-the-art performance and cross-architecture generalization capabilities in multimodal graph dataset condensation.", "date": "2025-11-25", "categories": []}
{"id": "2512.12932", "title": "Investigating Data Pruning for Pretraining Biological Foundation Models at Scale", "abstract": "Biological foundation models (BioFMs), pretrained on large-scale biological sequences, have recently shown strong potential in providing meaningful representations for diverse downstream bioinformatics tasks. However, such models often rely on millions to billions of training sequences and billions of parameters, resulting in prohibitive computational costs and significant barriers to reproducibility and accessibility, particularly for academic labs. To address these challenges, we investigate the feasibility of data pruning for BioFM pretraining and propose a post-hoc influence-guided data pruning framework tailored to biological domains. Our approach introduces a subset-based self-influence formulation that enables efficient estimation of sample importance at low computational cost, and builds upon it two simple yet effective selection strategies, namely Top-k Influence (Top I) and Coverage-Centric Influence (CCI). We empirically validate our method on two representative BioFMs, RNA-FM and ESM-C. For RNA, our framework consistently outperforms random selection baselines under an extreme pruning rate of over 99 percent, demonstrating its effectiveness. Furthermore, we show the generalizability of our framework on protein-related tasks using ESM-C. In particular, our coreset even outperforms random subsets that are ten times larger in both RNA and protein settings, revealing substantial redundancy in biological sequence datasets. These findings underscore the potential of influence-guided data pruning to substantially reduce the computational cost of BioFM pretraining, paving the way for more efficient, accessible, and sustainable biological AI research.", "date": "2025-12-15", "categories": []}
{"id": "2602.05400", "title": "OPUS: Towards Efficient and Principled Data Selection in Large Language Model Pre-training in Every Iteration", "abstract": "As high-quality public text approaches exhaustion, a phenomenon known as the Data Wall, pre-training is shifting from more tokens to better tokens. However, existing methods either rely on heuristic static filters that ignore training dynamics, or use dynamic yet optimizer-agnostic criteria based on raw gradients. We propose OPUS (Optimizer-induced Projected Utility Selection), a dynamic data selection framework that defines utility in the optimizer-induced update space. OPUS scores candidates by projecting their effective updates, shaped by modern optimizers, onto a target direction derived from a stable, in-distribution proxy. To ensure scalability, we employ Ghost technique with CountSketch for computational efficiency, and Boltzmann sampling for data diversity, incurring only 4.7\\% additional compute overhead. OPUS achieves remarkable results across diverse corpora, quality tiers, optimizers, and model scales. In pre-training of GPT-2 Large/XL on FineWeb and FineWeb-Edu with 30B tokens, OPUS outperforms industrial-level baselines and even full 200B-token training. Moreover, when combined with industrial-level static filters, OPUS further improves pre-training efficiency, even with lower-quality data. Furthermore, in continued pre-training of Qwen3-8B-Base on SciencePedia, OPUS achieves superior performance using only 0.5B tokens compared to full training with 3B tokens, demonstrating significant data efficiency gains in specialized domains.", "date": "2026-02-05", "categories": []}
{"id": "2602.07824", "title": "Data Darwinism Part I: Unlocking the Value of Scientific Data for Pre-training", "abstract": "Data quality determines foundation model performance, yet systematic processing frameworks are lacking. We introduce Data Darwinism, a ten-level taxonomy (L0-L9) that conceptualizes data-model co-evolution: advanced models produce superior data for next-generation systems. We validate this on scientific literature by constructing Darwin-Science, a 900B-token corpus (L0-L5). We identify a learnability gap in raw scientific text, which we bridge via L4 (Generative Refinement) and L5 (Cognitive Completion) using frontier LLMs to explicate reasoning and terminology.\n  To ensure rigorous attribution, we pre-trained daVinci-origin-3B/7B models from scratch, excluding scientific content to create contamination-free baselines. After 600B tokens of continued pre-training, Darwin-Science outperforms baselines by +2.12 (3B) and +2.95 (7B) points across 20+ benchmarks, rising to +5.60 and +8.40 points on domain-aligned tasks. Systematic progression to L5 yields a +1.36 total gain, confirming that higher-level processing unlocks latent data value. We release the Darwin-Science corpus and daVinci-origin models to enable principled, co-evolutionary development.", "date": "2026-02-08", "categories": []}
{"id": "2602.09003", "title": "Data Science and Technology Towards AGI Part I: Tiered Data Management", "abstract": "The development of artificial intelligence can be viewed as an evolution of data-driven learning paradigms, with successive shifts in data organization and utilization continuously driving advances in model capability. Current LLM research is dominated by a paradigm that relies heavily on unidirectional scaling of data size, increasingly encountering bottlenecks in data availability, acquisition cost, and training efficiency. In this work, we argue that the development of AGI is entering a new phase of data-model co-evolution, in which models actively guide data management while high-quality data, in turn, amplifies model capabilities. To implement this vision, we propose a tiered data management framework, designed to support the full LLM training lifecycle across heterogeneous learning objectives and cost constraints. Specifically, we introduce an L0-L4 tiered data management framework, ranging from raw uncurated resources to organized and verifiable knowledge. Importantly, LLMs are fully used in data management processes, such as quality scoring and content editing, to refine data across tiers. Each tier is characterized by distinct data properties, management strategies, and training roles, enabling data to be strategically allocated across LLM training stages, including pre-training, mid-training, and alignment. The framework balances data quality, acquisition cost, and marginal training benefit, providing a systematic approach to scalable and sustainable data management. We validate the effectiveness of the proposed framework through empirical studies, in which tiered datasets are constructed from raw corpora and used across multiple training phases. Experimental results demonstrate that tier-aware data utilization significantly improves training efficiency and model performance. To facilitate further research, we release our tiered datasets and processing tools to the community.", "date": "2026-02-09", "categories": []}
{"id": "2602.09008", "title": "ShapeCond: Fast Shapelet-Guided Dataset Condensation for Time Series Classification", "abstract": "Time series data supports many domains (e.g., finance and climate science), but its rapid growth strains storage and computation. Dataset condensation can alleviate this by synthesizing a compact training set that preserves key information. Yet most condensation methods are image-centric and often fail on time series because they miss time-series-specific temporal structure, especially local discriminative motifs such as shapelets. In this work, we propose ShapeCond, a novel and efficient condensation framework for time series classification that leverages shapelet-based dataset knowledge via a shapelet-guided optimization strategy. Our shapelet-assisted synthesis cost is independent of sequence length: longer series yield larger speedups in synthesis (e.g., 29$\\times$ faster over prior state-of-the-art method CondTSC for time-series condensation, and up to 10,000$\\times$ over naively using shapelets on the Sleep dataset with 3,000 timesteps). By explicitly preserving critical local patterns, ShapeCond improves downstream accuracy and consistently outperforms all prior state-of-the-art time series dataset condensation methods across extensive experiments. Code is available at https://github.com/lunaaa95/ShapeCond.", "date": "2026-02-09", "categories": []}
{"id": "2501.07681", "title": "Dataset Distillation as Pushforward Optimal Quantization", "abstract": "Dataset distillation aims to find a synthetic training set such that training on the synthetic data achieves similar performance to training on real data, with orders of magnitude less computational requirements. Existing methods can be broadly categorized as either bi-level optimization problems that have neural network training heuristics as the lower level problem, or disentangled methods that bypass the bi-level optimization by matching distributions of data. The latter method has the major advantages of speed and scalability in terms of size of both training and distilled datasets. We demonstrate that when equipped with an encoder-decoder structure, the empirically successful disentangled methods can be reformulated as an optimal quantization problem, where a finite set of points is found to approximate the underlying probability measure by minimizing the expected projection distance. In particular, we link existing disentangled dataset distillation methods to the classical optimal quantization and Wasserstein barycenter problems, demonstrating consistency of distilled datasets for diffusion-based generative priors. We propose Dataset Distillation by Optimal Quantization, based on clustering in a latent space. Compared to the previous SOTA method D\\textsuperscript{4}M, we achieve better performance and inter-model generalization on the ImageNet-1K dataset with trivial additional computation, and SOTA performance in higher image-per-class settings. Using the distilled noise initializations in a stronger diffusion transformer model, we obtain SOTA distillation performance on ImageNet-1K and its subsets, outperforming diffusion guidance methods.", "date": "2025-01-13", "categories": []}
{"id": "2505.18759", "title": "The Quest for Efficient Reasoning: A Data-Centric Benchmark to CoT Distillation", "abstract": "Data-centric distillation, including data augmentation, selection, and mixing, offers a promising path to creating smaller, more efficient student Large Language Models (LLMs) that retain strong reasoning abilities. However, there still lacks a comprehensive benchmark to systematically assess the effect of each distillation approach. This paper introduces DC-CoT, the first data-centric benchmark that investigates data manipulation in chain-of-thought (CoT) distillation from method, model and data perspectives. Utilizing various teacher models (e.g., o4-mini, Gemini-Pro, Claude-3.5) and student architectures (e.g., 3B, 7B parameters), we rigorously evaluate the impact of these data manipulations on student model performance across multiple reasoning datasets, with a focus on in-distribution (IID) and out-of-distribution (OOD) generalization, and cross-domain transfer. Our findings aim to provide actionable insights and establish best practices for optimizing CoT distillation through data-centric techniques, ultimately facilitating the development of more accessible and capable reasoning models. The codebase can be accessed at https://github.com/UNITES-Lab/Distillation-Bench", "date": "2025-05-24", "categories": []}
{"id": "2602.05251", "title": "TADS: Task-Aware Data Selection for Multi-Task Multimodal Pre-Training", "abstract": "Large-scale multimodal pre-trained models like CLIP rely heavily on high-quality training data, yet raw web-crawled datasets are often noisy, misaligned, and redundant, leading to inefficient training and suboptimal generalization. Existing data selection methods are either heuristic-based, suffering from bias and limited diversity, or data-driven but task-agnostic, failing to optimize for multi-task scenarios. To address these gaps, we introduce TADS (Task-Aware Data Selection), a novel framework for multi-task multimodal pre-training that integrates Intrinsic Quality, Task Relevance, and Distributional Diversity into a learnable value function. TADS employs a comprehensive quality assessment system with unimodal and cross-modal operators, quantifies task relevance via interpretable similarity vectors, and optimizes diversity through cluster-based weighting. A feedback-driven meta-learning mechanism adaptively refines the selection strategy based on proxy model performance across multiple downstream tasks. Experiments on CC12M demonstrate that TADS achieves superior zero-shot performance on benchmarks like ImageNet, CIFAR-100, MS-COCO, and Flickr30K, using only 36% of the data while outperforming baselines by an average of 1.0%. This highlights that TADS significantly enhances data efficiency by curating a high-utility subset that yields a much higher performance ceiling within the same computational constraints.", "date": "2026-02-05", "categories": []}
{"id": "2602.05616", "title": "Path-Guided Flow Matching for Dataset Distillation", "abstract": "Dataset distillation compresses large datasets into compact synthetic sets with comparable performance in training models. Despite recent progress on diffusion-based distillation, this type of method typically depends on heuristic guidance or prototype assignment, which comes with time-consuming sampling and trajectory instability and thus hurts downstream generalization especially under strong control or low IPC. We propose \\emph{Path-Guided Flow Matching (PGFM)}, the first flow matching-based framework for generative distillation, which enables fast deterministic synthesis by solving an ODE in a few steps. PGFM conducts flow matching in the latent space of a frozen VAE to learn class-conditional transport from Gaussian noise to data distribution. Particularly, we develop a continuous path-to-prototype guidance algorithm for ODE-consistent path control, which allows trajectories to reliably land on assigned prototypes while preserving diversity and efficiency. Extensive experiments across high-resolution benchmarks demonstrate that PGFM matches or surpasses prior diffusion-based distillation approaches with fewer steps of sampling while delivering competitive performance with remarkably improved efficiency, e.g., 7.6$\\times$ more efficient than the diffusion-based counterparts with 78\\% mode coverage.", "date": "2026-02-05", "categories": []}
{"id": "2602.04365", "title": "EXaMCaP: Subset Selection with Entropy Gain Maximization for Probing Capability Gains of Large Chart Understanding Training Sets", "abstract": "Recent works focus on synthesizing Chart Understanding (ChartU) training sets to inject advanced chart knowledge into Multimodal Large Language Models (MLLMs), where the sufficiency of the knowledge is typically verified by quantifying capability gains via the fine-tune-then-evaluate paradigm. However, full-set fine-tuning MLLMs to assess such gains incurs significant time costs, hindering the iterative refinement cycles of the ChartU dataset. Reviewing the ChartU dataset synthesis and data selection domains, we find that subsets can potentially probe the MLLMs' capability gains from full-set fine-tuning. Given that data diversity is vital for boosting MLLMs' performance and entropy reflects this feature, we propose EXaMCaP, which uses entropy gain maximization to select a subset. To obtain a high-diversity subset, EXaMCaP chooses the maximum-entropy subset from the large ChartU dataset. As enumerating all possible subsets is impractical, EXaMCaP iteratively selects samples to maximize the gain in set entropy relative to the current set, approximating the maximum-entropy subset of the full dataset. Experiments show that EXaMCaP outperforms baselines in probing the capability gains of the ChartU training set, along with its strong effectiveness across diverse subset sizes and compatibility with various MLLM architectures.", "date": "2026-02-04", "categories": []}
{"id": "2510.04838", "title": "Beyond Random: Automatic Inner-loop Optimization in Dataset Distillation", "abstract": "The growing demand for efficient deep learning has positioned dataset distillation as a pivotal technique for compressing training dataset while preserving model performance. However, existing inner-loop optimization methods for dataset distillation typically rely on random truncation strategies, which lack flexibility and often yield suboptimal results. In this work, we observe that neural networks exhibit distinct learning dynamics across different training stages-early, middle, and late-making random truncation ineffective. To address this limitation, we propose Automatic Truncated Backpropagation Through Time (AT-BPTT), a novel framework that dynamically adapts both truncation positions and window sizes according to intrinsic gradient behavior. AT-BPTT introduces three key components: (1) a probabilistic mechanism for stage-aware timestep selection, (2) an adaptive window sizing strategy based on gradient variation, and (3) a low-rank Hessian approximation to reduce computational overhead. Extensive experiments on CIFAR-10, CIFAR-100, Tiny-ImageNet, and ImageNet-1K show that AT-BPTT achieves state-of-the-art performance, improving accuracy by an average of 6.16% over baseline methods. Moreover, our approach accelerates inner-loop optimization by 3.9x while saving 63% memory cost.", "date": "2025-10-06", "categories": []}
{"id": "2602.02564", "title": "Label Curation Using Agentic AI", "abstract": "Data annotation is essential for supervised learning, yet producing accurate, unbiased, and scalable labels remains challenging as datasets grow in size and modality. Traditional human-centric pipelines are costly, slow, and prone to annotator variability, motivating reliability-aware automated annotation. We present AURA (Agentic AI for Unified Reliability Modeling and Annotation Aggregation), an agentic AI framework for large-scale, multi-modal data annotation. AURA coordinates multiple AI agents to generate and validate labels without requiring ground truth. At its core, AURA adapts a classical probabilistic model that jointly infers latent true labels and annotator reliability via confusion matrices, using Expectation-Maximization to reconcile conflicting annotations and aggregate noisy predictions. Across the four benchmark datasets evaluated, AURA achieves accuracy improvements of up to 5.8% over baseline. In more challenging settings with poor quality annotators, the improvement is up to 50% over baseline. AURA also accurately estimates the reliability of annotators, allowing assessment of annotator quality even without any pre-validation steps.", "date": "2026-01-30", "categories": []}
{"id": "2602.03123", "title": "Beyond Cropping and Rotation: Automated Evolution of Powerful Task-Specific Augmentations with Generative Models", "abstract": "Data augmentation has long been a cornerstone for reducing overfitting in vision models, with methods like AutoAugment automating the design of task-specific augmentations. Recent advances in generative models, such as conditional diffusion and few-shot NeRFs, offer a new paradigm for data augmentation by synthesizing data with significantly greater diversity and realism. However, unlike traditional augmentations like cropping or rotation, these methods introduce substantial changes that enhance robustness but also risk degrading performance if the augmentations are poorly matched to the task. In this work, we present EvoAug, an automated augmentation learning pipeline, which leverages these generative models alongside an efficient evolutionary algorithm to learn optimal task-specific augmentations. Our pipeline introduces a novel approach to image augmentation that learns stochastic augmentation trees that hierarchically compose augmentations, enabling more structured and adaptive transformations. We demonstrate strong performance across fine-grained classification and few-shot learning tasks. Notably, our pipeline discovers augmentations that align with domain knowledge, even in low-data settings. These results highlight the potential of learned generative augmentations, unlocking new possibilities for robust model training.", "date": "2026-02-03", "categories": []}
{"id": "2509.19743", "title": "Rectified Decoupled Dataset Distillation: A Closer Look for Fair and Comprehensive Evaluation", "abstract": "Dataset distillation aims to generate compact synthetic datasets that enable models trained on them to achieve performance comparable to those trained on full real datasets, while substantially reducing storage and computational costs. Early bi-level optimization methods (e.g., MTT) have shown promising results on small-scale datasets, but their scalability is limited by high computational overhead. To address this limitation, recent decoupled dataset distillation methods (e.g., SRe$^2$L) separate the teacher model pre-training from the synthetic data generation process. These methods also introduce random data augmentation and epoch-wise soft labels during the post-evaluation phase to improve performance and generalization. However, existing decoupled distillation methods suffer from inconsistent post-evaluation protocols, which hinders progress in the field. In this work, we propose Rectified Decoupled Dataset Distillation (RD$^3$), and systematically investigate how different post-evaluation settings affect test accuracy. We further examine whether the reported performance differences across existing methods reflect true methodological advances or stem from discrepancies in evaluation procedures. Our analysis reveals that much of the performance variation can be attributed to inconsistent evaluation rather than differences in the intrinsic quality of the synthetic data. In addition, we identify general strategies that improve the effectiveness of distilled datasets across settings. By establishing a standardized benchmark and rigorous evaluation protocol, RD$^3$ provides a foundation for fair and reproducible comparisons in future dataset distillation research.", "date": "2025-09-24", "categories": []}
{"id": "2602.00885", "title": "Reliability-Aware Determinantal Point Processes for Robust Informative Data Selection in Large Language Models", "abstract": "Informative data selection is a key requirement for large language models (LLMs) to minimize the amount of data required for fine-tuning, network distillation, and token pruning, enabling fast and efficient deployment, especially under computational and communication constraints. Traditional subset selection methods, including those based on Determinantal Point Processes (DPP), focus on maximizing diversity but assume that selected data batches are always available error-free. This presumption prohibits their use under partial storage outage, imperfect communication, and stochastic access failures. Furthermore, we show that the original formulation collapses under such conditions. To address this gap, we introduce ProbDPP, a novel reliability-aware implementation of k-DPP that accounts for probabilistic data access by recasting the objective function with a regularization term that remains well-posed and decomposes into a geometric diversity term and unreliability cost. The resulting objective facilitates robust selection of diverse data batches under uncertainty. Furthermore, we frame this reliability-aware diversity maximization as a combinatorial semi-bandit problem and propose a UCB-style algorithm to efficiently learn the unknown reliability online. Theoretical analysis provides regret bounds for the proposed approach, ensuring performance guarantees.", "date": "2026-01-31", "categories": []}
{"id": "2509.21479", "title": "Filtering with Confidence: When Data Augmentation Meets Conformal Prediction", "abstract": "With promising empirical performance across a wide range of applications, synthetic data augmentation appears a viable solution to data scarcity and the demands of increasingly data-intensive models. Its effectiveness lies in expanding the training set in a way that reduces estimator variance while introducing only minimal bias. Controlling this bias is therefore critical: effective data augmentation should generate diverse samples from the same underlying distribution as the training set, with minimal shifts. In this paper, we propose conformal data augmentation, a principled data filtering framework that leverages the power of conformal prediction to produce diverse synthetic data while filtering out poor-quality generations with provable risk control. Our method is simple to implement, requires no access to internal model logits, nor large-scale model retraining. We demonstrate the effectiveness of our approach across multiple tasks, including topic prediction, sentiment analysis, image classification, and fraud detection, showing consistent performance improvements of up to 40 percentage points (pp) in $F_1$ score over unaugmented baselines, and 4~pp over other filtered augmentation baselines.", "date": "2025-09-25", "categories": []}
{"id": "2505.17799", "title": "A Coreset Selection of Coreset Selection Literature: Introduction and Recent Advances", "abstract": "Coreset selection targets the challenge of finding a small, representative subset of a large dataset that preserves essential patterns for effective machine learning. Although several surveys have examined data reduction strategies before, most focus narrowly on either classical geometry-based methods or active learning techniques. In contrast, this survey presents a more comprehensive view by unifying three major lines of coreset research, namely, training-free, training-oriented, and label-free approaches, into a single taxonomy. We present subfields often overlooked by existing work, including submodular formulations, bilevel optimization, and recent progress in pseudo-labeling for unlabeled datasets. Additionally, we examine how pruning strategies influence generalization and neural scaling laws, offering new insights that are absent from prior reviews. Finally, we compare these methods under varying computational, robustness, and performance demands and highlight open challenges, such as robustness, outlier filtering, and adapting coreset selection to foundation models, for future research.", "date": "2025-05-23", "categories": []}
{"id": "2512.18834", "title": "Mix, MinHash, and Match: Cross-Source Agreement for Multilingual Pretraining Datasets", "abstract": "Multilingual data from the web is essential for LLM pretraining. Yet, scraping it is expensive, and research groups repeatedly crawl the same content. For example, we found that over 40\\% of tokens across major Arabic web corpora are duplicated between sources. In this work, we propose to use this wasteful redundancy as a quality signal to create high-quality pretraining datasets. Our key insight is that cross-source agreement functions as a free, model-free quality filter: content retained by multiple independent pipelines is more likely to represent high-quality text. Crucially, this signal requires no additional computation beyond standard deduplication, which is already performed at scale when pretraining language models. So, we propose MixMinMatch, a method that combines multiple existing web corpora, performs cross-dataset MinHash deduplication, and identifies documents independently recovered by multiple sources. We apply MixMinMatch to Arabic, Turkish, and Hindi, producing corpora that match or exceed the quality of the best single-source baselines, while providing up to 4$\\times$ more unique tokens. On Arabic, our matched subset achieves a 4.5\\% relative improvement over ArabicWeb24, while on Turkish, we improve over FineWeb-2 by 5.5\\%. We release the datasets at: https://huggingface.co/collections/AdaMLLab/mixminmatch", "date": "2025-12-21", "categories": []}
{"id": "2601.21296", "title": "Grounding and Enhancing Informativeness and Utility in Dataset Distillation", "abstract": "Dataset Distillation (DD) seeks to create a compact dataset from a large, real-world dataset. While recent methods often rely on heuristic approaches to balance efficiency and quality, the fundamental relationship between original and synthetic data remains underexplored. This paper revisits knowledge distillation-based dataset distillation within a solid theoretical framework. We introduce the concepts of Informativeness and Utility, capturing crucial information within a sample and essential samples in the training set, respectively. Building on these principles, we define optimal dataset distillation mathematically. We then present InfoUtil, a framework that balances informativeness and utility in synthesizing the distilled dataset. InfoUtil incorporates two key components: (1) game-theoretic informativeness maximization using Shapley Value attribution to extract key information from samples, and (2) principled utility maximization by selecting globally influential samples based on Gradient Norm. These components ensure that the distilled dataset is both informative and utility-optimized. Experiments demonstrate that our method achieves a 6.1\\% performance improvement over the previous state-of-the-art approach on ImageNet-1K dataset using ResNet-18.", "date": "2026-01-29", "categories": []}
{"id": "2410.00718", "title": "Pseudo-Nonlinear Data Augmentation: A Constrained Energy Minimization Viewpoint", "abstract": "We propose a simple yet novel data augmentation method for general data modalities based on energy-based modeling and principles from information geometry. Unlike most existing learning-based data augmentation methods, which rely on learning latent representations with generative models, our proposed framework enables an intuitive construction of a geometrically aware latent space that represents the structure of the data itself, supporting efficient and explicit encoding and decoding procedures. We then present and discuss how to design latent spaces that will subsequently control the augmentation with the proposed algorithm. Empirical results demonstrate that our data augmentation method achieves competitive performance in downstream tasks compared to other baselines, while offering fine-grained controllability that is lacking in the existing literature.", "date": "2024-10-01", "categories": []}
{"id": "2509.17543", "title": "Bilateral Distribution Compression: Reducing Both Data Size and Dimensionality", "abstract": "Existing distribution compression methods reduce the number of observations in a dataset by minimising the Maximum Mean Discrepancy (MMD) between original and compressed sets, but modern datasets are often large in both sample size and dimensionality. We propose Bilateral Distribution Compression (BDC), a two-stage framework that compresses along both axes while preserving the underlying distribution, with overall linear time and memory complexity in dataset size and dimension. Central to BDC is the Decoded MMD (DMMD), which we introduce to quantify the discrepancy between the original data and a compressed set decoded from a low-dimensional latent space. BDC proceeds by (i) learning a low-dimensional projection using the Reconstruction MMD (RMMD), and (ii) optimising a latent compressed set with the Encoded MMD (EMMD). We show that this procedure minimises the DMMD, guaranteeing that the compressed set faithfully represents the original distribution. Experiments show that BDC can achieve comparable or superior downstream task performance to ambient-space compression at substantially lower cost and with significantly higher rates of compression.", "date": "2025-09-22", "categories": []}
{"id": "2512.12210", "title": "EEG-DLite: Dataset Distillation for Efficient Large EEG Model Training", "abstract": "Large-scale EEG foundation models have shown strong generalization across a range of downstream tasks, but their training remains resource-intensive due to the volume and variable quality of EEG data. In this work, we introduce EEG-DLite, a data distillation framework that enables more efficient pre-training by selectively removing noisy and redundant samples from large EEG datasets. EEG-DLite begins by encoding EEG segments into compact latent representations using a self-supervised autoencoder, allowing sample selection to be performed efficiently and with reduced sensitivity to noise. Based on these representations, EEG-DLite filters out outliers and minimizes redundancy, resulting in a smaller yet informative subset that retains the diversity essential for effective foundation model training. Through extensive experiments, we demonstrate that training on only 5 percent of a 2,500-hour dataset curated with EEG-DLite yields performance comparable to, and in some cases better than, training on the full dataset across multiple downstream tasks. To our knowledge, this is the first systematic study of pre-training data distillation in the context of EEG foundation models. EEG-DLite provides a scalable and practical path toward more effective and efficient physiological foundation modeling. The code is available at https://github.com/t170815518/EEG-DLite.", "date": "2025-12-13", "categories": []}
{"id": "2512.15765", "title": "Data Valuation for LLM Fine-Tuning: Efficient Shapley Value Approximation via Language Model Arithmetic", "abstract": "Data is a critical asset for training large language models (LLMs), alongside compute resources and skilled workers. While some training data is publicly available, substantial investment is required to generate proprietary datasets, such as human preference annotations or to curate new ones from existing sources. As larger datasets generally yield better model performance, two natural questions arise. First, how can data owners make informed decisions about curation strategies and data sources investment? Second, how can multiple data owners collaboratively pool their resources to train superior models while fairly distributing the benefits? This problem, data valuation, which is not specific to large language models, has been addressed by the machine learning community through the lens of cooperative game theory, with the Shapley value being the prevalent solution concept. However, computing Shapley values is notoriously expensive for data valuation, typically requiring numerous model retrainings, which can become prohibitive for large machine learning models. In this work, we demonstrate that this computational challenge is dramatically simplified for LLMs trained with Direct Preference Optimization (DPO). We show how the specific mathematical structure of DPO enables scalable Shapley value computation. We believe this observation unlocks many applications at the intersection of data valuation and large language models.", "date": "2025-12-12", "categories": []}
{"id": "2510.17873", "title": "Auditing and Mitigating Bias in Gender Classification Algorithms: A Data-Centric Approach", "abstract": "Gender classification systems often inherit and amplify demographic imbalances in their training data. We first audit five widely used gender classification datasets, revealing that all suffer from significant intersectional underrepresentation. To measure the downstream impact of these flaws, we train identical MobileNetV2 classifiers on the two most balanced of these datasets, UTKFace and FairFace. Our fairness evaluation shows that even these models exhibit significant bias, misclassifying female faces at a higher rate than male faces and amplifying existing racial skew. To counter these data-induced biases, we construct BalancedFace, a new public dataset created by blending images from FairFace and UTKFace, supplemented with images from other collections to fill missing demographic gaps. It is engineered to equalize subgroup shares across 189 intersections of age, race, and gender using only real, unedited images. When a standard classifier is trained on BalancedFace, it reduces the maximum True Positive Rate gap across racial subgroups by over 50% and brings the average Disparate Impact score 63% closer to the ideal of 1.0 compared to the next-best dataset, all with a minimal loss of overall accuracy. These results underline the profound value of data-centric interventions and provide an openly available resource for fair gender classification research.", "date": "2025-10-17", "categories": []}
{"id": "2601.15829", "title": "Towards Realistic Remote Sensing Dataset Distillation with Discriminative Prototype-guided Diffusion", "abstract": "Recent years have witnessed the remarkable success of deep learning in remote sensing image interpretation, driven by the availability of large-scale benchmark datasets. However, this reliance on massive training data also brings two major challenges: (1) high storage and computational costs, and (2) the risk of data leakage, especially when sensitive categories are involved. To address these challenges, this study introduces the concept of dataset distillation into the field of remote sensing image interpretation for the first time. Specifically, we train a text-to-image diffusion model to condense a large-scale remote sensing dataset into a compact and representative distilled dataset. To improve the discriminative quality of the synthesized samples, we propose a classifier-driven guidance by injecting a classification consistency loss from a pre-trained model into the diffusion training process. Besides, considering the rich semantic complexity of remote sensing imagery, we further perform latent space clustering on training samples to select representative and diverse prototypes as visual style guidance, while using a visual language model to provide aggregated text descriptions. Experiments on three high-resolution remote sensing scene classification benchmarks show that the proposed method can distill realistic and diverse samples for downstream model training. Code and pre-trained models are available online (https://github.com/YonghaoXu/DPD).", "date": "2026-01-22", "categories": []}
{"id": "2601.13697", "title": "Uncertainty-Aware Gradient Signal-to-Noise Data Selection for Instruction Tuning", "abstract": "Instruction tuning is a standard paradigm for adapting large language models (LLMs), but modern instruction datasets are large, noisy, and redundant, making full-data fine-tuning costly and often unnecessary. Existing data selection methods either build expensive gradient datastores or assign static scores from a weak proxy, largely ignoring evolving uncertainty, and thus missing a key source of LLM interpretability. We propose GRADFILTERING, an objective-agnostic, uncertainty-aware data selection framework that utilizes a small GPT-2 proxy with a LoRA ensemble and aggregates per-example gradients into a Gradient Signal-to-Noise Ratio (G-SNR) utility. Our method matches or surpasses random subsets and strong baselines in most LLM-as-a-judge evaluations as well as in human assessment. Moreover, GRADFILTERING-selected subsets converge faster than competitive filters under the same compute budget, reflecting the benefit of uncertainty-aware scoring.", "date": "2026-01-20", "categories": []}
{"id": "2510.08955", "title": "Denoised Diffusion for Object-Focused Image Augmentation", "abstract": "Modern agricultural operations increasingly rely on integrated monitoring systems that combine multiple data sources for farm optimization. Aerial drone-based animal health monitoring serves as a key component but faces limited data availability, compounded by scene-specific issues such as small, occluded, or partially visible animals. Transfer learning approaches often fail to address this limitation due to the unavailability of large datasets that reflect specific farm conditions, including variations in animal breeds, environments, and behaviors. Therefore, there is a need for developing a problem-specific, animal-focused data augmentation strategy tailored to these unique challenges. To address this gap, we propose an object-focused data augmentation framework designed explicitly for animal health monitoring in constrained data settings. Our approach segments animals from backgrounds and augments them through transformations and diffusion-based synthesis to create realistic, diverse scenes that enhance animal detection and monitoring performance. Our initial experiments demonstrate that our augmented dataset yields superior performance compared to our baseline models on the animal detection task. By generating domain-specific data, our method empowers real-time animal health monitoring solutions even in data-scarce scenarios, bridging the gap between limited data and practical applicability.", "date": "2025-10-10", "categories": []}
{"id": "2601.10922", "title": "What Matters in Data Curation for Multimodal Reasoning? Insights from the DCVLR Challenge", "abstract": "We study data curation for multimodal reasoning through the NeurIPS 2025 Data Curation for Vision-Language Reasoning (DCVLR) challenge, which isolates dataset selection by fixing the model and training protocol. Using a compact curated dataset derived primarily from Walton Multimodal Cold Start, our submission placed first in the challenge. Through post-competition ablations, we show that difficulty-based example selection on an aligned base dataset is the dominant driver of performance gains. Increasing dataset size does not reliably improve mean accuracy under the fixed training recipe, but mainly reduces run-to-run variance, while commonly used diversity and synthetic augmentation heuristics provide no additional benefit and often degrade performance. These results characterize DCVLR as a saturation-regime evaluation and highlight the central role of alignment and difficulty in data-efficient multimodal reasoning.", "date": "2026-01-16", "categories": []}
{"id": "2601.10986", "title": "ZPD Detector: Data Selection via Capability-Difficulty Alignment for Large Language Models", "abstract": "As the cost of training large language models continues to increase and high-quality training data become increasingly scarce, selecting high-value samples or synthesizing effective training data under limited data budgets has emerged as a critical research problem. Most existing data selection methods rely on static criteria, such as difficulty, uncertainty, or heuristics, and fail to model the evolving relationship between the model and the data. Inspired by the educational theory of the Zone of Proximal Development (ZPD), we propose ZPD Detector, a data selection framework that adopts a bidirectional perspective between models and data by explicitly modeling the alignment between sample difficulty and the model's current capability. ZPD Detector integrates difficulty calibration, model capability estimation based on Item Response Theory (IRT), and a capability-difficulty matching score to dynamically identify the most informative samples at each learning stage, improving data utilization efficiency; moreover, this dynamic matching strategy provides new insights into training strategy design. All code and data will be released after our work be accepted to support reproducible researc", "date": "2026-01-16", "categories": []}
{"id": "2502.10510", "title": "MixMin: Finding Data Mixtures via Convex Minimization", "abstract": "Modern machine learning pipelines are increasingly combining and mixing data from diverse and disparate sources, e.g., pre-training large language models. Yet, finding the optimal data mixture is a challenging and open problem. We formalize this data mixing problem as a bi-level objective: the best mixture is the one that would lead to the best model for a downstream objective. Unfortunately, this objective is generally intractable. In this paper, we make the observation that the bi-level data mixing objective becomes convex as our model class becomes larger. We develop and study a gradient-based approach for optimizing this convex objective, which we call MixMin, and test it on language modeling and chemistry tasks. MixMin was the only method that uniformly improved the data mixture in all our experiments. With MixMin, we improved the data mixture using less than 0.2% additional compute for a pythia-410M model trained on 8.2B tokens, resulting between 1-5% relative improvement to negative log likelihood on PIQA, ARC Easy, SciQ, and OpenWebMath. Crucially, we found that MixMin mixtures for smaller models improved training of larger models, suggesting that MixMin mixtures may be scale-invariant. When mixing bioassay data to train an XGBoost model, we saw improvements to average precision scores of 0.03-0.15.", "date": "2025-02-14", "categories": []}
{"id": "2512.21011", "title": "Granular Ball Guided Masking: Structure-aware Data Augmentation", "abstract": "Deep learning models have achieved remarkable success in computer vision but still rely heavily on large-scale labeled data and tend to overfit when data is limited or distributions shift. Data augmentation -- particularly mask-based information dropping -- can enhance robustness by forcing models to explore complementary cues; however, existing approaches often lack structural awareness and risk discarding essential semantics. We propose Granular Ball Guided Masking (GBGM), a structure-aware augmentation strategy guided by Granular Ball Computing (GBC). GBGM adaptively preserves semantically rich, structurally important regions while suppressing redundant areas through a coarse-to-fine hierarchical masking process, producing augmentations that are both representative and discriminative. Extensive experiments on multiple benchmarks demonstrate consistent improvements not only in image classification and masked image reconstruction, but also in image tampering detection, validating the effectiveness and generalization of GBGM across both recognition and forensic scenarios. Simple and model-agnostic, GBGM integrates seamlessly into CNNs and Vision Transformers, offering a practical paradigm for structure-aware data augmentation.", "date": "2025-12-24", "categories": []}
{"id": "2601.09733", "title": "Closing the Data Loop: Using OpenDataArena to Engineer Superior Training Datasets", "abstract": "The construction of Supervised Fine-Tuning (SFT) datasets is a critical yet under-theorized stage in the post-training of Large Language Models (LLMs), as prevalent practices often rely on heuristic aggregation without a systematic understanding of how individual samples contribute to model performance. In this report, we propose a paradigm shift from ad-hoc curation to a closed-loop dataset engineering framework using OpenDataArena (ODA), which leverages value-anchored rankings and multi-dimensional analysis to transform value benchmarking into feedback signals guiding dataset construction. We instantiate this methodology through two new datasets: \\textbf{ODA-Math-460k}, a specialized mathematics reasoning dataset that utilizes a novel two-stage difficulty-aware pipeline to achieve State-of-the-Art (SOTA) results on benchmarks such as AIME and HMMT, and \\textbf{ODA-Mixture (100k \\& 500k)}, a series of multi-domain instruction datasets built via an ``Anchor-and-Patch'' strategy that outperforms significantly larger open-source baselines. Our empirical results demonstrate that ODA-driven datasets significantly improve both domain-specific reasoning and general utility while achieving superior data efficiency, validating a transition toward data-centric AI where transparent evaluation serves as the primary engine for engineering high-quality training data.", "date": "2025-12-30", "categories": []}
{"id": "2601.10090", "title": "Difficulty-guided Sampling: Bridging the Target Gap between Dataset Distillation and Downstream Tasks", "abstract": "In this paper, we propose difficulty-guided sampling (DGS) to bridge the target gap between the distillation objective and the downstream task, therefore improving the performance of dataset distillation. Deep neural networks achieve remarkable performance but have time and storage-consuming training processes. Dataset distillation is proposed to generate compact, high-quality distilled datasets, enabling effective model training while maintaining downstream performance. Existing approaches typically focus on features extracted from the original dataset, overlooking task-specific information, which leads to a target gap between the distillation objective and the downstream task. We propose leveraging characteristics that benefit the downstream training into data distillation to bridge this gap. Focusing on the downstream task of image classification, we introduce the concept of difficulty and propose DGS as a plug-in post-stage sampling module. Following the specific target difficulty distribution, the final distilled dataset is sampled from image pools generated by existing methods. We also propose difficulty-aware guidance (DAG) to explore the effect of difficulty in the generation process. Extensive experiments across multiple settings demonstrate the effectiveness of the proposed methods. It also highlights the broader potential of difficulty for diverse downstream tasks.", "date": "2026-01-15", "categories": []}
{"id": "2601.09531", "title": "Bipartite Mode Matching for Vision Training Set Search from a Hierarchical Data Server", "abstract": "We explore a situation in which the target domain is accessible, but real-time data annotation is not feasible. Instead, we would like to construct an alternative training set from a large-scale data server so that a competitive model can be obtained. For this problem, because the target domain usually exhibits distinct modes (i.e., semantic clusters representing data distribution), if the training set does not contain these target modes, the model performance would be compromised. While prior existing works improve algorithms iteratively, our research explores the often-overlooked potential of optimizing the structure of the data server. Inspired by the hierarchical nature of web search engines, we introduce a hierarchical data server, together with a bipartite mode matching algorithm (BMM) to align source and target modes. For each target mode, we look in the server data tree for the best mode match, which might be large or small in size. Through bipartite matching, we aim for all target modes to be optimally matched with source modes in a one-on-one fashion. Compared with existing training set search algorithms, we show that the matched server modes constitute training sets that have consistently smaller domain gaps with the target domain across object re-identification (re-ID) and detection tasks. Consequently, models trained on our searched training sets have higher accuracy than those trained otherwise. BMM allows data-centric unsupervised domain adaptation (UDA) orthogonal to existing model-centric UDA methods. By combining the BMM with existing UDA methods like pseudo-labeling, further improvement is observed.", "date": "2026-01-14", "categories": []}
{"id": "2502.13691", "title": "Is This Collection Worth My LLM's Time? Automatically Measuring Information Potential in Text Corpora", "abstract": "As large language models (LLMs) converge towards similar capabilities, the key to advancing their performance lies in identifying and incorporating valuable new information sources. However, evaluating which text collections are worth the substantial investment required for digitization, preprocessing, and integration into LLM systems remains a significant challenge. We present a novel approach to this challenge: an automated pipeline that evaluates the potential information gain from text collections without requiring model training or fine-tuning. Our method generates multiple choice questions (MCQs) from texts and measures an LLM's performance both with and without access to the source material. The performance gap between these conditions serves as a proxy for the collection's information potential. We validate our approach using five strategically selected datasets: EPFL PhD manuscripts, a private collection of Venetian historical records, two sets of Wikipedia articles on related topics, and a synthetic baseline dataset. Our results demonstrate that this method effectively identifies collections containing valuable novel information, providing a practical tool for prioritizing data acquisition and integration efforts.", "date": "2025-02-19", "categories": []}
{"id": "2502.14778", "title": "Harnessing PDF Data for Improving Japanese Large Multimodal Models", "abstract": "Large Multimodal Models (LMMs) have demonstrated strong performance in English, but their effectiveness in Japanese remains limited due to the lack of high-quality training data. Current Japanese LMMs often rely on translated English datasets, restricting their ability to capture Japan-specific cultural knowledge. To address this, we explore the potential of Japanese PDF data as a training resource, an area that remains largely underutilized. We introduce a fully automated pipeline that leverages pretrained models to extract image-text pairs from PDFs through layout analysis, OCR, and vision-language pairing, removing the need for manual annotation. Additionally, we construct instruction data from extracted image-text pairs to enrich the training data. To evaluate the effectiveness of PDF-derived data, we train Japanese LMMs and assess their performance on the Japanese LMM Benchmark. Our results demonstrate substantial improvements, with performance gains ranging from 2.1% to 13.8% on Heron-Bench. Further analysis highlights the impact of PDF-derived data on various factors, such as model size and language models, reinforcing its value as a multimodal resource for Japanese LMMs.", "date": "2025-02-20", "categories": []}
{"id": "2512.12976", "title": "Authors Should Label Their Own Documents", "abstract": "Third-party annotation is the status quo for labeling text, but egocentric information such as sentiment and belief can at best only be approximated by a third-person proxy. We introduce author labeling, an annotation technique where the writer of the document itself annotates the data at the moment of creation. We collaborate with a commercial chatbot with over 20,000 users to deploy an author labeling annotation system. This system identifies task-relevant queries, generates on-the-fly labeling questions, and records authors' answers in real time. We train and deploy an online-learning model architecture for product recommendation with author-labeled data to improve performance. We train our model to minimize the prediction error on questions generated for a set of predetermined subjective beliefs using author-labeled responses. Our model achieves a 537% improvement in click-through rate compared to an industry advertising baseline running concurrently. We then compare the quality and practicality of author labeling to three traditional annotation approaches for sentiment analysis and find author labeling to be higher quality, faster to acquire, and cheaper. These findings reinforce existing literature that annotations, especially for egocentric and subjective beliefs, are significantly higher quality when labeled by the author rather than a third party. To facilitate broader scientific adoption, we release an author labeling service for the research community at https://academic.echogroup.ai.", "date": "2025-12-15", "categories": []}
{"id": "2601.02727", "title": "Foreground-Aware Dataset Distillation via Dynamic Patch Selection", "abstract": "In this paper, we propose a foreground-aware dataset distillation method that enhances patch selection in a content-adaptive manner. With the rising computational cost of training large-scale deep models, dataset distillation has emerged as a promising approach for constructing compact synthetic datasets that retain the knowledge of their large original counterparts. However, traditional optimization-based methods often suffer from high computational overhead, memory constraints, and the generation of unrealistic, noise-like images with limited architectural generalization. Recent non-optimization methods alleviate some of these issues by constructing distilled data from real image patches, but the used rigid patch selection strategies can still discard critical information about the main objects. To solve this problem, we first leverage Grounded SAM2 to identify foreground objects and compute per-image foreground occupancy, from which we derive a category-wise patch decision threshold. Guided by these thresholds, we design a dynamic patch selection strategy that, for each image, either selects the most informative patch from multiple candidates or directly resizes the full image when the foreground dominates. This dual-path mechanism preserves more key information about the main objects while reducing redundant background content. Extensive experiments on multiple benchmarks show that the proposed method consistently improves distillation performance over existing approaches, producing more informative and representative distilled datasets and enhancing robustness across different architectures and image compositions.", "date": "2026-01-06", "categories": []}
{"id": "2601.03011", "title": "ReCCur: A Recursive Corner-Case Curation Framework for Robust Vision-Language Understanding in Open and Edge Scenarios", "abstract": "Corner cases are rare or extreme scenarios that drive real-world failures, but they are difficult to curate at scale: web data are noisy, labels are brittle, and edge deployments preclude large retraining. We present ReCCur (Recursive Corner-Case Curation), a low-compute framework that converts noisy web imagery into auditable fine-grained labels via a multi-agent recursive pipeline. First, large-scale data acquisition and filtering expands a domain vocabulary with a vision-language model (VLM), crawls the web, and enforces tri-modal (image, description, keyword) consistency with light human spot checks to yield refined candidates. Next, mixture-of-experts knowledge distillation uses complementary encoders (e.g., CLIP, DINOv2, BEiT) for kNN voting with dual-confidence activation and uncertainty sampling, converging to a high-precision set. Finally, region-evidence VLM adversarial labeling pairs a proposer (multi-granularity regions and semantic cues) with a validator (global and local chained consistency) to produce explainable labels and close the loop. On realistic corner-case scenarios (e.g., flooded-car inspection), ReCCur runs on consumer-grade GPUs, steadily improves purity and separability, and requires minimal human supervision, providing a practical substrate for downstream training and evaluation under resource constraints. Code and dataset will be released.", "date": "2026-01-06", "categories": []}
{"id": "2411.12115", "title": "Distill the Best, Ignore the Rest: Improving Dataset Distillation with Loss-Value-Based Pruning", "abstract": "Dataset distillation has gained significant interest in recent years, yet existing approaches typically distill from the entire dataset, potentially including non-beneficial samples. We introduce a novel \"Prune First, Distill After\" framework that systematically prunes datasets via loss-based sampling prior to distillation. By leveraging pruning before classical distillation techniques and generative priors, we create a representative core-set that leads to enhanced generalization for unseen architectures - a significant challenge of current distillation methods. More specifically, our proposed framework significantly boosts distilled quality, achieving up to a 5.2 percentage points accuracy increase even with substantial dataset pruning, i.e., removing 80% of the original dataset prior to distillation. Overall, our experimental results highlight the advantages of our easy-sample prioritization and cross-architecture robustness, paving the way for more effective and high-quality dataset distillation.", "date": "2024-11-18", "categories": []}
{"id": "2512.23239", "title": "RS-Prune: Training-Free Data Pruning at High Ratios for Efficient Remote Sensing Diffusion Foundation Models", "abstract": "Diffusion-based remote sensing (RS) generative foundation models are cruial for downstream tasks. However, these models rely on large amounts of globally representative data, which often contain redundancy, noise, and class imbalance, reducing training efficiency and preventing convergence. Existing RS diffusion foundation models typically aggregate multiple classification datasets or apply simplistic deduplication, overlooking the distributional requirements of generation modeling and the heterogeneity of RS imagery. To address these limitations, we propose a training-free, two-stage data pruning approach that quickly select a high-quality subset under high pruning ratios, enabling a preliminary foundation model to converge rapidly and serve as a versatile backbone for generation, downstream fine-tuning, and other applications. Our method jointly considers local information content with global scene-level diversity and representativeness. First, an entropy-based criterion efficiently removes low-information samples. Next, leveraging RS scene classification datasets as reference benchmarks, we perform scene-aware clustering with stratified sampling to improve clustering effectiveness while reducing computational costs on large-scale unlabeled data. Finally, by balancing cluster-level uniformity and sample representativeness, the method enables fine-grained selection under high pruning ratios while preserving overall diversity and representativeness. Experiments show that, even after pruning 85\\% of the training data, our method significantly improves convergence and generation quality. Furthermore, diffusion foundation models trained with our method consistently achieve state-of-the-art performance across downstream tasks, including super-resolution and semantic image synthesis. This data pruning paradigm offers practical guidance for developing RS generative foundation models.", "date": "2025-12-29", "categories": []}
{"id": "2512.13083", "title": "DiRe: Diversity-promoting Regularization for Dataset Condensation", "abstract": "In Dataset Condensation, the goal is to synthesize a small dataset that replicates the training utility of a large original dataset. Existing condensation methods synthesize datasets with significant redundancy, so there is a dire need to reduce redundancy and improve the diversity of the synthesized datasets. To tackle this, we propose an intuitive Diversity Regularizer (DiRe) composed of cosine similarity and Euclidean distance, which can be applied off-the-shelf to various state-of-the-art condensation methods. Through extensive experiments, we demonstrate that the addition of our regularizer improves state-of-the-art condensation methods on various benchmark datasets from CIFAR-10 to ImageNet-1K with respect to generalization and diversity metrics.", "date": "2025-12-15", "categories": []}
{"id": "2404.19557", "title": "Neural Dynamic Data Valuation: A Stochastic Optimal Control Approach", "abstract": "Data valuation has become a cornerstone of the modern data economy, where datasets function as tradable intellectual assets that drive decision-making, model training, and market transactions. Despite substantial progress, existing valuation methods remain limited by high computational cost, weak fairness guarantees, and poor interpretability, which hinder their deployment in large-scale, high-stakes applications. This paper introduces Neural Dynamic Data Valuation (NDDV), a new framework that formulates data valuation as a stochastic optimal control problem to capture the dynamic evolution of data utility over time. Unlike static combinatorial approaches, NDDV models data interactions through continuous trajectories that reflect both individual and collective learning dynamics.", "date": "2024-04-30", "categories": []}
{"id": "2512.10952", "title": "Hierarchical Dataset Selection for High-Quality Data Sharing", "abstract": "The success of modern machine learning hinges on access to high-quality training data. In many real-world scenarios, such as acquiring data from public repositories or sharing across institutions, data is naturally organized into discrete datasets that vary in relevance, quality, and utility. Selecting which repositories or institutions to search for useful datasets, and which datasets to incorporate into model training are therefore critical decisions, yet most existing methods select individual samples and treat all data as equally relevant, ignoring differences between datasets and their sources. In this work, we formalize the task of dataset selection: selecting entire datasets from a large, heterogeneous pool to improve downstream performance under resource constraints. We propose Dataset Selection via Hierarchies (DaSH), a dataset selection method that models utility at both dataset and group (e.g., collections, institutions) levels, enabling efficient generalization from limited observations. Across two public benchmarks (Digit-Five and DomainNet), DaSH outperforms state-of-the-art data selection baselines by up to 26.2% in accuracy, while requiring significantly fewer exploration steps. Ablations show DaSH is robust to low-resource settings and lack of relevant datasets, making it suitable for scalable and adaptive dataset selection in practical multi-source learning workflows.", "date": "2025-12-11", "categories": []}
{"id": "2512.21152", "title": "MODE: Multi-Objective Adaptive Coreset Selection", "abstract": "We present Mode(Multi-Objective adaptive Data Efficiency), a framework that dynamically combines coreset selection strategies based on their evolving contribution to model performance. Unlike static methods, \\mode adapts selection criteria to training phases: emphasizing class balance early, diversity during representation learning, and uncertainty at convergence. We show that MODE achieves (1-1/e)-approximation with O(n \\log n) complexity and demonstrates competitive accuracy while providing interpretable insights into data utility evolution. Experiments show \\mode reduces memory requirements", "date": "2025-12-24", "categories": []}
{"id": "2505.17266", "title": "Select2Reason: Efficient Instruction-Tuning Data Selection for Long-CoT Reasoning", "abstract": "A practical approach to activate long chain-of-thoughts reasoning ability in pre-trained large language models is to perform supervised fine-tuning on instruction datasets synthesized by strong Large Reasoning Models such as DeepSeek-R1, offering a cost-effective alternative to reinforcement learning. However, large-scale instruction sets with more than 100k samples incur significant training overhead, while effective strategies for automatic long-CoT instruction selection still remain unexplored. In this work, we propose Select2Reason, a novel and efficient instruction-tuning data selection framework for long-CoT reasoning. From the perspective of emergence of rethinking behaviors like self-correction and backtracking, we investigate common metrics that may determine the quality of long-CoT reasoning instructions. Select2Reason leverages a quantifier to estimate difficulty of question and jointly incorporates a reasoning trace length-based heuristic through a weighted scheme for ranking to prioritize high-utility examples. Empirical results on OpenR1-Math-220k demonstrate that fine-tuning LLM on only 10% of the data selected by Select2Reason achieves performance competitive with or superior to full-data tuning and open-source baseline OpenR1-Qwen-7B across three competition-level and six comprehensive mathematical benchmarks. Further experiments highlight the scalability in varying data size, efficiency during inference, and its adaptability to other instruction pools with minimal cost.", "date": "2025-05-22", "categories": []}
{"id": "2512.20577", "title": "Improving ML Training Data with Gold-Standard Quality Metrics", "abstract": "Hand-tagged training data is essential to many machine learning tasks. However, training data quality control has received little attention in the literature, despite data quality varying considerably with the tagging exercise. We propose methods to evaluate and enhance the quality of hand-tagged training data using statistical approaches to measure tagging consistency and agreement. We show that agreement metrics give more reliable results if recorded over multiple iterations of tagging, where declining variance in such recordings is an indicator of increasing data quality. We also show one way a tagging project can collect high-quality training data without requiring multiple tags for every work item, and that a tagger burn-in period may not be sufficient for minimizing tagger errors.", "date": "2025-12-23", "categories": []}
{"id": "2512.19363", "title": "From Points to Coalitions: Hierarchical Contrastive Shapley Values for Prioritizing Data Samples", "abstract": "How should we quantify the value of each training example when datasets are large, heterogeneous, and geometrically structured? Classical Data-Shapley answers in principle, but its O(n!) complexity and point-wise perspective are ill-suited to modern scales. We propose Hierarchical Contrastive Data Valuation (HCDV), a three-stage framework that (i) learns a contrastive, geometry-preserving representation, (ii) organizes the data into a balanced coarse-to-fine hierarchy of clusters, and (iii) assigns Shapley-style payoffs to coalitions via local Monte-Carlo games whose budgets are propagated downward. HCDV collapses the factorial burden to O(T sum_{l} K_{l}) = O(T K_max log n), rewards examples that sharpen decision boundaries, and regularizes outliers through curvature-based smoothness. We prove that HCDV approximately satisfies the four Shapley axioms with surplus loss O(eta log n), enjoys sub-Gaussian coalition deviation tilde O(1/sqrt{T}), and incurs at most k epsilon_infty regret for top-k selection. Experiments on four benchmarks--tabular, vision, streaming, and a 45M-sample CTR task--plus the OpenDataVal suite show that HCDV lifts accuracy by up to +5 pp, slashes valuation time by up to 100x, and directly supports tasks such as augmentation filtering, low-latency streaming updates, and fair marketplace payouts.", "date": "2025-12-22", "categories": []}
{"id": "2512.15675", "title": "Stylized Synthetic Augmentation further improves Corruption Robustness", "abstract": "This paper proposes a training data augmentation pipeline that combines synthetic image data with neural style transfer in order to address the vulnerability of deep vision models to common corruptions. We show that although applying style transfer on synthetic images degrades their quality with respect to the common Frechet Inception Distance (FID) metric, these images are surprisingly beneficial for model training. We conduct a systematic empirical analysis of the effects of both augmentations and their key hyperparameters on the performance of image classifiers. Our results demonstrate that stylization and synthetic data complement each other well and can be combined with popular rule-based data augmentation techniques such as TrivialAugment, while not working with others. Our method achieves state-of-the-art corruption robustness on several small-scale image classification benchmarks, reaching 93.54%, 74.9% and 50.86% robust accuracy on CIFAR-10-C, CIFAR-100-C and TinyImageNet-C, respectively", "date": "2025-12-17", "categories": []}
{"id": "2501.15108", "title": "Knowledge Hierarchy Guided Biological-Medical Dataset Distillation for Domain LLM Training", "abstract": "The rapid advancement of large language models (LLMs) in biological-medical applications has highlighted a gap between their potential and the limited scale and often low quality of available open-source annotated textual datasets. In addition, the inherent complexity of the biomedical knowledge hierarchy significantly hampers efforts to bridge this gap.Can LLMs themselves play a pivotal role in overcoming this limitation? Motivated by this question, we investigate this challenge in the present study.We propose a framework that automates the distillation of high-quality textual training data from the extensive scientific literature. Our approach self-evaluates and generates questions that are more closely aligned with the biomedical domain, guided by the biomedical knowledge hierarchy through medical subject headings (MeSH). This comprehensive framework establishes an automated workflow, thereby eliminating the need for manual intervention. Furthermore, we conducted comprehensive experiments to evaluate the impact of our framework-generated data on downstream language models of varying sizes. Our approach substantially improves question-answering tasks compared to pre-trained models from the life sciences domain and powerful close-source models represented by GPT-4. Notably, the generated AI-Ready dataset enabled the Llama3-70B base model to outperform GPT-4 using MedPrompt with multiple times the number of parameters. Detailed case studies and ablation experiments underscore the significance of each component within our framework", "date": "2025-01-25", "categories": []}
{"id": "2505.01225", "title": "Core-Set Selection for Data-efficient Land Cover Segmentation", "abstract": "The increasing accessibility of remotely sensed data and their potential to support large-scale decision-making have driven the development of deep learning models for many Earth Observation tasks. Traditionally, such models rely on large datasets. However, the common assumption that larger training datasets lead to better performance tends to overlook issues related to data redundancy, noise, and the computational cost of processing massive datasets. Effective solutions must therefore consider not only the quantity but also the quality of data. Towards this, in this paper, we introduce six basic core-set selection approaches -- that rely on imagery only, labels only, or a combination of both -- and investigate whether they can identify high-quality subsets of data capable of maintaining -- or even surpassing -- the performance achieved when using full datasets for remote sensing semantic segmentation. We benchmark such approaches against two traditional baselines on three widely used land-cover classification datasets (DFC2022, Vaihingen, and Potsdam) using two different architectures (SegFormer and U-Net), thus establishing a general baseline for future works. Our experiments show that all proposed methods consistently outperform the baselines across multiple subset sizes, with some approaches even selecting core sets that surpass training on all available data. Notably, on DFC2022, a selected subset comprising only 25% of the training data yields slightly higher SegFormer performance than training with the entire dataset. This result shows the importance and potential of data-centric learning for the remote sensing domain. The code is available at https://github.com/keillernogueira/data-centric-rs-classification/.", "date": "2025-05-02", "categories": []}
{"id": "2511.06463", "title": "Error Estimate and Convergence Analysis for Data Valuation", "abstract": "Data valuation quantifies data importance, but existing methods cannot ensure validity in a single training process. The neural dynamic data valuation (NDDV) method [3] addresses this limitation. Based on NDDV, we are the first to explore error estimation and convergence analysis in data valuation. Under Lipschitz and smoothness assumptions, we derive quadratic error bounds for loss differences that scale inversely with time steps and quadratically with control variations, ensuring stability. We also prove that the expected squared gradient norm for the training loss vanishes asymptotically, and that the meta loss converges sublinearly over iterations. In particular, NDDV achieves sublinear convergence.", "date": "2025-11-09", "categories": []}
{"id": "2512.16676", "title": "DataFlow: An LLM-Driven Framework for Unified Data Preparation and Workflow Automation in the Era of Data-Centric AI", "abstract": "The rapidly growing demand for high-quality data in Large Language Models (LLMs) has intensified the need for scalable, reliable, and semantically rich data preparation pipelines. However, current practices remain dominated by ad-hoc scripts and loosely specified workflows, which lack principled abstractions, hinder reproducibility, and offer limited support for model-in-the-loop data generation. To address these challenges, we present DataFlow, a unified and extensible LLM-driven data preparation framework. DataFlow is designed with system-level abstractions that enable modular, reusable, and composable data transformations, and provides a PyTorch-style pipeline construction API for building debuggable and optimizable dataflows. The framework consists of nearly 200 reusable operators and six domain-general pipelines spanning text, mathematical reasoning, code, Text-to-SQL, agentic RAG, and large-scale knowledge extraction. To further improve usability, we introduce DataFlow-Agent, which automatically translates natural-language specifications into executable pipelines via operator synthesis, pipeline planning, and iterative verification. Across six representative use cases, DataFlow consistently improves downstream LLM performance. Our math, code, and text pipelines outperform curated human datasets and specialized synthetic baselines, achieving up to +3\\% execution accuracy in Text-to-SQL over SynSQL, +7\\% average improvements on code benchmarks, and 1--3 point gains on MATH, GSM8K, and AIME. Moreover, a unified 10K-sample dataset produced by DataFlow enables base models to surpass counterparts trained on 1M Infinity-Instruct data. These results demonstrate that DataFlow provides a practical and high-performance substrate for reliable, reproducible, and scalable LLM data preparation, and establishes a system-level foundation for future data-centric AI development.", "date": "2025-12-18", "categories": []}
{"id": "2512.16905", "title": "Alchemist: Unlocking Efficiency in Text-to-Image Model Training via Meta-Gradient Data Selection", "abstract": "Recent advances in Text-to-Image (T2I) generative models, such as Imagen, Stable Diffusion, and FLUX, have led to remarkable improvements in visual quality. However, their performance is fundamentally limited by the quality of training data. Web-crawled and synthetic image datasets often contain low-quality or redundant samples, which lead to degraded visual fidelity, unstable training, and inefficient computation. Hence, effective data selection is crucial for improving data efficiency. Existing approaches rely on costly manual curation or heuristic scoring based on single-dimensional features in Text-to-Image data filtering. Although meta-learning based method has been explored in LLM, there is no adaptation for image modalities. To this end, we propose **Alchemist**, a meta-gradient-based framework to select a suitable subset from large-scale text-image data pairs. Our approach automatically learns to assess the influence of each sample by iteratively optimizing the model from a data-centric perspective. Alchemist consists of two key stages: data rating and data pruning. We train a lightweight rater to estimate each sample's influence based on gradient information, enhanced with multi-granularity perception. We then use the Shift-Gsampling strategy to select informative subsets for efficient model training. Alchemist is the first automatic, scalable, meta-gradient-based data selection framework for Text-to-Image model training. Experiments on both synthetic and web-crawled datasets demonstrate that Alchemist consistently improves visual quality and downstream performance. Training on an Alchemist-selected 50% of the data can outperform training on the full dataset.", "date": "2025-12-18", "categories": []}
{"id": "2512.13704", "title": "Adjudicator: Correcting Noisy Labels with a KG-Informed Council of LLM Agents", "abstract": "The performance of production machine learning systems is fundamentally limited by the quality of their training data. In high-stakes industrial applications, noisy labels can degrade performance and erode user trust. This paper presents Adjudicator, a system that addresses the critical data mining challenge of automatically identifying and correcting label noise and has been validated for production deployment. Adjudicator models this as a neuro-symbolic task, first constructing a dynamic Knowledge Graph (KG) to unify item context. This KG then informs a \"Council of Agents,\" a novel multi-agent Large Language Model architecture where specialized agents debate and vote on a label's validity. We validate our system on a 1,000-item balanced subset of the AlleNoise benchmark. Our KG-informed model achieves a 0.99 F1-score, significantly outperforming a single-LLM baseline (0.48 F1) and a non-KG council (0.59 F1). Our analysis reveals this is due to a Precision, achieved by a novel override logic that uses the KG to perfectly identify complex, structural errors (complete Recall) -- a class of errors that baselines fail to find. This result demonstrates a robust and explainable system for automated, high-precision data verification, serving as a vital proof-of-concept for generating golden datasets in strictly governed industrial environments.", "date": "2025-12-05", "categories": []}
{"id": "2512.14051", "title": "OpenDataArena: A Fair and Open Arena for Benchmarking Post-Training Dataset Value", "abstract": "The rapid evolution of Large Language Models (LLMs) is predicated on the quality and diversity of post-training datasets. However, a critical dichotomy persists: while models are rigorously benchmarked, the data fueling them remains a black box--characterized by opaque composition, uncertain provenance, and a lack of systematic evaluation. This opacity hinders reproducibility and obscures the causal link between data characteristics and model behaviors. To bridge this gap, we introduce OpenDataArena (ODA), a holistic and open platform designed to benchmark the intrinsic value of post-training data. ODA establishes a comprehensive ecosystem comprising four key pillars: (i) a unified training-evaluation pipeline that ensures fair, open comparisons across diverse models (e.g., Llama, Qwen) and domains; (ii) a multi-dimensional scoring framework that profiles data quality along tens of distinct axes; (iii) an interactive data lineage explorer to visualize dataset genealogy and dissect component sources; and (iv) a fully open-source toolkit for training, evaluation, and scoring to foster data research. Extensive experiments on ODA--covering over 120 training datasets across multiple domains on 22 benchmarks, validated by more than 600 training runs and 40 million processed data points--reveal non-trivial insights. Our analysis uncovers the inherent trade-offs between data complexity and task performance, identifies redundancy in popular benchmarks through lineage tracing, and maps the genealogical relationships across datasets. We release all results, tools, and configurations to democratize access to high-quality data evaluation. Rather than merely expanding a leaderboard, ODA envisions a shift from trial-and-error data curation to a principled science of Data-Centric AI, paving the way for rigorous studies on data mixing laws and the strategic composition of foundation models.", "date": "2025-12-16", "categories": []}
{"id": "2512.14230", "title": "Understanding the Gain from Data Filtering in Multimodal Contrastive Learning", "abstract": "The success of modern multimodal representation learning relies on internet-scale datasets. Due to the low quality of a large fraction of raw web data, data curation has become a critical step in the training pipeline. Filtering using a trained model (i.e., teacher-based filtering) has emerged as a successful solution, leveraging a pre-trained model to compute quality scores. To explain the empirical success of teacher-based filtering, we characterize the performance of filtered contrastive learning under the standard bimodal data generation model. Denoting $\\eta\\in(0,1]$ as the fraction of data with correctly matched modalities among $n$ paired samples, we utilize a linear contrastive learning setup to show a provable benefit of data filtering: $(i)$ the error without filtering is upper and lower bounded by $\\frac{1}{\\eta \\sqrt{n}}$, and $(ii)$ the error with teacher-based filtering is upper bounded by $\\frac{1}{\\sqrt{\\eta n}}$ in the large $\\eta$ regime, and by $\\frac{1}{\\sqrt{n}}$ in the small $\\eta$ regime.", "date": "2025-12-16", "categories": []}
{"id": "2512.14621", "title": "Distill Video Datasets into Images", "abstract": "Dataset distillation aims to synthesize compact yet informative datasets that allow models trained on them to achieve performance comparable to training on the full dataset. While this approach has shown promising results for image data, extending dataset distillation methods to video data has proven challenging and often leads to suboptimal performance. In this work, we first identify the core challenge in video set distillation as the substantial increase in learnable parameters introduced by the temporal dimension of video, which complicates optimization and hinders convergence. To address this issue, we observe that a single frame is often sufficient to capture the discriminative semantics of a video. Leveraging this insight, we propose Single-Frame Video set Distillation (SFVD), a framework that distills videos into highly informative frames for each class. Using differentiable interpolation, these frames are transformed into video sequences and matched with the original dataset, while updates are restricted to the frames themselves for improved optimization efficiency. To further incorporate temporal information, the distilled frames are combined with sampled real videos from real videos during the matching process through a channel reshaping layer. Extensive experiments on multiple benchmarks demonstrate that SFVD substantially outperforms prior methods, achieving improvements of up to 5.3% on MiniUCF, thereby offering a more effective solution.", "date": "2025-12-16", "categories": []}
{"id": "2503.22154", "title": "Dataset Distillation of 3D Point Clouds via Distribution Matching", "abstract": "Large-scale datasets are usually required to train deep neural networks, but it increases the computational complexity hindering the practical applications. Recently, dataset distillation for images and texts has been attracting a lot of attention, that reduces the original dataset to a synthetic dataset to alleviate the computational burden of training while preserving essential task-relevant information. However, the dataset distillation for 3D point clouds remains largely unexplored, as the point clouds exhibit fundamentally different characteristics from that of images, making the dataset distillation more challenging. In this paper, we propose a distribution matching-based distillation framework for 3D point clouds that jointly optimizes the geometric structures as well as the orientations of the synthetic 3D objects. To address the semantic misalignment caused by unordered indexing of points, we introduce a Semantically Aligned Distribution Matching loss computed on the sorted features in each channel. Moreover, to address the rotation variation, we jointly learn the optimal rotation angles while updating the synthetic dataset to better align with the original feature distribution. Extensive experiments on widely used benchmark datasets demonstrate that the proposed method consistently outperforms existing dataset distillation methods, achieving superior accuracy and strong cross-architecture generalization.", "date": "2025-03-28", "categories": []}
{"id": "2511.18858", "title": "Rethinking Long-tailed Dataset Distillation: A Uni-Level Framework with Unbiased Recovery and Relabeling", "abstract": "Dataset distillation creates a small distilled set that enables efficient training by capturing key information from the full dataset. While existing dataset distillation methods perform well on balanced datasets, they struggle under long-tailed distributions, where imbalanced class frequencies induce biased model representations and corrupt statistical estimates such as Batch Normalization (BN) statistics. In this paper, we rethink long-tailed dataset distillation by revisiting the limitations of trajectory-based methods, and instead adopt the statistical alignment perspective to jointly mitigate model bias and restore fair supervision. To this end, we introduce three dedicated components that enable unbiased recovery of distilled images and soft relabeling: (1) enhancing expert models (an observer model for recovery and a teacher model for relabeling) to enable reliable statistics estimation and soft-label generation; (2) recalibrating BN statistics via a full forward pass with dynamically adjusted momentum to reduce representation skew; (3) initializing synthetic images by incrementally selecting high-confidence and diverse augmentations via a multi-round mechanism that promotes coverage and diversity. Extensive experiments on four long-tailed benchmarks show consistent improvements over state-of-the-art methods across varying degrees of class imbalance. Notably, our approach improves top-1 accuracy by 15.6% on CIFAR-100-LT and 11.8% on Tiny-ImageNet-LT under IPC=10 and IF=10. Codes are available at https://github.com/2018cx/RLDD.", "date": "2025-11-24", "categories": []}
{"id": "2512.09422", "title": "InfoMotion: A Graph-Based Approach to Video Dataset Distillation for Echocardiography", "abstract": "Echocardiography plays a critical role in the diagnosis and monitoring of cardiovascular diseases as a non-invasive real-time assessment of cardiac structure and function. However, the growing scale of echocardiographic video data presents significant challenges in terms of storage, computation, and model training efficiency. Dataset distillation offers a promising solution by synthesizing a compact, informative subset of data that retains the key clinical features of the original dataset. In this work, we propose a novel approach for distilling a compact synthetic echocardiographic video dataset. Our method leverages motion feature extraction to capture temporal dynamics, followed by class-wise graph construction and representative sample selection using the Infomap algorithm. This enables us to select a diverse and informative subset of synthetic videos that preserves the essential characteristics of the original dataset. We evaluate our approach on the EchoNet-Dynamic datasets and achieve a test accuracy of \\(69.38\\%\\) using only \\(25\\) synthetic videos. These results demonstrate the effectiveness and scalability of our method for medical video dataset distillation.", "date": "2025-12-10", "categories": []}
{"id": "2512.12508", "title": "Generative Spatiotemporal Data Augmentation", "abstract": "We explore spatiotemporal data augmentation using video foundation models to diversify both camera viewpoints and scene dynamics. Unlike existing approaches based on simple geometric transforms or appearance perturbations, our method leverages off-the-shelf video diffusion models to generate realistic 3D spatial and temporal variations from a given image dataset. Incorporating these synthesized video clips as supplemental training data yields consistent performance gains in low-data settings, such as UAV-captured imagery where annotations are scarce. Beyond empirical improvements, we provide practical guidelines for (i) choosing an appropriate spatiotemporal generative setup, (ii) transferring annotations to synthetic frames, and (iii) addressing disocclusion - regions newly revealed and unlabeled in generated views. Experiments on COCO subsets and UAV-captured datasets show that, when applied judiciously, spatiotemporal augmentation broadens the data distribution along axes underrepresented by traditional and prior generative methods, offering an effective lever for improving model performance in data-scarce regimes.", "date": "2025-12-14", "categories": []}
{"id": "2512.13279", "title": "AIR: Post-training Data Selection for Reasoning via Attention Head Influence", "abstract": "LLMs achieve remarkable multi-step reasoning capabilities, yet effectively transferring these skills via post-training distillation remains challenging. Existing data selection methods, ranging from manual curation to heuristics based on length, entropy, or overall loss, fail to capture the causal importance of individual reasoning steps, limiting distillation efficiency. To address this, we propose Attention Influence for Reasoning (AIR), a principled, unsupervised and training-free framework that leverages mechanistic insights of the retrieval head to select high-value post-training data. AIR first identifies reasoning-critical attention heads of an off-the-shelf model, then constructs a weakened reference model with disabled head influence, and finally quantifies the resulting loss divergence as the Attention Influence Score. This score enables fine-grained assessment at both the step and sample levels, supporting step-level weighted fine-tuning and global sample selection. Experiments across multiple reasoning benchmarks show that AIR consistently improves reasoning accuracy, surpassing heuristic baselines and effectively isolating the most critical steps and samples. Our work establishes a mechanism-driven, data-efficient approach for reasoning distillation in LLMs.", "date": "2025-12-15", "categories": []}
{"id": "2512.10178", "title": "CIEGAD: Cluster-Conditioned Interpolative and Extrapolative Framework for Geometry-Aware and Domain-Aligned Data Augmentation", "abstract": "In practical deep learning deployment, the scarcity of data and the imbalance of label distributions often lead to semantically uncovered regions within the real-world data distribution, hindering model training and causing misclassification near class boundaries as well as unstable behaviors in peripheral areas. Although recent large language models (LLMs) show promise for data augmentation, an integrated framework that simultaneously achieves directional control of generation, domain alignment, and quality control has not yet been fully established. To address these challenges, we propose a Cluster-conditioned Interpolative and Extrapolative framework for Geometry-Aware and Domain-aligned data augmentation (CIEGAD), which systematically complements both in-distribution and out-of-distribution semantically uncovered regions. CIEGAD constructs domain profiles through cluster conditioning, allocates generation with a hierarchical frequency-geometric allocation integrating class frequency and geometric indicators, and finely controls generation directions via the coexistence of interpolative and extrapolative synthesis. It further performs quality control through geometry-constrained filtering combined with an LLM-as-a-Judge mechanism. Experiments on multiple classification tasks demonstrate that CIEGAD effectively extends the periphery of real-world data distributions while maintaining high alignment between generated and real-world data as well as semantic diversity. In particular, for long-tailed and multi-class classification tasks, CIEGAD consistently improves F1 and recall, validating the triple harmony of distributional consistency, diversity, and quality. These results indicate that CIEGAD serves as a practically oriented data augmentation framework that complements underrepresented regions while preserving alignment with real-world data.", "date": "2025-12-11", "categories": []}
{"id": "2512.05817", "title": "Utility Boundary of Dataset Distillation: Scaling and Configuration-Coverage Laws", "abstract": "Dataset distillation (DD) aims to construct compact synthetic datasets that allow models to achieve comparable performance to full-data training while substantially reducing storage and computation. Despite rapid empirical progress, its theoretical foundations remain limited: existing methods (gradient, distribution, trajectory matching) are built on heterogeneous surrogate objectives and optimization assumptions, which makes it difficult to analyze their common principles or provide general guarantees. Moreover, it is still unclear under what conditions distilled data can retain the effectiveness of full datasets when the training configuration, such as optimizer, architecture, or augmentation, changes. To answer these questions, we propose a unified theoretical framework, termed configuration--dynamics--error analysis, which reformulates major DD approaches under a common generalization-error perspective and provides two main results: (i) a scaling law that provides a single-configuration upper bound, characterizing how the error decreases as the distilled sample size increases and explaining the commonly observed performance saturation effect; and (ii) a coverage law showing that the required distilled sample size scales linearly with configuration diversity, with provably matching upper and lower bounds. In addition, our unified analysis reveals that various matching methods are interchangeable surrogates, reducing the same generalization error, clarifying why they can all achieve dataset distillation and providing guidance on how surrogate choices affect sample efficiency and robustness. Experiments across diverse methods and configurations empirically confirm the derived laws, advancing a theoretical foundation for DD and enabling theory-driven design of compact, configuration-robust dataset distillation.", "date": "2025-12-05", "categories": []}
{"id": "2512.08317", "title": "GeoDM: Geometry-aware Distribution Matching for Dataset Distillation", "abstract": "Dataset distillation aims to synthesize a compact subset of the original data, enabling models trained on it to achieve performance comparable to those trained on the original large dataset. Existing distribution-matching methods are confined to Euclidean spaces, making them only capture linear structures and overlook the intrinsic geometry of real data, e.g., curvature. However, high-dimensional data often lie on low-dimensional manifolds, suggesting that dataset distillation should have the distilled data manifold aligned with the original data manifold. In this work, we propose a geometry-aware distribution-matching framework, called \\textbf{GeoDM}, which operates in the Cartesian product of Euclidean, hyperbolic, and spherical manifolds, with flat, hierarchical, and cyclical structures all captured by a unified representation. To adapt to the underlying data geometry, we introduce learnable curvature and weight parameters for three kinds of geometries. At the same time, we design an optimal transport loss to enhance the distribution fidelity. Our theoretical analysis shows that the geometry-aware distribution matching in a product space yields a smaller generalization error bound than the Euclidean counterparts. Extensive experiments conducted on standard benchmarks demonstrate that our algorithm outperforms state-of-the-art data distillation methods and remains effective across various distribution-matching strategies for the single geometries.", "date": "2025-12-09", "categories": []}
{"id": "2505.06670", "title": "Video Dataset Condensation with Diffusion Models", "abstract": "In recent years, the rapid expansion of dataset sizes and the increasing complexity of deep learning models have significantly escalated the demand for computational resources, both for data storage and model training. Dataset distillation has emerged as a promising solution to address this challenge by generating a compact synthetic dataset that retains the essential information from a large real dataset. However, existing methods often suffer from limited performance, particularly in the video domain. In this paper, we focus on video dataset distillation. We begin by employing a video diffusion model to generate synthetic videos. Since the videos are generated only once, this significantly reduces computational costs. Next, we introduce the Video Spatio-Temporal U-Net (VST-UNet), a model designed to select a diverse and informative subset of videos that effectively captures the characteristics of the original dataset. To further optimize computational efficiency, we explore a training-free clustering algorithm, Temporal-Aware Cluster-based Distillation (TAC-DT), to select representative videos without requiring additional training overhead. We validate the effectiveness of our approach through extensive experiments on four benchmark datasets, demonstrating performance improvements of up to \\(10.61\\%\\) over the state-of-the-art. Our method consistently outperforms existing approaches across all datasets, establishing a new benchmark for video dataset distillation.", "date": "2025-05-10", "categories": []}
{"id": "2505.08605", "title": "Leveraging Multi-Modal Information to Enhance Dataset Distillation", "abstract": "Dataset distillation aims to create a small and highly representative synthetic dataset that preserves the essential information of a larger real dataset. Beyond reducing storage and computational costs, related approaches offer a promising avenue for privacy preservation in computer vision by eliminating the need to store or share sensitive real-world images. Existing methods focus solely on optimizing visual representations, overlooking the potential of multi-modal information. In this work, we propose a multi-modal dataset distillation framework that incorporates two key enhancements: caption-guided supervision and object-centric masking. To leverage textual information, we introduce two strategies: caption concatenation, which fuses caption embeddings with visual features during classification, and caption matching, which enforces semantic alignment between real and synthetic data through a caption-based loss. To improve data utility and reduce unnecessary background noise, we employ segmentation masks to isolate target objects and introduce two novel losses: masked feature alignment and masked gradient matching, both aimed at promoting object-centric learning. Extensive evaluations demonstrate that our approach improves downstream performance while promoting privacy protection by minimizing exposure to real data.", "date": "2025-05-13", "categories": []}
{"id": "2512.05987", "title": "Adaptive Dataset Quantization: A New Direction for Dataset Pruning", "abstract": "This paper addresses the challenges of storage and communication costs for large-scale datasets in resource-constrained edge devices by proposing a novel dataset quantization approach to reduce intra-sample redundancy. Unlike traditional dataset pruning and distillation methods that focus on inter-sample redundancy, the proposed method compresses each image by reducing redundant or less informative content within samples while preserving essential features. It first applies linear symmetric quantization to obtain an initial quantization range and scale for each sample. Then, an adaptive quantization allocation algorithm is introduced to distribute different quantization ratios for samples with varying precision requirements, maintaining a constant total compression ratio. The main contributions include: (1) being the first to use limited bits to represent datasets for storage reduction; (2) introducing a dataset-level quantization algorithm with adaptive ratio allocation; and (3) validating the method's effectiveness through extensive experiments on CIFAR-10, CIFAR-100, and ImageNet-1K. Results show that the method maintains model training performance while achieving significant dataset compression, outperforming traditional quantization and dataset pruning baselines under the same compression ratios.", "date": "2025-11-28", "categories": []}
{"id": "2409.02098", "title": "CRAFT Your Dataset: Task-Specific Synthetic Dataset Generation Through Corpus Retrieval and Augmentation", "abstract": "Building high-quality datasets for specialized tasks is a time-consuming and resource-intensive process that often requires specialized domain knowledge. We propose Corpus Retrieval and Augmentation for Fine-Tuning (CRAFT), a method for generating synthetic datasets, given a small number of user-written few-shots that demonstrate the task to be performed. Given these examples, CRAFT uses large-scale public web-crawled corpora and similarity-based document retrieval to find other relevant human-written documents. Lastly, instruction-tuned large language models (LLMs) augment the retrieved documents into custom-formatted task samples, which then can be used for fine-tuning. We demonstrate that CRAFT can efficiently generate large-scale task-specific training datasets for four diverse tasks: biology, medicine, and commonsense question-answering (QA), as well as summarization. Our experiments show that CRAFT-based models outperform or match general LLMs on QA tasks, while exceeding models trained on human-curated summarization data by 46 preference points. CRAFT outperforms other synthetic dataset generation methods such as Self- and Evol-Instruct, and remains robust even when the quality of the initial few-shots varies.", "date": "2024-09-03", "categories": []}
{"id": "2412.17961", "title": "Extending Graph Condensation to Multi-Label Datasets: A Benchmark Study", "abstract": "As graph data grows increasingly complicate, training graph neural networks (GNNs) on large-scale datasets presents significant challenges, including computational resource constraints, data redundancy, and transmission inefficiencies. While existing graph condensation techniques have shown promise in addressing these issues, they are predominantly designed for single-label datasets, where each node is associated with a single class label. However, many real-world applications, such as social network analysis and bioinformatics, involve multi-label graph datasets, where one node can have various related labels. To deal with this problem, we extends traditional graph condensation approaches to accommodate multi-label datasets by introducing modifications to synthetic dataset initialization and condensing optimization. Through experiments on eight real-world multi-label graph datasets, we prove the effectiveness of our method. In experiment, the GCond framework, combined with K-Center initialization and binary cross-entropy loss (BCELoss), achieves best performance in general. This benchmark for multi-label graph condensation not only enhances the scalability and efficiency of GNNs for multi-label graph data, but also offering substantial benefits for diverse real-world applications.", "date": "2024-12-23", "categories": []}
{"id": "2502.15177", "title": "Optimizing Product Provenance Verification using Data Valuation Methods", "abstract": "Determining and verifying product provenance remains a critical challenge in global supply chains, particularly as geopolitical conflicts and shifting borders create new incentives for misrepresentation of commodities, such as hiding the origin of illegally harvested timber or stolen agricultural products. Stable Isotope Ratio Analysis (SIRA), combined with Gaussian process regression-based isoscapes, has emerged as a powerful tool for geographic origin verification. While these models are now actively deployed in operational settings supporting regulators, certification bodies, and companies, they remain constrained by data scarcity and suboptimal dataset selection. In this work, we introduce a novel deployed data valuation framework designed to enhance the selection and utilization of training data for machine learning models applied in SIRA. By quantifying the marginal utility of individual samples using Shapley values, our method guides strategic, cost-effective, and robust sampling campaigns within active monitoring programs. By prioritizing high-informative samples, our approach improves model robustness and predictive accuracy across diverse datasets and geographies. Our framework has been implemented and validated in a live provenance verification system currently used by enforcement agencies, demonstrating tangible, real-world impact. Through extensive experiments and deployment in a live provenance verification system, we show that this system significantly enhances provenance verification, mitigates fraudulent trade practices, and strengthens regulatory enforcement of global supply chains.", "date": "2025-02-21", "categories": []}
{"id": "2512.03844", "title": "CoDA: From Text-to-Image Diffusion Models to Training-Free Dataset Distillation", "abstract": "Prevailing Dataset Distillation (DD) methods leveraging generative models confront two fundamental limitations. First, despite pioneering the use of diffusion models in DD and delivering impressive performance, the vast majority of approaches paradoxically require a diffusion model pre-trained on the full target dataset, undermining the very purpose of DD and incurring prohibitive training costs. Second, although some methods turn to general text-to-image models without relying on such target-specific training, they suffer from a significant distributional mismatch, as the web-scale priors encapsulated in these foundation models fail to faithfully capture the target-specific semantics, leading to suboptimal performance. To tackle these challenges, we propose Core Distribution Alignment (CoDA), a framework that enables effective DD using only an off-the-shelf text-to-image model. Our key idea is to first identify the \"intrinsic core distribution\" of the target dataset using a robust density-based discovery mechanism. We then steer the generative process to align the generated samples with this core distribution. By doing so, CoDA effectively bridges the gap between general-purpose generative priors and target semantics, yielding highly representative distilled datasets. Extensive experiments suggest that, without relying on a generative model specifically trained on the target dataset, CoDA achieves performance on par with or even superior to previous methods with such reliance across all benchmarks, including ImageNet-1K and its subsets. Notably, it establishes a new state-of-the-art accuracy of 60.4% at the 50-images-per-class (IPC) setup on ImageNet-1K. Our code is available on the project webpage: https://github.com/zzzlt422/CoDA", "date": "2025-12-03", "categories": []}
{"id": "2512.03967", "title": "Technical Report on Text Dataset Distillation", "abstract": "In the vision domain, dataset distillation arises as a technique to condense a large dataset into a smaller synthetic one that exhibits a similar result in the training process. While image data presents an extensive literature of distillation methods, text dataset distillation has fewer works in comparison. Text dataset distillation initially grew as an adaptation of efforts from the vision universe, as the particularities of the modality became clear obstacles, it rose into a separate branch of research. Several milestones mark the development of this area, such as the introduction of methods that use transformer models, the generation of discrete synthetic text, and the scaling to decoder-only models with over 1B parameters. Despite major advances in modern approaches, the field remains in a maturing phase, with room for improvement on benchmarking standardization, approaches to overcome the discrete nature of text, handling complex tasks, and providing explicit examples of real-world applications. In this report, we review past and recent advances in dataset distillation for text, highlighting different distillation strategies, key contributions, and general challenges.", "date": "2025-12-03", "categories": []}
{"id": "2512.02469", "title": "TGDD: Trajectory Guided Dataset Distillation with Balanced Distribution", "abstract": "Dataset distillation compresses large datasets into compact synthetic ones to reduce storage and computational costs. Among various approaches, distribution matching (DM)-based methods have attracted attention for their high efficiency. However, they often overlook the evolution of feature representations during training, which limits the expressiveness of synthetic data and weakens downstream performance. To address this issue, we propose Trajectory Guided Dataset Distillation (TGDD), which reformulates distribution matching as a dynamic alignment process along the model's training trajectory. At each training stage, TGDD captures evolving semantics by aligning the feature distribution between the synthetic and original dataset. Meanwhile, it introduces a distribution constraint regularization to reduce class overlap. This design helps synthetic data preserve both semantic diversity and representativeness, improving performance in downstream tasks. Without additional optimization overhead, TGDD achieves a favorable balance between performance and efficiency. Experiments on ten datasets demonstrate that TGDD achieves state-of-the-art performance, notably a 5.0% accuracy gain on high-resolution benchmarks.", "date": "2025-12-02", "categories": []}
{"id": "2506.01317", "title": "T-SHIRT: Token-Selective Hierarchical Data Selection for Instruction Tuning", "abstract": "Instruction tuning is essential for Large Language Models (LLMs) to effectively follow user instructions. To improve training efficiency and reduce data redundancy, recent works use LLM-based scoring functions, e.g., Instruction-Following Difficulty (IFD), to select high-quality instruction-tuning data with scores above a threshold. While these data selection methods often lead to models that can match or even exceed the performance of models trained on the full datasets, we identify two key limitations: (i) they assess quality at the sample level, ignoring token-level informativeness; and (ii) they overlook the robustness of the scoring method, often selecting a sample due to superficial lexical features instead of its true quality. In this work, we propose Token-Selective HIeRarchical Data Selection for Instruction Tuning (T-SHIRT), a novel data selection framework that introduces a new scoring method to include only informative tokens in quality evaluation and also promotes robust and reliable samples whose neighbors also show high quality with less local inconsistencies. We demonstrate that models instruction-tuned on a curated dataset (only 5% of the original size) using T-SHIRT can outperform those trained on the entire large-scale dataset by up to 5.48 points on average across eight benchmarks. Across various LLMs and training set scales, our method consistently surpasses existing state-of-the-art data selection techniques, while also remaining both cost-effective and highly efficient. For instance, by using GPT-2 for score computation, we are able to process a dataset of 52k samples in 40 minutes on a single GPU. Our code is available at https://github.com/Dynamite321/T-SHIRT.", "date": "2025-06-02", "categories": []}
{"id": "2510.06108", "title": "Influence Functions for Efficient Data Selection in Reasoning", "abstract": "Fine-tuning large language models (LLMs) on chain-of-thought (CoT) data shows that a small amount of high-quality data can outperform massive datasets. Yet, what constitutes \"quality\" remains ill-defined. Existing reasoning methods rely on indirect heuristics such as problem difficulty or trace length, while instruction-tuning has explored a broader range of automated selection strategies, but rarely in the context of reasoning. We propose to define reasoning data quality using influence functions, which measure the causal effect of individual CoT examples on downstream accuracy, and introduce influence-based pruning, which consistently outperforms perplexity and embedding-based baselines on math reasoning within a model family.", "date": "2025-10-07", "categories": []}
{"id": "2512.00130", "title": "Local and Global Context-and-Object-part-Aware Superpixel-based Data Augmentation for Deep Visual Recognition", "abstract": "Cutmix-based data augmentation, which uses a cut-and-paste strategy, has shown remarkable generalization capabilities in deep learning. However, existing methods primarily consider global semantics with image-level constraints, which excessively reduces attention to the discriminative local context of the class and leads to a performance improvement bottleneck. Moreover, existing methods for generating augmented samples usually involve cutting and pasting rectangular or square regions, resulting in a loss of object part information. To mitigate the problem of inconsistency between the augmented image and the generated mixed label, existing methods usually require double forward propagation or rely on an external pre-trained network for object centering, which is inefficient. To overcome the above limitations, we propose LGCOAMix, an efficient context-aware and object-part-aware superpixel-based grid blending method for data augmentation. To the best of our knowledge, this is the first time that a label mixing strategy using a superpixel attention approach has been proposed for cutmix-based data augmentation. It is the first instance of learning local features from discriminative superpixel-wise regions and cross-image superpixel contrasts. Extensive experiments on various benchmark datasets show that LGCOAMix outperforms state-of-the-art cutmix-based data augmentation methods on classification tasks, {and weakly supervised object location on CUB200-2011.} We have demonstrated the effectiveness of LGCOAMix not only for CNN networks, but also for Transformer networks. Source codes are available at https://github.com/DanielaPlusPlus/LGCOAMix.", "date": "2025-11-28", "categories": []}
{"id": "2512.00308", "title": "Optimizing Distributional Geometry Alignment with Optimal Transport for Generative Dataset Distillation", "abstract": "Dataset distillation seeks to synthesize a compact distilled dataset, enabling models trained on it to achieve performance comparable to models trained on the full dataset. Recent methods for large-scale datasets focus on matching global distributional statistics (e.g., mean and variance), but overlook critical instance-level characteristics and intraclass variations, leading to suboptimal generalization. We address this limitation by reformulating dataset distillation as an Optimal Transport (OT) distance minimization problem, enabling fine-grained alignment at both global and instance levels throughout the pipeline. OT offers a geometrically faithful framework for distribution matching. It effectively preserves local modes, intra-class patterns, and fine-grained variations that characterize the geometry of complex, high-dimensional distributions. Our method comprises three components tailored for preserving distributional geometry: (1) OT-guided diffusion sampling, which aligns latent distributions of real and distilled images; (2) label-image-aligned soft relabeling, which adapts label distributions based on the complexity of distilled image distributions; and (3) OT-based logit matching, which aligns the output of student models with soft-label distributions. Extensive experiments across diverse architectures and large-scale datasets demonstrate that our method consistently outperforms state-of-the-art approaches in an efficient manner, achieving at least 4% accuracy improvement under IPC=10 settings for each architecture on ImageNet-1K.", "date": "2025-11-29", "categories": []}
{"id": "2512.00536", "title": "Algorithmic Guarantees for Distilling Supervised and Offline RL Datasets", "abstract": "Given a training dataset, the goal of dataset distillation is to derive a synthetic dataset such that models trained on the latter perform as well as those trained on the training dataset. In this work, we develop and analyze an efficient dataset distillation algorithm for supervised learning, specifically regression in $\\mathbb{R}^d$, based on matching the losses on the training and synthetic datasets with respect to a fixed set of randomly sampled regressors without any model training. Our first key contribution is a novel performance guarantee proving that our algorithm needs only $\\tilde{O}(d^2)$ sampled regressors to derive a synthetic dataset on which the MSE loss of any bounded linear model is nearly the same as its MSE loss on the given training data. In particular, the model optimized on the synthetic data has close to minimum loss on the training data, thus performing nearly as well as the model optimized on the latter. Complementing this, we also prove a matching lower bound of $\\Omega(d^2)$ for the number of sampled regressors showing the tightness of our analysis.\n  Our second contribution is to extend our algorithm to offline RL dataset distillation by matching the Bellman loss, unlike previous works which used a behavioral cloning objective. This is the first such method which leverages both, the rewards and the next state information, available in offline RL datasets, without any policy model optimization. Our algorithm generates a synthetic dataset whose Bellman loss with respect to any linear action-value predictor is close to the latter's Bellman loss on the offline RL training dataset. Therefore, a policy associated with an action-value predictor optimized on the synthetic dataset performs nearly as well as that derived from the one optimized on the training data. We conduct experiments to validate our theoretical guarantees and observe performance gains.", "date": "2025-11-29", "categories": []}
{"id": "2512.02018", "title": "Data-Centric Visual Development for Self-Driving Labs", "abstract": "Self-driving laboratories offer a promising path toward reducing the labor-intensive, time-consuming, and often irreproducible workflows in the biological sciences. Yet their stringent precision requirements demand highly robust models whose training relies on large amounts of annotated data. However, this kind of data is difficult to obtain in routine practice, especially negative samples. In this work, we focus on pipetting, the most critical and precision sensitive action in SDLs. To overcome the scarcity of training data, we build a hybrid pipeline that fuses real and virtual data generation. The real track adopts a human-in-the-loop scheme that couples automated acquisition with selective human verification to maximize accuracy with minimal effort. The virtual track augments the real data using reference-conditioned, prompt-guided image generation, which is further screened and validated for reliability. Together, these two tracks yield a class-balanced dataset that enables robust bubble detection training. On a held-out real test set, a model trained entirely on automatically acquired real images reaches 99.6% accuracy, and mixing real and generated data during training sustains 99.4% accuracy while reducing collection and review load. Our approach offers a scalable and cost-effective strategy for supplying visual feedback data to SDL workflows and provides a practical solution to data scarcity in rare event detection and broader vision tasks.", "date": "2025-12-01", "categories": []}
{"id": "2405.17464", "title": "Data Valuation by Fusing Global and Local Statistical Information", "abstract": "Data valuation has garnered increasing attention in recent years, given the critical role of high-quality data in various applications. Among diverse data valuation approaches, Shapley value-based methods are predominant due to their strong theoretical grounding. However, the exact computation of Shapley values is often computationally prohibitive, prompting the development of numerous approximation techniques. Despite notable advancements, existing methods generally neglect the incorporation of value distribution information and fail to account for dynamic data conditions, thereby compromising their performance and application potential. In this paper, we highlight the crucial role of both global and local statistical properties of value distributions in the context of data valuation for machine learning. First, we conduct a comprehensive analysis of these distributions across various simulated and real-world datasets, uncovering valuable insights and key patterns. Second, we propose an enhanced data valuation method that fuses the explored distribution characteristics into two regularization terms to refine Shapley value estimation. The proposed regularizers can be seamlessly incorporated into various existing data valuation methods. Third, we introduce a novel approach for dynamic data valuation that infers updated data values without recomputing Shapley values, thereby significantly improving computational efficiency. Extensive experiments have been conducted across a range of tasks, including Shapley value estimation, value-based data addition and removal, mislabeled data detection, and dynamic data valuation. The results showcase the consistent effectiveness and efficiency of our proposed methodologies, affirming the significant potential of global and local value distributions in data valuation.", "date": "2024-05-23", "categories": []}
{"id": "2503.08805", "title": "Filter Like You Test: Data-Driven Data Filtering for CLIP Pretraining", "abstract": "We introduce Filter Like You Test (FLYT), an algorithm for curating large-scale vision-language datasets that learns the usefulness of each data point as a pretraining example. FLYT trains a scoring model that learns to weigh each example's features using gradient signals from downstream tasks training sets. Based on FLYT, we implement Mixing-FLYT (M-FLYT), which takes the per-example scores generated by different scoring methods as features, and learns to unify them into a single score. FLYT naturally produces a distribution over the training examples, which we leverage through Soft Cap Sampling (SCS), a strategy for obtaining a filtered pretraining dataset from per-example probabilities that samples examples while preventing over-representation through a repetition penalty. Using these methods, we achieve 40.1% ImageNet zero-shot accuracy on the DataComp medium scale filtering benchmark, a 2% absolute accuracy increase over all previous results and a 5.5% increase over results that - like us - use only public resources. Our approach also yields 37.7\\% on the average of 38 DataComp evaluation tasks, outperforming previous public-resource approaches by 0.4\\%.", "date": "2025-03-11", "categories": []}
{"id": "2511.21668", "title": "Through the telecom lens: Are all training samples important?", "abstract": "The rise of AI in telecommunications, from optimizing Radio Access Networks to managing user experience, has sharply increased data volumes and training demands. Telecom data is often noisy, high-dimensional, costly to store, process, and label. Despite Ai's critical role, standard workflows still assume all training samples contribute equally. On the other hand, next generation systems require AI models that are accurate, efficient, and sustainable.The paper questions the assumptions of equal importance by focusing on applying and analyzing the roles of individual samples in telecom training and assessing whether the proposed model optimizes computation and energy use. we perform sample-level gradient analysis across epochs to identify patterns of influence and redundancy in model learning. Based on this, we propose a sample importance framework thats electively prioritizes impactful data and reduces computation without compromising accuracy. Experiments on three real-world telecom datasets show that our method [reserves performance while reducing data needs and computational overhead while advancing the goals of sustainable AI in telecommunications.", "date": "2025-11-26", "categories": []}
{"id": "2509.02973", "title": "InstaDA: Augmenting Instance Segmentation Data with Dual-Agent System", "abstract": "Acquiring high-quality instance segmentation data is challenging due to the labor-intensive nature of the annotation process and significant class imbalances within datasets. Recent studies have utilized the integration of Copy-Paste and diffusion models to create more diverse datasets. However, these studies often lack deep collaboration between large language models (LLMs) and diffusion models, and underutilize the rich information within the existing training data. To address these limitations, we propose InstaDA, a novel, training-free Dual-Agent system designed to augment instance segmentation datasets. First, we introduce a Text-Agent (T-Agent) that enhances data diversity through collaboration between LLMs and diffusion models. This agent features a novel Prompt Rethink mechanism, which iteratively refines prompts based on the generated images. This process not only fosters collaboration but also increases image utilization and optimizes the prompts themselves. Additionally, we present an Image-Agent (I-Agent) aimed at enriching the overall data distribution. This agent augments the training set by generating new instances conditioned on the training images. To ensure practicality and efficiency, both agents operate as independent and automated workflows, enhancing usability. Experiments conducted on the LVIS 1.0 validation set indicate that InstaDA achieves significant improvements, with an increase of +4.0 in box average precision (AP) and +3.3 in mask AP compared to the baseline. Furthermore, it outperforms the leading model, DiverGen, by +0.3 in box AP and +0.1 in mask AP, with a notable +0.7 gain in box AP on common categories and mask AP gains of +0.2 on common categories and +0.5 on frequent categories.", "date": "2025-09-03", "categories": []}
{"id": "2309.16671", "title": "Demystifying CLIP Data", "abstract": "Contrastive Language-Image Pre-training (CLIP) is an approach that has advanced research and applications in computer vision, fueling modern recognition systems and generative models. We believe that the main ingredient to the success of CLIP is its data and not the model architecture or pre-training objective. However, CLIP only provides very limited information about its data and how it has been collected, leading to works that aim to reproduce CLIP's data by filtering with its model parameters. In this work, we intend to reveal CLIP's data curation approach and in our pursuit of making it open to the community introduce Metadata-Curated Language-Image Pre-training (MetaCLIP). MetaCLIP takes a raw data pool and metadata (derived from CLIP's concepts) and yields a balanced subset over the metadata distribution. Our experimental study rigorously isolates the model and training settings, concentrating solely on data. MetaCLIP applied to CommonCrawl with 400M image-text data pairs outperforms CLIP's data on multiple standard benchmarks. In zero-shot ImageNet classification, MetaCLIP achieves 70.8% accuracy, surpassing CLIP's 68.3% on ViT-B models. Scaling to 1B data, while maintaining the same training budget, attains 72.4%. Our observations hold across various model sizes, exemplified by ViT-H achieving 80.5%, without any bells-and-whistles. Curation code and training data distribution on metadata is made available at https://github.com/facebookresearch/MetaCLIP.", "date": "2023-09-28", "categories": []}
{"id": "2411.15349", "title": "Zero-Shot Coreset Selection via Iterative Subspace Sampling", "abstract": "Deep learning increasingly relies on massive data with substantial storage, annotation, and training costs. To reduce costs, coreset selection finds a representative subset of data to train models while ideally performing on par with the full data training. To maximize performance, current state-of-the-art coreset methods select data using dataset-specific ground truth labels and training. However, these methodological requirements prevent selection at scale on real-world, unlabeled data. To that end, this paper addresses the selection of coresets that achieve state-of-the-art performance but without using any labels or training on candidate data. Instead, our solution, Zero-Shot Coreset Selection via Iterative Subspace Sampling (ZCore), uses previously-trained foundation models to generate zero-shot, high-dimensional embedding spaces to interpret unlabeled data. ZCore then iteratively quantifies the relative value of all candidate data based on coverage and redundancy in numerous subspace distributions. Finally, ZCore selects a coreset sized for any data budget to train downstream models. We evaluate ZCore on four datasets and outperform several state-of-the-art label-based methods, especially at low data rates that provide the most substantial cost reduction. On ImageNet, ZCore selections for 10% training data achieve a downstream validation accuracy of 53.99%, which outperforms prior label-based methods and removes annotation and training costs for 1.15 million images. Our paper's code is publicly available at https://github.com/voxel51/zcore.", "date": "2024-11-22", "categories": []}
{"id": "2511.17890", "title": "Decoupled Audio-Visual Dataset Distillation", "abstract": "Audio-Visual Dataset Distillation aims to compress large-scale datasets into compact subsets while preserving the performance of the original data. However, conventional Distribution Matching (DM) methods struggle to capture intrinsic cross-modal alignment. Subsequent studies have attempted to introduce cross-modal matching, but two major challenges remain: (i) independently and randomly initialized encoders lead to inconsistent modality mapping spaces, increasing training difficulty; and (ii) direct interactions between modalities tend to damage modality-specific (private) information, thereby degrading the quality of the distilled data. To address these challenges, we propose DAVDD, a pretraining-based decoupled audio-visual distillation framework. DAVDD leverages a diverse pretrained bank to obtain stable modality features and uses a lightweight decoupler bank to disentangle them into common and private representations. To effectively preserve cross-modal structure, we further introduce Common Intermodal Matching together with a Sample-Distribution Joint Alignment strategy, ensuring that shared representations are aligned both at the sample level and the global distribution level. Meanwhile, private representations are entirely isolated from cross-modal interaction, safeguarding modality-specific cues throughout distillation. Extensive experiments across multiple benchmarks show that DAVDD achieves state-of-the-art results under all IPC settings, demonstrating the effectiveness of decoupled representation learning for high-quality audio-visual dataset distillation. Code will be released.", "date": "2025-11-22", "categories": []}
{"id": "2511.17914", "title": "Rectifying Soft-Label Entangled Bias in Long-Tailed Dataset Distillation", "abstract": "Dataset distillation compresses large-scale datasets into compact, highly informative synthetic data, significantly reducing storage and training costs. However, existing research primarily focuses on balanced datasets and struggles to perform under real-world long-tailed distributions. In this work, we emphasize the critical role of soft labels in long-tailed dataset distillation and uncover the underlying mechanisms contributing to performance degradation. Specifically, we derive an imbalance-aware generalization bound for model trained on distilled dataset. We then identify two primary sources of soft-label bias, which originate from the distillation model and the distilled images, through systematic perturbation of the data imbalance levels. To address this, we propose ADSA, an Adaptive Soft-label Alignment module that calibrates the entangled biases. This lightweight module integrates seamlessly into existing distillation pipelines and consistently improves performance. On ImageNet-1k-LT with EDC and IPC=50, ADSA improves tail-class accuracy by up to 11.8% and raises overall accuracy to 41.4%. Extensive experiments demonstrate that ADSA provides a robust and generalizable solution under limited label budgets and across a range of distillation techniques. Code is available at: https://github.com/j-cyoung/ADSA_DD.git.", "date": "2025-11-22", "categories": []}
{"id": "2511.18831", "title": "VideoCompressa: Data-Efficient Video Understanding via Joint Temporal Compression and Spatial Reconstruction", "abstract": "The scalability of video understanding models is increasingly limited by the prohibitive storage and computational costs of large-scale video datasets. While data synthesis has improved data efficiency in the image domain, its extension to video remains challenging due to pervasive temporal redundancy and complex spatiotemporal dynamics. In this work, we uncover a critical insight: the primary source of inefficiency in video datasets is not inter-sample redundancy, but intra-sample frame-level redundancy. To leverage this insight, we introduce VideoCompressa, a novel framework for video data synthesis that reframes the problem as dynamic latent compression. Specifically, VideoCompressa jointly optimizes a differentiable keyframe selector-implemented as a lightweight ConvNet with Gumbel-Softmax sampling-to identify the most informative frames, and a pretrained, frozen Variational Autoencoder (VAE) to compress these frames into compact, semantically rich latent codes. These latent representations are then fed into a compression network, enabling end-to-end backpropagation. Crucially, the keyframe selector and synthetic latent codes are co-optimized to maximize retention of task-relevant information. Experiments show that our method achieves unprecedented data efficiency: on UCF101 with ConvNets, VideoCompressa surpasses full-data training by 2.34\\% points using only 0.13\\% of the original data, with over 5800x speedup compared to traditional synthesis method. Moreover, when fine-tuning Qwen2.5-7B-VL on HMDB51, VideoCompressa matches full-data performance using just 0.41\\% of the training data-outperforming zero-shot baseline by 10.61\\%.", "date": "2025-11-24", "categories": []}
{"id": "2511.19274", "title": "Diffusion Reconstruction-based Data Likelihood Estimation for Core-Set Selection", "abstract": "Existing core-set selection methods predominantly rely on heuristic scoring signals such as training dynamics or model uncertainty, lacking explicit modeling of data likelihood. This omission may hinder the constructed subset from capturing subtle yet critical distributional structures that underpin effective model training. In this work, we propose a novel, theoretically grounded approach that leverages diffusion models to estimate data likelihood via reconstruction deviation induced by partial reverse denoising. Specifically, we establish a formal connection between reconstruction error and data likelihood, grounded in the Evidence Lower Bound (ELBO) of Markovian diffusion processes, thereby enabling a principled, distribution-aware scoring criterion for data selection. Complementarily, we introduce an efficient information-theoretic method to identify the optimal reconstruction timestep, ensuring that the deviation provides a reliable signal indicative of underlying data likelihood. Extensive experiments on ImageNet demonstrate that reconstruction deviation offers an effective scoring criterion, consistently outperforming existing baselines across selection ratios, and closely matching full-data training using only 50% of the data. Further analysis shows that the likelihood-informed nature of our score reveals informative insights in data selection, shedding light on the interplay between data distributional characteristics and model learning preferences.", "date": "2025-11-24", "categories": []}
{"id": "2511.16715", "title": "DDTime: Dataset Distillation with Spectral Alignment and Information Bottleneck for Time-Series Forecasting", "abstract": "Time-series forecasting is fundamental across many domains, yet training accurate models often requires large-scale datasets and substantial computational resources. Dataset distillation offers a promising alternative by synthesizing compact datasets that preserve the learning behavior of full data. However, extending dataset distillation to time-series forecasting is non-trivial due to two fundamental challenges: 1.temporal bias from strong autocorrelation, which leads to distorted value-term alignment between teacher and student models; and 2.insufficient diversity among synthetic samples, arising from the absence of explicit categorical priors to regularize trajectory variety.\n  In this work, we propose DDTime, a lightweight and plug-in distillation framework built upon first-order condensation decomposition. To tackle Challenge 1, it revisits value-term alignment through temporal statistics and introduces a frequency-domain alignment mechanism to mitigate autocorrelation-induced bias, ensuring spectral consistency and temporal fidelity. To address Challenge 2, we further design an inter-sample regularization inspired by the information bottleneck principle, which enhances diversity and maximizes information density across synthetic trajectories. The combined objective is theoretically compatible with a wide range of condensation paradigms and supports stable first-order optimization. Extensive experiments on 20 benchmark datasets and diverse forecasting architectures demonstrate that DDTime consistently outperforms existing distillation methods, achieving about 30% relative accuracy gains while introducing about 2.49% computational overhead. All code and distilled datasets will be released.", "date": "2025-11-20", "categories": []}
{"id": "2506.05281", "title": "Fast-DataShapley: Neural Modeling for Training Data Valuation", "abstract": "The value and copyright of training data are crucial in the artificial intelligence industry. Service platforms should protect data providers' legitimate rights and fairly reward them for their contributions. Shapley value, a potent tool for evaluating contributions, outperforms other methods in theory, but its computational overhead escalates exponentially with the number of data providers. Recent works based on Shapley values attempt to mitigate computation complexity by approximation algorithms. However, they need to retrain for each test sample, leading to intolerable costs. We propose Fast-DataShapley, a one-pass training method that leverages the weighted least squares characterization of the Shapley value to train a reusable explainer model with real-time reasoning speed. Given new test samples, no retraining is required to calculate the Shapley values of the training data. Additionally, we propose three methods with theoretical guarantees to reduce training overhead from two aspects: the approximate calculation of the utility function and the group calculation of the training data. We analyze time complexity to show the efficiency of our methods. The experimental evaluations on various image datasets demonstrate superior performance and efficiency compared to baselines. Specifically, the performance is improved to more than 2 times, and the explainer's training speed can be increased by two orders of magnitude.", "date": "2025-06-05", "categories": []}
{"id": "2511.16674", "title": "Dataset Distillation for Pre-Trained Self-Supervised Vision Models", "abstract": "The task of dataset distillation aims to find a small set of synthetic images such that training a model on them reproduces the performance of the same model trained on a much larger dataset of real samples. Existing distillation methods focus on synthesizing datasets that enable training randomly initialized models. In contrast, state-of-the-art vision approaches are increasingly building on large, pre-trained self-supervised models rather than training from scratch. In this paper, we investigate the problem of distilling datasets that enable us to optimally train linear probes on top of such large, pre-trained vision models. We introduce a method of dataset distillation for this task called Linear Gradient Matching that optimizes the synthetic images such that, when passed through a pre-trained feature extractor, they induce gradients in the linear classifier similar to those produced by the real data. Our method yields synthetic data that outperform all real-image baselines and, remarkably, generalize across pre-trained vision models, enabling us, for instance, to train a linear CLIP probe that performs competitively using a dataset distilled via a DINO backbone. Further, we show that our distilled datasets are exceptionally effective for fine-grained classification and provide a valuable tool for model interpretability, predicting, among other things, how similar two models' embedding spaces are under the platonic representation hypothesis or whether a model is sensitive to spurious correlations in adversarial datasets.", "date": "2025-11-20", "categories": []}
{"id": "2411.00388", "title": "Towards Data Valuation via Asymmetric Data Shapley", "abstract": "As data emerges as a vital driver of technological and economic advancements, a key challenge is accurately quantifying its value in algorithmic decision-making. The Shapley value, a well-established concept from cooperative game theory, has been widely adopted to assess the contribution of individual data sources in supervised machine learning. However, its symmetry axiom assumes all players in the cooperative game are homogeneous, which overlooks the complex structures and dependencies present in real-world datasets. To address this limitation, we extend the traditional data Shapley framework to asymmetric data Shapley, making it flexible enough to incorporate inherent structures within the datasets for structure-aware data valuation. We also introduce an efficient $k$-nearest neighbor-based algorithm for its exact computation. We demonstrate the practical applicability of our framework across various machine learning tasks and data market contexts. The code is available at: https://github.com/xzheng01/Asymmetric-Data-Shapley.", "date": "2024-11-01", "categories": []}
{"id": "2502.00198", "title": "Fairshare Data Pricing via Data Valuation for Large Language Models", "abstract": "Training data is the backbone of large language models (LLMs), yet today's data markets often operate under exploitative pricing -- sourcing data from marginalized groups with little pay or recognition. This paper introduces a theoretical framework for LLM data markets, modeling the strategic interactions between buyers (LLM builders) and sellers (human annotators). We begin with theoretical and empirical analysis showing how exploitative pricing drives high-quality sellers out of the market, degrading data quality and long-term model performance. Then we introduce fairshare, a pricing mechanism grounded in data valuation that quantifies each data's contribution. It aligns incentives by sustaining seller participation and optimizing utility for both buyers and sellers. Theoretically, we show that fairshare yields mutually optimal outcomes: maximizing long-term buyer utility and seller profit while sustaining market participation. Empirically when training open-source LLMs on complex NLP tasks, including math problems, medical diagnosis, and physical reasoning, fairshare boosts seller earnings and ensures a stable supply of high-quality data, while improving buyers' performance-per-dollar and long-term welfare. Our findings offer a concrete path toward fair, transparent, and economically sustainable data markets for LLM.", "date": "2025-01-31", "categories": []}
{"id": "2511.12988", "title": "UNSEEN: Enhancing Dataset Pruning from a Generalization Perspective", "abstract": "The growing scale of datasets in deep learning has introduced significant computational challenges. Dataset pruning addresses this challenge by constructing a compact but informative coreset from the full dataset with comparable performance. Previous approaches typically establish scoring metrics based on specific criteria to identify representative samples. However, these methods predominantly rely on sample scores obtained from the model's performance during the training (i.e., fitting) phase. As scoring models achieve near-optimal performance on training data, such fitting-centric approaches induce a dense distribution of sample scores within a narrow numerical range. This concentration reduces the distinction between samples and hinders effective selection. To address this challenge, we conduct dataset pruning from the perspective of generalization, i.e., scoring samples based on models not exposed to them during training. We propose a plug-and-play framework, UNSEEN, which can be integrated into existing dataset pruning methods. Additionally, conventional score-based methods are single-step and rely on models trained solely on the complete dataset, providing limited perspective on the importance of samples. To address this limitation, we scale UNSEEN to multi-step scenarios and propose an incremental selection technique through scoring models trained on varying coresets, and optimize the quality of the coreset dynamically. Extensive experiments demonstrate that our method significantly outperforms existing state-of-the-art (SOTA) methods on CIFAR-10, CIFAR-100, and ImageNet-1K. Notably, on ImageNet-1K, UNSEEN achieves lossless performance while reducing training data by 30\\%.", "date": "2025-11-17", "categories": []}
{"id": "2511.11945", "title": "Augmenting The Weather: A Hybrid Counterfactual-SMOTE Algorithm for Improving Crop Growth Prediction When Climate Changes", "abstract": "In recent years, humanity has begun to experience the catastrophic effects of climate change as economic sectors (such as agriculture) struggle with unpredictable and extreme weather events. Artificial Intelligence (AI) should help us handle these climate challenges but its most promising solutions are not good at dealing with climate-disrupted data; specifically, machine learning methods that work from historical data-distributions, are not good at handling out-of-distribution, outlier events. In this paper, we propose a novel data augmentation method, that treats the predictive problems around climate change as being, in part, due to class-imbalance issues; that is, prediction from historical datasets is difficult because, by definition, they lack sufficient minority-class instances of \"climate outlier events\". This novel data augmentation method -- called Counterfactual-Based SMOTE (CFA-SMOTE) -- combines an instance-based counterfactual method from Explainable AI (XAI) with the well-known class-imbalance method, SMOTE. CFA-SMOTE creates synthetic data-points representing outlier, climate-events that augment the dataset to improve predictive performance. We report comparative experiments using this CFA-SMOTE method, comparing it to benchmark counterfactual and class-imbalance methods under different conditions (i.e., class-imbalance ratios). The focal climate-change domain used relies on predicting grass growth on Irish dairy farms, during Europe-wide drought and forage crisis of 2018.", "date": "2025-11-14", "categories": []}
{"id": "2511.13106", "title": "Low-Level Dataset Distillation for Medical Image Enhancement", "abstract": "Medical image enhancement is clinically valuable, but existing methods require large-scale datasets to learn complex pixel-level mappings. However, the substantial training and storage costs associated with these datasets hinder their practical deployment. While dataset distillation (DD) can alleviate these burdens, existing methods mainly target high-level tasks, where multiple samples share the same label. This many-to-one mapping allows distilled data to capture shared semantics and achieve information compression. In contrast, low-level tasks involve a many-to-many mapping that requires pixel-level fidelity, making low-level DD an underdetermined problem, as a small distilled dataset cannot fully constrain the dense pixel-level mappings. To address this, we propose the first low-level DD method for medical image enhancement. We first leverage anatomical similarities across patients to construct the shared anatomical prior based on a representative patient, which serves as the initialization for the distilled data of different patients. This prior is then personalized for each patient using a Structure-Preserving Personalized Generation (SPG) module, which integrates patient-specific anatomical information into the distilled dataset while preserving pixel-level fidelity. For different low-level tasks, the distilled data is used to construct task-specific high- and low-quality training pairs. Patient-specific knowledge is injected into the distilled data by aligning the gradients computed from networks trained on the distilled pairs with those from the corresponding patient's raw data. Notably, downstream users cannot access raw patient data. Instead, only a distilled dataset containing abstract training information is shared, which excludes patient-specific details and thus preserves privacy.", "date": "2025-11-17", "categories": []}
{"id": "2511.13640", "title": "Data Value in the Age of Scaling: Understanding LLM Scaling Dynamics Under Real-Synthetic Data Mixtures", "abstract": "The rapid progress of large language models (LLMs) is fueled by the growing reliance on datasets that blend real and synthetic data. While synthetic data offers scalability and cost-efficiency, it often introduces systematic distributional discrepancies, particularly underrepresenting long-tail knowledge due to truncation effects from data generation mechanisms like top-p sampling, temperature scaling, and finite sampling. These discrepancies pose fundamental challenges in characterizing and evaluating the utility of mixed real-synthetic datasets. In this paper, we identify a three-phase scaling behavior characterized by two breakpoints that reflect transitions in model behavior across learning head and tail knowledge. We further derive an LLM generalization bound designed for real and synthetic mixtures, revealing several key factors that govern their generalization performance. Building on our theoretical findings, we propose an effective yet efficient data valuation method that scales to large-scale datasets. Comprehensive experiments across four tasks, including image classification, sentiment classification, instruction following, and complex reasoning, demonstrate that our method surpasses state-of-the-art baselines in data valuation with significantly low computational cost.", "date": "2025-11-17", "categories": []}
{"id": "2407.00482", "title": "Towards Formalizing Spuriousness of Biased Datasets Using Partial Information Decomposition", "abstract": "Spuriousness arises when there is an association between two or more variables in a dataset that are not causally related. In this work, we propose an explainability framework to preemptively disentangle the nature of such spurious associations in a dataset before model training. We leverage a body of work in information theory called Partial Information Decomposition (PID) to decompose the total information about the target into four non-negative quantities, namely unique information (in core and spurious features, respectively), redundant information, and synergistic information. Our framework helps anticipate when the core or spurious feature is indispensable, when either suffices, and when both are jointly needed for an optimal classifier trained on the dataset. Next, we leverage this decomposition to propose a novel measure of the spuriousness of a dataset. We arrive at this measure systematically by examining several candidate measures, and demonstrating what they capture and miss through intuitive canonical examples and counterexamples. Our framework Spurious Disentangler consists of segmentation, dimensionality reduction, and estimation modules, with capabilities to specifically handle high-dimensional image data efficiently. Finally, we also perform empirical evaluation to demonstrate the trends of unique, redundant, and synergistic information, as well as our proposed spuriousness measure across $6$ benchmark datasets under various experimental settings. We observe an agreement between our preemptive measure of dataset spuriousness and post-training model generalization metrics such as worst-group accuracy, further supporting our proposition. The code is available at https://github.com/Barproda/spuriousness-disentangler.", "date": "2024-06-29", "categories": []}
{"id": "2412.09959", "title": "Towards Consistent and Efficient Dataset Distillation via Diffusion-Driven Selection", "abstract": "Dataset distillation provides an effective approach to reduce memory and computational costs by optimizing a compact dataset that achieves performance comparable to the full original. However, for large-scale datasets and complex deep networks (e.g., ImageNet-1K with ResNet-101), the vast optimization space hinders distillation effectiveness, limiting practical applications. Recent methods leverage pre-trained diffusion models to directly generate informative images, thereby bypassing pixel-level optimization and achieving promising results. Nonetheless, these approaches often suffer from distribution shifts between the pre-trained diffusion prior and target datasets, as well as the need for multiple distillation steps under varying settings. To overcome these challenges, we propose a novel framework that is orthogonal to existing diffusion-based distillation techniques by utilizing the diffusion prior for patch selection rather than generation. Our method predicts noise from the diffusion model conditioned on input images and optional text prompts (with or without label information), and computes the associated loss for each image-patch pair. Based on the loss differences, we identify distinctive regions within the original images. Furthermore, we apply intra-class clustering and ranking on the selected patches to enforce diversity constraints. This streamlined pipeline enables a one-step distillation process. Extensive experiments demonstrate that our approach consistently outperforms state-of-the-art methods across various metrics and settings.", "date": "2024-12-13", "categories": []}
{"id": "2511.09488", "title": "AutoSynth: Automated Workflow Optimization for High-Quality Synthetic Dataset Generation via Monte Carlo Tree Search", "abstract": "Supervised fine-tuning (SFT) of large language models (LLMs) for specialized tasks requires high-quality datasets, but manual curation is prohibitively expensive. Synthetic data generation offers scalability, but its effectiveness relies on complex, multi-stage workflows, integrating prompt engineering and model orchestration. Existing automated workflow methods face a cold start problem: they require labeled datasets for reward modeling, which is especially problematic for subjective, open-ended tasks with no objective ground truth. We introduce AutoSynth, a framework that automates workflow discovery and optimization without reference datasets by reframing the problem as a Monte Carlo Tree Search guided by a novel dataset-free hybrid reward. This reward enables meta-learning through two LLM-as-judge components: one evaluates sample quality using dynamically generated task-specific metrics, and another assesses workflow code and prompt quality. Experiments on subjective educational tasks show that while expert-designed workflows achieve higher human preference rates (96-99% win rates vs. AutoSynth's 40-51%), models trained on AutoSynth-generated data dramatically outperform baselines (40-51% vs. 2-5%) and match or surpass expert workflows on certain metrics, suggesting discovery of quality dimensions beyond human intuition. These results are achieved while reducing human effort from 5-7 hours to just 30 minutes (>90% reduction). AutoSynth tackles the cold start issue in data-centric AI, offering a scalable, cost-effective method for subjective LLM tasks. Code: https://github.com/bisz9918-maker/AutoSynth.", "date": "2025-11-12", "categories": []}
{"id": "2410.04715", "title": "Selection of LLM Fine-Tuning Data based on Orthogonal Rules", "abstract": "High-quality training data is critical to the performance of large language models (LLMs). Recent work has explored using LLMs to rate and select data based on a small set of human-designed criteria (rules), but these approaches often rely heavily on heuristics, lack principled metrics for rule evaluation, and generalize poorly to new tasks. We propose a novel rule-based data selection framework that introduces a metric based on the orthogonality of rule score vectors to evaluate and select complementary rules. Our automated pipeline first uses LLMs to generate diverse rules covering multiple aspects of data quality, then rates samples according to these rules and applies the determinantal point process (DPP) to select the most independent rules. These rules are then used to score the full dataset, and high-scoring samples are selected for downstream tasks such as LLM fine-tuning. We evaluate our framework in two experiment setups: (1) alignment with ground-truth ratings and (2) performance of LLMs fine-tuned on the selected data. Experiments across IMDB, Medical, Math, and Code domains demonstrate that our DPP-based rule selection consistently improves both rating accuracy and downstream model performance over strong baselines.", "date": "2024-10-07", "categories": []}
{"id": "2511.08263", "title": "ImagebindDC: Compressing Multi-modal Data with Imagebind-based Condensation", "abstract": "Data condensation techniques aim to synthesize a compact dataset from a larger one to enable efficient model training, yet while successful in unimodal settings, they often fail in multimodal scenarios where preserving intricate inter-modal dependencies is crucial. To address this, we introduce ImageBindDC, a novel data condensation framework operating within the unified feature space of ImageBind. Our approach moves beyond conventional distribution-matching by employing a powerful Characteristic Function (CF) loss, which operates in the Fourier domain to facilitate a more precise statistical alignment via exact infinite moment matching. We design our objective to enforce three critical levels of distributional consistency: (i) uni-modal alignment, which matches the statistical properties of synthetic and real data within each modality; (ii) cross-modal alignment, which preserves pairwise semantics by matching the distributions of hybrid real-synthetic data pairs; and (iii) joint-modal alignment, which captures the complete multivariate data structure by aligning the joint distribution of real data pairs with their synthetic counterparts. Extensive experiments highlight the effectiveness of ImageBindDC: on the NYU-v2 dataset, a model trained on just 5 condensed datapoints per class achieves lossless performance comparable to one trained on the full dataset, achieving a new state-of-the-art with an 8.2\\% absolute improvement over the previous best method and more than 4$\\times$ less condensation time.", "date": "2025-11-11", "categories": []}
{"id": "2408.13850", "title": "Condensed Data Expansion Using Model Inversion for Knowledge Distillation", "abstract": "Condensed datasets offer a compact representation of larger datasets, but training models directly on them or using them to enhance model performance through knowledge distillation (KD) can result in suboptimal outcomes due to limited information. To address this, we propose a method that expands condensed datasets using model inversion, a technique for generating synthetic data based on the impressions of a pre-trained model on its training data. This approach is particularly well-suited for KD scenarios, as the teacher model is already pre-trained and retains knowledge of the original training data. By creating synthetic data that complements the condensed samples, we enrich the training set and better approximate the underlying data distribution, leading to improvements in student model accuracy during knowledge distillation. Our method demonstrates significant gains in KD accuracy compared to using condensed datasets alone and outperforms standard model inversion-based KD methods by up to 11.4% across various datasets and model architectures. Importantly, it remains effective even when using as few as one condensed sample per class, and can also enhance performance in few-shot scenarios where only limited real data samples are available.", "date": "2024-08-25", "categories": []}
{"id": "2511.03239", "title": "A Feedback-Control Framework for Efficient Dataset Collection from In-Vehicle Data Streams", "abstract": "Modern AI systems are increasingly constrained not by model capacity but by the quality and diversity of their data. Despite growing emphasis on data-centric AI, most datasets are still gathered in an open-loop manner which accumulates redundant samples without feedback from the current coverage. This results in inefficient storage, costly labeling, and limited generalization. To address this, this paper introduces Feedback Control Data Collection (FCDC), a paradigm that formulates data collection as a closed-loop control problem. FCDC continuously approximates the state of the collected data distribution using an online probabilistic model and adaptively regulates sample retention using based on feedback signals such as likelihood and Mahalanobis distance. Through this feedback mechanism, the system dynamically balances exploration and exploitation, maintains dataset diversity, and prevents redundancy from accumulating over time. In addition to demonstrating the controllability of FCDC on a synthetic dataset that converges toward a uniform distribution under Gaussian input assumption, experiments on real data streams show that FCDC produces more balanced datasets by 25.9% while reducing data storage by 39.8%. These results demonstrate that data collection itself can be actively controlled, transforming collection from a passive pipeline stage into a self-regulating, feedback-driven process at the core of data-centric AI.", "date": "2025-11-05", "categories": []}
{"id": "2511.06530", "title": "Better Datasets Start From RefineLab: Automatic Optimization for High-Quality Dataset Refinement", "abstract": "High-quality Question-Answer (QA) datasets are foundational for reliable Large Language Model (LLM) evaluation, yet even expert-crafted datasets exhibit persistent gaps in domain coverage, misaligned difficulty distributions, and factual inconsistencies. The recent surge in generative model-powered datasets has compounded these quality challenges. In this work, we introduce RefineLab, the first LLM-driven framework that automatically refines raw QA textual data into high-quality datasets under a controllable token-budget constraint. RefineLab takes a set of target quality attributes (such as coverage and difficulty balance) as refinement objectives, and performs selective edits within a predefined token budget to ensure practicality and efficiency. In essence, RefineLab addresses a constrained optimization problem: improving the quality of QA samples as much as possible while respecting resource limitations. With a set of available refinement operations (e.g., rephrasing, distractor replacement), RefineLab takes as input the original dataset, a specified set of target quality dimensions, and a token budget, and determines which refinement operations should be applied to each QA sample. This process is guided by an assignment module that selects optimal refinement strategies to maximize overall dataset quality while adhering to the budget constraint. Experiments demonstrate that RefineLab consistently narrows divergence from expert datasets across coverage, difficulty alignment, factual fidelity, and distractor quality. RefineLab pioneers a scalable, customizable path to reproducible dataset design, with broad implications for LLM evaluation.", "date": "2025-11-09", "categories": []}
{"id": "2511.07074", "title": "Importance-Aware Data Selection for Efficient LLM Instruction Tuning", "abstract": "Instruction tuning plays a critical role in enhancing the performance and efficiency of Large Language Models (LLMs). Its success depends not only on the quality of the instruction data but also on the inherent capabilities of the LLM itself. Some studies suggest that even a small amount of high-quality data can achieve instruction fine-tuning results that are on par with, or even exceed, those from using a full-scale dataset. However, rather than focusing solely on calculating data quality scores to evaluate instruction data, there is a growing need to select high-quality data that maximally enhances the performance of instruction tuning for a given LLM. In this paper, we propose the Model Instruction Weakness Value (MIWV) as a novel metric to quantify the importance of instruction data in enhancing model's capabilities. The MIWV metric is derived from the discrepancies in the model's responses when using In-Context Learning (ICL), helping identify the most beneficial data for enhancing instruction tuning performance. Our experimental results demonstrate that selecting only the top 1\\% of data based on MIWV can outperform training on the full dataset. Furthermore, this approach extends beyond existing research that focuses on data quality scoring for data selection, offering strong empirical evidence supporting the effectiveness of our proposed method.", "date": "2025-11-10", "categories": []}
{"id": "2511.05057", "title": "Role-SynthCLIP: A Role Play Driven Diverse Synthetic Data Approach", "abstract": "The effectiveness of Contrastive Language-Image Pre-training (CLIP) models critically depends on the semantic diversity and quality of their training data. However, while existing synthetic data generation methods primarily focus on increasing data volume, such emphasis often leads to limited semantic diversity and redundant or shallow captions. To address this limitation, we propose Role-SynthCLIP, a novel data synthesis framework that leverages multi-perspective role-playing prompts (e.g., a compositional analyst, an interpreter of image context) to guide Multimodal Large Language Models (MLLMs) in generating semantically diverse captions from distinct viewpoints. This mechanism enhances the semantic diversity and fine-grained image-text alignment of synthetic pairs, thereby improving caption expressiveness and accuracy while keeping the total number of image-text pairs unchanged. Experimental results demonstrate the effectiveness and efficiency of our method. A CLIP-B/16 model trained on only 1 million Role-SynthCLIP pairs achieves a Recall@1 of 64.1% on the MS COCO validation set, surpassing the best existing synthetic data baseline (trained on 5M pairs) by 2.8 percentage points. The code and trained models are released at https://github.com/huangfu170/Role-SynthCLIP.", "date": "2025-11-07", "categories": []}
{"id": "2511.03492", "title": "Why Less is More (Sometimes): A Theory of Data Curation", "abstract": "This paper introduces a theoretical framework to resolve a central paradox in modern machine learning: When is it better to use less data? This question has become critical as classical scaling laws suggesting ``more is more'' (Sun et al., 2025) are challenged by methods like LIMO (``less is more'') and s1 (Ye et al., 2025; Muenighoff et al., 2025), which achieve superior performance with small, aggressively curated datasets. Here, we study data curation strategies where an imperfect oracle selects the training examples according to their difficulty and correctness. Our results provide exact scaling law curves for test error under both label-agnostic and label-aware curation rules, revealing when and why keeping only a subset of data can improve generalization. In contrast to classical scaling laws, we show that under certain conditions, small curated datasets can outperform full datasets, and we provide analytical conditions for this by deriving precise phase transition curves tied to data size and quality. We validate these theoretical claims with empirical results on ImageNet, confirming our predictions about when curation improves accuracy and can even mitigate model collapse. Furthermore, our framework provides a principled explanation for the contradictory curation strategies recently observed in LLM mathematical reasoning.", "date": "2025-11-05", "categories": []}
{"id": "2511.02100", "title": "Geometric Data Valuation via Leverage Scores", "abstract": "Shapley data valuation provides a principled, axiomatic framework for assigning importance to individual datapoints, and has gained traction in dataset curation, pruning, and pricing. However, it is a combinatorial measure that requires evaluating marginal utility across all subsets of the data, making it computationally infeasible at scale. We propose a geometric alternative based on statistical leverage scores, which quantify each datapoint's structural influence in the representation space by measuring how much it extends the span of the dataset and contributes to the effective dimensionality of the training problem. We show that our scores satisfy the dummy, efficiency, and symmetry axioms of Shapley valuation and that extending them to \\emph{ridge leverage scores} yields strictly positive marginal gains that connect naturally to classical A- and D-optimal design criteria. We further show that training on a leverage-sampled subset produces a model whose parameters and predictive risk are within $O(\\varepsilon)$ of the full-data optimum, thereby providing a rigorous link between data valuation and downstream decision quality. Finally, we conduct an active learning experiment in which we empirically demonstrate that ridge-leverage sampling outperforms standard baselines without requiring access gradients or backward passes.", "date": "2025-11-03", "categories": []}
{"id": "2405.03869", "title": "Outlier Gradient Analysis: Efficiently Identifying Detrimental Training Samples for Deep Learning Models", "abstract": "A core data-centric learning challenge is the identification of training samples that are detrimental to model performance. Influence functions serve as a prominent tool for this task and offer a robust framework for assessing training data influence on model predictions. Despite their widespread use, their high computational cost associated with calculating the inverse of the Hessian matrix pose constraints, particularly when analyzing large-sized deep models. In this paper, we establish a bridge between identifying detrimental training samples via influence functions and outlier gradient detection. This transformation not only presents a straightforward and Hessian-free formulation but also provides insights into the role of the gradient in sample impact. Through systematic empirical evaluations, we first validate the hypothesis of our proposed outlier gradient analysis approach on synthetic datasets. We then demonstrate its effectiveness in detecting mislabeled samples in vision models and selecting data samples for improving performance of natural language processing transformer models. We also extend its use to influential sample identification for fine-tuning Large Language Models.", "date": "2024-05-06", "categories": []}
{"id": "2407.20299", "title": "Dataset Distillation for Offline Reinforcement Learning", "abstract": "Offline reinforcement learning often requires a quality dataset that we can train a policy on. However, in many situations, it is not possible to get such a dataset, nor is it easy to train a policy to perform well in the actual environment given the offline data. We propose using data distillation to train and distill a better dataset which can then be used for training a better policy model. We show that our method is able to synthesize a dataset where a model trained on it achieves similar performance to a model trained on the full dataset or a model trained using percentile behavioral cloning. Our project site is available at https://datasetdistillation4rl.github.io . We also provide our implementation at https://github.com/ggflow123/DDRL .", "date": "2024-07-29", "categories": []}
{"id": "2502.03656", "title": "A Study in Dataset Distillation for Image Super-Resolution", "abstract": "Dataset distillation aims to compress large datasets into compact yet highly informative subsets that preserve the training behavior of the original data. While this concept has gained traction in classification, its potential for image Super-Resolution (SR) remains largely untapped. In this work, we conduct the first systematic study of dataset distillation for SR, evaluating both pixel- and latent-space formulations. We show that a distilled dataset, occupying only 8.88% of the original size, can train SR models that retain nearly the same reconstruction fidelity as those trained on full datasets. Furthermore, we analyze how initialization strategies and distillation objectives affect efficiency, convergence, and visual quality. Our findings highlight the feasibility of SR dataset distillation and establish foundational insights for memory- and compute-efficient generative restoration models.", "date": "2025-02-05", "categories": []}
{"id": "2510.16806", "title": "Computational Budget Should Be Considered in Data Selection", "abstract": "Data selection improves computational efficiency by choosing informative subsets of training samples. However, existing methods ignore the compute budget, treating data selection and importance evaluation independently of compute budget constraints. Yet empirical studies show no algorithm can consistently outperform others (or even random selection) across varying budgets. We therefore argue that compute budget must be integral to data-selection strategies, since different budgets impose distinct requirements on data quantity, quality, and distribution for effective training. To this end, we propose a novel Computational budget-Aware Data Selection (CADS) method and naturally formulate it into a bilevel optimization framework, where the inner loop trains the model within the constraints of the computational budget on some selected subset of training data, while the outer loop optimizes data selection based on model evaluation. Our technical contributions lie in addressing two main challenges in solving this bilevel optimization problem: the expensive Hessian matrix estimation for outer-loop gradients and the computational burden of achieving inner-loop optimality during iterations. To solve the first issue, we propose a probabilistic reparameterization strategy and compute the gradient using a Hessian-free policy gradient estimator. To address the second challenge, we transform the inner optimization problem into a penalty term in the outer objective, further discovering that we only need to estimate the minimum of a one-dimensional loss to calculate the gradient, significantly improving efficiency. Extensive experiments show that our method achieves performance gains of up to 14.42% over baselines in vision and language benchmarks.", "date": "2025-10-19", "categories": []}
{"id": "2510.27253", "title": "Not All Instances Are Equally Valuable: Towards Influence-Weighted Dataset Distillation", "abstract": "Dataset distillation condenses large datasets into synthetic subsets, achieving performance comparable to training on the full dataset while substantially reducing storage and computation costs. Most existing dataset distillation methods assume that all real instances contribute equally to the process. In practice, real-world datasets contain both informative and redundant or even harmful instances, and directly distilling the full dataset without considering data quality can degrade model performance. In this work, we present Influence-Weighted Distillation IWD, a principled framework that leverages influence functions to explicitly account for data quality in the distillation process. IWD assigns adaptive weights to each instance based on its estimated impact on the distillation objective, prioritizing beneficial data while downweighting less useful or harmful ones. Owing to its modular design, IWD can be seamlessly integrated into diverse dataset distillation frameworks. Our empirical results suggest that integrating IWD tends to improve the quality of distilled datasets and enhance model performance, with accuracy gains of up to 7.8%.", "date": "2025-10-31", "categories": []}
{"id": "2510.26186", "title": "ConceptScope: Characterizing Dataset Bias via Disentangled Visual Concepts", "abstract": "Dataset bias, where data points are skewed to certain concepts, is ubiquitous in machine learning datasets. Yet, systematically identifying these biases is challenging without costly, fine-grained attribute annotations. We present ConceptScope, a scalable and automated framework for analyzing visual datasets by discovering and quantifying human-interpretable concepts using Sparse Autoencoders trained on representations from vision foundation models. ConceptScope categorizes concepts into target, context, and bias types based on their semantic relevance and statistical correlation to class labels, enabling class-level dataset characterization, bias identification, and robustness evaluation through concept-based subgrouping. We validate that ConceptScope captures a wide range of visual concepts, including objects, textures, backgrounds, facial attributes, emotions, and actions, through comparisons with annotated datasets. Furthermore, we show that concept activations produce spatial attributions that align with semantically meaningful image regions. ConceptScope reliably detects known biases (e.g., background bias in Waterbirds) and uncovers previously unannotated ones (e.g, co-occurring objects in ImageNet), offering a practical tool for dataset auditing and model diagnostics.", "date": "2025-10-30", "categories": []}
{"id": "2509.06463", "title": "Accelerate Scaling of LLM Finetuning via Quantifying the Coverage and Depth of Instruction Set", "abstract": "Scaling the amount of data used for supervied fine-tuning(SFT) does not guarantee the proportional gains in model performance, highlighting a critical need to understand what makes training samples effective. This work identifies two fundamental dataset properties that govern SFT scalability: \\textbf{semantic coverage}, or the breadth of task domains, and \\textbf{information depth}, or the richness of individual examples. We demonstrate that simple proxies for these properties explain the majority of validation loss variance in our experiments. In this work, we further propose the \\textbf{Information Landscape Approximation (ILA)}, a model-agnostic data selection framework that jointly optimizes for these two factors. ILA constructs compact subsets that approximate the informational value of large datasets. Empirical results show that models tuned on ILA-selected data achieve faster and more sustained performance improvements across diverse tasks and model sizes compared to existing methods, a phenomenon we term \\textbf{accelerated scaling}.", "date": "2025-09-08", "categories": []}
{"id": "2510.24262", "title": "UtilGen: Utility-Centric Generative Data Augmentation with Dual-Level Task Adaptation", "abstract": "Data augmentation using generative models has emerged as a powerful paradigm for enhancing performance in computer vision tasks. However, most existing augmentation approaches primarily focus on optimizing intrinsic data attributes -- such as fidelity and diversity -- to generate visually high-quality synthetic data, while often neglecting task-specific requirements. Yet, it is essential for data generators to account for the needs of downstream tasks, as training data requirements can vary significantly across different tasks and network architectures. To address these limitations, we propose UtilGen, a novel utility-centric data augmentation framework that adaptively optimizes the data generation process to produce task-specific, high-utility training data via downstream task feedback. Specifically, we first introduce a weight allocation network to evaluate the task-specific utility of each synthetic sample. Guided by these evaluations, UtilGen iteratively refines the data generation process using a dual-level optimization strategy to maximize the synthetic data utility: (1) model-level optimization tailors the generative model to the downstream task, and (2) instance-level optimization adjusts generation policies -- such as prompt embeddings and initial noise -- at each generation round. Extensive experiments on eight benchmark datasets of varying complexity and granularity demonstrate that UtilGen consistently achieves superior performance, with an average accuracy improvement of 3.87% over previous SOTA. Further analysis of data influence and distribution reveals that UtilGen produces more impactful and task-relevant synthetic data, validating the effectiveness of the paradigm shift from visual characteristics-centric to task utility-centric data augmentation.", "date": "2025-10-28", "categories": []}
{"id": "2505.17895", "title": "DataRater: Meta-Learned Dataset Curation", "abstract": "The quality of foundation models depends heavily on their training data. Consequently, great efforts have been put into dataset curation. Yet most approaches rely on manual tuning of coarse-grained mixtures of large buckets of data, or filtering by hand-crafted heuristics. An approach that is ultimately more scalable (let alone more satisfying) is to \\emph{learn} which data is actually valuable for training. This type of meta-learning could allow more sophisticated, fine-grained, and effective curation. Our proposed \\emph{DataRater} is an instance of this idea. It estimates the value of training on any particular data point. This is done by meta-learning using `meta-gradients', with the objective of improving training efficiency on held out data. In extensive experiments across a range of model scales and datasets, we find that using our DataRater to filter data is highly effective, resulting in significantly improved compute efficiency.", "date": "2025-05-23", "categories": []}
{"id": "2510.23160", "title": "ENTP: Enhancing Low-Quality SFT Data via Neural-Symbolic Text Purge-Mix", "abstract": "Supervised Fine-Tuning (SFT) adapts pre-trained Large Language Models (LLMs) to domain-specific instructions by training on a carefully curated subset of high-quality instruction-response pairs, typically drawn from a larger dataset that often contains many low-quality or noisy samples. However, existing quality-first paradigms often overlook valuable signals in discarded low-quality data and rely on imperfect quality filters. We introduce ENTP (Enhancing low-quality SFT data via Neural-symbolic Text Purge-Mix), a framework that revitalizes low-quality corpora through symbolic purification and neural reconstruction. The symbolic module identifies and prunes noisy samples based on statistical priors, while the neural component synthesizes enriched instruction-response pairs by leveraging latent representations and model knowledge. This neural-symbolic synergy enhances data informativeness and diversity. Experiments show that ENTP-augmented datasets, constructed exclusively from low-quality data, outperform 13 established data-selection baselines across five instruction-following benchmarks, and even surpass fine-tuning on the full original dataset (approximately 300K examples). Our results highlight the untapped potential of low-quality data and underscore the importance of intelligent purification and synthesis for efficient instruction alignment.", "date": "2025-10-27", "categories": []}
{"id": "2510.23284", "title": "DCMM-SQL: Automated Data-Centric Pipeline and Multi-Model Collaboration Training for Text-to-SQL Model", "abstract": "Text-to-SQL tasks have gained attractive improvements since the release of ChatGPT. Among them, agent-based frameworks have been widely used in this field. However, the impact of data-centric strategies on text-to-SQL tasks has rarely been explored. In this paper, we systemically design a fully automated data-centric pipeline for text-to-SQL tasks, including \\emph{adaptive data repair}, which can automatically find and fix errors in the training dataset; and \\emph{error data augmentation}, where we specifically diffuse and enhance erroneous data predicted by the initially trained models. Meanwhile, we propose a Multi-Model collaboration training schema, aiming to train multiple models with different augmented data, enabling them to possess distinct capabilities and work together to complement each other, because it has been found that the capability of a single fine-tuned model is very limited. Furthermore, we utilize an ensemble strategy to integrate the capabilities of multiple models to solve a multiple-choice question, aiming to further improve the accuracy of text-to-SQL tasks. The experiment results and ablation study have demonstrated the effectiveness of data-centric pipeline and Multi-Model(MM) interactive iterative strategies, achieving first place in lightweight text-to-SQL models (within 70B).", "date": "2025-10-27", "categories": []}
{"id": "2502.11546", "title": "DCAD-2000: A Multilingual Dataset across 2000+ Languages with Data Cleaning as Anomaly Detection", "abstract": "The rapid development of multilingual large language models (LLMs) highlights the need for high-quality, diverse, and well-curated multilingual datasets. In this paper, we introduce DCAD-2000 (Data Cleaning as Anomaly Detection), a large-scale multilingual corpus constructed from newly extracted Common Crawl data and existing multilingual sources. DCAD-2000 covers 2,282 languages, 46.72TB of text, and 8.63 billion documents, spanning 155 high- and medium-resource languages and 159 writing scripts. To overcome the limitations of existing data cleaning approaches, which rely on manually designed heuristic thresholds, we reframe data cleaning as an anomaly detection problem. This dynamic filtering paradigm substantially improves data quality by automatically identifying and removing noisy or anomalous content. By fine-tuning LLMs on DCAD-2000, we demonstrate notable improvements in data quality, robustness of the cleaning pipeline, and downstream performance, particularly for low-resource languages across multiple multilingual benchmarks.", "date": "2025-02-17", "categories": []}
{"id": "2508.01139", "title": "Dataset Condensation with Color Compensation", "abstract": "Dataset condensation always faces a constitutive trade-off: balancing performance and fidelity under extreme compression. Existing methods struggle with two bottlenecks: image-level selection methods (Coreset Selection, Dataset Quantization) suffer from inefficiency condensation, while pixel-level optimization (Dataset Distillation) introduces semantic distortion due to over-parameterization. With empirical observations, we find that a critical problem in dataset condensation is the oversight of color's dual role as an information carrier and a basic semantic representation unit. We argue that improving the colorfulness of condensed images is beneficial for representation learning. Motivated by this, we propose DC3: a Dataset Condensation framework with Color Compensation. After a calibrated selection strategy, DC3 utilizes the latent diffusion model to enhance the color diversity of an image rather than creating a brand-new one. Extensive experiments demonstrate the superior performance and generalization of DC3 that outperforms SOTA methods across multiple benchmarks. To the best of our knowledge, besides focusing on downstream tasks, DC3 is the first research to fine-tune pre-trained diffusion models with condensed datasets. The Frechet Inception Distance (FID) and Inception Score (IS) results prove that training networks with our high-quality datasets is feasible without model collapse or other degradation issues. Code and generated data are available at https://github.com/528why/Dataset-Condensation-with-Color-Compensation.", "date": "2025-08-02", "categories": []}
{"id": "2510.21286", "title": "Adaptive Data Selection for Multi-Layer Perceptron Training: A Sub-linear Value-Driven Method", "abstract": "Data selection is one of the fundamental problems in neural network training, particularly for multi-layer perceptrons (MLPs) where identifying the most valuable training samples from massive, multi-source, and heterogeneous data sources under budget constraints poses significant challenges. Existing data selection methods, including coreset construction, data Shapley values, and influence functions, suffer from critical limitations: they oversimplify nonlinear transformations, ignore informative intermediate representations in hidden layers, or fail to scale to larger MLPs due to high computational complexity. In response, we propose DVC (Data Value Contribution), a novel budget-aware method for evaluating and selecting data for MLP training that accounts for the dynamic evolution of network parameters during training. The DVC method decomposes data contribution into Layer Value Contribution (LVC) and Global Value Contribution (GVC), employing six carefully designed metrics and corresponding efficient algorithms to capture data characteristics across three dimensions--quality, relevance, and distributional diversity--at different granularities. DVC integrates these assessments with an Upper Confidence Bound (UCB) algorithm for adaptive source selection that balances exploration and exploitation. Extensive experiments across six datasets and eight baselines demonstrate that our method consistently outperforms existing approaches under various budget constraints, achieving superior accuracy and F1 scores. Our approach represents the first systematic treatment of hierarchical data evaluation for neural networks, providing both theoretical guarantees and practical advantages for large-scale machine learning systems.", "date": "2025-10-24", "categories": []}
{"id": "2505.24261", "title": "Taming Hyperparameter Sensitivity in Data Attribution: Practical Selection Without Costly Retraining", "abstract": "Data attribution methods, which quantify the influence of individual training data points on a machine learning model, have gained increasing popularity in data-centric applications in modern AI. Despite a recent surge of new methods developed in this space, the impact of hyperparameter tuning in these methods remains under-explored. In this work, we present the first large-scale empirical study to understand the hyperparameter sensitivity of common data attribution methods. Our results show that most methods are indeed sensitive to certain key hyperparameters. However, unlike typical machine learning algorithms -- whose hyperparameters can be tuned using computationally-cheap validation metrics -- evaluating data attribution performance often requires retraining models on subsets of training data, making such metrics prohibitively costly for hyperparameter tuning. This poses a critical open challenge for the practical application of data attribution methods. To address this challenge, we advocate for better theoretical understandings of hyperparameter behavior to inform efficient tuning strategies. As a case study, we provide a theoretical analysis of the regularization term that is critical in many variants of influence function methods. Building on this analysis, we propose a lightweight procedure for selecting the regularization value without model retraining, and validate its effectiveness across a range of standard data attribution benchmarks. Overall, our study identifies a fundamental yet overlooked challenge in the practical application of data attribution, and highlights the importance of careful discussion on hyperparameter selection in future method development.", "date": "2025-05-30", "categories": []}
{"id": "2508.21589", "title": "Middo: Model-Informed Dynamic Data Optimization for Enhanced LLM Fine-Tuning via Closed-Loop Learning", "abstract": "Supervised Fine-Tuning (SFT) Large Language Models (LLM) fundamentally rely on high-quality training data. While data selection and data synthesis are two common strategies to improve data quality, existing approaches often face limitations in static dataset curation that fail to adapt to evolving model capabilities. In this paper, we introduce Middo, a self-evolving Model-informed dynamic data optimization framework that uses model-aware data selection and context-preserving data refinement. Unlike conventional one-off filtering/synthesis methods, our framework establishes a closed-loop optimization system: (1) A self-referential diagnostic module proactively identifies suboptimal samples through tri-axial model signals - loss patterns (complexity), embedding cluster dynamics (diversity), and self-alignment scores (quality); (2) An adaptive optimization engine then transforms suboptimal samples into pedagogically valuable training points while preserving semantic integrity; (3) This optimization process continuously evolves with model capability through dynamic learning principles. Experiments on multiple benchmarks demonstrate that our Middo consistently enhances the quality of seed data and boosts LLM's performance with improving accuracy by 7.15% on average while maintaining the original dataset scale. This work establishes a new paradigm for sustainable LLM training through dynamic human-AI co-evolution of data and models. Our datasets, models, and code are publicly available at https://github.com/Word2VecT/Middo.", "date": "2025-08-29", "categories": []}
{"id": "2510.18909", "title": "Learning from the Best, Differently: A Diversity-Driven Rethinking on Data Selection", "abstract": "High-quality pre-training data is crutial for large language models, where quality captures factual reliability and semantic value, and diversity ensures broad coverage and distributional heterogeneity. Existing approaches typically rely on single or multiple-dimensional score-based selection. However, directly selecting top-scored data often degrades performance, and sampling from a broader range is required to recover results. The above non-monotonicity between dataset scores and downstream benchmark results reveals a fundamental bias: score-based methods collapse correlated dimensions, causing top-scored data to appear high-quality while systematically overlooking diversity. We argue that ensuring diversity requires decomposing correlated metrics into orthogonal feature dimensions, from which the top-scored data can be directly selected. Therefore, we proposed the Orthogonal Diversity-Aware Selection (ODiS) algorithm, which preserves both quality and diversity during data selection. First, ODiS evaluates data from multiple dimensions, covering language quality, knowledge quality, and comprehension difficulty. The multi-dimensional scores are then decorrelated via Principal Component Analysis (PCA), yielding orthogonal evaluation dimensions. For each dimension, a Roberta-based scorer is trained to regress the data onto PCA-projected scores, enabling scalable inference on large corpora. Finally, ODiS constructs the training dataset by selecting top-scored data within each orthogonal dimension, thereby ensuring both quality and diversity. Empirical results show that ODiS-selected data exhibit less than 2\\% inter-dimension overlap, confirming orthogonality between dimensions. More importantly, models trained with ODiS-selected data significantly outperform other baselines on downstream benchmarks, highlighting the necessity of orthogonal, diversity-aware data selection for LLMs.", "date": "2025-10-21", "categories": []}
{"id": "2502.02494", "title": "Analyzing Similarity Metrics for Data Selection for Language Model Pretraining", "abstract": "Measuring similarity between training examples is critical for curating high-quality and diverse pretraining datasets for language models. However, similarity is typically computed with a generic off-the-shelf embedding model that has been trained for tasks such as retrieval. Whether these embedding-based similarity metrics are well-suited for pretraining data selection remains largely unexplored. In this paper, we propose a new framework to assess the suitability of a similarity metric specifically for data curation in language model pretraining applications. Our framework's first evaluation criterion captures how well distances reflect generalization in pretraining loss between different training examples. Next, we use each embedding model to guide a standard diversity-based data curation algorithm and measure its utility by pretraining a language model on the selected data and evaluating downstream task performance. Finally, we evaluate the capabilities of embeddings to distinguish between examples from different data sources. With these evaluations, we demonstrate that standard off-the-shelf embedding models are not well-suited for the pretraining data curation setting, underperforming even remarkably simple embeddings that are extracted from models trained on the same pretraining corpus. Our experiments are performed on the Pile, for pretraining a 1.7B parameter language model on 200B tokens. We believe our analysis and evaluation framework serves as a foundation for the future design of embeddings that specifically reason about similarity in pretraining datasets.", "date": "2025-02-04", "categories": []}
{"id": "2507.11690", "title": "The Impact of Coreset Selection on Spurious Correlations and Group Robustness", "abstract": "Coreset selection methods have shown promise in reducing the training data size while maintaining model performance for data-efficient machine learning. However, as many datasets suffer from biases that cause models to learn spurious correlations instead of causal features, it is important to understand whether and how dataset reduction methods may perpetuate, amplify, or mitigate these biases. In this work, we conduct the first comprehensive analysis of the implications of data selection on the spurious bias levels of the selected coresets and the robustness of downstream models trained on them. We use an extensive experimental setting spanning ten different spurious correlations benchmarks, five score metrics to characterize sample importance/ difficulty, and five data selection policies across a broad range of coreset sizes. Thereby, we unravel a series of nontrivial nuances in interactions between sample difficulty and bias alignment, as well as dataset bias and resultant model robustness. For example, we find that selecting coresets using embedding-based sample characterization scores runs a comparatively lower risk of inadvertently exacerbating bias than selecting using characterizations based on learning dynamics. Most importantly, our analysis reveals that although some coreset selection methods could achieve lower bias levels by prioritizing difficult samples, they do not reliably guarantee downstream robustness.", "date": "2025-07-15", "categories": []}
{"id": "2510.17847", "title": "CoIDO: Efficient Data Selection for Visual Instruction Tuning via Coupled Importance-Diversity Optimization", "abstract": "Multimodal large language models (MLLMs) rely heavily on instruction tuning to align vision and language capabilities, yet the computational cost of training on large-scale datasets remains a major bottleneck. Existing data selection methods aim to mitigate this by selecting important and diverse subsets, but they often suffer from two critical drawbacks: high computational overhead from processing the entire dataset and suboptimal data selection due to separate treatment of importance and diversity.\n  We introduce CoIDO, a novel dual-objective framework that jointly optimizes data importance and diversity to overcome these challenges. Unlike existing approaches that require costly evaluations across the whole dataset, CoIDO employs a lightweight plug-in scorer. This scorer is trained on just a small random sample of data to learn the distribution of the candidate set, drastically reducing computational demands. By leveraging a homoscedastic uncertainty-based formulation, CoIDO effectively balances importance and diversity during training, enabling efficient and scalable data selection.\n  In our experiments, we trained the CoIDO scorer using only 20 percent of randomly sampled data. Once trained, CoIDO was applied to the entire dataset to select a 20 percent subset for instruction tuning. On the widely used LLaVA-1.5-7B model across ten downstream tasks, this selected subset achieved an impressive 98.2 percent of the performance of full-data fine-tuning, on average.", "date": "2025-10-11", "categories": []}
{"id": "2510.18143", "title": "Learning from Generalization Patterns: An Evaluation-Driven Approach to Enhanced Data Augmentation for Fine-Tuning Small Language Models", "abstract": "Small Language Models (SLMs) offer compelling advantages in deployment cost and latency, but their accuracy often lags behind larger models, particularly for complex domain-specific tasks. While supervised fine-tuning can help bridge this performance gap, it requires substantial manual effort in data preparation and iterative optimization. We present PaDA-Agent (Pattern-guided Data Augmentation Agent), an evaluation-driven approach that streamlines the data augmentation process for SLMs through coordinated operations. Unlike state-of-the-art approaches that focus on model training errors only and generating error-correcting samples, PaDA-Agent discovers failure patterns from the validation data via evaluations and drafts targeted data augmentation strategies aiming to directly reduce the generalization gap. Our experimental results demonstrate significant improvements over state-of-the-art LLM-based data augmentation approaches for Llama 3.2 1B Instruct model fine-tuning.", "date": "2025-10-20", "categories": []}
{"id": "2510.18583", "title": "CovMatch: Cross-Covariance Guided Multimodal Dataset Distillation with Trainable Text Encoder", "abstract": "Multimodal dataset distillation aims to synthesize a small set of image-text pairs that enables efficient training of large-scale vision-language models. While dataset distillation has shown promise in unimodal tasks, extending it to multimodal contrastive learning presents key challenges: learning cross-modal alignment and managing the high computational cost of large encoders. Prior approaches address scalability by freezing the text encoder and update only the image encoder and text projection layer. However, we find this severely limits semantic alignment and becomes a bottleneck for performance scaling. We propose CovMatch, a scalable dataset distillation framework that aligns the cross-covariance of real and synthetic features while regularizing feature distributions within each modality. Unlike prior approaches, CovMatch enables joint optimization of both encoders, leading to stronger cross-modal alignment and improved performance. Evaluated on Flickr30K and COCO, CovMatch outperforms state-of-the-art multimodal distillation methods and achieves up to 6.8% absolute gains in retrieval accuracy using only 500 synthetic pairs.", "date": "2025-10-21", "categories": []}
{"id": "2510.15970", "title": "Predict Training Data Quality via Its Geometry in Metric Space", "abstract": "High-quality training data is the foundation of machine learning and artificial intelligence, shaping how models learn and perform. Although much is known about what types of data are effective for training, the impact of the data's geometric structure on model performance remains largely underexplored. We propose that both the richness of representation and the elimination of redundancy within training data critically influence learning outcomes. To investigate this, we employ persistent homology to extract topological features from data within a metric space, thereby offering a principled way to quantify diversity beyond entropy-based measures. Our findings highlight persistent homology as a powerful tool for analyzing and enhancing the training data that drives AI systems.", "date": "2025-10-12", "categories": []}
{"id": "2510.16007", "title": "Layer-Aware Influence for Online Data Valuation Estimation", "abstract": "Data-centric learning emphasizes curating high-quality training samples to boost performance rather than designing new architectures. A central problem is to estimate the influence of training sample efficiently. Prior studies largely focus on static influence measured on a converged model, overlooking how data valuation dynamically changes during optimization. This omission neglects the dynamic nature of sample influence during optimization, especially in deep models. To address the computational burden of frequent influence estimation, we develop a layer-aware online estimator that requires only loss-to-output gradients. This design avoids parameter-level and full-network gradients while preserving ranking fidelity. Extensive experiments across LLM pretraining, fine-tuning, and image classification show our method improves accuracy with substantially lower time and memory cost, making dynamic data curation efficient and scalable in practice.", "date": "2025-10-14", "categories": []}
{"id": "2510.16207", "title": "Data-Centric AI for Tropical Agricultural Mapping: Challenges, Strategies and Scalable Solutions", "abstract": "Mapping agriculture in tropical areas through remote sensing presents unique challenges, including the lack of high-quality annotated data, the elevated costs of labeling, data variability, and regional generalisation. This paper advocates a Data-Centric Artificial Intelligence (DCAI) perspective and pipeline, emphasizing data quality and curation as key drivers for model robustness and scalability. It reviews and prioritizes techniques such as confident learning, core-set selection, data augmentation, and active learning. The paper highlights the readiness and suitability of 25 distinct strategies in large-scale agricultural mapping pipelines. The tropical context is of high interest, since high cloudiness, diverse crop calendars, and limited datasets limit traditional model-centric approaches. This tutorial outlines practical solutions as a data-centric approach for curating and training AI models better suited to the dynamic realities of tropical agriculture. Finally, we propose a practical pipeline using the 9 most mature and straightforward methods that can be applied to a large-scale tropical agricultural mapping project.", "date": "2025-10-17", "categories": []}
{"id": "2505.24623", "title": "Hyperbolic Dataset Distillation", "abstract": "To address the computational and storage challenges posed by large-scale datasets in deep learning, dataset distillation has been proposed to synthesize a compact dataset that replaces the original while maintaining comparable model performance. Unlike optimization-based approaches that require costly bi-level optimization, distribution matching (DM) methods improve efficiency by aligning the distributions of synthetic and original data, thereby eliminating nested optimization. DM achieves high computational efficiency and has emerged as a promising solution. However, existing DM methods, constrained to Euclidean space, treat data as independent and identically distributed points, overlooking complex geometric and hierarchical relationships. To overcome this limitation, we propose a novel hyperbolic dataset distillation method, termed HDD. Hyperbolic space, characterized by negative curvature and exponential volume growth with distance, naturally models hierarchical and tree-like structures. HDD embeds features extracted by a shallow network into the Lorentz hyperbolic space, where the discrepancy between synthetic and original data is measured by the hyperbolic (geodesic) distance between their centroids. By optimizing this distance, the hierarchical structure is explicitly integrated into the distillation process, guiding synthetic samples to gravitate towards the root-centric regions of the original data distribution while preserving their underlying geometric characteristics. Furthermore, we find that pruning in hyperbolic space requires only 20% of the distilled core set to retain model performance, while significantly improving training stability. To the best of our knowledge, this is the first work to incorporate the hyperbolic space into the dataset distillation process. The code is available at https://github.com/Guang000/HDD.", "date": "2025-05-30", "categories": []}
{"id": "2509.14777", "title": "Dataset Distillation for Super-Resolution without Class Labels and Pre-trained Models", "abstract": "Training deep neural networks has become increasingly demanding, requiring large datasets and significant computational resources, especially as model complexity advances. Data distillation methods, which aim to improve data efficiency, have emerged as promising solutions to this challenge. In the field of single image super-resolution (SISR), the reliance on large training datasets highlights the importance of these techniques. Recently, a generative adversarial network (GAN) inversion-based data distillation framework for SR was proposed, showing potential for better data utilization. However, the current method depends heavily on pre-trained SR networks and class-specific information, limiting its generalizability and applicability. To address these issues, we introduce a new data distillation approach for image SR that does not need class labels or pre-trained SR models. In particular, we first extract high-gradient patches and categorize images based on CLIP features, then fine-tune a diffusion model on the selected patches to learn their distribution and synthesize distilled training images. Experimental results show that our method achieves state-of-the-art performance while using significantly less training data and requiring less computational time. Specifically, when we train a baseline Transformer model for SR with only 0.68\\% of the original dataset, the performance drop is just 0.3 dB. In this case, diffusion model fine-tuning takes 4 hours, and SR model training completes within 1 hour, much shorter than the 11-hour training time with the full dataset.", "date": "2025-09-18", "categories": []}
{"id": "2510.05805", "title": "Improving Clinical Dataset Condensation with Mode Connectivity-based Trajectory Surrogates", "abstract": "Dataset condensation (DC) enables the creation of compact, privacy-preserving synthetic datasets that can match the utility of real patient records, supporting democratised access to highly regulated clinical data for developing downstream clinical models. State-of-the-art DC methods supervise synthetic data by aligning the training dynamics of models trained on real and those trained on synthetic data, typically using full stochastic gradient descent (SGD) trajectories as alignment targets; however, these trajectories are often noisy, high-curvature, and storage-intensive, leading to unstable gradients, slow convergence, and substantial memory overhead. We address these limitations by replacing full SGD trajectories with smooth, low-loss parametric surrogates, specifically quadratic B\\'ezier curves that connect the initial and final model states from real training trajectories. These mode-connected paths provide noise-free, low-curvature supervision signals that stabilise gradients, accelerate convergence, and eliminate the need for dense trajectory storage. We theoretically justify B\\'ezier-mode connections as effective surrogates for SGD paths and empirically show that the proposed method outperforms state-of-the-art condensation approaches across five clinical datasets, yielding condensed datasets that enable clinically effective model development.", "date": "2025-10-07", "categories": []}
{"id": "2510.15162", "title": "Train a Unified Multimodal Data Quality Classifier with Synthetic Data", "abstract": "The Multimodal Large Language Models (MLLMs) are continually pre-trained on a mixture of image-text caption data and interleaved document data, while the high-quality data filtering towards image-text interleaved document data is under-explored. We propose to train an efficient MLLM as a Unified Mulitmodal Data Quality Classifier to Filter both high-quality image-text caption and interleaved data (UniFilter). To address the challenge of collecting diverse labeled multimodal data, we introduce a semi-synthetic approach that leverages readily available raw images and generates corresponding text across four quality levels. This method enables efficient creation of sample-score pairs for both caption and interleaved document data to train UniFilter. We apply UniFilter to curate high-quality caption data from DataComp caption dataset and interleaved data from the OBELICS image-text interleaved dataset. MLLMs pre-trained on the filtered data demonstrate significantly enhanced capabilities compared to those trained on baseline-filtered data, achieving stronger zero-shot reasoning and in-context learning capabilities. After visual supervised fine-tuning, these UniFilter-induced MLLMs achieve stronger performance on various benchmarks, highlighting the downstream benefits of high-quality multimodal pre-training. We release the synthetic training data used for training UniFilter, the UniFilter model checkpoints, and the high-quality interleaved document subset OBELICS-HQ, curated by UniFilter, to the community for reproduction and further development.", "date": "2025-10-16", "categories": []}
{"id": "2510.13892", "title": "The Harder The Better: Maintaining Supervised Fine-tuning Generalization with Less but Harder Data", "abstract": "Large Language Models (LLMs) excel in general tasks, but adapting them to specialized domains relies on high-quality supervised fine-tuning (SFT) data. Although existing methods can identify subsets of high-quality data and reduce training cost to some extent, their selection process still suffers from over-reliance on LLMs' internal knowledge, weak interpretability, and limited generalization. To address these limitations, we propose THTB (The Harder The Better), a cognitive science-inspired framework for instruction data selection and annotation guidance. THTB prioritizes higher-level cognitive instructions by combining quality filtering with intrinsic and extrinsic hardness scoring, offering interpretable and quantifiable criteria for efficient SFT, both in data selection and annotation guidance. Experiments show that THTB enables models trained on only 5% of the data to outperform full-dataset training, while achieving superior generalization compared with LLM-only selection. In addition, THTB provides effective annotation guidance in vertical domains, enabling a model trained on just 2% of the data to surpass models trained on much larger datasets, demonstrating strong potential for domain adaptation. Our code, datasets, and models are available on https://github.com/DYJG-research/THTB.", "date": "2025-10-14", "categories": []}
{"id": "2405.09591", "title": "A Comprehensive Survey on Data Augmentation", "abstract": "Data augmentation is a series of techniques that generate high-quality artificial data by manipulating existing data samples. By leveraging data augmentation techniques, AI models can achieve significantly improved applicability in tasks involving scarce or imbalanced datasets, thereby substantially enhancing AI models' generalization capabilities. Existing literature surveys only focus on a certain type of specific modality data and categorize these methods from modality-specific and operation-centric perspectives, which lacks a consistent summary of data augmentation methods across multiple modalities and limits the comprehension of how existing data samples serve the data augmentation process. To bridge this gap, this survey proposes a more enlightening taxonomy that encompasses data augmentation techniques for different common data modalities by investigating how to take advantage of the intrinsic relationship between and within instances. Additionally, it categorizes data augmentation methods across five data modalities through a unified inductive approach.", "date": "2024-05-15", "categories": []}
{"id": "2411.19623", "title": "FairDD: Fair Dataset Distillation", "abstract": "Condensing large datasets into smaller synthetic counterparts has demonstrated its promise for image classification. However, previous research has overlooked a crucial concern in image recognition: ensuring that models trained on condensed datasets are unbiased towards protected attributes (PA), such as gender and race. Our investigation reveals that dataset distillation fails to alleviate the unfairness towards minority groups within original datasets. Moreover, this bias typically worsens in the condensed datasets due to their smaller size. To bridge the research gap, we propose a novel fair dataset distillation (FDD) framework, namely FairDD, which can be seamlessly applied to diverse matching-based DD approaches (DDs), requiring no modifications to their original architectures. The key innovation of FairDD lies in synchronously matching synthetic datasets to PA-wise groups of original datasets, rather than indiscriminate alignment to the whole distributions in vanilla DDs, dominated by majority groups. This synchronized matching allows synthetic datasets to avoid collapsing into majority groups and bootstrap their balanced generation to all PA groups. Consequently, FairDD could effectively regularize vanilla DDs to favor biased generation toward minority groups while maintaining the accuracy of target attributes. Theoretical analyses and extensive experimental evaluations demonstrate that FairDD significantly improves fairness compared to vanilla DDs, with a promising trade-off between fairness and accuracy. Its consistent superiority across diverse DDs, spanning Distribution and Gradient Matching, establishes it as a versatile FDD approach. Code is available at https://github.com/zqhang/FairDD.", "date": "2024-11-29", "categories": []}
{"id": "2503.02359", "title": "Add-One-In: Incremental Sample Selection for Large Language Models via a Choice-Based Greedy Paradigm", "abstract": "Selecting high-quality and diverse training samples from extensive datasets plays a crucial role in reducing training overhead and enhancing the performance of Large Language Models (LLMs). However, existing studies fall short in assessing the overall value of selected data, focusing primarily on individual quality, and struggle to strike an effective balance between ensuring diversity and minimizing data point traversals. Therefore, this paper introduces a novel choice-based sample selection framework that shifts the focus from evaluating individual sample quality to comparing the contribution value of different samples when incorporated into the subset. Thanks to the advanced language understanding capabilities of LLMs, we utilize LLMs to evaluate the value of each option during the selection process. Furthermore, we design a greedy sampling process where samples are incrementally added to the subset, thereby improving efficiency by eliminating the need for exhaustive traversal of the entire dataset with the limited budget. Extensive experiments demonstrate that selected data from our method not only surpasses the performance of the full dataset but also achieves competitive results with recent powerful studies, while requiring fewer selections. Moreover, we validate our approach on a larger medical dataset, highlighting its practical applicability in real-world applications. Our code and data are available at https://github.com/BIRlz/comperative_sample_selection.", "date": "2025-03-04", "categories": []}
{"id": "2505.14117", "title": "Collaborative Unlabeled Data Optimization", "abstract": "This paper pioneers a novel data-centric paradigm to maximize the utility of unlabeled data, tackling a critical question: How can we enhance the efficiency and sustainability of deep learning training by optimizing the data itself? We begin by identifying three key limitations in existing model-centric approaches, all rooted in a shared bottleneck: knowledge extracted from data is locked to model parameters, hindering its reusability and scalability. To this end, we propose CoOpt, a highly efficient, parallelized framework for collaborative unlabeled data optimization, thereby effectively encoding knowledge into the data itself. By distributing unlabeled data and leveraging publicly available task-agnostic models, CoOpt facilitates scalable, reusable, and sustainable training pipelines. Extensive experiments across diverse datasets and architectures demonstrate its efficacy and efficiency, achieving 13.6% and 6.8% improvements on Tiny-ImageNet and ImageNet-1K, respectively, with training speedups of $1.94 \\times $ and $1.2 \\times$.", "date": "2025-05-20", "categories": []}
{"id": "2510.06250", "title": "Scalable multilingual PII annotation for responsible AI in LLMs", "abstract": "As Large Language Models (LLMs) gain wider adoption, ensuring their reliable handling of Personally Identifiable Information (PII) across diverse regulatory contexts has become essential. This work introduces a scalable multilingual data curation framework designed for high-quality PII annotation across 13 underrepresented locales, covering approximately 336 locale-specific PII types. Our phased, human-in-the-loop annotation methodology combines linguistic expertise with rigorous quality assurance, leading to substantial improvements in recall and false positive rates from pilot, training, and production phases. By leveraging inter-annotator agreement metrics and root-cause analysis, the framework systematically uncovers and resolves annotation inconsistencies, resulting in high-fidelity datasets suitable for supervised LLM fine-tuning. Beyond reporting empirical gains, we highlight common annotator challenges in multilingual PII labeling and demonstrate how iterative, analytics-driven pipelines can enhance both annotation quality and downstream model reliability.", "date": "2025-10-03", "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": "2403.03881", "title": "Unlocking Dataset Distillation with Diffusion Models", "abstract": "Dataset distillation seeks to condense datasets into smaller but highly representative synthetic samples. While diffusion models now lead all generative benchmarks, current distillation methods avoid them and rely instead on GANs or autoencoders, or, at best, sampling from a fixed diffusion prior. This trend arises because naive backpropagation through the long denoising chain leads to vanishing gradients, which prevents effective synthetic sample optimization. To address this limitation, we introduce Latent Dataset Distillation with Diffusion Models (LD3M), the first method to learn gradient-based distilled latents and class embeddings end-to-end through a pre-trained latent diffusion model. A linearly decaying skip connection, injected from the initial noisy state into every reverse step, preserves the gradient signal across dozens of timesteps without requiring diffusion weight fine-tuning. Across multiple ImageNet subsets at 128x128 and 256x256, LD3M improves downstream accuracy by up to 4.8 percentage points (1 IPC) and 4.2 points (10 IPC) over the prior state-of-the-art. The code for LD3M is provided at https://github.com/Brian-Moser/prune_and_distill.", "date": "2024-03-06", "categories": []}
{"id": "2411.17063", "title": "Contrastive Graph Condensation: Advancing Data Versatility through Self-Supervised Learning", "abstract": "With the increasing computation of training graph neural networks (GNNs) on large-scale graphs, graph condensation (GC) has emerged as a promising solution to synthesize a compact, substitute graph of the large-scale original graph for efficient GNN training. However, existing GC methods predominantly employ classification as the surrogate task for optimization, thus excessively relying on node labels and constraining their utility in label-sparsity scenarios. More critically, this surrogate task tends to overfit class-specific information within the condensed graph, consequently restricting the generalization capabilities of GC for other downstream tasks. To address these challenges, we introduce Contrastive Graph Condensation (CTGC), which adopts a self-supervised surrogate task to extract critical, causal information from the original graph and enhance the cross-task generalizability of the condensed graph. Specifically, CTGC employs a dual-branch framework to disentangle the generation of the node attributes and graph structures, where a dedicated structural branch is designed to explicitly encode geometric information through nodes' positional embeddings. By implementing an alternating optimization scheme with contrastive loss terms, CTGC promotes the mutual enhancement of both branches and facilitates high-quality graph generation through the model inversion technique. Extensive experiments demonstrate that CTGC excels in handling various downstream tasks with a limited number of labels, consistently outperforming state-of-the-art GC methods.", "date": "2024-11-26", "categories": []}
{"id": "2510.05722", "title": "Data Factory with Minimal Human Effort Using VLMs", "abstract": "Generating enough and diverse data through augmentation offers an efficient solution to the time-consuming and labour-intensive process of collecting and annotating pixel-wise images. Traditional data augmentation techniques often face challenges in manipulating high-level semantic attributes, such as materials and textures. In contrast, diffusion models offer a robust alternative, by effectively utilizing text-to-image or image-to-image transformation. However, existing diffusion-based methods are either computationally expensive or compromise on performance. To address this issue, we introduce a novel training-free pipeline that integrates pretrained ControlNet and Vision-Language Models (VLMs) to generate synthetic images paired with pixel-level labels. This approach eliminates the need for manual annotations and significantly improves downstream tasks. To improve the fidelity and diversity, we add a Multi-way Prompt Generator, Mask Generator and High-quality Image Selection module. Our results on PASCAL-5i and COCO-20i present promising performance and outperform concurrent work for one-shot semantic segmentation.", "date": "2025-10-07", "categories": []}
{"id": "2510.03950", "title": "What Is The Performance Ceiling of My Classifier? Utilizing Category-Wise Influence Functions for Pareto Frontier Analysis", "abstract": "Data-centric learning seeks to improve model performance from the perspective of data quality, and has been drawing increasing attention in the machine learning community. Among its key tools, influence functions provide a powerful framework to quantify the impact of individual training samples on model predictions, enabling practitioners to identify detrimental samples and retrain models on a cleaner dataset for improved performance. However, most existing work focuses on the question: \"what data benefits the learning model?\" In this paper, we take a step further and investigate a more fundamental question: \"what is the performance ceiling of the learning model?\" Unlike prior studies that primarily measure improvement through overall accuracy, we emphasize category-wise accuracy and aim for Pareto improvements, ensuring that every class benefits, rather than allowing tradeoffs where some classes improve at the expense of others. To address this challenge, we propose category-wise influence functions and introduce an influence vector that quantifies the impact of each training sample across all categories. Leveraging these influence vectors, we develop a principled criterion to determine whether a model can still be improved, and further design a linear programming-based sample reweighting framework to achieve Pareto performance improvements. Through extensive experiments on synthetic datasets, vision, and text benchmarks, we demonstrate the effectiveness of our approach in estimating and achieving a model's performance improvement across multiple categories of interest.", "date": "2025-10-04", "categories": []}
{"id": "2510.00386", "title": "Train on Validation (ToV): Fast data selection with applications to fine-tuning", "abstract": "State-of-the-art machine learning often follows a two-stage process: $(i)$~pre-training on large, general-purpose datasets; $(ii)$~fine-tuning on task-specific data. In fine-tuning, selecting training examples that closely reflect the target distribution is crucial. However, it is often the case that only a few samples are available from the target distribution. Existing data selection methods treat these target samples as a validation set and estimate the effect of adding or removing a single sample from the training pool by performing inference on the validation set.\n  We propose a simpler and faster alternative that inverts the usual role of train and validation: we perform inference on the training pool before and after fine-tuning on the validation set. We then select samples whose predictions change the most. Our key insight is that the training samples most affected by fine-tuning on a small validation set tend to be the most beneficial for reducing test loss on the target distribution. Experiments on instruction tuning and named entity recognition tasks show that, in most cases, our method achieves lower test log-loss than state-of-the-art approaches. We support our findings with theoretical analysis.", "date": "2025-10-01", "categories": []}
{"id": "2510.00434", "title": "On-the-Fly Data Augmentation via Gradient-Guided and Sample-Aware Influence Estimation", "abstract": "Data augmentation has been widely employed to improve the generalization of deep neural networks. Most existing methods apply fixed or random transformations. However, we find that sample difficulty evolves along with the model's generalization capabilities in dynamic training environments. As a result, applying uniform or stochastic augmentations, without accounting for such dynamics, can lead to a mismatch between augmented data and the model's evolving training needs, ultimately degrading training effectiveness. To address this, we introduce SADA, a Sample-Aware Dynamic Augmentation that performs on-the-fly adjustment of augmentation strengths based on each sample's evolving influence on model optimization. Specifically, we estimate each sample's influence by projecting its gradient onto the accumulated model update direction and computing the temporal variance within a local training window. Samples with low variance, indicating stable and consistent influence, are augmented more strongly to emphasize diversity, while unstable samples receive milder transformations to preserve semantic fidelity and stabilize learning. Our method is lightweight, which does not require auxiliary models or policy tuning. It can be seamlessly integrated into existing training pipelines as a plug-and-play module. Experiments across various benchmark datasets and model architectures show consistent improvements of SADA, including +7.3\\% on fine-grained tasks and +4.3\\% on long-tailed datasets, highlighting the method's effectiveness and practicality.", "date": "2025-10-01", "categories": []}
{"id": "2509.26291", "title": "Representation-Based Data Quality Audits for Audio", "abstract": "Data quality issues such as off-topic samples, near duplicates, and label errors often limit the performance of audio-based systems. This paper addresses these issues by adapting SelfClean, a representation-to-rank data auditing framework, from the image to the audio domain. This approach leverages self-supervised audio representations to identify common data quality issues, creating ranked review lists that surface distinct issues within a single, unified process. The method is benchmarked on the ESC-50, GTZAN, and a proprietary industrial dataset, using both synthetic and naturally occurring corruptions. The results demonstrate that this framework achieves state-of-the-art ranking performance, often outperforming issue-specific baselines and enabling significant annotation savings by efficiently guiding human review.", "date": "2025-09-30", "categories": []}
{"id": "2412.02366", "title": "GenMix: Effective Data Augmentation with Generative Diffusion Model Image Editing", "abstract": "Data augmentation is widely used to enhance generalization in visual classification tasks. However, traditional methods struggle when source and target domains differ, as in domain adaptation, due to their inability to address domain gaps. This paper introduces GenMix, a generalizable prompt-guided generative data augmentation approach that enhances both in-domain and cross-domain image classification. Our technique leverages image editing to generate augmented images based on custom conditional prompts, designed specifically for each problem type. By blending portions of the input image with its edited generative counterpart and incorporating fractal patterns, our approach mitigates unrealistic images and label ambiguity, improving the performance and adversarial robustness of the resulting models. Efficacy of our method is established with extensive experiments on eight public datasets for general and fine-grained classification, in both in-domain and cross-domain settings. Additionally, we demonstrate performance improvements for self-supervised learning, learning with data scarcity, and adversarial robustness. As compared to the existing state-of-the-art methods, our technique achieves stronger performance across the board.", "date": "2024-12-03", "categories": []}
{"id": "2509.23461", "title": "Data-Efficient Training by Evolved Sampling", "abstract": "Data selection is designed to accelerate learning with preserved performance. To achieve this, a fundamental thought is to identify informative data samples with significant contributions to the training. In this work, we propose \\textbf{Evolved Sampling} (\\textbf{ES}), a simple yet effective framework for \\emph{dynamic} sampling along the training process. This method conducts \\em batch \\em level data selection based on the dynamics of losses and augmented \\emph{loss differences}, which enables flexible \\emph{frequency tuning}, and hence significantly reduces the back propagation time with maintained model performance. Due to its conciseness, ES is also readily extensible to incorporate \\em set \\em level data selection (to form ES with pruning, \\textbf{ESWP}) for further accelerations. As a plug-and-play framework, ES(WP) consistently achieves lossless training accelerations across various pre-training and post-training tasks, saving up to nearly 45\\% wall-clock time. Our results motivate further investigations on the data efficiency aspect of modern large-scale machine learning.", "date": "2025-09-27", "categories": []}
{"id": "2509.24177", "title": "High-Order Progressive Trajectory Matching for Medical Image Dataset Distillation", "abstract": "Medical image analysis faces significant challenges in data sharing due to privacy regulations and complex institutional protocols. Dataset distillation offers a solution to address these challenges by synthesizing compact datasets that capture essential information from real, large medical datasets. Trajectory matching has emerged as a promising methodology for dataset distillation; however, existing methods primarily focus on terminal states, overlooking crucial information in intermediate optimization states. We address this limitation by proposing a shape-wise potential that captures the geometric structure of parameter trajectories, and an easy-to-complex matching strategy that progressively addresses parameters based on their complexity. Experiments on medical image classification tasks demonstrate that our method improves distillation performance while preserving privacy and maintaining model accuracy comparable to training on the original datasets. Our code is available at https://github.com/Bian-jh/HoP-TM.", "date": "2025-09-29", "categories": []}
{"id": "2509.24420", "title": "A Data-Centric Perspective on the Influence of Image Data Quality in Machine Learning Models", "abstract": "In machine learning, research has traditionally focused on model development, with relatively less attention paid to training data. As model architectures have matured and marginal gains from further refinements diminish, data quality has emerged as a critical factor. However, systematic studies on evaluating and ensuring dataset quality in the image domain remain limited.\n  This study investigates methods for systematically assessing image dataset quality and examines how various image quality factors influence model performance. Using the publicly available and relatively clean CIFAKE dataset, we identify common quality issues and quantify their impact on training. Building on these findings, we develop a pipeline that integrates two community-developed tools, CleanVision and Fastdup. We analyze their underlying mechanisms and introduce several enhancements, including automatic threshold selection to detect problematic images without manual tuning.\n  Experimental results demonstrate that not all quality issues exert the same level of impact. While convolutional neural networks show resilience to certain distortions, they are particularly vulnerable to degradations that obscure critical visual features, such as blurring and severe downscaling. To assess the performance of existing tools and the effectiveness of our proposed enhancements, we formulate the detection of low-quality images as a binary classification task and use the F1 score as the evaluation metric. Our automatic thresholding method improves the F1 score from 0.6794 to 0.9468 under single perturbations and from 0.7447 to 0.8557 under dual perturbations. For near-duplicate detection, our deduplication strategy increases the F1 score from 0.4576 to 0.7928. These results underscore the effectiveness of our workflow and provide a foundation for advancing data quality assessment in image-based machine learning.", "date": "2025-09-29", "categories": []}
{"id": "2410.13674", "title": "Diffusion Curriculum: Synthetic-to-Real Data Curriculum via Image-Guided Diffusion", "abstract": "Low-quality or scarce data has posed significant challenges for training deep neural networks in practice. While classical data augmentation cannot contribute very different new data, diffusion models opens up a new door to build self-evolving AI by generating high-quality and diverse synthetic data through text-guided prompts. However, text-only guidance cannot control synthetic images' proximity to the original images, resulting in out-of-distribution data detrimental to the model performance. To overcome the limitation, we study image guidance to achieve a spectrum of interpolations between synthetic and real images. With stronger image guidance, the generated images are similar to the training data but hard to learn. While with weaker image guidance, the synthetic images will be easier for model but contribute to a larger distribution gap with the original data. The generated full spectrum of data enables us to build a novel \"Diffusion Curriculum (DisCL)\". DisCL adjusts the image guidance level of image synthesis for each training stage: It identifies and focuses on hard samples for the model and assesses the most effective guidance level of synthetic images to improve hard data learning. We apply DisCL to two challenging tasks: long-tail (LT) classification and learning from low-quality data. It focuses on lower-guidance images of high-quality to learn prototypical features as a warm-up of learning higher-guidance images that might be weak on diversity or quality. Extensive experiments showcase a gain of 2.7% and 2.1% in OOD and ID macro-accuracy when applying DisCL to iWildCam dataset. On ImageNet-LT, DisCL improves the base model's tail-class accuracy from 4.4% to 23.64% and leads to a 4.02% improvement in all-class accuracy.", "date": "2024-10-17", "categories": []}
{"id": "2504.05563", "title": "Do Data Valuations Make Good Data Prices?", "abstract": "As large language models increasingly rely on external data sources, compensating data contributors has become a central concern. But how should these payments be devised? We revisit data valuations from a $\\textit{market-design perspective}$ where payments serve to compensate data owners for the $\\textit{private}$ heterogeneous costs they incur for collecting and sharing data. We show that popular valuation methods-such as Leave-One-Out and Data Shapley-make for poor payments. They fail to ensure truthful reporting of the costs, leading to $\\textit{inefficient market}$ outcomes. To address this, we adapt well-established payment rules from mechanism design, namely Myerson and Vickrey-Clarke-Groves (VCG), to the data market setting. We show that Myerson payment is the minimal truthful mechanism, optimal from the buyer's perspective. Additionally, we identify a condition under which both data buyers and sellers are utility-satisfied, and the market achieves efficiency. Our findings highlight the importance of incorporating incentive compatibility into data valuation design, paving the way for more robust and efficient data markets. Our data market framework is readily applicable to real-world scenarios. We illustrate this with simulations of contributor compensation in an LLM based retrieval-augmented generation (RAG) marketplace tasked with challenging medical question answering.", "date": "2025-04-07", "categories": []}
{"id": "2504.10563", "title": "Data Augmentation Through Random Style Replacement", "abstract": "In this paper, we introduce a novel data augmentation technique that combines the advantages of style augmentation and random erasing by selectively replacing image subregions with style-transferred patches. Our approach first applies a random style transfer to training images, then randomly substitutes selected areas of these images with patches derived from the style-transferred versions. This method is able to seamlessly accommodate a wide range of existing style transfer algorithms and can be readily integrated into diverse data augmentation pipelines. By incorporating our strategy, the training process becomes more robust and less prone to overfitting. Comparative experiments demonstrate that, relative to previous style augmentation methods, our technique achieves superior performance and faster convergence.", "date": "2025-04-14", "categories": []}
{"id": "2509.15472", "title": "Efficient Multimodal Dataset Distillation via Generative Models", "abstract": "Dataset distillation aims to synthesize a small dataset from a large dataset, enabling the model trained on it to perform well on the original dataset. With the blooming of large language models and multimodal large language models, the importance of multimodal datasets, particularly image-text datasets, has grown significantly. However, existing multimodal dataset distillation methods are constrained by the Matching Training Trajectories algorithm, which significantly increases the computing resource requirement, and takes days to process the distillation. In this work, we introduce EDGE, a generative distillation method for efficient multimodal dataset distillation. Specifically, we identify two key challenges of distilling multimodal datasets with generative models: 1) The lack of correlation between generated images and captions. 2) The lack of diversity among generated samples. To address the aforementioned issues, we propose a novel generative model training workflow with a bi-directional contrastive loss and a diversity loss. Furthermore, we propose a caption synthesis strategy to further improve text-to-image retrieval performance by introducing more text information. Our method is evaluated on Flickr30K, COCO, and CC3M datasets, demonstrating superior performance and efficiency compared to existing approaches. Notably, our method achieves results 18x faster than the state-of-the-art method.", "date": "2025-09-18", "categories": []}
{"id": "2509.21748", "title": "SubZeroCore: A Submodular Approach with Zero Training for Coreset Selection", "abstract": "The goal of coreset selection is to identify representative subsets of datasets for efficient model training. Yet, existing approaches paradoxically require expensive training-based signals, e.g., gradients, decision boundary estimates or forgetting counts, computed over the entire dataset prior to pruning, which undermines their very purpose by requiring training on samples they aim to avoid. We introduce SubZeroCore, a novel, training-free coreset selection method that integrates submodular coverage and density into a single, unified objective. To achieve this, we introduce a sampling strategy based on a closed-form solution to optimally balance these objectives, guided by a single hyperparameter that explicitly controls the desired coverage for local density measures. Despite no training, extensive evaluations show that SubZeroCore matches training-based baselines and significantly outperforms them at high pruning rates, while dramatically reducing computational overhead. SubZeroCore also demonstrates superior robustness to label noise, highlighting its practical effectiveness and scalability for real-world scenarios.", "date": "2025-09-26", "categories": []}
{"id": "2509.22631", "title": "LABELING COPILOT: A Deep Research Agent for Automated Data Curation in Computer Vision", "abstract": "Curating high-quality, domain-specific datasets is a major bottleneck for deploying robust vision systems, requiring complex trade-offs between data quality, diversity, and cost when researching vast, unlabeled data lakes. We introduce Labeling Copilot, the first data curation deep research agent for computer vision. A central orchestrator agent, powered by a large multimodal language model, uses multi-step reasoning to execute specialized tools across three core capabilities: (1) Calibrated Discovery sources relevant, in-distribution data from large repositories; (2) Controllable Synthesis generates novel data for rare scenarios with robust filtering; and (3) Consensus Annotation produces accurate labels by orchestrating multiple foundation models via a novel consensus mechanism incorporating non-maximum suppression and voting. Our large-scale validation proves the effectiveness of Labeling Copilot's components. The Consensus Annotation module excels at object discovery: on the dense COCO dataset, it averages 14.2 candidate proposals per image-nearly double the 7.4 ground-truth objects-achieving a final annotation mAP of 37.1%. On the web-scale Open Images dataset, it navigated extreme class imbalance to discover 903 new bounding box categories, expanding its capability to over 1500 total. Concurrently, our Calibrated Discovery tool, tested at a 10-million sample scale, features an active learning strategy that is up to 40x more computationally efficient than alternatives with equivalent sample efficiency. These experiments validate that an agentic workflow with optimized, scalable tools provides a robust foundation for curating industrial-scale datasets.", "date": "2025-09-26", "categories": []}
{"id": "2505.16147", "title": "Losing is for Cherishing: Data Valuation Based on Machine Unlearning and Shapley Value", "abstract": "The proliferation of large models has intensified the need for efficient data valuation methods to quantify the contribution of individual data providers. Traditional approaches, such as game-theory-based Shapley value and influence-function-based techniques, face prohibitive computational costs or require access to full data and model training details, making them hardly achieve partial data valuation. To address this, we propose Unlearning Shapley, a novel framework that leverages machine unlearning to estimate data values efficiently. By unlearning target data from a pretrained model and measuring performance shifts on a reachable test set, our method computes Shapley values via Monte Carlo sampling, avoiding retraining and eliminating dependence on full data. Crucially, Unlearning Shapley supports both full and partial data valuation, making it scalable for large models (e.g., LLMs) and practical for data markets. Experiments on benchmark datasets and large-scale text corpora demonstrate that our approach matches the accuracy of state-of-the-art methods while reducing computational overhead by orders of magnitude. Further analysis confirms a strong correlation between estimated values and the true impact of data subsets, validating its reliability in real-world scenarios. This work bridges the gap between data valuation theory and practical deployment, offering a scalable, privacy-compliant solution for modern AI ecosystems.", "date": "2025-05-22", "categories": []}
{"id": "2509.18104", "title": "Data Valuation and Selection in a Federated Model Marketplace", "abstract": "In the era of Artificial Intelligence (AI), marketplaces have become essential platforms for facilitating the exchange of data products to foster data sharing. Model transactions provide economic solutions in data marketplaces that enhance data reusability and ensure the traceability of data ownership. To establish trustworthy data marketplaces, Federated Learning (FL) has emerged as a promising paradigm to enable collaborative learning across siloed datasets while safeguarding data privacy. However, effective data valuation and selection from heterogeneous sources in the FL setup remain key challenges. This paper introduces a comprehensive framework centered on a Wasserstein-based estimator tailored for FL. The estimator not only predicts model performance across unseen data combinations but also reveals the compatibility between data heterogeneity and FL aggregation algorithms. To ensure privacy, we propose a distributed method to approximate Wasserstein distance without requiring access to raw data. Furthermore, we demonstrate that model performance can be reliably extrapolated under the neural scaling law, enabling effective data selection without full-scale training. Extensive experiments across diverse scenarios, such as label skew, mislabeled, and unlabeled sources, show that our approach consistently identifies high-performing data combinations, paving the way for more reliable FL-based model marketplaces.", "date": "2025-09-09", "categories": []}
{"id": "2505.13300", "title": "DD-Ranking: Rethinking the Evaluation of Dataset Distillation", "abstract": "In recent years, dataset distillation has provided a reliable solution for data compression, where models trained on the resulting smaller synthetic datasets achieve performance comparable to those trained on the original datasets. To further improve the performance of synthetic datasets, various training pipelines and optimization objectives have been proposed, greatly advancing the field of dataset distillation. Recent decoupled dataset distillation methods introduce soft labels and stronger data augmentation during the post-evaluation phase and scale dataset distillation up to larger datasets (e.g., ImageNet-1K). However, this raises a question: Is accuracy still a reliable metric to fairly evaluate dataset distillation methods? Our empirical findings suggest that the performance improvements of these methods often stem from additional techniques rather than the inherent quality of the images themselves, with even randomly sampled images achieving superior results. Such misaligned evaluation settings severely hinder the development of DD. Therefore, we propose DD-Ranking, a unified evaluation framework, along with new general evaluation metrics to uncover the true performance improvements achieved by different methods. By refocusing on the actual information enhancement of distilled datasets, DD-Ranking provides a more comprehensive and fair evaluation standard for future research advancements.", "date": "2025-05-19", "categories": []}
{"id": "2509.16678", "title": "IPF-RDA: An Information-Preserving Framework for Robust Data Augmentation", "abstract": "Data augmentation is widely utilized as an effective technique to enhance the generalization performance of deep models. However, data augmentation may inevitably introduce distribution shifts and noises, which significantly constrain the potential and deteriorate the performance of deep networks. To this end, we propose a novel information-preserving framework, namely IPF-RDA, to enhance the robustness of data augmentations in this paper. IPF-RDA combines the proposal of (i) a new class-discriminative information estimation algorithm that identifies the points most vulnerable to data augmentation operations and corresponding importance scores; And (ii) a new information-preserving scheme that preserves the critical information in the augmented samples and ensures the diversity of augmented data adaptively. We divide data augmentation methods into three categories according to the operation types and integrate these approaches into our framework accordingly. After being integrated into our framework, the robustness of data augmentation methods can be enhanced and their full potential can be unleashed. Extensive experiments demonstrate that although being simple, IPF-RDA consistently improves the performance of numerous commonly used state-of-the-art data augmentation methods with popular deep models on a variety of datasets, including CIFAR-10, CIFAR-100, Tiny-ImageNet, CUHK03, Market1501, Oxford Flower, and MNIST, where its performance and scalability are stressed. The implementation is available at https://github.com/Jackbrocp/IPF-RDA.", "date": "2025-09-20", "categories": []}
{"id": "2509.14030", "title": "CrowdAgent: Multi-Agent Managed Multi-Source Annotation System", "abstract": "High-quality annotated data is a cornerstone of modern Natural Language Processing (NLP). While recent methods begin to leverage diverse annotation sources-including Large Language Models (LLMs), Small Language Models (SLMs), and human experts-they often focus narrowly on the labeling step itself. A critical gap remains in the holistic process control required to manage these sources dynamically, addressing complex scheduling and quality-cost trade-offs in a unified manner. Inspired by real-world crowdsourcing companies, we introduce CrowdAgent, a multi-agent system that provides end-to-end process control by integrating task assignment, data annotation, and quality/cost management. It implements a novel methodology that rationally assigns tasks, enabling LLMs, SLMs, and human experts to advance synergistically in a collaborative annotation workflow. We demonstrate the effectiveness of CrowdAgent through extensive experiments on six diverse multimodal classification tasks. The source code and video demo are available at https://github.com/QMMMS/CrowdAgent.", "date": "2025-09-17", "categories": []}
{"id": "2404.02353", "title": "Semantic Augmentation in Images using Language", "abstract": "Deep Learning models are incredibly data-hungry and require very large labeled datasets for supervised learning. As a consequence, these models often suffer from overfitting, limiting their ability to generalize to real-world examples. Recent advancements in diffusion models have enabled the generation of photorealistic images based on textual inputs. Leveraging the substantial datasets used to train these diffusion models, we propose a technique to utilize generated images to augment existing datasets. This paper explores various strategies for effective data augmentation to improve the out-of-domain generalization capabilities of deep learning models.", "date": "2024-04-02", "categories": []}
{"id": "2509.10367", "title": "A Discrepancy-Based Perspective on Dataset Condensation", "abstract": "Given a dataset of finitely many elements $\\mathcal{T} = \\{\\mathbf{x}_i\\}_{i = 1}^N$, the goal of dataset condensation (DC) is to construct a synthetic dataset $\\mathcal{S} = \\{\\tilde{\\mathbf{x}}_j\\}_{j = 1}^M$ which is significantly smaller ($M \\ll N$) such that a model trained from scratch on $\\mathcal{S}$ achieves comparable or even superior generalization performance to a model trained on $\\mathcal{T}$. Recent advances in DC reveal a close connection to the problem of approximating the data distribution represented by $\\mathcal{T}$ with a reduced set of points. In this work, we present a unified framework that encompasses existing DC methods and extend the task-specific notion of DC to a more general and formal definition using notions of discrepancy, which quantify the distance between probability distribution in different regimes. Our framework broadens the objective of DC beyond generalization, accommodating additional objectives such as robustness, privacy, and other desirable properties.", "date": "2025-09-12", "categories": []}
{"id": "2502.19279", "title": "CritiQ: Mining Data Quality Criteria from Human Preferences", "abstract": "Language model heavily depends on high-quality data for optimal performance. Existing approaches rely on manually designed heuristics, the perplexity of existing models, training classifiers, or careful prompt engineering, which require significant expert experience and human annotation effort while introduce biases. We introduce CritiQ, a novel data selection method that automatically mines criteria from human preferences for data quality with only ~30 human-annotated pairs and performs efficient data selection. The main component, CritiQ Flow, employs a manager agent to evolve quality criteria and worker agents to make pairwise judgments. We build a knowledge base that extracts quality criteria from previous work to boost CritiQ Flow. Compared to perplexity- and classifier- based methods, verbal criteria are more interpretable and possess reusable value. After deriving the criteria, we train the CritiQ Scorer to give quality scores and perform efficient data selection. We demonstrate the effectiveness of our method in the code, math, and logic domains, achieving high accuracy on human-annotated test sets. To validate the quality of the selected data, we continually train Llama 3.1 models and observe improved performance on downstream tasks compared to uniform sampling. Ablation studies validate the benefits of the knowledge base and the reflection process. We analyze how criteria evolve and the effectiveness of majority voting.", "date": "2025-02-26", "categories": []}
{"id": "2506.06656", "title": "Rescaled Influence Functions: Accurate Data Attribution in High Dimension", "abstract": "How does the training data affect a model's behavior? This is the question we seek to answer with data attribution. The leading practical approaches to data attribution are based on influence functions (IF). IFs utilize a first-order Taylor approximation to efficiently predict the effect of removing a set of samples from the training set without retraining the model, and are used in a wide variety of machine learning applications. However, especially in the high-dimensional regime (# params $\\geq \\Omega($# samples$)$), they are often imprecise and tend to underestimate the effect of sample removals, even for simple models such as logistic regression. We present rescaled influence functions (RIF), a new tool for data attribution which can be used as a drop-in replacement for influence functions, with little computational overhead but significant improvement in accuracy. We compare IF and RIF on a range of real-world datasets, showing that RIFs offer significantly better predictions in practice, and present a theoretical analysis explaining this improvement. Finally, we present a simple class of data poisoning attacks that would fool IF-based detections but would be detected by RIF.", "date": "2025-06-07", "categories": []}
{"id": "2509.07311", "title": "Does This Look Familiar to You? Knowledge Analysis via Model Internal Representations", "abstract": "Recent advances in large language models (LLMs) have been driven by pretraining, supervised fine tuning (SFT), and alignment tuning. Among these, SFT plays a crucial role in transforming a model 's general knowledge into structured responses tailored to specific tasks. However, there is no clearly established methodology for effective training data selection. Simply increasing the volume of data does not guarantee performance improvements, while preprocessing, sampling, and validation require substantial time and cost.\n  To address this issue, a variety of data selection methods have been proposed. Among them, knowledge based selection approaches identify suitable training data by analyzing the model 's responses. Nevertheless, these methods typically rely on prompt engineering, making them sensitive to variations and incurring additional costs for prompt design.\n  In this study, we propose Knowledge Analysis via Model Internal Representations (KAMIR), a novel approach that overcomes these limitations by analyzing data based on the model 's internal representations. KAMIR computes similarities between the hidden states of each layer (block) and the final hidden states for a given input to assess the data. Unlike prior methods that were largely limited to multiple choice tasks, KAMIR can be applied to a wide range of tasks such as machine reading comprehension and summarization. Moreover, it selects data useful for training based on the model 's familiarity with the input, even with a small dataset and a simple classifier architecture. Experiments across diverse task datasets demonstrate that training with less familiar data leads to better generalization performance.", "date": "2025-09-09", "categories": []}
{"id": "2501.18619", "title": "FAAGC: Feature Augmentation on Adaptive Geodesic Curve Based on the shape space theory", "abstract": "Deep learning models have been widely applied across various domains and industries. However, many fields still face challenges due to limited and insufficient data. This paper proposes a Feature Augmentation on Adaptive Geodesic Curve (FAAGC) method in the pre-shape space to increase data. In the pre-shape space, objects with identical shapes lie on a great circle. Thus, we project deep model representations into the pre-shape space and construct a geodesic curve, i.e., an arc of a great circle, for each class. Feature augmentation is then performed by sampling along these geodesic paths. Extensive experiments demonstrate that FAAGC improves classification accuracy under data-scarce conditions and generalizes well across various feature types.", "date": "2025-01-25", "categories": []}
{"id": "2509.06524", "title": "LAMDAS: LLM as an Implicit Classifier for Domain-specific Data Selection", "abstract": "Adapting large language models (LLMs) to specific domains often faces a critical bottleneck: the scarcity of high-quality, human-curated data. While large volumes of unchecked data are readily available, indiscriminately using them for fine-tuning risks introducing noise and degrading performance. Strategic data selection is thus crucial, requiring a method that is both accurate and efficient. Existing approaches, categorized as similarity-based and direct optimization methods, struggle to simultaneously achieve these goals. In this paper, we introduce LAMDAS (LLM As an iMplicit classifier for domain-specific DAta Selection), a novel approach that leverages the pre-trained LLM itself as an implicit classifier, thereby bypassing explicit feature engineering and computationally intensive optimization process. LAMDAS reframes data selection as a one-class classification problem, identifying candidate data that \"belongs\" to the target domain defined by a small reference dataset. Extensive experimental results demonstrate that LAMDAS not only exceeds the performance of full-data training using a fraction of the data but also outperforms nine state-of-the-art (SOTA) baselines under various scenarios. Furthermore, LAMDAS achieves the most compelling balance between performance gains and computational efficiency compared to all evaluated baselines.", "date": "2025-09-08", "categories": []}
{"id": "2312.03325", "title": "FAGC:Feature Augmentation on Geodesic Curve in the Pre-Shape Space", "abstract": "Due to the constraints on model performance imposed by the size of the training data, data augmentation has become an essential technique in deep learning. However, most existing data augmentation methods are affected by information loss and perform poorly in small-sample scenarios, which limits their application. To overcome the limitation, we propose a Feature Augmentation method on Geodesic Curve in the pre-shape space, called the FAGC. First, a pre-trained neural network model is employed to extract features from the input images. Then, the image features as a vector is projected into the pre-shape space by removing its position and scale information. In the pre-shape space, an optimal Geodesic curve is constructed to fit the feature vectors. Finally, new feature vectors are generated for model learning by interpolating along the constructed Geodesic curve. We conducted extensive experiments to demonstrate the effectiveness and versatility of the FAGC. The results demonstrate that applying the FAGC to deep learning or machine learning methods can significantly improve their performance in small-sample tasks.", "date": "2023-12-06", "categories": []}
{"id": "2509.02588", "title": "Sequential Hard Mining: a data-centric approach for Mitosis Detection", "abstract": "With a continuously growing availability of annotated datasets of mitotic figures in histology images, finding the best way to optimally use with this unprecedented amount of data to optimally train deep learning models has become a new challenge. Here, we build upon previously proposed approaches with a focus on efficient sampling of training data inspired by boosting techniques and present our candidate solutions for the two tracks of the MIDOG 2025 challenge.", "date": "2025-08-28", "categories": []}
{"id": "2504.03612", "title": "AIR: A Systematic Analysis of Annotations, Instructions, and Response Pairs in Preference Dataset", "abstract": "Preference learning is critical for aligning large language models (LLMs) with human values, yet its success hinges on high-quality datasets comprising three core components: Preference \\textbf{A}nnotations, \\textbf{I}nstructions, and \\textbf{R}esponse Pairs. Current approaches conflate these components, obscuring their individual impacts and hindering systematic optimization. In this work, we propose \\textbf{AIR}, a component-wise analysis framework that systematically isolates and optimizes each component while evaluating their synergistic effects. Through rigorous experimentation, AIR reveals actionable principles: annotation simplicity (point-wise generative scoring), instruction inference stability (variance-based filtering across LLMs), and response pair quality (moderate margins + high absolute scores). When combined, these principles yield +5.3 average gains over baseline method, even with only 14k high-quality pairs. Our work shifts preference dataset design from ad hoc scaling to component-aware optimization, offering a blueprint for efficient, reproducible alignment.", "date": "2025-04-04", "categories": []}
{"id": "2509.00083", "title": "Data Cartography for Detecting Memorization Hotspots and Guiding Data Interventions in Generative Models", "abstract": "Modern generative models risk overfitting and unintentionally memorizing rare training examples, which can be extracted by adversaries or inflate benchmark performance. We propose Generative Data Cartography (GenDataCarto), a data-centric framework that assigns each pretraining sample a difficulty score (early-epoch loss) and a memorization score (frequency of ``forget events''), then partitions examples into four quadrants to guide targeted pruning and up-/down-weighting. We prove that our memorization score lower-bounds classical influence under smoothness assumptions and that down-weighting high-memorization hotspots provably decreases the generalization gap via uniform stability bounds. Empirically, GenDataCarto reduces synthetic canary extraction success by over 40\\% at just 10\\% data pruning, while increasing validation perplexity by less than 0.5\\%. These results demonstrate that principled data interventions can dramatically mitigate leakage with minimal cost to generative performance.", "date": "2025-08-27", "categories": []}
{"id": "2509.00378", "title": "NoiseCutMix: A Novel Data Augmentation Approach by Mixing Estimated Noise in Diffusion Models", "abstract": "In this study, we propose a novel data augmentation method that introduces the concept of CutMix into the generation process of diffusion models, thereby exploiting both the ability of diffusion models to generate natural and high-resolution images and the characteristic of CutMix, which combines features from two classes to create diverse augmented data. Representative data augmentation methods for combining images from multiple classes include CutMix and MixUp. However, techniques like CutMix often result in unnatural boundaries between the two images due to contextual differences. Therefore, in this study, we propose a method, called NoiseCutMix, to achieve natural, high-resolution image generation featuring the fused characteristics of two classes by partially combining the estimated noise corresponding to two different classes in a diffusion model. In the classification experiments, we verified the effectiveness of the proposed method by comparing it with conventional data augmentation techniques that combine multiple classes, random image generation using Stable Diffusion, and combinations of these methods. Our codes are available at: https://github.com/shumpei-takezaki/NoiseCutMix", "date": "2025-08-30", "categories": []}
{"id": "2508.19659", "title": "SCAR: A Characterization Scheme for Multi-Modal Dataset", "abstract": "Foundation models exhibit remarkable generalization across diverse tasks, largely driven by the characteristics of their training data. Recent data-centric methods like pruning and compression aim to optimize training but offer limited theoretical insight into how data properties affect generalization, especially the data characteristics in sample scaling. Traditional perspectives further constrain progress by focusing predominantly on data quantity and training efficiency, often overlooking structural aspects of data quality. In this study, we introduce SCAR, a principled scheme for characterizing the intrinsic structural properties of datasets across four key measures: Scale, Coverage, Authenticity, and Richness. Unlike prior data-centric measures, SCAR captures stable characteristics that remain invariant under dataset scaling, providing a robust and general foundation for data understanding. Leveraging these structural properties, we introduce Foundation Data-a minimal subset that preserves the generalization behavior of the full dataset without requiring model-specific retraining. We model single-modality tasks as step functions and estimate the distribution of the foundation data size to capture step-wise generalization bias across modalities in the target multi-modal dataset. Finally, we develop a SCAR-guided data completion strategy based on this generalization bias, which enables efficient, modality-aware expansion of modality-specific characteristics in multimodal datasets. Experiments across diverse multi-modal datasets and model architectures validate the effectiveness of SCAR in predicting data utility and guiding data acquisition. Code is available at https://github.com/McAloma/SCAR.", "date": "2025-08-27", "categories": []}
{"id": "2508.17677", "title": "TiKMiX: Take Data Influence into Dynamic Mixture for Language Model Pre-training", "abstract": "The data mixture used in the pre-training of a language model is a cornerstone of its final performance. However, a static mixing strategy is suboptimal, as the model's learning preferences for various data domains shift dynamically throughout training. Crucially, observing these evolving preferences in a computationally efficient manner remains a significant challenge. To address this, we propose TiKMiX, a method that dynamically adjusts the data mixture according to the model's evolving preferences. TiKMiX introduces Group Influence, an efficient metric for evaluating the impact of data domains on the model. This metric enables the formulation of the data mixing problem as a search for an optimal, influence-maximizing distribution. We solve this via two approaches: TiKMiX-D for direct optimization, and TiKMiX-M, which uses a regression model to predict a superior mixture. We trained models with different numbers of parameters, on up to 1 trillion tokens. TiKMiX-D exceeds the performance of state-of-the-art methods like REGMIX while using just 20% of the computational resources. TiKMiX-M leads to an average performance gain of 2% across 9 downstream benchmarks. Our experiments reveal that a model's data preferences evolve with training progress and scale, and we demonstrate that dynamically adjusting the data mixture based on Group Influence, a direct measure of these preferences, significantly improves performance by mitigating the underdigestion of data seen with static ratios.", "date": "2025-08-25", "categories": []}
{"id": "2504.13227", "title": "DIDS: Domain Impact-aware Data Sampling for Large Language Model Training", "abstract": "Large language models (LLMs) are commonly trained on multi-domain datasets, where domain sampling strategies significantly impact model performance due to varying domain importance across downstream tasks. Existing approaches for optimizing domain-level sampling strategies struggle with maintaining intra-domain consistency and accurately measuring domain impact. In this paper, we present Domain Impact-aware Data Sampling (DIDS). To ensure intra-domain consistency, a gradient clustering algorithm is proposed to group training data based on their learning effects, where a proxy language model and dimensionality reduction are employed to reduce computational overhead. To accurately measure domain impact, we develop a Fisher Information Matrix (FIM) guided metric that quantifies how domain-specific parameter updates affect the model's output distributions on downstream tasks, with theoretical guarantees. Furthermore, to determine optimal sampling ratios, DIDS combines both the FIM-guided domain impact assessment and loss learning trajectories that indicate domain-specific potential, while accounting for diminishing marginal returns. Extensive experiments demonstrate that DIDS achieves 3.4% higher average performance while maintaining comparable training efficiency. The code is available at https://github.com/shiweijiezero/DIDS.", "date": "2025-04-17", "categories": []}
{"id": "2508.16255", "title": "Chunked Data Shapley: A Scalable Dataset Quality Assessment for Machine Learning", "abstract": "As the volume and diversity of available datasets continue to increase, assessing data quality has become crucial for reliable and efficient Machine Learning analytics. A modern, game-theoretic approach for evaluating data quality is the notion of Data Shapley which quantifies the value of individual data points within a dataset. State-of-the-art methods to scale the NP-hard Shapley computation also face severe challenges when applied to large-scale datasets, limiting their practical use. In this work, we present a Data Shapley approach to identify a dataset's high-quality data tuples, Chunked Data Shapley (C-DaSh). C-DaSh scalably divides the dataset into manageable chunks and estimates the contribution of each chunk using optimized subset selection and single-iteration stochastic gradient descent. This approach drastically reduces computation time while preserving high quality results. We empirically benchmark our method on diverse real-world classification and regression tasks, demonstrating that C-DaSh outperforms existing Shapley approximations in both computational efficiency (achieving speedups between 80x - 2300x) and accuracy in detecting low-quality data regions. Our method enables practical measurement of dataset quality on large tabular datasets, supporting both classification and regression pipelines.", "date": "2025-08-22", "categories": []}
{"id": "2508.15096", "title": "Nemotron-CC-Math: A 133 Billion-Token-Scale High Quality Math Pretraining Dataset", "abstract": "Pretraining large language models (LLMs) on high-quality, structured data such as mathematics and code substantially enhances reasoning capabilities. However, existing math-focused datasets built from Common Crawl suffer from degraded quality due to brittle extraction heuristics, lossy HTML-to-text conversion, and the failure to reliably preserve mathematical structure. In this work, we introduce Nemotron-CC-Math, a large-scale, high-quality mathematical corpus constructed from Common Crawl using a novel, domain-agnostic pipeline specifically designed for robust scientific text extraction.\n  Unlike previous efforts, our pipeline recovers math across various formats (e.g., MathJax, KaTeX, MathML) by leveraging layout-aware rendering with lynx and a targeted LLM-based cleaning stage. This approach preserves the structural integrity of equations and code blocks while removing boilerplate, standardizing notation into LaTeX representation, and correcting inconsistencies.\n  We collected a large, high-quality math corpus, namely Nemotron-CC-Math-3+ (133B tokens) and Nemotron-CC-Math-4+ (52B tokens). Notably, Nemotron-CC-Math-4+ not only surpasses all prior open math datasets-including MegaMath, FineMath, and OpenWebMath-but also contains 5.5 times more tokens than FineMath-4+, which was previously the highest-quality math pretraining dataset. When used to pretrain a Nemotron-T 8B model, our corpus yields +4.8 to +12.6 gains on MATH and +4.6 to +14.3 gains on MBPP+ over strong baselines, while also improving general-domain performance on MMLU and MMLU-Stem.\n  We present the first pipeline to reliably extract scientific content--including math--from noisy web-scale data, yielding measurable gains in math, code, and general reasoning, and setting a new state of the art among open math pretraining corpora. To support open-source efforts, we release our code and datasets.", "date": "2025-08-20", "categories": []}
{"id": "2508.14648", "title": "Understanding Data Influence with Differential Approximation", "abstract": "Data plays a pivotal role in the groundbreaking advancements in artificial intelligence. The quantitative analysis of data significantly contributes to model training, enhancing both the efficiency and quality of data utilization. However, existing data analysis tools often lag in accuracy. For instance, many of these tools even assume that the loss function of neural networks is convex. These limitations make it challenging to implement current methods effectively. In this paper, we introduce a new formulation to approximate a sample's influence by accumulating the differences in influence between consecutive learning steps, which we term Diff-In. Specifically, we formulate the sample-wise influence as the cumulative sum of its changes/differences across successive training iterations. By employing second-order approximations, we approximate these difference terms with high accuracy while eliminating the need for model convexity required by existing methods. Despite being a second-order method, Diff-In maintains computational complexity comparable to that of first-order methods and remains scalable. This efficiency is achieved by computing the product of the Hessian and gradient, which can be efficiently approximated using finite differences of first-order gradients. We assess the approximation accuracy of Diff-In both theoretically and empirically. Our theoretical analysis demonstrates that Diff-In achieves significantly lower approximation error compared to existing influence estimators. Extensive experiments further confirm its superior performance across multiple benchmark datasets in three data-centric tasks: data cleaning, data deletion, and coreset selection. Notably, our experiments on data pruning for large-scale vision-language pre-training show that Diff-In can scale to millions of data points and outperforms strong baselines.", "date": "2025-08-20", "categories": []}
{"id": "2508.14801", "title": "A Guide for Manual Annotation of Scientific Imagery: How to Prepare for Large Projects", "abstract": "Despite the high demand for manually annotated image data, managing complex and costly annotation projects remains under-discussed. This is partly due to the fact that leading such projects requires dealing with a set of diverse and interconnected challenges which often fall outside the expertise of specific domain experts, leaving practical guidelines scarce. These challenges range widely from data collection to resource allocation and recruitment, from mitigation of biases to effective training of the annotators. This paper provides a domain-agnostic preparation guide for annotation projects, with a focus on scientific imagery. Drawing from the authors' extensive experience in managing a large manual annotation project, it addresses fundamental concepts including success measures, annotation subjects, project goals, data availability, and essential team roles. Additionally, it discusses various human biases and recommends tools and technologies to improve annotation quality and efficiency. The goal is to encourage further research and frameworks for creating a comprehensive knowledge base to reduce the costs of manual annotation projects across various fields.", "date": "2025-08-20", "categories": []}
{"id": "2412.06293", "title": "Mastering Collaborative Multi-modal Data Selection: A Focus on Informativeness, Uniqueness, and Representativeness", "abstract": "Instruction tuning fine-tunes pre-trained Multi-modal Large Language Models (MLLMs) to handle real-world tasks. However, the rapid expansion of visual instruction datasets introduces data redundancy, leading to excessive computational costs. We propose a collaborative framework, DataTailor, which leverages three key principles--informativeness, uniqueness, and representativeness--for effective data selection. We argue that a valuable sample should be informative of the task, non-redundant, and represent the sample distribution (i.e., not an outlier). We further propose practical ways to score against each principle, which automatically adapts to a given dataset without tedious hyperparameter tuning. Comprehensive experiments on various benchmarks demonstrate that DataTailor achieves 101.3% of the performance of full-data fine-tuning with only 15% of the data, significantly reducing computational costs while maintaining superior results. This exemplifies the \"Less is More\" philosophy in MLLM development. The code and data is available in this \\href{https://github.com/Yuqifan1117/DataTailor}{URL}.", "date": "2024-12-09", "categories": []}
{"id": "2506.01701", "title": "Data Pruning by Information Maximization", "abstract": "In this paper, we present InfoMax, a novel data pruning method, also known as coreset selection, designed to maximize the information content of selected samples while minimizing redundancy. By doing so, InfoMax enhances the overall informativeness of the coreset. The information of individual samples is measured by importance scores, which capture their influence or difficulty in model learning. To quantify redundancy, we use pairwise sample similarities, based on the premise that similar samples contribute similarly to the learning process. We formalize the coreset selection problem as a discrete quadratic programming (DQP) task, with the objective of maximizing the total information content, represented as the sum of individual sample contributions minus the redundancies introduced by similar samples within the coreset. To ensure practical scalability, we introduce an efficient gradient-based solver, complemented by sparsification techniques applied to the similarity matrix and dataset partitioning strategies. This enables InfoMax to seamlessly scale to datasets with millions of samples. Extensive experiments demonstrate the superior performance of InfoMax in various data pruning tasks, including image classification, vision-language pre-training, and instruction tuning for large language models. Code is available at https://github.com/hrtan/InfoMax.", "date": "2025-06-02", "categories": []}
{"id": "2507.10904", "title": "Class-Proportional Coreset Selection for Difficulty-Separable Data", "abstract": "High-quality training data is essential for building reliable and efficient machine learning systems. One-shot coreset selection addresses this by pruning the dataset while maintaining or even improving model performance, often relying on training-dynamics-based data difficulty scores. However, most existing methods implicitly assume class-wise homogeneity in data difficulty, overlooking variation in data difficulty across different classes. In this work, we challenge this assumption by showing that, in domains such as network intrusion detection and medical imaging, data difficulty often clusters by class. We formalize this as class-difficulty separability and introduce the Class Difficulty Separability Coefficient (CDSC) as a quantitative measure. We demonstrate that high CDSC values correlate with performance degradation in class-agnostic coreset methods, which tend to overrepresent easy majority classes while neglecting rare but informative ones. To address this, we introduce class-proportional variants of multiple sampling strategies. Evaluated on five diverse datasets spanning security and medical domains, our methods consistently achieve state-of-the-art performance. For instance, on CTU-13, at an extreme 99% pruning rate, a class-proportional variant of Coverage-centric Coreset Selection (CCS-CP) shows remarkable stability, with accuracy dropping only 2.58%, precision 0.49%, and recall 0.19%. In contrast, the class-agnostic CCS baseline, the next best method, suffers sharper declines of 7.59% in accuracy, 4.57% in precision, and 4.11% in recall. We further show that aggressive pruning enhances generalization in noisy, imbalanced, and large-scale datasets. Our results underscore that explicitly modeling class-difficulty separability leads to more effective, robust, and generalizable data pruning, particularly in high-stakes scenarios.", "date": "2025-07-15", "categories": []}
{"id": "2508.09283", "title": "Distilling Reinforcement Learning into Single-Batch Datasets", "abstract": "Dataset distillation compresses a large dataset into a small synthetic dataset such that learning on the synthetic dataset approximates learning on the original. Training on the distilled dataset can be performed in as little as one step of gradient descent. We demonstrate that distillation is generalizable to different tasks by distilling reinforcement learning environments into one-batch supervised learning datasets. This demonstrates not only distillation's ability to compress a reinforcement learning task but also its ability to transform one learning modality (reinforcement learning) into another (supervised learning). We present a novel extension of proximal policy optimization for meta-learning and use it in distillation of a multi-dimensional extension of the classic cart-pole problem, all MuJoCo environments, and several Atari games. We demonstrate distillation's ability to compress complex RL environments into one-step supervised learning, explore RL distillation's generalizability across learner architectures, and demonstrate distilling an environment into the smallest-possible synthetic dataset.", "date": "2025-08-12", "categories": []}
{"id": "2508.09550", "title": "Exploring the Equivalence of Closed-Set Generative and Real Data Augmentation in Image Classification", "abstract": "In this paper, we address a key scientific problem in machine learning: Given a training set for an image classification task, can we train a generative model on this dataset to enhance the classification performance? (i.e., closed-set generative data augmentation). We start by exploring the distinctions and similarities between real images and closed-set synthetic images generated by advanced generative models. Through extensive experiments, we offer systematic insights into the effective use of closed-set synthetic data for augmentation. Notably, we empirically determine the equivalent scale of synthetic images needed for augmentation. In addition, we also show quantitative equivalence between the real data augmentation and open-set generative augmentation (generative models trained using data beyond the given training set). While it aligns with the common intuition that real images are generally preferred, our empirical formulation also offers a guideline to quantify the increased scale of synthetic data augmentation required to achieve comparable image classification performance. Our results on natural and medical image datasets further illustrate how this effect varies with the baseline training set size and the amount of synthetic data incorporated.", "date": "2025-08-13", "categories": []}
{"id": "2403.01306", "title": "ICC: Quantifying Image Caption Concreteness for Multimodal Dataset Curation", "abstract": "Web-scale training on paired text-image data is becoming increasingly central to multimodal learning, but is challenged by the highly noisy nature of datasets in the wild. Standard data filtering approaches succeed in removing mismatched text-image pairs, but permit semantically related but highly abstract or subjective text. These approaches lack the fine-grained ability to isolate the most concrete samples that provide the strongest signal for learning in a noisy dataset. In this work, we propose a new metric, image caption concreteness, that evaluates caption text without an image reference to measure its concreteness and relevancy for use in multimodal learning. Our approach leverages strong foundation models for measuring visual-semantic information loss in multimodal representations. We demonstrate that this strongly correlates with human evaluation of concreteness in both single-word and sentence-level texts. Moreover, we show that curation using ICC complements existing approaches: It succeeds in selecting the highest quality samples from multimodal web-scale datasets to allow for efficient training in resource-constrained settings.", "date": "2024-03-02", "categories": []}
{"id": "2508.07713", "title": "Detecting Mislabeled and Corrupted Data via Pointwise Mutual Information", "abstract": "Deep neural networks can memorize corrupted labels, making data quality critical for model performance, yet real-world datasets are frequently compromised by both label noise and input noise. This paper proposes a mutual information-based framework for data selection under hybrid noise scenarios that quantifies statistical dependencies between inputs and labels. We compute each sample's pointwise contribution to the overall mutual information and find that lower contributions indicate noisy or mislabeled instances. Empirical validation on MNIST with different synthetic noise settings demonstrates that the method effectively filters low-quality samples. Under label corruption, training on high-MI samples improves classification accuracy by up to 15\\% compared to random sampling. Furthermore, the method exhibits robustness to benign input modifications, preserving semantically valid data while filtering truly corrupted samples.", "date": "2025-08-11", "categories": []}
{"id": "2507.00038", "title": "Quality over Quantity: An Effective Large-Scale Data Reduction Strategy Based on Pointwise V-Information", "abstract": "In order to increase the effectiveness of model training, data reduction is essential to data-centric Artificial Intelligence (AI). It achieves this by locating the most instructive examples in massive datasets. To increase data quality and training efficiency, the main difficulty is choosing the best examples rather than the complete datasets. In this paper, we propose an effective data reduction strategy based on Pointwise V-Information (PVI). To enable a static method, we first use PVI to quantify instance difficulty and remove instances with low difficulty. Experiments show that classifier performance is maintained with only a 0.0001% to 0.76% decline in accuracy when 10%-30% of the data is removed. Second, we train the classifiers using a progressive learning strategy on examples sorted by increasing PVI, accelerating convergence and achieving a 0.8% accuracy gain over conventional training. Our findings imply that training a classifier on the chosen optimal subset may improve model performance and increase training efficiency when combined with an efficient data reduction strategy. Furthermore, we have adapted the PVI framework, which was previously limited to English datasets, to a variety of Chinese Natural Language Processing (NLP) tasks and base models, yielding insightful results for faster training and cross-lingual data reduction.", "date": "2025-06-19", "categories": []}
{"id": "2508.06227", "title": "Depth Jitter: Seeing through the Depth", "abstract": "Depth information is essential in computer vision, particularly in underwater imaging, robotics, and autonomous navigation. However, conventional augmentation techniques overlook depth aware transformations, limiting model robustness in real world depth variations. In this paper, we introduce Depth-Jitter, a novel depth-based augmentation technique that simulates natural depth variations to improve generalization. Our approach applies adaptive depth offsetting, guided by depth variance thresholds, to generate synthetic depth perturbations while preserving structural integrity. We evaluate Depth-Jitter on two benchmark datasets, FathomNet and UTDAC2020 demonstrating its impact on model stability under diverse depth conditions. Extensive experiments compare Depth-Jitter against traditional augmentation strategies such as ColorJitter, analyzing performance across varying learning rates, encoders, and loss functions. While Depth-Jitter does not always outperform conventional methods in absolute performance, it consistently enhances model stability and generalization in depth-sensitive environments. These findings highlight the potential of depth-aware augmentation for real-world applications and provide a foundation for further research into depth-based learning strategies. The proposed technique is publicly available to support advancements in depth-aware augmentation. The code is publicly available on \\href{https://github.com/mim-team/Depth-Jitter}{github}.", "date": "2025-08-08", "categories": []}
{"id": "2508.05089", "title": "Integrated Influence: Data Attribution with Baseline", "abstract": "As an effective approach to quantify how training samples influence test sample, data attribution is crucial for understanding data and model and further enhance the transparency of machine learning models. We find that prevailing data attribution methods based on leave-one-out (LOO) strategy suffer from the local-based explanation, as these LOO-based methods only perturb a single training sample, and overlook the collective influence in the training set. On the other hand, the lack of baseline in many data attribution methods reduces the flexibility of the explanation, e.g., failing to provide counterfactual explanations. In this paper, we propose Integrated Influence, a novel data attribution method that incorporates a baseline approach. Our method defines a baseline dataset, follows a data degeneration process to transition the current dataset to the baseline, and accumulates the influence of each sample throughout this process. We provide a solid theoretical framework for our method, and further demonstrate that popular methods, such as influence functions, can be viewed as special cases of our approach. Experimental results show that Integrated Influence generates more reliable data attributions compared to existing methods in both data attribution task and mislablled example identification task.", "date": "2025-08-07", "categories": []}
{"id": "2504.14194", "title": "Meta-rater: A Multi-dimensional Data Selection Method for Pre-training Language Models", "abstract": "The composition of pre-training datasets for large language models (LLMs) remains largely undisclosed, hindering transparency and efforts to optimize data quality, a critical driver of model performance. Current data selection methods, such as natural language quality assessments, diversity-based filters, and classifier-based approaches, are limited by single-dimensional evaluation or redundancy-focused strategies. To address these gaps, we propose four dimensions to evaluate data quality: professionalism, readability, reasoning, and cleanliness. We further introduce Meta-rater,a multi-dimensional data selection method that integrates these dimensions with existing quality metrics through learned optimal weightings. Meta-rater employs proxy models to train a regression model that predicts validation loss, enabling the identification of optimal combinations of quality scores. Experiments demonstrate that Meta-rater doubles convergence speed for 1.3B parameter models and improves downstream task performance by 3.23, with advantages that scale to models as large as 7.2B parameters. Our work establishes that holistic, multi-dimensional quality integration significantly outperforms conventional single-dimension approaches, offering a scalable paradigm for enhancing pre-training efficiency and model capability. To advance future research, we release scripts, data, and models at https://github.com/opendatalab/Meta-rater.", "date": "2025-04-19", "categories": []}
{"id": "2410.13453", "title": "Adaptive Augmentation Policy Optimization with LLM Feedback", "abstract": "Data augmentation is a critical component of deep learning pipelines, enhancing model generalization by increasing dataset diversity. Traditional augmentation strategies rely on manually designed transformations, stochastic sampling, or automated search-based approaches. Although automated methods improve performance, they often require extensive computational resources and are specifically designed for certain datasets. In this work, we propose a Large Language Model (LLM)-guided augmentation optimization strategy that refines augmentation policies based on model performance feedback. We propose two approaches: (1) LLM-Guided Augmentation Policy Optimization, where augmentation policies selected by LLM are refined iteratively across training cycles, and (2) Adaptive LLM-Guided Augmentation Policy Optimization, which adjusts policies at each iteration based on performance metrics. This in-training approach eliminates the need for full model retraining before getting LLM feedback, reducing computational costs while increasing performance. Our methodology employs an LLM to dynamically select augmentation transformations based on dataset characteristics, model architecture, and prior training performance. Leveraging LLMs' contextual knowledge, especially in domain-specific tasks like medical imaging, our method selects augmentations tailored to dataset characteristics and model performance. Experiments across domain-specific image classification datasets show consistent accuracy improvements over traditional methods.", "date": "2024-10-17", "categories": []}
{"id": "2507.21455", "title": "Boost Self-Supervised Dataset Distillation via Parameterization, Predefined Augmentation, and Approximation", "abstract": "Although larger datasets are crucial for training large deep models, the rapid growth of dataset size has brought a significant challenge in terms of considerable training costs, which even results in prohibitive computational expenses. Dataset Distillation becomes a popular technique recently to reduce the dataset size via learning a highly compact set of representative exemplars, where the model trained with these exemplars ideally should have comparable performance with respect to the one trained with the full dataset. While most of existing works upon dataset distillation focus on supervised datasets, we instead aim to distill images and their self-supervisedly trained representations into a distilled set. This procedure, named as Self-Supervised Dataset Distillation, effectively extracts rich information from real datasets, yielding the distilled sets with enhanced cross-architecture generalizability. Particularly, in order to preserve the key characteristics of original dataset more faithfully and compactly, several novel techniques are proposed: 1) we introduce an innovative parameterization upon images and representations via distinct low-dimensional bases, where the base selection for parameterization is experimentally shown to play a crucial role; 2) we tackle the instability induced by the randomness of data augmentation -- a key component in self-supervised learning but being underestimated in the prior work of self-supervised dataset distillation -- by utilizing predetermined augmentations; 3) we further leverage a lightweight network to model the connections among the representations of augmented views from the same image, leading to more compact pairs of distillation. Extensive experiments conducted on various datasets validate the superiority of our approach in terms of distillation efficiency, cross-architecture generalization, and transfer learning performance.", "date": "2025-07-29", "categories": []}
{"id": "2508.01264", "title": "Enhancing Diffusion-based Dataset Distillation via Adversary-Guided Curriculum Sampling", "abstract": "Dataset distillation aims to encapsulate the rich information contained in dataset into a compact distilled dataset but it faces performance degradation as the image-per-class (IPC) setting or image resolution grows larger. Recent advancements demonstrate that integrating diffusion generative models can effectively facilitate the compression of large-scale datasets while maintaining efficiency due to their superiority in matching data distribution and summarizing representative patterns. However, images sampled from diffusion models are always blamed for lack of diversity which may lead to information redundancy when multiple independent sampled images are aggregated as a distilled dataset. To address this issue, we propose Adversary-guided Curriculum Sampling (ACS), which partitions the distilled dataset into multiple curricula. For generating each curriculum, ACS guides diffusion sampling process by an adversarial loss to challenge a discriminator trained on sampled images, thus mitigating information overlap between curricula and fostering a more diverse distilled dataset. Additionally, as the discriminator evolves with the progression of curricula, ACS generates images from simpler to more complex, ensuring efficient and systematic coverage of target data informational spectrum. Extensive experiments demonstrate the effectiveness of ACS, which achieves substantial improvements of 4.1\\% on Imagewoof and 2.1\\% on ImageNet-1k over the state-of-the-art.", "date": "2025-08-02", "categories": []}
{"id": "2508.01749", "title": "Improving Noise Efficiency in Privacy-preserving Dataset Distillation", "abstract": "Modern machine learning models heavily rely on large datasets that often include sensitive and private information, raising serious privacy concerns. Differentially private (DP) data generation offers a solution by creating synthetic datasets that limit the leakage of private information within a predefined privacy budget; however, it requires a substantial amount of data to achieve performance comparable to models trained on the original data. To mitigate the significant expense incurred with synthetic data generation, Dataset Distillation (DD) stands out for its remarkable training and storage efficiency. This efficiency is particularly advantageous when integrated with DP mechanisms, curating compact yet informative synthetic datasets without compromising privacy. However, current state-of-the-art private DD methods suffer from a synchronized sampling-optimization process and the dependency on noisy training signals from randomly initialized networks. This results in the inefficient utilization of private information due to the addition of excessive noise. To address these issues, we introduce a novel framework that decouples sampling from optimization for better convergence and improves signal quality by mitigating the impact of DP noise through matching in an informative subspace. On CIFAR-10, our method achieves a \\textbf{10.0\\%} improvement with 50 images per class and \\textbf{8.3\\%} increase with just \\textbf{one-fifth} the distilled set size of previous state-of-the-art methods, demonstrating significant potential to advance privacy-preserving DD.", "date": "2025-08-03", "categories": []}
{"id": "2507.22186", "title": "SourceSplice: Source Selection for Machine Learning Tasks", "abstract": "Data quality plays a pivotal role in the predictive performance of machine learning (ML) tasks - a challenge amplified by the deluge of data sources available in modern organizations. Prior work in data discovery largely focus on metadata matching, semantic similarity or identifying tables that should be joined to answer a particular query, but do not consider source quality for high performance of the downstream ML task. This paper addresses the problem of determining the best subset of data sources that must be combined to construct the underlying training dataset for a given ML task. We propose SourceGrasp and SourceSplice, frameworks designed to efficiently select a suitable subset of sources that maximizes the utility of the downstream ML model. Both the algorithms rely on the core idea that sources (or their combinations) contribute differently to the task utility, and must be judiciously chosen. While SourceGrasp utilizes a metaheuristic based on a greediness criterion and randomization, the SourceSplice framework presents a source selection mechanism inspired from gene splicing - a core concept used in protein synthesis. We empirically evaluate our algorithms on three real-world datasets and synthetic datasets and show that, with significantly fewer subset explorations, SourceSplice effectively identifies subsets of data sources leading to high task utility. We also conduct studies reporting the sensitivity of SourceSplice to the decision choices under several settings.", "date": "2025-07-29", "categories": []}
{"id": "2506.23349", "title": "A case for data valuation transparency via DValCards", "abstract": "Following the rise in popularity of data-centric machine learning (ML), various data valuation methods have been proposed to quantify the contribution of each datapoint to desired ML model performance metrics (e.g., accuracy). Beyond the technical applications of data valuation methods (e.g., data cleaning, data acquisition, etc.), it has been suggested that within the context of data markets, data buyers might utilize such methods to fairly compensate data owners. Here we demonstrate that data valuation metrics are inherently biased and unstable under simple algorithmic design choices, resulting in both technical and ethical implications. By analyzing 9 tabular classification datasets and 6 data valuation methods, we illustrate how (1) common and inexpensive data pre-processing techniques can drastically alter estimated data values; (2) subsampling via data valuation metrics may increase class imbalance; and (3) data valuation metrics may undervalue underrepresented group data. Consequently, we argue in favor of increased transparency associated with data valuation in-the-wild and introduce the novel Data Valuation Cards (DValCards) framework towards this aim. The proliferation of DValCards will reduce misuse of data valuation metrics, including in data pricing, and build trust in responsible ML systems.", "date": "2025-06-29", "categories": []}
{"id": "2507.22360", "title": "GVD: Guiding Video Diffusion Model for Scalable Video Distillation", "abstract": "To address the larger computation and storage requirements associated with large video datasets, video dataset distillation aims to capture spatial and temporal information in a significantly smaller dataset, such that training on the distilled data has comparable performance to training on all of the data. We propose GVD: Guiding Video Diffusion, the first diffusion-based video distillation method. GVD jointly distills spatial and temporal features, ensuring high-fidelity video generation across diverse actions while capturing essential motion information. Our method's diverse yet representative distillations significantly outperform previous state-of-the-art approaches on the MiniUCF and HMDB51 datasets across 5, 10, and 20 Instances Per Class (IPC). Specifically, our method achieves 78.29 percent of the original dataset's performance using only 1.98 percent of the total number of frames in MiniUCF. Additionally, it reaches 73.83 percent of the performance with just 3.30 percent of the frames in HMDB51. Experimental results across benchmark video datasets demonstrate that GVD not only achieves state-of-the-art performance but can also generate higher resolution videos and higher IPC without significantly increasing computational cost.", "date": "2025-07-30", "categories": []}
{"id": "2507.17221", "title": "Dataset Distillation as Data Compression: A Rate-Utility Perspective", "abstract": "Driven by the ``scale-is-everything'' paradigm, modern machine learning increasingly demands ever-larger datasets and models, yielding prohibitive computational and storage requirements. Dataset distillation mitigates this by compressing an original dataset into a small set of synthetic samples, while preserving its full utility. Yet, existing methods either maximize performance under fixed storage budgets or pursue suitable synthetic data representations for redundancy removal, without jointly optimizing both objectives. In this work, we propose a joint rate-utility optimization method for dataset distillation. We parameterize synthetic samples as optimizable latent codes decoded by extremely lightweight networks. We estimate the Shannon entropy of quantized latents as the rate measure and plug any existing distillation loss as the utility measure, trading them off via a Lagrange multiplier. To enable fair, cross-method comparisons, we introduce bits per class (bpc), a precise storage metric that accounts for sample, label, and decoder parameter costs. On CIFAR-10, CIFAR-100, and ImageNet-128, our method achieves up to $170\\times$ greater compression than standard distillation at comparable accuracy. Across diverse bpc budgets, distillation losses, and backbone architectures, our approach consistently establishes better rate-utility trade-offs.", "date": "2025-07-23", "categories": []}
{"id": "2507.16729", "title": "Improving Model Classification by Optimizing the Training Dataset", "abstract": "In the era of data-centric AI, the ability to curate high-quality training data is as crucial as model design. Coresets offer a principled approach to data reduction, enabling efficient learning on large datasets through importance sampling. However, conventional sensitivity-based coreset construction often falls short in optimizing for classification performance metrics, e.g., $F1$ score, focusing instead on loss approximation. In this work, we present a systematic framework for tuning the coreset generation process to enhance downstream classification quality. Our method introduces new tunable parameters--including deterministic sampling, class-wise allocation, and refinement via active sampling, beyond traditional sensitivity scores. Through extensive experiments on diverse datasets and classifiers, we demonstrate that tuned coresets can significantly outperform both vanilla coresets and full dataset training on key classification metrics, offering an effective path towards better and more efficient model training.", "date": "2025-07-22", "categories": []}
{"id": "2507.14344", "title": "Influence Functions for Preference Dataset Pruning", "abstract": "Language models are commonly fine-tuned via reinforcement learning to alter their behavior or elicit new capabilities. Datasets used for these purposes, and particularly human preference datasets, are often noisy. The relatively small size post-training datasets, combined with parameter-efficient fine-tuning methods, enable the use of influence functions approximations to detect and prune training examples that are harmful to performance on a validation set. In this work, we adapt the TL;DR dataset for reward model training to demonstrate how conjugate-gradient approximated influence functions can be used to filter datasets. In our experiments, influence function filtering yields a small retraining accuracy uplift of 1.5% after removing 10% of training examples. We also show that gradient similarity outperforms influence functions for detecting helpful training examples. This suggests that local curvature is important for detecting harmful training examples, but less so for identifying helpful examples.", "date": "2025-07-18", "categories": []}
{"id": "2507.15061", "title": "WebShaper: Agentically Data Synthesizing via Information-Seeking Formalization", "abstract": "The advent of Large Language Model (LLM)-powered agents has revolutionized artificial intelligence by enabling solutions to complex, open-ended tasks through web-based information-seeking (IS) capabilities. The scarcity of high-quality training data has limited the development of IS agents. Existing approaches typically adopt an information-driven paradigm that first collects web data and then generates questions based on the retrieval. However, this may lead to inconsistency between information structure and reasoning structure, question and answer. To mitigate, we propose a formalization-driven IS data synthesis framework WebShaper to construct a dataset. WebShaper systematically formalizes IS tasks through set theory. Central to the formalization is the concept of Knowledge Projections (KP), which enables precise control over reasoning structure by KP operation compositions. During synthesis, we begin by creating seed tasks, then use a multi-step expansion process. At each step, an agentic Expander expands the current formal question more complex with retrieval and validation tools based on our formalization. We train our model on the synthesized dataset. Experiment results demonstrate that WebShaper achieves state-of-the-art performance among open-sourced IS agents on GAIA and WebWalkerQA benchmarks.", "date": "2025-07-20", "categories": []}
{"id": "2507.03331", "title": "Task-Specific Generative Dataset Distillation with Difficulty-Guided Sampling", "abstract": "To alleviate the reliance of deep neural networks on large-scale datasets, dataset distillation aims to generate compact, high-quality synthetic datasets that can achieve comparable performance to the original dataset. The integration of generative models has significantly advanced this field. However, existing approaches primarily focus on aligning the distilled dataset with the original one, often overlooking task-specific information that can be critical for optimal downstream performance. In this paper, focusing on the downstream task of classification, we propose a task-specific sampling strategy for generative dataset distillation that incorporates the concept of difficulty to consider the requirements of the target task better. The final dataset is sampled from a larger image pool with a sampling distribution obtained by matching the difficulty distribution of the original dataset. A logarithmic transformation is applied as a pre-processing step to correct for distributional bias. The results of extensive experiments demonstrate the effectiveness of our method and suggest its potential for enhancing performance on other downstream tasks. The code is available at https://github.com/SumomoTaku/DiffGuideSamp.", "date": "2025-07-04", "categories": []}
{"id": "2507.12750", "title": "Multimodal-Guided Dynamic Dataset Pruning for Robust and Efficient Data-Centric Learning", "abstract": "Modern deep models are trained on large real-world datasets, where data quality varies and redundancy is common. Data-centric approaches such as dataset pruning have shown promise in improving training efficiency and model performance. However, most existing methods rely on static heuristics or task-specific metrics, limiting their robustness and generalizability across domains. In this work, we introduce a dynamic dataset pruning framework that adaptively selects training samples based on both task-driven difficulty and cross-modality semantic consistency. By incorporating supervision from pretrained multimodal foundation models, our approach captures training dynamics while effectively filtering out uninformative samples. Our work highlights the potential of integrating cross-modality alignment for robust sample selection, advancing data-centric learning toward more efficient and robust practices across application domains.", "date": "2025-07-17", "categories": []}
{"id": "2502.10341", "title": "Organize the Web: Constructing Domains Enhances Pre-Training Data Curation", "abstract": "Modern language models are trained on large, unstructured datasets consisting of trillions of tokens and obtained by crawling the web. The unstructured nature makes it difficult to reason about their contents and develop systematic approaches to data curation. In this paper, we unpack monolithic web corpora by developing taxonomies of their contents and organizing them into domains. We introduce WebOrganizer, a framework for organizing web pages in terms of both their topic and format. Using these two complementary notions of domains, we automatically annotate pre-training data by distilling annotations from a large language model into efficient classifiers. This allows us to study how data from different domains should be mixed to improve models on downstream tasks, and we show that we can combine insights about effective topics and formats to further boost performance. We demonstrate that our domain mixing also improves existing methods that select data based on quality. Furthermore, we study and compare how quality-based methods will implicitly change the domain mixture. Overall, our work demonstrates that constructing and mixing domains provides a valuable complement to quality-based data curation methods, opening new avenues for effective and insightful pre-training data curation.", "date": "2025-02-14", "categories": []}
{"id": "2507.12466", "title": "Language Models Improve When Pretraining Data Matches Target Tasks", "abstract": "Every data selection method inherently has a target. In practice, these targets often emerge implicitly through benchmark-driven iteration: researchers develop selection strategies, train models, measure benchmark performance, then refine accordingly. This raises a natural question: what happens when we make this optimization explicit? To explore this, we propose benchmark-targeted ranking (BETR), a simple method that selects pretraining documents based on similarity to benchmark training examples. BETR embeds benchmark examples and a sample of pretraining documents in a shared space, scores this sample by similarity to benchmarks, then trains a lightweight classifier to predict these scores for the full corpus. We compare data selection methods by training over 500 models spanning $10^{19}$ to $10^{22}$ FLOPs and fitting scaling laws to them. From this, we find that simply aligning pretraining data to evaluation benchmarks using BETR achieves a 2.1x compute multiplier over DCLM-Baseline (4.7x over unfiltered data) and improves performance on 9 out of 10 tasks across all scales. BETR also generalizes well: when targeting a diverse set of benchmarks disjoint from our evaluation suite, it still matches or outperforms baselines. Our scaling analysis further reveals a clear trend: larger models require less aggressive filtering. Overall, our findings show that directly matching pretraining data to target tasks precisely shapes model capabilities and highlight that optimal selection strategies must adapt to model scale.", "date": "2025-07-16", "categories": []}
{"id": "2412.06136", "title": "AIDE: Attribute-Guided MultI-Hop Data Expansion for Data Scarcity in Task-Specific Fine-tuning", "abstract": "Fine-tuning large language models (LLMs) for specific tasks requires diverse, high-quality training data. However, obtaining sufficient relevant data remains a significant challenge. Existing data synthesis methods either depend on extensive seed datasets or struggle to balance task relevance and data diversity. To address these challenges, we propose Attribute-guided multI-hop Data Expansion (AIDE), a novel data synthesis framework that uses a multi-hop process to expand very few seed data points while ensuring data diversity and task relevance. AIDE extracts the main topic and key knowledge attributes from the seeds to guide the synthesis steps. The process repeats for K hops, using the generated data as seeds. To prevent irrelevant data generation as the hop depth increases, AIDE incorporates a residual connection mechanism. Our empirical results show that AIDE enables fine-tuning of Mistral-7B, Llama-3.1-8B and Llama-3.2-3B from 10 seeds, surpassing the models fine-tuned on human curated data. Furthermore, AIDE outperforms state-of-the-art data synthesis methods, such as Evol-Instruct, by over 30% in task-specific fine-tuning. Code is available at https://github.com/Code4Graph/AIDE.", "date": "2024-12-09", "categories": []}
{"id": "2405.11467", "title": "AdaAugment: A Tuning-Free and Adaptive Approach to Enhance Data Augmentation", "abstract": "Data augmentation (DA) is widely employed to improve the generalization performance of deep models. However, most existing DA methods employ augmentation operations with fixed or random magnitudes throughout the training process. While this fosters data diversity, it can also inevitably introduce uncontrolled variability in augmented data, which could potentially cause misalignment with the evolving training status of the target models. Both theoretical and empirical findings suggest that this misalignment increases the risks of both underfitting and overfitting. To address these limitations, we propose AdaAugment, an innovative and tuning-free adaptive augmentation method that leverages reinforcement learning to dynamically and adaptively adjust augmentation magnitudes for individual training samples based on real-time feedback from the target network. Specifically, AdaAugment features a dual-model architecture consisting of a policy network and a target network, which are jointly optimized to adapt augmentation magnitudes in accordance with the model's training progress effectively. The policy network optimizes the variability within the augmented data, while the target network utilizes the adaptively augmented samples for training. These two networks are jointly optimized and mutually reinforce each other. Extensive experiments across benchmark datasets and deep architectures demonstrate that AdaAugment consistently outperforms other state-of-the-art DA methods in effectiveness while maintaining remarkable efficiency. Code is available at https://github.com/Jackbrocp/AdaAugment.", "date": "2024-05-19", "categories": []}
{"id": "2504.07426", "title": "Conditional Data Synthesis Augmentation", "abstract": "Reliable machine learning and statistical analysis rely on diverse, well-distributed training data. However, real-world datasets are often limited in size and exhibit underrepresentation across key subpopulations, leading to biased predictions and reduced performance, particularly in supervised tasks such as classification. To address these challenges, we propose Conditional Data Synthesis Augmentation (CoDSA), a novel framework that leverages generative models, such as diffusion models, to synthesize high-fidelity data for improving model performance across multimodal domains including tabular, textual, and image data. CoDSA generates synthetic samples that faithfully capture the conditional distributions of the original data, with a focus on under-sampled or high-interest regions. Through transfer learning, CoDSA fine-tunes pre-trained generative models to enhance the realism of synthetic data and increase sample density in sparse areas. This process preserves inter-modal relationships, mitigates data imbalance, improves domain adaptation, and boosts generalization. We also introduce a theoretical framework that quantifies the statistical accuracy improvements enabled by CoDSA as a function of synthetic sample volume and targeted region allocation, providing formal guarantees of its effectiveness. Extensive experiments demonstrate that CoDSA consistently outperforms non-adaptive augmentation strategies and state-of-the-art baselines in both supervised and unsupervised settings.", "date": "2025-04-10", "categories": []}
{"id": "2405.09150", "title": "Curriculum Dataset Distillation", "abstract": "Most dataset distillation methods struggle to accommodate large-scale datasets due to their substantial computational and memory requirements. Recent research has begun to explore scalable disentanglement methods. However, there are still performance bottlenecks and room for optimization in this direction. In this paper, we present a curriculum-based dataset distillation framework aiming to harmonize performance and scalability. This framework strategically distills synthetic images, adhering to a curriculum that transitions from simple to complex. By incorporating curriculum evaluation, we address the issue of previous methods generating images that tend to be homogeneous and simplistic, doing so at a manageable computational cost. Furthermore, we introduce adversarial optimization towards synthetic images to further improve their representativeness and safeguard against their overfitting to the neural network involved in distilling. This enhances the generalization capability of the distilled images across various neural network architectures and also increases their robustness to noise. Extensive experiments demonstrate that our framework sets new benchmarks in large-scale dataset distillation, achieving substantial improvements of 11.1\\% on Tiny-ImageNet, 9.0\\% on ImageNet-1K, and 7.3\\% on ImageNet-21K. Our distilled datasets and code are available at https://github.com/MIV-XJTU/CUDD.", "date": "2024-05-15", "categories": []}
{"id": "2312.01991", "title": "Shapley-Based Data Valuation with Mutual Information: A Key to Modified K-Nearest Neighbors", "abstract": "The K-Nearest Neighbors (KNN) algorithm is widely used for classification and regression; however, it suffers from limitations, including the equal treatment of all samples. We propose Information-Modified KNN (IM-KNN), a novel approach that leverages Mutual Information ($I$) and Shapley values to assign weighted values to neighbors, thereby bridging the gap in treating all samples with the same value and weight. On average, IM-KNN improves the accuracy, precision, and recall of traditional KNN by 16.80%, 17.08%, and 16.98%, respectively, across 12 benchmark datasets. Experiments on four large-scale datasets further highlight IM-KNN's robustness to noise, imbalanced data, and skewed distributions.", "date": "2023-12-04", "categories": []}
{"id": "2506.11315", "title": "Sampling Imbalanced Data with Multi-objective Bilevel Optimization", "abstract": "Two-class classification problems are often characterized by an imbalance between the number of majority and minority datapoints resulting in poor classification of the minority class in particular. Traditional approaches, such as reweighting the loss function or na\\\"ive resampling, risk overfitting and subsequently fail to improve classification because they do not consider the diversity between majority and minority datasets. Such consideration is infeasible because there is no metric that can measure the impact of imbalance on the model. To obviate these challenges, we make two key contributions. First, we introduce MOODS~(Multi-Objective Optimization for Data Sampling), a novel multi-objective bilevel optimization framework that guides both synthetic oversampling and majority undersampling. Second, we introduce a validation metric -- `$\\epsilon/ \\delta$ non-overlapping diversification metric' -- that quantifies the goodness of a sampling method towards model performance. With this metric we experimentally demonstrate state-of-the-art performance with improvement in diversity driving a $1-15 \\%$ increase in $F1$ scores.", "date": "2025-06-12", "categories": []}
{"id": "2506.22637", "title": "CaO$_2$: Rectifying Inconsistencies in Diffusion-Based Dataset Distillation", "abstract": "The recent introduction of diffusion models in dataset distillation has shown promising potential in creating compact surrogate datasets for large, high-resolution target datasets, offering improved efficiency and performance over traditional bi-level/uni-level optimization methods. However, current diffusion-based dataset distillation approaches overlook the evaluation process and exhibit two critical inconsistencies in the distillation process: (1) Objective Inconsistency, where the distillation process diverges from the evaluation objective, and (2) Condition Inconsistency, leading to mismatches between generated images and their corresponding conditions. To resolve these issues, we introduce Condition-aware Optimization with Objective-guided Sampling (CaO$_2$), a two-stage diffusion-based framework that aligns the distillation process with the evaluation objective. The first stage employs a probability-informed sample selection pipeline, while the second stage refines the corresponding latent representations to improve conditional likelihood. CaO$_2$ achieves state-of-the-art performance on ImageNet and its subsets, surpassing the best-performing baselines by an average of 2.3% accuracy.", "date": "2025-06-27", "categories": []}
{"id": "2507.03253", "title": "RefineX: Learning to Refine Pre-training Data at Scale from Expert-Guided Programs", "abstract": "The foundational capabilities of large language models (LLMs) are deeply influenced by the quality of their pre-training corpora. However, enhancing data quality at scale remains a significant challenge, primarily due to the trade-off between refinement effectiveness and processing efficiency. While rule-based filtering remains the dominant paradigm, it typically operates at the document level and lacks the granularity needed to refine specific content within documents. Inspired by emerging work such as ProX, we propose $\\textbf{RefineX}$, a novel framework for large-scale, surgical refinement of pre-training data through programmatic editing tasks. RefineX enables efficient and fine-grained data refinement while reliably preserving the diversity and naturalness of raw text. The core strength of RefineX lies in distilling high-quality, expert-guided end-to-end refinement results into minimal edit-based deletion programs. This high-precision distillation pipeline is used to train an efficient and reliable refine model that can systematically improve every instance in the corpus at scale. We evaluate RefineX across from-scratch pre-training at multiple model scales and find that it consistently outperforms models trained on raw, filtered, or alternatively refined data across diverse downstream tasks. On the 750M model, RefineX yields 2.6%-7.2% average gains on lighteval tasks, and achieves comparable performance using significantly fewer training tokens. Further analysis shows that RefineX reliably enhances text quality with both high efficiency and precision, outperforming prior approaches such as end-to-end generation and Prox-C. These results position RefineX as a scalable, effective, and reliable solution for optimizing pre-training data in modern LLM pipelines.", "date": "2025-07-04", "categories": []}
{"id": "2412.01460", "title": "A Comprehensive Study of Shapley Value in Data Analytics", "abstract": "Over the recent years, Shapley value (SV), a solution concept from cooperative game theory, has found numerous applications in data analytics (DA). This paper presents the first comprehensive study of SV used throughout the DA workflow, clarifying the key variables in defining DA-applicable SV and the essential functionalities that SV can provide for data scientists. We condense four primary challenges of using SV in DA, namely computation efficiency, approximation error, privacy preservation, and interpretability, disentangle the resolution techniques from existing arts in this field, then analyze and discuss the techniques w.r.t. each challenge and the potential conflicts between challenges.We also implement SVBench, a modular and extensible open-source framework for developing SV applications in different DA tasks, and conduct extensive evaluations to validate our analyses and discussions. Based on the qualitative and quantitative results, we identify the limitations of current efforts for applying SV to DA and highlight the directions of future research and engineering.", "date": "2024-12-02", "categories": []}
{"id": "2502.16152", "title": "DUPRE: Data Utility Prediction for Efficient Data Valuation", "abstract": "Data valuation is increasingly used in machine learning (ML) to decide the fair compensation for data owners and identify valuable or harmful data for improving ML models. Cooperative game theory-based data valuation, such as Data Shapley, requires evaluating the data utility (e.g., validation accuracy) and retraining the ML model for multiple data subsets. While most existing works on efficient estimation of the Shapley values have focused on reducing the number of subsets to evaluate, our framework, \\texttt{DUPRE}, takes an alternative yet complementary approach that reduces the cost per subset evaluation by predicting data utilities instead of evaluating them by model retraining. Specifically, given the evaluated data utilities of some data subsets, \\texttt{DUPRE} fits a \\emph{Gaussian process} (GP) regression model to predict the utility of every other data subset. Our key contribution lies in the design of our GP kernel based on the sliced Wasserstein distance between empirical data distributions. In particular, we show that the kernel is valid and positive semi-definite, encodes prior knowledge of similarities between different data subsets, and can be efficiently computed. We empirically verify that \\texttt{DUPRE} introduces low prediction error and speeds up data valuation for various ML models, datasets, and utility functions.", "date": "2025-02-22", "categories": []}
{"id": "2507.03004", "title": "CLUES: Collaborative High-Quality Data Selection for LLMs via Training Dynamics", "abstract": "Recent research has highlighted the importance of data quality in scaling large language models (LLMs). However, automated data quality control faces unique challenges in collaborative settings where sharing is not allowed directly between data silos. To tackle this issue, this paper proposes a novel data quality control technique based on the notion of data influence on the training dynamics of LLMs, that high quality data are more likely to have similar training dynamics to the anchor dataset. We then leverage the influence of the training dynamics to select high-quality data from different private domains, with centralized model updates on the server side in a collaborative training fashion by either model merging or federated learning. As for the data quality indicator, we compute the per-sample gradients with respect to the private data and the anchor dataset, and use the trace of the accumulated inner products as a measurement of data quality. In addition, we develop a quality control evaluation tailored for collaborative settings with heterogeneous domain data. Experiments show that training on the high-quality data selected by our method can often outperform other data selection methods for collaborative fine-tuning of LLMs, across diverse private domain datasets, in medical, multilingual and financial settings. Our code is released at github.com/Ryan0v0/CLUES.", "date": "2025-07-02", "categories": []}
{"id": "2507.03648", "title": "Disentangling the Roles of Representation and Selection in Data Pruning", "abstract": "Data pruning, selecting small but impactful subsets, offers a promising way to efficiently scale NLP model training. However, existing methods often involve many different design choices, which have not been systematically studied. This limits future developments. In this work, we decompose data pruning into two key components: the data representation and the selection algorithm, and we systematically analyze their influence on the selection of instances. Our theoretical and empirical results highlight the crucial role of representations: better representations, e.g., training gradients, generally lead to a better selection of instances, regardless of the chosen selection algorithm. Furthermore, different selection algorithms excel in different settings, and none consistently outperforms the others. Moreover, the selection algorithms do not always align with their intended objectives: for example, algorithms designed for the same objective can select drastically different instances, highlighting the need for careful evaluation.", "date": "2025-07-04", "categories": []}
{"id": "2507.04529", "title": "A Data-Driven Novelty Score for Diverse In-Vehicle Data Recording", "abstract": "High-quality datasets are essential for training robust perception systems in autonomous driving. However, real-world data collection is often biased toward common scenes and objects, leaving novel cases underrepresented. This imbalance hinders model generalization and compromises safety. The core issue is the curse of rarity. Over time, novel events occur infrequently, and standard logging methods fail to capture them effectively. As a result, large volumes of redundant data are stored, while critical novel cases are diluted, leading to biased datasets. This work presents a real-time data selection method focused on object-level novelty detection to build more balanced and diverse datasets. The method assigns a data-driven novelty score to image frames using a novel dynamic Mean Shift algorithm. It models normal content based on mean and covariance statistics to identify frames with novel objects, discarding those with redundant elements. The main findings show that reducing the training dataset size with this method can improve model performance, whereas higher redundancy tends to degrade it. Moreover, as data redundancy increases, more aggressive filtering becomes both possible and beneficial. While random sampling can offer some gains, it often leads to overfitting and unpredictability in outcomes. The proposed method supports real-time deployment with 32 frames per second and is constant over time. By continuously updating the definition of normal content, it enables efficient detection of novelties in a continuous data stream.", "date": "2025-07-06", "categories": []}
{"id": "2507.04619", "title": "Information-Guided Diffusion Sampling for Dataset Distillation", "abstract": "Dataset distillation aims to create a compact dataset that retains essential information while maintaining model performance. Diffusion models (DMs) have shown promise for this task but struggle in low images-per-class (IPC) settings, where generated samples lack diversity. In this paper, we address this issue from an information-theoretic perspective by identifying two key types of information that a distilled dataset must preserve: ($i$) prototype information $\\mathrm{I}(X;Y)$, which captures label-relevant features; and ($ii$) contextual information $\\mathrm{H}(X | Y)$, which preserves intra-class variability. Here, $(X,Y)$ represents the pair of random variables corresponding to the input data and its ground truth label, respectively. Observing that the required contextual information scales with IPC, we propose maximizing $\\mathrm{I}(X;Y) + \\beta \\mathrm{H}(X | Y)$ during the DM sampling process, where $\\beta$ is IPC-dependent. Since directly computing $\\mathrm{I}(X;Y)$ and $\\mathrm{H}(X | Y)$ is intractable, we develop variational estimations to tightly lower-bound these quantities via a data-driven approach. Our approach, information-guided diffusion sampling (IGDS), seamlessly integrates with diffusion models and improves dataset distillation across all IPC settings. Experiments on Tiny ImageNet and ImageNet subsets show that IGDS significantly outperforms existing methods, particularly in low-IPC regimes. The code will be released upon acceptance.", "date": "2025-07-07", "categories": []}
{"id": "2412.00420", "title": "TAROT: Targeted Data Selection via Optimal Transport", "abstract": "We propose TAROT, a targeted data selection framework grounded in optimal transport theory. Previous targeted data selection methods primarily rely on influence-based greedy heuristics to enhance domain-specific performance. While effective on limited, unimodal data (i.e., data following a single pattern), these methods struggle as target data complexity increases. Specifically, in multimodal distributions, these heuristics fail to account for multiple inherent patterns, leading to suboptimal data selection. This work identifies two primary factors contributing to this limitation: (i) the disproportionate impact of dominant feature components in high-dimensional influence estimation, and (ii) the restrictive linear additive assumptions inherent in greedy selection strategies. To address these challenges, TAROT incorporates whitened feature distance to mitigate dominant feature bias, providing a more reliable measure of data influence. Building on this, TAROT uses whitened feature distance to quantify and minimize the optimal transport distance between the selected data and target domains. Notably, this minimization also facilitates the estimation of optimal selection ratios. We evaluate TAROT across multiple tasks, including semantic segmentation, motion prediction, and instruction tuning. Results consistently show that TAROT outperforms state-of-the-art methods, highlighting its versatility across various deep learning tasks. Code is available at https://github.com/vita-epfl/TAROT.", "date": "2024-11-30", "categories": []}
{"id": "2502.05673", "title": "The Evolution of Dataset Distillation: Toward Scalable and Generalizable Solutions", "abstract": "Dataset distillation, which condenses large-scale datasets into compact synthetic representations, has emerged as a critical solution for training modern deep learning models efficiently. While prior surveys focus on developments before 2023, this work comprehensively reviews recent advances, emphasizing scalability to large-scale datasets such as ImageNet-1K and ImageNet-21K. We categorize progress into a few key methodologies: trajectory matching, gradient matching, distribution matching, scalable generative approaches, and decoupling optimization mechanisms. As a comprehensive examination of recent dataset distillation advances, this survey highlights breakthrough innovations: the SRe2L framework for efficient and effective condensation, soft label strategies that significantly enhance model accuracy, and lossless distillation techniques that maximize compression while maintaining performance. Beyond these methodological advancements, we address critical challenges, including robustness against adversarial and backdoor attacks, effective handling of non-IID data distributions. Additionally, we explore emerging applications in video and audio processing, multi-modal learning, medical imaging, and scientific computing, highlighting its domain versatility. By offering extensive performance comparisons and actionable research directions, this survey equips researchers and practitioners with practical insights to advance efficient and generalizable dataset distillation, paving the way for future innovations.", "date": "2025-02-08", "categories": []}
{"id": "2506.23799", "title": "KAIROS: Scalable Model-Agnostic Data Valuation", "abstract": "Training data increasingly shapes not only model accuracy but also regulatory compliance and market valuation of AI assets. Yet existing valuation methods remain inadequate: model-based techniques depend on a single fitted model and inherit its biases, while algorithm-based approaches such as Data Shapley require costly retrainings at web scale. Recent Wasserstein-based model-agnostic methods rely on approximations that misrank examples relative to their true leave-one-out (LOO) utility. We introduce KAIROS, a scalable, model-agnostic valuation framework that assigns each example a distributional influence score: its contribution to the Maximum Mean Discrepancy (MMD) between the empirical training distribution and a clean reference set. Unlike Wasserstein surrogates, our MMD-based influence admits a closed-form solution that faithfully approximates the exact LOO ranking within $O(1/N^2)$ error, requires no retraining, and naturally extends to conditional kernels for unified label- and feature-error detection. Moreover, KAIROS supports efficient online updates: when a new batch of size m arrives, all scores can be updated in $O(mN)$ time, delivering up to 50x speedup without compromising ranking quality. Empirical evaluations on noise, mislabeling, and poisoning benchmarks show that KAIROS consistently outperforms state-of-the-art model-, Shapley-, and Wasserstein-based baselines in both accuracy and runtime. We provide rigorous theoretical guarantees, including symmetry for reproducible rankings and density-separation for interpretable thresholds.", "date": "2025-06-30", "categories": []}
{"id": "2311.18531", "title": "Dataset Distillation via the Wasserstein Metric", "abstract": "Dataset Distillation (DD) aims to generate a compact synthetic dataset that enables models to achieve performance comparable to training on the full large dataset, significantly reducing computational costs. Drawing from optimal transport theory, we introduce WMDD (Wasserstein Metric-based Dataset Distillation), a straightforward yet powerful method that employs the Wasserstein metric to enhance distribution matching.\n  We compute the Wasserstein barycenter of features from a pretrained classifier to capture essential characteristics of the original data distribution. By optimizing synthetic data to align with this barycenter in feature space and leveraging per-class BatchNorm statistics to preserve intra-class variations, WMDD maintains the efficiency of distribution matching approaches while achieving state-of-the-art results across various high-resolution datasets. Our extensive experiments demonstrate WMDD's effectiveness and adaptability, highlighting its potential for advancing machine learning applications at scale.", "date": "2023-11-30", "categories": []}
{"id": "2504.05250", "title": "PEAKS: Selecting Key Training Examples Incrementally via Prediction Error Anchored by Kernel Similarity", "abstract": "As deep learning continues to be driven by ever-larger datasets, understanding which examples are most important for generalization has become a critical question. While progress in data selection continues, emerging applications require studying this problem in dynamic contexts. To bridge this gap, we pose the Incremental Data Selection (IDS) problem, where examples arrive as a continuous stream, and need to be selected without access to the full data source. In this setting, the learner must incrementally build a training dataset of predefined size while simultaneously learning the underlying task. We find that in IDS, the impact of a new sample on the model state depends fundamentally on both its geometric relationship in the feature space and its prediction error. Leveraging this insight, we propose PEAKS (Prediction Error Anchored by Kernel Similarity), an efficient data selection method tailored for IDS. Our comprehensive evaluations demonstrate that PEAKS consistently outperforms existing selection strategies. Furthermore, PEAKS yields increasingly better performance returns than random selection as training data size grows on real-world datasets. The code is available at https://github.com/BurakGurbuz97/PEAKS.", "date": "2025-04-07", "categories": []}
{"id": "2507.00049", "title": "AdaDeDup: Adaptive Hybrid Data Pruning for Efficient Large-Scale Object Detection Training", "abstract": "The computational burden and inherent redundancy of large-scale datasets challenge the training of contemporary machine learning models. Data pruning offers a solution by selecting smaller, informative subsets, yet existing methods struggle: density-based approaches can be task-agnostic, while model-based techniques may introduce redundancy or prove computationally prohibitive. We introduce Adaptive De-Duplication (AdaDeDup), a novel hybrid framework that synergistically integrates density-based pruning with model-informed feedback in a cluster-adaptive manner. AdaDeDup first partitions data and applies an initial density-based pruning. It then employs a proxy model to evaluate the impact of this initial pruning within each cluster by comparing losses on kept versus pruned samples. This task-aware signal adaptively adjusts cluster-specific pruning thresholds, enabling more aggressive pruning in redundant clusters while preserving critical data in informative ones. Extensive experiments on large-scale object detection benchmarks (Waymo, COCO, nuScenes) using standard models (BEVFormer, Faster R-CNN) demonstrate AdaDeDup's advantages. It significantly outperforms prominent baselines, substantially reduces performance degradation (e.g., over 54% versus random sampling on Waymo), and achieves near-original model performance while pruning 20% of data, highlighting its efficacy in enhancing data efficiency for large-scale model training. Code is open-sourced.", "date": "2025-06-24", "categories": []}
{"id": "2507.00054", "title": "Enhancing Reasoning Capabilities in SLMs with Reward Guided Dataset Distillation", "abstract": "The push to compress and impart the proficiency of Large Language Models (LLMs) into more deployable and efficient Small Language Models (SLMs) has benefited from improvements in knowledge distillation (KD) techniques. These techniques allow a smaller student model to learn from a more capable and larger teacher model's responses. However, distillation often revolves around the student model merely copying the teacher's in-distribution responses, limiting its generalisability. This limitation is amplified on reasoning tasks and can be computationally expensive. In this study, we propose AdvDistill, a reward-guided dataset distillation framework. We utilise multiple generations (responses) from a teacher for each prompt and assign rewards based on rule-based verifiers. These varying and normally distributed rewards serve as weights when training student models. Our methods and their subsequent behavioural analysis demonstrate a significant improvement in student model performance for mathematical and complex reasoning tasks, showcasing the efficacy and benefits of incorporating a rewarding mechanism in dataset distillation processes.", "date": "2025-06-25", "categories": []}
{"id": "2506.23580", "title": "Dataset Distillation via Vision-Language Category Prototype", "abstract": "Dataset distillation (DD) condenses large datasets into compact yet informative substitutes, preserving performance comparable to the original dataset while reducing storage, transmission costs, and computational consumption. However, previous DD methods mainly focus on distilling information from images, often overlooking the semantic information inherent in the data. The disregard for context hinders the model's generalization ability, particularly in tasks involving complex datasets, which may result in illogical outputs or the omission of critical objects. In this study, we integrate vision-language methods into DD by introducing text prototypes to distill language information and collaboratively synthesize data with image prototypes, thereby enhancing dataset distillation performance. Notably, the text prototypes utilized in this study are derived from descriptive text information generated by an open-source large language model. This framework demonstrates broad applicability across datasets without pre-existing text descriptions, expanding the potential of dataset distillation beyond traditional image-based approaches. Compared to other methods, the proposed approach generates logically coherent images containing target objects, achieving state-of-the-art validation performance and demonstrating robust generalization. Source code and generated data are available in https://github.com/zou-yawen/Dataset-Distillation-via-Vision-Language-Category-Prototype/", "date": "2025-06-30", "categories": []}
{"id": "2506.23674", "title": "Partial Forward Blocking: A Novel Data Pruning Paradigm for Lossless Training Acceleration", "abstract": "The ever-growing size of training datasets enhances the generalization capability of modern machine learning models but also incurs exorbitant computational costs. Existing data pruning approaches aim to accelerate training by removing those less important samples. However, they often rely on gradients or proxy models, leading to prohibitive additional costs of gradient back-propagation and proxy model training. In this paper, we propose Partial Forward Blocking (PFB), a novel framework for lossless training acceleration. The efficiency of PFB stems from its unique adaptive pruning pipeline: sample importance is assessed based on features extracted from the shallow layers of the target model. Less important samples are then pruned, allowing only the retained ones to proceed with the subsequent forward pass and loss back-propagation. This mechanism significantly reduces the computational overhead of deep-layer forward passes and back-propagation for pruned samples, while also eliminating the need for auxiliary backward computations and proxy model training. Moreover, PFB introduces probability density as an indicator of sample importance. Combined with an adaptive distribution estimation module, our method dynamically prioritizes relatively rare samples, aligning with the constantly evolving training state. Extensive experiments demonstrate the significant superiority of PFB in performance and speed. On ImageNet, PFB achieves a 0.5% accuracy improvement and 33% training time reduction with 40% data pruned.", "date": "2025-06-30", "categories": []}
{"id": "2506.24125", "title": "FADRM: Fast and Accurate Data Residual Matching for Dataset Distillation", "abstract": "Residual connection has been extensively studied and widely applied at the model architecture level. However, its potential in the more challenging data-centric approaches remains unexplored. In this work, we introduce the concept of Data Residual Matching for the first time, leveraging data-level skip connections to facilitate data generation and mitigate data information vanishing. This approach maintains a balance between newly acquired knowledge through pixel space optimization and existing core local information identification within raw data modalities, specifically for the dataset distillation task. Furthermore, by incorporating optimization-level refinements, our method significantly improves computational efficiency, achieving superior performance while reducing training time and peak GPU memory usage by 50%. Consequently, the proposed method Fast and Accurate Data Residual Matching for Dataset Distillation (FADRM) establishes a new state-of-the-art, demonstrating substantial improvements over existing methods across multiple dataset benchmarks in both efficiency and effectiveness. For instance, with ResNet-18 as the student model and a 0.8% compression ratio on ImageNet-1K, the method achieves 47.7% test accuracy in single-model dataset distillation and 50.0% in multi-model dataset distillation, surpassing RDED by +5.7% and outperforming state-of-the-art multi-model approaches, EDC and CV-DD, by +1.4% and +4.0%. Code is available at: https://github.com/Jiacheng8/FADRM.", "date": "2025-06-30", "categories": []}
{"id": "2506.14473", "title": "Foundation Model Insights and a Multi-Model Approach for Superior Fine-Grained One-shot Subset Selection", "abstract": "One-shot subset selection serves as an effective tool to reduce deep learning training costs by identifying an informative data subset based on the information extracted by an information extractor (IE). Traditional IEs, typically pre-trained on the target dataset, are inherently dataset-dependent. Foundation models (FMs) offer a promising alternative, potentially mitigating this limitation. This work investigates two key questions: (1) Can FM-based subset selection outperform traditional IE-based methods across diverse datasets? (2) Do all FMs perform equally well as IEs for subset selection? Extensive experiments uncovered surprising insights: FMs consistently outperform traditional IEs on fine-grained datasets, whereas their advantage diminishes on coarse-grained datasets with noisy labels. Motivated by these finding, we propose RAM-APL (RAnking Mean-Accuracy of Pseudo-class Labels), a method tailored for fine-grained image datasets. RAM-APL leverages multiple FMs to enhance subset selection by exploiting their complementary strengths. Our approach achieves state-of-the-art performance on fine-grained datasets, including Oxford-IIIT Pet, Food-101, and Caltech-UCSD Birds-200-2011.", "date": "2025-06-17", "categories": []}
{"id": "2506.21545", "title": "Data Efficacy for Language Model Training", "abstract": "Data is fundamental to the training of language models (LM). Recent research has been dedicated to data efficiency, which aims to maximize performance by selecting a minimal or optimal subset of training data. Techniques such as data filtering, sampling, and selection play a crucial role in this area. To complement it, we define Data Efficacy, which focuses on maximizing performance by optimizing the organization of training data and remains relatively underexplored. This work introduces a general paradigm, DELT, for considering data efficacy in LM training, which highlights the significance of training data organization. DELT comprises three components: Data Scoring, Data Selection, and Data Ordering. Among these components, we design Learnability-Quality Scoring (LQS), as a new instance of Data Scoring, which considers both the learnability and quality of each data sample from the gradient consistency perspective. We also devise Folding Ordering (FO), as a novel instance of Data Ordering, which addresses issues such as model forgetting and data distribution bias. Comprehensive experiments validate the data efficacy in LM training, which demonstrates the following: Firstly, various instances of the proposed DELT enhance LM performance to varying degrees without increasing the data scale and model size. Secondly, among these instances, the combination of our proposed LQS for data scoring and Folding for data ordering achieves the most significant improvement. Lastly, data efficacy can be achieved together with data efficiency by applying data selection. Therefore, we believe that data efficacy is a promising foundational area in LM training.", "date": "2025-06-26", "categories": []}
{"id": "2506.21037", "title": "RL-Selector: Reinforcement Learning-Guided Data Selection via Redundancy Assessment", "abstract": "Modern deep architectures often rely on large-scale datasets, but training on these datasets incurs high computational and storage overhead. Real-world datasets often contain substantial redundancies, prompting the need for more data-efficient training paradigms. Data selection has shown promise to mitigate redundancy by identifying the most representative samples, thereby reducing training costs without compromising performance. Existing methods typically rely on static scoring metrics or pretrained models, overlooking the combined effect of selected samples and their evolving dynamics during training. We introduce the concept of epsilon-sample cover, which quantifies sample redundancy based on inter-sample relationships, capturing the intrinsic structure of the dataset. Based on this, we reformulate data selection as a reinforcement learning (RL) process and propose RL-Selector, where a lightweight RL agent optimizes the selection policy by leveraging epsilon-sample cover derived from evolving dataset distribution as a reward signal. Extensive experiments across benchmark datasets and diverse architectures demonstrate that our method consistently outperforms existing state-of-the-art baselines. Models trained with our selected datasets show enhanced generalization performance with improved training efficiency.", "date": "2025-06-26", "categories": []}
{"id": "2506.19261", "title": "Automated Image Recognition Framework", "abstract": "While the efficacy of deep learning models heavily relies on data, gathering and annotating data for specific tasks, particularly when addressing novel or sensitive subjects lacking relevant datasets, poses significant time and resource challenges. In response to this, we propose a novel Automated Image Recognition (AIR) framework that harnesses the power of generative AI. AIR empowers end-users to synthesize high-quality, pre-annotated datasets, eliminating the necessity for manual labeling. It also automatically trains deep learning models on the generated datasets with robust image recognition performance. Our framework includes two main data synthesis processes, AIR-Gen and AIR-Aug. The AIR-Gen enables end-users to seamlessly generate datasets tailored to their specifications. To improve image quality, we introduce a novel automated prompt engineering module that leverages the capabilities of large language models. We also introduce a distribution adjustment algorithm to eliminate duplicates and outliers, enhancing the robustness and reliability of generated datasets. On the other hand, the AIR-Aug enhances a given dataset, thereby improving the performance of deep classifier models. AIR-Aug is particularly beneficial when users have limited data for specific tasks. Through comprehensive experiments, we demonstrated the efficacy of our generated data in training deep learning models and showcased the system's potential to provide image recognition models for a wide range of objects. We also conducted a user study that achieved an impressive score of 4.4 out of 5.0, underscoring the AI community's positive perception of AIR.", "date": "2025-06-24", "categories": []}
{"id": "2503.17195", "title": "TreeSynth: Synthesizing Diverse Data from Scratch via Tree-Guided Subspace Partitioning", "abstract": "Model customization necessitates high-quality and diverse datasets, but acquiring such data remains time-consuming and labor-intensive. Despite the great potential of large language models (LLMs) for data synthesis, current approaches are constrained by limited seed data, model biases, and low-variation prompts, resulting in limited diversity and biased distributions with the increase of data scales. To tackle this challenge, we introduce TREESYNTH, a tree-guided subspace-based data synthesis approach inspired by decision trees. It constructs a spatial partitioning tree to recursively divide a task-specific full data space (i.e., root node) into numerous atomic subspaces (i.e., leaf nodes) with mutually exclusive and exhaustive attributes to ensure both distinctiveness and comprehensiveness before synthesizing samples within each atomic subspace. This globally dividing-and-synthesizing method finally collects subspace samples into a comprehensive dataset, effectively circumventing repetition and space collapse to ensure the diversity of large-scale data synthesis. Furthermore, the spatial partitioning tree enables sample allocation into atomic subspaces, allowing the rebalancing of existing datasets for more balanced and comprehensive distributions. Empirically, extensive experiments across diverse benchmarks consistently demonstrate the superior data diversity, model performance, and robust scalability of TREESYNTH compared to both human-crafted datasets and peer data synthesis methods, with an average performance gain reaching 10%. Besides, the consistent improvements of TREESYNTH-balanced datasets highlight its efficacious application to redistribute existing datasets for more comprehensive coverage and the induced performance enhancement. The code is available at https://github.com/cpa2001/TreeSynth.", "date": "2025-03-21", "categories": []}
{"id": "2410.11215", "title": "A CLIP-Powered Framework for Robust and Generalizable Data Selection", "abstract": "Large-scale datasets have been pivotal to the advancements of deep learning models in recent years, but training on such large datasets invariably incurs substantial storage and computational overhead. Meanwhile, real-world datasets often contain redundant and noisy data, imposing a negative impact on training efficiency and model performance. Data selection has shown promise in identifying the most representative samples from the entire dataset, which aims to minimize the performance gap with reduced training costs. Existing works typically rely on single-modality information to assign importance scores for individual samples, which may lead to inaccurate assessments, especially when dealing with noisy or corrupted samples. To address this limitation, we propose a novel CLIP-powered data selection framework that leverages multimodal information for more robust and generalizable sample selection. Specifically, our framework consists of three key modules-dataset adaptation, sample scoring, and selection optimization-that together harness extensive pre-trained multimodal knowledge to comprehensively assess sample influence and optimize the selection results through multi-objective optimization. Extensive experiments demonstrate that our approach consistently outperforms existing state-of-the-art baselines on various benchmark datasets. Notably, our method effectively removes noisy or damaged samples from the dataset, enabling it to achieve even higher performance with less data. This indicates that it is not only a way to accelerate training but can also improve overall data quality.", "date": "2024-10-15", "categories": []}
{"id": "2502.14709", "title": "Group-Level Data Selection for Efficient Pretraining", "abstract": "In this paper, we introduce Group-MATES, an efficient group-level data selection approach to optimize the speed-quality frontier of language model pretraining. Specifically, Group-MATES parameterizes costly group-level selection with a relational data influence model. To train this model, we sample training trajectories of the language model and collect oracle data influences alongside. The relational data influence model approximates the oracle data influence by weighting individual influence with relationships among training data. To enable efficient selection with our relational data influence model, we partition the dataset into small clusters using relationship weights and select data within each cluster independently. Experiments on DCLM 400M-4x, 1B-1x, and 3B-1x show that Group-MATES achieves 3.5%-9.4% relative performance gains over random selection across 22 downstream tasks, nearly doubling the improvements achieved by state-of-the-art individual data selection baselines. Furthermore, Group-MATES reduces the number of tokens required to reach a certain downstream performance by up to 1.75x, substantially elevating the speed-quality frontier. Further analyses highlight the critical role of relationship weights in the relational data influence model and the effectiveness of our cluster-based inference. Our code is open-sourced at https://github.com/facebookresearch/Group-MATES.", "date": "2025-02-20", "categories": []}
{"id": "2506.09010", "title": "Effective Data Pruning through Score Extrapolation", "abstract": "Training advanced machine learning models demands massive datasets, resulting in prohibitive computational costs. To address this challenge, data pruning techniques identify and remove redundant training samples while preserving model performance. Yet, existing pruning techniques predominantly require a full initial training pass to identify removable samples, negating any efficiency benefits for single training runs. To overcome this limitation, we introduce a novel importance score extrapolation framework that requires training on only a small subset of data. We present two initial approaches in this framework - k-nearest neighbors and graph neural networks - to accurately predict sample importance for the entire dataset using patterns learned from this minimal subset. We demonstrate the effectiveness of our approach for 2 state-of-the-art pruning methods (Dynamic Uncertainty and TDDS), 4 different datasets (CIFAR-10, CIFAR-100, Places-365, and ImageNet), and 3 training paradigms (supervised, unsupervised, and adversarial). Our results indicate that score extrapolation is a promising direction to scale expensive score calculation methods, such as pruning, data attribution, or other tasks.", "date": "2025-06-10", "categories": []}
{"id": "2506.14111", "title": "Essential-Web v1.0: 24T tokens of organized web data", "abstract": "Data plays the most prominent role in how language models acquire skills and knowledge. The lack of massive, well-organized pre-training datasets results in costly and inaccessible data pipelines. We present Essential-Web v1.0, a 24-trillion-token dataset in which every document is annotated with a twelve-category taxonomy covering topic, format, content complexity, and quality. Taxonomy labels are produced by EAI-Distill-0.5b, a fine-tuned 0.5b-parameter model that achieves an annotator agreement within 3% of Qwen2.5-32B-Instruct. With nothing more than SQL-style filters, we obtain competitive web-curated datasets in math (-8.0% relative to SOTA), web code (+14.3%), STEM (+24.5%) and medical (+8.6%). Essential-Web v1.0 is available on HuggingFace: https://huggingface.co/datasets/EssentialAI/essential-web-v1.0", "date": "2025-06-17", "categories": []}
{"id": "2506.16393", "title": "From LLM-anation to LLM-orchestrator: Coordinating Small Models for Data Labeling", "abstract": "Although the annotation paradigm based on Large Language Models (LLMs) has made significant breakthroughs in recent years, its actual deployment still has two core bottlenecks: first, the cost of calling commercial APIs in large-scale annotation is very expensive; second, in scenarios that require fine-grained semantic understanding, such as sentiment classification and toxicity classification, the annotation accuracy of LLMs is even lower than that of Small Language Models (SLMs) dedicated to this field. To address these problems, we propose a new paradigm of multi-model cooperative annotation and design a fully automatic annotation framework AutoAnnotator based on this. Specifically, AutoAnnotator consists of two layers. The upper-level meta-controller layer uses the generation and reasoning capabilities of LLMs to select SLMs for annotation, automatically generate annotation code and verify difficult samples; the lower-level task-specialist layer consists of multiple SLMs that perform annotation through multi-model voting. In addition, we use the difficult samples obtained by the secondary review of the meta-controller layer as the reinforcement learning set and fine-tune the SLMs in stages through a continual learning strategy, thereby improving the generalization of SLMs. Extensive experiments show that AutoAnnotator outperforms existing open-source/API LLMs in zero-shot, one-shot, CoT, and majority voting settings. Notably, AutoAnnotator reduces the annotation cost by 74.15% compared to directly annotating with GPT-3.5-turbo, while still improving the accuracy by 6.21%. Project page: https://github.com/Zhaiyuan-Ji/AutoAnnotator.", "date": "2025-06-19", "categories": []}
{"id": "2506.15365", "title": "FedWSIDD: Federated Whole Slide Image Classification via Dataset Distillation", "abstract": "Federated learning (FL) has emerged as a promising approach for collaborative medical image analysis, enabling multiple institutions to build robust predictive models while preserving sensitive patient data. In the context of Whole Slide Image (WSI) classification, FL faces significant challenges, including heterogeneous computational resources across participating medical institutes and privacy concerns. To address these challenges, we propose FedWSIDD, a novel FL paradigm that leverages dataset distillation (DD) to learn and transmit synthetic slides. On the server side, FedWSIDD aggregates synthetic slides from participating centres and distributes them across all centres. On the client side, we introduce a novel DD algorithm tailored to histopathology datasets which incorporates stain normalisation into the distillation process to generate a compact set of highly informative synthetic slides. These synthetic slides, rather than model parameters, are transmitted to the server. After communication, the received synthetic slides are combined with original slides for local tasks. Extensive experiments on multiple WSI classification tasks, including CAMELYON16 and CAMELYON17, demonstrate that FedWSIDD offers flexibility for heterogeneous local models, enhances local WSI classification performance, and preserves patient privacy. This makes it a highly effective solution for complex WSI classification tasks. The code is available at FedWSIDD.", "date": "2025-06-18", "categories": []}
{"id": "2506.12619", "title": "Semivalue-based data valuation is arbitrary and gameable", "abstract": "The game-theoretic notion of the semivalue offers a popular framework for credit attribution and data valuation in machine learning. Semivalues have been proposed for a variety of high-stakes decisions involving data, such as determining contributor compensation, acquiring data from external sources, or filtering out low-value datapoints. In these applications, semivalues depend on the specification of a utility function that maps subsets of data to a scalar score. While it is broadly agreed that this utility function arises from a composition of a learning algorithm and a performance metric, its actual instantiation involves numerous subtle modeling choices. We argue that this underspecification leads to varying degrees of arbitrariness in semivalue-based valuations. Small, but arguably reasonable changes to the utility function can induce substantial shifts in valuations across datapoints. Moreover, these valuation methodologies are also often gameable: low-cost adversarial strategies exist to exploit this ambiguity and systematically redistribute value among datapoints. Through theoretical constructions and empirical examples, we demonstrate that a bad-faith valuator can manipulate utility specifications to favor preferred datapoints, and that a good-faith valuator is left without principled guidance to justify any particular specification. These vulnerabilities raise ethical and epistemic concerns about the use of semivalues in several applications. We conclude by highlighting the burden of justification that semivalue-based approaches place on modelers and discuss important considerations for identifying appropriate uses.", "date": "2025-06-14", "categories": []}
{"id": "2502.06905", "title": "Lightweight Dataset Pruning without Full Training via Example Difficulty and Prediction Uncertainty", "abstract": "Recent advances in deep learning rely heavily on massive datasets, leading to substantial storage and training costs. Dataset pruning aims to alleviate this demand by discarding redundant examples. However, many existing methods require training a model with a full dataset over a large number of epochs before being able to prune the dataset, which ironically makes the pruning process more expensive than just training the model on the entire dataset. To overcome this limitation, we introduce a Difficulty and Uncertainty-Aware Lightweight (DUAL) score, which aims to identify important samples from the early training stage by considering both example difficulty and prediction uncertainty. To address a catastrophic accuracy drop at an extreme pruning, we further propose a ratio-adaptive sampling using Beta distribution. Experiments on various datasets and learning scenarios such as image classification with label noise and image corruption, and model architecture generalization demonstrate the superiority of our method over previous state-of-the-art (SOTA) approaches. Specifically, on ImageNet-1k, our method reduces the time cost for pruning to 66% compared to previous methods while achieving a SOTA, specifically 60% test accuracy at a 90% pruning ratio. On CIFAR datasets, the time cost is reduced to just 15% while maintaining SOTA performance.", "date": "2025-02-10", "categories": []}
{"id": "2502.08211", "title": "Quality over Quantity: Boosting Data Efficiency Through Ensembled Multimodal Data Curation", "abstract": "In an era overwhelmed by vast amounts of data, the effective curation of web-crawl datasets is essential for optimizing model performance. This paper tackles the challenges associated with the unstructured and heterogeneous nature of such datasets. Traditional heuristic curation methods often inadequately capture complex features, resulting in biases and the exclusion of relevant data. We introduce an advanced, learning-driven approach, Ensemble Curation Of DAta ThroUgh Multimodal Operators (EcoDatum), incorporating a novel quality-guided deduplication method to ensure balanced feature distributions. EcoDatum strategically integrates various unimodal and multimodal data curation operators within a weak supervision ensemble framework, utilizing automated optimization to score each data point effectively. EcoDatum, which significantly improves the data curation quality and efficiency, outperforms existing state-of-the-art (SOTA) techniques, ranked 1st on the DataComp leaderboard, with an average performance score of 0.182 across 38 diverse evaluation datasets. This represents a 28% improvement over the DataComp baseline method, demonstrating its effectiveness in improving dataset curation and model training efficiency.", "date": "2025-02-12", "categories": []}
{"id": "2506.10952", "title": "Domain2Vec: Vectorizing Datasets to Find the Optimal Data Mixture without Training", "abstract": "We introduce~\\textsc{Domain2Vec}, a novel approach that decomposes any dataset into a linear combination of several \\emph{meta-domains}, a new concept designed to capture the key underlying features of datasets. \\textsc{Domain2Vec} maintains a vocabulary of meta-domains and uses a classifier to decompose any given dataset into a domain vector that corresponds to a distribution over this vocabulary. These domain vectors enable the identification of the optimal data mixture for language model (LM) pretraining in a training-free manner under the \\emph{\\textbf{D}istribution \\textbf{A}lignment \\textbf{A}ssumption} (DA$^{2}$), which suggests that when the data distributions of the training set and the validation set are better aligned, a lower validation loss is achieved. Moreover, \\textsc{Domain2vec} can be seamlessly integrated into previous works to model the relationship between domain vectors and LM performance, greatly enhancing the efficiency and scalability of previous methods. Extensive experiments demonstrate that \\textsc{Domain2Vec} helps find the data mixture that enhances downstream task performance with minimal computational overhead. Specifically, \\textsc{Domain2Vec} achieves the same validation loss on Pile-CC using only $51.5\\%$ of the computation required when training on the original mixture of The Pile dataset. Under equivalent compute budget, \\textsc{Domain2Vec} improves downstream performance by an average of $2.83\\%$.", "date": "2025-06-12", "categories": []}
{"id": "2302.07944", "title": "Effective Data Augmentation With Diffusion Models", "abstract": "Data augmentation is one of the most prevalent tools in deep learning, underpinning many recent advances, including those from classification, generative models, and representation learning. The standard approach to data augmentation combines simple transformations like rotations and flips to generate new images from existing ones. However, these new images lack diversity along key semantic axes present in the data. Current augmentations cannot alter the high-level semantic attributes, such as animal species present in a scene, to enhance the diversity of data. We address the lack of diversity in data augmentation with image-to-image transformations parameterized by pre-trained text-to-image diffusion models. Our method edits images to change their semantics using an off-the-shelf diffusion model, and generalizes to novel visual concepts from a few labelled examples. We evaluate our approach on few-shot image classification tasks, and on a real-world weed recognition task, and observe an improvement in accuracy in tested domains.", "date": "2023-02-07", "categories": []}
{"id": "2406.11011", "title": "Data Shapley in One Training Run", "abstract": "Data Shapley provides a principled framework for attributing data's contribution within machine learning contexts. However, existing approaches require re-training models on different data subsets, which is computationally intensive, foreclosing their application to large-scale models. Furthermore, they produce the same attribution score for any models produced by running the learning algorithm, meaning they cannot perform targeted attribution towards a specific model obtained from a single run of the algorithm. This paper introduces In-Run Data Shapley, which addresses these limitations by offering scalable data attribution for a target model of interest. In its most efficient implementation, our technique incurs negligible additional runtime compared to standard model training. This dramatic efficiency improvement makes it possible to perform data attribution for the foundation model pretraining stage for the first time. We present several case studies that offer fresh insights into pretraining data's contribution and discuss their implications for copyright in generative AI and pretraining data curation.", "date": "2024-06-16", "categories": []}
{"id": "2408.12483", "title": "Not All Samples Should Be Utilized Equally: Towards Understanding and Improving Dataset Distillation", "abstract": "Dataset Distillation (DD) aims to synthesize a small dataset capable of performing comparably to the original dataset. Despite the success of numerous DD methods, theoretical exploration of this area remains unaddressed. In this paper, we take an initial step towards understanding various matching-based DD methods from the perspective of sample difficulty. We begin by empirically examining sample difficulty, measured by gradient norm, and observe that different matching-based methods roughly correspond to specific difficulty tendencies. We then extend the neural scaling laws of data pruning to DD to theoretically explain these matching-based methods. Our findings suggest that prioritizing the synthesis of easier samples from the original dataset can enhance the quality of distilled datasets, especially in low IPC (image-per-class) settings. Based on our empirical observations and theoretical analysis, we introduce the Sample Difficulty Correction (SDC) approach, designed to predominantly generate easier samples to achieve higher dataset quality. Our SDC can be seamlessly integrated into existing methods as a plugin with minimal code adjustments. Experimental results demonstrate that adding SDC generates higher-quality distilled datasets across 7 distillation methods and 6 datasets.", "date": "2024-08-22", "categories": []}
{"id": "2411.19946", "title": "DELT: A Simple Diversity-driven EarlyLate Training for Dataset Distillation", "abstract": "Recent advances in dataset distillation have led to solutions in two main directions. The conventional batch-to-batch matching mechanism is ideal for small-scale datasets and includes bi-level optimization methods on models and syntheses, such as FRePo, RCIG, and RaT-BPTT, as well as other methods like distribution matching, gradient matching, and weight trajectory matching. Conversely, batch-to-global matching typifies decoupled methods, which are particularly advantageous for large-scale datasets. This approach has garnered substantial interest within the community, as seen in SRe$^2$L, G-VBSM, WMDD, and CDA. A primary challenge with the second approach is the lack of diversity among syntheses within each class since samples are optimized independently and the same global supervision signals are reused across different synthetic images. In this study, we propose a new Diversity-driven EarlyLate Training (DELT) scheme to enhance the diversity of images in batch-to-global matching with less computation. Our approach is conceptually simple yet effective, it partitions predefined IPC samples into smaller subtasks and employs local optimizations to distill each subset into distributions from distinct phases, reducing the uniformity induced by the unified optimization process. These distilled images from the subtasks demonstrate effective generalization when applied to the entire task. We conduct extensive experiments on CIFAR, Tiny-ImageNet, ImageNet-1K, and its sub-datasets. Our approach outperforms the previous state-of-the-art by 2$\\sim$5% on average across different datasets and IPCs (images per class), increasing diversity per class by more than 5% while reducing synthesis time by up to 39.3% for enhancing the training efficiency. Code is available at: https://github.com/VILA-Lab/DELT.", "date": "2024-11-29", "categories": []}
{"id": "2506.06853", "title": "Curvature Enhanced Data Augmentation for Regression", "abstract": "Deep learning models with a large number of parameters, often referred to as over-parameterized models, have achieved exceptional performance across various tasks. Despite concerns about overfitting, these models frequently generalize well to unseen data, thanks to effective regularization techniques, with data augmentation being among the most widely used. While data augmentation has shown great success in classification tasks using label-preserving transformations, its application in regression problems has received less attention. Recently, a novel \\emph{manifold learning} approach for generating synthetic data was proposed, utilizing a first-order approximation of the data manifold. Building on this foundation, we present a theoretical framework and practical tools for approximating and sampling general data manifolds. Furthermore, we introduce the Curvature-Enhanced Manifold Sampling (CEMS) method for regression tasks. CEMS leverages a second-order representation of the data manifold to enable efficient sampling and reconstruction of new data points. Extensive evaluations across multiple datasets and comparisons with state-of-the-art methods demonstrate that CEMS delivers superior performance in both in-distribution and out-of-distribution scenarios, while introducing only minimal computational overhead. Code is available at https://github.com/azencot-group/CEMS.", "date": "2025-06-07", "categories": []}
{"id": "2407.11784", "title": "Data-Juicer Sandbox: A Feedback-Driven Suite for Multimodal Data-Model Co-development", "abstract": "The emergence of multimodal large models has advanced artificial intelligence, introducing unprecedented levels of performance and functionality. However, optimizing these models remains challenging due to historically isolated paths of model-centric and data-centric developments, leading to suboptimal outcomes and inefficient resource utilization. In response, we present a new sandbox suite tailored for integrated data-model co-development. This sandbox provides a feedback-driven experimental platform, enabling cost-effective iteration and guided refinement of both data and models. Our proposed ``Probe-Analyze-Refine'' workflow, validated through practical use cases on multimodal tasks such as image-text pre-training with CLIP, image-to-text generation with LLaVA-like models, and text-to-video generation with DiT-based models, yields transferable and notable performance boosts, such as topping the VBench leaderboard. A comprehensive set of over 100 experiments demonstrated the suite's usability and extensibility, while also uncovering insights into the interplay between data quality, diversity, model behavior, and computational costs. All codes, datasets, and models are open-sourced to foster future research and applications that would otherwise be infeasible due to the lack of a dedicated co-development infrastructure.", "date": "2024-07-16", "categories": []}
{"id": "2412.10032", "title": "Single-Pass Object-Focused Data Selection", "abstract": "While unlabeled image data is often plentiful, the costs of high-quality labels pose an important practical challenge: Which images should one select for labeling to use the annotation budget for a particular target task most effectively? To address this problem, we focus on single-pass data selection, which refers to the process of selecting all data to be annotated at once before training a downstream model. Prior methods for single-pass data selection rely on image-level representations and fail to reliably outperform random selection for object detection and segmentation. We propose Object-Focused Data Selection (OFDS) which leverages object-level features from foundation models and ensures semantic coverage of all target classes. In extensive experiments across tasks and target domains, OFDS consistently outperforms random selection and all baselines. The best results for constrained annotation budgets are obtained by combining human labels from OFDS with autolabels from foundation models. Moreover, using OFDS to select the initial labeled set for active learning yields consistent improvements", "date": "2024-12-13", "categories": []}
{"id": "2502.16733", "title": "Coreset Selection via LLM-based Concept Bottlenecks", "abstract": "Coreset Selection (CS) aims to identify a subset of the training dataset that achieves model performance comparable to using the entire dataset. Many state-of-the-art CS methods select coresets using scores whose computation requires training the downstream model on the entire dataset first and recording changes in the model's behavior on samples as it trains (training dynamics). These scores are inefficient to compute and hard to interpret, as they do not indicate whether a sample is difficult to learn in general or only for a specific downstream model. Our work addresses these challenges by proposing a score that computes a sample's difficulty using human-understandable textual attributes (concepts) independent of any downstream model. Specifically, we measure the alignment between a sample's visual features and concept bottlenecks, derived via large language models, by training a linear concept bottleneck layer and computing the sample's difficulty score using it.We then use stratified sampling based on this score to generate a coreset of the dataset.Crucially, our score is efficiently computable without training the downstream model on the full dataset even once, leads to high-performing coresets for various downstream models, and is computable even for an unlabeled dataset. Through experiments on CIFAR-10/100, and ImageNet-1K, we show that our coresets outperform random subsets, even at high pruning rates, and achieve model performance comparable to or better than coresets found by training dynamics-based methods.", "date": "2025-02-23", "categories": []}
{"id": "2505.22813", "title": "X-Factor: Quality Is a Dataset-Intrinsic Property", "abstract": "In the universal quest to optimize machine-learning classifiers, three factors -- model architecture, dataset size, and class balance -- have been shown to influence test-time performance but do not fully account for it. Previously, evidence was presented for an additional factor that can be referred to as dataset quality, but it was unclear whether this was actually a joint property of the dataset and the model architecture, or an intrinsic property of the dataset itself. If quality is truly dataset-intrinsic and independent of model architecture, dataset size, and class balance, then the same datasets should perform better (or worse) regardless of these other factors. To test this hypothesis, here we create thousands of datasets, each controlled for size and class balance, and use them to train classifiers with a wide range of architectures, from random forests and support-vector machines to deep networks. We find that classifier performance correlates strongly by subset across architectures ($R^2=0.79$), supporting quality as an intrinsic property of datasets independent of dataset size and class balance and of model architecture. Digging deeper, we find that dataset quality appears to be an emergent property of something more fundamental: the quality of datasets' constituent classes. Thus, quality joins size, class balance, and model architecture as an independent correlate of performance and a separate target for optimizing machine-learning-based classification.", "date": "2025-05-28", "categories": []}
{"id": "2506.04026", "title": "On the Usage of Gaussian Process for Efficient Data Valuation", "abstract": "In machine learning, knowing the impact of a given datum on model training is a fundamental task referred to as Data Valuation. Building on previous works from the literature, we have designed a novel canonical decomposition allowing practitioners to analyze any data valuation method as the combination of two parts: a utility function that captures characteristics from a given model and an aggregation procedure that merges such information. We also propose to use Gaussian Processes as a means to easily access the utility function on ``sub-models'', which are models trained on a subset of the training set. The strength of our approach stems from both its theoretical grounding in Bayesian theory, and its practical reach, by enabling fast estimation of valuations thanks to efficient update formulae.", "date": "2025-06-04", "categories": []}
{"id": "2504.00589", "title": "Efficient Annotator Reliability Assessment with EffiARA", "abstract": "Data annotation is an essential component of the machine learning pipeline; it is also a costly and time-consuming process. With the introduction of transformer-based models, annotation at the document level is increasingly popular; however, there is no standard framework for structuring such tasks. The EffiARA annotation framework is, to our knowledge, the first project to support the whole annotation pipeline, from understanding the resources required for an annotation task to compiling the annotated dataset and gaining insights into the reliability of individual annotators as well as the dataset as a whole. The framework's efficacy is supported by two previous studies: one improving classification performance through annotator-reliability-based soft-label aggregation and sample weighting, and the other increasing the overall agreement among annotators through removing identifying and replacing an unreliable annotator. This work introduces the EffiARA Python package and its accompanying webtool, which provides an accessible graphical user interface for the system. We open-source the EffiARA Python package at https://github.com/MiniEggz/EffiARA and the webtool is publicly accessible at https://effiara.gate.ac.uk.", "date": "2025-04-01", "categories": []}
{"id": "2504.13077", "title": "Effective Dual-Region Augmentation for Reduced Reliance on Large Amounts of Labeled Data", "abstract": "This paper introduces a novel dual-region augmentation approach designed to reduce reliance on large-scale labeled datasets while improving model robustness and adaptability across diverse computer vision tasks, including source-free domain adaptation (SFDA) and person re-identification (ReID). Our method performs targeted data transformations by applying random noise perturbations to foreground objects and spatially shuffling background patches. This effectively increases the diversity of the training data, improving model robustness and generalization. Evaluations on the PACS dataset for SFDA demonstrate that our augmentation strategy consistently outperforms existing methods, achieving significant accuracy improvements in both single-target and multi-target adaptation settings. By augmenting training data through structured transformations, our method enables model generalization across domains, providing a scalable solution for reducing reliance on manually annotated datasets. Furthermore, experiments on Market-1501 and DukeMTMC-reID datasets validate the effectiveness of our approach for person ReID, surpassing traditional augmentation techniques. The code is available at https://github.com/PrasannaPulakurthi/Foreground-Background-Augmentation", "date": "2025-04-17", "categories": []}
{"id": "2506.01789", "title": "Datasheets Aren't Enough: DataRubrics for Automated Quality Metrics and Accountability", "abstract": "High-quality datasets are fundamental to training and evaluating machine learning models, yet their creation-especially with accurate human annotations-remains a significant challenge. Many dataset paper submissions lack originality, diversity, or rigorous quality control, and these shortcomings are often overlooked during peer review. Submissions also frequently omit essential details about dataset construction and properties. While existing tools such as datasheets aim to promote transparency, they are largely descriptive and do not provide standardized, measurable methods for evaluating data quality. Similarly, metadata requirements at conferences promote accountability but are inconsistently enforced. To address these limitations, this position paper advocates for the integration of systematic, rubric-based evaluation metrics into the dataset review process-particularly as submission volumes continue to grow. We also explore scalable, cost-effective methods for synthetic data generation, including dedicated tools and LLM-as-a-judge approaches, to support more efficient evaluation. As a call to action, we introduce DataRubrics, a structured framework for assessing the quality of both human- and model-generated datasets. Leveraging recent advances in LLM-based evaluation, DataRubrics offers a reproducible, scalable, and actionable solution for dataset quality assessment, enabling both authors and reviewers to uphold higher standards in data-centric research. We also release code to support reproducibility of LLM-based evaluations at https://github.com/datarubrics/datarubrics.", "date": "2025-06-02", "categories": []}
{"id": "2506.01987", "title": "Equally Critical: Samples, Targets, and Their Mappings in Datasets", "abstract": "Data inherently possesses dual attributes: samples and targets. For targets, knowledge distillation has been widely employed to accelerate model convergence, primarily relying on teacher-generated soft target supervision. Conversely, recent advancements in data-efficient learning have emphasized sample optimization techniques, such as dataset distillation, while neglected the critical role of target. This dichotomy motivates our investigation into understanding how both sample and target collectively influence training dynamic. To address this gap, we first establish a taxonomy of existing paradigms through the lens of sample-target interactions, categorizing them into distinct sample-to-target mapping strategies. Building upon this foundation, we then propose a novel unified loss framework to assess their impact on training efficiency. Through extensive empirical studies on our proposed strategies, we comprehensively analyze how variations in target and sample types, quantities, and qualities influence model training, providing six key insights to enhance training efficacy.", "date": "2025-05-17", "categories": []}
{"id": "2506.02021", "title": "Dynamic-Aware Video Distillation: Optimizing Temporal Resolution Based on Video Semantics", "abstract": "With the rapid development of vision tasks and the scaling on datasets and models, redundancy reduction in vision datasets has become a key area of research. To address this issue, dataset distillation (DD) has emerged as a promising approach to generating highly compact synthetic datasets with significantly less redundancy while preserving essential information. However, while DD has been extensively studied for image datasets, DD on video datasets remains underexplored. Video datasets present unique challenges due to the presence of temporal information and varying levels of redundancy across different classes. Existing DD approaches assume a uniform level of temporal redundancy across all different video semantics, which limits their effectiveness on video datasets. In this work, we propose Dynamic-Aware Video Distillation (DAViD), a Reinforcement Learning (RL) approach to predict the optimal Temporal Resolution of the synthetic videos. A teacher-in-the-loop reward function is proposed to update the RL agent policy. To the best of our knowledge, this is the first study to introduce adaptive temporal resolution based on video semantics in video dataset distillation. Our approach significantly outperforms existing DD methods, demonstrating substantial improvements in performance. This work paves the way for future research on more efficient and semantic-adaptive video dataset distillation research.", "date": "2025-05-28", "categories": []}
{"id": "2311.16646", "title": "Rethinking Backdoor Attacks on Dataset Distillation: A Kernel Method Perspective", "abstract": "Dataset distillation offers a potential means to enhance data efficiency in deep learning. Recent studies have shown its ability to counteract backdoor risks present in original training samples. In this study, we delve into the theoretical aspects of backdoor attacks and dataset distillation based on kernel methods. We introduce two new theory-driven trigger pattern generation methods specialized for dataset distillation. Following a comprehensive set of analyses and experiments, we show that our optimization-based trigger design framework informs effective backdoor attacks on dataset distillation. Notably, datasets poisoned by our designed trigger prove resilient against conventional backdoor attack detection and mitigation methods. Our empirical results validate that the triggers developed using our approaches are proficient at executing resilient backdoor attacks.", "date": "2023-11-28", "categories": []}
{"id": "2412.02595", "title": "Nemotron-CC: Transforming Common Crawl into a Refined Long-Horizon Pretraining Dataset", "abstract": "Recent English Common Crawl datasets like FineWeb-Edu and DCLM achieved significant benchmark gains via aggressive model-based filtering, but at the cost of removing 90% of data. This limits their suitability for long token horizon training, such as 15T tokens for Llama 3.1. In this paper, we show how to achieve better trade-offs between accuracy and data quantity by a combination of classifier ensembling, synthetic data rephrasing, and reduced reliance on heuristic filters. When training 8B parameter models for 1T tokens, using a high-quality subset of our data improves MMLU by 5.6 over DCLM, demonstrating the efficacy of our methods for boosting accuracies over a relatively short token horizon. Furthermore, our full 6.3T token dataset matches DCLM on MMLU, but contains four times more unique real tokens than DCLM. This unlocks state-of-the-art training over a long token horizon: an 8B parameter model trained for 15T tokens, of which 7.2T came from our dataset, is better than the Llama 3.1 8B model: +5 on MMLU, +3.1 on ARC-Challenge, and +0.5 on average across ten diverse tasks. The dataset is available at https://data.commoncrawl.org/contrib/Nemotron/Nemotron-CC/index.html", "date": "2024-12-03", "categories": []}
{"id": "2502.17184", "title": "Measuring Data Diversity for Instruction Tuning: A Systematic Analysis and A Reliable Metric", "abstract": "Data diversity is crucial for the instruction tuning of large language models. Existing studies have explored various diversity-aware data selection methods to construct high-quality datasets and enhance model performance. However, the fundamental problem of precisely defining and measuring data diversity remains underexplored, limiting clear guidance for data engineering. To address this, we systematically analyze 11 existing diversity measurement methods by evaluating their correlation with model performance through extensive fine-tuning experiments. Our results indicate that a reliable diversity measure should properly account for both inter-sample differences and the information density in the sample space. Building on this, we propose NovelSum, a new diversity metric based on sample-level \"novelty.\" Experiments on both simulated and real-world data show that NovelSum accurately captures diversity variations and achieves a 0.97 correlation with instruction-tuned model performance, highlighting its value in guiding data engineering practices. With NovelSum as an optimization objective, we further develop a greedy, diversity-oriented data selection strategy that outperforms existing approaches, validating both the effectiveness and practical significance of our metric. The code is available at https://github.com/UmeanNever/NovelSum.", "date": "2025-02-24", "categories": []}
{"id": "2505.22232", "title": "Judging Quality Across Languages: A Multilingual Approach to Pretraining Data Filtering with Language Models", "abstract": "High-quality multilingual training data is essential for effectively pretraining large language models (LLMs). Yet, the availability of suitable open-source multilingual datasets remains limited. Existing state-of-the-art datasets mostly rely on heuristic filtering methods, restricting both their cross-lingual transferability and scalability. Here, we introduce JQL, a systematic approach that efficiently curates diverse and high-quality multilingual data at scale while significantly reducing computational demands. JQL distills LLMs' annotation capabilities into lightweight annotators based on pretrained multilingual embeddings. These models exhibit robust multilingual and cross-lingual performance, even for languages and scripts unseen during training. Evaluated empirically across 35 languages, the resulting annotation pipeline substantially outperforms current heuristic filtering methods like Fineweb2. JQL notably enhances downstream model training quality and increases data retention rates. Our research provides practical insights and valuable resources for multilingual data curation, raising the standards of multilingual dataset development.", "date": "2025-05-28", "categories": []}
{"id": "2506.01586", "title": "Multi-Modal Dataset Distillation in the Wild", "abstract": "Recent multi-modal models have shown remarkable versatility in real-world applications. However, their rapid development encounters two critical data challenges. First, the training process requires large-scale datasets, leading to substantial storage and computational costs. Second, these data are typically web-crawled with inevitable noise, i.e., partially mismatched pairs, severely degrading model performance. To these ends, we propose Multi-modal dataset Distillation in the Wild, i.e., MDW, the first framework to distill noisy multi-modal datasets into compact clean ones for effective and efficient model training. Specifically, MDW introduces learnable fine-grained correspondences during distillation and adaptively optimizes distilled data to emphasize correspondence-discriminative regions, thereby enhancing distilled data's information density and efficacy. Moreover, to capture robust cross-modal correspondence prior knowledge from real data, MDW proposes dual-track collaborative learning to avoid the risky data noise, alleviating information loss with certifiable noise tolerance. Extensive experiments validate MDW's theoretical and empirical efficacy with remarkable scalability, surpassing prior methods by over 15% across various compression ratios, highlighting its appealing practicality for applications with diverse efficacy and resource needs.", "date": "2025-06-02", "categories": []}
{"id": "2505.24165", "title": "Tag-Evol: Achieving Efficient Instruction Evolving via Tag Injection", "abstract": "Evol-Instruct has made significant improvements as a data synthesis method in several areas. Existing methods typically rely on a fixed set of strategies to evolve, which require manual design and are monolithic in form. In addition, iterative evolution also makes the acquisition of hard samples expensive. In view of this, we propose the Tag-Evol framework, a more diverse and efficient instruction evolving method. Specifically, Tag-Evol uses diverse and specific knowledge tags as strategies to achieve controlled evolution by injecting different combinations of tags into the original instructions. Experiments with multiple backbones in diverse domain benchmarks show that the proposed method generates significantly better evolved data than other methods. Furthermore, we conduct a thorough analysis of the evolved data, demonstrating that Tag-Evol is not only efficient but also generates more diverse and challenging data.", "date": "2025-05-30", "categories": []}
{"id": "2405.18253", "title": "Proper Dataset Valuation by Pointwise Mutual Information", "abstract": "Data plays a central role in advancements in modern artificial intelligence, with high-quality data emerging as a key driver of model performance. This has prompted the development of principled and effective data curation methods in recent years. However, existing methods largely rely on heuristics, and whether they are truly effective remains unclear. For instance, standard evaluation methods that assess a trained model's performance on specific benchmarks may incentivize assigning high scores to data that merely resembles the test set. This issue exemplifies Goodhart's law: when a measure becomes a target, it ceases to be a good measure. To address this issue, we propose an information-theoretic framework for evaluating data curation methods. We define dataset quality in terms of its informativeness about the true model parameters, formalized using the Blackwell ordering of informativeness. Under this ordering, Blackwell's theorem ensures that more informative data yields optimal models with lower expected loss on the true underlying distribution. To measure informativeness, we show that the Blackwell order can be determined by the Shannon mutual information between the curated data and the test data. To estimate this mutual information, we introduce a novel method that trains Bayesian models on embedded datasets and computes mutual information from the posteriors of model parameters. Experiments on real-world data demonstrate that our mutual information-based evaluation assigns appropriately lower scores to data curation strategies that reduce dataset informativeness, while traditional test score-based evaluation methods may favor data curation strategies that overfit to the test set but compromise the training data's informativeness.", "date": "2024-05-28", "categories": []}
{"id": "2505.23223", "title": "Daunce: Data Attribution through Uncertainty Estimation", "abstract": "Training data attribution (TDA) methods aim to identify which training examples influence a model's predictions on specific test data most. By quantifying these influences, TDA supports critical applications such as data debugging, curation, and valuation. Gradient-based TDA methods rely on gradients and second-order information, limiting their applicability at scale. While recent random projection-based methods improve scalability, they often suffer from degraded attribution accuracy. Motivated by connections between uncertainty and influence functions, we introduce Daunce - a simple yet effective data attribution approach through uncertainty estimation. Our method operates by fine-tuning a collection of perturbed models and computing the covariance of per-example losses across these models as the attribution score. Daunce is scalable to large language models (LLMs) and achieves more accurate attribution compared to existing TDA methods. We validate Daunce on tasks ranging from vision tasks to LLM fine-tuning, and further demonstrate its compatibility with black-box model access. Applied to OpenAI's GPT models, our method achieves, to our knowledge, the first instance of data attribution on proprietary LLMs.", "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": "2505.21692", "title": "What Data Enables Optimal Decisions? An Exact Characterization for Linear Optimization", "abstract": "We study the fundamental question of how informative a dataset is for solving a given decision-making task. In our setting, the dataset provides partial information about unknown parameters that influence task outcomes. Focusing on linear programs, we characterize when a dataset is sufficient to recover an optimal decision, given an uncertainty set on the cost vector. Our main contribution is a sharp geometric characterization that identifies the directions of the cost vector that matter for optimality, relative to the task constraints and uncertainty set. We further develop a practical algorithm that, for a given task, constructs a minimal or least-costly sufficient dataset. Our results reveal that small, well-chosen datasets can often fully determine optimal decisions -- offering a principled foundation for task-aware data selection.", "date": "2025-05-27", "categories": []}
{"id": "2505.21958", "title": "Resolving Knowledge Conflicts in Domain-specific Data Selection: A Case Study on Medical Instruction-tuning", "abstract": "Domain-specific instruction-tuning has become the defacto standard for improving the performance of large language models (LLMs) in specialized applications, e.g., medical question answering. Since the instruction-tuning dataset might contain redundant or low-quality data, data selection (DS) is usually required to maximize the data efficiency. Despite the successes in the general domain, current DS methods often struggle to select the desired data for domain-specific instruction-tuning. One of the main reasons is that they neglect the impact of knowledge conflicts, i.e., the discrepancy between LLMs' pretrained knowledge and context knowledge of instruction data, which could damage LLMs' prior abilities and lead to hallucination. To this end, we propose a simple-yet-effective Knowledge-aware Data Selection (namely KDS) framework to select the domain-specific instruction-tuning data that meets LLMs' actual needs. The core of KDS is to leverage two knowledge-aware metrics for quantitatively measuring knowledge conflicts from two aspects: context-memory knowledge alignment and intra-memory knowledge consistency. By filtering the data with large knowledge conflicts and sampling the high-quality and diverse data, KDS can effectively stimulate the LLMs' abilities and achieve better domain-specific performance. Taking the medical domain as the testbed, we conduct extensive experiments and empirically prove that KDS surpasses the other baselines and brings significant and consistent performance gains among all LLMs. More encouragingly, KDS effectively improves the model generalization and alleviates the hallucination problem.", "date": "2025-05-28", "categories": []}
{"id": "2505.22387", "title": "DAM: Domain-Aware Module for Multi-Domain Dataset Condensation", "abstract": "Dataset Condensation (DC) has emerged as a promising solution to mitigate the computational and storage burdens associated with training deep learning models. However, existing DC methods largely overlook the multi-domain nature of modern datasets, which are increasingly composed of heterogeneous images spanning multiple domains. In this paper, we extend DC and introduce Multi-Domain Dataset Condensation (MDDC), which aims to condense data that generalizes across both single-domain and multi-domain settings. To this end, we propose the Domain-Aware Module (DAM), a training-time module that embeds domain-related features into each synthetic image via learnable spatial masks. As explicit domain labels are mostly unavailable in real-world datasets, we employ frequency-based pseudo-domain labeling, which leverages low-frequency amplitude statistics. DAM is only active during the condensation process, thus preserving the same images per class (IPC) with prior methods. Experiments show that DAM consistently improves in-domain, out-of-domain, and cross-architecture performance over baseline dataset condensation methods.", "date": "2025-05-28", "categories": []}
{"id": "2410.12458", "title": "The Best of Both Worlds: Bridging Quality and Diversity in Data Selection with Bipartite Graph", "abstract": "The performance of large language models (LLMs) is strongly influenced by the quality and diversity of data used during supervised fine-tuning (SFT). However, current data selection methods often prioritize one aspect over the other, resulting in suboptimal training outcomes. To address this, we formulate data selection as a set cover problem and present GraphFilter, a novel approach that balances both quality and diversity in data selection. GraphFilter models the dataset as a bipartite graph connecting sentences to their constituent n-grams, then employs a priority function that combines quality and diversity metrics multiplicatively. GraphFilter iteratively selects sentences with the highest priority, removes covered n-grams from the bipartite graph, and recomputes priorities to reflect the changing data landscape. We validate GraphFilter using three model backbones across six widely-used benchmarks, demonstrating that it outperforms nine existing baselines in both model performance and computational efficiency. Further analysis shows that our design choices lead to more effective subset selection, underscores the value of instruction diversity, and provides insights into how quality and diversity interact with different subset sizes.", "date": "2024-10-16", "categories": []}
{"id": "2505.20380", "title": "GRAPE: Optimize Data Mixture for Group Robust Multi-target Adaptive Pretraining", "abstract": "The performance of large language models (LLMs) across diverse downstream applications is fundamentally governed by the quality and composition of their pretraining corpora. Existing domain reweighting algorithms primarily optimize data mixtures for a single target task, thereby resulting in models that overfit to specialized objectives while exhibiting substantial performance degradation on other benchmarks. This paper introduces Group Robust Multi-target Adaptive PrEtraining (GRAPE), a novel multi-source-multi-target domain reweighting framework designed to calibrate pretraining data mixtures for robust performance across multiple target tasks simultaneously. GRAPE dynamically adjusts sampling weights across source domains (domain weights) while concurrently modulating task weights that quantify the relative importance of each individual target task. This adaptive process prioritizes tasks based on their learning difficulty throughout training. We formulate this interleaved reweighting mechanism as a minimax optimization problem: The inner maximization adjusts task weights leveraging group distributed-robust-optimization (DRO), where those tasks demonstrating the least improvement under the current data mixture are prioritized with higher weights; The outer minimization then optimizes domain weights to maximize loss reduction on the prioritized tasks. Experiments on ClimbLab and SlimPajama datasets demonstrate that GRAPE consistently outperforms baseline methods in terms of reasoning performance across 6 benchmarks. Furthermore, when applied to multilingual targets, GRAPE effectively identifies optimal training mixtures from mainstream languages, achieving superior language modeling capabilities across 8 low-resource target languages.", "date": "2025-05-26", "categories": []}
{"id": "2505.20694", "title": "Temporal Saliency-Guided Distillation: A Scalable Framework for Distilling Video Datasets", "abstract": "Dataset distillation (DD) has emerged as a powerful paradigm for dataset compression, enabling the synthesis of compact surrogate datasets that approximate the training utility of large-scale ones. While significant progress has been achieved in distilling image datasets, extending DD to the video domain remains challenging due to the high dimensionality and temporal complexity inherent in video data. Existing video distillation (VD) methods often suffer from excessive computational costs and struggle to preserve temporal dynamics, as na\\\"ive extensions of image-based approaches typically lead to degraded performance. In this paper, we propose a novel uni-level video dataset distillation framework that directly optimizes synthetic videos with respect to a pre-trained model. To address temporal redundancy and enhance motion preservation, we introduce a temporal saliency-guided filtering mechanism that leverages inter-frame differences to guide the distillation process, encouraging the retention of informative temporal cues while suppressing frame-level redundancy. Extensive experiments on standard video benchmarks demonstrate that our method achieves state-of-the-art performance, bridging the gap between real and distilled video data and offering a scalable solution for video dataset compression.", "date": "2025-05-27", "categories": []}
{"id": "2505.20807", "title": "Simple yet Effective Graph Distillation via Clustering", "abstract": "Despite plentiful successes achieved by graph representation learning in various domains, the training of graph neural networks (GNNs) still remains tenaciously challenging due to the tremendous computational overhead needed for sizable graphs in practice. Recently, graph data distillation (GDD), which seeks to distill large graphs into compact and informative ones, has emerged as a promising technique to enable efficient GNN training. However, most existing GDD works rely on heuristics that align model gradients or representation distributions on condensed and original graphs, leading to compromised result quality, expensive training for distilling large graphs, or both. Motivated by this, this paper presents an efficient and effective GDD approach, ClustGDD. Under the hood, ClustGDD resorts to synthesizing the condensed graph and node attributes through fast and theoretically-grounded clustering that minimizes the within-cluster sum of squares and maximizes the homophily on the original graph. The fundamental idea is inspired by our empirical and theoretical findings unveiling the connection between clustering and empirical condensation quality using Fr\\'echet Inception Distance, a well-known quality metric for synthetic images. Furthermore, to mitigate the adverse effects caused by the homophily-based clustering, ClustGDD refines the nodal attributes of the condensed graph with a small augmentation learned via class-aware graph sampling and consistency loss. Our extensive experiments exhibit that GNNs trained over condensed graphs output by ClustGDD consistently achieve superior or comparable performance to state-of-the-art GDD methods in terms of node classification on five benchmark datasets, while being orders of magnitude faster.", "date": "2025-05-27", "categories": []}
{"id": "2403.01944", "title": "Fourier-basis Functions to Bridge Augmentation Gap: Rethinking Frequency Augmentation in Image Classification", "abstract": "Computer vision models normally witness degraded performance when deployed in real-world scenarios, due to unexpected changes in inputs that were not accounted for during training. Data augmentation is commonly used to address this issue, as it aims to increase data variety and reduce the distribution gap between training and test data. However, common visual augmentations might not guarantee extensive robustness of computer vision models. In this paper, we propose Auxiliary Fourier-basis Augmentation (AFA), a complementary technique targeting augmentation in the frequency domain and filling the augmentation gap left by visual augmentations. We demonstrate the utility of augmentation via Fourier-basis additive noise in a straightforward and efficient adversarial setting. Our results show that AFA benefits the robustness of models against common corruptions, OOD generalization, and consistency of performance of models against increasing perturbations, with negligible deficit to the standard performance of models. It can be seamlessly integrated with other augmentation techniques to further boost performance. Code and models can be found at: https://github.com/nis-research/afa-augment", "date": "2024-03-04", "categories": []}
{"id": "2502.00494", "title": "Data Overvaluation Attack and Truthful Data Valuation in Federated Learning", "abstract": "In collaborative machine learning (CML), data valuation, i.e., evaluating the contribution of each client's data to the machine learning model, has become a critical task for incentivizing and selecting positive data contributions. However, existing studies often assume that clients engage in data valuation truthfully, overlooking the practical motivation for clients to exaggerate their contributions. To unlock this threat, this paper introduces the data overvaluation attack, enabling strategic clients to have their data significantly overvalued in federated learning, a widely adopted paradigm for decentralized CML. Furthermore, we propose a Bayesian truthful data valuation metric, named Truth-Shapley. Truth-Shapley is the unique metric that guarantees some promising axioms for data valuation while ensuring that clients' optimal strategy is to perform truthful data valuation under certain conditions. Our experiments demonstrate the vulnerability of existing data valuation metrics to the proposed attack and validate the robustness and effectiveness of Truth-Shapley.", "date": "2025-02-01", "categories": []}
{"id": "2505.18358", "title": "CONCORD: Concept-Informed Diffusion for Dataset Distillation", "abstract": "Dataset distillation (DD) has witnessed significant progress in creating small datasets that encapsulate rich information from large original ones. Particularly, methods based on generative priors show promising performance, while maintaining computational efficiency and cross-architecture generalization. However, the generation process lacks explicit controllability for each sample. Previous distillation methods primarily match the real distribution from the perspective of the entire dataset, whereas overlooking concept completeness at the instance level. The missing or incorrectly represented object details cannot be efficiently compensated due to the constrained sample amount typical in DD settings. To this end, we propose incorporating the concept understanding of large language models (LLMs) to perform Concept-Informed Diffusion (CONCORD) for dataset distillation. Specifically, distinguishable and fine-grained concepts are retrieved based on category labels to inform the denoising process and refine essential object details. By integrating these concepts, the proposed method significantly enhances both the controllability and interpretability of the distilled image generation, without relying on pre-trained classifiers. We demonstrate the efficacy of CONCORD by achieving state-of-the-art performance on ImageNet-1K and its subsets. The code implementation is released in https://github.com/vimar-gu/CONCORD.", "date": "2025-05-23", "categories": []}
{"id": "2505.18399", "title": "Taming Diffusion for Dataset Distillation with High Representativeness", "abstract": "Recent deep learning models demand larger datasets, driving the need for dataset distillation to create compact, cost-efficient datasets while maintaining performance. Due to the powerful image generation capability of diffusion, it has been introduced to this field for generating distilled images. In this paper, we systematically investigate issues present in current diffusion-based dataset distillation methods, including inaccurate distribution matching, distribution deviation with random noise, and separate sampling. Building on this, we propose D^3HR, a novel diffusion-based framework to generate distilled datasets with high representativeness. Specifically, we adopt DDIM inversion to map the latents of the full dataset from a low-normality latent domain to a high-normality Gaussian domain, preserving information and ensuring structural consistency to generate representative latents for the distilled dataset. Furthermore, we propose an efficient sampling scheme to better align the representative latents with the high-normality Gaussian distribution. Our comprehensive experiments demonstrate that D^3HR can achieve higher accuracy across different model architectures compared with state-of-the-art baselines in dataset distillation. Source code: https://github.com/lin-zhao-resoLve/D3HR.", "date": "2025-05-23", "categories": []}
{"id": "2505.18963", "title": "MGD$^3$: Mode-Guided Dataset Distillation using Diffusion Models", "abstract": "Dataset distillation has emerged as an effective strategy, significantly reducing training costs and facilitating more efficient model deployment. Recent advances have leveraged generative models to distill datasets by capturing the underlying data distribution. Unfortunately, existing methods require model fine-tuning with distillation losses to encourage diversity and representativeness. However, these methods do not guarantee sample diversity, limiting their performance. We propose a mode-guided diffusion model leveraging a pre-trained diffusion model without the need to fine-tune with distillation losses. Our approach addresses dataset diversity in three stages: Mode Discovery to identify distinct data modes, Mode Guidance to enhance intra-class diversity, and Stop Guidance to mitigate artifacts in synthetic samples that affect performance. Our approach outperforms state-of-the-art methods, achieving accuracy gains of 4.4%, 2.9%, 1.6%, and 1.6% on ImageNette, ImageIDC, ImageNet-100, and ImageNet-1K, respectively. Our method eliminates the need for fine-tuning diffusion models with distillation losses, significantly reducing computational costs. Our code is available on the project webpage: https://jachansantiago.github.io/mode-guided-distillation/", "date": "2025-05-25", "categories": []}
{"id": "2505.19051", "title": "Efficient Data Selection at Scale via Influence Distillation", "abstract": "Effective data selection is critical for efficient training of modern Large Language Models (LLMs). This paper introduces Influence Distillation, a novel, mathematically-justified framework for data selection that employs second-order information to optimally weight training samples. By distilling each sample's influence on a target distribution, our method assigns model-specific weights that are used to select training data for LLM fine-tuning, guiding it toward strong performance on the target domain. We derive these optimal weights for both Gradient Descent and Adam optimizers. To ensure scalability and reduce computational cost, we propose a $\\textit{landmark-based approximation}$: influence is precisely computed for a small subset of \"landmark\" samples and then efficiently propagated to all other samples to determine their weights. We validate Influence Distillation by applying it to instruction tuning on the Tulu V2 dataset, targeting a range of tasks including GSM8k, SQuAD, and MMLU, across several models from the Llama and Qwen families. Experiments show that Influence Distillation matches or outperforms state-of-the-art performance while achieving up to $3.5\\times$ faster selection.", "date": "2025-05-25", "categories": []}
{"id": "2505.19469", "title": "Diversity-Driven Generative Dataset Distillation Based on Diffusion Model with Self-Adaptive Memory", "abstract": "Dataset distillation enables the training of deep neural networks with comparable performance in significantly reduced time by compressing large datasets into small and representative ones. Although the introduction of generative models has made great achievements in this field, the distributions of their distilled datasets are not diverse enough to represent the original ones, leading to a decrease in downstream validation accuracy. In this paper, we present a diversity-driven generative dataset distillation method based on a diffusion model to solve this problem. We introduce self-adaptive memory to align the distribution between distilled and real datasets, assessing the representativeness. The degree of alignment leads the diffusion model to generate more diverse datasets during the distillation process. Extensive experiments show that our method outperforms existing state-of-the-art methods in most situations, proving its ability to tackle dataset distillation tasks.", "date": "2025-05-26", "categories": []}
{"id": "2505.17245", "title": "Extending Dataset Pruning to Object Detection: A Variance-based Approach", "abstract": "Dataset pruning -- selecting a small yet informative subset of training data -- has emerged as a promising strategy for efficient machine learning, offering significant reductions in computational cost and storage compared to alternatives like dataset distillation. While pruning methods have shown strong performance in image classification, their extension to more complex computer vision tasks, particularly object detection, remains relatively underexplored. In this paper, we present the first principled extension of classification pruning techniques to the object detection domain, to the best of our knowledge. We identify and address three key challenges that hinder this transition: the Object-Level Attribution Problem, the Scoring Strategy Problem, and the Image-Level Aggregation Problem. To overcome these, we propose tailored solutions, including a novel scoring method called Variance-based Prediction Score (VPS). VPS leverages both Intersection over Union (IoU) and confidence scores to effectively identify informative training samples specific to detection tasks. Extensive experiments on PASCAL VOC and MS COCO demonstrate that our approach consistently outperforms prior dataset pruning methods in terms of mean Average Precision (mAP). We also show that annotation count and class distribution shift can influence detection performance, but selecting informative examples is a more critical factor than dataset size or balance. Our work bridges dataset pruning and object detection, paving the way for dataset pruning in complex vision tasks.", "date": "2025-05-22", "categories": []}
{"id": "2502.00761", "title": "FIRE: Flexible Integration of Data Quality Ratings for Effective Pre-Training", "abstract": "Selecting high-quality data can improve the pretraining efficiency of large language models (LLMs). Existing methods generally rely on heuristic techniques or single quality signals, limiting their ability to evaluate data quality comprehensively. In this work, we propose FIRE, a flexible and scalable framework for integrating multiple data quality raters, which allows for a comprehensive assessment of data quality across various dimensions. FIRE aligns multiple quality signals into a unified space, and integrates diverse data quality raters to provide a comprehensive quality signal for each data point. Further, we introduce a progressive data selection scheme based on FIRE that iteratively refines the selection of high-quality data points. Extensive experiments show that FIRE outperforms other data selection methods and significantly boosts pretrained model performance across a wide range of downstream tasks, while requiring less than 37.5\\% of the training data needed by the Random baseline to reach the target performance.", "date": "2025-02-02", "categories": []}
{"id": "2504.08364", "title": "DRIP: DRop unImportant data Points -- Enhancing Machine Learning Efficiency with Grad-CAM-Based Real-Time Data Prioritization for On-Device Training", "abstract": "Selecting data points for model training is critical in machine learning. Effective selection methods can reduce the labeling effort, optimize on-device training for embedded systems with limited data storage, and enhance the model performance. This paper introduces a novel algorithm that uses Grad-CAM to make online decisions about retaining or discarding data points. Optimized for embedded devices, the algorithm computes a unique DRIP Score to quantify the importance of each data point. This enables dynamic decision-making on whether a data point should be stored for potential retraining or discarded without compromising model performance. Experimental evaluations on four benchmark datasets demonstrate that our approach can match or even surpass the accuracy of models trained on the entire dataset, all while achieving storage savings of up to 39\\%. To our knowledge, this is the first algorithm that makes online decisions about data point retention without requiring access to the entire dataset.", "date": "2025-04-11", "categories": []}
{"id": "2505.15267", "title": "Contrastive Learning-Enhanced Trajectory Matching for Small-Scale Dataset Distillation", "abstract": "Deploying machine learning models in resource-constrained environments, such as edge devices or rapid prototyping scenarios, increasingly demands distillation of large datasets into significantly smaller yet informative synthetic datasets. Current dataset distillation techniques, particularly Trajectory Matching methods, optimize synthetic data so that the model's training trajectory on synthetic samples mirrors that on real data. While demonstrating efficacy on medium-scale synthetic datasets, these methods fail to adequately preserve semantic richness under extreme sample scarcity. To address this limitation, we propose a novel dataset distillation method integrating contrastive learning during image synthesis. By explicitly maximizing instance-level feature discrimination, our approach produces more informative and diverse synthetic samples, even when dataset sizes are significantly constrained. Experimental results demonstrate that incorporating contrastive learning substantially enhances the performance of models trained on very small-scale synthetic datasets. This integration not only guides more effective feature representation but also significantly improves the visual fidelity of the synthesized images. Experimental results demonstrate that our method achieves notable performance improvements over existing distillation techniques, especially in scenarios with extremely limited synthetic data.", "date": "2025-05-21", "categories": []}
{"id": "2407.08151", "title": "Enrich the content of the image Using Context-Aware Copy Paste", "abstract": "Data augmentation remains a widely utilized technique in deep learning, particularly in tasks such as image classification, semantic segmentation, and object detection. Among them, Copy-Paste is a simple yet effective method and gain great attention recently. However, existing Copy-Paste often overlook contextual relevance between source and target images, resulting in inconsistencies in generated outputs. To address this challenge, we propose a context-aware approach that integrates Bidirectional Latent Information Propagation (BLIP) for content extraction from source images. By matching extracted content information with category information, our method ensures cohesive integration of target objects using Segment Anything Model (SAM) and You Only Look Once (YOLO). This approach eliminates the need for manual annotation, offering an automated and user-friendly solution. Experimental evaluations across diverse datasets demonstrate the effectiveness of our method in enhancing data diversity and generating high-quality pseudo-images across various computer vision tasks.", "date": "2024-07-11", "categories": []}
{"id": "2505.14705", "title": "Beyond Modality Collapse: Representations Blending for Multimodal Dataset Distillation", "abstract": "Multimodal Dataset Distillation (MDD) seeks to condense large-scale image-text datasets into compact surrogates while retaining their effectiveness for cross-modal learning. Despite recent progress, existing MDD approaches often suffer from \\textit{\\textbf{Modality Collapse}}, characterized by over-concentrated intra-modal representations and enlarged distributional gap across modalities. In this paper, at the first time, we identify this issue as stemming from a fundamental conflict between the over-compression behavior inherent in dataset distillation and the cross-modal supervision imposed by contrastive objectives. To alleviate modality collapse, we introduce \\textbf{RepBlend}, a novel MDD framework that weakens overdominant cross-modal supervision via representation blending, thereby significantly enhancing intra-modal diversity. Additionally, we observe that current MDD methods impose asymmetric supervision across modalities, resulting in biased optimization. To address this, we propose symmetric projection trajectory matching, which synchronizes the optimization dynamics using modality-specific projection heads, thereby promoting balanced supervision and enhancing cross-modal alignment. Experiments on Flickr-30K and MS-COCO show that RepBlend consistently outperforms prior state-of-the-art MDD methods, achieving significant gains in retrieval performance (e.g., +9.4 IR@10, +6.3 TR@10 under the 100-pair setting) and offering up to 6.7$\\times$ distillation speedup.", "date": "2025-05-16", "categories": []}
{"id": "2505.15303", "title": "Laplace Sample Information: Data Informativeness Through a Bayesian Lens", "abstract": "Accurately estimating the informativeness of individual samples in a dataset is an important objective in deep learning, as it can guide sample selection, which can improve model efficiency and accuracy by removing redundant or potentially harmful samples. We propose Laplace Sample Information (LSI) measure of sample informativeness grounded in information theory widely applicable across model architectures and learning settings. LSI leverages a Bayesian approximation to the weight posterior and the KL divergence to measure the change in the parameter distribution induced by a sample of interest from the dataset. We experimentally show that LSI is effective in ordering the data with respect to typicality, detecting mislabeled samples, measuring class-wise informativeness, and assessing dataset difficulty. We demonstrate these capabilities of LSI on image and text data in supervised and unsupervised settings. Moreover, we show that LSI can be computed efficiently through probes and transfers well to the training of large models.", "date": "2025-05-21", "categories": []}
{"id": "2412.09945", "title": "Going Beyond Feature Similarity: Effective Dataset Distillation based on Class-Aware Conditional Mutual Information", "abstract": "Dataset distillation (DD) aims to minimize the time and memory consumption needed for training deep neural networks on large datasets, by creating a smaller synthetic dataset that has similar performance to that of the full real dataset. However, current dataset distillation methods often result in synthetic datasets that are excessively difficult for networks to learn from, due to the compression of a substantial amount of information from the original data through metrics measuring feature similarity, e,g., distribution matching (DM). In this work, we introduce conditional mutual information (CMI) to assess the class-aware complexity of a dataset and propose a novel method by minimizing CMI. Specifically, we minimize the distillation loss while constraining the class-aware complexity of the synthetic dataset by minimizing its empirical CMI from the feature space of pre-trained networks, simultaneously. Conducting on a thorough set of experiments, we show that our method can serve as a general regularization method to existing DD methods and improve the performance and training efficiency.", "date": "2024-12-13", "categories": []}
{"id": "2505.05327", "title": "RICo: Refined In-Context Contribution for Automatic Instruction-Tuning Data Selection", "abstract": "Data selection for instruction tuning is crucial for improving the performance of large language models (LLMs) while reducing training costs. In this paper, we propose Refined Contribution Measurement with In-Context Learning (RICo), a novel gradient-free method that quantifies the fine-grained contribution of individual samples to both task-level and global-level model performance. RICo enables more accurate identification of high-contribution data, leading to better instruction tuning. We further introduce a lightweight selection paradigm trained on RICo scores, enabling scalable data selection with a strictly linear inference complexity. Extensive experiments on three LLMs across 12 benchmarks and 5 pairwise evaluation sets demonstrate the effectiveness of RICo. Remarkably, on LLaMA3.1-8B, models trained on 15% of RICo-selected data outperform full datasets by 5.42% points and exceed the best performance of widely used selection methods by 2.06% points. We further analyze high-contribution samples selected by RICo, which show both diverse tasks and appropriate difficulty levels, rather than just the hardest ones.", "date": "2025-05-08", "categories": []}
{"id": "2505.06647", "title": "Dataset Distillation with Probabilistic Latent Features", "abstract": "As deep learning models grow in complexity and the volume of training data increases, reducing storage and computational costs becomes increasingly important. Dataset distillation addresses this challenge by synthesizing a compact set of synthetic data that can effectively replace the original dataset in downstream classification tasks. While existing methods typically rely on mapping data from pixel space to the latent space of a generative model, we propose a novel stochastic approach that models the joint distribution of latent features. This allows our method to better capture spatial structures and produce diverse synthetic samples, which benefits model training. Specifically, we introduce a low-rank multivariate normal distribution parameterized by a lightweight network. This design maintains low computational complexity and is compatible with various matching networks used in dataset distillation. After distillation, synthetic images are generated by feeding the learned latent features into a pretrained generator. These synthetic images are then used to train classification models, and performance is evaluated on real test set. We validate our method on several benchmarks, including ImageNet subsets, CIFAR-10, and the MedMNIST histopathological dataset. Our approach achieves state-of-the-art cross architecture performance across a range of backbone architectures, demonstrating its generality and effectiveness.", "date": "2025-05-10", "categories": []}
{"id": "2505.11703", "title": "LoFT: LoRA-fused Training Dataset Generation with Few-shot Guidance", "abstract": "Despite recent advances in text-to-image generation, using synthetically generated data seldom brings a significant boost in performance for supervised learning. Oftentimes, synthetic datasets do not faithfully recreate the data distribution of real data, i.e., they lack the fidelity or diversity needed for effective downstream model training. While previous work has employed few-shot guidance to address this issue, existing methods still fail to capture and generate features unique to specific real images. In this paper, we introduce a novel dataset generation framework named LoFT, LoRA-Fused Training-data Generation with Few-shot Guidance. Our method fine-tunes LoRA weights on individual real images and fuses them at inference time, producing synthetic images that combine the features of real images for improved diversity and fidelity of generated data. We evaluate the synthetic data produced by LoFT on 10 datasets, using 8 to 64 real images per class as guidance and scaling up to 1000 images per class. Our experiments show that training on LoFT-generated data consistently outperforms other synthetic dataset methods, significantly increasing accuracy as the dataset size increases. Additionally, our analysis demonstrates that LoFT generates datasets with high fidelity and sufficient diversity, which contribute to the performance improvement. The code is available at https://github.com/ExplainableML/LoFT.", "date": "2025-05-16", "categories": []}
{"id": "2505.11813", "title": "SGD-Mix: Enhancing Domain-Specific Image Classification with Label-Preserving Data Augmentation", "abstract": "Data augmentation for domain-specific image classification tasks often struggles to simultaneously address diversity, faithfulness, and label clarity of generated data, leading to suboptimal performance in downstream tasks. While existing generative diffusion model-based methods aim to enhance augmentation, they fail to cohesively tackle these three critical aspects and often overlook intrinsic challenges of diffusion models, such as sensitivity to model characteristics and stochasticity under strong transformations. In this paper, we propose a novel framework that explicitly integrates diversity, faithfulness, and label clarity into the augmentation process. Our approach employs saliency-guided mixing and a fine-tuned diffusion model to preserve foreground semantics, enrich background diversity, and ensure label consistency, while mitigating diffusion model limitations. Extensive experiments across fine-grained, long-tail, few-shot, and background robustness tasks demonstrate our method's superior performance over state-of-the-art approaches.", "date": "2025-05-17", "categories": []}
{"id": "2503.01052", "title": "ALinFiK: Learning to Approximate Linearized Future Influence Kernel for Scalable Third-Party LLM Data Valuation", "abstract": "Large Language Models (LLMs) heavily rely on high-quality training data, making data valuation crucial for optimizing model performance, especially when working within a limited budget. In this work, we aim to offer a third-party data valuation approach that benefits both data providers and model developers. We introduce a linearized future influence kernel (LinFiK), which assesses the value of individual data samples in improving LLM performance during training. We further propose ALinFiK, a learning strategy to approximate LinFiK, enabling scalable data valuation. Our comprehensive evaluations demonstrate that this approach surpasses existing baselines in effectiveness and efficiency, demonstrating significant scalability advantages as LLM parameters increase.", "date": "2025-03-02", "categories": []}
{"id": "2412.09712", "title": "Investigating the Impact of Balancing, Filtering, and Complexity on Predictive Multiplicity: A Data-Centric Perspective", "abstract": "The Rashomon effect presents a significant challenge in model selection. It occurs when multiple models achieve similar performance on a dataset but produce different predictions, resulting in predictive multiplicity. This is especially problematic in high-stakes environments, where arbitrary model outcomes can have serious consequences. Traditional model selection methods prioritize accuracy and fail to address this issue. Factors such as class imbalance and irrelevant variables further complicate the situation, making it harder for models to provide trustworthy predictions. Data-centric AI approaches can mitigate these problems by prioritizing data optimization, particularly through preprocessing techniques. However, recent studies suggest preprocessing methods may inadvertently inflate predictive multiplicity. This paper investigates how data preprocessing techniques like balancing and filtering methods impact predictive multiplicity and model stability, considering the complexity of the data. We conduct the experiments on 21 real-world datasets, applying various balancing and filtering techniques, and assess the level of predictive multiplicity introduced by these methods by leveraging the Rashomon effect. Additionally, we examine how filtering techniques reduce redundancy and enhance model generalization. The findings provide insights into the relationship between balancing methods, data complexity, and predictive multiplicity, demonstrating how data-centric AI strategies can improve model performance.", "date": "2024-12-12", "categories": []}
{"id": "2505.03809", "title": "When Dynamic Data Selection Meets Data Augmentation", "abstract": "Dynamic data selection aims to accelerate training with lossless performance. However, reducing training data inherently limits data diversity, potentially hindering generalization. While data augmentation is widely used to enhance diversity, it is typically not optimized in conjunction with selection. As a result, directly combining these techniques fails to fully exploit their synergies. To tackle the challenge, we propose a novel online data training framework that, for the first time, unifies dynamic data selection and augmentation, achieving both training efficiency and enhanced performance. Our method estimates each sample's joint distribution of local density and multimodal semantic consistency, allowing for the targeted selection of augmentation-suitable samples while suppressing the inclusion of noisy or ambiguous data. This enables a more significant reduction in dataset size without sacrificing model generalization. Experimental results demonstrate that our method outperforms existing state-of-the-art approaches on various benchmark datasets and architectures, e.g., reducing 50\\% training costs on ImageNet-1k with lossless performance. Furthermore, our approach enhances noise resistance and improves model robustness, reinforcing its practical utility in real-world scenarios.", "date": "2025-05-02", "categories": []}
{"id": "2505.06356", "title": "Understanding and Mitigating Toxicity in Image-Text Pretraining Datasets: A Case Study on LLaVA", "abstract": "Pretraining datasets are foundational to the development of multimodal models, yet they often have inherent biases and toxic content from the web-scale corpora they are sourced from. In this paper, we investigate the prevalence of toxicity in LLaVA image-text pretraining dataset, examining how harmful content manifests in different modalities. We present a comprehensive analysis of common toxicity categories and propose targeted mitigation strategies, resulting in the creation of a refined toxicity-mitigated dataset. This dataset removes 7,531 of toxic image-text pairs in the LLaVA pre-training dataset. We offer guidelines for implementing robust toxicity detection pipelines. Our findings underscore the need to actively identify and filter toxic content - such as hate speech, explicit imagery, and targeted harassment - to build more responsible and equitable multimodal systems. The toxicity-mitigated dataset is open source and is available for further research.", "date": "2025-05-09", "categories": []}
{"id": "2402.05356", "title": "Exploring Learning Complexity for Efficient Downstream Dataset Pruning", "abstract": "  The ever-increasing fine-tuning cost of large-scale pre-trained models gives\nrise to the importance of dataset pruning, which aims to reduce dataset size\nwhile maintaining task performance. However, existing dataset pruning methods\nrequire training on the entire dataset, which is impractical for large-scale\npre-trained models. In this paper, we propose a straightforward, novel, and\ntraining-free hardness score named Distorting-based Learning Complexity (DLC),\nto identify informative images and instructions from the downstream dataset\nefficiently. Our method is motivated by the observation that easy samples\nlearned faster can also be learned with fewer parameters. Specifically, we\ndefine the Learning Complexity to quantify sample hardness and utilize a\nlightweight weights masking process for fast estimation, instead of the costly\nSGD optimization. Based on DLC, we further design a flexible under-sampling\nwith randomness (dubbed FlexRand), replacing the top-K strategy, to alleviate\nthe severe subset distribution shift. Extensive experiments with downstream\nimage and instructions dataset pruning benchmarks demonstrate the effectiveness\nand efficiency of the proposed approach. In the images pruning benchmark, DLC\nsignificantly reduces the pruning time by 35x while establishing\nstate-of-the-art performance with FlexRand.\n", "date": "2024-02-08", "categories": []}
{"id": "2505.03790", "title": "A Time-Series Data Augmentation Model through Diffusion and Transformer\n  Integration", "abstract": "  With the development of Artificial Intelligence, numerous real-world tasks\nhave been accomplished using technology integrated with deep learning. To\nachieve optimal performance, deep neural networks typically require large\nvolumes of data for training. Although advances in data augmentation have\nfacilitated the acquisition of vast datasets, most of this data is concentrated\nin domains like images and speech. However, there has been relatively less\nfocus on augmenting time-series data. To address this gap and generate a\nsubstantial amount of time-series data, we propose a simple and effective\nmethod that combines the Diffusion and Transformer models. By utilizing an\nadjusted diffusion denoising model to generate a large volume of initial\ntime-step action data, followed by employing a Transformer model to predict\nsubsequent actions, and incorporating a weighted loss function to achieve\nconvergence, the method demonstrates its effectiveness. Using the performance\nimprovement of the model after applying augmented data as a benchmark, and\ncomparing the results with those obtained without data augmentation or using\ntraditional data augmentation methods, this approach shows its capability to\nproduce high-quality augmented data.\n", "date": "2025-05-01", "categories": []}
{"id": "2505.05427", "title": "Ultra-FineWeb: Efficient Data Filtering and Verification for\n  High-Quality LLM Training Data", "abstract": "  Data quality has become a key factor in enhancing model performance with the\nrapid development of large language models (LLMs). Model-driven data filtering\nhas increasingly become a primary approach for acquiring high-quality data.\nHowever, it still faces two main challenges: (1) the lack of an efficient data\nverification strategy makes it difficult to provide timely feedback on data\nquality; and (2) the selection of seed data for training classifiers lacks\nclear criteria and relies heavily on human expertise, introducing a degree of\nsubjectivity. To address the first challenge, we introduce an efficient\nverification strategy that enables rapid evaluation of the impact of data on\nLLM training with minimal computational cost. To tackle the second challenge,\nwe build upon the assumption that high-quality seed data is beneficial for LLM\ntraining, and by integrating the proposed verification strategy, we optimize\nthe selection of positive and negative samples and propose an efficient data\nfiltering pipeline. This pipeline not only improves filtering efficiency,\nclassifier quality, and robustness, but also significantly reduces experimental\nand inference costs. In addition, to efficiently filter high-quality data, we\nemploy a lightweight classifier based on fastText, and successfully apply the\nfiltering pipeline to two widely-used pre-training corpora, FineWeb and Chinese\nFineWeb datasets, resulting in the creation of the higher-quality Ultra-FineWeb\ndataset. Ultra-FineWeb contains approximately 1 trillion English tokens and 120\nbillion Chinese tokens. Empirical results demonstrate that the LLMs trained on\nUltra-FineWeb exhibit significant performance improvements across multiple\nbenchmark tasks, validating the effectiveness of our pipeline in enhancing both\ndata quality and training efficiency.\n", "date": "2025-05-08", "categories": []}
{"id": "2504.20500", "title": "UniDetox: Universal Detoxification of Large Language Models via Dataset\n  Distillation", "abstract": "  We present UniDetox, a universally applicable method designed to mitigate\ntoxicity across various large language models (LLMs). Previous detoxification\nmethods are typically model-specific, addressing only individual models or\nmodel families, and require careful hyperparameter tuning due to the trade-off\nbetween detoxification efficacy and language modeling performance. In contrast,\nUniDetox provides a detoxification technique that can be universally applied to\na wide range of LLMs without the need for separate model-specific tuning.\nSpecifically, we propose a novel and efficient dataset distillation technique\nfor detoxification using contrastive decoding. This approach distills\ndetoxifying representations in the form of synthetic text data, enabling\nuniversal detoxification of any LLM through fine-tuning with the distilled\ntext. Our experiments demonstrate that the detoxifying text distilled from\nGPT-2 can effectively detoxify larger models, including OPT, Falcon, and\nLLaMA-2. Furthermore, UniDetox eliminates the need for separate hyperparameter\ntuning for each model, as a single hyperparameter configuration can be\nseamlessly applied across different models. Additionally, analysis of the\ndetoxifying text reveals a reduction in politically biased content, providing\ninsights into the attributes necessary for effective detoxification of LLMs.\n", "date": "2025-04-29", "categories": []}
{"id": "2504.20644", "title": "Combatting Dimensional Collapse in LLM Pre-Training Data via Diversified\n  File Selection", "abstract": "  Selecting high-quality pre-training data for large language models (LLMs) is\ncrucial for enhancing their overall performance under limited computation\nbudget, improving both training and sample efficiency. Recent advancements in\nfile selection primarily rely on using an existing or trained proxy model to\nassess the similarity of samples to a target domain, such as high quality\nsources BookCorpus and Wikipedia. However, upon revisiting these methods, the\ndomain-similarity selection criteria demonstrates a diversity dilemma,\ni.e.dimensional collapse in the feature space, improving performance on the\ndomain-related tasks but causing severe degradation on generic performance. To\nprevent collapse and enhance diversity, we propose a DiverSified File selection\nalgorithm (DiSF), which selects the most decorrelated text files in the feature\nspace. We approach this with a classical greedy algorithm to achieve more\nuniform eigenvalues in the feature covariance matrix of the selected texts,\nanalyzing its approximation to the optimal solution under a formulation of\n$\\gamma$-weakly submodular optimization problem. Empirically, we establish a\nbenchmark and conduct extensive experiments on the TinyLlama architecture with\nmodels from 120M to 1.1B parameters. Evaluating across nine tasks from the\nHarness framework, DiSF demonstrates a significant improvement on overall\nperformance. Specifically, DiSF saves 98.5% of 590M training files in\nSlimPajama, outperforming the full-data pre-training within a 50B training\nbudget, and achieving about 1.5x training efficiency and 5x data efficiency.\n", "date": "2025-04-29", "categories": []}
{"id": "2504.16511", "title": "QuaDMix: Quality-Diversity Balanced Data Selection for Efficient LLM\n  Pretraining", "abstract": "  Quality and diversity are two critical metrics for the training data of large\nlanguage models (LLMs), positively impacting performance. Existing studies\noften optimize these metrics separately, typically by first applying quality\nfiltering and then adjusting data proportions. However, these approaches\noverlook the inherent trade-off between quality and diversity, necessitating\ntheir joint consideration. Given a fixed training quota, it is essential to\nevaluate both the quality of each data point and its complementary effect on\nthe overall dataset. In this paper, we introduce a unified data selection\nframework called QuaDMix, which automatically optimizes the data distribution\nfor LLM pretraining while balancing both quality and diversity. Specifically,\nwe first propose multiple criteria to measure data quality and employ domain\nclassification to distinguish data points, thereby measuring overall diversity.\nQuaDMix then employs a unified parameterized data sampling function that\ndetermines the sampling probability of each data point based on these quality\nand diversity related labels. To accelerate the search for the optimal\nparameters involved in the QuaDMix framework, we conduct simulated experiments\non smaller models and use LightGBM for parameters searching, inspired by the\nRegMix method. Our experiments across diverse models and datasets demonstrate\nthat QuaDMix achieves an average performance improvement of 7.2% across\nmultiple benchmarks. These results outperform the independent strategies for\nquality and diversity, highlighting the necessity and ability to balance data\nquality and diversity.\n", "date": "2025-04-23", "categories": []}
{"id": "2504.17132", "title": "Latent Video Dataset Distillation", "abstract": "  Dataset distillation has demonstrated remarkable effectiveness in\nhigh-compression scenarios for image datasets. While video datasets inherently\ncontain greater redundancy, existing video dataset distillation methods\nprimarily focus on compression in the pixel space, overlooking advances in the\nlatent space that have been widely adopted in modern text-to-image and\ntext-to-video models. In this work, we bridge this gap by introducing a novel\nvideo dataset distillation approach that operates in the latent space using a\nstate-of-the-art variational encoder. Furthermore, we employ a diversity-aware\ndata selection strategy to select both representative and diverse samples.\nAdditionally, we introduce a simple, training-free method to further compress\nthe distilled latent dataset. By combining these techniques, our approach\nachieves a new state-of-the-art performance in dataset distillation,\noutperforming prior methods on all datasets, e.g. on HMDB51 IPC 1, we achieve a\n2.6% performance increase; on MiniUCF IPC 5, we achieve a 7.8% performance\nincrease. Our code is available at\nhttps://github.com/liningresearch/Latent_Video_Dataset_Distillation.\n", "date": "2025-04-23", "categories": []}
{"id": "2504.16268", "title": "Boosting KNNClassifier Performance with Opposition-Based Data\n  Transformation", "abstract": "  In this paper, we introduce a novel data transformation framework based on\nOpposition-Based Learning (OBL) to boost the performance of traditional\nclassification algorithms. Originally developed to accelerate convergence in\noptimization tasks, OBL is leveraged here to generate synthetic opposite\nsamples that enrich the training data and improve decision boundary formation.\nWe explore three OBL variants Global OBL, Class-Wise OBL, and Localized\nClass-Wise OBL and integrate them with K-Nearest Neighbors (KNN). Extensive\nexperiments conducted on 26 heterogeneous and high-dimensional datasets\ndemonstrate that OBL-enhanced classifiers consistently outperform the basic\nKNN. These findings underscore the potential of OBL as a lightweight yet\npowerful data transformation strategy for enhancing classification performance,\nespecially in complex or sparse learning environments.\n", "date": "2025-04-22", "categories": []}
{"id": "2409.17685", "title": "Feature-to-Image Data Augmentation: Improving Model Feature Extraction\n  with Cluster-Guided Synthetic Samples", "abstract": "  One of the growing trends in machine learning is the use of data generation\ntechniques, since the performance of machine learning models is dependent on\nthe quantity of the training dataset. However, in many real-world applications,\nparticularly in medical and low-resource domains, collecting large datasets is\nchallenging due to resource constraints, which leads to overfitting and poor\ngeneralization. This study introduces FICAug, a novel feature-to-image data\naugmentation framework designed to improve model generalization under limited\ndata conditions by generating structured synthetic samples.\n  FICAug first operates in the feature space, where original data are clustered\nusing the k-means algorithm. Within pure-label clusters, synthetic data are\ngenerated through Gaussian sampling to increase diversity while maintaining\nlabel consistency. These synthetic features are then projected back into the\nimage domain using a generative neural network, and a convolutional neural\nnetwork is trained on the reconstructed images to learn enhanced\nrepresentations.\n  Experimental results demonstrate that FICAug significantly improves\nclassification accuracy. In feature space, it achieved a cross-validation\naccuracy of 84.09%, while training a ResNet-18 model on the reconstructed\nimages further boosted performance to 88.63%, illustrating the effectiveness of\nthe proposed framework in extracting new and task-relevant features.\n", "date": "2024-09-26", "categories": []}
{"id": "2504.09818", "title": "Transferable text data distillation by trajectory matching", "abstract": "  In the realm of large language model (LLM), as the size of large models\nincreases, it also brings higher training costs. There is a urgent need to\nminimize the data size in LLM training. Compared with data selection method,\nthe data distillation method aims to synthesize a small number of data samples\nto achieve the training effect of the full data set and has better flexibility.\nDespite its successes in computer vision, the discreteness of text data has\nhitherto stymied its exploration in natural language processing (NLP). In this\nwork, we proposed a method that involves learning pseudo prompt data based on\ntrajectory matching and finding its nearest neighbor ID to achieve\ncross-architecture transfer. During the distillation process, we introduce a\nregularization loss to improve the robustness of our distilled data. To our\nbest knowledge, this is the first data distillation work suitable for text\ngeneration tasks such as instruction tuning. Evaluations on two benchmarks,\nincluding ARC-Easy and MMLU instruction tuning datasets, established the\nsuperiority of our distillation approach over the SOTA data selection method\nLESS. Furthermore, our method demonstrates a good transferability over LLM\nstructures (i.e., OPT to Llama).\n", "date": "2025-04-14", "categories": []}
{"id": "2409.05202", "title": "A Survey on Mixup Augmentations and Beyond", "abstract": "  As Deep Neural Networks have achieved thrilling breakthroughs in the past\ndecade, data augmentations have garnered increasing attention as regularization\ntechniques when massive labeled data are unavailable. Among existing\naugmentations, Mixup and relevant data-mixing methods that convexly combine\nselected samples and the corresponding labels are widely adopted because they\nyield high performances by generating data-dependent virtual data while easily\nmigrating to various domains. This survey presents a comprehensive review of\nfoundational mixup methods and their applications. We first elaborate on the\ntraining pipeline with mixup augmentations as a unified framework containing\nmodules. A reformulated framework could contain various mixup methods and give\nintuitive operational procedures. Then, we systematically investigate the\napplications of mixup augmentations on vision downstream tasks, various data\nmodalities, and some analysis \\& theorems of mixup. Meanwhile, we conclude the\ncurrent status and limitations of mixup research and point out further work for\neffective and efficient mixup augmentations. This survey can provide\nresearchers with the current state of the art in mixup methods and provide some\ninsights and guidance roles in the mixup arena. An online project with this\nsurvey is available at https://github.com/Westlake-AI/Awesome-Mixup.\n", "date": "2024-09-08", "categories": []}
{"id": "2504.16277", "title": "DataS^3: Dataset Subset Selection for Specialization", "abstract": "  In many real-world machine learning (ML) applications (e.g. detecting broken\nbones in x-ray images, detecting species in camera traps), in practice models\nneed to perform well on specific deployments (e.g. a specific hospital, a\nspecific national park) rather than the domain broadly. However, deployments\noften have imbalanced, unique data distributions. Discrepancy between the\ntraining distribution and the deployment distribution can lead to suboptimal\nperformance, highlighting the need to select deployment-specialized subsets\nfrom the available training data. We formalize dataset subset selection for\nspecialization (DS3): given a training set drawn from a general distribution\nand a (potentially unlabeled) query set drawn from the desired\ndeployment-specific distribution, the goal is to select a subset of the\ntraining data that optimizes deployment performance.\n  We introduce DataS^3; the first dataset and benchmark designed specifically\nfor the DS3 problem. DataS^3 encompasses diverse real-world application\ndomains, each with a set of distinct deployments to specialize in. We conduct a\ncomprehensive study evaluating algorithms from various families--including\ncoresets, data filtering, and data curation--on DataS^3, and find that\ngeneral-distribution methods consistently fail on deployment-specific tasks.\nAdditionally, we demonstrate the existence of manually curated\n(deployment-specific) expert subsets that outperform training on all available\ndata with accuracy gains up to 51.3 percent. Our benchmark highlights the\ncritical role of tailored dataset curation in enhancing performance and\ntraining efficiency on deployment-specific distributions, which we posit will\nonly become more important as global, public datasets become available across\ndomains and ML models are deployed in the real world.\n", "date": "2025-04-22", "categories": []}
{"id": "2504.16430", "title": "MAGIC: Near-Optimal Data Attribution for Deep Learning", "abstract": "  The goal of predictive data attribution is to estimate how adding or removing\na given set of training datapoints will affect model predictions. In convex\nsettings, this goal is straightforward (i.e., via the infinitesimal jackknife).\nIn large-scale (non-convex) settings, however, existing methods are far less\nsuccessful -- current methods' estimates often only weakly correlate with\nground truth. In this work, we present a new data attribution method (MAGIC)\nthat combines classical methods and recent advances in metadifferentiation to\n(nearly) optimally estimate the effect of adding or removing training data on\nmodel predictions.\n", "date": "2025-04-23", "categories": []}
{"id": "2504.16668", "title": "Efficient Data Valuation Approximation in Federated Learning: A\n  Sampling-based Approach", "abstract": "  Federated learning paradigm to utilize datasets across multiple data\nproviders. In FL, cross-silo data providers often hesitate to share their\nhigh-quality dataset unless their data value can be fairly assessed. Shapley\nvalue (SV) has been advocated as the standard metric for data valuation in FL\ndue to its desirable properties. However, the computational overhead of SV is\nprohibitive in practice, as it inherently requires training and evaluating an\nFL model across an exponential number of dataset combinations. Furthermore,\nexisting solutions fail to achieve high accuracy and efficiency, making\npractical use of SV still out of reach, because they ignore choosing suitable\ncomputation scheme for approximation framework and overlook the property of\nutility function in FL. We first propose a unified stratified-sampling\nframework for two widely-used schemes. Then, we analyze and choose the more\npromising scheme under the FL linear regression assumption. After that, we\nidentify a phenomenon termed key combinations, where only limited dataset\ncombinations have a high-impact on final data value. Building on these\ninsights, we propose a practical approximation algorithm, IPSS, which\nstrategically selects high-impact dataset combinations rather than evaluating\nall possible combinations, thus substantially reducing time cost with minor\napproximation error. Furthermore, we conduct extensive evaluations on the FL\nbenchmark datasets to demonstrate that our proposed algorithm outperforms a\nseries of representative baselines in terms of efficiency and effectiveness.\n", "date": "2025-04-23", "categories": []}
{"id": "2406.11794", "title": "DataComp-LM: In search of the next generation of training sets for\n  language models", "abstract": "  We introduce DataComp for Language Models (DCLM), a testbed for controlled\ndataset experiments with the goal of improving language models. As part of\nDCLM, we provide a standardized corpus of 240T tokens extracted from Common\nCrawl, effective pretraining recipes based on the OpenLM framework, and a broad\nsuite of 53 downstream evaluations. Participants in the DCLM benchmark can\nexperiment with data curation strategies such as deduplication, filtering, and\ndata mixing at model scales ranging from 412M to 7B parameters. As a baseline\nfor DCLM, we conduct extensive experiments and find that model-based filtering\nis key to assembling a high-quality training set. The resulting dataset,\nDCLM-Baseline enables training a 7B parameter language model from scratch to\n64% 5-shot accuracy on MMLU with 2.6T training tokens. Compared to MAP-Neo, the\nprevious state-of-the-art in open-data language models, DCLM-Baseline\nrepresents a 6.6 percentage point improvement on MMLU while being trained with\n40% less compute. Our baseline model is also comparable to Mistral-7B-v0.3 and\nLlama 3 8B on MMLU (63% & 66%), and performs similarly on an average of 53\nnatural language understanding tasks while being trained with 6.6x less compute\nthan Llama 3 8B. Our results highlight the importance of dataset design for\ntraining language models and offer a starting point for further research on\ndata curation.\n", "date": "2024-06-17", "categories": []}
{"id": "2411.05735", "title": "Aioli: A Unified Optimization Framework for Language Model Data Mixing", "abstract": "  Language model performance depends on identifying the optimal mixture of data\ngroups to train on (e.g., law, code, math). Prior work has proposed a diverse\nset of methods to efficiently learn mixture proportions, ranging from fitting\nregression models over training runs to dynamically updating proportions\nthroughout training. Surprisingly, we find that no existing method consistently\noutperforms a simple stratified sampling baseline in terms of average test\nperplexity. To understand this inconsistency, we unify existing methods into a\nstandard framework, showing they are equivalent to solving a common\noptimization problem: minimize average loss subject to a method-specific mixing\nlaw -- an implicit assumption on the relationship between loss and mixture\nproportions. This framework suggests that measuring the fidelity of a method's\nmixing law can offer insights into its performance. Empirically, we find that\nexisting methods set their mixing law parameters inaccurately, resulting in the\ninconsistent mixing performance we observe. Using this insight, we derive a new\nonline method named Aioli, which directly estimates the mixing law parameters\nthroughout training and uses them to dynamically adjust proportions. Aioli\noutperforms stratified sampling on 6 out of 6 datasets by an average of 0.27\ntest perplexity points, whereas existing methods fail to consistently beat\nstratified sampling, doing up to 6.9 points worse. Moreover, in a practical\nsetting where proportions are learned on shorter runs due to computational\nconstraints, Aioli can dynamically adjust these proportions over the full\ntraining run, consistently improving performance over existing methods by up to\n12.012 test perplexity points.\n", "date": "2024-11-08", "categories": []}
{"id": "2504.00695", "title": "ToReMi: Topic-Aware Data Reweighting for Dynamic Pre-Training Data\n  Selection", "abstract": "  Pre-training large language models (LLMs) necessitates enormous diverse\ntextual corpora, making effective data selection a key challenge for balancing\ncomputational resources and model performance. Current methodologies primarily\nemphasize data quality metrics and mixing proportions, yet they fail to\nadequately capture the underlying semantic connections between training samples\nand quality disparities within individual domains. We introduce ToReMi\n(Topic-based Reweighting for Model improvement), a novel two-stage framework\nthat dynamically adjusts training sample weights according to their topical\nassociations and observed learning patterns. Our comprehensive experiments\nreveal that ToReMi variants consistently achieve superior performance over\nconventional pre-training approaches, demonstrating accelerated perplexity\nreduction across multiple domains and enhanced capabilities on downstream\nevaluation tasks. Code is available at https://github.com/zxx000728/ToReMi.\n", "date": "2025-04-01", "categories": []}
{"id": "2504.14826", "title": "Distribution-aware Dataset Distillation for Efficient Image Restoration", "abstract": "  With the exponential increase in image data, training an image restoration\nmodel is laborious. Dataset distillation is a potential solution to this\nproblem, yet current distillation techniques are a blank canvas in the field of\nimage restoration. To fill this gap, we propose the Distribution-aware Dataset\nDistillation method (TripleD), a new framework that extends the principles of\ndataset distillation to image restoration. Specifically, TripleD uses a\npre-trained vision Transformer to extract features from images for complexity\nevaluation, and the subset (the number of samples is much smaller than the\noriginal training set) is selected based on complexity. The selected subset is\nthen fed through a lightweight CNN that fine-tunes the image distribution to\nalign with the distribution of the original dataset at the feature level. To\nefficiently condense knowledge, the training is divided into two stages. Early\nstages focus on simpler, low-complexity samples to build foundational\nknowledge, while later stages select more complex and uncertain samples as the\nmodel matures. Our method achieves promising performance on multiple image\nrestoration tasks, including multi-task image restoration, all-in-one image\nrestoration, and ultra-high-definition image restoration tasks. Note that we\ncan train a state-of-the-art image restoration model on an\nultra-high-definition (4K resolution) dataset using only one consumer-grade GPU\nin less than 8 hours (500 savings in computing resources and immeasurable\ntraining time).\n", "date": "2025-04-21", "categories": []}
{"id": "2504.14875", "title": "ReSpec: Relevance and Specificity Grounded Online Filtering for Learning\n  on Video-Text Data Streams", "abstract": "  The rapid growth of video-text data presents challenges in storage and\ncomputation during training. Online learning, which processes streaming data in\nreal-time, offers a promising solution to these issues while also allowing\nswift adaptations in scenarios demanding real-time responsiveness. One strategy\nto enhance the efficiency and effectiveness of learning involves identifying\nand prioritizing data that enhances performance on target downstream tasks. We\npropose Relevance and Specificity-based online filtering framework (ReSpec)\nthat selects data based on four criteria: (i) modality alignment for clean\ndata, (ii) task relevance for target focused data, (iii) specificity for\ninformative and detailed data, and (iv) efficiency for low-latency processing.\nRelevance is determined by the probabilistic alignment of incoming data with\ndownstream tasks, while specificity employs the distance to a root embedding\nrepresenting the least specific data as an efficient proxy for informativeness.\nBy establishing reference points from target task data, ReSpec filters incoming\ndata in real-time, eliminating the need for extensive storage and compute.\nEvaluating on large-scale datasets WebVid2M and VideoCC3M, ReSpec attains\nstate-of-the-art performance on five zeroshot video retrieval tasks, using as\nlittle as 5% of the data while incurring minimal compute. The source code is\navailable at https://github.com/cdjkim/ReSpec.\n", "date": "2025-04-21", "categories": []}
{"id": "2504.15022", "title": "LLMs as Data Annotators: How Close Are We to Human Performance", "abstract": "  In NLP, fine-tuning LLMs is effective for various applications but requires\nhigh-quality annotated data. However, manual annotation of data is\nlabor-intensive, time-consuming, and costly. Therefore, LLMs are increasingly\nused to automate the process, often employing in-context learning (ICL) in\nwhich some examples related to the task are given in the prompt for better\nperformance. However, manually selecting context examples can lead to\ninefficiencies and suboptimal model performance. This paper presents\ncomprehensive experiments comparing several LLMs, considering different\nembedding models, across various datasets for the Named Entity Recognition\n(NER) task. The evaluation encompasses models with approximately $7$B and $70$B\nparameters, including both proprietary and non-proprietary models. Furthermore,\nleveraging the success of Retrieval-Augmented Generation (RAG), it also\nconsiders a method that addresses the limitations of ICL by automatically\nretrieving contextual examples, thereby enhancing performance. The results\nhighlight the importance of selecting the appropriate LLM and embedding model,\nunderstanding the trade-offs between LLM sizes and desired performance, and the\nnecessity to direct research efforts towards more challenging datasets.\n", "date": "2025-04-21", "categories": []}
{"id": "2504.13586", "title": "How to Achieve Higher Accuracy with Less Training Points?", "abstract": "  In the era of large-scale model training, the extensive use of available\ndatasets has resulted in significant computational inefficiencies. To tackle\nthis issue, we explore methods for identifying informative subsets of training\ndata that can achieve comparable or even superior model performance. We propose\na technique based on influence functions to determine which training samples\nshould be included in the training set. We conducted empirical evaluations of\nour method on binary classification tasks utilizing logistic regression models.\nOur approach demonstrates performance comparable to that of training on the\nentire dataset while using only 10% of the data. Furthermore, we found that our\nmethod achieved even higher accuracy when trained with just 60% of the data.\n", "date": "2025-04-18", "categories": []}
{"id": "2504.13835", "title": "MIG: Automatic Data Selection for Instruction Tuning by Maximizing\n  Information Gain in Semantic Space", "abstract": "  Data quality and diversity are key to the construction of effective\ninstruction-tuning datasets. % With the increasing availability of open-source\ninstruction-tuning datasets, it is advantageous to automatically select\nhigh-quality and diverse subsets from a vast amount of data. % Existing methods\ntypically prioritize instance quality and use heuristic rules to maintain\ndiversity. % However, this absence of a comprehensive view of the entire\ncollection often leads to suboptimal results. % Moreover, heuristic rules\ngenerally focus on distance or clustering within the embedding space, which\nfails to accurately capture the intent of complex instructions in the semantic\nspace. % To bridge this gap, we propose a unified method for quantifying the\ninformation content of datasets. This method models the semantic space by\nconstructing a label graph and quantifies diversity based on the distribution\nof information within the graph. % Based on such a measurement, we further\nintroduce an efficient sampling method that selects data samples iteratively to\n\\textbf{M}aximize the \\textbf{I}nformation \\textbf{G}ain (MIG) in semantic\nspace. % Experiments on various datasets and base models demonstrate that MIG\nconsistently outperforms state-of-the-art methods. % Notably, the model\nfine-tuned with 5\\% Tulu3 data sampled by MIG achieves comparable performance\nto the official SFT model trained on the full dataset, with improvements of\n+5.73\\% on AlpacaEval and +6.89\\% on Wildbench.\n", "date": "2025-04-18", "categories": []}
{"id": "2206.04846", "title": "Masked Autoencoders are Robust Data Augmentors", "abstract": "  Deep neural networks are capable of learning powerful representations to\ntackle complex vision tasks but expose undesirable properties like the\nover-fitting issue. To this end, regularization techniques like image\naugmentation are necessary for deep neural networks to generalize well.\nNevertheless, most prevalent image augmentation recipes confine themselves to\noff-the-shelf linear transformations like scale, flip, and colorjitter. Due to\ntheir hand-crafted property, these augmentations are insufficient to generate\ntruly hard augmented examples. In this paper, we propose a novel perspective of\naugmentation to regularize the training process. Inspired by the recent success\nof applying masked image modeling to self-supervised learning, we adopt the\nself-supervised masked autoencoder to generate the distorted view of the input\nimages. We show that utilizing such model-based nonlinear transformation as\ndata augmentation can improve high-level recognition tasks. We term the\nproposed method as \\textbf{M}ask-\\textbf{R}econstruct \\textbf{A}ugmentation\n(MRA). The extensive experiments on various image classification benchmarks\nverify the effectiveness of the proposed augmentation. Specifically, MRA\nconsistently enhances the performance on supervised, semi-supervised as well as\nfew-shot classification.\n", "date": "2022-06-10", "categories": []}
{"id": "2304.10701", "title": "GMValuator: Similarity-based Data Valuation for Generative Models", "abstract": "  Data valuation plays a crucial role in machine learning. Existing data\nvaluation methods, mainly focused on discriminative models, overlook generative\nmodels that have gained attention recently. In generative models, data\nvaluation measures the impact of training data on generated datasets. Very few\nexisting attempts at data valuation methods designed for deep generative models\neither concentrate on specific models or lack robustness in their outcomes.\nMoreover, efficiency still reveals vulnerable shortcomings. We formulate the\ndata valuation problem in generative models from a similarity matching\nperspective to bridge the gaps. Specifically, we introduce Generative Model\nValuator (GMValuator), the first training-free and model-agnostic approach to\nproviding data valuation for image generation tasks. It empowers efficient data\nvaluation through our innovative similarity matching module, calibrates biased\ncontributions by incorporating image quality assessment, and attributes credits\nto all training samples based on their contributions to the generated samples.\nAdditionally, we introduce four evaluation criteria for assessing data\nvaluation methods in generative models. GMValuator is extensively evaluated on\nbenchmark and high-resolution datasets and various mainstream generative\narchitectures to demonstrate its effectiveness.\n", "date": "2023-04-21", "categories": []}
{"id": "2410.18194", "title": "ZIP-FIT: Embedding-Free Data Selection via Compression-Based Alignment", "abstract": "  Data selection is crucial for optimizing language model (LM) performance on\nspecific tasks, yet most existing methods fail to effectively consider the\ntarget task distribution.\n  Current approaches either ignore task-specific requirements entirely or rely\non approximations that fail to capture the nuanced patterns needed for tasks\nlike Autoformalization or code generation.\n  Methods that do consider the target distribution often rely on simplistic,\nsometimes noisy, representations, like hashed n-gram features, which can lead\nto collisions and introduce noise.\n  We introduce ZIP-FIT, a data selection framework that uses gzip compression\nto directly measure alignment between potential training data and the target\ntask distribution.\n  In extensive evaluations on Autoformalization and Python code generation,\nZIP-FIT significantly outperforms leading baselines like DSIR and D4.\n  Models trained on ZIP-FIT-selected data achieve their lowest cross-entropy\nloss up to 85.1\\% faster than baselines, demonstrating that better task\nalignment leads to more efficient learning.\n  In addition, ZIP-FIT performs selection up to 65.8\\% faster than DSIR and two\norders of magnitude faster than D4.\n  Notably, ZIP-FIT shows that smaller, well-aligned datasets often outperform\nlarger but less targeted ones, demonstrating that a small amount of higher\nquality data is superior to a large amount of lower quality data.\n  Our results imply that task-aware data selection is crucial for efficient\ndomain adaptation, and that compression offers a principled way to measure task\nalignment.\n  By showing that targeted data selection can dramatically improve\ntask-specific performance, our work provides new insights into the relationship\nbetween data quality, task alignment, and model learning efficiency.\n", "date": "2024-10-23", "categories": []}
{"id": "2503.18626", "title": "Generative Dataset Distillation using Min-Max Diffusion Model", "abstract": "  In this paper, we address the problem of generative dataset distillation that\nutilizes generative models to synthesize images. The generator may produce any\nnumber of images under a preserved evaluation time. In this work, we leverage\nthe popular diffusion model as the generator to compute a surrogate dataset,\nboosted by a min-max loss to control the dataset's diversity and\nrepresentativeness during training. However, the diffusion model is\ntime-consuming when generating images, as it requires an iterative generation\nprocess. We observe a critical trade-off between the number of image samples\nand the image quality controlled by the diffusion steps and propose Diffusion\nStep Reduction to achieve optimal performance. This paper details our\ncomprehensive method and its performance. Our model achieved $2^{nd}$ place in\nthe generative track of \\href{https://www.dd-challenge.com/#/}{The First\nDataset Distillation Challenge of ECCV2024}, demonstrating its superior\nperformance.\n", "date": "2025-03-24", "categories": []}
{"id": "2504.09341", "title": "Minority Reports: Balancing Cost and Quality in Ground Truth Data\n  Annotation", "abstract": "  High-quality data annotation is an essential but laborious and costly aspect\nof developing machine learning-based software. We explore the inherent tradeoff\nbetween annotation accuracy and cost by detecting and removing minority reports\n-- instances where annotators provide incorrect responses -- that indicate\nunnecessary redundancy in task assignments. We propose an approach to prune\npotentially redundant annotation task assignments before they are executed by\nestimating the likelihood of an annotator disagreeing with the majority vote\nfor a given task. Our approach is informed by an empirical analysis over\ncomputer vision datasets annotated by a professional data annotation platform,\nwhich reveals that the likelihood of a minority report event is dependent\nprimarily on image ambiguity, worker variability, and worker fatigue.\nSimulations over these datasets show that we can reduce the number of\nannotations required by over 60% with a small compromise in label quality,\nsaving approximately 6.6 days-equivalent of labor. Our approach provides\nannotation service platforms with a method to balance cost and dataset quality.\nMachine learning practitioners can tailor annotation accuracy levels according\nto specific application needs, thereby optimizing budget allocation while\nmaintaining the data quality necessary for critical settings like autonomous\ndriving technology.\n", "date": "2025-04-12", "categories": []}
{"id": "2403.10045", "title": "Towards Adversarially Robust Dataset Distillation by Curvature\n  Regularization", "abstract": "  Dataset distillation (DD) allows datasets to be distilled to fractions of\ntheir original size while preserving the rich distributional information, so\nthat models trained on the distilled datasets can achieve a comparable accuracy\nwhile saving significant computational loads. Recent research in this area has\nbeen focusing on improving the accuracy of models trained on distilled\ndatasets. In this paper, we aim to explore a new perspective of DD. We study\nhow to embed adversarial robustness in distilled datasets, so that models\ntrained on these datasets maintain the high accuracy and meanwhile acquire\nbetter adversarial robustness. We propose a new method that achieves this goal\nby incorporating curvature regularization into the distillation process with\nmuch less computational overhead than standard adversarial training. Extensive\nempirical experiments suggest that our method not only outperforms standard\nadversarial training on both accuracy and robustness with less computation\noverhead but is also capable of generating robust distilled datasets that can\nwithstand various adversarial attacks. Our implementation is available at:\nhttps://github.com/yumozi/GUARD.\n", "date": "2024-03-15", "categories": []}
{"id": "2408.07107", "title": "A Self-Supervised Paradigm for Data-Efficient Medical Foundation Model\n  Pre-training: V-information Optimization Framework", "abstract": "  Self-supervised pre-training medical foundation models on large-scale\ndatasets demonstrate exceptional performance. Recent research challenges this\ncommon paradigm by introducing data-effective learning approaches,\ndemonstrating that merely increasing pre-training data volume does not\nnecessarily improve model performance. However, current methods still have\nunclear standards and the underlying theoretical foundation remains unknown. In\nthis paper, as the first attempt to address this limitation, we introduce\nV-information into self-supervised pre-training of foundation models to provide\na theoretical foundation for sample selection. Our derivation confirms that by\noptimizing V-information, sample selection can be framed as an optimization\nproblem where choosing diverse and challenging samples enhances model\nperformance even under limited training data. Under this guidance, we develop\nan optimized data-effective learning method (OptiDEL) to optimize V-information\nin real-world medical domains by generating more diverse and harder samples. We\ncompare the OptiDEL method with state-of-the-art approaches finding that\nOptiDEL consistently outperforms existing approaches across eight different\ndatasets, with foundation models trained on only 5% of the pre-training data\nachieving up to 6.2% higher mIoU than those trained on the full dataset.\nRemarkably, OptiDEL demonstrates an average improvement of 4.7% mIoU over\ncompeting methods while using 20x less training data.\n", "date": "2024-08-13", "categories": []}
{"id": "2504.00564", "title": "Geometric Median Matching for Robust k-Subset Selection from Noisy Data", "abstract": "  Data pruning -- the combinatorial task of selecting a small and\nrepresentative subset from a large dataset, is crucial for mitigating the\nenormous computational costs associated with training data-hungry modern deep\nlearning models at scale. Since large scale data collections are invariably\nnoisy, developing data pruning strategies that remain robust even in the\npresence of corruption is critical in practice. However, existing data pruning\nmethods often fail under high corruption rates due to their reliance on\nempirical mean estimation, which is highly sensitive to outliers.\n  In response, we propose Geometric Median (GM) Matching, a novel k-subset\nselection strategy that leverages Geometric Median -- a robust estimator with\nan optimal breakdown point of 1/2; to enhance resilience against noisy data.\nOur method iteratively selects a k-subset such that the mean of the subset\napproximates the GM of the (potentially) noisy dataset, ensuring robustness\neven under arbitrary corruption. We provide theoretical guarantees, showing\nthat GM Matching enjoys an improved O(1/k) convergence rate -- a quadratic\nimprovement over random sampling, even under arbitrary corruption. Extensive\nexperiments across image classification and image generation tasks demonstrate\nthat GM Matching consistently outperforms existing pruning approaches,\nparticularly in high-corruption settings and at high pruning rates; making it a\nstrong baseline for robust data pruning.\n", "date": "2025-04-01", "categories": []}
{"id": "2410.02116", "title": "Dataset Distillation via Knowledge Distillation: Towards Efficient\n  Self-Supervised Pre-Training of Deep Networks", "abstract": "  Dataset distillation (DD) generates small synthetic datasets that can\nefficiently train deep networks with a limited amount of memory and compute.\nDespite the success of DD methods for supervised learning, DD for\nself-supervised pre-training of deep models has remained unaddressed.\nPre-training on unlabeled data is crucial for efficiently generalizing to\ndownstream tasks with limited labeled data. In this work, we propose the first\neffective DD method for SSL pre-training. First, we show, theoretically and\nempirically, that naive application of supervised DD methods to SSL fails, due\nto the high variance of the SSL gradient. Then, we address this issue by\nrelying on insights from knowledge distillation (KD) literature. Specifically,\nwe train a small student model to match the representations of a larger teacher\nmodel trained with SSL. Then, we generate a small synthetic dataset by matching\nthe training trajectories of the student models. As the KD objective has\nconsiderably lower variance than SSL, our approach can generate synthetic\ndatasets that can successfully pre-train high-quality encoders. Through\nextensive experiments, we show that our distilled sets lead to up to 13% higher\naccuracy than prior work, on a variety of downstream tasks, in the presence of\nlimited labeled data. Code at https://github.com/BigML-CS-UCLA/MKDT.\n", "date": "2024-10-03", "categories": []}
{"id": "2502.14050", "title": "Diversity-driven Data Selection for Language Model Tuning through Sparse\n  Autoencoder", "abstract": "  Instruction tuning data are often quantity-saturated due to the large volume\nof data collection and fast model iteration, leaving data selection important\nbut underexplored. Existing quality-driven data selection methods, such as LIMA\n(NeurIPS 2023 \\citep{zhou2024lima}) and AlpaGasus (ICLR 2024\n\\citep{chenalpagasus}) generally ignore the equal importance of data diversity\nand complexity. In this work, we aim to design a diversity-aware data selection\nstrategy and creatively propose using sparse autoencoders (SAEs) to tackle the\nchallenge of data diversity measure. In addition, SAEs can also provide more\ninterpretability of model behavior and explain, e.g., the surprising\neffectiveness of selecting the longest response (ICML 2024 \\citep{zhaolong}).\nUsing effective data selection, we experimentally prove that models trained on\nour selected data can outperform other methods in terms of model capabilities,\nreduce training cost, and potentially gain more control over model behaviors.\nWe prove that SAEs can serve as a good alternative to diversity measure and\ndesign our method to be scalable for potential industrial large-scale pruning,\nand we will also release our trained SAEs for use by the broader community.\n", "date": "2025-02-19", "categories": []}
{"id": "2410.17193", "title": "Emphasizing Discriminative Features for Dataset Distillation in Complex\n  Scenarios", "abstract": "  Dataset distillation has demonstrated strong performance on simple datasets\nlike CIFAR, MNIST, and TinyImageNet but struggles to achieve similar results in\nmore complex scenarios. In this paper, we propose EDF (emphasizes the\ndiscriminative features), a dataset distillation method that enhances key\ndiscriminative regions in synthetic images using Grad-CAM activation maps. Our\napproach is inspired by a key observation: in simple datasets, high-activation\nareas typically occupy most of the image, whereas in complex scenarios, the\nsize of these areas is much smaller. Unlike previous methods that treat all\npixels equally when synthesizing images, EDF uses Grad-CAM activation maps to\nenhance high-activation areas. From a supervision perspective, we downplay\nsupervision signals that have lower losses, as they contain common patterns.\nAdditionally, to help the DD community better explore complex scenarios, we\nbuild the Complex Dataset Distillation (Comp-DD) benchmark by meticulously\nselecting sixteen subsets, eight easy and eight hard, from ImageNet-1K. In\nparticular, EDF consistently outperforms SOTA results in complex scenarios,\nsuch as ImageNet-1K subsets. Hopefully, more researchers will be inspired and\nencouraged to improve the practicality and efficacy of DD. Our code and\nbenchmark will be made public at https://github.com/NUS-HPC-AI-Lab/EDF.\n", "date": "2024-10-22", "categories": []}
{"id": "2410.17579", "title": "Bonsai: Gradient-free Graph Condensation for Node Classification", "abstract": "  Graph condensation has emerged as a promising avenue to enable scalable\ntraining of GNNs by compressing the training dataset while preserving essential\ngraph characteristics. Our study uncovers significant shortcomings in current\ngraph condensation techniques. First, the majority of the algorithms\nparadoxically require training on the full dataset to perform condensation.\nSecond, due to their gradient-emulating approach, these methods require fresh\ncondensation for any change in hyperparameters or GNN architecture, limiting\ntheir flexibility and reusability. Finally, they fail to achieve substantial\nsize reduction due to synthesizing fully-connected, edge-weighted graphs. To\naddress these challenges, we present Bonsai, a novel graph condensation method\nempowered by the observation that \\textit{computation trees} form the\nfundamental processing units of message-passing GNNs. Bonsai condenses datasets\nby encoding a careful selection of \\textit{exemplar} trees that maximize the\nrepresentation of all computation trees in the training set. This unique\napproach imparts Bonsai as the first linear-time, model-agnostic graph\ncondensation algorithm for node classification that outperforms existing\nbaselines across $7$ real-world datasets on accuracy, while being $22$ times\nfaster on average. Bonsai is grounded in rigorous mathematical guarantees on\nthe adopted approximation strategies making it robust to GNN architectures,\ndatasets, and parameters.\n", "date": "2024-10-23", "categories": []}
{"id": "2310.00093", "title": "DataDAM: Efficient Dataset Distillation with Attention Matching", "abstract": "  Researchers have long tried to minimize training costs in deep learning while\nmaintaining strong generalization across diverse datasets. Emerging research on\ndataset distillation aims to reduce training costs by creating a small\nsynthetic set that contains the information of a larger real dataset and\nultimately achieves test accuracy equivalent to a model trained on the whole\ndataset. Unfortunately, the synthetic data generated by previous methods are\nnot guaranteed to distribute and discriminate as well as the original training\ndata, and they incur significant computational costs. Despite promising\nresults, there still exists a significant performance gap between models\ntrained on condensed synthetic sets and those trained on the whole dataset. In\nthis paper, we address these challenges using efficient Dataset Distillation\nwith Attention Matching (DataDAM), achieving state-of-the-art performance while\nreducing training costs. Specifically, we learn synthetic images by matching\nthe spatial attention maps of real and synthetic data generated by different\nlayers within a family of randomly initialized neural networks. Our method\noutperforms the prior methods on several datasets, including CIFAR10/100,\nTinyImageNet, ImageNet-1K, and subsets of ImageNet-1K across most of the\nsettings, and achieves improvements of up to 6.5% and 4.1% on CIFAR100 and\nImageNet-1K, respectively. We also show that our high-quality distilled images\nhave practical benefits for downstream applications, such as continual learning\nand neural architecture search.\n", "date": "2023-09-29", "categories": []}
{"id": "2405.01373", "title": "ATOM: Attention Mixer for Efficient Dataset Distillation", "abstract": "  Recent works in dataset distillation seek to minimize training expenses by\ngenerating a condensed synthetic dataset that encapsulates the information\npresent in a larger real dataset. These approaches ultimately aim to attain\ntest accuracy levels akin to those achieved by models trained on the entirety\nof the original dataset. Previous studies in feature and distribution matching\nhave achieved significant results without incurring the costs of bi-level\noptimization in the distillation process. Despite their convincing efficiency,\nmany of these methods suffer from marginal downstream performance improvements,\nlimited distillation of contextual information, and subpar cross-architecture\ngeneralization. To address these challenges in dataset distillation, we propose\nthe ATtentiOn Mixer (ATOM) module to efficiently distill large datasets using a\nmixture of channel and spatial-wise attention in the feature matching process.\nSpatial-wise attention helps guide the learning process based on consistent\nlocalization of classes in their respective images, allowing for distillation\nfrom a broader receptive field. Meanwhile, channel-wise attention captures the\ncontextual information associated with the class itself, thus making the\nsynthetic image more informative for training. By integrating both types of\nattention, our ATOM module demonstrates superior performance across various\ncomputer vision datasets, including CIFAR10/100 and TinyImagenet. Notably, our\nmethod significantly improves performance in scenarios with a low number of\nimages per class, thereby enhancing its potential. Furthermore, we maintain the\nimprovement in cross-architectures and applications such as neural architecture\nsearch.\n", "date": "2024-05-02", "categories": []}
{"id": "2503.18195", "title": "Shapley-Guided Utility Learning for Effective Graph Inference Data\n  Valuation", "abstract": "  Graph Neural Networks (GNNs) have demonstrated remarkable performance in\nvarious graph-based machine learning tasks, yet evaluating the importance of\nneighbors of testing nodes remains largely unexplored due to the challenge of\nassessing data importance without test labels. To address this gap, we propose\nShapley-Guided Utility Learning (SGUL), a novel framework for graph inference\ndata valuation. SGUL innovatively combines transferable data-specific and\nmodelspecific features to approximate test accuracy without relying on ground\ntruth labels. By incorporating Shapley values as a preprocessing step and using\nfeature Shapley values as input, our method enables direct optimization of\nShapley value prediction while reducing computational demands. SGUL overcomes\nkey limitations of existing methods, including poor generalization to unseen\ntest-time structures and indirect optimization. Experiments on diverse graph\ndatasets demonstrate that SGUL consistently outperforms existing baselines in\nboth inductive and transductive settings. SGUL offers an effective, efficient,\nand interpretable approach for quantifying the value of test-time neighbors.\n", "date": "2025-03-23", "categories": []}
{"id": "2503.18267", "title": "Enhancing Dataset Distillation via Non-Critical Region Refinement", "abstract": "  Dataset distillation has become a popular method for compressing large\ndatasets into smaller, more efficient representations while preserving critical\ninformation for model training. Data features are broadly categorized into two\ntypes: instance-specific features, which capture unique, fine-grained details\nof individual examples, and class-general features, which represent shared,\nbroad patterns across a class. However, previous approaches often struggle to\nbalance these features-some focus solely on class-general patterns, neglecting\nfiner instance details, while others prioritize instance-specific features,\noverlooking the shared characteristics essential for class-level understanding.\nIn this paper, we introduce the Non-Critical Region Refinement Dataset\nDistillation (NRR-DD) method, which preserves instance-specific details and\nfine-grained regions in synthetic data while enriching non-critical regions\nwith class-general information. This approach enables models to leverage all\npixel information, capturing both feature types and enhancing overall\nperformance. Additionally, we present Distance-Based Representative (DBR)\nknowledge transfer, which eliminates the need for soft labels in training by\nrelying on the distance between synthetic data predictions and one-hot encoded\nlabels. Experimental results show that NRR-DD achieves state-of-the-art\nperformance on both small- and large-scale datasets. Furthermore, by storing\nonly two distances per instance, our method delivers comparable results across\nvarious settings. The code is available at\nhttps://github.com/tmtuan1307/NRR-DD.\n", "date": "2025-03-24", "categories": []}
{"id": "2503.18872", "title": "Curriculum Coarse-to-Fine Selection for High-IPC Dataset Distillation", "abstract": "  Dataset distillation (DD) excels in synthesizing a small number of images per\nclass (IPC) but struggles to maintain its effectiveness in high-IPC settings.\nRecent works on dataset distillation demonstrate that combining distilled and\nreal data can mitigate the effectiveness decay. However, our analysis of the\ncombination paradigm reveals that the current one-shot and independent\nselection mechanism induces an incompatibility issue between distilled and real\nimages. To address this issue, we introduce a novel curriculum coarse-to-fine\nselection (CCFS) method for efficient high-IPC dataset distillation. CCFS\nemploys a curriculum selection framework for real data selection, where we\nleverage a coarse-to-fine strategy to select appropriate real data based on the\ncurrent synthetic dataset in each curriculum. Extensive experiments validate\nCCFS, surpassing the state-of-the-art by +6.6\\% on CIFAR-10, +5.8\\% on\nCIFAR-100, and +3.4\\% on Tiny-ImageNet under high-IPC settings. Notably, CCFS\nachieves 60.2\\% test accuracy on ResNet-18 with a 20\\% compression ratio of\nTiny-ImageNet, closely matching full-dataset training with only 0.3\\%\ndegradation. Code: https://github.com/CYDaaa30/CCFS.\n", "date": "2025-03-24", "categories": []}
{"id": "2410.18639", "title": "Diffusion Attribution Score: Evaluating Training Data Influence in\n  Diffusion Models", "abstract": "  As diffusion models become increasingly popular, the misuse of copyrighted\nand private images has emerged as a major concern. One promising solution to\nmitigate this issue is identifying the contribution of specific training\nsamples in generative models, a process known as data attribution. Existing\ndata attribution methods for diffusion models typically quantify the\ncontribution of a training sample by evaluating the change in diffusion loss\nwhen the sample is included or excluded from the training process. However, we\nargue that the direct usage of diffusion loss cannot represent such a\ncontribution accurately due to the calculation of diffusion loss. Specifically,\nthese approaches measure the divergence between predicted and ground truth\ndistributions, which leads to an indirect comparison between the predicted\ndistributions and cannot represent the variances between model behaviors. To\naddress these issues, we aim to measure the direct comparison between predicted\ndistributions with an attribution score to analyse the training sample\nimportance, which is achieved by Diffusion Attribution Score (\\textit{DAS}).\nUnderpinned by rigorous theoretical analysis, we elucidate the effectiveness of\nDAS. Additionally, we explore strategies to accelerate DAS calculations,\nfacilitating its application to large-scale diffusion models. Our extensive\nexperiments across various datasets and diffusion models demonstrate that DAS\nsignificantly surpasses previous benchmarks in terms of the linear\ndata-modelling score, establishing new state-of-the-art performance. Code is\navailable at \\hyperlink{here}{https://github.com/Jinxu-Lin/DAS}.\n", "date": "2024-10-24", "categories": []}
{"id": "2403.16952", "title": "Data Mixing Laws: Optimizing Data Mixtures by Predicting Language\n  Modeling Performance", "abstract": "  Pretraining data of large language models composes multiple domains (e.g.,\nweb texts, academic papers, codes), whose mixture proportions crucially impact\nthe competence of outcome models. While existing endeavors rely on heuristics\nor qualitative strategies to tune the proportions, we discover the quantitative\npredictability of model performance regarding the mixture proportions in\nfunction forms, which we refer to as the data mixing laws. Fitting such\nfunctions on sample mixtures unveils model performance on unseen mixtures\nbefore actual runs, thus guiding the selection of an ideal data mixture.\nFurthermore, we propose nested use of the scaling laws of training steps, model\nsizes, and our data mixing law to enable predicting the performance of large\nmodels trained on massive data under various mixtures with only small-scale\ntraining. Moreover, experimental results verify that our method effectively\noptimizes the training mixture of a 1B model trained for 100B tokens in\nRedPajama, reaching a performance comparable to the one trained for 48% more\nsteps on the default mixture. Extending the application of data mixing laws to\ncontinual training accurately predicts the critical mixture proportion that\navoids catastrophic forgetting and outlooks the potential for dynamic data\nschedules\n", "date": "2024-03-25", "categories": []}
{"id": "2302.04391", "title": "The Re-Label Method For Data-Centric Machine Learning", "abstract": "  In industry deep learning application, our manually labeled data has a\ncertain number of noisy data. To solve this problem and achieve more than 90\nscore in dev dataset, we present a simple method to find the noisy data and\nre-label the noisy data by human, given the model predictions as references in\nhuman labeling. In this paper, we illustrate our idea for a broad set of deep\nlearning tasks, includes classification, sequence tagging, object detection,\nsequence generation, click-through rate prediction. The dev dataset evaluation\nresults and human evaluation results verify our idea.\n", "date": "2023-02-09", "categories": []}
{"id": "2311.01434", "title": "Tailoring Mixup to Data for Calibration", "abstract": "  Among all data augmentation techniques proposed so far, linear interpolation\nof training samples, also called Mixup, has found to be effective for a large\npanel of applications. Along with improved predictive performance, Mixup is\nalso a good technique for improving calibration. However, mixing data\ncarelessly can lead to manifold mismatch, i.e., synthetic data lying outside\noriginal class manifolds, which can deteriorate calibration. In this work, we\nshow that the likelihood of assigning a wrong label with mixup increases with\nthe distance between data to mix. To this end, we propose to dynamically change\nthe underlying distributions of interpolation coefficients depending on the\nsimilarity between samples to mix, and define a flexible framework to do so\nwithout losing in diversity. We provide extensive experiments for\nclassification and regression tasks, showing that our proposed method improves\npredictive performance and calibration of models, while being much more\nefficient.\n", "date": "2023-11-02", "categories": []}
{"id": "2406.01130", "title": "SAVA: Scalable Learning-Agnostic Data Valuation", "abstract": "  Selecting data for training machine learning models is crucial since large,\nweb-scraped, real datasets contain noisy artifacts that affect the quality and\nrelevance of individual data points. These noisy artifacts will impact model\nperformance. We formulate this problem as a data valuation task, assigning a\nvalue to data points in the training set according to how similar or dissimilar\nthey are to a clean and curated validation set. Recently, LAVA demonstrated the\nuse of optimal transport (OT) between a large noisy training dataset and a\nclean validation set, to value training data efficiently, without the\ndependency on model performance. However, the LAVA algorithm requires the\nentire dataset as an input, this limits its application to larger datasets.\nInspired by the scalability of stochastic (gradient) approaches which carry out\ncomputations on batches of data points instead of the entire dataset, we\nanalogously propose SAVA, a scalable variant of LAVA with its computation on\nbatches of data points. Intuitively, SAVA follows the same scheme as LAVA which\nleverages the hierarchically defined OT for data valuation. However, while LAVA\nprocesses the whole dataset, SAVA divides the dataset into batches of data\npoints, and carries out the OT problem computation on those batches. Moreover,\nour theoretical derivations on the trade-off of using entropic regularization\nfor OT problems include refinements of prior work. We perform extensive\nexperiments, to demonstrate that SAVA can scale to large datasets with millions\nof data points and does not trade off data valuation performance.\n", "date": "2024-06-03", "categories": []}
{"id": "2406.05704", "title": "Hierarchical Features Matter: A Deep Exploration of Progressive\n  Parameterization Method for Dataset Distillation", "abstract": "  Dataset distillation is an emerging dataset reduction method, which condenses\nlarge-scale datasets while maintaining task accuracy. Current parameterization\nmethods achieve enhanced performance under extremely high compression ratio by\noptimizing determined synthetic dataset in informative feature domain. However,\nthey limit themselves to a fixed optimization space for distillation,\nneglecting the diverse guidance across different informative latent spaces. To\novercome this limitation, we propose a novel parameterization method dubbed\nHierarchical Parameterization Distillation (H-PD), to systematically explore\nhierarchical feature within provided feature space (e.g., layers within\npre-trained generative adversarial networks). We verify the correctness of our\ninsights by applying the hierarchical optimization strategy on GAN-based\nparameterization method. In addition, we introduce a novel class-relevant\nfeature distance metric to alleviate the computational burden associated with\nsynthetic dataset evaluation, bridging the gap between synthetic and original\ndatasets. Experimental results demonstrate that the proposed H-PD achieves a\nsignificant performance improvement under various settings with equivalent time\nconsumption, and even surpasses current generative distillation using diffusion\nmodels under extreme compression ratios IPC=1 and IPC=10.\n", "date": "2024-06-09", "categories": []}
{"id": "2408.11852", "title": "Fast Training Dataset Attribution via In-Context Learning", "abstract": "  We investigate the use of in-context learning and prompt engineering to\nestimate the contributions of training data in the outputs of instruction-tuned\nlarge language models (LLMs). We propose two novel approaches: (1) a\nsimilarity-based approach that measures the difference between LLM outputs with\nand without provided context, and (2) a mixture distribution model approach\nthat frames the problem of identifying contribution scores as a matrix\nfactorization task. Our empirical comparison demonstrates that the mixture\nmodel approach is more robust to retrieval noise in in-context learning,\nproviding a more reliable estimation of data contributions.\n", "date": "2024-08-14", "categories": []}
{"id": "2408.14506", "title": "Distilling Long-tailed Datasets", "abstract": "  Dataset distillation aims to synthesize a small, information-rich dataset\nfrom a large one for efficient model training. However, existing dataset\ndistillation methods struggle with long-tailed datasets, which are prevalent in\nreal-world scenarios. By investigating the reasons behind this unexpected\nresult, we identified two main causes: 1) The distillation process on\nimbalanced datasets develops biased gradients, leading to the synthesis of\nsimilarly imbalanced distilled datasets. 2) The experts trained on such\ndatasets perform suboptimally on tail classes, resulting in misguided\ndistillation supervision and poor-quality soft-label initialization. To address\nthese issues, we first propose Distribution-agnostic Matching to avoid directly\nmatching the biased expert trajectories. It reduces the distance between the\nstudent and the biased expert trajectories and prevents the tail class bias\nfrom being distilled to the synthetic dataset. Moreover, we improve the\ndistillation guidance with Expert Decoupling, which jointly matches the\ndecoupled backbone and classifier to improve the tail class performance and\ninitialize reliable soft labels. This work pioneers the field of long-tailed\ndataset distillation, marking the first effective effort to distill long-tailed\ndatasets.\n", "date": "2024-08-24", "categories": []}
{"id": "2503.14559", "title": "Squeeze Out Tokens from Sample for Finer-Grained Data Governance", "abstract": "  Widely observed data scaling laws, in which error falls off as a power of the\ntraining size, demonstrate the diminishing returns of unselective data\nexpansion. Hence, data governance is proposed to downsize datasets through\npruning non-informative samples. Yet, isolating the impact of a specific sample\non overall model performance is challenging, due to the vast computation\nrequired for tryout all sample combinations. Current data governors circumvent\nthis complexity by estimating sample contributions through heuristic-derived\nscalar scores, thereby discarding low-value ones. Despite thorough sample\nsieving, retained samples contain substantial undesired tokens intrinsically,\nunderscoring the potential for further compression and purification. In this\nwork, we upgrade data governance from a 'sieving' approach to a 'juicing' one.\nInstead of scanning for least-flawed samples, our dual-branch DataJuicer\napplies finer-grained intra-sample governance. It squeezes out informative\ntokens and boosts image-text alignments. Specifically, the vision branch\nretains salient image patches and extracts relevant object classes, while the\ntext branch incorporates these classes to enhance captions. Consequently,\nDataJuicer yields more refined datasets through finer-grained governance.\nExtensive experiments across datasets demonstrate that DataJuicer significantly\noutperforms existing DataSieve in image-text retrieval, classification, and\ndense visual reasoning.\n", "date": "2025-03-18", "categories": []}
{"id": "2503.13935", "title": "SCORE: Soft Label Compression-Centric Dataset Condensation via Coding\n  Rate Optimization", "abstract": "  Dataset Condensation (DC) aims to obtain a condensed dataset that allows\nmodels trained on the condensed dataset to achieve performance comparable to\nthose trained on the full dataset. Recent DC approaches increasingly focus on\nencoding knowledge into realistic images with soft labeling, for their\nscalability to ImageNet-scale datasets and strong capability of cross-domain\ngeneralization. However, this strong performance comes at a substantial storage\ncost which could significantly exceed the storage cost of the original dataset.\nWe argue that the three key properties to alleviate this performance-storage\ndilemma are informativeness, discriminativeness, and compressibility of the\ncondensed data. Towards this end, this paper proposes a \\textbf{S}oft label\ncompression-centric dataset condensation framework using \\textbf{CO}ding\n\\textbf{R}at\\textbf{E} (SCORE). SCORE formulates dataset condensation as a\nmin-max optimization problem, which aims to balance the three key properties\nfrom an information-theoretic perspective. In particular, we theoretically\ndemonstrate that our coding rate-inspired objective function is submodular, and\nits optimization naturally enforces low-rank structure in the soft label set\ncorresponding to each condensed data. Extensive experiments on large-scale\ndatasets, including ImageNet-1K and Tiny-ImageNet, demonstrate that SCORE\noutperforms existing methods in most cases. Even with 30$\\times$ compression of\nsoft labels, performance decreases by only 5.5\\% and 2.7\\% for ImageNet-1K with\nIPC 10 and 50, respectively. Code will be released upon paper acceptance.\n", "date": "2025-03-18", "categories": []}
{"id": "2503.14112", "title": "Condensing Action Segmentation Datasets via Generative Network Inversion", "abstract": "  This work presents the first condensation approach for procedural video\ndatasets used in temporal action segmentation. We propose a condensation\nframework that leverages generative prior learned from the dataset and network\ninversion to condense data into compact latent codes with significant storage\nreduced across temporal and channel aspects. Orthogonally, we propose sampling\ndiverse and representative action sequences to minimize video-wise redundancy.\nOur evaluation on standard benchmarks demonstrates consistent effectiveness in\ncondensing TAS datasets and achieving competitive performances. Specifically,\non the Breakfast dataset, our approach reduces storage by over 500$\\times$\nwhile retaining 83% of the performance compared to training with the full\ndataset. Furthermore, when applied to a downstream incremental learning task,\nit yields superior performance compared to the state-of-the-art.\n", "date": "2025-03-18", "categories": []}
{"id": "2503.12069", "title": "Robust Dataset Distillation by Matching Adversarial Trajectories", "abstract": "  Dataset distillation synthesizes compact datasets that enable models to\nachieve performance comparable to training on the original large-scale\ndatasets. However, existing distillation methods overlook the robustness of the\nmodel, resulting in models that are vulnerable to adversarial attacks when\ntrained on distilled data. To address this limitation, we introduce the task of\n``robust dataset distillation\", a novel paradigm that embeds adversarial\nrobustness into the synthetic datasets during the distillation process. We\npropose Matching Adversarial Trajectories (MAT), a method that integrates\nadversarial training into trajectory-based dataset distillation. MAT\nincorporates adversarial samples during trajectory generation to obtain robust\ntraining trajectories, which are then used to guide the distillation process.\nAs experimentally demonstrated, even through natural training on our distilled\ndataset, models can achieve enhanced adversarial robustness while maintaining\ncompetitive accuracy compared to existing distillation methods. Our work\nhighlights robust dataset distillation as a new and important research\ndirection and provides a strong baseline for future research to bridge the gap\nbetween efficient training and adversarial robustness.\n", "date": "2025-03-15", "categories": []}
{"id": "2503.13383", "title": "Cream of the Crop: Harvesting Rich, Scalable and Transferable\n  Multi-Modal Data for Instruction Fine-Tuning", "abstract": "  The hypothesis that pretrained large language models (LLMs) necessitate only\nminimal supervision during the fine-tuning (SFT) stage (Zhou et al., 2024) has\nbeen substantiated by recent advancements in data curation and selection\nresearch. However, their stability and generalizability are compromised due to\nthe vulnerability to experimental setups and validation protocols, falling\nshort of surpassing random sampling (Diddee & Ippolito, 2024; Xia et al.,\n2024b). Built upon LLMs, multi-modal LLMs (MLLMs), combined with the sheer\ntoken volume and heightened heterogeneity of data sources, amplify both the\nsignificance and complexity of data selection.\n  To harvest multi-modal instructional data in a robust and efficient manner,\nwe re-define the granularity of the quality metric by decomposing it into 14\nvision-language-related capabilities, and introduce multi-modal rich scorers to\nevaluate the capabilities of each data candidate. To promote diversity, in\nlight of the inherent objective of the alignment stage, we take interaction\nstyle as diversity indicator and use a multi-modal rich styler to identify data\ninstruction patterns. In doing so, our multi-modal rich scorers and styler\n(mmSSR) guarantee that high-scoring information is conveyed to users in\ndiversified forms. Free from embedding-based clustering or greedy sampling,\nmmSSR efficiently scales to millions of data with varying budget constraints,\nsupports customization for general or specific capability acquisition, and\nfacilitates training-free generalization to new domains for curation. Across\n10+ experimental settings, validated by 14 multi-modal benchmarks, we\ndemonstrate consistent improvements over random sampling, baseline strategies\nand state-of-the-art selection methods, achieving 99.1% of full performance\nwith only 30% of the 2.6M data.\n", "date": "2025-03-17", "categories": []}
{"id": "2312.05327", "title": "Better, Not Just More: Data-Centric Machine Learning for Earth\n  Observation", "abstract": "  Recent developments and research in modern machine learning have led to\nsubstantial improvements in the geospatial field. Although numerous deep\nlearning architectures and models have been proposed, the majority of them have\nbeen solely developed on benchmark datasets that lack strong real-world\nrelevance. Furthermore, the performance of many methods has already saturated\non these datasets. We argue that a shift from a model-centric view to a\ncomplementary data-centric perspective is necessary for further improvements in\naccuracy, generalization ability, and real impact on end-user applications.\nFurthermore, considering the entire machine learning cycle-from problem\ndefinition to model deployment with feedback-is crucial for enhancing machine\nlearning models that can be reliable in unforeseen situations. This work\npresents a definition as well as a precise categorization and overview of\nautomated data-centric learning approaches for geospatial data. It highlights\nthe complementary role of data-centric learning with respect to model-centric\nin the larger machine learning deployment cycle. We review papers across the\nentire geospatial field and categorize them into different groups. A set of\nrepresentative experiments shows concrete implementation examples. These\nexamples provide concrete steps to act on geospatial data with data-centric\nmachine learning approaches.\n", "date": "2023-12-08", "categories": []}
{"id": "2503.10687", "title": "Context-guided Responsible Data Augmentation with Diffusion Models", "abstract": "  Generative diffusion models offer a natural choice for data augmentation when\ntraining complex vision models. However, ensuring reliability of their\ngenerative content as augmentation samples remains an open challenge. Despite a\nnumber of techniques utilizing generative images to strengthen model training,\nit remains unclear how to utilize the combination of natural and generative\nimages as a rich supervisory signal for effective model induction. In this\nregard, we propose a text-to-image (T2I) data augmentation method, named\nDiffCoRe-Mix, that computes a set of generative counterparts for a training\nsample with an explicitly constrained diffusion model that leverages\nsample-based context and negative prompting for a reliable augmentation sample\ngeneration. To preserve key semantic axes, we also filter out undesired\ngenerative samples in our augmentation process. To that end, we propose a\nhard-cosine filtration in the embedding space of CLIP. Our approach\nsystematically mixes the natural and generative images at pixel and patch\nlevels. We extensively evaluate our technique on ImageNet-1K,Tiny ImageNet-200,\nCIFAR-100, Flowers102, CUB-Birds, Stanford Cars, and Caltech datasets,\ndemonstrating a notable increase in performance across the board, achieving up\nto $\\sim 3\\%$ absolute gain for top-1 accuracy over the state-of-the-art\nmethods, while showing comparable computational overhead. Our code is publicly\navailable at https://github.com/khawar-islam/DiffCoRe-Mix\n", "date": "2025-03-12", "categories": []}
{"id": "2503.10698", "title": "Ordered Semantically Diverse Sampling for Textual Data", "abstract": "  The goal of diversity sampling is to select a representative subset of data\nin a way that maximizes information contained in the subset while keeping its\ncardinality small. We introduce the ordered diverse sampling problem based on a\nnew metric that measures the diversity in an ordered list of samples. We\npresent a novel approach for generating ordered diverse samples for textual\ndata that uses principal components on the embedding vectors. The proposed\napproach is simple and compared with existing approaches using the new metric.\nWe transform standard text classification benchmarks into benchmarks for\nordered diverse sampling. Our empirical evaluation shows that prevailing\napproaches perform 6% to 61% worse than our method while also being more time\ninefficient. Ablation studies show how the parts of the new approach contribute\nto the overall metrics.\n", "date": "2025-03-12", "categories": []}
{"id": "2410.06215", "title": "DataEnvGym: Data Generation Agents in Teacher Environments with Student\n  Feedback", "abstract": "  The process of creating training data to teach models is currently driven by\nhumans, who manually analyze model weaknesses and plan how to create data that\nimproves a student model. Approaches using LLMs as annotators reduce human\neffort, but still require humans to interpret feedback from evaluations and\ncontrol the LLM to produce data the student needs. Automating this\nlabor-intensive process by creating autonomous data generation agents - or\nteachers - is desirable, but requires environments that can simulate the\nfeedback-driven, iterative, closed loop of data creation. To enable rapid,\nscalable testing for such agents and their modules, we introduce DataEnvGym, a\ntestbed of teacher environments for data generation agents. DataEnvGym frames\ndata generation as a sequential decision-making task, involving an agent\nconsisting of a data generation policy (which generates a plan for creating\ntraining data) and a data generation engine (which transforms the plan into\ndata), inside an environment that provides student feedback. The agent's goal\nis to improve student performance. Students are iteratively trained and\nevaluated on generated data, and their feedback (in the form of errors or weak\nskills) is reported to the agent after each iteration. DataEnvGym includes\nmultiple teacher environment instantiations across 3 levels of structure in the\nstate representation and action space. More structured environments are based\non inferred skills and offer more interpretability and curriculum control. We\nsupport 4 domains (math, code, VQA, and tool-use) and test multiple students\nand teachers. Example agents in our teaching environments can iteratively\nimprove students across tasks and settings. Moreover, we show that environments\nteach different skill levels and test variants of key modules, pointing to\nfuture work in improving data generation agents, engines, and feedback\nmechanisms.\n", "date": "2024-10-08", "categories": []}
{"id": "2412.21197", "title": "A Large-Scale Study on Video Action Dataset Condensation", "abstract": "  Recently, dataset condensation has made significant progress in the image\ndomain. Unlike images, videos possess an additional temporal dimension, which\nharbors considerable redundant information, making condensation even more\ncrucial. However, video dataset condensation still remains an underexplored\narea. We aim to bridge this gap by providing a large-scale study with\nsystematic design and fair comparison. Specifically, our work delves into three\nkey aspects to provide valuable empirical insights: (1) temporal processing of\nvideo data, (2) the evaluation protocol for video dataset condensation, and (3)\nadaptation of condensation algorithms to the space-time domain. From this\nstudy, we derive several intriguing observations: (i) labeling methods greatly\ninfluence condensation performance, (ii) simple sliding-window sampling is\neffective for temporal processing, and (iii) dataset distillation methods\nperform better in challenging scenarios, while sample selection methods excel\nin easier ones. Furthermore, we propose a unified evaluation protocol for the\nfair comparison of different condensation algorithms and achieve\nstate-of-the-art results on four widely-used action recognition datasets:\nHMDB51, UCF101, SSv2 and K400. Our code is available at\nhttps://github.com/MCG-NJU/Video-DC.\n", "date": "2024-12-30", "categories": []}
{"id": "2503.07998", "title": "Efficient Dataset Distillation through Low-Rank Space Sampling", "abstract": "  Huge amount of data is the key of the success of deep learning, however,\nredundant information impairs the generalization ability of the model and\nincreases the burden of calculation. Dataset Distillation (DD) compresses the\noriginal dataset into a smaller but representative subset for high-quality data\nand efficient training strategies. Existing works for DD generate synthetic\nimages by treating each image as an independent entity, thereby overlooking the\ncommon features among data. This paper proposes a dataset distillation method\nbased on Matching Training Trajectories with Low-rank Space Sampling(MTT-LSS),\nwhich uses low-rank approximations to capture multiple low-dimensional manifold\nsubspaces of the original data. The synthetic data is represented by basis\nvectors and shared dimension mappers from these subspaces, reducing the cost of\ngenerating individual data points while effectively minimizing information\nredundancy. The proposed method is tested on CIFAR-10, CIFAR-100, and SVHN\ndatasets, and outperforms the baseline methods by an average of 9.9%.\n", "date": "2025-03-11", "categories": []}
{"id": "2312.13440", "title": "MGAug: Multimodal Geometric Augmentation in Latent Spaces of Image\n  Deformations", "abstract": "  Geometric transformations have been widely used to augment the size of\ntraining images. Existing methods often assume a unimodal distribution of the\nunderlying transformations between images, which limits their power when data\nwith multimodal distributions occur. In this paper, we propose a novel model,\nMultimodal Geometric Augmentation (MGAug), that for the first time generates\naugmenting transformations in a multimodal latent space of geometric\ndeformations. To achieve this, we first develop a deep network that embeds the\nlearning of latent geometric spaces of diffeomorphic transformations (a.k.a.\ndiffeomorphisms) in a variational autoencoder (VAE). A mixture of multivariate\nGaussians is formulated in the tangent space of diffeomorphisms and serves as a\nprior to approximate the hidden distribution of image transformations. We then\naugment the original training dataset by deforming images using randomly\nsampled transformations from the learned multimodal latent space of VAE. To\nvalidate the efficiency of our model, we jointly learn the augmentation\nstrategy with two distinct domain-specific tasks: multi-class classification on\n2D synthetic datasets and segmentation on real 3D brain magnetic resonance\nimages (MRIs). We also compare MGAug with state-of-the-art transformation-based\nimage augmentation algorithms. Experimental results show that our proposed\napproach outperforms all baselines by significantly improved prediction\naccuracy. Our code is publicly available at\nhttps://github.com/tonmoy-hossain/MGAug.\n", "date": "2023-12-20", "categories": []}
{"id": "2409.05816", "title": "Improving Pretraining Data Using Perplexity Correlations", "abstract": "  Quality pretraining data is often seen as the key to high-performance\nlanguage models. However, progress in understanding pretraining data has been\nslow due to the costly pretraining runs required for data selection\nexperiments. We present a framework that avoids these costs and selects\nhigh-quality pretraining data without any LLM training of our own. Our work is\nbased on a simple observation: LLM losses on many pretraining texts are\ncorrelated with downstream benchmark performance, and selecting\nhigh-correlation documents is an effective pretraining data selection method.\nWe build a new statistical framework for data selection centered around\nestimates of perplexity-benchmark correlations and perform data selection using\na sample of 90 LLMs taken from the Open LLM Leaderboard on texts from tens of\nthousands of web domains. In controlled pretraining experiments at the 160M\nparameter scale on 8 benchmarks, our approach outperforms DSIR on every\nbenchmark, while matching the best data selector found in DataComp-LM, a\nhand-engineered bigram classifier. We have now also updated this paper to\ninclude results from preregistered experiments with new pretraining data on an\naggregation of 22 benchmarks up to the 1.4B scale, showing increasing\nimprovements of our method over others with more scale. A pip package with full\ndocumentation can be found here:\nhttps://github.com/TristanThrush/perplexity-correlations.\n", "date": "2024-09-09", "categories": []}
{"id": "2503.04835", "title": "Distilling Dataset into Neural Field", "abstract": "  Utilizing a large-scale dataset is essential for training high-performance\ndeep learning models, but it also comes with substantial computation and\nstorage costs. To overcome these challenges, dataset distillation has emerged\nas a promising solution by compressing the large-scale dataset into a smaller\nsynthetic dataset that retains the essential information needed for training.\nThis paper proposes a novel parameterization framework for dataset\ndistillation, coined Distilling Dataset into Neural Field (DDiF), which\nleverages the neural field to store the necessary information of the\nlarge-scale dataset. Due to the unique nature of the neural field, which takes\ncoordinates as input and output quantity, DDiF effectively preserves the\ninformation and easily generates various shapes of data. We theoretically\nconfirm that DDiF exhibits greater expressiveness than some previous literature\nwhen the utilized budget for a single synthetic instance is the same. Through\nextensive experiments, we demonstrate that DDiF achieves superior performance\non several benchmark datasets, extending beyond the image domain to include\nvideo, audio, and 3D voxel. We release the code at\nhttps://github.com/aailab-kaist/DDiF.\n", "date": "2025-03-05", "categories": []}
{"id": "2503.05047", "title": "Biases in Large Language Model-Elicited Text: A Case Study in Natural\n  Language Inference", "abstract": "  We test whether NLP datasets created with Large Language Models (LLMs)\ncontain annotation artifacts and social biases like NLP datasets elicited from\ncrowd-source workers. We recreate a portion of the Stanford Natural Language\nInference corpus using GPT-4, Llama-2 70b for Chat, and Mistral 7b Instruct. We\ntrain hypothesis-only classifiers to determine whether LLM-elicited NLI\ndatasets contain annotation artifacts. Next, we use pointwise mutual\ninformation to identify the words in each dataset that are associated with\ngender, race, and age-related terms. On our LLM-generated NLI datasets,\nfine-tuned BERT hypothesis-only classifiers achieve between 86-96% accuracy.\nOur analyses further characterize the annotation artifacts and stereotypical\nbiases in LLM-generated datasets.\n", "date": "2025-03-06", "categories": []}
{"id": "2405.14736", "title": "GIFT: Unlocking Full Potential of Labels in Distilled Dataset at\n  Near-zero Cost", "abstract": "  Recent advancements in dataset distillation have demonstrated the significant\nbenefits of employing soft labels generated by pre-trained teacher models. In\nthis paper, we introduce a novel perspective by emphasizing the full\nutilization of labels. We first conduct a comprehensive comparison of various\nloss functions for soft label utilization in dataset distillation, revealing\nthat the model trained on the synthetic dataset exhibits high sensitivity to\nthe choice of loss function for soft label utilization. This finding highlights\nthe necessity of a universal loss function for training models on synthetic\ndatasets. Building on these insights, we introduce an extremely simple yet\nsurprisingly effective plug-and-play approach, GIFT, which encompasses soft\nlabel refinement and a cosine similarity-based loss function to efficiently\nleverage full label information. Extensive experiments indicate that GIFT\nconsistently enhances state-of-the-art dataset distillation methods across\nvarious dataset scales, without incurring additional computational costs.\nImportantly, GIFT significantly enhances cross-optimizer generalization, an\narea previously overlooked. For instance, on ImageNet-1K with IPC = 10, GIFT\nenhances the state-of-the-art method RDED by 30.8% in cross-optimizer\ngeneralization. Our code is available at https://github.com/LINs-lab/GIFT.\n", "date": "2024-05-23", "categories": []}
{"id": "2503.04357", "title": "scDD: Latent Codes Based scRNA-seq Dataset Distillation with Foundation\n  Model Knowledge", "abstract": "  Single-cell RNA sequencing (scRNA-seq) technology has profiled hundreds of\nmillions of human cells across organs, diseases, development and perturbations\nto date. However, the high-dimensional sparsity, batch effect noise, category\nimbalance, and ever-increasing data scale of the original sequencing data pose\nsignificant challenges for multi-center knowledge transfer, data fusion, and\ncross-validation between scRNA-seq datasets. To address these barriers, (1) we\nfirst propose a latent codes-based scRNA-seq dataset distillation framework\nnamed scDD, which transfers and distills foundation model knowledge and\noriginal dataset information into a compact latent space and generates\nsynthetic scRNA-seq dataset by a generator to replace the original dataset.\nThen, (2) we propose a single-step conditional diffusion generator named SCDG,\nwhich perform single-step gradient back-propagation to help scDD optimize\ndistillation quality and avoid gradient decay caused by multi-step\nback-propagation. Meanwhile, SCDG ensures the scRNA-seq data characteristics\nand inter-class discriminability of the synthetic dataset through flexible\nconditional control and generation quality assurance. Finally, we propose a\ncomprehensive benchmark to evaluate the performance of scRNA-seq dataset\ndistillation in different data analysis tasks. It is validated that our\nproposed method can achieve 7.61% absolute and 15.70% relative improvement over\nprevious state-of-the-art methods on average task.\n", "date": "2025-03-06", "categories": []}
{"id": "2408.06927", "title": "Breaking Class Barriers: Efficient Dataset Distillation via Inter-Class\n  Feature Compensator", "abstract": "  Dataset distillation has emerged as a technique aiming to condense\ninformative features from large, natural datasets into a compact and synthetic\nform. While recent advancements have refined this technique, its performance is\nbottlenecked by the prevailing class-specific synthesis paradigm. Under this\nparadigm, synthetic data is optimized exclusively for a pre-assigned one-hot\nlabel, creating an implicit class barrier in feature condensation. This leads\nto inefficient utilization of the distillation budget and oversight of\ninter-class feature distributions, which ultimately limits the effectiveness\nand efficiency, as demonstrated in our analysis. To overcome these constraints,\nthis paper presents the Inter-class Feature Compensator (INFER), an innovative\ndistillation approach that transcends the class-specific data-label framework\nwidely utilized in current dataset distillation methods. Specifically, INFER\nleverages a Universal Feature Compensator (UFC) to enhance feature integration\nacross classes, enabling the generation of multiple additional synthetic\ninstances from a single UFC input. This significantly improves the efficiency\nof the distillation budget. Moreover, INFER enriches inter-class interactions\nduring the distillation, thereby enhancing the effectiveness and\ngeneralizability of the distilled data. By allowing for the linear\ninterpolation of labels similar to those in the original dataset, INFER\nmeticulously optimizes the synthetic data and dramatically reduces the size of\nsoft labels in the synthetic dataset to almost zero, establishing a new\nbenchmark for efficiency and effectiveness in dataset distillation. In\npractice, INFER demonstrates state-of-the-art performance across benchmark\ndatasets. For instance, in the ipc = 50 setting on ImageNet-1k with the same\ncompression level, it outperforms SRe2L by 34.5% using ResNet18.\n", "date": "2024-08-13", "categories": []}
{"id": "2412.03881", "title": "Weak-to-Strong Generalization Through the Data-Centric Lens", "abstract": "  The weak-to-strong generalization phenomenon is the driver for important\nmachine learning applications including highly data-efficient learning and,\nmost recently, performing superalignment. While decades of research have\nresulted in numerous algorithms that produce strong empirical performance,\nunderstanding what aspects of data enable weak-to-strong generalization has\nbeen understudied. We propose a simple data-centric mechanism that\ncharacterizes weak-to-strong generalization: the overlap density. Intuitively,\ngeneralization tracks the number of points that contain overlaps, i.e., both\neasy patterns (learnable by a weak model) and challenging patterns (only\nlearnable by a stronger model), as with such points, weak predictions can be\nused to learn challenging patterns by stronger models. We provide a practical\noverlap detection algorithm to find such points in datasets and leverage them\nto learn, among multiple sources of data, which to query when seeking to\nmaximize overlap density and thereby enhance weak-to-strong generalization. We\npresent a theoretical result showing that the generalization benefit is a\nfunction of the overlap density and a regret bound for our data selection\nalgorithm. Empirically, we validate the mechanism and the overlap detection\nalgorithm on a wide array of settings.\n", "date": "2024-12-05", "categories": []}
{"id": "2403.08632", "title": "A Decade's Battle on Dataset Bias: Are We There Yet?", "abstract": "  We revisit the \"dataset classification\" experiment suggested by Torralba &\nEfros (2011) a decade ago, in the new era with large-scale, diverse, and\nhopefully less biased datasets as well as more capable neural network\narchitectures. Surprisingly, we observe that modern neural networks can achieve\nexcellent accuracy in classifying which dataset an image is from: e.g., we\nreport 84.7% accuracy on held-out validation data for the three-way\nclassification problem consisting of the YFCC, CC, and DataComp datasets. Our\nfurther experiments show that such a dataset classifier could learn semantic\nfeatures that are generalizable and transferable, which cannot be explained by\nmemorization. We hope our discovery will inspire the community to rethink\nissues involving dataset bias.\n", "date": "2024-03-13", "categories": []}
{"id": "2409.20164", "title": "Erase, then Redraw: A Novel Data Augmentation Approach for Free Space\n  Detection Using Diffusion Model", "abstract": "  Data augmentation is one of the most common tools in deep learning,\nunderpinning many recent advances including tasks such as classification,\ndetection, and semantic segmentation. The standard approach to data\naugmentation involves simple transformations like rotation and flipping to\ngenerate new images. However, these new images often lack diversity along the\nmain semantic dimensions within the data. Traditional data augmentation methods\ncannot alter high-level semantic attributes such as the presence of vehicles,\ntrees, and buildings in a scene to enhance data diversity. In recent years, the\nrapid development of generative models has injected new vitality into the field\nof data augmentation. In this paper, we address the lack of diversity in data\naugmentation for road detection task by using a pre-trained text-to-image\ndiffusion model to parameterize image-to-image transformations. Our method\ninvolves editing images using these diffusion models to change their semantics.\nIn essence, we achieve this goal by erasing instances of real objects from the\noriginal dataset and generating new instances with similar semantics in the\nerased regions using the diffusion model, thereby expanding the original\ndataset. We evaluate our approach on the KITTI road dataset and achieve the\nbest results compared to other data augmentation methods, which demonstrates\nthe effectiveness of our proposed development.\n", "date": "2024-09-30", "categories": []}
{"id": "2410.02268", "title": "Structural-Entropy-Based Sample Selection for Efficient and Effective\n  Learning", "abstract": "  Sample selection improves the efficiency and effectiveness of machine\nlearning models by providing informative and representative samples. Typically,\nsamples can be modeled as a sample graph, where nodes are samples and edges\nrepresent their similarities. Most existing methods are based on local\ninformation, such as the training difficulty of samples, thereby overlooking\nglobal information, such as connectivity patterns. This oversight can result in\nsuboptimal selection because global information is crucial for ensuring that\nthe selected samples well represent the structural properties of the graph. To\naddress this issue, we employ structural entropy to quantify global information\nand losslessly decompose it from the whole graph to individual nodes using the\nShapley value. Based on the decomposition, we present\n$\\textbf{S}$tructural-$\\textbf{E}$ntropy-based sample $\\textbf{S}$election\n($\\textbf{SES}$), a method that integrates both global and local information to\nselect informative and representative samples. SES begins by constructing a\n$k$NN-graph among samples based on their similarities. It then measures sample\nimportance by combining structural entropy (global metric) with training\ndifficulty (local metric). Finally, SES applies importance-biased blue noise\nsampling to select a set of diverse and representative samples. Comprehensive\nexperiments on three learning scenarios -- supervised learning, active\nlearning, and continual learning -- clearly demonstrate the effectiveness of\nour method.\n", "date": "2024-10-03", "categories": []}
{"id": "2503.01836", "title": "CrowdSelect: Synthetic Instruction Data Selection with Multi-LLM Wisdom", "abstract": "  Distilling advanced Large Language Models' instruction-following capabilities\ninto smaller models using a selected subset has become a mainstream approach in\nmodel training. While existing synthetic instruction data selection strategies\nrely mainly on single-dimensional signals (i.e., reward scores, model\nperplexity), they fail to capture the complexity of instruction-following\nacross diverse fields. Therefore, we investigate more diverse signals to\ncapture comprehensive instruction-response pair characteristics and propose\nthree foundational metrics that leverage Multi-LLM wisdom, informed by (1)\ndiverse LLM responses and (2) reward model assessment. Building upon base\nmetrics, we propose CrowdSelect, an integrated metric incorporating a\nclustering-based approach to maintain response diversity. Our comprehensive\nexperiments demonstrate that our foundation metrics consistently improve\nperformance across 4 base models on MT-bench and Arena-Hard. CrowdSelect,\nefficiently incorporating all metrics, achieves state-of-the-art performance in\nboth Full and LoRA fine-tuning, showing improvements of 4.81% on Arena-Hard and\n11.1% on MT-bench with Llama-3.2-3b-instruct. We hope our findings will bring\nvaluable insights for future research in this direction. Code are available at\nhttps://github.com/listentm/crowdselect.\n", "date": "2025-03-03", "categories": []}
{"id": "2502.20653", "title": "Dataset Distillation with Neural Characteristic Function: A Minmax\n  Perspective", "abstract": "  Dataset distillation has emerged as a powerful approach for reducing data\nrequirements in deep learning. Among various methods, distribution\nmatching-based approaches stand out for their balance of computational\nefficiency and strong performance. However, existing distance metrics used in\ndistribution matching often fail to accurately capture distributional\ndifferences, leading to unreliable measures of discrepancy. In this paper, we\nreformulate dataset distillation as a minmax optimization problem and introduce\nNeural Characteristic Function Discrepancy (NCFD), a comprehensive and\ntheoretically grounded metric for measuring distributional differences. NCFD\nleverages the Characteristic Function (CF) to encapsulate full distributional\ninformation, employing a neural network to optimize the sampling strategy for\nthe CF's frequency arguments, thereby maximizing the discrepancy to enhance\ndistance estimation. Simultaneously, we minimize the difference between real\nand synthetic data under this optimized NCFD measure. Our approach, termed\nNeural Characteristic Function Matching (\\mymethod{}), inherently aligns the\nphase and amplitude of neural features in the complex plane for both real and\nsynthetic data, achieving a balance between realism and diversity in synthetic\nsamples. Experiments demonstrate that our method achieves significant\nperformance gains over state-of-the-art methods on both low- and\nhigh-resolution datasets. Notably, we achieve a 20.5\\% accuracy boost on\nImageSquawk. Our method also reduces GPU memory usage by over 300$\\times$ and\nachieves 20$\\times$ faster processing speeds compared to state-of-the-art\nmethods. To the best of our knowledge, this is the first work to achieve\nlossless compression of CIFAR-100 on a single NVIDIA 2080 Ti GPU using only 2.3\nGB of memory.\n", "date": "2025-02-28", "categories": []}
{"id": "2502.19917", "title": "Picking the Cream of the Crop: Visual-Centric Data Selection with\n  Collaborative Agents", "abstract": "  To improve Multimodal Large Language Models' (MLLMs) ability to process\nimages and complex instructions, researchers predominantly curate large-scale\nvisual instruction tuning datasets, which are either sourced from existing\nvision tasks or synthetically generated using LLMs and image descriptions.\nHowever, they often suffer from critical flaws, including misaligned\ninstruction-image pairs and low-quality images. Such issues hinder training\nefficiency and limit performance improvements, as models waste resources on\nnoisy or irrelevant data with minimal benefit to overall capability. To address\nthis issue, we propose a \\textbf{Vi}sual-Centric \\textbf{S}election approach\nvia \\textbf{A}gents Collaboration (ViSA), which centers on image quality\nassessment and image-instruction relevance evaluation. Specifically, our\napproach consists of 1) an image information quantification method via visual\nagents collaboration to select images with rich visual information, and 2) a\nvisual-centric instruction quality assessment method to select high-quality\ninstruction data related to high-quality images. Finally, we reorganize 80K\ninstruction data from large open-source datasets. Extensive experiments\ndemonstrate that ViSA outperforms or is comparable to current state-of-the-art\nmodels on seven benchmarks, using only 2.5\\% of the original data, highlighting\nthe efficiency of our data selection approach. Moreover, we conduct ablation\nstudies to validate the effectiveness of each component of our method. The code\nis available at https://github.com/HITsz-TMG/ViSA.\n", "date": "2025-02-27", "categories": []}
{"id": "2501.18578", "title": "R.I.P.: Better Models by Survival of the Fittest Prompts", "abstract": "  Training data quality is one of the most important drivers of final model\nquality. In this work, we introduce a method for evaluating data integrity\nbased on the assumption that low-quality input prompts result in high variance\nand low quality responses. This is achieved by measuring the rejected response\nquality and the reward gap between the chosen and rejected preference pair. Our\nmethod, Rejecting Instruction Preferences (RIP) can be used to filter prompts\nfrom existing training sets, or to make high quality synthetic datasets,\nyielding large performance gains across various benchmarks compared to\nunfiltered data. Using Llama 3.1-8B-Instruct, RIP improves AlpacaEval2 LC Win\nRate by 9.4%, Arena-Hard by 8.7%, and WildBench by 9.9%. Using Llama\n3.3-70B-Instruct, RIP improves Arena-Hard from 67.5 to 82.9, which is from 18th\nplace to 6th overall in the leaderboard.\n", "date": "2025-01-30", "categories": []}
{"id": "2502.18691", "title": "Enhancing Image Classification with Augmentation: Data Augmentation\n  Techniques for Improved Image Classification", "abstract": "  Convolutional Neural Networks (CNNs) serve as the workhorse of deep learning,\nfinding applications in various fields that rely on images. Given sufficient\ndata, they exhibit the capacity to learn a wide range of concepts across\ndiverse settings. However, a notable limitation of CNNs is their susceptibility\nto overfitting when trained on small datasets. The augmentation of such\ndatasets can significantly enhance CNN performance by introducing additional\ndata points for learning. In this study, we explore the effectiveness of 11\ndifferent sets of data augmentation techniques, which include three novel sets\nproposed in this work. The first set of data augmentation employs pairwise\nchannel transfer, transferring Red, Green, Blue, Hue, and Saturation values\nfrom randomly selected images in the database to all images in the dataset. The\nsecond set introduces a novel occlusion approach, where objects in the images\nare occluded by randomly selected objects from the dataset. The third set\ninvolves a novel masking approach, using vertical, horizontal, circular, and\ncheckered masks to occlude portions of the images. In addition to these novel\ntechniques, we investigate other existing augmentation methods, including\nrotation, horizontal and vertical flips, resizing, translation, blur, color\njitter, and random erasing, and their effects on accuracy and overfitting. We\nfine-tune a base EfficientNet-B0 model for each augmentation method and conduct\na comparative analysis to showcase their efficacy. For the evaluation and\ncomparison of these augmentation techniques, we utilize the Caltech-101\ndataset. The ensemble of image augmentation techniques proposed emerges as the\nmost effective on the Caltech-101 dataset. The results demonstrate that diverse\ndata augmentation techniques present a viable means of enhancing datasets for\nimproved image classification.\n", "date": "2025-02-25", "categories": []}
{"id": "2306.02071", "title": "DU-Shapley: A Shapley Value Proxy for Efficient Dataset Valuation", "abstract": "  We consider the dataset valuation problem, that is, the problem of\nquantifying the incremental gain, to some relevant pre-defined utility of a\nmachine learning task, of aggregating an individual dataset to others. The\nShapley value is a natural tool to perform dataset valuation due to its formal\naxiomatic justification, which can be combined with Monte Carlo integration to\novercome the computational tractability challenges. Such generic approximation\nmethods, however, remain expensive in some cases. In this paper, we exploit the\nknowledge about the structure of the dataset valuation problem to devise more\nefficient Shapley value estimators. We propose a novel approximation, referred\nto as discrete uniform Shapley, which is expressed as an expectation under a\ndiscrete uniform distribution with support of reasonable size. We justify the\nrelevancy of the proposed framework via asymptotic and non-asymptotic\ntheoretical guarantees and illustrate its benefits via an extensive set of\nnumerical experiments.\n", "date": "2023-06-03", "categories": []}
{"id": "2312.03996", "title": "Stable Diffusion for Data Augmentation in COCO and Weed Datasets", "abstract": "  Generative models have increasingly impacted various tasks, from computer\nvision to interior design and beyond. Stable Diffusion, a powerful diffusion\nmodel, enables the creation of high-resolution images with intricate details\nfrom text prompts or reference images. An intriguing challenge lies in\nimproving performance for small datasets with image-sparse categories. This\nstudy explores the effectiveness of Stable Diffusion by evaluating seven common\ncategories and three widespread weed species. Synthetic images were generated\nusing three Stable Diffusion-based techniques: Image-to-Image Translation,\nDreamBooth, and ControlNet, each with distinct focuses. Classification and\ndetection tasks were then performed on these synthetic images, and their\nperformance was compared to models trained on original images. Promising\nresults were achieved for certain classes, demonstrating the potential of\nStable Diffusion in enhancing image-sparse datasets. This foundational study\nmay accelerate the adaptation of diffusion models across various domains.\n", "date": "2023-12-07", "categories": []}
{"id": "2406.04273", "title": "ELFS: Label-Free Coreset Selection with Proxy Training Dynamics", "abstract": "  High-quality human-annotated data is crucial for modern deep learning\npipelines, yet the human annotation process is both costly and time-consuming.\nGiven a constrained human labeling budget, selecting an informative and\nrepresentative data subset for labeling can significantly reduce human\nannotation effort. Well-performing state-of-the-art (SOTA) coreset selection\nmethods require ground truth labels over the whole dataset, failing to reduce\nthe human labeling burden. Meanwhile, SOTA label-free coreset selection methods\ndeliver inferior performance due to poor geometry-based difficulty scores. In\nthis paper, we introduce ELFS (Effective Label-Free Coreset Selection), a novel\nlabel-free coreset selection method. ELFS significantly improves label-free\ncoreset selection by addressing two challenges: 1) ELFS utilizes deep\nclustering to estimate training dynamics-based data difficulty scores without\nground truth labels; 2) Pseudo-labels introduce a distribution shift in the\ndata difficulty scores, and we propose a simple but effective double-end\npruning method to mitigate bias on calculated scores. We evaluate ELFS on four\nvision benchmarks and show that, given the same vision encoder, ELFS\nconsistently outperforms SOTA label-free baselines. For instance, when using\nSwAV as the encoder, ELFS outperforms D2 by up to 10.2% in accuracy on\nImageNet-1K. We make our code publicly available on GitHub.\n", "date": "2024-06-06", "categories": []}
{"id": "2407.14429", "title": "Dataset Distillation in Medical Imaging: A Feasibility Study", "abstract": "  Data sharing in the medical image analysis field has potential yet remains\nunderappreciated. The aim is often to share datasets efficiently with other\nsites to train models effectively. One possible solution is to avoid\ntransferring the entire dataset while still achieving similar model\nperformance. Recent progress in data distillation within computer science\noffers promising prospects for sharing medical data efficiently without\nsignificantly compromising model effectiveness. However, it remains uncertain\nwhether these methods would be applicable to medical imaging, since medical and\nnatural images are distinct fields. Moreover, it is intriguing to consider what\nlevel of performance could be achieved with these methods. To answer these\nquestions, we conduct investigations on a variety of leading data distillation\nmethods, in different contexts of medical imaging. We evaluate the feasibility\nof these methods with extensive experiments in two aspects: 1) Assess the\nimpact of data distillation across multiple datasets characterized by minor or\ngreat variations. 2) Explore the indicator to predict the distillation\nperformance. Our extensive experiments across multiple medical datasets reveal\nthat data distillation can significantly reduce dataset size while maintaining\ncomparable model performance to that achieved with the full dataset, suggesting\nthat a small, representative sample of images can serve as a reliable indicator\nof distillation success. This study demonstrates that data distillation is a\nviable method for efficient and secure medical data sharing, with the potential\nto facilitate enhanced collaborative research and clinical applications.\n", "date": "2024-07-19", "categories": []}
{"id": "2502.15834", "title": "Challenges of Multi-Modal Coreset Selection for Depth Prediction", "abstract": "  Coreset selection methods are effective in accelerating training and reducing\nmemory requirements but remain largely unexplored in applied multimodal\nsettings. We adapt a state-of-the-art (SoTA) coreset selection technique for\nmultimodal data, focusing on the depth prediction task. Our experiments with\nembedding aggregation and dimensionality reduction approaches reveal the\nchallenges of extending unimodal algorithms to multimodal scenarios,\nhighlighting the need for specialized methods to better capture inter-modal\nrelationships.\n", "date": "2025-02-20", "categories": []}
{"id": "2502.15950", "title": "Optimizing Pre-Training Data Mixtures with Mixtures of Data Expert\n  Models", "abstract": "  We propose a method to optimize language model pre-training data mixtures\nthrough efficient approximation of the cross-entropy loss corresponding to each\ncandidate mixture via a Mixture of Data Experts (MDE). We use this\napproximation as a source of additional features in a regression model, trained\nfrom observations of model loss for a small number of mixtures.\n  Experiments with Transformer decoder-only language models in the range of 70M\nto 1B parameters on the SlimPajama dataset show that our method achieves\nsignificantly better performance than approaches that train regression models\nusing only the mixture rates as input features. Combining this improved\noptimization method with an objective that takes into account cross-entropy on\nend task data leads to superior performance on few-shot downstream evaluations.\n  We also provide theoretical insights on why aggregation of data expert\npredictions can provide good approximations to model losses for data mixtures.\n", "date": "2025-02-21", "categories": []}
{"id": "2502.16703", "title": "Subsampling Graphs with GNN Performance Guarantees", "abstract": "  How can we subsample graph data so that a graph neural network (GNN) trained\non the subsample achieves performance comparable to training on the full\ndataset? This question is of fundamental interest, as smaller datasets reduce\nlabeling costs, storage requirements, and computational resources needed for\ntraining. Selecting an effective subset is challenging: a poorly chosen\nsubsample can severely degrade model performance, and empirically testing\nmultiple subsets for quality obviates the benefits of subsampling. Therefore,\nit is critical that subsampling comes with guarantees on model performance. In\nthis work, we introduce new subsampling methods for graph datasets that\nleverage the Tree Mover's Distance to reduce both the number of graphs and the\nsize of individual graphs. To our knowledge, our approach is the first that is\nsupported by rigorous theoretical guarantees: we prove that training a GNN on\nthe subsampled data results in a bounded increase in loss compared to training\non the full dataset. Unlike existing methods, our approach is both\nmodel-agnostic, requiring minimal assumptions about the GNN architecture, and\nlabel-agnostic, eliminating the need to label the full training set. This\nenables subsampling early in the model development pipeline (before data\nannotation, model selection, and hyperparameter tuning) reducing costs and\nresources needed for storage, labeling, and training. We validate our\ntheoretical results with experiments showing that our approach outperforms\nexisting subsampling methods across multiple datasets.\n", "date": "2025-02-23", "categories": []}
{"id": "2502.15245", "title": "Steganographic Embeddings as an Effective Data Augmentation", "abstract": "  Image Steganography is a cryptographic technique that embeds secret\ninformation into an image, ensuring the hidden data remains undetectable to the\nhuman eye while preserving the image's original visual integrity. Least\nSignificant Bit (LSB) Steganography achieves this by replacing the k least\nsignificant bits of an image with the k most significant bits of a secret\nimage, maintaining the appearance of the original image while simultaneously\nencoding the essential elements of the hidden data. In this work, we shift away\nfrom conventional applications of steganography in deep learning and explore\nits potential from a new angle. We present experimental results on CIFAR-10\nshowing that LSB Steganography, when used as a data augmentation strategy for\ndownstream computer vision tasks such as image classification, can\nsignificantly improve the training efficiency of deep neural networks. It can\nalso act as an implicit, uniformly discretized piecewise linear approximation\nof color augmentations such as (brightness, contrast, hue, and saturation),\nwithout introducing additional training overhead through a new joint image\ntraining regime that disregards the need for tuning sensitive augmentation\nhyperparameters.\n", "date": "2025-02-21", "categories": []}
{"id": "2502.15588", "title": "Improving the Scaling Laws of Synthetic Data with Deliberate Practice", "abstract": "  Inspired by the principle of deliberate practice in human learning, we\npropose Deliberate Practice for Synthetic Data Generation (DP), a novel\nframework that improves sample efficiency through dynamic synthetic data\ngeneration. Prior work has shown that scaling synthetic data is inherently\nchallenging, as naively adding new data leads to diminishing returns. To\naddress this, pruning has been identified as a key mechanism for improving\nscaling, enabling models to focus on the most informative synthetic samples.\nRather than generating a large dataset and pruning it afterward, DP efficiently\napproximates the direct generation of informative samples. We theoretically\nshow how training on challenging, informative examples improves scaling laws\nand empirically validate that DP achieves better scaling performance with\nsignificantly fewer training samples and iterations. On ImageNet-100, DP\ngenerates 3.4x fewer samples and requires six times fewer iterations, while on\nImageNet-1k, it generates 8x fewer samples with a 30 percent reduction in\niterations, all while achieving superior performance compared to prior work.\n", "date": "2025-02-21", "categories": []}
{"id": "2406.09408", "title": "Data Attribution for Text-to-Image Models by Unlearning Synthesized\n  Images", "abstract": "  The goal of data attribution for text-to-image models is to identify the\ntraining images that most influence the generation of a new image. Influence is\ndefined such that, for a given output, if a model is retrained from scratch\nwithout the most influential images, the model would fail to reproduce the same\noutput. Unfortunately, directly searching for these influential images is\ncomputationally infeasible, since it would require repeatedly retraining models\nfrom scratch. In our work, we propose an efficient data attribution method by\nsimulating unlearning the synthesized image. We achieve this by increasing the\ntraining loss on the output image, without catastrophic forgetting of other,\nunrelated concepts. We then identify training images with significant loss\ndeviations after the unlearning process and label these as influential. We\nevaluate our method with a computationally intensive but \"gold-standard\"\nretraining from scratch and demonstrate our method's advantages over previous\nmethods.\n", "date": "2024-06-13", "categories": []}
{"id": "2502.13198", "title": "Enhancing Machine Learning Performance through Intelligent Data Quality\n  Assessment: An Unsupervised Data-centric Framework", "abstract": "  Poor data quality limits the advantageous power of Machine Learning (ML) and\nweakens high-performing ML software systems. Nowadays, data are more prone to\nthe risk of poor quality due to their increasing volume and complexity.\nTherefore, tedious and time-consuming work goes into data preparation and\nimprovement before moving further in the ML pipeline. To address this\nchallenge, we propose an intelligent data-centric evaluation framework that can\nidentify high-quality data and improve the performance of an ML system. The\nproposed framework combines the curation of quality measurements and\nunsupervised learning to distinguish high- and low-quality data. The framework\nis designed to integrate flexible and general-purpose methods so that it is\ndeployed in various domains and applications. To validate the outcomes of the\ndesigned framework, we implemented it in a real-world use case from the field\nof analytical chemistry, where it is tested on three datasets of anti-sense\noligonucleotides. A domain expert is consulted to identify the relevant quality\nmeasurements and evaluate the outcomes of the framework. The results show that\nthe quality-centric data evaluation framework identifies the characteristics of\nhigh-quality data that guide the conduct of efficient laboratory experiments\nand consequently improve the performance of the ML system.\n", "date": "2025-02-18", "categories": []}
{"id": "2502.13514", "title": "Shall Your Data Strategy Work? Perform a Swift Study", "abstract": "  This work presents a swift method to assess the efficacy of particular types\nof instruction-tuning data, utilizing just a handful of probe examples and\neliminating the need for model retraining. This method employs the idea of\ngradient-based data influence estimation, analyzing the gradient projections of\nprobe examples from the chosen strategy onto evaluation examples to assess its\nadvantages. Building upon this method, we conducted three swift studies to\ninvestigate the potential of Chain-of-thought (CoT) data, query clarification\ndata, and response evaluation data in enhancing model generalization.\nSubsequently, we embarked on a validation study to corroborate the findings of\nthese swift studies. In this validation study, we developed training datasets\ntailored to each studied strategy and compared model performance with and\nwithout the use of these datasets. The results of the validation study aligned\nwith the findings of the swift studies, validating the efficacy of our proposed\nmethod.\n", "date": "2025-02-19", "categories": []}
{"id": "2412.06303", "title": "DSAI: Unbiased and Interpretable Latent Feature Extraction for\n  Data-Centric AI", "abstract": "  Large language models (LLMs) often struggle to objectively identify latent\ncharacteristics in large datasets due to their reliance on pre-trained\nknowledge rather than actual data patterns. To address this data grounding\nissue, we propose Data Scientist AI (DSAI), a framework that enables unbiased\nand interpretable feature extraction through a multi-stage pipeline with\nquantifiable prominence metrics for evaluating extracted features. On synthetic\ndatasets with known ground-truth features, DSAI demonstrates high recall in\nidentifying expert-defined features while faithfully reflecting the underlying\ndata. Applications on real-world datasets illustrate the framework's practical\nutility in uncovering meaningful patterns with minimal expert oversight,\nsupporting use cases such as interpretable classification.\n  The title of our paper is chosen from multiple candidates based on\nDSAI-generated criteria.\n", "date": "2024-12-09", "categories": []}
{"id": "2501.14253", "title": "Distributionally Robust Coreset Selection under Covariate Shift", "abstract": "  Coreset selection, which involves selecting a small subset from an existing\ntraining dataset, is an approach to reducing training data, and various\napproaches have been proposed for this method. In practical situations where\nthese methods are employed, it is often the case that the data distributions\ndiffer between the development phase and the deployment phase, with the latter\nbeing unknown. Thus, it is challenging to select an effective subset of\ntraining data that performs well across all deployment scenarios. We therefore\npropose Distributionally Robust Coreset Selection (DRCS). DRCS theoretically\nderives an estimate of the upper bound for the worst-case test error, assuming\nthat the future covariate distribution may deviate within a defined range from\nthe training distribution. Furthermore, by selecting instances in a way that\nsuppresses the estimate of the upper bound for the worst-case test error, DRCS\nachieves distributionally robust training instance selection. This study is\nprimarily applicable to convex training computation, but we demonstrate that it\ncan also be applied to deep learning under appropriate approximations. In this\npaper, we focus on covariate shift, a type of data distribution shift, and\ndemonstrate the effectiveness of DRCS through experiments.\n", "date": "2025-01-24", "categories": []}
{"id": "2502.12494", "title": "EDGE: Efficient Data Selection for LLM Agents via Guideline\n  Effectiveness", "abstract": "  Large Language Models (LLMs) have shown remarkable capabilities as AI agents.\nHowever, existing methods for enhancing LLM-agent abilities often lack a focus\non data quality, leading to inefficiencies and suboptimal results in both\nfine-tuning and prompt engineering. To address this issue, we introduce EDGE, a\nnovel approach for identifying informative samples without needing golden\nanswers. We propose the Guideline Effectiveness (GE) metric, which selects\nchallenging samples by measuring the impact of human-provided guidelines in\nmulti-turn interaction tasks. A low GE score indicates that the human expertise\nrequired for a sample is missing from the guideline, making the sample more\ninformative. By selecting samples with low GE scores, we can improve the\nefficiency and outcomes of both prompt engineering and fine-tuning processes\nfor LLMs. Extensive experiments validate the performance of our method. Our\nmethod achieves competitive results on the HotpotQA and WebShop and datasets,\nrequiring 75\\% and 50\\% less data, respectively, while outperforming existing\nmethods. We also provide a fresh perspective on the data quality of LLM-agent\nfine-tuning.\n", "date": "2025-02-18", "categories": []}
{"id": "2502.12607", "title": "Generalized Kernel Inducing Points by Duality Gap for Dataset\n  Distillation", "abstract": "  We propose Duality Gap KIP (DGKIP), an extension of the Kernel Inducing\nPoints (KIP) method for dataset distillation. While existing dataset\ndistillation methods often rely on bi-level optimization, DGKIP eliminates the\nneed for such optimization by leveraging duality theory in convex programming.\nThe KIP method has been introduced as a way to avoid bi-level optimization;\nhowever, it is limited to the squared loss and does not support other loss\nfunctions (e.g., cross-entropy or hinge loss) that are more suitable for\nclassification tasks. DGKIP addresses this limitation by exploiting an upper\nbound on parameter changes after dataset distillation using the duality gap,\nenabling its application to a wider range of loss functions. We also\ncharacterize theoretical properties of DGKIP by providing upper bounds on the\ntest error and prediction consistency after dataset distillation. Experimental\nresults on standard benchmarks such as MNIST and CIFAR-10 demonstrate that\nDGKIP retains the efficiency of KIP while offering broader applicability and\nrobust performance.\n", "date": "2025-02-18", "categories": []}
{"id": "2502.08828", "title": "A Survey on Data-Centric AI: Tabular Learning from Reinforcement\n  Learning and Generative AI Perspective", "abstract": "  Tabular data is one of the most widely used data formats across various\ndomains such as bioinformatics, healthcare, and marketing. As artificial\nintelligence moves towards a data-centric perspective, improving data quality\nis essential for enhancing model performance in tabular data-driven\napplications. This survey focuses on data-driven tabular data optimization,\nspecifically exploring reinforcement learning (RL) and generative approaches\nfor feature selection and feature generation as fundamental techniques for\nrefining data spaces. Feature selection aims to identify and retain the most\ninformative attributes, while feature generation constructs new features to\nbetter capture complex data patterns. We systematically review existing\ngenerative methods for tabular data engineering, analyzing their latest\nadvancements, real-world applications, and respective strengths and\nlimitations. This survey emphasizes how RL-based and generative techniques\ncontribute to the automation and intelligence of feature engineering. Finally,\nwe summarize the existing challenges and discuss future research directions,\naiming to provide insights that drive continued innovation in this field.\n", "date": "2025-02-12", "categories": []}
{"id": "2502.10489", "title": "LiveVal: Time-aware Data Valuation via Adaptive Reference Points", "abstract": "  Time-aware data valuation enhances training efficiency and model robustness,\nas early detection of harmful samples could prevent months of wasted\ncomputation. However, existing methods rely on model retraining or convergence\nassumptions or fail to capture long-term training dynamics.\n  We propose LiveVal, an efficient time-aware data valuation method with three\nkey designs:\n  1) seamless integration with SGD training for efficient data contribution\nmonitoring; 2) reference-based valuation with normalization for reliable\nbenchmark establishment; and 3) adaptive reference point selection for\nreal-time updating with optimized memory usage.\n  We establish theoretical guarantees for LiveVal's stability and prove that\nits valuations are bounded and directionally aligned with optimization\nprogress. Extensive experiments demonstrate that LiveVal provides efficient\ndata valuation across different modalities and model scales, achieving 180\nspeedup over traditional methods while maintaining robust detection\nperformance.\n", "date": "2025-02-14", "categories": []}
{"id": "2502.10723", "title": "A Mathematics Framework of Artificial Shifted Population Risk and Its\n  Further Understanding Related to Consistency Regularization", "abstract": "  Data augmentation is an important technique in training deep neural networks\nas it enhances their ability to generalize and remain robust. While data\naugmentation is commonly used to expand the sample size and act as a\nconsistency regularization term, there is a lack of research on the\nrelationship between them. To address this gap, this paper introduces a more\ncomprehensive mathematical framework for data augmentation. Through this\nframework, we establish that the expected risk of the shifted population is the\nsum of the original population risk and a gap term, which can be interpreted as\na consistency regularization term. The paper also provides a theoretical\nunderstanding of this gap, highlighting its negative effects on the early\nstages of training. We also propose a method to mitigate these effects. To\nvalidate our approach, we conducted experiments using same data augmentation\ntechniques and computing resources under several scenarios, including standard\ntraining, out-of-distribution, and imbalanced classification. The results\ndemonstrate that our methods surpass compared methods under all scenarios in\nterms of generalization ability and convergence stability. We provide our code\nimplementation at the following link: https://github.com/ydlsfhll/ASPR.\n", "date": "2025-02-15", "categories": []}
{"id": "2502.11062", "title": "Beyond Similarity: A Gradient-based Graph Method for Instruction Tuning\n  Data Selection", "abstract": "  Large language models (LLMs) have shown great potential across various\nindustries due to their remarkable ability to generalize through instruction\ntuning. However, the limited availability of domain-specific data significantly\nhampers their performance on specialized tasks. While existing methods\nprimarily focus on selecting training data from general datasets that are\nsimilar to the target domain, they often fail to consider the joint\ndistribution of instructions, resulting in inefficient learning and suboptimal\nknowledge transfer. To address these challenges, we introduce G2IS\n(Gradient-based Graph Instruction Selection), a novel method that constructs a\nmixed gradient-based instruction graph to capture the joint distribution and\ninterdependencies between instructions. By accounting for the relationships\nbetween instructions, G2IS improves domain adaptation efficiency. Additionally,\nwe propose a gradient walk algorithm to refine the data selection process,\nenhancing both training effectiveness and efficiency. Our experiments\ndemonstrate that G2IS outperforms traditional methods across various domain\nadaptation tasks, yielding significant performance gains, particularly in\ncomplex, data-scarce scenarios. These results underscore the potential of G2IS\nin advancing the development of large, domain-specific models.\n", "date": "2025-02-16", "categories": []}
{"id": "2404.05579", "title": "DRoP: Distributionally Robust Data Pruning", "abstract": "  In the era of exceptionally data-hungry models, careful selection of the\ntraining data is essential to mitigate the extensive costs of deep learning.\nData pruning offers a solution by removing redundant or uninformative samples\nfrom the dataset, which yields faster convergence and improved neural scaling\nlaws. However, little is known about its impact on classification bias of the\ntrained models. We conduct the first systematic study of this effect and reveal\nthat existing data pruning algorithms can produce highly biased classifiers. We\npresent theoretical analysis of the classification risk in a mixture of\nGaussians to argue that choosing appropriate class pruning ratios, coupled with\nrandom pruning within classes has potential to improve worst-class performance.\nWe thus propose DRoP, a distributionally robust approach to pruning and\nempirically demonstrate its performance on standard computer vision benchmarks.\nIn sharp contrast to existing algorithms, our proposed method continues\nimproving distributional robustness at a tolerable drop of average performance\nas we prune more from the datasets.\n", "date": "2024-04-08", "categories": []}
{"id": "2410.00759", "title": "Targeted synthetic data generation for tabular data via hardness\n  characterization", "abstract": "  Data augmentation via synthetic data generation has been shown to be\neffective in improving model performance and robustness in the context of\nscarce or low-quality data. Using the data valuation framework to statistically\nidentify beneficial and detrimental observations, we introduce a simple\naugmentation pipeline that generates only high-value training points based on\nhardness characterization, in a computationally efficient manner. We first\nempirically demonstrate via benchmarks on real data that Shapley-based data\nvaluation methods perform comparably with learning-based methods in hardness\ncharacterization tasks, while offering significant computational advantages.\nThen, we show that synthetic data generators trained on the hardest points\noutperform non-targeted data augmentation on a number of tabular datasets. Our\napproach improves the quality of out-of-sample predictions and it is\ncomputationally more efficient compared to non-targeted methods.\n", "date": "2024-10-01", "categories": []}
{"id": "2407.06120", "title": "Sketchy Moment Matching: Toward Fast and Provable Data Selection for\n  Finetuning", "abstract": "  We revisit data selection in a modern context of finetuning from a\nfundamental perspective. Extending the classical wisdom of variance\nminimization in low dimensions to high-dimensional finetuning, our\ngeneralization analysis unveils the importance of additionally reducing bias\ninduced by low-rank approximation. Inspired by the variance-bias tradeoff in\nhigh dimensions from the theory, we introduce Sketchy Moment Matching (SkMM), a\nscalable data selection scheme with two stages. (i) First, the bias is\ncontrolled using gradient sketching that explores the finetuning parameter\nspace for an informative low-dimensional subspace $\\mathcal{S}$; (ii) then the\nvariance is reduced over $\\mathcal{S}$ via moment matching between the original\nand selected datasets. Theoretically, we show that gradient sketching is fast\nand provably accurate: selecting $n$ samples by reducing variance over\n$\\mathcal{S}$ preserves the fast-rate generalization $O(\\dim(\\mathcal{S})/n)$,\nindependent of the parameter dimension. Empirically, we concretize the\nvariance-bias balance via synthetic experiments and demonstrate the\neffectiveness of SkMM for finetuning in real vision tasks.\n", "date": "2024-07-08", "categories": []}
{"id": "2502.05027", "title": "Trust-Aware Diversion for Data-Effective Distillation", "abstract": "  Dataset distillation compresses a large dataset into a small synthetic subset\nthat retains essential information. Existing methods assume that all samples\nare perfectly labeled, limiting their real-world applications where incorrect\nlabels are ubiquitous. These mislabeled samples introduce untrustworthy\ninformation into the dataset, which misleads model optimization in dataset\ndistillation. To tackle this issue, we propose a Trust-Aware Diversion (TAD)\ndataset distillation method. Our proposed TAD introduces an iterative dual-loop\noptimization framework for data-effective distillation. Specifically, the outer\nloop divides data into trusted and untrusted spaces, redirecting distillation\ntoward trusted samples to guarantee trust in the distillation process. This\nstep minimizes the impact of mislabeled samples on dataset distillation. The\ninner loop maximizes the distillation objective by recalibrating untrusted\nsamples, thus transforming them into valuable ones for distillation. This\ndual-loop iteratively refines and compensates for each other, gradually\nexpanding the trusted space and shrinking the untrusted space. Experiments\ndemonstrate that our method can significantly improve the performance of\nexisting dataset distillation methods on three widely used benchmarks (CIFAR10,\nCIFAR100, and Tiny ImageNet) in three challenging mislabeled settings\n(symmetric, asymmetric, and real-world).\n", "date": "2025-02-07", "categories": []}
{"id": "2502.02854", "title": "TD3: Tucker Decomposition Based Dataset Distillation Method for\n  Sequential Recommendation", "abstract": "  In the era of data-centric AI, the focus of recommender systems has shifted\nfrom model-centric innovations to data-centric approaches. The success of\nmodern AI models is built on large-scale datasets, but this also results in\nsignificant training costs. Dataset distillation has emerged as a key solution,\ncondensing large datasets to accelerate model training while preserving model\nperformance. However, condensing discrete and sequentially correlated user-item\ninteractions, particularly with extensive item sets, presents considerable\nchallenges. This paper introduces \\textbf{TD3}, a novel \\textbf{T}ucker\n\\textbf{D}ecomposition based \\textbf{D}ataset \\textbf{D}istillation method\nwithin a meta-learning framework, designed for sequential recommendation. TD3\ndistills a fully expressive \\emph{synthetic sequence summary} from original\ndata. To efficiently reduce computational complexity and extract refined latent\npatterns, Tucker decomposition decouples the summary into four factors:\n\\emph{synthetic user latent factor}, \\emph{temporal dynamics latent factor},\n\\emph{shared item latent factor}, and a \\emph{relation core} that models their\ninterconnections. Additionally, a surrogate objective in bi-level optimization\nis proposed to align feature spaces extracted from models trained on both\noriginal data and synthetic sequence summary beyond the na\\\"ive performance\nmatching approach. In the \\emph{inner-loop}, an augmentation technique allows\nthe learner to closely fit the synthetic summary, ensuring an accurate update\nof it in the \\emph{outer-loop}. To accelerate the optimization process and\naddress long dependencies, RaT-BPTT is employed for bi-level optimization.\nExperiments and analyses on multiple public datasets have confirmed the\nsuperiority and cross-architecture generalizability of the proposed designs.\nCodes are released at https://github.com/USTC-StarTeam/TD3.\n", "date": "2025-02-05", "categories": []}
{"id": "2502.04229", "title": "Dark Distillation: Backdooring Distilled Datasets without Accessing Raw\n  Data", "abstract": "  Dataset distillation (DD) enhances training efficiency and reduces bandwidth\nby condensing large datasets into smaller synthetic ones. It enables models to\nachieve performance comparable to those trained on the raw full dataset and has\nbecome a widely adopted method for data sharing. However, security concerns in\nDD remain underexplored. Existing studies typically assume that malicious\nbehavior originates from dataset owners during the initial distillation\nprocess, where backdoors are injected into raw datasets. In contrast, this work\nis the first to address a more realistic and concerning threat: attackers may\nintercept the dataset distribution process, inject backdoors into the distilled\ndatasets, and redistribute them to users. While distilled datasets were\npreviously considered resistant to backdoor attacks, we demonstrate that they\nremain vulnerable to such attacks. Furthermore, we show that attackers do not\neven require access to any raw data to inject the backdoors successfully.\nSpecifically, our approach reconstructs conceptual archetypes for each class\nfrom the model trained on the distilled dataset. Backdoors are then injected\ninto these archetypes to update the distilled dataset. Moreover, we ensure the\nupdated dataset not only retains the backdoor but also preserves the original\noptimization trajectory, thus maintaining the knowledge of the raw dataset. To\nachieve this, a hybrid loss is designed to integrate backdoor information along\nthe benign optimization trajectory, ensuring that previously learned\ninformation is not forgotten. Extensive experiments demonstrate that distilled\ndatasets are highly vulnerable to backdoor attacks, with risks pervasive across\nvarious raw datasets, distillation methods, and downstream training strategies.\nMoreover, our attack method is efficient, capable of synthesizing a malicious\ndistilled dataset in under one minute in certain cases.\n", "date": "2025-02-06", "categories": []}
{"id": "2502.01703", "title": "QLESS: A Quantized Approach for Data Valuation and Selection in Large\n  Language Model Fine-Tuning", "abstract": "  Fine-tuning large language models (LLMs) is often constrained by the\ncomputational costs of processing massive datasets. We propose \\textbf{QLESS}\n(Quantized Low-rank Gradient Similarity Search), which integrates gradient\nquantization with the LESS framework to enable memory-efficient data valuation\nand selection. QLESS employs a two-step compression process: first, it obtains\nlow-dimensional gradient representations through LoRA-based random projection;\nthen, it quantizes these gradients to low-bitwidth representations. Experiments\non multiple LLM architectures (LLaMA, Mistral, Qwen) and benchmarks (MMLU, BBH,\nTyDiQA) show that QLESS achieves comparable data selection performance to LESS\nwhile reducing memory usage by up to 16x. Even 1-bit gradient quantization\npreserves data valuation quality. These findings underscore QLESS as a\npractical, scalable approach to identifying informative examples within strict\nmemory constraints.\n", "date": "2025-02-03", "categories": []}
{"id": "2502.01865", "title": "Enhancing Generalization via Sharpness-Aware Trajectory Matching for\n  Dataset Condensation", "abstract": "  Dataset condensation aims to synthesize datasets with a few representative\nsamples that can effectively represent the original datasets. This enables\nefficient training and produces models with performance close to those trained\non the original sets. Most existing dataset condensation methods conduct\ndataset learning under the bilevel (inner- and outer-loop) based optimization.\nHowever, the preceding methods perform with limited dataset generalization due\nto the notoriously complicated loss landscape and expensive time-space\ncomplexity of the inner-loop unrolling of bilevel optimization. These issues\ndeteriorate when the datasets are learned via matching the trajectories of\nnetworks trained on the real and synthetic datasets with a long horizon\ninner-loop. To address these issues, we introduce Sharpness-Aware Trajectory\nMatching (SATM), which enhances the generalization capability of learned\nsynthetic datasets by optimising the sharpness of the loss landscape and\nobjective simultaneously. Moreover, our approach is coupled with an efficient\nhypergradient approximation that is mathematically well-supported and\nstraightforward to implement along with controllable computational overhead.\nEmpirical evaluations of SATM demonstrate its effectiveness across various\napplications, including in-domain benchmarks and out-of-domain settings.\nMoreover, its easy-to-implement properties afford flexibility, allowing it to\nintegrate with other advanced sharpness-aware minimizers. Our code will be\nreleased.\n", "date": "2025-02-03", "categories": []}
{"id": "2402.04676", "title": "Group Distributionally Robust Dataset Distillation with Risk\n  Minimization", "abstract": "  Dataset distillation (DD) has emerged as a widely adopted technique for\ncrafting a synthetic dataset that captures the essential information of a\ntraining dataset, facilitating the training of accurate neural models. Its\napplications span various domains, including transfer learning, federated\nlearning, and neural architecture search. The most popular methods for\nconstructing the synthetic data rely on matching the convergence properties of\ntraining the model with the synthetic dataset and the training dataset.\nHowever, using the empirical loss as the criterion must be thought of as\nauxiliary in the same sense that the training set is an approximate substitute\nfor the population distribution, and the latter is the data of interest. Yet\ndespite its popularity, an aspect that remains unexplored is the relationship\nof DD to its generalization, particularly across uncommon subgroups. That is,\nhow can we ensure that a model trained on the synthetic dataset performs well\nwhen faced with samples from regions with low population density? Here, the\nrepresentativeness and coverage of the dataset become salient over the\nguaranteed training error at inference. Drawing inspiration from\ndistributionally robust optimization, we introduce an algorithm that combines\nclustering with the minimization of a risk measure on the loss to conduct DD.\nWe provide a theoretical rationale for our approach and demonstrate its\neffective generalization and robustness across subgroups through numerical\nexperiments. The source code is available at\nhttps://github.com/Mming11/RobustDatasetDistillation.\n", "date": "2024-02-07", "categories": []}
{"id": "2407.11004", "title": "The ALCHEmist: Automated Labeling 500x CHEaper Than LLM Data Annotators", "abstract": "  Large pretrained models can be used as annotators, helping replace or augment\ncrowdworkers and enabling distilling generalist models into smaller specialist\nmodels. Unfortunately, this comes at a cost: employing top-of-the-line models\noften requires paying thousands of dollars for API calls, while the resulting\ndatasets are static and challenging to audit. To address these challenges, we\npropose a simple alternative: rather than directly querying labels from\npretrained models, we task models to generate programs that can produce labels.\nThese programs can be stored and applied locally, re-used and extended, and\ncost orders of magnitude less. Our system, Alchemist, obtains comparable to or\nbetter performance than large language model-based annotation in a range of\ntasks for a fraction of the cost: on average, improvements amount to a 12.9%\nenhancement while the total labeling costs across all datasets are reduced by a\nfactor of approximately 500x.\n", "date": "2024-06-25", "categories": []}
{"id": "2411.04637", "title": "Hands-On Tutorial: Labeling with LLM and Human-in-the-Loop", "abstract": "  Training and deploying machine learning models relies on a large amount of\nhuman-annotated data. As human labeling becomes increasingly expensive and\ntime-consuming, recent research has developed multiple strategies to speed up\nannotation and reduce costs and human workload: generating synthetic training\ndata, active learning, and hybrid labeling. This tutorial is oriented toward\npractical applications: we will present the basics of each strategy, highlight\ntheir benefits and limitations, and discuss in detail real-life case studies.\nAdditionally, we will walk through best practices for managing human annotators\nand controlling the quality of the final dataset. The tutorial includes a\nhands-on workshop, where attendees will be guided in implementing a hybrid\nannotation setup. This tutorial is designed for NLP practitioners from both\nresearch and industry backgrounds who are involved in or interested in\noptimizing data labeling projects.\n", "date": "2024-11-07", "categories": []}
{"id": "2501.15409", "title": "TdAttenMix: Top-Down Attention Guided Mixup", "abstract": "  CutMix is a data augmentation strategy that cuts and pastes image patches to\nmixup training data. Existing methods pick either random or salient areas which\nare often inconsistent to labels, thus misguiding the training model. By our\nknowledge, we integrate human gaze to guide cutmix for the first time. Since\nhuman attention is driven by both high-level recognition and low-level clues,\nwe propose a controllable Top-down Attention Guided Module to obtain a general\nartificial attention which balances top-down and bottom-up attention. The\nproposed TdATttenMix then picks the patches and adjust the label mixing ratio\nthat focuses on regions relevant to the current label. Experimental results\ndemonstrate that our TdAttenMix outperforms existing state-of-the-art mixup\nmethods across eight different benchmarks. Additionally, we introduce a new\nmetric based on the human gaze and use this metric to investigate the issue of\nimage-label inconsistency. Project page:\n\\url{https://github.com/morning12138/TdAttenMix}\n", "date": "2025-01-26", "categories": []}
{"id": "2501.15963", "title": "Evaluating Data Influence in Meta Learning", "abstract": "  As one of the most fundamental models, meta learning aims to effectively\naddress few-shot learning challenges. However, it still faces significant\nissues related to the training data, such as training inefficiencies due to\nnumerous low-contribution tasks in large datasets and substantial noise from\nincorrect labels. Thus, training data attribution methods are needed for meta\nlearning. However, the dual-layer structure of mata learning complicates the\nmodeling of training data contributions because of the interdependent influence\nbetween meta-parameters and task-specific parameters, making existing data\ninfluence evaluation tools inapplicable or inaccurate. To address these\nchallenges, based on the influence function, we propose a general data\nattribution evaluation framework for meta-learning within the bilevel\noptimization framework. Our approach introduces task influence functions\n(task-IF) and instance influence functions (instance-IF) to accurately assess\nthe impact of specific tasks and individual data points in closed forms. This\nframework comprehensively models data contributions across both the inner and\nouter training processes, capturing the direct effects of data points on\nmeta-parameters as well as their indirect influence through task-specific\nparameters. We also provide several strategies to enhance computational\nefficiency and scalability. Experimental results demonstrate the framework's\neffectiveness in training data evaluation via several downstream tasks.\n", "date": "2025-01-27", "categories": []}
{"id": "2312.06254", "title": "Modyn: Data-Centric Machine Learning Pipeline Orchestration", "abstract": "  In real-world machine learning (ML) pipelines, datasets are continuously\ngrowing. Models must incorporate this new training data to improve\ngeneralization and adapt to potential distribution shifts. The cost of model\nretraining is proportional to how frequently the model is retrained and how\nmuch data it is trained on, which makes the naive approach of retraining from\nscratch each time impractical.\n  We present Modyn, a data-centric end-to-end machine learning platform.\nModyn's ML pipeline abstraction enables users to declaratively describe\npolicies for continuously training a model on a growing dataset. Modyn\npipelines allow users to apply data selection policies (to reduce the number of\ndata points) and triggering policies (to reduce the number of trainings). Modyn\nexecutes and orchestrates these continuous ML training pipelines. The system is\nopen-source and comes with an ecosystem of benchmark datasets, models, and\ntooling. We formally discuss how to measure the performance of ML pipelines by\nintroducing the concept of composite models, enabling fair comparison of\npipelines with different data selection and triggering policies. We empirically\nanalyze how various data selection and triggering policies impact model\naccuracy, and also show that Modyn enables high throughput training with\nsample-level data selection.\n", "date": "2023-12-11", "categories": []}
{"id": "2411.11081", "title": "The Promises and Pitfalls of LLM Annotations in Dataset Labeling: a Case\n  Study on Media Bias Detection", "abstract": "  High annotation costs from hiring or crowdsourcing complicate the creation of\nlarge, high-quality datasets needed for training reliable text classifiers.\nRecent research suggests using Large Language Models (LLMs) to automate the\nannotation process, reducing these costs while maintaining data quality. LLMs\nhave shown promising results in annotating downstream tasks like hate speech\ndetection and political framing. Building on the success in these areas, this\nstudy investigates whether LLMs are viable for annotating the complex task of\nmedia bias detection and whether a downstream media bias classifier can be\ntrained on such data. We create annolexical, the first large-scale dataset for\nmedia bias classification with over 48000 synthetically annotated examples. Our\nclassifier, fine-tuned on this dataset, surpasses all of the annotator LLMs by\n5-9 percent in Matthews Correlation Coefficient (MCC) and performs close to or\noutperforms the model trained on human-labeled data when evaluated on two media\nbias benchmark datasets (BABE and BASIL). This study demonstrates how our\napproach significantly reduces the cost of dataset creation in the media bias\ndomain and, by extension, the development of classifiers, while our subsequent\nbehavioral stress-testing reveals some of its current limitations and\ntrade-offs.\n", "date": "2024-11-17", "categories": []}
{"id": "2412.14809", "title": "ResoFilter: Fine-grained Synthetic Data Filtering for Large Language\n  Models through Data-Parameter Resonance Analysis", "abstract": "  Large language models (LLMs) have shown remarkable effectiveness across\nvarious domains, with data augmentation methods utilizing GPT for synthetic\ndata generation becoming prevalent. However, the quality and utility of\naugmented data remain questionable, and current methods lack clear metrics for\nevaluating data characteristics. To address these challenges, we propose\nResoFilter, a novel method that integrates models, data, and tasks to refine\ndatasets. ResoFilter leverages the fine-tuning process to obtain Data-Parameter\nfeatures for data selection, offering improved interpretability by representing\ndata characteristics through model weights. Our experiments demonstrate that\nResoFilter achieves comparable results to full-scale fine-tuning using only\nhalf the data in mathematical tasks and exhibits strong generalization across\ndifferent models and domains. This method provides valuable insights for\nconstructing synthetic datasets and evaluating high-quality data, offering a\npromising solution for enhancing data augmentation techniques and improving\ntraining dataset quality for LLMs. For reproducibility, we will release our\ncode and data upon acceptance.\n", "date": "2024-12-19", "categories": []}
{"id": "2405.13707", "title": "Rethinking and Accelerating Graph Condensation: A Training-Free Approach\n  with Class Partition", "abstract": "  The increasing prevalence of large-scale graphs poses a significant challenge\nfor graph neural network training, attributed to their substantial\ncomputational requirements. In response, graph condensation (GC) emerges as a\npromising data-centric solution aiming to substitute the large graph with a\nsmall yet informative condensed graph to facilitate data-efficient GNN\ntraining. However, existing GC methods suffer from intricate optimization\nprocesses, necessitating excessive computing resources and training time. In\nthis paper, we revisit existing GC optimization strategies and identify two\npervasive issues therein: (1) various GC optimization strategies converge to\ncoarse-grained class-level node feature matching between the original and\ncondensed graphs; (2) existing GC methods rely on a Siamese graph network\narchitecture that requires time-consuming bi-level optimization with iterative\ngradient computations. To overcome these issues, we propose a training-free GC\nframework termed Class-partitioned Graph Condensation (CGC), which refines the\nnode distribution matching from the class-to-class paradigm into a novel\nclass-to-node paradigm, transforming the GC optimization into a class partition\nproblem which can be efficiently solved by any clustering methods. Moreover,\nCGC incorporates a pre-defined graph structure to enable a closed-form solution\nfor condensed node features, eliminating the need for back-and-forth gradient\ndescent in existing GC approaches. Extensive experiments demonstrate that CGC\nachieves an exceedingly efficient condensation process with advanced accuracy.\nCompared with the state-of-the-art GC methods, CGC condenses the Ogbn-products\ngraph within 30 seconds, achieving a speedup ranging from $10^2$X to $10^4$X\nand increasing accuracy by up to 4.2%.\n", "date": "2024-05-22", "categories": []}
{"id": "2407.01492", "title": "RegMix: Data Mixture as Regression for Language Model Pre-training", "abstract": "  The data mixture for large language model pre-training significantly impacts\nperformance, yet how to determine an effective mixture remains unclear. We\npropose RegMix to automatically identify a high-performing data mixture by\nformulating it as a regression task. RegMix trains many small models on diverse\ndata mixtures, uses regression to predict performance of unseen mixtures, and\napplies the best predicted mixture to train a large-scale model with orders of\nmagnitude more compute. To empirically validate RegMix, we train 512 models\nwith 1M parameters for 1B tokens to fit the regression model and predict the\nbest data mixture. Using this mixture we train a 1B parameter model for 25B\ntokens (i.e. 1000x larger and 25x longer) which we find performs best among 64\ncandidate 1B parameter models with other mixtures. Furthermore, RegMix\nconsistently outperforms human selection in experiments involving models up to\n7B models trained on 100B tokens, while matching or exceeding DoReMi using just\n10% of the computational resources. Our experiments also show that (1) Data\nmixtures significantly impact performance; (2) Web corpora rather than data\nperceived as high-quality like Wikipedia have the strongest positive\ncorrelation with downstream performance; (3) Domains interact in complex ways\noften contradicting common sense, thus automatic approaches like RegMix are\nneeded; (4) Data mixture effects transcend scaling laws. Our code is available\nat https://github.com/sail-sg/regmix.\n", "date": "2024-07-01", "categories": []}
{"id": "2501.13905", "title": "On Learning Representations for Tabular Data Distillation", "abstract": "  Dataset distillation generates a small set of information-rich instances from\na large dataset, resulting in reduced storage requirements, privacy or\ncopyright risks, and computational costs for downstream modeling, though much\nof the research has focused on the image data modality. We study tabular data\ndistillation, which brings in novel challenges such as the inherent feature\nheterogeneity and the common use of non-differentiable learning models (such as\ndecision tree ensembles and nearest-neighbor predictors). To mitigate these\nchallenges, we present $\\texttt{TDColER}$, a tabular data distillation\nframework via column embeddings-based representation learning. To evaluate this\nframework, we also present a tabular data distillation benchmark, ${{\\sf \\small\nTDBench}}$. Based on an elaborate evaluation on ${{\\sf \\small TDBench}}$,\nresulting in 226,890 distilled datasets and 548,880 models trained on them, we\ndemonstrate that $\\texttt{TDColER}$ is able to boost the distilled data quality\nof off-the-shelf distillation schemes by 0.5-143% across 7 different tabular\nlearning models.\n", "date": "2025-01-23", "categories": []}
{"id": "2406.11730", "title": "CHG Shapley: Efficient Data Valuation and Selection towards Trustworthy\n  Machine Learning", "abstract": "  Understanding the decision-making process of machine learning models is\ncrucial for ensuring trustworthy machine learning. Data Shapley, a landmark\nstudy on data valuation, advances this understanding by assessing the\ncontribution of each datum to model performance. However, the\nresource-intensive and time-consuming nature of multiple model retraining poses\nchallenges for applying Data Shapley to large datasets. To address this, we\npropose the CHG (compound of Hardness and Gradient) utility function, which\napproximates the utility of each data subset on model performance in every\ntraining epoch. By deriving the closed-form Shapley value for each data point\nusing the CHG utility function, we reduce the computational complexity to that\nof a single model retraining, achieving a quadratic improvement over existing\nmarginal contribution-based methods. We further leverage CHG Shapley for\nreal-time data selection, conducting experiments across three settings:\nstandard datasets, label noise datasets, and class imbalance datasets. These\nexperiments demonstrate its effectiveness in identifying high-value and noisy\ndata. By enabling efficient data valuation, CHG Shapley promotes trustworthy\nmodel training through a novel data-centric perspective. Our codes are\navailable at https://github.com/caihuaiguang/CHG-Shapley-for-Data-Valuation and\nhttps://github.com/caihuaiguang/CHG-Shapley-for-Data-Selection.\n", "date": "2024-06-17", "categories": []}
{"id": "2406.10485", "title": "A Label is Worth a Thousand Images in Dataset Distillation", "abstract": "  Data $\\textit{quality}$ is a crucial factor in the performance of machine\nlearning models, a principle that dataset distillation methods exploit by\ncompressing training datasets into much smaller counterparts that maintain\nsimilar downstream performance. Understanding how and why data distillation\nmethods work is vital not only for improving these methods but also for\nrevealing fundamental characteristics of \"good\" training data. However, a major\nchallenge in achieving this goal is the observation that distillation\napproaches, which rely on sophisticated but mostly disparate methods to\ngenerate synthetic data, have little in common with each other. In this work,\nwe highlight a largely overlooked aspect common to most of these methods: the\nuse of soft (probabilistic) labels. Through a series of ablation experiments,\nwe study the role of soft labels in depth. Our results reveal that the main\nfactor explaining the performance of state-of-the-art distillation methods is\nnot the specific techniques used to generate synthetic data but rather the use\nof soft labels. Furthermore, we demonstrate that not all soft labels are\ncreated equal; they must contain $\\textit{structured information}$ to be\nbeneficial. We also provide empirical scaling laws that characterize the\neffectiveness of soft labels as a function of images-per-class in the distilled\ndataset and establish an empirical Pareto frontier for data-efficient learning.\nCombined, our findings challenge conventional wisdom in dataset distillation,\nunderscore the importance of soft labels in learning, and suggest new\ndirections for improving distillation methods. Code for all experiments is\navailable at https://github.com/sunnytqin/no-distillation.\n", "date": "2024-06-15", "categories": []}
{"id": "2501.10555", "title": "Towards Data-Centric AI: A Comprehensive Survey of Traditional,\n  Reinforcement, and Generative Approaches for Tabular Data Transformation", "abstract": "  Tabular data is one of the most widely used formats across industries,\ndriving critical applications in areas such as finance, healthcare, and\nmarketing. In the era of data-centric AI, improving data quality and\nrepresentation has become essential for enhancing model performance,\nparticularly in applications centered around tabular data. This survey examines\nthe key aspects of tabular data-centric AI, emphasizing feature selection and\nfeature generation as essential techniques for data space refinement. We\nprovide a systematic review of feature selection methods, which identify and\nretain the most relevant data attributes, and feature generation approaches,\nwhich create new features to simplify the capture of complex data patterns.\nThis survey offers a comprehensive overview of current methodologies through an\nanalysis of recent advancements, practical applications, and the strengths and\nlimitations of these techniques. Finally, we outline open challenges and\nsuggest future perspectives to inspire continued innovation in this field.\n", "date": "2025-01-17", "categories": []}
{"id": "2501.10893", "title": "Learn-by-interact: A Data-Centric Framework for Self-Adaptive Agents in\n  Realistic Environments", "abstract": "  Autonomous agents powered by large language models (LLMs) have the potential\nto enhance human capabilities, assisting with digital tasks from sending emails\nto performing data analysis. The abilities of existing LLMs at such tasks are\noften hindered by the lack of high-quality agent data from the corresponding\nenvironments they interact with. We propose Learn-by-interact, a data-centric\nframework to adapt LLM agents to any given environments without human\nannotations. Learn-by-interact synthesizes trajectories of agent-environment\ninteractions based on documentations, and constructs instructions by\nsummarizing or abstracting the interaction histories, a process called backward\nconstruction. We assess the quality of our synthetic data by using them in both\ntraining-based scenarios and training-free in-context learning (ICL), where we\ncraft innovative retrieval approaches optimized for agents. Extensive\nexperiments on SWE-bench, WebArena, OSWorld and Spider2-V spanning across\nrealistic coding, web, and desktop environments show the effectiveness of\nLearn-by-interact in various downstream agentic tasks -- baseline results are\nimproved by up to 12.2\\% for ICL with Claude-3.5 and 19.5\\% for training with\nCodestral-22B. We further demonstrate the critical role of backward\nconstruction, which provides up to 14.0\\% improvement for training. Our\nablation studies demonstrate the efficiency provided by our synthesized data in\nICL and the superiority of our retrieval pipeline over alternative approaches\nlike conventional retrieval-augmented generation (RAG). We expect that\nLearn-by-interact will serve as a foundation for agent data synthesis as LLMs\nare increasingly deployed at real-world environments.\n", "date": "2025-01-18", "categories": []}
{"id": "2501.12147", "title": "Improving Influence-based Instruction Tuning Data Selection for Balanced\n  Learning of Diverse Capabilities", "abstract": "  Selecting appropriate training data is crucial for effective instruction\nfine-tuning of large language models (LLMs), which aims to (1) elicit strong\ncapabilities, and (2) achieve balanced performance across a diverse range of\ntasks. Influence-based methods show promise in achieving (1) by estimating the\ncontribution of each training example to the model's predictions, but often\nstruggle with (2). Our systematic investigation reveals that this\nunderperformance can be attributed to an inherent bias where certain tasks\nintrinsically have greater influence than others. As a result, data selection\nis often biased towards these tasks, not only hurting the model's performance\non others but also, counterintuitively, harms performance on these\nhigh-influence tasks themselves.\n  As a remedy, we propose BIDS, a Balanced and Influential Data Selection\nalgorithm. BIDS first normalizes influence scores of the training data, and\nthen iteratively balances data selection by choosing the training example with\nthe highest influence on the most underrepresented task. Experiments with both\nLlama-3 and Mistral-v0.3 on seven benchmarks spanning five diverse capabilities\nshow that BIDS consistently outperforms both state-of-the-art influence-based\nalgorithms and other non-influence-based selection frameworks. Surprisingly,\ntraining on a 15% subset selected by BIDS can even outperform full-dataset\ntraining with a much more balanced performance. Our analysis further highlights\nthe importance of both instance-level normalization and iterative optimization\nof selected data for balanced learning of diverse capabilities.\n", "date": "2025-01-21", "categories": []}
{"id": "2404.13733", "title": "Elucidating the Design Space of Dataset Condensation", "abstract": "  Dataset condensation, a concept within data-centric learning, efficiently\ntransfers critical attributes from an original dataset to a synthetic version,\nmaintaining both diversity and realism. This approach significantly improves\nmodel training efficiency and is adaptable across multiple application areas.\nPrevious methods in dataset condensation have faced challenges: some incur high\ncomputational costs which limit scalability to larger datasets (e.g., MTT,\nDREAM, and TESLA), while others are restricted to less optimal design spaces,\nwhich could hinder potential improvements, especially in smaller datasets\n(e.g., SRe2L, G-VBSM, and RDED). To address these limitations, we propose a\ncomprehensive design framework that includes specific, effective strategies\nlike implementing soft category-aware matching and adjusting the learning rate\nschedule. These strategies are grounded in empirical evidence and theoretical\nbacking. Our resulting approach, Elucidate Dataset Condensation (EDC),\nestablishes a benchmark for both small and large-scale dataset condensation. In\nour testing, EDC achieves state-of-the-art accuracy, reaching 48.6% on\nImageNet-1k with a ResNet-18 model at an IPC of 10, which corresponds to a\ncompression ratio of 0.78%. This performance exceeds those of SRe2L, G-VBSM,\nand RDED by margins of 27.3%, 17.2%, and 6.6%, respectively.\n", "date": "2024-04-21", "categories": []}
{"id": "2501.09504", "title": "HydraMix: Multi-Image Feature Mixing for Small Data Image Classification", "abstract": "  Training deep neural networks requires datasets with a large number of\nannotated examples. The collection and annotation of these datasets is not only\nextremely expensive but also faces legal and privacy problems. These factors\nare a significant limitation for many real-world applications. To address this,\nwe introduce HydraMix, a novel architecture that generates new image\ncompositions by mixing multiple different images from the same class. HydraMix\nlearns the fusion of the content of various images guided by a\nsegmentation-based mixing mask in feature space and is optimized via a\ncombination of unsupervised and adversarial training. Our data augmentation\nscheme allows the creation of models trained from scratch on very small\ndatasets. We conduct extensive experiments on ciFAIR-10, STL-10, and\nciFAIR-100. Additionally, we introduce a novel text-image metric to assess the\ngenerality of the augmented datasets. Our results show that HydraMix\noutperforms existing state-of-the-art methods for image classification on small\ndatasets.\n", "date": "2025-01-16", "categories": []}
{"id": "2501.04861", "title": "LayerMix: Enhanced Data Augmentation through Fractal Integration for\n  Robust Deep Learning", "abstract": "  Deep learning models have demonstrated remarkable performance across various\ncomputer vision tasks, yet their vulnerability to distribution shifts remains a\ncritical challenge. Despite sophisticated neural network architectures,\nexisting models often struggle to maintain consistent performance when\nconfronted with Out-of-Distribution (OOD) samples, including natural\ncorruptions, adversarial perturbations, and anomalous patterns. We introduce\nLayerMix, an innovative data augmentation approach that systematically enhances\nmodel robustness through structured fractal-based image synthesis. By\nmeticulously integrating structural complexity into training datasets, our\nmethod generates semantically consistent synthetic samples that significantly\nimprove neural network generalization capabilities. Unlike traditional\naugmentation techniques that rely on random transformations, LayerMix employs a\nstructured mixing pipeline that preserves original image semantics while\nintroducing controlled variability. Extensive experiments across multiple\nbenchmark datasets, including CIFAR-10, CIFAR-100, ImageNet-200, and\nImageNet-1K demonstrate LayerMixs superior performance in classification\naccuracy and substantially enhances critical Machine Learning (ML) safety\nmetrics, including resilience to natural image corruptions, robustness against\nadversarial attacks, improved model calibration and enhanced prediction\nconsistency. LayerMix represents a significant advancement toward developing\nmore reliable and adaptable artificial intelligence systems by addressing the\nfundamental challenges of deep learning generalization. The code is available\nat https://github.com/ahmadmughees/layermix.\n", "date": "2025-01-08", "categories": []}
{"id": "2501.06405", "title": "FocusDD: Real-World Scene Infusion for Robust Dataset Distillation", "abstract": "  Dataset distillation has emerged as a strategy to compress real-world\ndatasets for efficient training. However, it struggles with large-scale and\nhigh-resolution datasets, limiting its practicality. This paper introduces a\nnovel resolution-independent dataset distillation method Focus ed Dataset\nDistillation (FocusDD), which achieves diversity and realism in distilled data\nby identifying key information patches, thereby ensuring the generalization\ncapability of the distilled dataset across different network architectures.\nSpecifically, FocusDD leverages a pre-trained Vision Transformer (ViT) to\nextract key image patches, which are then synthesized into a single distilled\nimage. These distilled images, which capture multiple targets, are suitable not\nonly for classification tasks but also for dense tasks such as object\ndetection. To further improve the generalization of the distilled dataset, each\nsynthesized image is augmented with a downsampled view of the original image.\nExperimental results on the ImageNet-1K dataset demonstrate that, with 100\nimages per class (IPC), ResNet50 and MobileNet-v2 achieve validation accuracies\nof 71.0% and 62.6%, respectively, outperforming state-of-the-art methods by\n2.8% and 4.7%. Notably, FocusDD is the first method to use distilled datasets\nfor object detection tasks. On the COCO2017 dataset, with an IPC of 50,\nYOLOv11n and YOLOv11s achieve 24.4% and 32.1% mAP, respectively, further\nvalidating the effectiveness of our approach.\n", "date": "2025-01-11", "categories": []}
{"id": "2501.07314", "title": "FinerWeb-10BT: Refining Web Data with LLM-Based Line-Level Filtering", "abstract": "  Data quality is crucial for training Large Language Models (LLMs).\nTraditional heuristic filters often miss low-quality text or mistakenly remove\nvaluable content. In this paper, we introduce an LLM-based line-level filtering\nmethod to enhance training data quality. We use GPT-4o mini to label a\n20,000-document sample from FineWeb at the line level, allowing the model to\ncreate descriptive labels for low-quality lines. These labels are grouped into\nnine main categories, and we train a DeBERTa-v3 classifier to scale the\nfiltering to a 10B-token subset of FineWeb. To test the impact of our\nfiltering, we train GPT-2 models on both the original and the filtered\ndatasets. The results show that models trained on the filtered data achieve\nhigher accuracy on the HellaSwag benchmark and reach their performance targets\nfaster, even with up to 25\\% less data. This demonstrates that LLM-based\nline-level filtering can significantly improve data quality and training\nefficiency for LLMs. We release our quality-annotated dataset, FinerWeb-10BT,\nand the codebase to support further work in this area.\n", "date": "2025-01-13", "categories": []}
{"id": "2407.04103", "title": "Advances in Diffusion Models for Image Data Augmentation: A Review of\n  Methods, Models, Evaluation Metrics and Future Research Directions", "abstract": "  Image data augmentation constitutes a critical methodology in modern computer\nvision tasks, since it can facilitate towards enhancing the diversity and\nquality of training datasets; thereby, improving the performance and robustness\nof machine learning models in downstream tasks. In parallel, augmentation\napproaches can also be used for editing/modifying a given image in a context-\nand semantics-aware way. Diffusion Models (DMs), which comprise one of the most\nrecent and highly promising classes of methods in the field of generative\nArtificial Intelligence (AI), have emerged as a powerful tool for image data\naugmentation, capable of generating realistic and diverse images by learning\nthe underlying data distribution. The current study realizes a systematic,\ncomprehensive and in-depth review of DM-based approaches for image\naugmentation, covering a wide range of strategies, tasks and applications. In\nparticular, a comprehensive analysis of the fundamental principles, model\narchitectures and training strategies of DMs is initially performed.\nSubsequently, a taxonomy of the relevant image augmentation methods is\nintroduced, focusing on techniques regarding semantic manipulation,\npersonalization and adaptation, and application-specific augmentation tasks.\nThen, performance assessment methodologies and respective evaluation metrics\nare analyzed. Finally, current challenges and future research directions in the\nfield are discussed.\n", "date": "2024-07-04", "categories": []}
{"id": "2501.04155", "title": "MM-GEN: Enhancing Task Performance Through Targeted Multimodal Data\n  Curation", "abstract": "  Vision-language models (VLMs) are highly effective but often underperform on\nspecialized tasks; for example, Llava-1.5 struggles with chart and diagram\nunderstanding due to scarce task-specific training data. Existing training\ndata, sourced from general-purpose datasets, fails to capture the nuanced\ndetails needed for these tasks. We introduce MM-Gen, a scalable method that\ngenerates task-specific, high-quality synthetic text for candidate images by\nleveraging stronger models. MM-Gen employs a three-stage targeted process:\npartitioning data into subgroups, generating targeted text based on task\ndescriptions, and filtering out redundant and outlier data. Fine-tuning VLMs\nwith data generated by MM-Gen leads to significant performance gains, including\n29% on spatial reasoning and 15% on diagram understanding for Llava-1.5 (7B).\nCompared to human-curated caption data, MM-Gen achieves up to 1.6x better\nimprovements for the original models, proving its effectiveness in enhancing\ntask-specific VLM performance and bridging the gap between general-purpose\ndatasets and specialized requirements. Code available at\nhttps://github.com/sjoshi804/MM-Gen.\n", "date": "2025-01-07", "categories": []}
{"id": "2501.04202", "title": "Generative Dataset Distillation Based on Self-knowledge Distillation", "abstract": "  Dataset distillation is an effective technique for reducing the cost and\ncomplexity of model training while maintaining performance by compressing large\ndatasets into smaller, more efficient versions. In this paper, we present a\nnovel generative dataset distillation method that can improve the accuracy of\naligning prediction logits. Our approach integrates self-knowledge distillation\nto achieve more precise distribution matching between the synthetic and\noriginal data, thereby capturing the overall structure and relationships within\nthe data. To further improve the accuracy of alignment, we introduce a\nstandardization step on the logits before performing distribution matching,\nensuring consistency in the range of logits. Through extensive experiments, we\ndemonstrate that our method outperforms existing state-of-the-art methods,\nresulting in superior distillation performance.\n", "date": "2025-01-08", "categories": []}
{"id": "2309.04195", "title": "Towards Mitigating Architecture Overfitting on Distilled Datasets", "abstract": "  Dataset distillation methods have demonstrated remarkable performance for\nneural networks trained with very limited training data. However, a significant\nchallenge arises in the form of \\textit{architecture overfitting}: the\ndistilled training dataset synthesized by a specific network architecture\n(i.e., training network) generates poor performance when trained by other\nnetwork architectures (i.e., test networks), especially when the test networks\nhave a larger capacity than the training network. This paper introduces a\nseries of approaches to mitigate this issue. Among them, DropPath renders the\nlarge model to be an implicit ensemble of its sub-networks, and knowledge\ndistillation ensures each sub-network acts similarly to the small but\nwell-performing teacher network. These methods, characterized by their\nsmoothing effects, significantly mitigate architecture overfitting. We conduct\nextensive experiments to demonstrate the effectiveness and generality of our\nmethods. Particularly, across various scenarios involving different tasks and\ndifferent sizes of distilled data, our approaches significantly mitigate\narchitecture overfitting. Furthermore, our approaches achieve comparable or\neven superior performance when the test network is larger than the training\nnetwork.\n", "date": "2023-09-08", "categories": []}
{"id": "2501.02004", "title": "General Information Metrics for Improving AI Model Training Efficiency", "abstract": "  To address the growing size of AI model training data and the lack of a\nuniversal data selection methodology-factors that significantly drive up\ntraining costs -- this paper presents the General Information Metrics\nEvaluation (GIME) method. GIME leverages general information metrics from\nObjective Information Theory (OIT), including volume, delay, scope,\ngranularity, variety, duration, sampling rate, aggregation, coverage,\ndistortion, and mismatch to optimize dataset selection for training purposes.\nComprehensive experiments conducted across diverse domains, such as CTR\nPrediction, Civil Case Prediction, and Weather Forecasting, demonstrate that\nGIME effectively preserves model performance while substantially reducing both\ntraining time and costs. Additionally, applying GIME within the Judicial AI\nProgram led to a remarkable 39.56% reduction in total model training expenses,\nunderscoring its potential to support efficient and sustainable AI development.\n", "date": "2025-01-02", "categories": []}
{"id": "2501.02432", "title": "Swift Cross-Dataset Pruning: Enhancing Fine-Tuning Efficiency in Natural\n  Language Understanding", "abstract": "  Dataset pruning aims to select a subset of a dataset for efficient model\ntraining. While data efficiency in natural language processing has primarily\nfocused on within-corpus scenarios during model pre-training, efficient dataset\npruning for task-specific fine-tuning across diverse datasets remains\nchallenging due to variability in dataset sizes, data distributions, class\nimbalance and label spaces. Current cross-dataset pruning techniques for\nfine-tuning often rely on computationally expensive sample ranking processes,\ntypically requiring full dataset training or reference models. We address this\ngap by proposing Swift Cross-Dataset Pruning (SCDP). Specifically, our approach\nuses TF-IDF embeddings with geometric median to rapidly evaluate sample\nimportance. We then apply dataset size-adaptive pruning to ensure diversity:\nfor smaller datasets, we retain samples far from the geometric median, while\nfor larger ones, we employ distance-based stratified pruning. Experimental\nresults on six diverse datasets demonstrate the effectiveness of our method,\nspanning various tasks and scales while significantly reducing computational\nresources. Source code is available at:\nhttps://github.com/he-y/NLP-Dataset-Pruning\n", "date": "2025-01-05", "categories": []}
{"id": "2410.11303", "title": "TSDS: Data Selection for Task-Specific Model Finetuning", "abstract": "  Finetuning foundation models for specific tasks is an emerging paradigm in\nmodern machine learning. The efficacy of task-specific finetuning largely\ndepends on the selection of appropriate training data. We present TSDS\n(Task-Specific Data Selection), a framework to select data for task-specific\nmodel finetuning, guided by a small but representative set of examples from the\ntarget task. To do so, we formulate data selection for task-specific finetuning\nas an optimization problem with a distribution alignment loss based on optimal\ntransport to capture the discrepancy between the selected data and the target\ndistribution. In addition, we add a regularizer to encourage the diversity of\nthe selected data and incorporate kernel density estimation into the\nregularizer to reduce the negative effects of near-duplicates among the\ncandidate data. We connect our optimization problem to nearest neighbor search\nand design efficient algorithms to compute the optimal solution based on\napproximate nearest neighbor search techniques. We evaluate our method on data\nselection for both continued pretraining and instruction tuning of language\nmodels. We show that instruction tuning using data selected by our method with\na 1% selection ratio often outperforms using the full dataset and beats the\nbaseline selection methods by 1.5 points in F1 score on average.\n", "date": "2024-10-15", "categories": []}
{"id": "2412.18140", "title": "An Instrumental Value for Data Production and its Application to Data\n  Pricing", "abstract": "  How much value does a dataset or a data production process have to an agent\nwho wishes to use the data to assist decision-making? This is a fundamental\nquestion towards understanding the value of data as well as further pricing of\ndata. This paper develops an approach for capturing the instrumental value of\ndata production processes, which takes two key factors into account: (a) the\ncontext of the agent's decision-making problem; (b) prior data or information\nthe agent already possesses. We ''micro-found'' our valuation concepts by\nshowing how they connect to classic notions of information design and signals\nin information economics. When instantiated in the domain of Bayesian linear\nregression, our value naturally corresponds to information gain. Based on our\ndesigned data value, we then study a basic monopoly pricing setting with a\nbuyer looking to purchase from a seller some labeled data of a certain feature\ndirection in order to improve a Bayesian regression model. We show that when\nthe seller has the ability to fully customize any data request, she can extract\nthe first-best revenue (i.e., full surplus) from any population of buyers,\ni.e., achieving first-degree price discrimination. If the seller can only sell\ndata that are derived from an existing data pool, this limits her ability to\ncustomize, and achieving first-best revenue becomes generally impossible.\nHowever, we design a mechanism that achieves seller revenue at most $\\log\n(\\kappa)$ less than the first-best revenue, where $\\kappa$ is the condition\nnumber associated with the data matrix. A corollary of this result is that the\nseller can extract the first-best revenue in the multi-armed bandits special\ncase.\n", "date": "2024-12-24", "categories": []}
{"id": "2412.16250", "title": "Training-free Heterogeneous Graph Condensation via Data Selection", "abstract": "  Efficient training of large-scale heterogeneous graphs is of paramount\nimportance in real-world applications. However, existing approaches typically\nexplore simplified models to mitigate resource and time overhead, neglecting\nthe crucial aspect of simplifying large-scale heterogeneous graphs from the\ndata-centric perspective. Addressing this gap, HGCond introduces graph\ncondensation (GC) in heterogeneous graphs and generates a small condensed graph\nfor efficient model training. Despite its efficacy in graph generation, HGCond\nencounters two significant limitations. The first is low effectiveness, HGCond\nexcessively relies on the simplest relay model for the condensation procedure,\nwhich restricts the ability to exert powerful Heterogeneous Graph Neural\nNetworks (HGNNs) with flexible condensation ratio and limits the generalization\nability. The second is low efficiency, HGCond follows the existing GC methods\ndesigned for homogeneous graphs and leverages the sophisticated optimization\nparadigm, resulting in a time-consuming condensing procedure. In light of these\nchallenges, we present the first Training \\underline{Free} Heterogeneous Graph\nCondensation method, termed FreeHGC, facilitating both efficient and\nhigh-quality generation of heterogeneous condensed graphs. Specifically, we\nreformulate the heterogeneous graph condensation problem as a data selection\nissue, offering a new perspective for assessing and condensing representative\nnodes and edges in the heterogeneous graphs. By leveraging rich meta-paths, we\nintroduce a new, high-quality heterogeneous data selection criterion to select\ntarget-type nodes. Furthermore, two training-free condensation strategies for\nheterogeneous graphs are designed to condense and synthesize other-types nodes\neffectively.\n", "date": "2024-12-20", "categories": []}
{"id": "2412.16895", "title": "Adaptive Dataset Quantization", "abstract": "  Contemporary deep learning, characterized by the training of cumbersome\nneural networks on massive datasets, confronts substantial computational\nhurdles. To alleviate heavy data storage burdens on limited hardware resources,\nnumerous dataset compression methods such as dataset distillation (DD) and\ncoreset selection have emerged to obtain a compact but informative dataset\nthrough synthesis or selection for efficient training. However, DD involves an\nexpensive optimization procedure and exhibits limited generalization across\nunseen architectures, while coreset selection is limited by its low data keep\nratio and reliance on heuristics, hindering its practicality and feasibility.\nTo address these limitations, we introduce a newly versatile framework for\ndataset compression, namely Adaptive Dataset Quantization (ADQ). Specifically,\nwe first identify the sub-optimal performance of naive Dataset Quantization\n(DQ), which relies on uniform sampling and overlooks the varying importance of\neach generated bin. Subsequently, we propose a novel adaptive sampling strategy\nthrough the evaluation of generated bins' representativeness score, diversity\nscore and importance score, where the former two scores are quantified by the\ntexture level and contrastive learning-based techniques, respectively.\nExtensive experiments demonstrate that our method not only exhibits superior\ngeneralization capability across different architectures, but also attains\nstate-of-the-art results, surpassing DQ by average 3\\% on various datasets.\n", "date": "2024-12-22", "categories": []}
{"id": "2412.17008", "title": "Data value estimation on private gradients", "abstract": "  For gradient-based machine learning (ML) methods commonly adopted in practice\nsuch as stochastic gradient descent, the de facto differential privacy (DP)\ntechnique is perturbing the gradients with random Gaussian noise. Data\nvaluation attributes the ML performance to the training data and is widely used\nin privacy-aware applications that require enforcing DP such as data pricing,\ncollaborative ML, and federated learning (FL). Can existing data valuation\nmethods still be used when DP is enforced via gradient perturbations? We show\nthat the answer is no with the default approach of injecting i.i.d.~random\nnoise to the gradients because the estimation uncertainty of the data value\nestimation paradoxically linearly scales with more estimation budget, producing\nestimates almost like random guesses. To address this issue, we propose to\ninstead inject carefully correlated noise to provably remove the linear scaling\nof estimation uncertainty w.r.t.~the budget. We also empirically demonstrate\nthat our method gives better data value estimates on various ML tasks and is\napplicable to use cases including dataset valuation and~FL.\n", "date": "2024-12-22", "categories": []}
{"id": "2412.17069", "title": "Optimizing Data Curation through Spectral Analysis and Joint Batch\n  Selection (SALN)", "abstract": "  In modern deep learning models, long training times and large datasets\npresent significant challenges to both efficiency and scalability. Effective\ndata curation and sample selection are crucial for optimizing the training\nprocess of deep neural networks. This paper introduces SALN, a method designed\nto prioritize and select samples within each batch rather than from the entire\ndataset. By utilizing jointly selected batches, SALN enhances training\nefficiency compared to independent batch selection. The proposed method applies\na spectral analysis-based heuristic to identify the most informative data\npoints within each batch, improving both training speed and accuracy. The SALN\nalgorithm significantly reduces training time and enhances accuracy when\ncompared to traditional batch prioritization or standard training procedures.\nIt demonstrates up to an 8x reduction in training time and up to a 5\\% increase\nin accuracy over standard training methods. Moreover, SALN achieves better\nperformance and shorter training times compared to Google's JEST method\ndeveloped by DeepMind.\n", "date": "2024-12-22", "categories": []}
{"id": "2412.17355", "title": "Bi-Directional Multi-Scale Graph Dataset Condensation via Information\n  Bottleneck", "abstract": "  Dataset condensation has significantly improved model training efficiency,\nbut its application on devices with different computing power brings new\nrequirements for different data sizes. Thus, condensing multiple scale graphs\nsimultaneously is the core of achieving efficient training in different\non-device scenarios. Existing efficient works for multi-scale graph dataset\ncondensation mainly perform efficient approximate computation in scale order\n(large-to-small or small-to-large scales). However, for non-Euclidean\nstructures of sparse graph data, these two commonly used paradigms for\nmulti-scale graph dataset condensation have serious scaling down degradation\nand scaling up collapse problems of a graph. The main bottleneck of the above\nparadigms is whether the effective information of the original graph is fully\npreserved when consenting to the primary sub-scale (the first of multiple\nscales), which determines the condensation effect and consistency of all\nscales. In this paper, we proposed a novel GNN-centric Bi-directional\nMulti-Scale Graph Dataset Condensation (BiMSGC) framework, to explore unifying\nparadigms by operating on both large-to-small and small-to-large for\nmulti-scale graph condensation. Based on the mutual information theory, we\nestimate an optimal ``meso-scale'' to obtain the minimum necessary dense graph\npreserving the maximum utility information of the original graph, and then we\nachieve stable and consistent ``bi-directional'' condensation learning by\noptimizing graph eigenbasis matching with information bottleneck on other\nscales. Encouraging empirical results on several datasets demonstrates the\nsignificant superiority of the proposed framework in graph condensation at\ndifferent scales.\n", "date": "2024-12-23", "categories": []}
{"id": "2405.19547", "title": "CLIPLoss and Norm-Based Data Selection Methods for Multimodal\n  Contrastive Learning", "abstract": "  Data selection has emerged as a core issue for large-scale visual-language\nmodel pretaining (e.g., CLIP), particularly with noisy web-curated datasets.\nThree main data selection approaches are: (1) leveraging external non-CLIP\nmodels to aid data selection, (2) training new CLIP-style embedding models that\nare more effective at selecting high-quality data than the original OpenAI CLIP\nmodel, and (3) designing better metrics or strategies universally applicable to\nany CLIP embedding without requiring specific model properties (e.g., CLIPScore\nis one popular metric). While the first two approaches have been extensively\nstudied, the third remains under-explored. In this paper, we advance the third\napproach by proposing two new methods. Firstly, instead of classical CLIP\nscores that only consider the alignment between two modalities from a single\nsample, we introduce surrogate-CLIPLoss (s-CLIPLoss), a CLIP loss-inspired\nmethod that adds the alignment between one sample and its contrastive pairs as\nan extra normalization term for better quality measurement. Secondly, when\ndownstream tasks are known, we propose a new norm-based metric, NormSim, to\nmeasure the similarity between pretraining data and target data. We test our\nmethods on the data selection benchmark, DataComp~\\cite{gadre2023datacomp}.\nCompared to the best baseline using only OpenAI's CLIP-L/14, our methods\nachieve a 5.3\\% improvement on ImageNet-1k and a 2.8\\% improvement on 38\ndownstream evaluation tasks. Moreover, both s-CLIPLoss and NormSim are\ncompatible with existing techniques. By combining our methods with the current\nbest methods DFN and HYPE, we can boost average performance on downstream tasks\nby 0.9\\%, achieving a new state-of-the-art on the DataComp-medium benchmark.\n", "date": "2024-05-29", "categories": []}
{"id": "2412.15358", "title": "Dataset Augmentation by Mixing Visual Concepts", "abstract": "  This paper proposes a dataset augmentation method by fine-tuning pre-trained\ndiffusion models. Generating images using a pre-trained diffusion model with\ntextual conditioning often results in domain discrepancy between real data and\ngenerated images. We propose a fine-tuning approach where we adapt the\ndiffusion model by conditioning it with real images and novel text embeddings.\nWe introduce a unique procedure called Mixing Visual Concepts (MVC) where we\ncreate novel text embeddings from image captions. The MVC enables us to\ngenerate multiple images which are diverse and yet similar to the real data\nenabling us to perform effective dataset augmentation. We perform comprehensive\nqualitative and quantitative evaluations with the proposed dataset augmentation\napproach showcasing both coarse-grained and finegrained changes in generated\nimages. Our approach outperforms state-of-the-art augmentation techniques on\nbenchmark classification tasks.\n", "date": "2024-12-19", "categories": []}
{"id": "2310.13008", "title": "DavIR: Data Selection via Implicit Reward for Large Language Models", "abstract": "  We introduce DavIR, a model-based data selection method for post-training\nLarge Language Models. DavIR generalizes Reducible Holdout Loss to core-set\nselection problem of causal language modeling, and quantifies the learnability\nof a given datum with respect to a pre-trained LLM based on relative reduction\nin loss during fine-tuning, a metric we show to be closely related to the\nimplicit reward model described in Direct Preference Optimization (DPO). We\nshow that 6% of Alpaca dataset selected with DavIR can steer both the LLaMA and\nGemma model family to produce superior performance compared to the same models\ntrained on the full 52K dataset. We also show that Alpaca dataset compressed\nwith DavIR can be combined with GSM8K dataset to effectively balance\nopen-domain freeform QA and mathematical reasoning capabilities. Finally, we\napply the DavIR objective to DPO and develop a normalized DavIR-DPO objective\nwhich improves alignment performance of Zephyr-7B-SFT model by 8% (relative) on\nAlpacaEval, compared against training on vanilla DPO objective.\n", "date": "2023-10-16", "categories": []}
{"id": "2407.02112", "title": "A Data-Centric Perspective on Evaluating Machine Learning Models for\n  Tabular Data", "abstract": "  Tabular data is prevalent in real-world machine learning applications, and\nnew models for supervised learning of tabular data are frequently proposed.\nComparative studies assessing the performance of models typically consist of\nmodel-centric evaluation setups with overly standardized data preprocessing.\nThis paper demonstrates that such model-centric evaluations are biased, as\nreal-world modeling pipelines often require dataset-specific preprocessing and\nfeature engineering. Therefore, we propose a data-centric evaluation framework.\nWe select 10 relevant datasets from Kaggle competitions and implement\nexpert-level preprocessing pipelines for each dataset. We conduct experiments\nwith different preprocessing pipelines and hyperparameter optimization (HPO)\nregimes to quantify the impact of model selection, HPO, feature engineering,\nand test-time adaptation. Our main findings are: 1. After dataset-specific\nfeature engineering, model rankings change considerably, performance\ndifferences decrease, and the importance of model selection reduces. 2. Recent\nmodels, despite their measurable progress, still significantly benefit from\nmanual feature engineering. This holds true for both tree-based models and\nneural networks. 3. While tabular data is typically considered static, samples\nare often collected over time, and adapting to distribution shifts can be\nimportant even in supposedly static data. These insights suggest that research\nefforts should be directed toward a data-centric perspective, acknowledging\nthat tabular data requires feature engineering and often exhibits temporal\ncharacteristics. Our framework is available under:\nhttps://github.com/atschalz/dc_tabeval.\n", "date": "2024-07-02", "categories": []}
{"id": "2412.04158", "title": "LossVal: Efficient Data Valuation for Neural Networks", "abstract": "  Assessing the importance of individual training samples is a key challenge in\nmachine learning. Traditional approaches retrain models with and without\nspecific samples, which is computationally expensive and ignores dependencies\nbetween data points. We introduce LossVal, an efficient data valuation method\nthat computes importance scores during neural network training by embedding a\nself-weighting mechanism into loss functions like cross-entropy and mean\nsquared error. LossVal reduces computational costs, making it suitable for\nlarge datasets and practical applications. Experiments on classification and\nregression tasks across multiple datasets show that LossVal effectively\nidentifies noisy samples and is able to distinguish helpful from harmful\nsamples. We examine the gradient calculation of LossVal to highlight its\nadvantages. The source code is available at:\nhttps://github.com/twibiral/LossVal\n", "date": "2024-12-05", "categories": []}
{"id": "2412.09990", "title": "Small Language Model as Data Prospector for Large Language Model", "abstract": "  The quality of instruction data directly affects the performance of\nfine-tuned Large Language Models (LLMs). Previously, \\cite{li2023one} proposed\n\\texttt{NUGGETS}, which identifies and selects high-quality quality data from a\nlarge dataset by identifying those individual instruction examples that can\nsignificantly improve the performance of different tasks after being learnt as\none-shot instances. In this work, we propose \\texttt{SuperNUGGETS}, an improved\nvariant of \\texttt{NUGGETS} optimised for efficiency and performance. Our\n\\texttt{SuperNUGGETS} uses a small language model (SLM) instead of a large\nlanguage model (LLM) to filter the data for outstanding one-shot instances and\nrefines the predefined set of tests. The experimental results show that the\nperformance of \\texttt{SuperNUGGETS} only decreases by 1-2% compared to\n\\texttt{NUGGETS}, but the efficiency can be increased by a factor of 58.\nCompared to the original \\texttt{NUGGETS}, our \\texttt{SuperNUGGETS} has a\nhigher utility value due to the significantly lower resource consumption.\n", "date": "2024-12-13", "categories": []}
{"id": "2412.09538", "title": "Capturing the Temporal Dependence of Training Data Influence", "abstract": "  Traditional data influence estimation methods, like influence function,\nassume that learning algorithms are permutation-invariant with respect to\ntraining data. However, modern training paradigms, especially for foundation\nmodels using stochastic algorithms and multi-stage curricula, are sensitive to\ndata ordering, thus violating this assumption. This mismatch renders influence\nfunctions inadequate for answering a critical question in machine learning: How\ncan we capture the dependence of data influence on the optimization trajectory\nduring training? To address this gap, we formalize the concept of\ntrajectory-specific leave-one-out (LOO) influence, which quantifies the impact\nof removing a data point from a specific iteration during training, accounting\nfor the exact sequence of data encountered and the model's optimization\ntrajectory. However, exactly evaluating the trajectory-specific LOO presents a\nsignificant computational challenge. To address this, we propose data value\nembedding, a novel technique enabling efficient approximation of\ntrajectory-specific LOO. Specifically, we compute a training data embedding\nthat encapsulates the cumulative interactions between data and the evolving\nmodel parameters. The LOO can then be efficiently approximated through a simple\ndot-product between the data value embedding and the gradient of the given test\ndata. As data value embedding captures training data ordering, it offers\nvaluable insights into model training dynamics. In particular, we uncover\ndistinct phases of data influence, revealing that data points in the early and\nlate stages of training exert a greater impact on the final model. These\ninsights translate into actionable strategies for managing the computational\noverhead of data selection by strategically timing the selection process,\npotentially opening new avenues in data curation research.\n", "date": "2024-12-12", "categories": []}
{"id": "2411.11329", "title": "Color-Oriented Redundancy Reduction in Dataset Distillation", "abstract": "  Dataset Distillation (DD) is designed to generate condensed representations\nof extensive image datasets, enhancing training efficiency. Despite recent\nadvances, there remains considerable potential for improvement, particularly in\naddressing the notable redundancy within the color space of distilled images.\nIn this paper, we propose AutoPalette, a framework that minimizes color\nredundancy at the individual image and overall dataset levels, respectively. At\nthe image level, we employ a palette network, a specialized neural network, to\ndynamically allocate colors from a reduced color space to each pixel. The\npalette network identifies essential areas in synthetic images for model\ntraining and consequently assigns more unique colors to them. At the dataset\nlevel, we develop a color-guided initialization strategy to minimize redundancy\namong images. Representative images with the least replicated color patterns\nare selected based on the information gain. A comprehensive performance study\ninvolving various datasets and evaluation scenarios is conducted, demonstrating\nthe superior performance of our proposed color-aware DD compared to existing DD\nmethods. The code is available at\n\\url{https://github.com/KeViNYuAn0314/AutoPalette}.\n", "date": "2024-11-18", "categories": []}
{"id": "2412.04668", "title": "Diffusion-Augmented Coreset Expansion for Scalable Dataset Distillation", "abstract": "  With the rapid scaling of neural networks, data storage and communication\ndemands have intensified. Dataset distillation has emerged as a promising\nsolution, condensing information from extensive datasets into a compact set of\nsynthetic samples by solving a bilevel optimization problem. However, current\nmethods face challenges in computational efficiency, particularly with\nhigh-resolution data and complex architectures. Recently,\nknowledge-distillation-based dataset condensation approaches have made this\nprocess more computationally feasible. Yet, with the recent developments of\ngenerative foundation models, there is now an opportunity to achieve even\ngreater compression, enhance the quality of distilled data, and introduce\nvaluable diversity into the data representation. In this work, we propose a\ntwo-stage solution. First, we compress the dataset by selecting only the most\ninformative patches to form a coreset. Next, we leverage a generative\nfoundation model to dynamically expand this compressed set in real-time,\nenhancing the resolution of these patches and introducing controlled\nvariability to the coreset. Our extensive experiments demonstrate the\nrobustness and efficiency of our approach across a range of dataset\ndistillation benchmarks. We demonstrate a significant improvement of over 10%\ncompared to the state-of-the-art on several large-scale dataset distillation\nbenchmarks. The code will be released soon.\n", "date": "2024-12-05", "categories": []}
{"id": "2412.04748", "title": "Decomposed Distribution Matching in Dataset Condensation", "abstract": "  Dataset Condensation (DC) aims to reduce deep neural networks training\nefforts by synthesizing a small dataset such that it will be as effective as\nthe original large dataset. Conventionally, DC relies on a costly bi-level\noptimization which prohibits its practicality. Recent research formulates DC as\na distribution matching problem which circumvents the costly bi-level\noptimization. However, this efficiency sacrifices the DC performance. To\ninvestigate this performance degradation, we decomposed the dataset\ndistribution into content and style. Our observations indicate two major\nshortcomings of: 1) style discrepancy between original and condensed data, and\n2) limited intra-class diversity of condensed dataset. We present a simple yet\neffective method to match the style information between original and condensed\ndata, employing statistical moments of feature maps as well-established style\nindicators. Moreover, we enhance the intra-class diversity by maximizing the\nKullback-Leibler divergence within each synthetic class, i.e., content. We\ndemonstrate the efficacy of our method through experiments on diverse datasets\nof varying size and resolution, achieving improvements of up to 4.1% on\nCIFAR10, 4.2% on CIFAR100, 4.3% on TinyImageNet, 2.0% on ImageNet-1K, 3.3% on\nImageWoof, 2.5% on ImageNette, and 5.5% in continual learning accuracy.\n", "date": "2024-12-06", "categories": []}
{"id": "2412.04025", "title": "Exploring the Influence of Label Aggregation on Minority Voices:\n  Implications for Dataset Bias and Model Training", "abstract": "  Resolving disagreement in manual annotation typically consists of removing\nunreliable annotators and using a label aggregation strategy such as majority\nvote or expert opinion to resolve disagreement. These may have the side-effect\nof silencing or under-representing minority but equally valid opinions. In this\npaper, we study the impact of standard label aggregation strategies on minority\nopinion representation in sexism detection. We investigate the quality and\nvalue of minority annotations, and then examine their effect on the class\ndistributions in gold labels, as well as how this affects the behaviour of\nmodels trained on the resulting datasets. Finally, we discuss the potential\nbiases introduced by each method and how they can be amplified by the models.\n", "date": "2024-12-05", "categories": []}
{"id": "2412.03398", "title": "RedStone: Curating General, Code, Math, and QA Data for Large Language\n  Models", "abstract": "  Pre-training Large Language Models (LLMs) on high-quality, meticulously\ncurated datasets is widely recognized as critical for enhancing their\nperformance and generalization capabilities. This study explores the untapped\npotential of Common Crawl as a comprehensive and flexible resource for\npre-training LLMs, addressing both general-purpose language understanding and\nspecialized domain knowledge. We introduce RedStone, an innovative and scalable\npipeline engineered to extract and process data from Common Crawl, facilitating\nthe creation of extensive and varied pre-training datasets. Unlike traditional\ndatasets, which often require expensive curation and domain-specific expertise,\nRedStone leverages the breadth of Common Crawl to deliver datasets tailored to\na wide array of domains. In this work, we exemplify its capability by\nconstructing pre-training datasets across multiple fields, including general\nlanguage understanding, code, mathematics, and question-answering tasks. The\nflexibility of RedStone allows for easy adaptation to other specialized\ndomains, significantly lowering the barrier to creating valuable\ndomain-specific datasets. Our findings demonstrate that Common Crawl, when\nharnessed through effective pipelines like RedStone, can serve as a rich,\nrenewable source of pre-training data, unlocking new avenues for domain\nadaptation and knowledge discovery in LLMs. This work also underscores the\nimportance of innovative data acquisition strategies and highlights the role of\nweb-scale data as a powerful resource in the continued evolution of LLMs.\nRedStone code and data samples will be publicly available at\n\\url{https://aka.ms/redstone}.\n", "date": "2024-12-04", "categories": []}
{"id": "2412.01876", "title": "Understanding Bias in Large-Scale Visual Datasets", "abstract": "  A recent study has shown that large-scale visual datasets are very biased:\nthey can be easily classified by modern neural networks. However, the concrete\nforms of bias among these datasets remain unclear. In this study, we propose a\nframework to identify the unique visual attributes distinguishing these\ndatasets. Our approach applies various transformations to extract semantic,\nstructural, boundary, color, and frequency information from datasets, and\nassess how much each type of information reflects their bias. We further\ndecompose their semantic bias with object-level analysis, and leverage natural\nlanguage methods to generate detailed, open-ended descriptions of each\ndataset's characteristics. Our work aims to help researchers understand the\nbias in existing large-scale pre-training datasets, and build more diverse and\nrepresentative ones in the future. Our project page and code are available at\nhttp://boyazeng.github.io/understand_bias .\n", "date": "2024-12-02", "categories": []}
{"id": "2412.00111", "title": "Video Set Distillation: Information Diversification and Temporal\n  Densification", "abstract": "  The rapid development of AI models has led to a growing emphasis on enhancing\ntheir capabilities for complex input data such as videos. While large-scale\nvideo datasets have been introduced to support this growth, the unique\nchallenges of reducing redundancies in video \\textbf{sets} have not been\nexplored. Compared to image datasets or individual videos, video \\textbf{sets}\nhave a two-layer nested structure, where the outer layer is the collection of\nindividual videos, and the inner layer contains the correlations among\nframe-level data points to provide temporal information. Video \\textbf{sets}\nhave two dimensions of redundancies: within-sample and inter-sample\nredundancies. Existing methods like key frame selection, dataset pruning or\ndataset distillation are not addressing the unique challenge of video sets\nsince they aimed at reducing redundancies in only one of the dimensions. In\nthis work, we are the first to study Video Set Distillation, which synthesizes\noptimized video data by jointly addressing within-sample and inter-sample\nredundancies. Our Information Diversification and Temporal Densification (IDTD)\nmethod jointly reduces redundancies across both dimensions. This is achieved\nthrough a Feature Pool and Feature Selectors mechanism to preserve inter-sample\ndiversity, alongside a Temporal Fusor that maintains temporal information\ndensity within synthesized videos. Our method achieves state-of-the-art results\nin Video Dataset Distillation, paving the way for more effective redundancy\nreduction and efficient AI model training on video datasets.\n", "date": "2024-11-28", "categories": []}
{"id": "2312.14626", "title": "DSAP: Analyzing Bias Through Demographic Comparison of Datasets", "abstract": "  In the last few years, Artificial Intelligence systems have become\nincreasingly widespread. Unfortunately, these systems can share many biases\nwith human decision-making, including demographic biases. Often, these biases\ncan be traced back to the data used for training, where large uncurated\ndatasets have become the norm. Despite our knowledge of these biases, we still\nlack general tools to detect and quantify them, as well as to compare the\nbiases in different datasets. Thus, in this work, we propose DSAP (Demographic\nSimilarity from Auxiliary Profiles), a two-step methodology for comparing the\ndemographic composition of two datasets. DSAP can be deployed in three key\napplications: to detect and characterize demographic blind spots and bias\nissues across datasets, to measure dataset demographic bias in single datasets,\nand to measure dataset demographic shift in deployment scenarios. An essential\nfeature of DSAP is its ability to robustly analyze datasets without explicit\ndemographic labels, offering simplicity and interpretability for a wide range\nof situations. To show the usefulness of the proposed methodology, we consider\nthe Facial Expression Recognition task, where demographic bias has previously\nbeen found. The three applications are studied over a set of twenty datasets\nwith varying properties. The code is available at\nhttps://github.com/irisdominguez/DSAP.\n", "date": "2023-12-22", "categories": []}
{"id": "2411.11924", "title": "Dataset Distillers Are Good Label Denoisers In the Wild", "abstract": "  Learning from noisy data has become essential for adapting deep learning\nmodels to real-world applications. Traditional methods often involve first\nevaluating the noise and then applying strategies such as discarding noisy\nsamples, re-weighting, or re-labeling. However, these methods can fall into a\nvicious cycle when the initial noise evaluation is inaccurate, leading to\nsuboptimal performance. To address this, we propose a novel approach that\nleverages dataset distillation for noise removal. This method avoids the\nfeedback loop common in existing techniques and enhances training efficiency,\nwhile also providing strong privacy protection through offline processing. We\nrigorously evaluate three representative dataset distillation methods (DATM,\nDANCE, and RCIG) under various noise conditions, including symmetric noise,\nasymmetric noise, and real-world natural noise. Our empirical findings reveal\nthat dataset distillation effectively serves as a denoising tool in random\nnoise scenarios but may struggle with structured asymmetric noise patterns,\nwhich can be absorbed into the distilled samples. Additionally, clean but\nchallenging samples, such as those from tail classes in imbalanced datasets,\nmay undergo lossy compression during distillation. Despite these challenges,\nour results highlight that dataset distillation holds significant promise for\nrobust model training, especially in high-privacy environments where noise is\nprevalent. The source code is available at https://github.com/Kciiiman/DD_LNL.\n", "date": "2024-11-18", "categories": []}
{"id": "2311.18838", "title": "Dataset Distillation via Curriculum Data Synthesis in Large Data Era", "abstract": "  Dataset distillation or condensation aims to generate a smaller but\nrepresentative subset from a large dataset, which allows a model to be trained\nmore efficiently, meanwhile evaluating on the original testing data\ndistribution to achieve decent performance. Previous decoupled methods like\nSRe$^2$L simply use a unified gradient update scheme for synthesizing data from\nGaussian noise, while, we notice that the initial several update iterations\nwill determine the final outline of synthesis, thus an improper gradient update\nstrategy may dramatically affect the final generation quality. To address this,\nwe introduce a simple yet effective global-to-local gradient refinement\napproach enabled by curriculum data augmentation ($\\texttt{CDA}$) during data\nsynthesis. The proposed framework achieves the current published highest\naccuracy on both large-scale ImageNet-1K and 21K with 63.2% under IPC (Images\nPer Class) 50 and 36.1% under IPC 20, using a regular input resolution of\n224$\\times$224 with faster convergence speed and less synthetic time. The\nproposed model outperforms the current state-of-the-art methods like SRe$^2$L,\nTESLA, and MTT by more than 4% Top-1 accuracy on ImageNet-1K/21K and for the\nfirst time, reduces the gap to its full-data training counterparts to less than\nabsolute 15%. Moreover, this work represents the inaugural success in dataset\ndistillation on the larger-scale ImageNet-21K dataset under the standard\n224$\\times$224 resolution. Our code and distilled ImageNet-21K dataset of 20\nIPC, 2K recovery budget are available at\nhttps://github.com/VILA-Lab/SRe2L/tree/main/CDA.\n", "date": "2023-11-30", "categories": []}
{"id": "2312.00486", "title": "REDUCR: Robust Data Downsampling Using Class Priority Reweighting", "abstract": "  Modern machine learning models are becoming increasingly expensive to train\nfor real-world image and text classification tasks, where massive web-scale\ndata is collected in a streaming fashion. To reduce the training cost, online\nbatch selection techniques have been developed to choose the most informative\ndatapoints. However, these techniques can suffer from poor worst-class\ngeneralization performance due to class imbalance and distributional shifts.\nThis work introduces REDUCR, a robust and efficient data downsampling method\nthat uses class priority reweighting. REDUCR reduces the training data while\npreserving worst-class generalization performance. REDUCR assigns priority\nweights to datapoints in a class-aware manner using an online learning\nalgorithm. We demonstrate the data efficiency and robust performance of REDUCR\non vision and text classification tasks. On web-scraped datasets with\nimbalanced class distributions, REDUCR significantly improves worst-class test\naccuracy (and average accuracy), surpassing state-of-the-art methods by around\n15%.\n", "date": "2023-12-01", "categories": []}
{"id": "2411.16128", "title": "CIA: Controllable Image Augmentation Framework Based on Stable Diffusion", "abstract": "  Computer vision tasks such as object detection and segmentation rely on the\navailability of extensive, accurately annotated datasets. In this work, We\npresent CIA, a modular pipeline, for (1) generating synthetic images for\ndataset augmentation using Stable Diffusion, (2) filtering out low quality\nsamples using defined quality metrics, (3) forcing the existence of specific\npatterns in generated images using accurate prompting and ControlNet. In order\nto show how CIA can be used to search for an optimal augmentation pipeline of\ntraining data, we study human object detection in a data constrained scenario,\nusing YOLOv8n on COCO and Flickr30k datasets. We have recorded significant\nimprovement using CIA-generated images, approaching the performances obtained\nwhen doubling the amount of real images in the dataset. Our findings suggest\nthat our modular framework can significantly enhance object detection systems,\nand make it possible for future research to be done on data-constrained\nscenarios. The framework is available at: github.com/multitel-ai/CIA.\n", "date": "2024-11-25", "categories": []}
{"id": "2409.13548", "title": "Data Diet: Can Trimming PET/CT Datasets Enhance Lesion Segmentation?", "abstract": "  In this work, we describe our approach to compete in the autoPET3 datacentric\ntrack. While conventional wisdom suggests that larger datasets lead to better\nmodel performance, recent studies indicate that excluding certain training\nsamples can enhance model accuracy. We find that in the autoPETIII dataset, a\nmodel that is trained on the entire dataset exhibits undesirable\ncharacteristics by producing a large number of false positives particularly for\nPSMA-PETs. We counteract this by removing the easiest samples from the training\ndataset as measured by the model loss before retraining from scratch. Using the\nproposed approach we manage to drive down the false negative volume and improve\nupon the baseline model in both false negative volume and dice score on the\npreliminary test set. Code and pre-trained models are available at\ngithub.com/alexanderjaus/autopet3_datadiet.\n", "date": "2024-09-20", "categories": []}
{"id": "2411.14497", "title": "Star-Agents: Automatic Data Optimization with LLM Agents for Instruction\n  Tuning", "abstract": "  The efficacy of large language models (LLMs) on downstream tasks usually\nhinges on instruction tuning, which relies critically on the quality of\ntraining data. Unfortunately, collecting high-quality and diverse data is both\nexpensive and time-consuming. To mitigate this issue, we propose a novel\nStar-Agents framework, which automates the enhancement of data quality across\ndatasets through multi-agent collaboration and assessment. The framework adopts\na three-pronged strategy. It initially generates diverse instruction data with\nmultiple LLM agents through a bespoke sampling method. Subsequently, the\ngenerated data undergo a rigorous evaluation using a dual-model method that\nassesses both difficulty and quality. Finaly, the above process evolves in a\ndynamic refinement phase, where more effective LLMs are prioritized, enhancing\nthe overall data quality. Our empirical studies, including instruction tuning\nexperiments with models such as Pythia and LLaMA, demonstrate the effectiveness\nof the proposed framework. Optimized datasets have achieved substantial\nimprovements, with an average increase of 12% and notable gains in specific\nmetrics, such as a 40% improvement in Fermi, as evidenced by benchmarks like\nMT-bench, Vicuna bench, and WizardLM testset.\n", "date": "2024-11-21", "categories": []}
{"id": "2310.04987", "title": "Data-centric Graph Learning: A Survey", "abstract": "  The history of artificial intelligence (AI) has witnessed the significant\nimpact of high-quality data on various deep learning models, such as ImageNet\nfor AlexNet and ResNet. Recently, instead of designing more complex neural\narchitectures as model-centric approaches, the attention of AI community has\nshifted to data-centric ones, which focuses on better processing data to\nstrengthen the ability of neural models. Graph learning, which operates on\nubiquitous topological data, also plays an important role in the era of deep\nlearning. In this survey, we comprehensively review graph learning approaches\nfrom the data-centric perspective, and aim to answer three crucial questions:\n(1) when to modify graph data, (2) what part of the graph data needs\nmodification to unlock the potential of various graph models, and (3) how to\nsafeguard graph models from problematic data influence. Accordingly, we propose\na novel taxonomy based on the stages in the graph learning pipeline, and\nhighlight the processing methods for different data structures in the graph\ndata, i.e., topology, feature and label. Furthermore, we analyze some potential\nproblems embedded in graph data and discuss how to solve them in a data-centric\nmanner. Finally, we provide some promising future directions for data-centric\ngraph learning.\n", "date": "2023-10-08", "categories": []}
{"id": "2408.16266", "title": "Inversion Circle Interpolation: Diffusion-based Image Augmentation for\n  Data-scarce Classification", "abstract": "  Data Augmentation (DA), i.e., synthesizing faithful and diverse samples to\nexpand the original training set, is a prevalent and effective strategy to\nimprove the performance of various data-scarce tasks. With the powerful image\ngeneration ability, diffusion-based DA has shown strong performance gains on\ndifferent image classification benchmarks. In this paper, we analyze today's\ndiffusion-based DA methods, and argue that they cannot take account of both\nfaithfulness and diversity, which are two critical keys for generating\nhigh-quality samples and boosting classification performance. To this end, we\npropose a novel Diffusion-based DA method: Diff-II. Specifically, it consists\nof three steps: 1) Category concepts learning: Learning concept embeddings for\neach category. 2) Inversion interpolation: Calculating the inversion for each\nimage, and conducting circle interpolation for two randomly sampled inversions\nfrom the same category. 3) Two-stage denoising: Using different prompts to\ngenerate synthesized images in a coarse-to-fine manner. Extensive experiments\non various data-scarce image classification tasks (e.g., few-shot, long-tailed,\nand out-of-distribution classification) have demonstrated its effectiveness\nover state-of-the-art diffusion-based DA methods.\n", "date": "2024-08-29", "categories": []}
{"id": "2411.12841", "title": "Data-to-Model Distillation: Data-Efficient Learning Framework", "abstract": "  Dataset distillation aims to distill the knowledge of a large-scale real\ndataset into small yet informative synthetic data such that a model trained on\nit performs as well as a model trained on the full dataset. Despite recent\nprogress, existing dataset distillation methods often struggle with\ncomputational efficiency, scalability to complex high-resolution datasets, and\ngeneralizability to deep architectures. These approaches typically require\nretraining when the distillation ratio changes, as knowledge is embedded in raw\npixels. In this paper, we propose a novel framework called Data-to-Model\nDistillation (D2M) to distill the real dataset's knowledge into the learnable\nparameters of a pre-trained generative model by aligning rich representations\nextracted from real and generated images. The learned generative model can then\nproduce informative training images for different distillation ratios and deep\narchitectures. Extensive experiments on 15 datasets of varying resolutions show\nD2M's superior performance, re-distillation efficiency, and cross-architecture\ngeneralizability. Our method effectively scales up to high-resolution 128x128\nImageNet-1K. Furthermore, we verify D2M's practical benefits for downstream\napplications in neural architecture search.\n", "date": "2024-11-19", "categories": []}
{"id": "2409.17612", "title": "Diversity-Driven Synthesis: Enhancing Dataset Distillation through\n  Directed Weight Adjustment", "abstract": "  The sharp increase in data-related expenses has motivated research into\ncondensing datasets while retaining the most informative features. Dataset\ndistillation has thus recently come to the fore. This paradigm generates\nsynthetic datasets that are representative enough to replace the original\ndataset in training a neural network. To avoid redundancy in these synthetic\ndatasets, it is crucial that each element contains unique features and remains\ndiverse from others during the synthesis stage. In this paper, we provide a\nthorough theoretical and empirical analysis of diversity within synthesized\ndatasets. We argue that enhancing diversity can improve the parallelizable yet\nisolated synthesizing approach. Specifically, we introduce a novel method that\nemploys dynamic and directed weight adjustment techniques to modulate the\nsynthesis process, thereby maximizing the representativeness and diversity of\neach synthetic instance. Our method ensures that each batch of synthetic data\nmirrors the characteristics of a large, varying subset of the original dataset.\nExtensive experiments across multiple datasets, including CIFAR, Tiny-ImageNet,\nand ImageNet-1K, demonstrate the superior performance of our method,\nhighlighting its effectiveness in producing diverse and representative\nsynthetic datasets with minimal computational expense. Our code is available at\nhttps://github.com/AngusDujw/Diversity-Driven-Synthesis.https://github.com/AngusDujw/Diversity-Driven-Synthesis.\n", "date": "2024-09-26", "categories": []}
{"id": "2402.18191", "title": "Clustering and Ranking: Diversity-preserved Instruction Selection\n  through Expert-aligned Quality Estimation", "abstract": "  With contributions from the open-source community, a vast amount of\ninstruction tuning (IT) data has emerged. Given the significant resource\nallocation required for training and evaluating models, it is advantageous to\nhave an efficient method for selecting high-quality IT data. However, existing\nmethods for instruction data selection have limitations such as relying on\nfragile external APIs, being affected by biases in GPT models, or reducing the\ndiversity of the selected instruction dataset. In this paper, we propose an\nindustrial-friendly, expert-aligned and diversity-preserved instruction data\nselection method: Clustering and Ranking (CaR). CaR employs a two-step process:\nfirst, it ranks instruction pairs using a high-accuracy (84.25%) scoring model\naligned with expert preferences; second, it preserves dataset diversity through\nclustering. In our experiment, CaR efficiently selected a mere 1.96% of\nAlpaca's IT data, yet the resulting AlpaCaR model surpassed Alpaca's\nperformance by an average of 32.1% in GPT-4 evaluations. Moreover, we find that\ndata selecting is a consistent paradigm whether the pre-trained model is more\ncapable or the model parameters scaling up. Our approach employs compact models\nwith 550M parameters and incurs just 11.2% of the financial outlay of current\nmethods, enhancing its industrial deployability.\n", "date": "2024-02-28", "categories": []}
{"id": "2411.09126", "title": "SCAN: Bootstrapping Contrastive Pre-training for Data Efficiency", "abstract": "  While contrastive pre-training is widely employed, its data efficiency\nproblem has remained relatively under-explored thus far. Existing methods often\nrely on static coreset selection algorithms to pre-identify important data for\ntraining. However, this static nature renders them unable to dynamically track\nthe data usefulness throughout pre-training, leading to subpar pre-trained\nmodels. To address this challenge, our paper introduces a novel dynamic\nbootstrapping dataset pruning method. It involves pruning data preparation\nfollowed by dataset mutation operations, both of which undergo iterative and\ndynamic updates. We apply this method to two prevalent contrastive pre-training\nframeworks: \\textbf{CLIP} and \\textbf{MoCo}, representing vision-language and\nvision-centric domains, respectively. In particular, we individually pre-train\nseven CLIP models on two large-scale image-text pair datasets, and two MoCo\nmodels on the ImageNet dataset, resulting in a total of 16 pre-trained models.\nWith a data pruning rate of 30-35\\% across all 16 models, our method exhibits\nonly marginal performance degradation (less than \\textbf{1\\%} on average)\ncompared to corresponding models trained on the full dataset counterparts\nacross various downstream datasets, and also surpasses several baselines with a\nlarge performance margin. Additionally, the byproduct from our method, \\ie\ncoresets derived from the original datasets after pre-training, also\ndemonstrates significant superiority in terms of downstream performance over\nother static coreset selection approaches.\n", "date": "2024-11-14", "categories": []}
{"id": "2411.09265", "title": "BEARD: Benchmarking the Adversarial Robustness for Dataset Distillation", "abstract": "  Dataset Distillation (DD) is an emerging technique that compresses\nlarge-scale datasets into significantly smaller synthesized datasets while\npreserving high test performance and enabling the efficient training of large\nmodels. However, current research primarily focuses on enhancing evaluation\naccuracy under limited compression ratios, often overlooking critical security\nconcerns such as adversarial robustness. A key challenge in evaluating this\nrobustness lies in the complex interactions between distillation methods, model\narchitectures, and adversarial attack strategies, which complicate standardized\nassessments. To address this, we introduce BEARD, an open and unified benchmark\ndesigned to systematically assess the adversarial robustness of DD methods,\nincluding DM, IDM, and BACON. BEARD encompasses a variety of adversarial\nattacks (e.g., FGSM, PGD, C&W) on distilled datasets like CIFAR-10/100 and\nTinyImageNet. Utilizing an adversarial game framework, it introduces three key\nmetrics: Robustness Ratio (RR), Attack Efficiency Ratio (AE), and Comprehensive\nRobustness-Efficiency Index (CREI). Our analysis includes unified benchmarks,\nvarious Images Per Class (IPC) settings, and the effects of adversarial\ntraining. Results are available on the BEARD Leaderboard, along with a library\nproviding model and dataset pools to support reproducible research. Access the\ncode at BEARD.\n", "date": "2024-11-14", "categories": []}
{"id": "2409.18055", "title": "Visual Data Diagnosis and Debiasing with Concept Graphs", "abstract": "  The widespread success of deep learning models today is owed to the curation\nof extensive datasets significant in size and complexity. However, such models\nfrequently pick up inherent biases in the data during the training process,\nleading to unreliable predictions. Diagnosing and debiasing datasets is thus a\nnecessity to ensure reliable model performance. In this paper, we present\nConBias, a novel framework for diagnosing and mitigating Concept co-occurrence\nBiases in visual datasets. ConBias represents visual datasets as knowledge\ngraphs of concepts, enabling meticulous analysis of spurious concept\nco-occurrences to uncover concept imbalances across the whole dataset.\nMoreover, we show that by employing a novel clique-based concept balancing\nstrategy, we can mitigate these imbalances, leading to enhanced performance on\ndownstream tasks. Extensive experiments show that data augmentation based on a\nbalanced concept distribution augmented by Conbias improves generalization\nperformance across multiple datasets compared to state-of-the-art methods.\n", "date": "2024-09-26", "categories": []}
{"id": "2312.02548", "title": "GeNIe: Generative Hard Negative Images Through Diffusion", "abstract": "  Data augmentation is crucial in training deep models, preventing them from\noverfitting to limited data. Recent advances in generative AI, e.g., diffusion\nmodels, have enabled more sophisticated augmentation techniques that produce\ndata resembling natural images. We introduce GeNIe a novel augmentation method\nwhich leverages a latent diffusion model conditioned on a text prompt to\ncombine two contrasting data points (an image from the source category and a\ntext prompt from the target category) to generate challenging augmentations. To\nachieve this, we adjust the noise level (equivalently, number of diffusion\niterations) to ensure the generated image retains low-level and background\nfeatures from the source image while representing the target category,\nresulting in a hard negative sample for the source category. We further\nautomate and enhance GeNIe by adaptively adjusting the noise level selection on\na per image basis (coined as GeNIe-Ada), leading to further performance\nimprovements. Our extensive experiments, in both few-shot and long-tail\ndistribution settings, demonstrate the effectiveness of our novel augmentation\nmethod and its superior performance over the prior art. Our code is available\nat: https://github.com/UCDvision/GeNIe\n", "date": "2023-12-05", "categories": []}
{"id": "2411.03768", "title": "A Bayesian Approach to Data Point Selection", "abstract": "  Data point selection (DPS) is becoming a critical topic in deep learning due\nto the ease of acquiring uncurated training data compared to the difficulty of\nobtaining curated or processed data. Existing approaches to DPS are\npredominantly based on a bi-level optimisation (BLO) formulation, which is\ndemanding in terms of memory and computation, and exhibits some theoretical\ndefects regarding minibatches. Thus, we propose a novel Bayesian approach to\nDPS. We view the DPS problem as posterior inference in a novel Bayesian model\nwhere the posterior distributions of the instance-wise weights and the main\nneural network parameters are inferred under a reasonable prior and likelihood\nmodel. We employ stochastic gradient Langevin MCMC sampling to learn the main\nnetwork and instance-wise weights jointly, ensuring convergence even with\nminibatches. Our update equation is comparable to the widely used SGD and much\nmore efficient than existing BLO-based methods. Through controlled experiments\nin both the vision and language domains, we present the proof-of-concept.\nAdditionally, we demonstrate that our method scales effectively to large\nlanguage models and facilitates automated per-task optimization for instruction\nfine-tuning datasets.\n", "date": "2024-11-06", "categories": []}
{"id": "2411.02592", "title": "Decoupled Data Augmentation for Improving Image Classification", "abstract": "  Recent advancements in image mixing and generative data augmentation have\nshown promise in enhancing image classification. However, these techniques face\nthe challenge of balancing semantic fidelity with diversity. Specifically,\nimage mixing involves interpolating two images to create a new one, but this\npixel-level interpolation can compromise fidelity. Generative augmentation uses\ntext-to-image generative models to synthesize or modify images, often limiting\ndiversity to avoid generating out-of-distribution data that potentially affects\naccuracy. We propose that this fidelity-diversity dilemma partially stems from\nthe whole-image paradigm of existing methods. Since an image comprises the\nclass-dependent part (CDP) and the class-independent part (CIP), where each\npart has fundamentally different impacts on the image's fidelity, treating\ndifferent parts uniformly can therefore be misleading. To address this\nfidelity-diversity dilemma, we introduce Decoupled Data Augmentation (De-DA),\nwhich resolves the dilemma by separating images into CDPs and CIPs and handling\nthem adaptively. To maintain fidelity, we use generative models to modify real\nCDPs under controlled conditions, preserving semantic consistency. To enhance\ndiversity, we replace the image's CIP with inter-class variants, creating\ndiverse CDP-CIP combinations. Additionally, we implement an online randomized\ncombination strategy during training to generate numerous distinct CDP-CIP\ncombinations cost-effectively. Comprehensive empirical evaluations validate the\neffectiveness of our method.\n", "date": "2024-10-29", "categories": []}
{"id": "2410.15919", "title": "Are Large-scale Soft Labels Necessary for Large-scale Dataset\n  Distillation?", "abstract": "  In ImageNet-condensation, the storage for auxiliary soft labels exceeds that\nof the condensed dataset by over 30 times. However, are large-scale soft labels\nnecessary for large-scale dataset distillation? In this paper, we first\ndiscover that the high within-class similarity in condensed datasets\nnecessitates the use of large-scale soft labels. This high within-class\nsimilarity can be attributed to the fact that previous methods use samples from\ndifferent classes to construct a single batch for batch normalization (BN)\nmatching. To reduce the within-class similarity, we introduce class-wise\nsupervision during the image synthesizing process by batching the samples\nwithin classes, instead of across classes. As a result, we can increase\nwithin-class diversity and reduce the size of required soft labels. A key\nbenefit of improved image diversity is that soft label compression can be\nachieved through simple random pruning, eliminating the need for complex\nrule-based strategies. Experiments validate our discoveries. For example, when\ncondensing ImageNet-1K to 200 images per class, our approach compresses the\nrequired soft labels from 113 GB to 2.8 GB (40x compression) with a 2.6%\nperformance gain. Code is available at:\nhttps://github.com/he-y/soft-label-pruning-for-dataset-distillation\n", "date": "2024-10-21", "categories": []}
{"id": "2410.18634", "title": "Little Giants: Synthesizing High-Quality Embedding Data at Scale", "abstract": "  Synthetic data generation has become an increasingly popular way of training\nmodels without the need for large, manually labeled datasets. For tasks like\ntext embedding, synthetic data offers diverse and scalable training examples,\nsignificantly reducing the cost of human annotation. However, most current\napproaches rely heavily on proprietary models like GPT-4, which are expensive\nand inefficient for generating large-scale embedding data. In this paper, we\nintroduce SPEED, a framework that aligns open-source small models (8B) to\nefficiently generate large-scale synthetic embedding data. Through supervised\nfine-tuning, preference optimization, and self-improvement, SPEED enables small\nopen-source models to produce high-quality data. Remarkably, SPEED uses only\nless than 1/10 of the GPT API calls, outperforming the state-of-the-art\nembedding model E5_mistral when both are trained solely on their synthetic\ndata. Using this efficient generator, we conduct a comprehensive study on how\nvarious factors within the alignment pipeline impact data quality and reveal\nthe scaling law for synthetic embedding data.\n", "date": "2024-10-24", "categories": []}
{"id": "2407.17453", "title": "VILA$^2$: VILA Augmented VILA", "abstract": "  While visual language model architectures and training infrastructures\nadvance rapidly, data curation remains under-explored where quantity and\nquality become a bottleneck. Existing work either crawls extra Internet data\nwith a loose guarantee of quality or distills from black-box proprietary\nmodels, e.g., GPT-4V / Gemini that are API frequency and performance bounded.\nThis work enables a VLM to improve itself via data enhancement, exploiting its\ngenerative nature. We introduce a simple yet effective VLM augmentation scheme\nthat includes a self-augment step and a specialist-augment step to iteratively\nimprove data quality and hence, model performance. In the self-augment step,\nthe instruction-finetuned VLM recaptions its pretraining caption datasets and\nthen retrains from scratch leveraging refined data. Without any expensive\nhuman-in-the-loop annotation, we observe improvements in data quality and\ndownstream accuracy boosts with three self-augmentation rounds -- a viable free\nlunch to the current VLM training recipe. When self-augmentation saturates, we\naugment the caption diversity by leveraging specialty skills picked up from\ninstruction finetuning. We finetune VLM specialists from the self-augmented VLM\nwith domain-specific experts, including spatial, grounding, and OCR, to fuse\ntask-aware synthetic data into the pretraining stage. Data quality improvements\nand hallucination reductions are cross-checked by VLM (GPT-4V, Gemini) and\nhuman judges. Combining self-augmentation and specialist-augmented training,\nVILA$^2$ consistently improves the accuracy on a wide range of benchmarks over\nthe prior art, producing a reusable pretraining dataset that is 300x more\ncost-efficient than human labeling.\n", "date": "2024-07-24", "categories": []}
{"id": "2411.00611", "title": "Small coresets via negative dependence: DPPs, linear statistics, and\n  concentration", "abstract": "  Determinantal point processes (DPPs) are random configurations of points with\ntunable negative dependence. Because sampling is tractable, DPPs are natural\ncandidates for subsampling tasks, such as minibatch selection or coreset\nconstruction. A \\emph{coreset} is a subset of a (large) training set, such that\nminimizing an empirical loss averaged over the coreset is a controlled\nreplacement for the intractable minimization of the original empirical loss.\nTypically, the control takes the form of a guarantee that the average loss over\nthe coreset approximates the total loss uniformly across the parameter space.\nRecent work has provided significant empirical support in favor of using DPPs\nto build randomized coresets, coupled with interesting theoretical results that\nare suggestive but leave some key questions unanswered. In particular, the\ncentral question of whether the cardinality of a DPP-based coreset is\nfundamentally smaller than one based on independent sampling remained open. In\nthis paper, we answer this question in the affirmative, demonstrating that\n\\emph{DPPs can provably outperform independently drawn coresets}. In this vein,\nwe contribute a conceptual understanding of coreset loss as a \\emph{linear\nstatistic} of the (random) coreset. We leverage this structural observation to\nconnect the coresets problem to a more general problem of concentration\nphenomena for linear statistics of DPPs, wherein we obtain \\emph{effective\nconcentration inequalities that extend well-beyond the state-of-the-art},\nencompassing general non-projection, even non-symmetric kernels. The latter\nhave been recently shown to be of interest in machine learning beyond coresets,\nbut come with a limited theoretical toolbox, to the extension of which our\nresult contributes. Finally, we are also able to address the coresets problem\nfor vector-valued objective functions, a novelty in the coresets literature.\n", "date": "2024-11-01", "categories": []}
{"id": "2411.00745", "title": "Private, Augmentation-Robust and Task-Agnostic Data Valuation Approach\n  for Data Marketplace", "abstract": "  Evaluating datasets in data marketplaces, where the buyer aim to purchase\nvaluable data, is a critical challenge. In this paper, we introduce an\ninnovative task-agnostic data valuation method called PriArTa which is an\napproach for computing the distance between the distribution of the buyer's\nexisting dataset and the seller's dataset, allowing the buyer to determine how\neffectively the new data can enhance its dataset. PriArTa is\ncommunication-efficient, enabling the buyer to evaluate datasets without\nneeding access to the entire dataset from each seller. Instead, the buyer\nrequests that sellers perform specific preprocessing on their data and then\nsend back the results. Using this information and a scoring metric, the buyer\ncan evaluate the dataset. The preprocessing is designed to allow the buyer to\ncompute the score while preserving the privacy of each seller's dataset,\nmitigating the risk of information leakage before the purchase. A key feature\nof PriArTa is its robustness to common data transformations, ensuring\nconsistent value assessment and reducing the risk of purchasing redundant data.\nThe effectiveness of PriArTa is demonstrated through experiments on real-world\nimage datasets, showing its ability to perform privacy-preserving,\naugmentation-robust data valuation in data marketplaces.\n", "date": "2024-11-01", "categories": []}
{"id": "2406.11271", "title": "MINT-1T: Scaling Open-Source Multimodal Data by 10x: A Multimodal\n  Dataset with One Trillion Tokens", "abstract": "  Multimodal interleaved datasets featuring free-form interleaved sequences of\nimages and text are crucial for training frontier large multimodal models\n(LMMs). Despite the rapid progression of open-source LMMs, there remains a\npronounced scarcity of large-scale, diverse open-source multimodal interleaved\ndatasets. In response, we introduce MINT-1T, the most extensive and diverse\nopen-source Multimodal INTerleaved dataset to date. MINT-1T comprises one\ntrillion text tokens and 3.4 billion images, a 10x scale-up from existing\nopen-source datasets. Additionally, we include previously untapped sources such\nas PDFs and ArXiv papers. As scaling multimodal interleaved datasets requires\nsubstantial engineering effort, sharing the data curation process and releasing\nthe dataset greatly benefits the community. Our experiments show that LMMs\ntrained on MINT-1T rival the performance of models trained on the previous\nleading dataset, OBELICS. Our data and code will be released at\nhttps://github.com/mlfoundations/MINT-1T.\n", "date": "2024-06-17", "categories": []}
{"id": "2406.17557", "title": "The FineWeb Datasets: Decanting the Web for the Finest Text Data at\n  Scale", "abstract": "  The performance of a large language model (LLM) depends heavily on the\nquality and size of its pretraining dataset. However, the pretraining datasets\nfor state-of-the-art open LLMs like Llama 3 and Mixtral are not publicly\navailable and very little is known about how they were created. In this work,\nwe introduce FineWeb, a 15-trillion token dataset derived from 96 Common Crawl\nsnapshots that produces better-performing LLMs than other open pretraining\ndatasets. To advance the understanding of how best to curate high-quality\npretraining datasets, we carefully document and ablate all of the design\nchoices used in FineWeb, including in-depth investigations of deduplication and\nfiltering strategies. In addition, we introduce FineWeb-Edu, a 1.3-trillion\ntoken collection of educational text filtered from FineWeb. LLMs pretrained on\nFineWeb-Edu exhibit dramatically better performance on knowledge- and\nreasoning-intensive benchmarks like MMLU and ARC. Along with our datasets, we\npublicly release our data curation codebase and all of the models trained\nduring our ablation experiments.\n", "date": "2024-06-25", "categories": []}
{"id": "2405.00705", "title": "SHED: Shapley-Based Automated Dataset Refinement for Instruction\n  Fine-Tuning", "abstract": "  The pre-trained Large Language Models (LLMs) can be adapted for many\ndownstream tasks and tailored to align with human preferences through\nfine-tuning. Recent studies have discovered that LLMs can achieve desirable\nperformance with only a small amount of high-quality data, suggesting that a\nlarge amount of the data in these extensive datasets is redundant or even\nharmful. Identifying high-quality data from vast datasets to curate small yet\neffective datasets has emerged as a critical challenge. In this paper, we\nintroduce SHED, an automated dataset refinement framework based on Shapley\nvalue for instruction fine-tuning. SHED eliminates the need for human\nintervention or the use of commercial LLMs. Moreover, the datasets curated\nthrough SHED exhibit transferability, indicating they can be reused across\ndifferent LLMs with consistently high performance. We conduct extensive\nexperiments to evaluate the datasets curated by SHED. The results demonstrate\nSHED's superiority over state-of-the-art methods across various tasks and LLMs;\nnotably, datasets comprising only 10% of the original data selected by SHED\nachieve performance comparable to or surpassing that of the full datasets.\n", "date": "2024-04-23", "categories": []}
{"id": "2408.03572", "title": "2D-OOB: Attributing Data Contribution Through Joint Valuation Framework", "abstract": "  Data valuation has emerged as a powerful framework for quantifying each\ndatum's contribution to the training of a machine learning model. However, it\nis crucial to recognize that the quality of cells within a single data point\ncan vary greatly in practice. For example, even in the case of an abnormal data\npoint, not all cells are necessarily noisy. The single scalar score assigned by\nexisting data valuation methods blurs the distinction between noisy and clean\ncells of a data point, making it challenging to interpret the data values. In\nthis paper, we propose 2D-OOB, an out-of-bag estimation framework for jointly\ndetermining helpful (or detrimental) samples as well as the particular cells\nthat drive them. Our comprehensive experiments demonstrate that 2D-OOB achieves\nstate-of-the-art performance across multiple use cases while being\nexponentially faster. Specifically, 2D-OOB shows promising results in detecting\nand rectifying fine-grained outliers at the cell level, and localizing backdoor\ntriggers in data poisoning attacks.\n", "date": "2024-08-07", "categories": []}
{"id": "2305.17048", "title": "Intrinsic Self-Supervision for Data Quality Audits", "abstract": "  Benchmark datasets in computer vision often contain off-topic images, near\nduplicates, and label errors, leading to inaccurate estimates of model\nperformance. In this paper, we revisit the task of data cleaning and formalize\nit as either a ranking problem, which significantly reduces human inspection\neffort, or a scoring problem, which allows for automated decisions based on\nscore distributions. We find that a specific combination of context-aware\nself-supervised representation learning and distance-based indicators is\neffective in finding issues without annotation biases. This methodology, which\nwe call SelfClean, surpasses state-of-the-art performance in detecting\noff-topic images, near duplicates, and label errors within widely-used image\ndatasets, such as ImageNet-1k, Food-101N, and STL-10, both for synthetic issues\nand real contamination. We apply the detailed method to multiple image\nbenchmarks, identify up to 16% of issues, and confirm an improvement in\nevaluation reliability upon cleaning. The official implementation can be found\nat: https://github.com/Digital-Dermatology/SelfClean.\n", "date": "2023-05-26", "categories": []}
{"id": "2402.13454", "title": "Theoretical Analysis of Submodular Information Measures for Targeted\n  Data Subset Selection", "abstract": "  With increasing volume of data being used across machine learning tasks, the\ncapability to target specific subsets of data becomes more important. To aid in\nthis capability, the recently proposed Submodular Mutual Information (SMI) has\nbeen effectively applied across numerous tasks in literature to perform\ntargeted subset selection with the aid of a exemplar query set. However, all\nsuch works are deficient in providing theoretical guarantees for SMI in terms\nof its sensitivity to a subset's relevance and coverage of the targeted data.\nFor the first time, we provide such guarantees by deriving similarity-based\nbounds on quantities related to relevance and coverage of the targeted data.\nWith these bounds, we show that the SMI functions, which have empirically shown\nsuccess in multiple applications, are theoretically sound in achieving good\nquery relevance and query coverage.\n", "date": "2024-02-21", "categories": []}
{"id": "2301.09902", "title": "Investigating Labeler Bias in Face Annotation for Machine Learning", "abstract": "  In a world increasingly reliant on artificial intelligence, it is more\nimportant than ever to consider the ethical implications of artificial\nintelligence on humanity. One key under-explored challenge is labeler bias,\nwhich can create inherently biased datasets for training and subsequently lead\nto inaccurate or unfair decisions in healthcare, employment, education, and law\nenforcement. Hence, we conducted a study to investigate and measure the\nexistence of labeler bias using images of people from different ethnicities and\nsexes in a labeling task. Our results show that participants possess\nstereotypes that influence their decision-making process and that labeler\ndemographics impact assigned labels. We also discuss how labeler bias\ninfluences datasets and, subsequently, the models trained on them. Overall, a\nhigh degree of transparency must be maintained throughout the entire artificial\nintelligence training process to identify and correct biases in the data as\nearly as possible.\n", "date": "2023-01-24", "categories": []}
{"id": "2410.18678", "title": "Ali-AUG: Innovative Approaches to Labeled Data Augmentation using\n  One-Step Diffusion Model", "abstract": "  This paper introduces Ali-AUG, a novel single-step diffusion model for\nefficient labeled data augmentation in industrial applications. Our method\naddresses the challenge of limited labeled data by generating synthetic,\nlabeled images with precise feature insertion. Ali-AUG utilizes a stable\ndiffusion architecture enhanced with skip connections and LoRA modules to\nefficiently integrate masks and images, ensuring accurate feature placement\nwithout affecting unrelated image content. Experimental validation across\nvarious industrial datasets demonstrates Ali-AUG's superiority in generating\nhigh-quality, defect-enhanced images while maintaining rapid single-step\ninference. By offering precise control over feature insertion and minimizing\nrequired training steps, our technique significantly enhances data augmentation\ncapabilities, providing a powerful tool for improving the performance of deep\nlearning models in scenarios with limited labeled data. Ali-AUG is especially\nuseful for use cases like defective product image generation to train AI-based\nmodels to improve their ability to detect defects in manufacturing processes.\nUsing different data preparation strategies, including Classification Accuracy\nScore (CAS) and Naive Augmentation Score (NAS), we show that Ali-AUG improves\nmodel performance by 31% compared to other augmentation methods and by 45%\ncompared to models without data augmentation. Notably, Ali-AUG reduces training\ntime by 32% and supports both paired and unpaired datasets, enhancing\nflexibility in data preparation.\n", "date": "2024-10-24", "categories": []}
{"id": "2406.02131", "title": "CondTSF: One-line Plugin of Dataset Condensation for Time Series\n  Forecasting", "abstract": "  Dataset condensation is a newborn technique that generates a small dataset\nthat can be used in training deep neural networks to lower training costs. The\nobjective of dataset condensation is to ensure that the model trained with the\nsynthetic dataset can perform comparably to the model trained with full\ndatasets. However, existing methods predominantly concentrate on classification\ntasks, posing challenges in their adaptation to time series forecasting\n(TS-forecasting). This challenge arises from disparities in the evaluation of\nsynthetic data. In classification, the synthetic data is considered\nwell-distilled if the model trained with the full dataset and the model trained\nwith the synthetic dataset yield identical labels for the same input,\nregardless of variations in output logits distribution. Conversely, in\nTS-forecasting, the effectiveness of synthetic data distillation is determined\nby the distance between predictions of the two models. The synthetic data is\ndeemed well-distilled only when all data points within the predictions are\nsimilar. Consequently, TS-forecasting has a more rigorous evaluation\nmethodology compared to classification. To mitigate this gap, we theoretically\nanalyze the optimization objective of dataset condensation for TS-forecasting\nand propose a new one-line plugin of dataset condensation designated as Dataset\nCondensation for Time Series Forecasting (CondTSF) based on our analysis.\nPlugging CondTSF into previous dataset condensation methods facilitates a\nreduction in the distance between the predictions of the model trained with the\nfull dataset and the model trained with the synthetic dataset, thereby\nenhancing performance. We conduct extensive experiments on eight commonly used\ntime series datasets. CondTSF consistently improves the performance of all\nprevious dataset condensation methods across all datasets, particularly at low\ncondensing ratios.\n", "date": "2024-06-04", "categories": []}
{"id": "2410.16710", "title": "Influential Language Data Selection via Gradient Trajectory Pursuit", "abstract": "  Curating a desirable dataset for training has been the core of building\nhighly capable large language models (Touvron et al., 2023; Achiam et al.,\n2023; Team et al.,2024). Gradient influence scores (Pruthi et al., 2020; Xia et\nal., 2024) are shown to be correlated with model performance and are commonly\nused as the criterion for data selection. However, existing methods are built\nupon either individual sample rankings or inefficient matching process, leading\nto suboptimal performance or scaling up issues.In this paper, we propose\nGradient Trajectory Pursuit (GTP), an algorithm that performs pursuit of\ngradient trajectories via jointly selecting data points under an L0-norm\nregularized objective. The proposed algorithm highlights: (1) joint selection\ninstead of independent top-k selection, which automatically de-duplicates\nsamples; (2) higher efficiency with compressive sampling processes, which can\nbe further sped up using a distributed framework. In the experiments, we\ndemonstrate the algorithm in both in-domain and target-domain selection\nbenchmarks and show that it outperforms top-k selection and competitive\nalgorithms consistently, for example, our algorithm chooses as low as 0.5% data\nto achieve full performance on the targeted instruction tuning tasks\n", "date": "2024-10-22", "categories": []}
{"id": "2410.09940", "title": "Generalized Group Data Attribution", "abstract": "  Data Attribution (DA) methods quantify the influence of individual training\ndata points on model outputs and have broad applications such as\nexplainability, data selection, and noisy label identification. However,\nexisting DA methods are often computationally intensive, limiting their\napplicability to large-scale machine learning models. To address this\nchallenge, we introduce the Generalized Group Data Attribution (GGDA)\nframework, which computationally simplifies DA by attributing to groups of\ntraining points instead of individual ones. GGDA is a general framework that\nsubsumes existing attribution methods and can be applied to new DA techniques\nas they emerge. It allows users to optimize the trade-off between efficiency\nand fidelity based on their needs. Our empirical results demonstrate that GGDA\napplied to popular DA methods such as Influence Functions, TracIn, and TRAK\nresults in upto 10x-50x speedups over standard DA methods while gracefully\ntrading off attribution fidelity. For downstream applications such as dataset\npruning and noisy label identification, we demonstrate that GGDA significantly\nimproves computational efficiency and maintains effectiveness, enabling\npractical applications in large-scale machine learning scenarios that were\npreviously infeasible.\n", "date": "2024-10-13", "categories": []}
{"id": "2410.14254", "title": "RAZOR: Refining Accuracy by Zeroing Out Redundancies", "abstract": "  In many application domains, the proliferation of sensors and devices is\ngenerating vast volumes of data, imposing significant pressure on existing data\nanalysis and data mining techniques. Nevertheless, an increase in data volume\ndoes not inherently imply an increase in informational content, as a\nsubstantial portion may be redundant or represent noise. This challenge is\nparticularly evident in the deep learning domain, where the utility of\nadditional data is contingent on its informativeness. In the absence of such,\nlarger datasets merely exacerbate the computational cost and complexity of the\nlearning process. To address these challenges, we propose RAZOR, a novel\ninstance selection technique designed to extract a significantly smaller yet\nsufficiently informative subset from a larger set of instances without\ncompromising the learning process. RAZOR has been specifically engineered to be\nrobust, efficient, and scalable, making it suitable for large-scale datasets.\nUnlike many techniques in the literature, RAZOR is capable of operating in both\nsupervised and unsupervised settings. Experimental results demonstrate that\nRAZOR outperforms recent state-of-the-art techniques in terms of both\neffectiveness and efficiency.\n", "date": "2024-10-18", "categories": []}
{"id": "2410.13311", "title": "Enhancing Dataset Distillation via Label Inconsistency Elimination and\n  Learning Pattern Refinement", "abstract": "  Dataset Distillation (DD) seeks to create a condensed dataset that, when used\nto train a model, enables the model to achieve performance similar to that of a\nmodel trained on the entire original dataset. It relieves the model training\nfrom processing massive data and thus reduces the computation resources,\nstorage, and time costs. This paper illustrates our solution that ranks 1st in\nthe ECCV-2024 Data Distillation Challenge (track 1). Our solution, Modified\nDifficulty-Aligned Trajectory Matching (M-DATM), introduces two key\nmodifications to the original state-of-the-art method DATM: (1) the soft labels\nlearned by DATM do not achieve one-to-one correspondence with the counterparts\ngenerated by the official evaluation script, so we remove the soft labels\ntechnique to alleviate such inconsistency; (2) since the removal of soft labels\nmakes it harder for the synthetic dataset to learn late trajectory information,\nparticularly on Tiny ImageNet, we reduce the matching range, allowing the\nsynthetic data to concentrate more on the easier patterns. In the final\nevaluation, our M-DATM achieved accuracies of 0.4061 and 0.1831 on the\nCIFAR-100 and Tiny ImageNet datasets, ranking 1st in the Fixed Images Per Class\n(IPC) Track.\n", "date": "2024-10-17", "categories": []}
{"id": "2410.13464", "title": "IterSelectTune: An Iterative Training Framework for Efficient\n  Instruction-Tuning Data Selection", "abstract": "  As large language models (LLMs) continue to advance, instruction tuning has\nbecome critical for improving their ability to generate accurate and\ncontextually appropriate responses. Although numerous instruction-tuning\ndatasets have been developed to enhance LLM performance, selecting high-quality\ninstruction data from large source datasets typically demands significant human\neffort. In this work, we introduce $\\textbf{IterSelectTune}$, an efficient,\ncost-effective iterative training policy for selecting high-quality instruction\ndata with no human involvement and limited reliance on GPT-4. By fine-tuning on\napproximately 20\\% of the source data, our method consistently outperforms\nmodels fine-tuned on the full dataset across multiple benchmarks and public\ntest datasets. These results highlight the effectiveness of our approach in\nenhancing LLM performance while reducing the computational resources required\nfor instruction tuning.\n", "date": "2024-10-17", "categories": []}
{"id": "2410.13761", "title": "GDeR: Safeguarding Efficiency, Balancing, and Robustness via\n  Prototypical Graph Pruning", "abstract": "  Training high-quality deep models necessitates vast amounts of data,\nresulting in overwhelming computational and memory demands. Recently, data\npruning, distillation, and coreset selection have been developed to streamline\ndata volume by retaining, synthesizing, or selecting a small yet informative\nsubset from the full set. Among these methods, data pruning incurs the least\nadditional training cost and offers the most practical acceleration benefits.\nHowever, it is the most vulnerable, often suffering significant performance\ndegradation with imbalanced or biased data schema, thus raising concerns about\nits accuracy and reliability in on-device deployment. Therefore, there is a\nlooming need for a new data pruning paradigm that maintains the efficiency of\nprevious practices while ensuring balance and robustness. Unlike the fields of\ncomputer vision and natural language processing, where mature solutions have\nbeen developed to address these issues, graph neural networks (GNNs) continue\nto struggle with increasingly large-scale, imbalanced, and noisy datasets,\nlacking a unified dataset pruning solution. To achieve this, we introduce a\nnovel dynamic soft-pruning method, GDeR, designed to update the training\n``basket'' during the process using trainable prototypes. GDeR first constructs\na well-modeled graph embedding hypersphere and then samples\n\\textit{representative, balanced, and unbiased subsets} from this embedding\nspace, which achieves the goal we called Graph Training Debugging. Extensive\nexperiments on five datasets across three GNN backbones, demonstrate that GDeR\n(I) achieves or surpasses the performance of the full dataset with 30%~50%\nfewer training samples, (II) attains up to a 2.81x lossless training speedup,\nand (III) outperforms state-of-the-art pruning methods in imbalanced training\nand noisy training scenarios by 0.3%~4.3% and 3.6%~7.8%, respectively.\n", "date": "2024-10-17", "categories": []}
{"id": "2409.00284", "title": "Reframing Data Value for Large Language Models Through the Lens of\n  Plausibility", "abstract": "  Data valuation seeks to answer the important question, \"How much is this data\nworth?\" Existing data valuation methods have largely focused on discriminative\nmodels, primarily examining data value through the lens of its utility in\ntraining. However, with the push for ever-larger language models, relying on\nvaluation methods that require training becomes increasingly expensive and\ndependent on specific techniques. We propose an alternative perspective on the\ndata value problem for language models, centering around the plausibility of\nthe data. We posit that data holds lesser value if it can be plausibly\ngenerated by the model itself. Starting from some intuitive criteria that align\nwith our notions of valuable data, we develop a novel value function that is\ncomputationally tractable and derived from first principles with provable\nproperties. We conduct a theoretical analysis of our value function and\nevaluate it across multiple scenarios and datasets.\n", "date": "2024-08-30", "categories": []}
{"id": "2410.11820", "title": "Adaptive Data Optimization: Dynamic Sample Selection with Scaling Laws", "abstract": "  The composition of pretraining data is a key determinant of foundation\nmodels' performance, but there is no standard guideline for allocating a\nlimited computational budget across different data sources. Most current\napproaches either rely on extensive experiments with smaller models or dynamic\ndata adjustments that also require proxy models, both of which significantly\nincrease the workflow complexity and computational overhead. In this paper, we\nintroduce Adaptive Data Optimization (ADO), an algorithm that optimizes data\ndistributions in an online fashion, concurrent with model training. Unlike\nexisting techniques, ADO does not require external knowledge, proxy models, or\nmodifications to the model update. Instead, ADO uses per-domain scaling laws to\nestimate the learning potential of each domain during training and adjusts the\ndata mixture accordingly, making it more scalable and easier to integrate.\nExperiments demonstrate that ADO can achieve comparable or better performance\nthan prior methods while maintaining computational efficiency across different\ncomputation scales, offering a practical solution for dynamically adjusting\ndata distribution without sacrificing flexibility or increasing costs. Beyond\nits practical benefits, ADO also provides a new perspective on data collection\nstrategies via scaling laws.\n", "date": "2024-10-15", "categories": []}
{"id": "2403.13322", "title": "DD-RobustBench: An Adversarial Robustness Benchmark for Dataset\n  Distillation", "abstract": "  Dataset distillation is an advanced technique aimed at compressing datasets\ninto significantly smaller counterparts, while preserving formidable training\nperformance. Significant efforts have been devoted to promote evaluation\naccuracy under limited compression ratio while overlooked the robustness of\ndistilled dataset. In this work, we introduce a comprehensive benchmark that,\nto the best of our knowledge, is the most extensive to date for evaluating the\nadversarial robustness of distilled datasets in a unified way. Our benchmark\nsignificantly expands upon prior efforts by incorporating a wider range of\ndataset distillation methods, including the latest advancements such as TESLA\nand SRe2L, a diverse array of adversarial attack methods, and evaluations\nacross a broader and more extensive collection of datasets such as ImageNet-1K.\nMoreover, we assessed the robustness of these distilled datasets against\nrepresentative adversarial attack algorithms like PGD and AutoAttack, while\nexploring their resilience from a frequency perspective. We also discovered\nthat incorporating distilled data into the training batches of the original\ndataset can yield to improvement of robustness.\n", "date": "2024-03-20", "categories": []}
{"id": "2408.03360", "title": "Prioritize Alignment in Dataset Distillation", "abstract": "  Dataset Distillation aims to compress a large dataset into a significantly\nmore compact, synthetic one without compromising the performance of the trained\nmodels. To achieve this, existing methods use the agent model to extract\ninformation from the target dataset and embed it into the distilled dataset.\nConsequently, the quality of extracted and embedded information determines the\nquality of the distilled dataset. In this work, we find that existing methods\nintroduce misaligned information in both information extraction and embedding\nstages. To alleviate this, we propose Prioritize Alignment in Dataset\nDistillation (PAD), which aligns information from the following two\nperspectives. 1) We prune the target dataset according to the compressing ratio\nto filter the information that can be extracted by the agent model. 2) We use\nonly deep layers of the agent model to perform the distillation to avoid\nexcessively introducing low-level information. This simple strategy effectively\nfilters out misaligned information and brings non-trivial improvement for\nmainstream matching-based distillation algorithms. Furthermore, built on\ntrajectory matching, \\textbf{PAD} achieves remarkable improvements on various\nbenchmarks, achieving state-of-the-art performance.\n", "date": "2024-08-06", "categories": []}
{"id": "2410.09642", "title": "RepMatch: Quantifying Cross-Instance Similarities in Representation\n  Space", "abstract": "  Advances in dataset analysis techniques have enabled more sophisticated\napproaches to analyzing and characterizing training data instances, often\ncategorizing data based on attributes such as ``difficulty''. In this work, we\nintroduce RepMatch, a novel method that characterizes data through the lens of\nsimilarity. RepMatch quantifies the similarity between subsets of training\ninstances by comparing the knowledge encoded in models trained on them,\novercoming the limitations of existing analysis methods that focus solely on\nindividual instances and are restricted to within-dataset analysis. Our\nframework allows for a broader evaluation, enabling similarity comparisons\nacross arbitrary subsets of instances, supporting both dataset-to-dataset and\ninstance-to-dataset analyses. We validate the effectiveness of RepMatch across\nmultiple NLP tasks, datasets, and models. Through extensive experimentation, we\ndemonstrate that RepMatch can effectively compare datasets, identify more\nrepresentative subsets of a dataset (that lead to better performance than\nrandomly selected subsets of equivalent size), and uncover heuristics\nunderlying the construction of some challenge datasets.\n", "date": "2024-10-12", "categories": []}
{"id": "2410.08432", "title": "MYCROFT: Towards Effective and Efficient External Data Augmentation", "abstract": "  Machine learning (ML) models often require large amounts of data to perform\nwell. When the available data is limited, model trainers may need to acquire\nmore data from external sources. Often, useful data is held by private entities\nwho are hesitant to share their data due to propriety and privacy concerns.\nThis makes it challenging and expensive for model trainers to acquire the data\nthey need to improve model performance. To address this challenge, we propose\nMycroft, a data-efficient method that enables model trainers to evaluate the\nrelative utility of different data sources while working with a constrained\ndata-sharing budget. By leveraging feature space distances and gradient\nmatching, Mycroft identifies small but informative data subsets from each\nowner, allowing model trainers to maximize performance with minimal data\nexposure. Experimental results across four tasks in two domains show that\nMycroft converges rapidly to the performance of the full-information baseline,\nwhere all data is shared. Moreover, Mycroft is robust to noise and can\neffectively rank data owners by utility. Mycroft can pave the way for\ndemocratized training of high performance ML models.\n", "date": "2024-10-11", "categories": []}
{"id": "2410.05639", "title": "DecorateLM: Data Engineering through Corpus Rating, Tagging, and Editing\n  with Language Models", "abstract": "  The performance of Large Language Models (LLMs) is substantially influenced\nby the pretraining corpus, which consists of vast quantities of unsupervised\ndata processed by the models. Despite its critical role in model performance,\nensuring the quality of this data is challenging due to its sheer volume and\nthe absence of sample-level quality annotations and enhancements. In this\npaper, we introduce DecorateLM, a data engineering method designed to refine\nthe pretraining corpus through data rating, tagging and editing. Specifically,\nDecorateLM rates texts against quality criteria, tags texts with hierarchical\nlabels, and edits texts into a more formalized format. Due to the massive size\nof the pretraining corpus, adopting an LLM for decorating the entire corpus is\nless efficient. Therefore, to balance performance with efficiency, we curate a\nmeticulously annotated training corpus for DecorateLM using a large language\nmodel and distill data engineering expertise into a compact 1.2 billion\nparameter small language model (SLM). We then apply DecorateLM to enhance 100\nbillion tokens of the training corpus, selecting 45 billion tokens that\nexemplify high quality and diversity for the further training of another 1.2\nbillion parameter LLM. Our results demonstrate that employing such high-quality\ndata can significantly boost model performance, showcasing a powerful approach\nto enhance the quality of the pretraining corpus.\n", "date": "2024-10-08", "categories": []}
{"id": "2410.07579", "title": "Teddy: Efficient Large-Scale Dataset Distillation via\n  Taylor-Approximated Matching", "abstract": "  Dataset distillation or condensation refers to compressing a large-scale\ndataset into a much smaller one, enabling models trained on this synthetic\ndataset to generalize effectively on real data. Tackling this challenge, as\ndefined, relies on a bi-level optimization algorithm: a novel model is trained\nin each iteration within a nested loop, with gradients propagated through an\nunrolled computation graph. However, this approach incurs high memory and time\ncomplexity, posing difficulties in scaling up to large datasets such as\nImageNet. Addressing these concerns, this paper introduces Teddy, a\nTaylor-approximated dataset distillation framework designed to handle\nlarge-scale dataset and enhance efficiency. On the one hand, backed up by\ntheoretical analysis, we propose a memory-efficient approximation derived from\nTaylor expansion, which transforms the original form dependent on multi-step\ngradients to a first-order one. On the other hand, rather than repeatedly\ntraining a novel model in each iteration, we unveil that employing a pre-cached\npool of weak models, which can be generated from a single base model, enhances\nboth time efficiency and performance concurrently, particularly when dealing\nwith large-scale datasets. Extensive experiments demonstrate that the proposed\nTeddy attains state-of-the-art efficiency and performance on the Tiny-ImageNet\nand original-sized ImageNet-1K dataset, notably surpassing prior methods by up\nto 12.8%, while reducing 46.6% runtime. Our code will be available at\nhttps://github.com/Lexie-YU/Teddy.\n", "date": "2024-10-10", "categories": []}
{"id": "2410.05578", "title": "Swift Sampler: Efficient Learning of Sampler by 10 Parameters", "abstract": "  Data selection is essential for training deep learning models. An effective\ndata sampler assigns proper sampling probability for training data and helps\nthe model converge to a good local minimum with high performance. Previous\nstudies in data sampling are mainly based on heuristic rules or learning\nthrough a huge amount of time-consuming trials. In this paper, we propose an\nautomatic \\textbf{swift sampler} search algorithm, \\textbf{SS}, to explore\nautomatically learning effective samplers efficiently. In particular,\n\\textbf{SS} utilizes a novel formulation to map a sampler to a low dimension of\nhyper-parameters and uses an approximated local minimum to quickly examine the\nquality of a sampler. Benefiting from its low computational expense,\n\\textbf{SS} can be applied on large-scale data sets with high efficiency.\nComprehensive experiments on various tasks demonstrate that \\textbf{SS} powered\nsampling can achieve obvious improvements (e.g., 1.5\\% on ImageNet) and\ntransfer among different neural networks. Project page:\nhttps://github.com/Alexander-Yao/Swift-Sampler.\n", "date": "2024-10-08", "categories": []}
{"id": "2410.06550", "title": "Investigating Cost-Efficiency of LLM-Generated Training Data for\n  Conversational Semantic Frame Analysis", "abstract": "  Recent studies have demonstrated that few-shot learning allows LLMs to\ngenerate training data for supervised models at a low cost. However, the\nquality of LLM-generated data may not entirely match that of human-labeled\ndata. This raises a crucial question: how should one balance the trade-off\nbetween the higher quality but more expensive human data and the lower quality\nyet substantially cheaper LLM-generated data? In this paper, we synthesized\ntraining data for conversational semantic frame analysis using GPT-4 and\nexamined how to allocate budgets optimally to achieve the best performance. Our\nexperiments, conducted across various budget levels, reveal that optimal\ncost-efficiency is achieved by combining both human and LLM-generated data\nacross a wide range of budget levels. Notably, as the budget decreases, a\nhigher proportion of LLM-generated data becomes more preferable.\n", "date": "2024-10-09", "categories": []}
{"id": "2209.04851", "title": "OpenMixup: Open Mixup Toolbox and Benchmark for Visual Representation\n  Learning", "abstract": "  Mixup augmentation has emerged as a widely used technique for improving the\ngeneralization ability of deep neural networks (DNNs). However, the lack of\nstandardized implementations and benchmarks has impeded recent progress,\nresulting in poor reproducibility, unfair comparisons, and conflicting\ninsights. In this paper, we introduce OpenMixup, the first mixup augmentation\ncodebase, and benchmark for visual representation learning. Specifically, we\ntrain 18 representative mixup baselines from scratch and rigorously evaluate\nthem across 11 image datasets of varying scales and granularity, ranging from\nfine-grained scenarios to complex non-iconic scenes. We also open-source our\nmodular codebase, including a collection of popular vision backbones,\noptimization strategies, and analysis toolkits, which not only supports the\nbenchmarking but enables broader mixup applications beyond classification, such\nas self-supervised learning and regression tasks. Through experiments and\nempirical analysis, we gain observations and insights on mixup\nperformance-efficiency trade-offs, generalization, and optimization behaviors,\nand thereby identify preferred choices for different needs. To the best of our\nknowledge, OpenMixup has facilitated several recent studies. We believe this\nwork can further advance reproducible mixup augmentation research and thereby\nlay a solid ground for future progress in the community. The source code and\nuser documents are available at \\url{https://github.com/Westlake-AI/openmixup}.\n", "date": "2022-09-11", "categories": []}
{"id": "2410.04386", "title": "Data Distribution Valuation", "abstract": "  Data valuation is a class of techniques for quantitatively assessing the\nvalue of data for applications like pricing in data marketplaces. Existing data\nvaluation methods define a value for a discrete dataset. However, in many use\ncases, users are interested in not only the value of the dataset, but that of\nthe distribution from which the dataset was sampled. For example, consider a\nbuyer trying to evaluate whether to purchase data from different vendors. The\nbuyer may observe (and compare) only a small preview sample from each vendor,\nto decide which vendor's data distribution is most useful to the buyer and\npurchase. The core question is how should we compare the values of data\ndistributions from their samples? Under a Huber characterization of the data\nheterogeneity across vendors, we propose a maximum mean discrepancy (MMD)-based\nvaluation method which enables theoretically principled and actionable policies\nfor comparing data distributions from samples. We empirically demonstrate that\nour method is sample-efficient and effective in identifying valuable data\ndistributions against several existing baselines, on multiple real-world\ndatasets (e.g., network intrusion detection, credit card fraud detection) and\ndownstream applications (classification, regression).\n", "date": "2024-10-06", "categories": []}
{"id": "2410.04555", "title": "$\\texttt{dattri}$: A Library for Efficient Data Attribution", "abstract": "  Data attribution methods aim to quantify the influence of individual training\nsamples on the prediction of artificial intelligence (AI) models. As training\ndata plays an increasingly crucial role in the modern development of\nlarge-scale AI models, data attribution has found broad applications in\nimproving AI performance and safety. However, despite a surge of new data\nattribution methods being developed recently, there lacks a comprehensive\nlibrary that facilitates the development, benchmarking, and deployment of\ndifferent data attribution methods. In this work, we introduce\n$\\texttt{dattri}$, an open-source data attribution library that addresses the\nabove needs. Specifically, $\\texttt{dattri}$ highlights three novel design\nfeatures. Firstly, $\\texttt{dattri}$ proposes a unified and easy-to-use API,\nallowing users to integrate different data attribution methods into their\nPyTorch-based machine learning pipeline with a few lines of code changed.\nSecondly, $\\texttt{dattri}$ modularizes low-level utility functions that are\ncommonly used in data attribution methods, such as Hessian-vector product,\ninverse-Hessian-vector product or random projection, making it easier for\nresearchers to develop new data attribution methods. Thirdly, $\\texttt{dattri}$\nprovides a comprehensive benchmark framework with pre-trained models and ground\ntruth annotations for a variety of benchmark settings, including generative AI\nsettings. We have implemented a variety of state-of-the-art efficient data\nattribution methods that can be applied to large-scale neural network models,\nand will continuously update the library in the future. Using the developed\n$\\texttt{dattri}$ library, we are able to perform a comprehensive and fair\nbenchmark analysis across a wide range of data attribution methods. The source\ncode of $\\texttt{dattri}$ is available at https://github.com/TRAIS-Lab/dattri.\n", "date": "2024-10-06", "categories": []}
{"id": "2410.05057", "title": "SELECT: A Large-Scale Benchmark of Data Curation Strategies for Image\n  Classification", "abstract": "  Data curation is the problem of how to collect and organize samples into a\ndataset that supports efficient learning. Despite the centrality of the task,\nlittle work has been devoted towards a large-scale, systematic comparison of\nvarious curation methods. In this work, we take steps towards a formal\nevaluation of data curation strategies and introduce SELECT, the first\nlarge-scale benchmark of curation strategies for image classification.\n  In order to generate baseline methods for the SELECT benchmark, we create a\nnew dataset, ImageNet++, which constitutes the largest superset of ImageNet-1K\nto date. Our dataset extends ImageNet with 5 new training-data shifts, each\napproximately the size of ImageNet-1K itself, and each assembled using a\ndistinct curation strategy. We evaluate our data curation baselines in two\nways: (i) using each training-data shift to train identical image\nclassification models from scratch (ii) using the data itself to fit a\npretrained self-supervised representation.\n  Our findings show interesting trends, particularly pertaining to recent\nmethods for data curation such as synthetic data generation and lookup based on\nCLIP embeddings. We show that although these strategies are highly competitive\nfor certain tasks, the curation strategy used to assemble the original\nImageNet-1K dataset remains the gold standard. We anticipate that our benchmark\ncan illuminate the path for new methods to further reduce the gap. We release\nour checkpoints, code, documentation, and a link to our dataset at\nhttps://github.com/jimmyxu123/SELECT.\n", "date": "2024-10-07", "categories": []}
{"id": "2410.05103", "title": "MetaDD: Boosting Dataset Distillation with Neural Network\n  Architecture-Invariant Generalization", "abstract": "  Dataset distillation (DD) entails creating a refined, compact distilled\ndataset from a large-scale dataset to facilitate efficient training. A\nsignificant challenge in DD is the dependency between the distilled dataset and\nthe neural network (NN) architecture used. Training a different NN architecture\nwith a distilled dataset distilled using a specific architecture often results\nin diminished trainning performance for other architectures. This paper\nintroduces MetaDD, designed to enhance the generalizability of DD across\nvarious NN architectures. Specifically, MetaDD partitions distilled data into\nmeta features (i.e., the data's common characteristics that remain consistent\nacross different NN architectures) and heterogeneous features (i.e., the data's\nunique feature to each NN architecture). Then, MetaDD employs an\narchitecture-invariant loss function for multi-architecture feature alignment,\nwhich increases meta features and reduces heterogeneous features in distilled\ndata. As a low-memory consumption component, MetaDD can be seamlessly\nintegrated into any DD methodology. Experimental results demonstrate that\nMetaDD significantly improves performance across various DD methods. On the\nDistilled Tiny-Imagenet with Sre2L (50 IPC), MetaDD achieves cross-architecture\nNN accuracy of up to 30.1\\%, surpassing the second-best method (GLaD) by 1.7\\%.\n", "date": "2024-10-07", "categories": []}
{"id": "2408.03560", "title": "In2Core: Leveraging Influence Functions for Coreset Selection in\n  Instruction Finetuning of Large Language Models", "abstract": "  Despite advancements, fine-tuning Large Language Models (LLMs) remains costly\ndue to the extensive parameter count and substantial data requirements for\nmodel generalization. Accessibility to computing resources remains a barrier\nfor the open-source community. To address this challenge, we propose the\nIn2Core algorithm, which selects a coreset by analyzing the correlation between\ntraining and evaluation samples with a trained model. Notably, we assess the\nmodel's internal gradients to estimate this relationship, aiming to rank the\ncontribution of each training point. To enhance efficiency, we propose an\noptimization to compute influence functions with a reduced number of layers\nwhile achieving similar accuracy. By applying our algorithm to instruction\nfine-tuning data of LLMs, we can achieve similar performance with just 50% of\nthe training data. Meantime, using influence functions to analyze model\ncoverage to certain testing samples could provide a reliable and interpretable\nsignal on the training set's coverage of those test points.\n", "date": "2024-08-07", "categories": []}
{"id": "2410.02512", "title": "SAFLEX: Self-Adaptive Augmentation via Feature Label Extrapolation", "abstract": "  Data augmentation, a cornerstone technique in deep learning, is crucial in\nenhancing model performance, especially with scarce labeled data. While\ntraditional techniques are effective, their reliance on hand-crafted methods\nlimits their applicability across diverse data types and tasks. Although modern\nlearnable augmentation methods offer increased adaptability, they are\ncomputationally expensive and challenging to incorporate within prevalent\naugmentation workflows. In this work, we present a novel, efficient method for\ndata augmentation, effectively bridging the gap between existing augmentation\nstrategies and emerging datasets and learning tasks. We introduce SAFLEX\n(Self-Adaptive Augmentation via Feature Label EXtrapolation), which learns the\nsample weights and soft labels of augmented samples provided by any given\nupstream augmentation pipeline, using a specifically designed efficient bilevel\noptimization algorithm. Remarkably, SAFLEX effectively reduces the noise and\nlabel errors of the upstream augmentation pipeline with a marginal\ncomputational cost. As a versatile module, SAFLEX excels across diverse\ndatasets, including natural and medical images and tabular data, showcasing its\nprowess in few-shot learning and out-of-distribution generalization. SAFLEX\nseamlessly integrates with common augmentation strategies like RandAug, CutMix,\nand those from large pre-trained generative models like stable diffusion and is\nalso compatible with frameworks such as CLIP's fine-tuning. Our findings\nhighlight the potential to adapt existing augmentation pipelines for new data\ntypes and tasks, signaling a move towards more adaptable and resilient training\nframeworks.\n", "date": "2024-10-03", "categories": []}
{"id": "2406.17238", "title": "Generative Expansion of Small Datasets: An Expansive Graph Approach", "abstract": "  Limited data availability in machine learning significantly impacts\nperformance and generalization. Traditional augmentation methods enhance\nmoderately sufficient datasets. GANs struggle with convergence when generating\ndiverse samples. Diffusion models, while effective, have high computational\ncosts. We introduce an Expansive Synthesis model generating large-scale,\ninformation-rich datasets from minimal samples. It uses expander graph mappings\nand feature interpolation to preserve data distribution and feature\nrelationships. The model leverages neural networks' non-linear latent space,\ncaptured by a Koopman operator, to create a linear feature space for dataset\nexpansion. An autoencoder with self-attention layers and optimal transport\nrefines distributional consistency. We validate by comparing classifiers\ntrained on generated data to those trained on original datasets. Results show\ncomparable performance, demonstrating the model's potential to augment training\ndata effectively. This work advances data generation, addressing scarcity in\nmachine learning applications.\n", "date": "2024-06-25", "categories": []}
{"id": "2403.13893", "title": "DAVED: Data Acquisition via Experimental Design for Data Markets", "abstract": "  The acquisition of training data is crucial for machine learning\napplications. Data markets can increase the supply of data, particularly in\ndata-scarce domains such as healthcare, by incentivizing potential data\nproviders to join the market. A major challenge for a data buyer in such a\nmarket is choosing the most valuable data points from a data seller. Unlike\nprior work in data valuation, which assumes centralized data access, we propose\na federated approach to the data acquisition problem that is inspired by linear\nexperimental design. Our proposed data acquisition method achieves lower\nprediction error without requiring labeled validation data and can be optimized\nin a fast and federated procedure. The key insight of our work is that a method\nthat directly estimates the benefit of acquiring data for test set prediction\nis particularly compatible with a decentralized market setting.\n", "date": "2024-03-20", "categories": []}
{"id": "2409.19532", "title": "Video DataFlywheel: Resolving the Impossible Data Trinity in\n  Video-Language Understanding", "abstract": "  Recently, video-language understanding has achieved great success through\nlarge-scale pre-training. However, data scarcity remains a prevailing\nchallenge. This study quantitatively reveals an \"impossible trinity\" among data\nquantity, diversity, and quality in pre-training datasets. Recent efforts seek\nto refine large-scale, diverse ASR datasets compromised by low quality through\nsynthetic annotations. These methods successfully leverage useful information\nin multimodal video content (frames, tags, ASR transcripts, etc.) to refine the\noriginal annotations. Nevertheless, they struggle to mitigate noise within\nsynthetic annotations and lack scalability as the dataset size expands. To\naddress these issues, we introduce the Video DataFlywheel framework, which\niteratively refines video annotations with improved noise control methods. For\niterative refinement, we first leverage a video-language model to generate\nsynthetic annotations, resulting in a refined dataset. Then, we pre-train on it\nand fine-tune on human refinement examples for a stronger model. These\nprocesses are repeated for continuous improvement. For noise control, we\npresent AdaTaiLr, a novel noise control method that requires weaker assumptions\non noise distribution, thereby proving more effective in large datasets with\ntheoretical guarantees. The combination of iterative refinement and AdaTaiLr\ncan achieve better scalability in video-language understanding. Extensive\nexperiments show that our framework outperforms existing data refinement\nbaselines, delivering a 3% performance boost and improving dataset quality with\nminimal diversity loss. Furthermore, our refined dataset facilitates\nsignificant improvements in various video-language understanding tasks,\nincluding video question answering and text-video retrieval.\n", "date": "2024-09-29", "categories": []}
{"id": "2402.04924", "title": "Two Trades is not Baffled: Condensing Graph via Crafting Rational\n  Gradient Matching", "abstract": "  Training on large-scale graphs has achieved remarkable results in graph\nrepresentation learning, but its cost and storage have raised growing concerns.\nAs one of the most promising directions, graph condensation methods address\nthese issues by employing gradient matching, aiming to condense the full graph\ninto a more concise yet information-rich synthetic set. Though encouraging,\nthese strategies primarily emphasize matching directions of the gradients,\nwhich leads to deviations in the training trajectories. Such deviations are\nfurther magnified by the differences between the condensation and evaluation\nphases, culminating in accumulated errors, which detrimentally affect the\nperformance of the condensed graphs. In light of this, we propose a novel graph\ncondensation method named \\textbf{C}raf\\textbf{T}ing \\textbf{R}ationa\\textbf{L}\ntrajectory (\\textbf{CTRL}), which offers an optimized starting point closer to\nthe original dataset's feature distribution and a more refined strategy for\ngradient matching. Theoretically, CTRL can effectively neutralize the impact of\naccumulated errors on the performance of condensed graphs. We provide extensive\nexperiments on various graph datasets and downstream tasks to support the\neffectiveness of CTRL. Code is released at\nhttps://github.com/NUS-HPC-AI-Lab/CTRL.\n", "date": "2024-02-07", "categories": []}
{"id": "2409.16949", "title": "DALDA: Data Augmentation Leveraging Diffusion Model and LLM with\n  Adaptive Guidance Scaling", "abstract": "  In this paper, we present an effective data augmentation framework leveraging\nthe Large Language Model (LLM) and Diffusion Model (DM) to tackle the\nchallenges inherent in data-scarce scenarios. Recently, DMs have opened up the\npossibility of generating synthetic images to complement a few training images.\nHowever, increasing the diversity of synthetic images also raises the risk of\ngenerating samples outside the target distribution. Our approach addresses this\nissue by embedding novel semantic information into text prompts via LLM and\nutilizing real images as visual prompts, thus generating semantically rich\nimages. To ensure that the generated images remain within the target\ndistribution, we dynamically adjust the guidance weight based on each image's\nCLIPScore to control the diversity. Experimental results show that our method\nproduces synthetic images with enhanced diversity while maintaining adherence\nto the target distribution. Consequently, our approach proves to be more\nefficient in the few-shot setting on several benchmarks. Our code is available\nat https://github.com/kkyuhun94/dalda .\n", "date": "2024-09-25", "categories": []}
{"id": "2409.16239", "title": "Label-Augmented Dataset Distillation", "abstract": "  Traditional dataset distillation primarily focuses on image representation\nwhile often overlooking the important role of labels. In this study, we\nintroduce Label-Augmented Dataset Distillation (LADD), a new dataset\ndistillation framework enhancing dataset distillation with label augmentations.\nLADD sub-samples each synthetic image, generating additional dense labels to\ncapture rich semantics. These dense labels require only a 2.5% increase in\nstorage (ImageNet subsets) with significant performance benefits, providing\nstrong learning signals. Our label generation strategy can complement existing\ndataset distillation methods for significantly enhancing their training\nefficiency and performance. Experimental results demonstrate that LADD\noutperforms existing methods in terms of computational overhead and accuracy.\nWith three high-performance dataset distillation algorithms, LADD achieves\nremarkable gains by an average of 14.9% in accuracy. Furthermore, the\neffectiveness of our method is proven across various datasets, distillation\nhyperparameters, and algorithms. Finally, our method improves the\ncross-architecture robustness of the distilled dataset, which is important in\nthe application scenario.\n", "date": "2024-09-24", "categories": []}
{"id": "2409.13860", "title": "SSE: Multimodal Semantic Data Selection and Enrichment for\n  Industrial-scale Data Assimilation", "abstract": "  In recent years, the data collected for artificial intelligence has grown to\nan unmanageable amount. Particularly within industrial applications, such as\nautonomous vehicles, model training computation budgets are being exceeded\nwhile model performance is saturating -- and yet more data continues to pour\nin. To navigate the flood of data, we propose a framework to select the most\nsemantically diverse and important dataset portion. Then, we further\nsemantically enrich it by discovering meaningful new data from a massive\nunlabeled data pool. Importantly, we can provide explainability by leveraging\nfoundation models to generate semantics for every data point. We quantitatively\nshow that our Semantic Selection and Enrichment framework (SSE) can a)\nsuccessfully maintain model performance with a smaller training dataset and b)\nimprove model performance by enriching the smaller dataset without exceeding\nthe original dataset size. Consequently, we demonstrate that semantic diversity\nis imperative for optimal data selection and model performance.\n", "date": "2024-09-20", "categories": []}
{"id": "2409.13915", "title": "Data Pruning via Separability, Integrity, and Model Uncertainty-Aware\n  Importance Sampling", "abstract": "  This paper improves upon existing data pruning methods for image\nclassification by introducing a novel pruning metric and pruning procedure\nbased on importance sampling. The proposed pruning metric explicitly accounts\nfor data separability, data integrity, and model uncertainty, while the\nsampling procedure is adaptive to the pruning ratio and considers both\nintra-class and inter-class separation to further enhance the effectiveness of\npruning. Furthermore, the sampling method can readily be applied to other\npruning metrics to improve their performance. Overall, the proposed approach\nscales well to high pruning ratio and generalizes better across different\nclassification models, as demonstrated by experiments on four benchmark\ndatasets, including the fine-grained classification scenario.\n", "date": "2024-09-20", "categories": []}
{"id": "2409.14538", "title": "Towards Model-Agnostic Dataset Condensation by Heterogeneous Models", "abstract": "  Abstract. The advancement of deep learning has coincided with the\nproliferation of both models and available data. The surge in dataset sizes and\nthe subsequent surge in computational requirements have led to the development\nof the Dataset Condensation (DC). While prior studies have delved into\ngenerating synthetic images through methods like distribution alignment and\ntraining trajectory tracking for more efficient model training, a significant\nchallenge arises when employing these condensed images practically. Notably,\nthese condensed images tend to be specific to particular models, constraining\ntheir versatility and practicality. In response to this limitation, we\nintroduce a novel method, Heterogeneous Model Dataset Condensation (HMDC),\ndesigned to produce universally applicable condensed images through cross-model\ninteractions. To address the issues of gradient magnitude difference and\nsemantic distance in models when utilizing heterogeneous models, we propose the\nGradient Balance Module (GBM) and Mutual Distillation (MD) with the\nSpatialSemantic Decomposition method. By balancing the contribution of each\nmodel and maintaining their semantic meaning closely, our approach overcomes\nthe limitations associated with model-specific condensed images and enhances\nthe broader utility. The source code is available in\nhttps://github.com/KHU-AGI/HMDC.\n", "date": "2024-09-22", "categories": []}
{"id": "2409.15028", "title": "Region Mixup", "abstract": "  This paper introduces a simple extension of mixup (Zhang et al., 2018) data\naugmentation to enhance generalization in visual recognition tasks. Unlike the\nvanilla mixup method, which blends entire images, our approach focuses on\ncombining regions from multiple images.\n", "date": "2024-09-23", "categories": []}
{"id": "2407.07805", "title": "SUMix: Mixup with Semantic and Uncertain Information", "abstract": "  Mixup data augmentation approaches have been applied for various tasks of\ndeep learning to improve the generalization ability of deep neural networks.\nSome existing approaches CutMix, SaliencyMix, etc. randomly replace a patch in\none image with patches from another to generate the mixed image. Similarly, the\ncorresponding labels are linearly combined by a fixed ratio $\\lambda$ by l. The\nobjects in two images may be overlapped during the mixing process, so some\nsemantic information is corrupted in the mixed samples. In this case, the mixed\nimage does not match the mixed label information. Besides, such a label may\nmislead the deep learning model training, which results in poor performance. To\nsolve this problem, we proposed a novel approach named SUMix to learn the\nmixing ratio as well as the uncertainty for the mixed samples during the\ntraining process. First, we design a learnable similarity function to compute\nan accurate mix ratio. Second, an approach is investigated as a regularized\nterm to model the uncertainty of the mixed samples. We conduct experiments on\nfive image benchmarks, and extensive experimental results imply that our method\nis capable of improving the performance of classifiers with different\ncutting-based mixup approaches. The source code is available at\nhttps://github.com/JinXins/SUMix.\n", "date": "2024-07-10", "categories": []}
{"id": "2409.12112", "title": "Pareto Data Framework: Steps Towards Resource-Efficient Decision Making\n  Using Minimum Viable Data (MVD)", "abstract": "  This paper introduces the Pareto Data Framework, an approach for identifying\nand selecting the Minimum Viable Data (MVD) required for enabling machine\nlearning applications on constrained platforms such as embedded systems, mobile\ndevices, and Internet of Things (IoT) devices. We demonstrate that strategic\ndata reduction can maintain high performance while significantly reducing\nbandwidth, energy, computation, and storage costs. The framework identifies\nMinimum Viable Data (MVD) to optimize efficiency across resource-constrained\nenvironments without sacrificing performance. It addresses common inefficient\npractices in an IoT application such as overprovisioning of sensors and\noverprecision, and oversampling of signals, proposing scalable solutions for\noptimal sensor selection, signal extraction and transmission, and data\nrepresentation. An experimental methodology demonstrates effective acoustic\ndata characterization after downsampling, quantization, and truncation to\nsimulate reduced-fidelity sensors and network and storage constraints; results\nshows that performance can be maintained up to 95\\% with sample rates reduced\nby 75\\% and bit depths and clip length reduced by 50\\% which translates into\nsubstantial cost and resource reduction. These findings have implications on\nthe design and development of constrained systems. The paper also discusses\nbroader implications of the framework, including the potential to democratize\nadvanced AI technologies across IoT applications and sectors such as\nagriculture, transportation, and manufacturing to improve access and multiply\nthe benefits of data-driven insights.\n", "date": "2024-09-18", "categories": []}
{"id": "2409.11378", "title": "Diversify and Conquer: Diversity-Centric Data Selection with Iterative\n  Refinement", "abstract": "  Finetuning large language models on instruction data is crucial for enhancing\npre-trained knowledge and improving instruction-following capabilities. As\ninstruction datasets proliferate, selecting optimal data for effective training\nbecomes increasingly important. This work addresses the question: How can we\ndetermine the optimal subset of data for effective training? While existing\nresearch often emphasizes local criteria like instance quality for subset\nselection, we argue that a global approach focused on data diversity is more\ncritical. Our method employs k-means clustering to ensure the selected subset\neffectively represents the full dataset. We propose an iterative refinement\nmethod inspired by active learning techniques to resample instances from\nclusters, reassessing each cluster's importance and sampling weight in every\ntraining iteration. This approach reduces the effect of outliers and\nautomatically filters out clusters containing low-quality data. Through\nextensive evaluation across natural language reasoning, general world\nknowledge, code and math reasoning tasks, and by fine-tuning models from\nvarious families, we observe consistent improvements, achieving a 7% increase\nover random selection and a 3.8% improvement over state-of-the-art sampling\nmethods. Our work highlights the significance of diversity-first sampling when\nfinetuning LLMs to enhance performance across a broad array of evaluation\ntasks. Our code is available at\nhttps://github.com/for-ai/iterative-data-selection.\n", "date": "2024-09-17", "categories": []}
{"id": "2409.10120", "title": "Data-Centric Strategies for Overcoming PET/CT Heterogeneity: Insights\n  from the AutoPET III Lesion Segmentation Challenge", "abstract": "  The third autoPET challenge introduced a new data-centric task this year,\nshifting the focus from model development to improving metastatic lesion\nsegmentation on PET/CT images through data quality and handling strategies. In\nresponse, we developed targeted methods to enhance segmentation performance\ntailored to the characteristics of PET/CT imaging. Our approach encompasses two\nkey elements. First, to address potential alignment errors between CT and PET\nmodalities as well as the prevalence of punctate lesions, we modified the\nbaseline data augmentation scheme and extended it with misalignment\naugmentation. This adaptation aims to improve segmentation accuracy,\nparticularly for tiny metastatic lesions. Second, to tackle the variability in\nimage dimensions significantly affecting the prediction time, we implemented a\ndynamic ensembling and test-time augmentation (TTA) strategy. This method\noptimizes the use of ensembling and TTA within a 5-minute prediction time\nlimit, effectively leveraging the generalization potential for both small and\nlarge images. Both of our solutions are designed to be robust across different\ntracers and institutional settings, offering a general, yet imaging-specific\napproach to the multi-tracer and multi-institutional challenges of the\ncompetition. We made the challenge repository with our modifications publicly\navailable at \\url{https://github.com/MIC-DKFZ/miccai2024_autopet3_datacentric}.\n", "date": "2024-09-16", "categories": []}
{"id": "2409.06290", "title": "EntAugment: Entropy-Driven Adaptive Data Augmentation Framework for\n  Image Classification", "abstract": "  Data augmentation (DA) has been widely used to improve the generalization of\ndeep neural networks. While existing DA methods have proven effective, they\noften rely on augmentation operations with random magnitudes to each sample.\nHowever, this approach can inadvertently introduce noise, induce distribution\nshifts, and increase the risk of overfitting. In this paper, we propose\nEntAugment, a tuning-free and adaptive DA framework. Unlike previous work,\nEntAugment dynamically assesses and adjusts the augmentation magnitudes for\neach sample during training, leveraging insights into both the inherent\ncomplexities of training samples and the evolving status of deep models.\nSpecifically, in EntAugment, the magnitudes are determined by the information\nentropy derived from the probability distribution obtained by applying the\nsoftmax function to the model's output. In addition, to further enhance the\nefficacy of EntAugment, we introduce a novel entropy regularization term,\nEntLoss, which complements the EntAugment approach. Theoretical analysis\nfurther demonstrates that EntLoss, compared to traditional cross-entropy loss,\nachieves closer alignment between the model distributions and underlying\ndataset distributions. Moreover, EntAugment and EntLoss can be utilized\nseparately or jointly. We conduct extensive experiments across multiple image\nclassification tasks and network architectures with thorough comparisons of\nexisting DA methods. Importantly, the proposed methods outperform others\nwithout introducing any auxiliary models or noticeable extra computational\ncosts, highlighting both effectiveness and efficiency. Code is available at\nhttps://github.com/Jackbrocp/EntAugment.\n", "date": "2024-09-10", "categories": []}
{"id": "2404.01067", "title": "Exploring the Mystery of Influential Data for Mathematical Reasoning", "abstract": "  Selecting influential data for fine-tuning on downstream tasks is a key\nfactor for both performance and computation efficiency. Recent works have shown\nthat training with only limited data can show a superior performance on general\ntasks. However, the feasibility on mathematical reasoning tasks has not been\nvalidated. To go further, there exist two open questions for mathematical\nreasoning: how to select influential data and what is an influential data\ncomposition. For the former one, we propose a Quality-aware Diverse Selection\n(QaDS) strategy adaptable for mathematical reasoning. A comparison with other\nselection strategies validates the superiority of QaDS. For the latter one, we\nfirst enlarge our setting and explore the influential data composition. We\nconduct a series of experiments and highlight: scaling up reasoning data, and\ntraining with general data selected by QaDS is helpful. Then, we define our\noptimal mixture as OpenMathMix, an influential data mixture with open-source\ndata selected by QaDS. With OpenMathMix, we achieve a state-of-the-art 48.8%\naccuracy on MATH with 7B base model. Additionally, we showcase the use of QaDS\nin creating efficient fine-tuning mixtures with various selection ratios, and\nanalyze the quality of a wide range of open-source datasets, which can perform\nas a reference for future works on mathematical reasoning tasks.\n", "date": "2024-04-01", "categories": []}
{"id": "2409.04820", "title": "FreeAugment: Data Augmentation Search Across All Degrees of Freedom", "abstract": "  Data augmentation has become an integral part of deep learning, as it is\nknown to improve the generalization capabilities of neural networks. Since the\nmost effective set of image transformations differs between tasks and domains,\nautomatic data augmentation search aims to alleviate the extreme burden of\nmanually finding the optimal image transformations. However, current methods\nare not able to jointly optimize all degrees of freedom: (1) the number of\ntransformations to be applied, their (2) types, (3) order, and (4) magnitudes.\nMany existing methods risk picking the same transformation more than once,\nlimit the search to two transformations only, or search for the number of\ntransformations exhaustively or iteratively in a myopic manner. Our approach,\nFreeAugment, is the first to achieve global optimization of all four degrees of\nfreedom simultaneously, using a fully differentiable method. It efficiently\nlearns the number of transformations and a probability distribution over their\npermutations, inherently refraining from redundant repetition while sampling.\nOur experiments demonstrate that this joint learning of all degrees of freedom\nsignificantly improves performance, achieving state-of-the-art results on\nvarious natural image benchmarks and beyond across other domains. Project page\nat https://tombekor.github.io/FreeAugment-web\n", "date": "2024-09-07", "categories": []}
{"id": "2409.03929", "title": "Data-Efficient Generation for Dataset Distillation", "abstract": "  While deep learning techniques have proven successful in image-related tasks,\nthe exponentially increased data storage and computation costs become a\nsignificant challenge. Dataset distillation addresses these challenges by\nsynthesizing only a few images for each class that encapsulate all essential\ninformation. Most current methods focus on matching. The problems lie in the\nsynthetic images not being human-readable and the dataset performance being\ninsufficient for downstream learning tasks. Moreover, the distillation time can\nquickly get out of bounds when the number of synthetic images per class\nincreases even slightly. To address this, we train a class conditional latent\ndiffusion model capable of generating realistic synthetic images with labels.\nThe sampling time can be reduced to several tens of images per seconds. We\ndemonstrate that models can be effectively trained using only a small set of\nsynthetic images and evaluated on a large real test set. Our approach achieved\nrank \\(1\\) in The First Dataset Distillation Challenge at ECCV 2024 on the\nCIFAR100 and TinyImageNet datasets.\n", "date": "2024-09-05", "categories": []}
{"id": "2312.06726", "title": "Filter & Align: Leveraging Human Knowledge to Curate Image-Text Data", "abstract": "  The increasing availability of image-text pairs has largely fueled the rapid\nadvancement in vision-language foundation models. However, the vast scale of\nthese datasets inevitably introduces significant variability in data quality,\nwhich can adversely affect the model performance. This highlights the critical\nrole of data filtering, not only to enhance training efficiency but also to\nimprove overall data quality. Existing methods typically rely on metrics such\nas CLIP Score and BLIP Score, which are derived from pre-trained models.\nHowever, these models are often trained on uncurated, noisy datasets, which can\nperpetuate errors and misalignments in the filtered dataset. We present a novel\nalgorithm that incorporates human knowledge on image-text alignment to guide\nfiltering vast corpus of web-crawled image-text datasets into a compact and\nhigh-quality form. To systemically capture human preferences on image-text\nalignments, we collect a diverse image-text dataset where each image is\nassociated with multiple captions from various sources, and establish a\ncomprehensive set of both subjective and objective criteria for critically\nguiding the alignment assessment from labelers. Additionally, we train a reward\nmodel on these human-preference annotations to internalize the nuanced human\nunderstanding of image-text alignment. The resulting reward model thus can act\nas a human-like referee to filter image-text pairs. Extensive experiments\ndemonstrate that we can maintain, sometimes even improve, model performance\nwhile compressing the image-text datasets up to ~90%. An impressive example is\nthat, by aggressively reducing the total training sample from 130M to only\n15.5M, our BLIP-B/16 models consistently show an average improvement of 2.9% on\nretrieval tasks and 11.5% on captioning tasks compared to full-size-dataset\ncounterparts.\n", "date": "2023-12-11", "categories": []}
{"id": "2409.00547", "title": "Data Augmentation for Image Classification using Generative AI", "abstract": "  Scaling laws dictate that the performance of AI models is proportional to the\namount of available data. Data augmentation is a promising solution to\nexpanding the dataset size. Traditional approaches focused on augmentation\nusing rotation, translation, and resizing. Recent approaches use generative AI\nmodels to improve dataset diversity. However, the generative methods struggle\nwith issues such as subject corruption and the introduction of irrelevant\nartifacts. In this paper, we propose the Automated Generative Data Augmentation\n(AGA). The framework combines the utility of large language models (LLMs),\ndiffusion models, and segmentation models to augment data. AGA preserves\nforeground authenticity while ensuring background diversity. Specific\ncontributions include: i) segment and superclass based object extraction, ii)\nprompt diversity with combinatorial complexity using prompt decomposition, and\niii) affine subject manipulation. We evaluate AGA against state-of-the-art\n(SOTA) techniques on three representative datasets, ImageNet, CUB, and\niWildCam. The experimental evaluation demonstrates an accuracy improvement of\n15.6% and 23.5% for in and out-of-distribution data compared to baseline\nmodels, respectively. There is also a 64.3% improvement in SIC score compared\nto the baselines.\n", "date": "2024-08-31", "categories": []}
{"id": "2409.01410", "title": "Dataset Distillation from First Principles: Integrating Core Information\n  Extraction and Purposeful Learning", "abstract": "  Dataset distillation (DD) is an increasingly important technique that focuses\non constructing a synthetic dataset capable of capturing the core information\nin training data to achieve comparable performance in models trained on the\nlatter. While DD has a wide range of applications, the theory supporting it is\nless well evolved. New methods of DD are compared on a common set of\nbenchmarks, rather than oriented towards any particular learning task. In this\nwork, we present a formal model of DD, arguing that a precise characterization\nof the underlying optimization problem must specify the inference task\nassociated with the application of interest. Without this task-specific focus,\nthe DD problem is under-specified, and the selection of a DD algorithm for a\nparticular task is merely heuristic. Our formalization reveals novel\napplications of DD across different modeling environments. We analyze existing\nDD methods through this broader lens, highlighting their strengths and\nlimitations in terms of accuracy and faithfulness to optimal DD operation.\nFinally, we present numerical results for two case studies important in\ncontemporary settings. Firstly, we address a critical challenge in medical data\nanalysis: merging the knowledge from different datasets composed of\nintersecting, but not identical, sets of features, in order to construct a\nlarger dataset in what is usually a small sample setting. Secondly, we consider\nout-of-distribution error across boundary conditions for physics-informed\nneural networks (PINNs), showing the potential for DD to provide more\nphysically faithful data. By establishing this general formulation of DD, we\naim to establish a new research paradigm by which DD can be understood and from\nwhich new DD techniques can arise.\n", "date": "2024-09-02", "categories": []}
{"id": "2408.16871", "title": "GSTAM: Efficient Graph Distillation with Structural Attention-Matching", "abstract": "  Graph distillation has emerged as a solution for reducing large graph\ndatasets to smaller, more manageable, and informative ones. Existing methods\nprimarily target node classification, involve computationally intensive\nprocesses, and fail to capture the true distribution of the full graph dataset.\nTo address these issues, we introduce Graph Distillation with Structural\nAttention Matching (GSTAM), a novel method for condensing graph classification\ndatasets. GSTAM leverages the attention maps of GNNs to distill structural\ninformation from the original dataset into synthetic graphs. The structural\nattention-matching mechanism exploits the areas of the input graph that GNNs\nprioritize for classification, effectively distilling such information into the\nsynthetic graphs and improving overall distillation performance. Comprehensive\nexperiments demonstrate GSTAM's superiority over existing methods, achieving\n0.45% to 6.5% better performance in extreme condensation ratios, highlighting\nits potential use in advancing distillation for graph classification tasks\n(Code available at https://github.com/arashrasti96/GSTAM).\n", "date": "2024-08-29", "categories": []}
{"id": "2408.16236", "title": "Neural Spectral Decomposition for Dataset Distillation", "abstract": "  In this paper, we propose Neural Spectrum Decomposition, a generic\ndecomposition framework for dataset distillation. Unlike previous methods, we\nconsider the entire dataset as a high-dimensional observation that is low-rank\nacross all dimensions. We aim to discover the low-rank representation of the\nentire dataset and perform distillation efficiently. Toward this end, we learn\na set of spectrum tensors and transformation matrices, which, through simple\nmatrix multiplication, reconstruct the data distribution. Specifically, a\nspectrum tensor can be mapped back to the image space by a transformation\nmatrix, and efficient information sharing during the distillation learning\nprocess is achieved through pairwise combinations of different spectrum vectors\nand transformation matrices. Furthermore, we integrate a trajectory matching\noptimization method guided by a real distribution. Our experimental results\ndemonstrate that our approach achieves state-of-the-art performance on\nbenchmarks, including CIFAR10, CIFAR100, Tiny Imagenet, and ImageNet Subset.\nOur code are available at \\url{https://github.com/slyang2021/NSD}.\n", "date": "2024-08-29", "categories": []}
{"id": "2408.16261", "title": "Evaluating Time-Series Training Dataset through Lens of Spectrum in Deep\n  State Space Models", "abstract": "  This study investigates a method to evaluate time-series datasets in terms of\nthe performance of deep neural networks (DNNs) with state space models (deep\nSSMs) trained on the dataset. SSMs have attracted attention as components\ninside DNNs to address time-series data. Since deep SSMs have powerful\nrepresentation capacities, training datasets play a crucial role in solving a\nnew task. However, the effectiveness of training datasets cannot be known until\ndeep SSMs are actually trained on them. This can increase the cost of data\ncollection for new tasks, as a trial-and-error process of data collection and\ntime-consuming training are needed to achieve the necessary performance. To\nadvance the practical use of deep SSMs, the metric of datasets to estimate the\nperformance early in the training can be one key element. To this end, we\nintroduce the concept of data evaluation methods used in system identification.\nIn system identification of linear dynamical systems, the effectiveness of\ndatasets is evaluated by using the spectrum of input signals. We introduce this\nconcept to deep SSMs, which are nonlinear dynamical systems. We propose the\nK-spectral metric, which is the sum of the top-K spectra of signals inside deep\nSSMs, by focusing on the fact that each layer of a deep SSM can be regarded as\na linear dynamical system. Our experiments show that the K-spectral metric has\na large absolute value of the correlation coefficient with the performance and\ncan be used to evaluate the quality of training datasets.\n", "date": "2024-08-29", "categories": []}
{"id": "2408.16268", "title": "UDD: Dataset Distillation via Mining Underutilized Regions", "abstract": "  Dataset distillation synthesizes a small dataset such that a model trained on\nthis set approximates the performance of the original dataset. Recent studies\non dataset distillation focused primarily on the design of the optimization\nprocess, with methods such as gradient matching, feature alignment, and\ntraining trajectory matching. However, little attention has been given to the\nissue of underutilized regions in synthetic images. In this paper, we propose\nUDD, a novel approach to identify and exploit the underutilized regions to make\nthem informative and discriminate, and thus improve the utilization of the\nsynthetic dataset. Technically, UDD involves two underutilized regions\nsearching policies for different conditions, i.e., response-based policy and\ndata jittering-based policy. Compared with previous works, such two policies\nare utilization-sensitive, equipping with the ability to dynamically adjust the\nunderutilized regions during the training process. Additionally, we analyze the\ncurrent model optimization problem and design a category-wise feature\ncontrastive loss, which can enhance the distinguishability of different\ncategories and alleviate the shortcomings of the existing multi-formation\nmethods. Experimentally, our method improves the utilization of the synthetic\ndataset and outperforms the state-of-the-art methods on various datasets, such\nas MNIST, FashionMNIST, SVHN, CIFAR-10, and CIFAR-100. For example, the\nimprovements on CIFAR-10 and CIFAR-100 are 4.0\\% and 3.7\\% over the next best\nmethod with IPC=1, by mining the underutilized regions.\n", "date": "2024-08-29", "categories": []}
{"id": "2401.16553", "title": "SelectLLM: Can LLMs Select Important Instructions to Annotate?", "abstract": "  Instruction tuning benefits from large and diverse datasets; however,\ncreating such datasets involves a high cost of human labeling. While synthetic\ndatasets generated by large language models (LLMs) have partly solved this\nissue, they often contain low-quality data. One effective solution is\nselectively annotating unlabelled instructions, especially given the relative\nease of acquiring unlabeled instructions or texts from various sources.\nHowever, how to select unlabelled instructions is not well-explored, especially\nin the context of LLMs. Therefore, we introduce SelectLLM, an alternative\nframework that leverages the capabilities of LLMs to select unlabeled\ninstructions more effectively. Specifically, SelectLLM consists of two key\nsteps: Coreset-based clustering of unlabelled instructions for enlarging\ndiversity and prompting of LLM to identify the most beneficial instructions\nwithin each cluster. We evaluate SelectLLM on AlpacaEval2 and MT-Bench,\ndemonstrating its ability to outperform state-of-the-art methods like\nAlpagasus. In addition, we compare the performance and compatibility of\nSelectLLM with various LLMs, such as ChatGPT, LLaMA-3.1-70B, and Gemma-2-27b.\nSelectLLM's adaptability and robustness are further evidenced by its ability to\nmaintain high performance across both human and synthetic datasets. All code\nand data are publicly available (https://github.com/minnesotanlp/select-llm).\n", "date": "2024-01-29", "categories": []}
{"id": "2408.13426", "title": "Optimal Layer Selection for Latent Data Augmentation", "abstract": "  While data augmentation (DA) is generally applied to input data, several\nstudies have reported that applying DA to hidden layers in neural networks,\ni.e., feature augmentation, can improve performance. However, in previous\nstudies, the layers to which DA is applied have not been carefully considered,\noften being applied randomly and uniformly or only to a specific layer, leaving\nroom for arbitrariness. Thus, in this study, we investigated the trends of\nsuitable layers for applying DA in various experimental configurations, e.g.,\ntraining from scratch, transfer learning, various dataset settings, and\ndifferent models. In addition, to adjust the suitable layers for DA\nautomatically, we propose the adaptive layer selection (AdaLASE) method, which\nupdates the ratio to perform DA for each layer based on the gradient descent\nmethod during training. The experimental results obtained on several image\nclassification datasets indicate that the proposed AdaLASE method altered the\nratio as expected and achieved high overall test accuracy.\n", "date": "2024-08-24", "categories": []}
{"id": "2408.13788", "title": "3D-VirtFusion: Synthetic 3D Data Augmentation through Generative\n  Diffusion Models and Controllable Editing", "abstract": "  Data augmentation plays a crucial role in deep learning, enhancing the\ngeneralization and robustness of learning-based models. Standard approaches\ninvolve simple transformations like rotations and flips for generating extra\ndata. However, these augmentations are limited by their initial dataset,\nlacking high-level diversity. Recently, large models such as language models\nand diffusion models have shown exceptional capabilities in perception and\ncontent generation. In this work, we propose a new paradigm to automatically\ngenerate 3D labeled training data by harnessing the power of pretrained large\nfoundation models. For each target semantic class, we first generate 2D images\nof a single object in various structure and appearance via diffusion models and\nchatGPT generated text prompts. Beyond texture augmentation, we propose a\nmethod to automatically alter the shape of objects within 2D images.\nSubsequently, we transform these augmented images into 3D objects and construct\nvirtual scenes by random composition. This method can automatically produce a\nsubstantial amount of 3D scene data without the need of real data, providing\nsignificant benefits in addressing few-shot learning challenges and mitigating\nlong-tailed class imbalances. By providing a flexible augmentation approach,\nour work contributes to enhancing 3D data diversity and advancing model\ncapabilities in scene understanding tasks.\n", "date": "2024-08-25", "categories": []}
{"id": "2408.14381", "title": "Learning Tree-Structured Composition of Data Augmentation", "abstract": "  Data augmentation is widely used for training a neural network given little\nlabeled data. A common practice of augmentation training is applying a\ncomposition of multiple transformations sequentially to the data. Existing\naugmentation methods such as RandAugment randomly sample from a list of\npre-selected transformations, while methods such as AutoAugment apply advanced\nsearch to optimize over an augmentation set of size $k^d$, which is the number\nof transformation sequences of length $d$, given a list of $k$ transformations.\n  In this paper, we design efficient algorithms whose running time complexity\nis much faster than the worst-case complexity of $O(k^d)$, provably. We propose\na new algorithm to search for a binary tree-structured composition of $k$\ntransformations, where each tree node corresponds to one transformation. The\nbinary tree generalizes sequential augmentations, such as the SimCLR\naugmentation scheme for contrastive learning. Using a top-down, recursive\nsearch procedure, our algorithm achieves a runtime complexity of $O(2^d k)$,\nwhich is much faster than $O(k^d)$ as $k$ increases above $2$. We apply our\nalgorithm to tackle data distributions with heterogeneous subpopulations by\nsearching for one tree in each subpopulation and then learning a weighted\ncombination, resulting in a forest of trees.\n  We validate our proposed algorithms on numerous graph and image datasets,\nincluding a multi-label graph classification dataset we collected. The dataset\nexhibits significant variations in the sizes of graphs and their average\ndegrees, making it ideal for studying data augmentation. We show that our\napproach can reduce the computation cost by 43% over existing search methods\nwhile improving performance by 4.3%. The tree structures can be used to\ninterpret the relative importance of each transformation, such as identifying\nthe important transformations on small vs. large graphs.\n", "date": "2024-08-26", "categories": []}
{"id": "2408.12659", "title": "Disentangled Structural and Featural Representation for Task-Agnostic\n  Graph Valuation", "abstract": "  With the emergence of data marketplaces, the demand for methods to assess the\nvalue of data has increased significantly. While numerous techniques have been\nproposed for this purpose, none have specifically addressed graphs as the main\ndata modality. Graphs are widely used across various fields, ranging from\nchemical molecules to social networks. In this study, we break down graphs into\ntwo main components: structural and featural, and we focus on evaluating data\nwithout relying on specific task-related metrics, making it applicable in\npractical scenarios where validation requirements may be lacking. We introduce\na novel framework called blind message passing, which aligns the seller's and\nbuyer's graphs using a shared node permutation based on graph matching. This\nallows us to utilize the graph Wasserstein distance to quantify the differences\nin the structural distribution of graph datasets, called the structural\ndisparities. We then consider featural aspects of buyers' and sellers' graphs\nfor data valuation and capture their statistical similarities and differences,\nreferred to as relevance and diversity, respectively. Our approach ensures that\nbuyers and sellers remain unaware of each other's datasets. Our experiments on\nreal datasets demonstrate the effectiveness of our approach in capturing the\nrelevance, diversity, and structural disparities of seller data for buyers,\nparticularly in graph-based data valuation scenarios.\n", "date": "2024-08-22", "categories": []}
{"id": "2308.07545", "title": "Vision-Language Dataset Distillation", "abstract": "  Dataset distillation methods reduce large-scale datasets to smaller sets of\nsynthetic data, preserving sufficient information to quickly train a new model\nfrom scratch. However, prior work on dataset distillation has focused\nexclusively on image classification datasets, whereas modern large-scale\ndatasets are primarily vision-language datasets. In this work, we design the\nfirst vision-language dataset distillation method, building on the idea of\ntrajectory matching. A key challenge is that vision-language datasets do not\nhave a set of discrete classes. To overcome this, our proposed method jointly\ndistills image-text pairs in a contrastive formulation. Further, we leverage\nLow-Rank Adaptation (LoRA) matching to enable more efficient and effective\ntrajectory matching in complex modern vision-language models. Since there are\nno existing baselines, we compare our distillation approach with three adapted\nvision-language coreset selection methods. We demonstrate significant\nimprovements on the challenging Flickr30K and COCO retrieval benchmarks: for\nexample, on Flickr30K, the best coreset selection method selecting 1000\nimage-text pairs for training achieves only 5.6% image-to-text retrieval\naccuracy (i.e., recall@1); in contrast, our dataset distillation almost doubles\nthat to 9.9% with just 100 training pairs, an order of magnitude fewer.\n", "date": "2023-08-15", "categories": []}
{"id": "2408.09709", "title": "Dataset Distillation for Histopathology Image Classification", "abstract": "  Deep neural networks (DNNs) have exhibited remarkable success in the field of\nhistopathology image analysis. On the other hand, the contemporary trend of\nemploying large models and extensive datasets has underscored the significance\nof dataset distillation, which involves compressing large-scale datasets into a\ncondensed set of synthetic samples, offering distinct advantages in improving\ntraining efficiency and streamlining downstream applications. In this work, we\nintroduce a novel dataset distillation algorithm tailored for histopathology\nimage datasets (Histo-DD), which integrates stain normalisation and model\naugmentation into the distillation progress. Such integration can substantially\nenhance the compatibility with histopathology images that are often\ncharacterised by high colour heterogeneity. We conduct a comprehensive\nevaluation of the effectiveness of the proposed algorithm and the generated\nhistopathology samples in both patch-level and slide-level classification\ntasks. The experimental results, carried out on three publicly available WSI\ndatasets, including Camelyon16, TCGA-IDH, and UniToPath, demonstrate that the\nproposed Histo-DD can generate more informative synthetic patches than previous\ncoreset selection and patch sampling methods. Moreover, the synthetic samples\ncan preserve discriminative information, substantially reduce training efforts,\nand exhibit architecture-agnostic properties. These advantages indicate that\nsynthetic samples can serve as an alternative to large-scale datasets.\n", "date": "2024-08-19", "categories": []}
{"id": "2401.17542", "title": "A Medical Data-Effective Learning Benchmark for Highly Efficient\n  Pre-training of Foundation Models", "abstract": "  Foundation models, pre-trained on massive datasets, have achieved\nunprecedented generalizability. However, is it truly necessary to involve such\nvast amounts of data in pre-training, consuming extensive computational\nresources? This paper introduces data-effective learning, aiming to use data in\nthe most impactful way to pre-train foundation models. This involves strategies\nthat focus on data quality rather than quantity, ensuring the data used for\ntraining has high informational value. Data-effective learning plays a profound\nrole in accelerating foundation model training, reducing computational costs,\nand saving data storage, which is very important as the volume of medical data\nin recent years has grown beyond many people's expectations. However, due to\nthe lack of standards and comprehensive benchmarks, research on medical\ndata-effective learning is poorly studied. To address this gap, our paper\nintroduces a comprehensive benchmark specifically for evaluating data-effective\nlearning in the medical field. This benchmark includes a dataset with millions\nof data samples from 31 medical centers (DataDEL), a baseline method for\ncomparison (MedDEL), and a new evaluation metric (NormDEL) to objectively\nmeasure data-effective learning performance. Our extensive experimental results\nshow the baseline MedDEL can achieve performance comparable to the original\nlarge dataset with only 5% of the data. Establishing such an open\ndata-effective learning benchmark is crucial for the medical foundation model\nresearch community because it facilitates efficient data use, promotes\ncollaborative breakthroughs, and fosters the development of cost-effective,\nscalable, and impactful healthcare solutions.\n", "date": "2024-01-31", "categories": []}
{"id": "2408.08610", "title": "Generative Dataset Distillation Based on Diffusion Model", "abstract": "  This paper presents our method for the generative track of The First Dataset\nDistillation Challenge at ECCV 2024. Since the diffusion model has become the\nmainstay of generative models because of its high-quality generative effects,\nwe focus on distillation methods based on the diffusion model. Considering that\nthe track can only generate a fixed number of images in 10 minutes using a\ngenerative model for CIFAR-100 and Tiny-ImageNet datasets, we need to use a\ngenerative model that can generate images at high speed. In this study, we\nproposed a novel generative dataset distillation method based on Stable\nDiffusion. Specifically, we use the SDXL-Turbo model which can generate images\nat high speed and quality. Compared to other diffusion models that can only\ngenerate images per class (IPC) = 1, our method can achieve an IPC = 10 for\nTiny-ImageNet and an IPC = 20 for CIFAR-100, respectively. Additionally, to\ngenerate high-quality distilled datasets for CIFAR-100 and Tiny-ImageNet, we\nuse the class information as text prompts and post data augmentation for the\nSDXL-Turbo model. Experimental results show the effectiveness of the proposed\nmethod, and we achieved third place in the generative track of the ECCV 2024 DD\nChallenge. Codes are available at https://github.com/Guang000/BANKO.\n", "date": "2024-08-16", "categories": []}
{"id": "2408.08201", "title": "Heavy Labels Out! Dataset Distillation with Label Space Lightening", "abstract": "  Dataset distillation or condensation aims to condense a large-scale training\ndataset into a much smaller synthetic one such that the training performance of\ndistilled and original sets on neural networks are similar. Although the number\nof training samples can be reduced substantially, current state-of-the-art\nmethods heavily rely on enormous soft labels to achieve satisfactory\nperformance. As a result, the required storage can be comparable even to\noriginal datasets, especially for large-scale ones. To solve this problem,\ninstead of storing these heavy labels, we propose a novel label-lightening\nframework termed HeLlO aiming at effective image-to-label projectors, with\nwhich synthetic labels can be directly generated online from synthetic images.\nSpecifically, to construct such projectors, we leverage prior knowledge in\nopen-source foundation models, e.g., CLIP, and introduce a LoRA-like\nfine-tuning strategy to mitigate the gap between pre-trained and target\ndistributions, so that original models for soft-label generation can be\ndistilled into a group of low-rank matrices. Moreover, an effective image\noptimization method is proposed to further mitigate the potential error between\nthe original and distilled label generators. Extensive experiments demonstrate\nthat with only about 0.003% of the original storage required for a complete set\nof soft labels, we achieve comparable performance to current state-of-the-art\ndataset distillation methods on large-scale datasets. Our code will be\navailable.\n", "date": "2024-08-15", "categories": []}
{"id": "2408.08310", "title": "ScalingFilter: Assessing Data Quality through Inverse Utilization of\n  Scaling Laws", "abstract": "  High-quality data is crucial for the pre-training performance of large\nlanguage models. Unfortunately, existing quality filtering methods rely on a\nknown high-quality dataset as reference, which can introduce potential bias and\ncompromise diversity. In this paper, we propose ScalingFilter, a novel approach\nthat evaluates text quality based on the perplexity difference between two\nlanguage models trained on the same data, thereby eliminating the influence of\nthe reference dataset in the filtering process. An theoretical analysis shows\nthat ScalingFilter is equivalent to an inverse utilization of scaling laws.\nThrough training models with 1.3B parameters on the same data source processed\nby various quality filters, we find ScalingFilter can improve zero-shot\nperformance of pre-trained models in downstream tasks. To assess the bias\nintroduced by quality filtering, we introduce semantic diversity, a metric of\nutilizing text embedding models for semantic representations. Extensive\nexperiments reveal that semantic diversity is a reliable indicator of dataset\ndiversity, and ScalingFilter achieves an optimal balance between downstream\nperformance and semantic diversity.\n", "date": "2024-08-15", "categories": []}
{"id": "2307.10988", "title": "On minimizing the training set fill distance in machine learning\n  regression", "abstract": "  For regression tasks one often leverages large datasets for training\npredictive machine learning models. However, using large datasets may not be\nfeasible due to computational limitations or high data labelling costs.\nTherefore, suitably selecting small training sets from large pools of\nunlabelled data points is essential to maximize model performance while\nmaintaining efficiency. In this work, we study Farthest Point Sampling (FPS), a\ndata selection approach that aims to minimize the fill distance of the selected\nset. We derive an upper bound for the maximum expected prediction error,\nconditional to the location of the unlabelled data points, that linearly\ndepends on the training set fill distance. For empirical validation, we perform\nexperiments using two regression models on three datasets. We empirically show\nthat selecting a training set by aiming to minimize the fill distance, thereby\nminimizing our derived bound, significantly reduces the maximum prediction\nerror of various regression models, outperforming alternative sampling\napproaches by a large margin. Furthermore, we show that selecting training sets\nwith the FPS can also increase model stability for the specific case of\nGaussian kernel regression approaches.\n", "date": "2023-07-20", "categories": []}
{"id": "2403.07854", "title": "Distilling the Knowledge in Data Pruning", "abstract": "  With the increasing size of datasets used for training neural networks, data\npruning becomes an attractive field of research. However, most current data\npruning algorithms are limited in their ability to preserve accuracy compared\nto models trained on the full data, especially in high pruning regimes. In this\npaper we explore the application of data pruning while incorporating knowledge\ndistillation (KD) when training on a pruned subset. That is, rather than\nrelying solely on ground-truth labels, we also use the soft predictions from a\nteacher network pre-trained on the complete data. By integrating KD into\ntraining, we demonstrate significant improvement across datasets, pruning\nmethods, and on all pruning fractions. We first establish a theoretical\nmotivation for employing self-distillation to improve training on pruned data.\nThen, we empirically make a compelling and highly practical observation: using\nKD, simple random pruning is comparable or superior to sophisticated pruning\nmethods across all pruning regimes. On ImageNet for example, we achieve\nsuperior accuracy despite training on a random subset of only 50% of the data.\nAdditionally, we demonstrate a crucial connection between the pruning factor\nand the optimal knowledge distillation weight. This helps mitigate the impact\nof samples with noisy labels and low-quality images retained by typical pruning\nalgorithms. Finally, we make an intriguing observation: when using lower\npruning fractions, larger teachers lead to accuracy degradation, while\nsurprisingly, employing teachers with a smaller capacity than the student's may\nimprove results. Our code will be made available.\n", "date": "2024-03-12", "categories": []}
{"id": "2307.09165", "title": "Towards Trustworthy Dataset Distillation", "abstract": "  Efficiency and trustworthiness are two eternal pursuits when applying deep\nlearning in real-world applications. With regard to efficiency, dataset\ndistillation (DD) endeavors to reduce training costs by distilling the large\ndataset into a tiny synthetic dataset. However, existing methods merely\nconcentrate on in-distribution (InD) classification in a closed-world setting,\ndisregarding out-of-distribution (OOD) samples. On the other hand, OOD\ndetection aims to enhance models' trustworthiness, which is always\ninefficiently achieved in full-data settings. For the first time, we\nsimultaneously consider both issues and propose a novel paradigm called\nTrustworthy Dataset Distillation (TrustDD). By distilling both InD samples and\noutliers, the condensed datasets are capable of training models competent in\nboth InD classification and OOD detection. To alleviate the requirement of real\noutlier data, we further propose to corrupt InD samples to generate\npseudo-outliers, namely Pseudo-Outlier Exposure (POE). Comprehensive\nexperiments on various settings demonstrate the effectiveness of TrustDD, and\nPOE surpasses the state-of-the-art method Outlier Exposure (OE). Compared with\nthe preceding DD, TrustDD is more trustworthy and applicable to open-world\nscenarios. Our code is available at https://github.com/mashijie1028/TrustDD\n", "date": "2023-07-18", "categories": []}
{"id": "2408.02266", "title": "One-Shot Collaborative Data Distillation", "abstract": "  Large machine-learning training datasets can be distilled into small\ncollections of informative synthetic data samples. These synthetic sets support\nefficient model learning and reduce the communication cost of data sharing.\nThus, high-fidelity distilled data can support the efficient deployment of\nmachine learning applications in distributed network environments. A naive way\nto construct a synthetic set in a distributed environment is to allow each\nclient to perform local data distillation and to merge local distillations at a\ncentral server. However, the quality of the resulting set is impaired by\nheterogeneity in the distributions of the local data held by clients. To\novercome this challenge, we introduce the first collaborative data distillation\ntechnique, called CollabDM, which captures the global distribution of the data\nand requires only a single round of communication between client and server.\nOur method outperforms the state-of-the-art one-shot learning method on skewed\ndata in distributed learning environments. We also show the promising practical\nbenefits of our method when applied to attack detection in 5G networks.\n", "date": "2024-08-05", "categories": []}
{"id": "2305.18381", "title": "Distill Gold from Massive Ores: Bi-level Data Pruning towards Efficient\n  Dataset Distillation", "abstract": "  Data-efficient learning has garnered significant attention, especially given\nthe current trend of large multi-modal models. Recently, dataset distillation\nhas become an effective approach by synthesizing data samples that are\nessential for network training. However, it remains to be explored which\nsamples are essential for the dataset distillation process itself. In this\nwork, we study the data efficiency and selection for the dataset distillation\ntask. By re-formulating the dynamics of distillation, we provide insight into\nthe inherent redundancy in the real dataset, both theoretically and\nempirically. We propose to use the empirical loss value as a static data\npruning criterion. To further compensate for the variation of the data value in\ntraining, we find the most contributing samples based on their causal effects\non the distillation. The proposed selection strategy can efficiently exploit\nthe training dataset, outperform the previous SOTA distillation algorithms, and\nconsistently enhance the distillation algorithms, even on much larger-scale and\nmore heterogeneous datasets, e.g., full ImageNet-1K and Kinetics-400. We\nbelieve this paradigm will open up new avenues in the dynamics of distillation\nand pave the way for efficient dataset distillation. Our code is available on\nhttps://github.com/silicx/GoldFromOres-BiLP.\n", "date": "2023-05-28", "categories": []}
{"id": "2408.02714", "title": "MDM: Advancing Multi-Domain Distribution Matching for Automatic\n  Modulation Recognition Dataset Synthesis", "abstract": "  Recently, deep learning technology has been successfully introduced into\nAutomatic Modulation Recognition (AMR) tasks. However, the success of deep\nlearning is all attributed to the training on large-scale datasets. Such a\nlarge amount of data brings huge pressure on storage, transmission and model\ntraining. In order to solve the problem of large amount of data, some\nresearchers put forward the method of data distillation, which aims to compress\nlarge training data into smaller synthetic datasets to maintain its\nperformance. While numerous data distillation techniques have been developed\nwithin the realm of image processing, the unique characteristics of signals set\nthem apart. Signals exhibit distinct features across various domains,\nnecessitating specialized approaches for their analysis and processing. To this\nend, a novel dataset distillation method--Multi-domain Distribution Matching\n(MDM) is proposed. MDM employs the Discrete Fourier Transform (DFT) to\ntranslate timedomain signals into the frequency domain, and then uses a model\nto compute distribution matching losses between the synthetic and real\ndatasets, considering both the time and frequency domains. Ultimately, these\ntwo losses are integrated to update the synthetic dataset. We conduct extensive\nexperiments on three AMR datasets. Experimental results show that, compared\nwith baseline methods, our method achieves better performance under the same\ncompression ratio. Furthermore, we conduct crossarchitecture generalization\nexperiments on several models, and the experimental results show that our\nsynthetic datasets can generalize well on other unseen models.\n", "date": "2024-08-05", "categories": []}
{"id": "2408.02891", "title": "Diverse Generation while Maintaining Semantic Coordination: A\n  Diffusion-Based Data Augmentation Method for Object Detection", "abstract": "  Recent studies emphasize the crucial role of data augmentation in enhancing\nthe performance of object detection models. However,existing methodologies\noften struggle to effectively harmonize dataset diversity with semantic\ncoordination.To bridge this gap, we introduce an innovative augmentation\ntechnique leveraging pre-trained conditional diffusion models to mediate this\nbalance. Our approach encompasses the development of a Category Affinity\nMatrix, meticulously designed to enhance dataset diversity, and a Surrounding\nRegion Alignment strategy, which ensures the preservation of semantic\ncoordination in the augmented images. Extensive experimental evaluations\nconfirm the efficacy of our method in enriching dataset diversity while\nseamlessly maintaining semantic coordination. Our method yields substantial\naverage improvements of +1.4AP, +0.9AP, and +3.4AP over existing alternatives\non three distinct object detection models, respectively.\n", "date": "2024-08-06", "categories": []}
{"id": "2408.02919", "title": "Data Checklist: On Unit-Testing Datasets with Usable Information", "abstract": "  Model checklists (Ribeiro et al., 2020) have emerged as a useful tool for\nunderstanding the behavior of LLMs, analogous to unit-testing in software\nengineering. However, despite datasets being a key determinant of model\nbehavior, evaluating datasets, e.g., for the existence of annotation artifacts,\nis largely done ad hoc, once a problem in model behavior has already been found\ndownstream. In this work, we take a more principled approach to unit-testing\ndatasets by proposing a taxonomy based on the V-information literature. We call\na collection of such unit tests a data checklist. Using a checklist, not only\nare we able to recover known artifacts in well-known datasets such as SNLI, but\nwe also discover previously unknown artifacts in preference datasets for LLM\nalignment. Data checklists further enable a new kind of data filtering, which\nwe use to improve the efficacy and data efficiency of preference alignment.\n", "date": "2024-08-06", "categories": []}
{"id": "2402.16827", "title": "A Survey on Data Selection for Language Models", "abstract": "  A major factor in the recent success of large language models is the use of\nenormous and ever-growing text datasets for unsupervised pre-training. However,\nnaively training a model on all available data may not be optimal (or\nfeasible), as the quality of available text data can vary. Filtering out data\ncan also decrease the carbon footprint and financial costs of training models\nby reducing the amount of training required. Data selection methods aim to\ndetermine which candidate data points to include in the training dataset and\nhow to appropriately sample from the selected data points. The promise of\nimproved data selection methods has caused the volume of research in the area\nto rapidly expand. However, because deep learning is mostly driven by empirical\nevidence and experimentation on large-scale data is expensive, few\norganizations have the resources for extensive data selection research.\nConsequently, knowledge of effective data selection practices has become\nconcentrated within a few organizations, many of which do not openly share\ntheir findings and methodologies. To narrow this gap in knowledge, we present a\ncomprehensive review of existing literature on data selection methods and\nrelated research areas, providing a taxonomy of existing approaches. By\ndescribing the current landscape of research, this work aims to accelerate\nprogress in data selection by establishing an entry point for new and\nestablished researchers. Additionally, throughout this review we draw attention\nto noticeable holes in the literature and conclude the paper by proposing\npromising avenues for future research.\n", "date": "2024-02-26", "categories": []}
{"id": "2407.15724", "title": "Beyond Size and Class Balance: Alpha as a New Dataset Quality Metric for\n  Deep Learning", "abstract": "  In deep learning, achieving high performance on image classification tasks\nrequires diverse training sets. However, the current best\npractice$\\unicode{x2013}$maximizing dataset size and class\nbalance$\\unicode{x2013}$does not guarantee dataset diversity. We hypothesized\nthat, for a given model architecture, model performance can be improved by\nmaximizing diversity more directly. To test this hypothesis, we introduce a\ncomprehensive framework of diversity measures from ecology that generalizes\nfamiliar quantities like Shannon entropy by accounting for similarities among\nimages. (Size and class balance emerge as special cases.) Analyzing thousands\nof subsets from seven medical datasets showed that the best correlates of\nperformance were not size or class balance but $A$$\\unicode{x2013}$\"big\nalpha\"$\\unicode{x2013}$a set of generalized entropy measures interpreted as the\neffective number of image-class pairs in the dataset, after accounting for\nimage similarities. One of these, $A_0$, explained 67% of the variance in\nbalanced accuracy, vs. 54% for class balance and just 39% for size. The best\npair of measures was size-plus-$A_1$ (79%), which outperformed\nsize-plus-class-balance (74%). Subsets with the largest $A_0$ performed up to\n16% better than those with the largest size (median improvement, 8%). We\npropose maximizing $A$ as a way to improve deep learning performance in medical\nimaging.\n", "date": "2024-07-22", "categories": []}
{"id": "2306.11670", "title": "GIO: Gradient Information Optimization for Training Dataset Selection", "abstract": "  It is often advantageous to train models on a subset of the available train\nexamples, because the examples are of variable quality or because one would\nlike to train with fewer examples, without sacrificing performance. We present\nGradient Information Optimization (GIO), a scalable, task-agnostic approach to\nthis data selection problem that requires only a small set of (unlabeled)\nexamples representing a target distribution. GIO begins from a natural,\ninformation-theoretic objective that is intractable in practice. Our\ncontribution is in showing that it can be made highly scalable through a simple\nrelaxation of the objective and a highly efficient implementation. In\nexperiments with machine translation, spelling correction, and image\nrecognition, we show that GIO delivers outstanding results with very small\ntrain sets. These findings are robust to different representation models and\nhyperparameters for GIO itself. GIO is task- and domain-agnostic and can be\napplied out-of-the-box to new datasets and domains. We open source a\npip-installable implementation of the algorithm as \"pip install grad-info-opt\".\n", "date": "2023-06-20", "categories": []}
{"id": "2406.01805", "title": "TabMDA: Tabular Manifold Data Augmentation for Any Classifier using\n  Transformers with In-context Subsetting", "abstract": "  Tabular data is prevalent in many critical domains, yet it is often\nchallenging to acquire in large quantities. This scarcity usually results in\npoor performance of machine learning models on such data. Data augmentation, a\ncommon strategy for performance improvement in vision and language tasks,\ntypically underperforms for tabular data due to the lack of explicit symmetries\nin the input space. To overcome this challenge, we introduce TabMDA, a novel\nmethod for manifold data augmentation on tabular data. This method utilises a\npre-trained in-context model, such as TabPFN, to map the data into an embedding\nspace. TabMDA performs label-invariant transformations by encoding the data\nmultiple times with varied contexts. This process explores the learned\nembedding space of the underlying in-context models, thereby enlarging the\ntraining dataset. TabMDA is a training-free method, making it applicable to any\nclassifier. We evaluate TabMDA on five standard classifiers and observe\nsignificant performance improvements across various tabular datasets. Our\nresults demonstrate that TabMDA provides an effective way to leverage\ninformation from pre-trained in-context models to enhance the performance of\ndownstream classifiers. Code is available at\nhttps://github.com/AdrianBZG/TabMDA.\n", "date": "2024-06-03", "categories": []}
{"id": "2407.19180", "title": "Data Processing Techniques for Modern Multimodal Models", "abstract": "  Data processing plays an significant role in current multimodal model\ntraining. In this paper. we provide an comprehensive review of common data\nprocessing techniques used in modern multimodal model training with a focus on\ndiffusion models and multimodal large language models (MLLMs). We summarized\nall techniques into four categories: data quality, data quantity, data\ndistribution and data safety. We further present our findings in the choice of\ndata process methods in different type of models. This study aims to provide\nguidance to multimodal models developers with effective data processing\ntechniques.\n", "date": "2024-07-27", "categories": []}
{"id": "2407.19373", "title": "Uncertainty Quantification of Data Shapley via Statistical Inference", "abstract": "  As data plays an increasingly pivotal role in decision-making, the emergence\nof data markets underscores the growing importance of data valuation. Within\nthe machine learning landscape, Data Shapley stands out as a widely embraced\nmethod for data valuation. However, a limitation of Data Shapley is its\nassumption of a fixed dataset, contrasting with the dynamic nature of\nreal-world applications where data constantly evolves and expands. This paper\nestablishes the relationship between Data Shapley and infinite-order\nU-statistics and addresses this limitation by quantifying the uncertainty of\nData Shapley with changes in data distribution from the perspective of\nU-statistics. We make statistical inferences on data valuation to obtain\nconfidence intervals for the estimations. We construct two different algorithms\nto estimate this uncertainty and provide recommendations for their applicable\nsituations. We also conduct a series of experiments on various datasets to\nverify asymptotic normality and propose a practical trading scenario enabled by\nthis method.\n", "date": "2024-07-28", "categories": []}
{"id": "2402.05280", "title": "No Dimensional Sampling Coresets for Classification", "abstract": "  We refine and generalize what is known about coresets for classification\nproblems via the sensitivity sampling framework. Such coresets seek the\nsmallest possible subsets of input data, so one can optimize a loss function on\nthe coreset and ensure approximation guarantees with respect to the original\ndata. Our analysis provides the first no dimensional coresets, so the size does\nnot depend on the dimension. Moreover, our results are general, apply for\ndistributional input and can use iid samples, so provide sample complexity\nbounds, and work for a variety of loss functions. A key tool we develop is a\nRadamacher complexity version of the main sensitivity sampling approach, which\ncan be of independent interest.\n", "date": "2024-02-07", "categories": []}
{"id": "2405.18347", "title": "Dataset Growth", "abstract": "  Deep learning benefits from the growing abundance of available data.\nMeanwhile, efficiently dealing with the growing data scale has become a\nchallenge. Data publicly available are from different sources with various\nqualities, and it is impractical to do manual cleaning against noise and\nredundancy given today's data scale. There are existing techniques for\ncleaning/selecting the collected data. However, these methods are mainly\nproposed for offline settings that target one of the cleanness and redundancy\nproblems. In practice, data are growing exponentially with both problems. This\nleads to repeated data curation with sub-optimal efficiency. To tackle this\nchallenge, we propose InfoGrowth, an efficient online algorithm for data\ncleaning and selection, resulting in a growing dataset that keeps up to date\nwith awareness of cleanliness and diversity. InfoGrowth can improve data\nquality/efficiency on both single-modal and multi-modal tasks, with an\nefficient and scalable design. Its framework makes it practical for real-world\ndata engines.\n", "date": "2024-05-28", "categories": []}
{"id": "2406.04284", "title": "What is Dataset Distillation Learning?", "abstract": "  Dataset distillation has emerged as a strategy to overcome the hurdles\nassociated with large datasets by learning a compact set of synthetic data that\nretains essential information from the original dataset. While distilled data\ncan be used to train high performing models, little is understood about how the\ninformation is stored. In this study, we posit and answer three questions about\nthe behavior, representativeness, and point-wise information content of\ndistilled data. We reveal distilled data cannot serve as a substitute for real\ndata during training outside the standard evaluation setting for dataset\ndistillation. Additionally, the distillation process retains high task\nperformance by compressing information related to the early training dynamics\nof real models. Finally, we provide an framework for interpreting distilled\ndata and reveal that individual distilled data points contain meaningful\nsemantic information. This investigation sheds light on the intricate nature of\ndistilled data, providing a better understanding on how they can be effectively\nutilized.\n", "date": "2024-06-06", "categories": []}
{"id": "2407.10446", "title": "DDFAD: Dataset Distillation Framework for Audio Data", "abstract": "  Deep neural networks (DNNs) have achieved significant success in numerous\napplications. The remarkable performance of DNNs is largely attributed to the\navailability of massive, high-quality training datasets. However, processing\nsuch massive training data requires huge computational and storage resources.\nDataset distillation is a promising solution to this problem, offering the\ncapability to compress a large dataset into a smaller distilled dataset. The\nmodel trained on the distilled dataset can achieve comparable performance to\nthe model trained on the whole dataset.\n  While dataset distillation has been demonstrated in image data, none have\nexplored dataset distillation for audio data. In this work, for the first time,\nwe propose a Dataset Distillation Framework for Audio Data (DDFAD).\nSpecifically, we first propose the Fused Differential MFCC (FD-MFCC) as\nextracted features for audio data. After that, the FD-MFCC is distilled through\nthe matching training trajectory distillation method. Finally, we propose an\naudio signal reconstruction algorithm based on the Griffin-Lim Algorithm to\nreconstruct the audio signal from the distilled FD-MFCC. Extensive experiments\ndemonstrate the effectiveness of DDFAD on various audio datasets. In addition,\nwe show that DDFAD has promising application prospects in many applications,\nsuch as continual learning and neural architecture search.\n", "date": "2024-07-15", "categories": []}
{"id": "2407.15138", "title": "D$^4$M: Dataset Distillation via Disentangled Diffusion Model", "abstract": "  Dataset distillation offers a lightweight synthetic dataset for fast network\ntraining with promising test accuracy. To imitate the performance of the\noriginal dataset, most approaches employ bi-level optimization and the\ndistillation space relies on the matching architecture. Nevertheless, these\napproaches either suffer significant computational costs on large-scale\ndatasets or experience performance decline on cross-architectures. We advocate\nfor designing an economical dataset distillation framework that is independent\nof the matching architectures. With empirical observations, we argue that\nconstraining the consistency of the real and synthetic image spaces will\nenhance the cross-architecture generalization. Motivated by this, we introduce\nDataset Distillation via Disentangled Diffusion Model (D$^4$M), an efficient\nframework for dataset distillation. Compared to architecture-dependent methods,\nD$^4$M employs latent diffusion model to guarantee consistency and incorporates\nlabel information into category prototypes. The distilled datasets are\nversatile, eliminating the need for repeated generation of distinct datasets\nfor various architectures. Through comprehensive experiments, D$^4$M\ndemonstrates superior performance and robust generalization, surpassing the\nSOTA methods across most aspects.\n", "date": "2024-07-21", "categories": []}
{"id": "2407.15235", "title": "TAGCOS: Task-agnostic Gradient Clustered Coreset Selection for\n  Instruction Tuning Data", "abstract": "  Instruction tuning has achieved unprecedented success in NLP, turning large\nlanguage models into versatile chatbots. However, the increasing variety and\nvolume of instruction datasets demand significant computational resources. To\naddress this, it is essential to extract a small and highly informative subset\n(i.e., Coreset) that achieves comparable performance to the full dataset.\nAchieving this goal poses non-trivial challenges: 1) data selection requires\naccurate data representations that reflect the training samples' quality, 2)\nconsidering the diverse nature of instruction datasets, and 3) ensuring the\nefficiency of the coreset selection algorithm for large models. To address\nthese challenges, we propose Task-Agnostic Gradient Clustered COreset Selection\n(TAGCOS). Specifically, we leverage sample gradients as the data\nrepresentations, perform clustering to group similar data, and apply an\nefficient greedy algorithm for coreset selection. Experimental results show\nthat our algorithm, selecting only 5% of the data, surpasses other unsupervised\nmethods and achieves performance close to that of the full dataset.\n", "date": "2024-07-21", "categories": []}
{"id": "2310.07506", "title": "Leveraging Hierarchical Feature Sharing for Efficient Dataset\n  Condensation", "abstract": "  Given a real-world dataset, data condensation (DC) aims to synthesize a small\nsynthetic dataset that captures the knowledge of a natural dataset while being\nusable for training models with comparable accuracy. Recent works propose to\nenhance DC with data parameterization, which condenses data into very compact\nparameterized data containers instead of images. The intuition behind data\nparameterization is to encode shared features of images to avoid additional\nstorage costs. In this paper, we recognize that images share common features in\na hierarchical way due to the inherent hierarchical structure of the\nclassification system, which is overlooked by current data parameterization\nmethods. To better align DC with this hierarchical nature and encourage more\nefficient information sharing inside data containers, we propose a novel data\nparameterization architecture, Hierarchical Memory Network (HMN). HMN stores\ncondensed data in a three-tier structure, representing the dataset-level,\nclass-level, and instance-level features. Another helpful property of the\nhierarchical architecture is that HMN naturally ensures good independence among\nimages despite achieving information sharing. This enables instance-level\npruning for HMN to reduce redundant information, thereby further minimizing\nredundancy and enhancing performance. We evaluate HMN on five public datasets\nand show that our proposed method outperforms all baselines.\n", "date": "2023-10-11", "categories": []}
{"id": "2407.14245", "title": "Dataset Distillation by Automatic Training Trajectories", "abstract": "  Dataset Distillation is used to create a concise, yet informative, synthetic\ndataset that can replace the original dataset for training purposes. Some\nleading methods in this domain prioritize long-range matching, involving the\nunrolling of training trajectories with a fixed number of steps (NS) on the\nsynthetic dataset to align with various expert training trajectories. However,\ntraditional long-range matching methods possess an overfitting-like problem,\nthe fixed step size NS forces synthetic dataset to distortedly conform seen\nexpert training trajectories, resulting in a loss of generality-especially to\nthose from unencountered architecture. We refer to this as the Accumulated\nMismatching Problem (AMP), and propose a new approach, Automatic Training\nTrajectories (ATT), which dynamically and adaptively adjusts trajectory length\nNS to address the AMP. Our method outperforms existing methods particularly in\ntests involving cross-architectures. Moreover, owing to its adaptive nature, it\nexhibits enhanced stability in the face of parameter variations.\n", "date": "2024-07-19", "categories": []}
{"id": "2402.09739", "title": "QuRating: Selecting High-Quality Data for Training Language Models", "abstract": "  Selecting high-quality pre-training data is important for creating capable\nlanguage models, but existing methods rely on simple heuristics. We introduce\nQuRating, a method for selecting pre-training data that can capture human\nintuitions about data quality. In this paper, we investigate four qualities -\nwriting style, required expertise, facts & trivia, and educational value - and\nfind that LLMs are able to discern these qualities, especially when making\npairwise judgments of texts. We train a QuRater model to learn scalar ratings\nfrom pairwise judgments, and use it to annotate a 260B training corpus with\nquality ratings for each of the four criteria. In our experiments, we select\n30B tokens according to the different quality ratings and train 1.3B-parameter\nlanguage models on the selected data. We find that it is important to balance\nquality and diversity. When we sample using quality ratings as logits over\ndocuments, our models obtain lower perplexity and stronger in-context learning\nperformance than baselines. Our best model is based on educational value and\nperforms similarly to a model trained with uniform sampling for 50% more steps.\nBeyond data selection, we use the quality ratings to construct a training\ncurriculum which improves performance without changing the training dataset. We\nextensively analyze the quality ratings and discuss their characteristics,\nbiases, and wider implications.\n", "date": "2024-02-15", "categories": []}
{"id": "2407.12793", "title": "Data Collection and Labeling Techniques for Machine Learning", "abstract": "  Data collection and labeling are critical bottlenecks in the deployment of\nmachine learning applications. With the increasing complexity and diversity of\napplications, the need for efficient and scalable data collection and labeling\ntechniques has become paramount. This paper provides a review of the\nstate-of-the-art methods in data collection, data labeling, and the improvement\nof existing data and models. By integrating perspectives from both the machine\nlearning and data management communities, we aim to provide a holistic view of\nthe current landscape and identify future research directions.\n", "date": "2024-06-19", "categories": []}
{"id": "2404.17507", "title": "HYPE: Hyperbolic Entailment Filtering for Underspecified Images and\n  Texts", "abstract": "  In an era where the volume of data drives the effectiveness of\nself-supervised learning, the specificity and clarity of data semantics play a\ncrucial role in model training. Addressing this, we introduce HYPerbolic\nEntailment filtering (HYPE), a novel methodology designed to meticulously\nextract modality-wise meaningful and well-aligned data from extensive, noisy\nimage-text pair datasets. Our approach leverages hyperbolic embeddings and the\nconcept of entailment cones to evaluate and filter out samples with meaningless\nor underspecified semantics, focusing on enhancing the specificity of each data\nsample. HYPE not only demonstrates a significant improvement in filtering\nefficiency but also sets a new state-of-the-art in the DataComp benchmark when\ncombined with existing filtering techniques. This breakthrough showcases the\npotential of HYPE to refine the data selection process, thereby contributing to\nthe development of more accurate and efficient self-supervised learning models.\nAdditionally, the image specificity $\\epsilon_{i}$ can be independently applied\nto induce an image-only dataset from an image-text or image-only data pool for\ntraining image-only self-supervised models and showed superior performance when\ncompared to the dataset induced by CLIP score.\n", "date": "2024-04-26", "categories": []}
{"id": "2407.10910", "title": "DataDream: Few-shot Guided Dataset Generation", "abstract": "  While text-to-image diffusion models have been shown to achieve\nstate-of-the-art results in image synthesis, they have yet to prove their\neffectiveness in downstream applications. Previous work has proposed to\ngenerate data for image classifier training given limited real data access.\nHowever, these methods struggle to generate in-distribution images or depict\nfine-grained features, thereby hindering the generalization of classification\nmodels trained on synthetic datasets. We propose DataDream, a framework for\nsynthesizing classification datasets that more faithfully represents the real\ndata distribution when guided by few-shot examples of the target classes.\nDataDream fine-tunes LoRA weights for the image generation model on the few\nreal images before generating the training data using the adapted model. We\nthen fine-tune LoRA weights for CLIP using the synthetic data to improve\ndownstream image classification over previous approaches on a large variety of\ndatasets. We demonstrate the efficacy of DataDream through extensive\nexperiments, surpassing state-of-the-art classification accuracy with few-shot\ndata across 7 out of 10 datasets, while being competitive on the other 3.\nAdditionally, we provide insights into the impact of various factors, such as\nthe number of real-shot and generated images as well as the fine-tuning compute\non model performance. The code is available at\nhttps://github.com/ExplainableML/DataDream.\n", "date": "2024-07-15", "categories": []}
{"id": "2406.08418", "title": "OmniCorpus: A Unified Multimodal Corpus of 10 Billion-Level Images\n  Interleaved with Text", "abstract": "  Image-text interleaved data, consisting of multiple images and texts arranged\nin a natural document format, aligns with the presentation paradigm of internet\ndata and closely resembles human reading habits. Recent studies have shown that\nsuch data aids multimodal in-context learning and maintains the capabilities of\nlarge language models during multimodal fine-tuning. However, the limited scale\nand diversity of current image-text interleaved data restrict the development\nof multimodal large language models. In this paper, we introduce OmniCorpus, a\n10 billion-scale image-text interleaved dataset. Using an efficient data\nengine, we filter and extract large-scale high-quality documents, which contain\n8.6 billion images and 1,696 billion text tokens. Compared to counterparts\n(e.g., MMC4, OBELICS), our dataset 1) has 15 times larger scales while\nmaintaining good data quality; 2) features more diverse sources, including both\nEnglish and non-English websites as well as video-centric websites; 3) is more\nflexible, easily degradable from an image-text interleaved format to pure text\ncorpus and image-text pairs. Through comprehensive analysis and experiments, we\nvalidate the quality, usability, and effectiveness of the proposed dataset. We\nhope this could provide a solid data foundation for future multimodal model\nresearch. Code and data are released at\nhttps://github.com/OpenGVLab/OmniCorpus.\n", "date": "2024-06-12", "categories": []}
{"id": "2407.06645", "title": "Entropy Law: The Story Behind Data Compression and LLM Performance", "abstract": "  Data is the cornerstone of large language models (LLMs), but not all data is\nuseful for model learning. Carefully selected data can better elicit the\ncapabilities of LLMs with much less computational overhead. Most methods\nconcentrate on evaluating the quality of individual samples in data selection,\nwhile the combinatorial effects among samples are neglected. Even if each\nsample is of perfect quality, their combinations may be suboptimal in teaching\nLLMs due to their intrinsic homogeneity or contradiction. In this paper, we aim\nto uncover the underlying relationships between LLM performance and data\nselection. Inspired by the information compression nature of LLMs, we uncover\nan ``entropy law'' that connects LLM performance with data compression ratio\nand first-epoch training loss, which reflect the information redundancy of a\ndataset and the mastery of inherent knowledge encoded in this dataset,\nrespectively. Through both theoretical deduction and empirical evaluation, we\nfind that model performance is negatively correlated to the compression ratio\nof training data, which usually yields a lower training loss. Based on the\nfindings of the entropy law, we propose a quite efficient and universal data\nselection method named \\textbf{ZIP} for training LLMs, which aim to prioritize\ndata subsets exhibiting a low compression ratio. Based on a multi-stage\nalgorithm that selects diverse data in a greedy manner, we can obtain a good\ndata subset with satisfactory diversity. Extensive experiments have been\nconducted to validate the entropy law and the superiority of ZIP across\ndifferent LLM backbones and alignment stages. We also present an interesting\napplication of entropy law that can detect potential performance risks at the\nbeginning of model training.\n", "date": "2024-07-09", "categories": []}
{"id": "2407.08113", "title": "FYI: Flip Your Images for Dataset Distillation", "abstract": "  Dataset distillation synthesizes a small set of images from a large-scale\nreal dataset such that synthetic and real images share similar behavioral\nproperties (e.g, distributions of gradients or features) during a training\nprocess. Through extensive analyses on current methods and real datasets,\ntogether with empirical observations, we provide in this paper two important\nthings to share for dataset distillation. First, object parts that appear on\none side of a real image are highly likely to appear on the opposite side of\nanother image within a dataset, which we call the bilateral equivalence.\nSecond, the bilateral equivalence enforces synthetic images to duplicate\ndiscriminative parts of objects on both the left and right sides of the images,\nlimiting the recognition of subtle differences between objects. To address this\nproblem, we introduce a surprisingly simple yet effective technique for dataset\ndistillation, dubbed FYI, that enables distilling rich semantics of real images\ninto synthetic ones. To this end, FYI embeds a horizontal flipping technique\ninto distillation processes, mitigating the influence of the bilateral\nequivalence, while capturing more details of objects. Experiments on\nCIFAR-10/100, Tiny-ImageNet, and ImageNet demonstrate that FYI can be\nseamlessly integrated into several state-of-the-art methods, without modifying\ntraining objectives and network architectures, and it improves the performance\nremarkably.\n", "date": "2024-07-11", "categories": []}
{"id": "2406.06609", "title": "Mitigating Bias in Dataset Distillation", "abstract": "  Dataset Distillation has emerged as a technique for compressing large\ndatasets into smaller synthetic counterparts, facilitating downstream training\ntasks. In this paper, we study the impact of bias inside the original dataset\non the performance of dataset distillation. With a comprehensive empirical\nevaluation on canonical datasets with color, corruption and background biases,\nwe found that color and background biases in the original dataset will be\namplified through the distillation process, resulting in a notable decline in\nthe performance of models trained on the distilled dataset, while corruption\nbias is suppressed through the distillation process. To reduce bias\namplification in dataset distillation, we introduce a simple yet highly\neffective approach based on a sample reweighting scheme utilizing kernel\ndensity estimation. Empirical results on multiple real-world and synthetic\ndatasets demonstrate the effectiveness of the proposed method. Notably, on\nCMNIST with 5% bias-conflict ratio and IPC 50, our method achieves 91.5% test\naccuracy compared to 23.8% from vanilla DM, boosting the performance by 67.7%,\nwhereas applying state-of-the-art debiasing method on the same dataset only\nachieves 53.7% accuracy. Our findings highlight the importance of addressing\nbiases in dataset distillation and provide a promising avenue to address bias\namplification in the process.\n", "date": "2024-06-06", "categories": []}
{"id": "2406.13536", "title": "Image Distillation for Safe Data Sharing in Histopathology", "abstract": "  Histopathology can help clinicians make accurate diagnoses, determine disease\nprognosis, and plan appropriate treatment strategies. As deep learning\ntechniques prove successful in the medical domain, the primary challenges\nbecome limited data availability and concerns about data sharing and privacy.\nFederated learning has addressed this challenge by training models locally and\nupdating parameters on a server. However, issues, such as domain shift and\nbias, persist and impact overall performance. Dataset distillation presents an\nalternative approach to overcoming these challenges. It involves creating a\nsmall synthetic dataset that encapsulates essential information, which can be\nshared without constraints. At present, this paradigm is not practicable as\ncurrent distillation approaches only generate non human readable\nrepresentations and exhibit insufficient performance for downstream learning\ntasks. We train a latent diffusion model and construct a new distilled\nsynthetic dataset with a small number of human readable synthetic images.\nSelection of maximally informative synthetic images is done via graph community\nanalysis of the representation space. We compare downstream classification\nmodels trained on our synthetic distillation data to models trained on real\ndata and reach performances suitable for practical application.\n", "date": "2024-06-19", "categories": []}
{"id": "2407.07268", "title": "Dataset Quantization with Active Learning based Adaptive Sampling", "abstract": "  Deep learning has made remarkable progress recently, largely due to the\navailability of large, well-labeled datasets. However, the training on such\ndatasets elevates costs and computational demands. To address this, various\ntechniques like coreset selection, dataset distillation, and dataset\nquantization have been explored in the literature. Unlike traditional\ntechniques that depend on uniform sample distributions across different\nclasses, our research demonstrates that maintaining performance is feasible\neven with uneven distributions. We find that for certain classes, the variation\nin sample quantity has a minimal impact on performance. Inspired by this\nobservation, an intuitive idea is to reduce the number of samples for stable\nclasses and increase the number of samples for sensitive classes to achieve a\nbetter performance with the same sampling ratio. Then the question arises: how\ncan we adaptively select samples from a dataset to achieve optimal performance?\nIn this paper, we propose a novel active learning based adaptive sampling\nstrategy, Dataset Quantization with Active Learning based Adaptive Sampling\n(DQAS), to optimize the sample selection. In addition, we introduce a novel\npipeline for dataset quantization, utilizing feature space from the final stage\nof dataset quantization to generate more precise dataset bins. Our\ncomprehensive evaluations on the multiple datasets show that our approach\noutperforms the state-of-the-art dataset compression methods.\n", "date": "2024-07-09", "categories": []}
{"id": "2402.09288", "title": "EcoVal: An Efficient Data Valuation Framework for Machine Learning", "abstract": "  Quantifying the value of data within a machine learning workflow can play a\npivotal role in making more strategic decisions in machine learning\ninitiatives. The existing Shapley value based frameworks for data valuation in\nmachine learning are computationally expensive as they require considerable\namount of repeated training of the model to obtain the Shapley value. In this\npaper, we introduce an efficient data valuation framework EcoVal, to estimate\nthe value of data for machine learning models in a fast and practical manner.\nInstead of directly working with individual data sample, we determine the value\nof a cluster of similar data points. This value is further propagated amongst\nall the member cluster points. We show that the overall value of the data can\nbe determined by estimating the intrinsic and extrinsic value of each data.\nThis is enabled by formulating the performance of a model as a\n\\textit{production function}, a concept which is popularly used to estimate the\namount of output based on factors like labor and capital in a traditional free\neconomic market. We provide a formal proof of our valuation technique and\nelucidate the principles and mechanisms that enable its accelerated\nperformance. We demonstrate the real-world applicability of our method by\nshowcasing its effectiveness for both in-distribution and out-of-sample data.\nThis work addresses one of the core challenges of efficient data valuation at\nscale in machine learning models. The code is available at\n\\underline{https://github.com/respai-lab/ecoval}.\n", "date": "2024-02-14", "categories": []}
{"id": "2405.12915", "title": "G-DIG: Towards Gradient-based Diverse and High-quality Instruction Data\n  Selection for Machine Translation", "abstract": "  Large Language Models (LLMs) have demonstrated remarkable abilities in\ngeneral scenarios. Instruction finetuning empowers them to align with humans in\nvarious tasks. Nevertheless, the Diversity and Quality of the instruction data\nremain two main challenges for instruction finetuning. With regard to this, in\nthis paper, we propose a novel gradient-based method to automatically select\nhigh-quality and diverse instruction finetuning data for machine translation.\nOur key innovation centers around analyzing how individual training examples\ninfluence the model during training. Specifically, we select training examples\nthat exert beneficial influences on the model as high-quality ones by means of\nInfluence Function plus a small high-quality seed dataset. Moreover, to enhance\nthe diversity of the training data we maximize the variety of influences they\nhave on the model by clustering on their gradients and resampling. Extensive\nexperiments on WMT22 and FLORES translation tasks demonstrate the superiority\nof our methods, and in-depth analysis further validates their effectiveness and\ngeneralization.\n", "date": "2024-05-21", "categories": []}
{"id": "2407.03463", "title": "Precision at Scale: Domain-Specific Datasets On-Demand", "abstract": "  In the realm of self-supervised learning (SSL), conventional wisdom has\ngravitated towards the utility of massive, general domain datasets for\npretraining robust backbones. In this paper, we challenge this idea by\nexploring if it is possible to bridge the scale between general-domain datasets\nand (traditionally smaller) domain-specific datasets to reduce the current\nperformance gap. More specifically, we propose Precision at Scale (PaS), a\nnovel method for the autonomous creation of domain-specific datasets on-demand.\nThe modularity of the PaS pipeline enables leveraging state-of-the-art\nfoundational and generative models to create a collection of images of any\ngiven size belonging to any given domain with minimal human intervention.\nExtensive analysis in two complex domains, proves the superiority of PaS\ndatasets over existing traditional domain-specific datasets in terms of\ndiversity, scale, and effectiveness in training visual transformers and\nconvolutional neural networks. Most notably, we prove that automatically\ngenerated domain-specific datasets lead to better pretraining than large-scale\nsupervised datasets such as ImageNet-1k and ImageNet-21k. Concretely, models\ntrained on domain-specific datasets constructed by PaS pipeline, beat\nImageNet-1k pretrained backbones by at least 12% in all the considered domains\nand classification tasks and lead to better food domain performance than\nsupervised ImageNet-21k pretrain while being 12 times smaller. Code repository:\nhttps://github.com/jesusmolrdv/Precision-at-Scale/\n", "date": "2024-07-03", "categories": []}
{"id": "2407.00407", "title": "SHADE: Semantic Hypernym Annotator for Domain-specific Entities -- DnD\n  Domain Use Case", "abstract": "  Manual data annotation is an important NLP task but one that takes\nconsiderable amount of resources and effort. In spite of the costs, labeling\nand categorizing entities is essential for NLP tasks such as semantic\nevaluation. Even though annotation can be done by non-experts in most cases,\ndue to the fact that this requires human labor, the process is costly. Another\nmajor challenge encountered in data annotation is maintaining the annotation\nconsistency. Annotation efforts are typically carried out by teams of multiple\nannotators. The annotations need to maintain the consistency in relation to\nboth the domain truth and annotation format while reducing human errors.\nAnnotating a specialized domain that deviates significantly from the general\ndomain, such as fantasy literature, will see a lot of human error and annotator\ndisagreement. So it is vital that proper guidelines and error reduction\nmechanisms are enforced. One such way to enforce these constraints is using a\nspecialized application. Such an app can ensure that the notations are\nconsistent, and the labels can be pre-defined or restricted reducing the room\nfor errors. In this paper, we present SHADE, an annotation software that can be\nused to annotate entities in the high fantasy literature domain. Specifically\nin Dungeons and Dragons lore extracted from the Forgotten Realms Fandom Wiki.\n", "date": "2024-06-29", "categories": []}
{"id": "2406.19614", "title": "A Survey on Data Quality Dimensions and Tools for Machine Learning", "abstract": "  Machine learning (ML) technologies have become substantial in practically all\naspects of our society, and data quality (DQ) is critical for the performance,\nfairness, robustness, safety, and scalability of ML models. With the large and\ncomplex data in data-centric AI, traditional methods like exploratory data\nanalysis (EDA) and cross-validation (CV) face challenges, highlighting the\nimportance of mastering DQ tools. In this survey, we review 17 DQ evaluation\nand improvement tools in the last 5 years. By introducing the DQ dimensions,\nmetrics, and main functions embedded in these tools, we compare their strengths\nand limitations and propose a roadmap for developing open-source DQ tools for\nML. Based on the discussions on the challenges and emerging trends, we further\nhighlight the potential applications of large language models (LLMs) and\ngenerative AI in DQ evaluation and improvement for ML. We believe this\ncomprehensive survey can enhance understanding of DQ in ML and could drive\nprogress in data-centric AI. A complete list of the literature investigated in\nthis survey is available on GitHub at:\nhttps://github.com/haihua0913/awesome-dq4ml.\n", "date": "2024-06-28", "categories": []}
{"id": "2406.19827", "title": "Towards Stable and Storage-efficient Dataset Distillation: Matching\n  Convexified Trajectory", "abstract": "  The rapid evolution of deep learning and large language models has led to an\nexponential growth in the demand for training data, prompting the development\nof Dataset Distillation methods to address the challenges of managing large\ndatasets. Among these, Matching Training Trajectories (MTT) has been a\nprominent approach, which replicates the training trajectory of an expert\nnetwork on real data with a synthetic dataset. However, our investigation found\nthat this method suffers from three significant limitations: 1. Instability of\nexpert trajectory generated by Stochastic Gradient Descent (SGD); 2. Low\nconvergence speed of the distillation process; 3. High storage consumption of\nthe expert trajectory. To address these issues, we offer a new perspective on\nunderstanding the essence of Dataset Distillation and MTT through a simple\ntransformation of the objective function, and introduce a novel method called\nMatching Convexified Trajectory (MCT), which aims to provide better guidance\nfor the student trajectory. MCT leverages insights from the linearized dynamics\nof Neural Tangent Kernel methods to create a convex combination of expert\ntrajectories, guiding the student network to converge rapidly and stably. This\ntrajectory is not only easier to store, but also enables a continuous sampling\nstrategy during distillation, ensuring thorough learning and fitting of the\nentire expert trajectory. Comprehensive experiments across three public\ndatasets validate the superiority of MCT over traditional MTT methods.\n", "date": "2024-06-28", "categories": []}
{"id": "2406.18561", "title": "SelMatch: Effectively Scaling Up Dataset Distillation via\n  Selection-Based Initialization and Partial Updates by Trajectory Matching", "abstract": "  Dataset distillation aims to synthesize a small number of images per class\n(IPC) from a large dataset to approximate full dataset training with minimal\nperformance loss. While effective in very small IPC ranges, many distillation\nmethods become less effective, even underperforming random sample selection, as\nIPC increases. Our examination of state-of-the-art trajectory-matching based\ndistillation methods across various IPC scales reveals that these methods\nstruggle to incorporate the complex, rare features of harder samples into the\nsynthetic dataset even with the increased IPC, resulting in a persistent\ncoverage gap between easy and hard test samples. Motivated by such\nobservations, we introduce SelMatch, a novel distillation method that\neffectively scales with IPC. SelMatch uses selection-based initialization and\npartial updates through trajectory matching to manage the synthetic dataset's\ndesired difficulty level tailored to IPC scales. When tested on CIFAR-10/100\nand TinyImageNet, SelMatch consistently outperforms leading selection-only and\ndistillation-only methods across subset ratios from 5% to 30%.\n", "date": "2024-05-28", "categories": []}
{"id": "2312.05598", "title": "Boosting the Cross-Architecture Generalization of Dataset Distillation\n  through an Empirical Study", "abstract": "  The poor cross-architecture generalization of dataset distillation greatly\nweakens its practical significance. This paper attempts to mitigate this issue\nthrough an empirical study, which suggests that the synthetic datasets undergo\nan inductive bias towards the distillation model. Therefore, the evaluation\nmodel is strictly confined to having similar architectures of the distillation\nmodel. We propose a novel method of EvaLuation with distillation Feature (ELF),\nwhich utilizes features from intermediate layers of the distillation model for\nthe cross-architecture evaluation. In this manner, the evaluation model learns\nfrom bias-free knowledge therefore its architecture becomes unfettered while\nretaining performance. By performing extensive experiments, we successfully\nprove that ELF can well enhance the cross-architecture generalization of\ncurrent DD methods. Code of this project is at\n\\url{https://github.com/Lirui-Zhao/ELF}.\n", "date": "2023-12-09", "categories": []}
{"id": "2406.17711", "title": "Data curation via joint example selection further accelerates multimodal\n  learning", "abstract": "  Data curation is an essential component of large-scale pretraining. In this\nwork, we demonstrate that jointly selecting batches of data is more effective\nfor learning than selecting examples independently. Multimodal contrastive\nobjectives expose the dependencies between data and thus naturally yield\ncriteria for measuring the joint learnability of a batch. We derive a simple\nand tractable algorithm for selecting such batches, which significantly\naccelerate training beyond individually-prioritized data points. As performance\nimproves by selecting from larger super-batches, we also leverage recent\nadvances in model approximation to reduce the associated computational\noverhead. As a result, our approach--multimodal contrastive learning with joint\nexample selection (JEST)--surpasses state-of-the-art models with up to\n13$\\times$ fewer iterations and 10$\\times$ less computation. Essential to the\nperformance of JEST is the ability to steer the data selection process towards\nthe distribution of smaller, well-curated datasets via pretrained reference\nmodels, exposing the level of data curation as a new dimension for neural\nscaling laws.\n", "date": "2024-06-25", "categories": []}
{"id": "2406.15042", "title": "Behaviour Distillation", "abstract": "  Dataset distillation aims to condense large datasets into a small number of\nsynthetic examples that can be used as drop-in replacements when training new\nmodels. It has applications to interpretability, neural architecture search,\nprivacy, and continual learning. Despite strong successes in supervised\ndomains, such methods have not yet been extended to reinforcement learning,\nwhere the lack of a fixed dataset renders most distillation methods unusable.\nFilling the gap, we formalize behaviour distillation, a setting that aims to\ndiscover and then condense the information required for training an expert\npolicy into a synthetic dataset of state-action pairs, without access to expert\ndata. We then introduce Hallucinating Datasets with Evolution Strategies\n(HaDES), a method for behaviour distillation that can discover datasets of just\nfour state-action pairs which, under supervised learning, train agents to\ncompetitive performance levels in continuous control tasks. We show that these\ndatasets generalize out of distribution to training policies with a wide range\nof architectures and hyperparameters. We also demonstrate application to a\ndownstream task, namely training multi-task agents in a zero-shot fashion.\nBeyond behaviour distillation, HaDES provides significant improvements in\nneuroevolution for RL over previous approaches and achieves SoTA results on one\nstandard supervised dataset distillation task. Finally, we show that\nvisualizing the synthetic datasets can provide human-interpretable task\ninsights.\n", "date": "2024-06-21", "categories": []}
{"id": "2402.05011", "title": "Navigating Complexity: Toward Lossless Graph Condensation via Expanding\n  Window Matching", "abstract": "  Graph condensation aims to reduce the size of a large-scale graph dataset by\nsynthesizing a compact counterpart without sacrificing the performance of Graph\nNeural Networks (GNNs) trained on it, which has shed light on reducing the\ncomputational cost for training GNNs. Nevertheless, existing methods often fall\nshort of accurately replicating the original graph for certain datasets,\nthereby failing to achieve the objective of lossless condensation. To\nunderstand this phenomenon, we investigate the potential reasons and reveal\nthat the previous state-of-the-art trajectory matching method provides biased\nand restricted supervision signals from the original graph when optimizing the\ncondensed one. This significantly limits both the scale and efficacy of the\ncondensed graph. In this paper, we make the first attempt toward\n\\textit{lossless graph condensation} by bridging the previously neglected\nsupervision signals. Specifically, we employ a curriculum learning strategy to\ntrain expert trajectories with more diverse supervision signals from the\noriginal graph, and then effectively transfer the information into the\ncondensed graph with expanding window matching. Moreover, we design a loss\nfunction to further extract knowledge from the expert trajectories. Theoretical\nanalysis justifies the design of our method and extensive experiments verify\nits superiority across different datasets. Code is released at\nhttps://github.com/NUS-HPC-AI-Lab/GEOM.\n", "date": "2024-02-07", "categories": []}
{"id": "2406.10914", "title": "First-Order Manifold Data Augmentation for Regression Learning", "abstract": "  Data augmentation (DA) methods tailored to specific domains generate\nsynthetic samples by applying transformations that are appropriate for the\ncharacteristics of the underlying data domain, such as rotations on images and\ntime warping on time series data. In contrast, domain-independent approaches,\ne.g. mixup, are applicable to various data modalities, and as such they are\ngeneral and versatile. While regularizing classification tasks via DA is a\nwell-explored research topic, the effect of DA on regression problems received\nless attention. To bridge this gap, we study the problem of domain-independent\naugmentation for regression, and we introduce FOMA: a new data-driven\ndomain-independent data augmentation method. Essentially, our approach samples\nnew examples from the tangent planes of the train distribution. Augmenting data\nin this way aligns with the network tendency towards capturing the dominant\nfeatures of its input signals. We evaluate FOMA on in-distribution\ngeneralization and out-of-distribution robustness benchmarks, and we show that\nit improves the generalization of several neural architectures. We also find\nthat strong baselines based on mixup are less effective in comparison to our\napproach. Our code is publicly available\nathttps://github.com/azencot-group/FOMA.\n", "date": "2024-06-16", "categories": []}
{"id": "2306.05175", "title": "Large-scale Dataset Pruning with Dynamic Uncertainty", "abstract": "  The state of the art of many learning tasks, e.g., image classification, is\nadvanced by collecting larger datasets and then training larger models on them.\nAs the outcome, the increasing computational cost is becoming unaffordable. In\nthis paper, we investigate how to prune the large-scale datasets, and thus\nproduce an informative subset for training sophisticated deep models with\nnegligible performance drop. We propose a simple yet effective dataset pruning\nmethod by exploring both the prediction uncertainty and training dynamics. We\nstudy dataset pruning by measuring the variation of predictions during the\nwhole training process on large-scale datasets, i.e., ImageNet-1K and\nImageNet-21K, and advanced models, i.e., Swin Transformer and ConvNeXt.\nExtensive experimental results indicate that our method outperforms the state\nof the art and achieves 25% lossless pruning ratio on both ImageNet-1K and\nImageNet-21K. The code and pruned datasets are available at\nhttps://github.com/BAAI-DCAI/Dataset-Pruning.\n", "date": "2023-06-08", "categories": []}
{"id": "2406.09860", "title": "Dataset Condensation with Latent Quantile Matching", "abstract": "  Dataset condensation (DC) methods aim to learn a smaller synthesized dataset\nwith informative data records to accelerate the training of machine learning\nmodels. Current distribution matching (DM) based DC methods learn a synthesized\ndataset by matching the mean of the latent embeddings between the synthetic and\nthe real dataset. However two distributions with the same mean can still be\nvastly different. In this work we demonstrate the shortcomings of using Maximum\nMean Discrepancy to match latent distributions i.e. the weak matching power and\nlack of outlier regularization. To alleviate these shortcomings we propose our\nnew method: Latent Quantile Matching (LQM) which matches the quantiles of the\nlatent embeddings to minimize the goodness of fit test statistic between two\ndistributions. Empirical experiments on both image and graph-structured\ndatasets show that LQM matches or outperforms previous state of the art in\ndistribution matching based DC. Moreover we show that LQM improves the\nperformance in continual graph learning (CGL) setting where memory efficiency\nand privacy can be important. Our work sheds light on the application of DM\nbased DC for CGL.\n", "date": "2024-06-14", "categories": []}
{"id": "2402.04333", "title": "LESS: Selecting Influential Data for Targeted Instruction Tuning", "abstract": "  Instruction tuning has unlocked powerful capabilities in large language\nmodels (LLMs), effectively using combined datasets to develop generalpurpose\nchatbots. However, real-world applications often require a specialized suite of\nskills (e.g., reasoning). The challenge lies in identifying the most relevant\ndata from these extensive datasets to effectively develop specific\ncapabilities, a setting we frame as targeted instruction tuning. We propose\nLESS, an optimizer-aware and practically efficient algorithm to effectively\nestimate data influences and perform Low-rank gradiEnt Similarity Search for\ninstruction data selection. Crucially, LESS adapts existing influence\nformulations to work with the Adam optimizer and variable-length instruction\ndata. LESS first constructs a highly reusable and transferable gradient\ndatastore with low-dimensional gradient features and then selects examples\nbased on their similarity to few-shot examples embodying a specific capability.\nExperiments show that training on a LESS-selected 5% of the data can often\noutperform training on the full dataset across diverse downstream tasks.\nFurthermore, the selected data is highly transferable: smaller models can be\nleveraged to select useful data for larger models and models from different\nfamilies. Our qualitative analysis shows that our method goes beyond surface\nform cues to identify data that exemplifies the necessary reasoning skills for\nthe intended downstream application.\n", "date": "2024-02-06", "categories": []}
{"id": "2406.07583", "title": "Situated Ground Truths: Enhancing Bias-Aware AI by Situating Data Labels\n  with SituAnnotate", "abstract": "  In the contemporary world of AI and data-driven applications, supervised\nmachines often derive their understanding, which they mimic and reproduce,\nthrough annotations--typically conveyed in the form of words or labels.\nHowever, such annotations are often divorced from or lack contextual\ninformation, and as such hold the potential to inadvertently introduce biases\nwhen subsequently used for training. This paper introduces SituAnnotate, a\nnovel ontology explicitly crafted for 'situated grounding,' aiming to anchor\nthe ground truth data employed in training AI systems within the contextual and\nculturally-bound situations from which those ground truths emerge. SituAnnotate\noffers an ontology-based approach to structured and context-aware data\nannotation, addressing potential bias issues associated with isolated\nannotations. Its representational power encompasses situational context,\nincluding annotator details, timing, location, remuneration schemes, annotation\nroles, and more, ensuring semantic richness. Aligned with the foundational\nDolce Ultralight ontology, it provides a robust and consistent framework for\nknowledge representation. As a method to create, query, and compare label-based\ndatasets, SituAnnotate empowers downstream AI systems to undergo training with\nexplicit consideration of context and cultural bias, laying the groundwork for\nenhanced system interpretability and adaptability, and enabling AI models to\nalign with a multitude of cultural contexts and viewpoints.\n", "date": "2024-06-10", "categories": []}
{"id": "2406.08249", "title": "Dataset Enhancement with Instance-Level Augmentations", "abstract": "  We present a method for expanding a dataset by incorporating knowledge from\nthe wide distribution of pre-trained latent diffusion models. Data\naugmentations typically incorporate inductive biases about the image formation\nprocess into the training (e.g. translation, scaling, colour changes, etc.).\nHere, we go beyond simple pixel transformations and introduce the concept of\ninstance-level data augmentation by repainting parts of the image at the level\nof object instances. The method combines a conditional diffusion model with\ndepth and edge maps control conditioning to seamlessly repaint individual\nobjects inside the scene, being applicable to any segmentation or detection\ndataset. Used as a data augmentation method, it improves the performance and\ngeneralization of the state-of-the-art salient object detection, semantic\nsegmentation and object detection models. By redrawing all privacy-sensitive\ninstances (people, license plates, etc.), the method is also applicable for\ndata anonymization. We also release fully synthetic and anonymized expansions\nfor popular datasets: COCO, Pascal VOC and DUTS.\n", "date": "2024-06-12", "categories": []}
{"id": "2403.07294", "title": "Graph Data Condensation via Self-expressive Graph Structure\n  Reconstruction", "abstract": "  With the increasing demands of training graph neural networks (GNNs) on\nlarge-scale graphs, graph data condensation has emerged as a critical technique\nto relieve the storage and time costs during the training phase. It aims to\ncondense the original large-scale graph to a much smaller synthetic graph while\npreserving the essential information necessary for efficiently training a\ndownstream GNN. However, existing methods concentrate either on optimizing node\nfeatures exclusively or endeavor to independently learn node features and the\ngraph structure generator. They could not explicitly leverage the information\nof the original graph structure and failed to construct an interpretable graph\nstructure for the synthetic dataset. To address these issues, we introduce a\nnovel framework named \\textbf{G}raph Data \\textbf{C}ondensation via\n\\textbf{S}elf-expressive Graph Structure \\textbf{R}econstruction\n(\\textbf{GCSR}). Our method stands out by (1) explicitly incorporating the\noriginal graph structure into the condensing process and (2) capturing the\nnuanced interdependencies between the condensed nodes by reconstructing an\ninterpretable self-expressive graph structure. Extensive experiments and\ncomprehensive analysis validate the efficacy of the proposed method across\ndiverse GNN models and datasets. Our code is available at\n\\url{https://github.com/zclzcl0223/GCSR}.\n", "date": "2024-03-12", "categories": []}
{"id": "2406.03793", "title": "Low-Rank Similarity Mining for Multimodal Dataset Distillation", "abstract": "  Though dataset distillation has witnessed rapid development in recent years,\nthe distillation of multimodal data, e.g., image-text pairs, poses unique and\nunder-explored challenges. Unlike unimodal data, image-text contrastive\nlearning (ITC) data lack inherent categorization and should instead place\ngreater emphasis on modality correspondence. In this work, we propose Low-Rank\nSimilarity Mining (LoRS) for multimodal dataset distillation, that concurrently\ndistills a ground truth similarity matrix with image-text pairs, and leverages\nlow-rank factorization for efficiency and scalability. The proposed approach\nbrings significant improvement to the existing algorithms, marking a\nsignificant contribution to the field of visual-language dataset distillation.\nWe advocate adopting LoRS as a foundational synthetic data setup for image-text\ndataset distillation. Our code is available at\nhttps://github.com/silicx/LoRS_Distill.\n", "date": "2024-06-06", "categories": []}
{"id": "2403.06741", "title": "Distribution-Aware Data Expansion with Diffusion Models", "abstract": "  The scale and quality of a dataset significantly impact the performance of\ndeep models. However, acquiring large-scale annotated datasets is both a costly\nand time-consuming endeavor. To address this challenge, dataset expansion\ntechnologies aim to automatically augment datasets, unlocking the full\npotential of deep models. Current data expansion techniques include image\ntransformation and image synthesis methods. Transformation-based methods\nintroduce only local variations, leading to limited diversity. In contrast,\nsynthesis-based methods generate entirely new content, greatly enhancing\ninformativeness. However, existing synthesis methods carry the risk of\ndistribution deviations, potentially degrading model performance with\nout-of-distribution samples. In this paper, we propose DistDiff, a\ntraining-free data expansion framework based on the distribution-aware\ndiffusion model. DistDiff constructs hierarchical prototypes to approximate the\nreal data distribution, optimizing latent data points within diffusion models\nwith hierarchical energy guidance. We demonstrate its capability to generate\ndistribution-consistent samples, significantly improving data expansion tasks.\nDistDiff consistently enhances accuracy across a diverse range of datasets\ncompared to models trained solely on original data. Furthermore, our approach\nconsistently outperforms existing synthesis-based techniques and demonstrates\ncompatibility with widely adopted transformation-based augmentation methods.\nAdditionally, the expanded dataset exhibits robustness across various\narchitectural frameworks. Our code is available at\nhttps://github.com/haoweiz23/DistDiff\n", "date": "2024-03-11", "categories": []}
{"id": "2406.02612", "title": "Is Data Valuation Learnable and Interpretable?", "abstract": "  Measuring the value of individual samples is critical for many data-driven\ntasks, e.g., the training of a deep learning model. Recent literature witnesses\nthe substantial efforts in developing data valuation methods. The primary data\nvaluation methodology is based on the Shapley value from game theory, and\nvarious methods are proposed along this path. {Even though Shapley value-based\nvaluation has solid theoretical basis, it is entirely an experiment-based\napproach and no valuation model has been constructed so far.} In addition,\ncurrent data valuation methods ignore the interpretability of the output\nvalues, despite an interptable data valuation method is of great helpful for\napplications such as data pricing. This study aims to answer an important\nquestion: is data valuation learnable and interpretable? A learned valuation\nmodel have several desirable merits such as fixed number of parameters and\nknowledge reusability. An intrepretable data valuation model can explain why a\nsample is valuable or invaluable. To this end, two new data value modeling\nframeworks are proposed, in which a multi-layer perception~(MLP) and a new\nregression tree are utilized as specific base models for model training and\ninterpretability, respectively. Extensive experiments are conducted on\nbenchmark datasets. {The experimental results provide a positive answer for the\nquestion.} Our study opens up a new technical path for the assessing of data\nvalues. Large data valuation models can be built across many different\ndata-driven tasks, which can promote the widespread application of data\nvaluation.\n", "date": "2024-06-03", "categories": []}
{"id": "2406.03057", "title": "BWS: Best Window Selection Based on Sample Scores for Data Pruning\n  across Broad Ranges", "abstract": "  Data subset selection aims to find a smaller yet informative subset of a\nlarge dataset that can approximate the full-dataset training, addressing\nchallenges associated with training neural networks on large-scale datasets.\nHowever, existing methods tend to specialize in either high or low selection\nratio regimes, lacking a universal approach that consistently achieves\ncompetitive performance across a broad range of selection ratios. We introduce\na universal and efficient data subset selection method, Best Window Selection\n(BWS), by proposing a method to choose the best window subset from samples\nordered based on their difficulty scores. This approach offers flexibility by\nallowing the choice of window intervals that span from easy to difficult\nsamples. Furthermore, we provide an efficient mechanism for selecting the best\nwindow subset by evaluating its quality using kernel ridge regression. Our\nexperimental results demonstrate the superior performance of BWS compared to\nother baselines across a broad range of selection ratios over datasets,\nincluding CIFAR-10/100 and ImageNet, and the scenarios involving training from\nrandom initialization or fine-tuning of pre-trained models.\n", "date": "2024-06-05", "categories": []}
{"id": "2212.10888", "title": "A Survey of Mix-based Data Augmentation: Taxonomy, Methods,\n  Applications, and Explainability", "abstract": "  Data augmentation (DA) is indispensable in modern machine learning and deep\nneural networks. The basic idea of DA is to construct new training data to\nimprove the model's generalization by adding slightly disturbed versions of\nexisting data or synthesizing new data. This survey comprehensively reviews a\ncrucial subset of DA techniques, namely Mix-based Data Augmentation (MixDA),\nwhich generates novel samples by combining multiple examples. In contrast to\ntraditional DA approaches that operate on single samples or entire datasets,\nMixDA stands out due to its effectiveness, simplicity, flexibility,\ncomputational efficiency, theoretical foundation, and broad applicability. We\nbegin by introducing a novel taxonomy that categorizes MixDA into Mixup-based,\nCutmix-based, and mixture approaches based on a hierarchical perspective of the\ndata mixing operation. Subsequently, we provide an in-depth review of various\nMixDA techniques, focusing on their underlying motivations. Owing to its\nversatility, MixDA has penetrated a wide range of applications, which we also\nthoroughly investigate in this survey. Moreover, we delve into the underlying\nmechanisms of MixDA's effectiveness by examining its impact on model\ngeneralization and calibration while providing insights into the model's\nbehavior by analyzing the inherent properties of MixDA. Finally, we\nrecapitulate the critical findings and fundamental challenges of current MixDA\nstudies while outlining the potential directions for future works. Different\nfrom previous related surveys that focus on DA approaches in specific domains\n(e.g., CV and NLP) or only review a limited subset of MixDA studies, we are the\nfirst to provide a systematical survey of MixDA, covering its taxonomy,\nmethodology, application, and explainability. Furthermore, we provide promising\ndirections for researchers interested in this exciting area.\n", "date": "2022-12-21", "categories": []}
{"id": "2312.10302", "title": "One-Shot Learning as Instruction Data Prospector for Large Language\n  Models", "abstract": "  Contemporary practices in instruction tuning often hinge on enlarging data\nscaling without a clear strategy for ensuring data quality, inadvertently\nintroducing noise that may compromise model performance. To address this\nchallenge, we introduce \\textsc{Nuggets}, a novel and efficient methodology\nthat leverages one-shot learning to discern and select high-quality instruction\ndata from extensive datasets. \\textsc{Nuggets} assesses the potential of\nindividual instruction examples to act as effective one-shot learning\ninstances, thereby identifying those that can significantly improve performance\nacross diverse tasks. \\textsc{Nuggets} utilizes a scoring system based on the\nimpact of candidate examples on the perplexity of a diverse anchor set,\nfacilitating the selection of the most advantageous data for instruction\ntuning. Through comprehensive evaluations on two benchmarks, including MT-Bench\nand Alpaca-Eval, we show that instruction tuning with the top 1\\% of examples\ncurated by \\textsc{Nuggets} substantially outperforms conventional methods\nemploying the entire dataset.\n", "date": "2023-12-16", "categories": []}
{"id": "2406.01063", "title": "DANCE: Dual-View Distribution Alignment for Dataset Condensation", "abstract": "  Dataset condensation addresses the problem of data burden by learning a small\nsynthetic training set that preserves essential knowledge from the larger real\ntraining set. To date, the state-of-the-art (SOTA) results are often yielded by\noptimization-oriented methods, but their inefficiency hinders their application\nto realistic datasets. On the other hand, the Distribution-Matching (DM)\nmethods show remarkable efficiency but sub-optimal results compared to\noptimization-oriented methods. In this paper, we reveal the limitations of\ncurrent DM-based methods from the inner-class and inter-class views, i.e.,\nPersistent Training and Distribution Shift. To address these problems, we\npropose a new DM-based method named Dual-view distribution AligNment for\ndataset CondEnsation (DANCE), which exploits a few pre-trained models to\nimprove DM from both inner-class and inter-class views. Specifically, from the\ninner-class view, we construct multiple \"middle encoders\" to perform pseudo\nlong-term distribution alignment, making the condensed set a good proxy of the\nreal one during the whole training process; while from the inter-class view, we\nuse the expert models to perform distribution calibration, ensuring the\nsynthetic data remains in the real class region during condensing. Experiments\ndemonstrate the proposed method achieves a SOTA performance while maintaining\ncomparable efficiency with the original DM across various scenarios. Source\ncodes are available at https://github.com/Hansong-Zhang/DANCE.\n", "date": "2024-06-03", "categories": []}
{"id": "2406.01112", "title": "BACON: Bayesian Optimal Condensation Framework for Dataset Distillation", "abstract": "  Dataset Distillation (DD) aims to distill knowledge from extensive datasets\ninto more compact ones while preserving performance on the test set, thereby\nreducing storage costs and training expenses. However, existing methods often\nsuffer from computational intensity, particularly exhibiting suboptimal\nperformance with large dataset sizes due to the lack of a robust theoretical\nframework for analyzing the DD problem. To address these challenges, we propose\nthe BAyesian optimal CONdensation framework (BACON), which is the first work to\nintroduce the Bayesian theoretical framework to the literature of DD. This\nframework provides theoretical support for enhancing the performance of DD.\nFurthermore, BACON formulates the DD problem as the minimization of the\nexpected risk function in joint probability distributions using the Bayesian\nframework. Additionally, by analyzing the expected risk function for optimal\ncondensation, we derive a numerically feasible lower bound based on specific\nassumptions, providing an approximate solution for BACON. We validate BACON\nacross several datasets, demonstrating its superior performance compared to\nexisting state-of-the-art methods. For instance, under the IPC-10 setting,\nBACON achieves a 3.46% accuracy gain over the IDM method on the CIFAR-10\ndataset and a 3.10% gain on the TinyImageNet dataset. Our extensive experiments\nconfirm the effectiveness of BACON and its seamless integration with existing\nmethods, thereby enhancing their performance for the DD task. Code and\ndistilled datasets are available at BACON.\n", "date": "2024-06-03", "categories": []}
{"id": "2406.01417", "title": "Mixup Augmentation with Multiple Interpolations", "abstract": "  Mixup and its variants form a popular class of data augmentation\ntechniques.Using a random sample pair, it generates a new sample by linear\ninterpolation of the inputs and labels. However, generating only one single\ninterpolation may limit its augmentation ability. In this paper, we propose a\nsimple yet effective extension called multi-mix, which generates multiple\ninterpolations from a sample pair. With an ordered sequence of generated\nsamples, multi-mix can better guide the training process than standard mixup.\nMoreover, theoretically, this can also reduce the stochastic gradient variance.\nExtensive experiments on a number of synthetic and large-scale data sets\ndemonstrate that multi-mix outperforms various mixup variants and\nnon-mixup-based baselines in terms of generalization, robustness, and\ncalibration.\n", "date": "2024-06-03", "categories": []}
{"id": "2403.12166", "title": "The Power of Few: Accelerating and Enhancing Data Reweighting with\n  Coreset Selection", "abstract": "  As machine learning tasks continue to evolve, the trend has been to gather\nlarger datasets and train increasingly larger models. While this has led to\nadvancements in accuracy, it has also escalated computational costs to\nunsustainable levels. Addressing this, our work aims to strike a delicate\nbalance between computational efficiency and model accuracy, a persisting\nchallenge in the field. We introduce a novel method that employs core subset\nselection for reweighting, effectively optimizing both computational time and\nmodel performance. By focusing on a strategically selected coreset, our\napproach offers a robust representation, as it efficiently minimizes the\ninfluence of outliers. The re-calibrated weights are then mapped back to and\npropagated across the entire dataset. Our experimental results substantiate the\neffectiveness of this approach, underscoring its potential as a scalable and\nprecise solution for model training.\n", "date": "2024-03-18", "categories": []}
{"id": "2405.20456", "title": "Scaling Laws for the Value of Individual Data Points in Machine Learning", "abstract": "  Recent works have shown that machine learning models improve at a predictable\nrate with the total amount of training data, leading to scaling laws that\ndescribe the relationship between error and dataset size. These scaling laws\ncan help design a model's training dataset, but they typically take an\naggregate view of the data by only considering the dataset's size. We introduce\na new perspective by investigating scaling behavior for the value of individual\ndata points: we find that a data point's contribution to model's performance\nshrinks predictably with the size of the dataset in a log-linear manner.\nInterestingly, there is significant variability in the scaling exponent among\ndifferent data points, indicating that certain points are more valuable in\nsmall datasets while others are relatively more useful as a part of large\ndatasets. We provide learning theory to support our scaling law, and we observe\nempirically that it holds across diverse model classes. We further propose a\nmaximum likelihood estimator and an amortized estimator to efficiently learn\nthe individualized scaling behaviors from a small number of noisy observations\nper data point. Using our estimators, we provide insights into factors that\ninfluence the scaling behavior of different data points. Finally, we\ndemonstrate applications of the individualized scaling laws to data valuation\nand data subset selection. Overall, our work represents a first step towards\nunderstanding and utilizing scaling properties for the value of individual data\npoints.\n", "date": "2024-05-30", "categories": []}
{"id": "2405.20541", "title": "Perplexed by Perplexity: Perplexity-Based Data Pruning With Small\n  Reference Models", "abstract": "  In this work, we investigate whether small language models can determine\nhigh-quality subsets of large-scale text datasets that improve the performance\nof larger language models. While existing work has shown that pruning based on\nthe perplexity of a larger model can yield high-quality data, we investigate\nwhether smaller models can be used for perplexity-based pruning and how pruning\nis affected by the domain composition of the data being pruned. We demonstrate\nthat for multiple dataset compositions, perplexity-based pruning of pretraining\ndata can \\emph{significantly} improve downstream task performance: pruning\nbased on perplexities computed with a 125 million parameter model improves the\naverage performance on downstream tasks of a 3 billion parameter model by up to\n2.04 and achieves up to a $1.45\\times$ reduction in pretraining steps to reach\ncommensurate baseline performance. Furthermore, we demonstrate that such\nperplexity-based data pruning also yields downstream performance gains in the\nover-trained and data-constrained regimes.\n", "date": "2024-05-30", "categories": []}
{"id": "2311.13613", "title": "Spanning Training Progress: Temporal Dual-Depth Scoring (TDDS) for\n  Enhanced Dataset Pruning", "abstract": "  Dataset pruning aims to construct a coreset capable of achieving performance\ncomparable to the original, full dataset. Most existing dataset pruning methods\nrely on snapshot-based criteria to identify representative samples, often\nresulting in poor generalization across various pruning and cross-architecture\nscenarios. Recent studies have addressed this issue by expanding the scope of\ntraining dynamics considered, including factors such as forgetting event and\nprobability change, typically using an averaging approach. However, these works\nstruggle to integrate a broader range of training dynamics without overlooking\nwell-generalized samples, which may not be sufficiently highlighted in an\naveraging manner. In this study, we propose a novel dataset pruning method\ntermed as Temporal Dual-Depth Scoring (TDDS), to tackle this problem. TDDS\nutilizes a dual-depth strategy to achieve a balance between incorporating\nextensive training dynamics and identifying representative samples for dataset\npruning. In the first depth, we estimate the series of each sample's individual\ncontributions spanning the training progress, ensuring comprehensive\nintegration of training dynamics. In the second depth, we focus on the\nvariability of the sample-wise contributions identified in the first depth to\nhighlight well-generalized samples. Extensive experiments conducted on CIFAR\nand ImageNet datasets verify the superiority of TDDS over previous SOTA\nmethods. Specifically on CIFAR-100, our method achieves 54.51% accuracy with\nonly 10% training data, surpassing random selection by 7.83% and other\ncomparison methods by at least 12.69%.\n", "date": "2023-11-22", "categories": []}
{"id": "2405.17489", "title": "On the Inflation of KNN-Shapley Value", "abstract": "  Shapley value-based data valuation methods, originating from cooperative game\ntheory, quantify the usefulness of each individual sample by considering its\ncontribution to all possible training subsets. Despite their extensive\napplications, these methods encounter the challenge of value inflation - while\nsamples with negative Shapley values are detrimental, some with positive values\ncan also be harmful. This challenge prompts two fundamental questions: the\nsuitability of zero as a threshold for distinguishing detrimental from\nbeneficial samples and the determination of an appropriate threshold. To\naddress these questions, we focus on KNN-Shapley and propose Calibrated\nKNN-Shapley (CKNN-Shapley), which calibrates zero as the threshold to\ndistinguish detrimental samples from beneficial ones by mitigating the negative\neffects of small-sized training subsets. Through extensive experiments, we\ndemonstrate the effectiveness of CKNN-Shapley in alleviating data valuation\ninflation, detecting detrimental samples, and assessing data quality. We also\nextend our approach beyond conventional classification settings, applying it to\ndiverse and practical scenarios such as learning with mislabeled data, online\nlearning with stream data, and active learning for label annotation.\n", "date": "2024-05-25", "categories": []}
{"id": "2405.17535", "title": "Calibrated Dataset Condensation for Faster Hyperparameter Search", "abstract": "  Dataset condensation can be used to reduce the computational cost of training\nmultiple models on a large dataset by condensing the training dataset into a\nsmall synthetic set. State-of-the-art approaches rely on matching the model\ngradients between the real and synthetic data. However, there is no theoretical\nguarantee of the generalizability of the condensed data: data condensation\noften generalizes poorly across hyperparameters/architectures in practice. This\npaper considers a different condensation objective specifically geared toward\nhyperparameter search. We aim to generate a synthetic validation dataset so\nthat the validation-performance rankings of the models, with different\nhyperparameters, on the condensed and original datasets are comparable. We\npropose a novel hyperparameter-calibrated dataset condensation (HCDC)\nalgorithm, which obtains the synthetic validation dataset by matching the\nhyperparameter gradients computed via implicit differentiation and efficient\ninverse Hessian approximation. Experiments demonstrate that the proposed\nframework effectively maintains the validation-performance rankings of models\nand speeds up hyperparameter/architecture search for tasks on both images and\ngraphs.\n", "date": "2024-05-27", "categories": []}
{"id": "2405.14881", "title": "DiffuseMix: Label-Preserving Data Augmentation with Diffusion Models", "abstract": "  Recently, a number of image-mixing-based augmentation techniques have been\nintroduced to improve the generalization of deep neural networks. In these\ntechniques, two or more randomly selected natural images are mixed together to\ngenerate an augmented image. Such methods may not only omit important portions\nof the input images but also introduce label ambiguities by mixing images\nacross labels resulting in misleading supervisory signals. To address these\nlimitations, we propose DiffuseMix, a novel data augmentation technique that\nleverages a diffusion model to reshape training images, supervised by our\nbespoke conditional prompts. First, concatenation of a partial natural image\nand its generated counterpart is obtained which helps in avoiding the\ngeneration of unrealistic images or label ambiguities. Then, to enhance\nresilience against adversarial attacks and improves safety measures, a randomly\nselected structural pattern from a set of fractal images is blended into the\nconcatenated image to form the final augmented image for training. Our\nempirical results on seven different datasets reveal that DiffuseMix achieves\nsuperior performance compared to existing state-of the-art methods on tasks\nincluding general classification,fine-grained classification, fine-tuning, data\nscarcity, and adversarial robustness. Augmented datasets and codes are\navailable here: https://diffusemix.github.io/\n", "date": "2024-04-05", "categories": []}
{"id": "2405.13866", "title": "Koopcon: A new approach towards smarter and less complex learning", "abstract": "  In the era of big data, the sheer volume and complexity of datasets pose\nsignificant challenges in machine learning, particularly in image processing\ntasks. This paper introduces an innovative Autoencoder-based Dataset\nCondensation Model backed by Koopman operator theory that effectively packs\nlarge datasets into compact, information-rich representations. Inspired by the\npredictive coding mechanisms of the human brain, our model leverages a novel\napproach to encode and reconstruct data, maintaining essential features and\nlabel distributions. The condensation process utilizes an autoencoder neural\nnetwork architecture, coupled with Optimal Transport theory and Wasserstein\ndistance, to minimize the distributional discrepancies between the original and\nsynthesized datasets. We present a two-stage implementation strategy: first,\ncondensing the large dataset into a smaller synthesized subset; second,\nevaluating the synthesized data by training a classifier and comparing its\nperformance with a classifier trained on an equivalent subset of the original\ndata. Our experimental results demonstrate that the classifiers trained on\ncondensed data exhibit comparable performance to those trained on the original\ndatasets, thus affirming the efficacy of our condensation model. This work not\nonly contributes to the reduction of computational resources but also paves the\nway for efficient data handling in constrained environments, marking a\nsignificant step forward in data-efficient machine learning.\n", "date": "2024-05-22", "categories": []}
{"id": "2405.13954", "title": "What is Your Data Worth to GPT? LLM-Scale Data Valuation with Influence\n  Functions", "abstract": "  Large language models (LLMs) are trained on a vast amount of human-written\ndata, but data providers often remain uncredited. In response to this issue,\ndata valuation (or data attribution), which quantifies the contribution or\nvalue of each data to the model output, has been discussed as a potential\nsolution. Nevertheless, applying existing data valuation methods to recent LLMs\nand their vast training datasets has been largely limited by prohibitive\ncompute and memory costs. In this work, we focus on influence functions, a\npopular gradient-based data valuation method, and significantly improve its\nscalability with an efficient gradient projection strategy called LoGra that\nleverages the gradient structure in backpropagation. We then provide a\ntheoretical motivation of gradient projection approaches to influence functions\nto promote trust in the data valuation process. Lastly, we lower the barrier to\nimplementing data valuation systems by introducing LogIX, a software package\nthat can transform existing training code into data valuation code with minimal\neffort. In our data valuation experiments, LoGra achieves competitive accuracy\nagainst more expensive baselines while showing up to 6,500x improvement in\nthroughput and 5x reduction in GPU memory usage when applied to\nLlama3-8B-Instruct and the 1B-token dataset.\n", "date": "2024-05-22", "categories": []}
{"id": "2405.08217", "title": "Data Valuation with Gradient Similarity", "abstract": "  High-quality data is crucial for accurate machine learning and actionable\nanalytics, however, mislabeled or noisy data is a common problem in many\ndomains. Distinguishing low- from high-quality data can be challenging, often\nrequiring expert knowledge and considerable manual intervention. Data Valuation\nalgorithms are a class of methods that seek to quantify the value of each\nsample in a dataset based on its contribution or importance to a given\npredictive task. These data values have shown an impressive ability to identify\nmislabeled observations, and filtering low-value data can boost machine\nlearning performance. In this work, we present a simple alternative to existing\nmethods, termed Data Valuation with Gradient Similarity (DVGS). This approach\ncan be easily applied to any gradient descent learning algorithm, scales well\nto large datasets, and performs comparably or better than baseline valuation\nmethods for tasks such as corrupted label discovery and noise quantification.\nWe evaluate the DVGS method on tabular, image and RNA expression datasets to\nshow the effectiveness of the method across domains. Our approach has the\nability to rapidly and accurately identify low-quality data, which can reduce\nthe need for expert knowledge and manual intervention in data cleaning tasks.\n", "date": "2024-05-13", "categories": []}
{"id": "2402.00195", "title": "Dataset Condensation Driven Machine Unlearning", "abstract": "  The current trend in data regulation requirements and privacy-preserving\nmachine learning has emphasized the importance of machine unlearning. The naive\napproach to unlearning training data by retraining over the complement of the\nforget samples is susceptible to computational challenges. These challenges\nhave been effectively addressed through a collection of techniques falling\nunder the umbrella of machine unlearning. However, there still exists a lack of\nsufficiency in handling persistent computational challenges in harmony with the\nutility and privacy of unlearned model. We attribute this to the lack of work\non improving the computational complexity of approximate unlearning from the\nperspective of the training dataset. In this paper, we aim to fill this gap by\nintroducing dataset condensation as an essential component of machine\nunlearning in the context of image classification. To achieve this goal, we\npropose new dataset condensation techniques and an innovative unlearning scheme\nthat strikes a balance between machine unlearning privacy, utility, and\nefficiency. Furthermore, we present a novel and effective approach to\ninstrumenting machine unlearning and propose its application in defending\nagainst membership inference and model inversion attacks. Additionally, we\nexplore a new application of our approach, which involves removing data from\n`condensed model', which can be employed to quickly train any arbitrary model\nwithout being influenced by unlearning samples. The corresponding code is\navailable at \\href{https://github.com/algebraicdianuj/DC_U}{URL}.\n", "date": "2024-01-31", "categories": []}
{"id": "2405.06893", "title": "ADLDA: A Method to Reduce the Harm of Data Distribution Shift in Data\n  Augmentation", "abstract": "  This study introduces a novel data augmentation technique, ADLDA, aimed at\nmitigating the negative impact of data distribution shifts caused by the data\naugmentation process in computer vision task. ADLDA partitions augmented data\ninto distinct subdomains and incorporates domain labels, combined with domain\nadaptation techniques, to optimize data representation in the model's feature\nspace. Experimental results demonstrate that ADLDA significantly enhances model\nperformance across multiple datasets, particularly in neural network\narchitectures with complex feature extraction layers. Furthermore, ADLDA\nimproves the model's ability to locate and recognize key features, showcasing\npotential in object recognition and image segmentation tasks. This paper's\ncontribution provides an effective data augmentation regularization method for\nthe field of computer vision aiding in the enhancement of robustness and\naccuracy in deep learning models.\n", "date": "2024-05-11", "categories": []}
{"id": "2311.05304", "title": "Data Valuation and Detections in Federated Learning", "abstract": "  Federated Learning (FL) enables collaborative model training while preserving\nthe privacy of raw data. A challenge in this framework is the fair and\nefficient valuation of data, which is crucial for incentivizing clients to\ncontribute high-quality data in the FL task. In scenarios involving numerous\ndata clients within FL, it is often the case that only a subset of clients and\ndatasets are pertinent to a specific learning task, while others might have\neither a negative or negligible impact on the model training process. This\npaper introduces a novel privacy-preserving method for evaluating client\ncontributions and selecting relevant datasets without a pre-specified training\nalgorithm in an FL task. Our proposed approach FedBary, utilizes Wasserstein\ndistance within the federated context, offering a new solution for data\nvaluation in the FL framework. This method ensures transparent data valuation\nand efficient computation of the Wasserstein barycenter and reduces the\ndependence on validation datasets. Through extensive empirical experiments and\ntheoretical analyses, we demonstrate the potential of this data valuation\nmethod as a promising avenue for FL research.\n", "date": "2023-11-09", "categories": []}
{"id": "2309.12689", "title": "AMPLIFY:Attention-based Mixup for Performance Improvement and Label\n  Smoothing in Transformer", "abstract": "  Mixup is an effective data augmentation method that generates new augmented\nsamples by aggregating linear combinations of different original samples.\nHowever, if there are noises or aberrant features in the original samples,\nMixup may propagate them to the augmented samples, leading to over-sensitivity\nof the model to these outliers . To solve this problem, this paper proposes a\nnew Mixup method called AMPLIFY. This method uses the Attention mechanism of\nTransformer itself to reduce the influence of noises and aberrant values in the\noriginal samples on the prediction results, without increasing additional\ntrainable parameters, and the computational cost is very low, thereby avoiding\nthe problem of high resource consumption in common Mixup methods such as\nSentence Mixup . The experimental results show that, under a smaller\ncomputational resource cost, AMPLIFY outperforms other Mixup methods in text\nclassification tasks on 7 benchmark datasets, providing new ideas and new ways\nto further improve the performance of pre-trained models based on the Attention\nmechanism, such as BERT, ALBERT, RoBERTa, and GPT. Our code can be obtained at\nhttps://github.com/kiwi-lilo/AMPLIFY.\n", "date": "2023-09-22", "categories": []}
{"id": "2405.03875", "title": "Rethinking Data Shapley for Data Selection Tasks: Misleads and Merits", "abstract": "  Data Shapley provides a principled approach to data valuation and plays a\ncrucial role in data-centric machine learning (ML) research. Data selection is\nconsidered a standard application of Data Shapley. However, its data selection\nperformance has shown to be inconsistent across settings in the literature.\nThis study aims to deepen our understanding of this phenomenon. We introduce a\nhypothesis testing framework and show that Data Shapley's performance can be no\nbetter than random selection without specific constraints on utility functions.\nWe identify a class of utility functions, monotonically transformed modular\nfunctions, within which Data Shapley optimally selects data. Based on this\ninsight, we propose a heuristic for predicting Data Shapley's effectiveness in\ndata selection tasks. Our experiments corroborate these findings, adding new\ninsights into when Data Shapley may or may not succeed.\n", "date": "2024-05-06", "categories": []}
{"id": "2405.02830", "title": "You Only Need Half: Boosting Data Augmentation by Using Partial Content", "abstract": "  We propose a novel data augmentation method termed You Only Need hAlf (YONA),\nwhich simplifies the augmentation process. YONA bisects an image, substitutes\none half with noise, and applies data augmentation techniques to the remaining\nhalf. This method reduces the redundant information in the original image,\nencourages neural networks to recognize objects from incomplete views, and\nsignificantly enhances neural networks' robustness. YONA is distinguished by\nits properties of parameter-free, straightforward application, enhancing\nvarious existing data augmentation strategies, and thereby bolstering neural\nnetworks' robustness without additional computational cost. To demonstrate\nYONA's efficacy, extensive experiments were carried out. These experiments\nconfirm YONA's compatibility with diverse data augmentation methods and neural\nnetwork architectures, yielding substantial improvements in CIFAR\nclassification tasks, sometimes outperforming conventional image-level data\naugmentation methods. Furthermore, YONA markedly increases the resilience of\nneural networks to adversarial attacks. Additional experiments exploring YONA's\nvariants conclusively show that masking half of an image optimizes performance.\nThe code is available at https://github.com/HansMoe/YONA.\n", "date": "2024-05-05", "categories": []}
{"id": "2405.00348", "title": "Practical Dataset Distillation Based on Deep Support Vectors", "abstract": "  Conventional dataset distillation requires significant computational\nresources and assumes access to the entire dataset, an assumption impractical\nas it presumes all data resides on a central server. In this paper, we focus on\ndataset distillation in practical scenarios with access to only a fraction of\nthe entire dataset. We introduce a novel distillation method that augments the\nconventional process by incorporating general model knowledge via the addition\nof Deep KKT (DKKT) loss. In practical settings, our approach showed improved\nperformance compared to the baseline distribution matching distillation method\non the CIFAR-10 dataset. Additionally, we present experimental evidence that\nDeep Support Vectors (DSVs) offer unique information to the original\ndistillation, and their integration results in enhanced performance.\n", "date": "2024-05-01", "categories": []}
{"id": "2210.03921", "title": "Data Selection: A General Principle for Building Small Interpretable\n  Models", "abstract": "  We present convincing empirical evidence for an effective and general\nstrategy for building accurate small models. Such models are attractive for\ninterpretability and also find use in resource-constrained environments. The\nstrategy is to learn the training distribution and sample accordingly from the\nprovided training data. The distribution learning algorithm is not a\ncontribution of this work; our contribution is a rigorous demonstration of the\nbroad utility of this strategy in various practical settings. We apply it to\nthe tasks of (1) building cluster explanation trees, (2) prototype-based\nclassification, and (3) classification using Random Forests, and show that it\nimproves the accuracy of decades-old weak traditional baselines to be\ncompetitive with specialized modern techniques.\n  This strategy is also versatile wrt the notion of model size. In the first\ntwo tasks, model size is considered to be number of leaves in the tree and the\nnumber of prototypes respectively. In the final task involving Random Forests,\nthe strategy is shown to be effective even when model size comprises of more\nthan one factor: number of trees and their maximum depth.\n  Positive results using multiple datasets are presented that are shown to be\nstatistically significant.\n", "date": "2022-10-08", "categories": []}
{"id": "2306.12189", "title": "Annotating Ambiguous Images: General Annotation Strategy for\n  High-Quality Data with Real-World Biomedical Validation", "abstract": "  In the field of image classification, existing methods often struggle with\nbiased or ambiguous data, a prevalent issue in real-world scenarios. Current\nstrategies, including semi-supervised learning and class blending, offer\npartial solutions but lack a definitive resolution. Addressing this gap, our\npaper introduces a novel strategy for generating high-quality labels in\nchallenging datasets. Central to our approach is a clearly designed flowchart,\nbased on a broad literature review, which enables the creation of reliable\nlabels. We validate our methodology through a rigorous real-world test case in\nthe biomedical field, specifically in deducing height reduction from vertebral\nimaging. Our empirical study, leveraging over 250,000 annotations, demonstrates\nthe effectiveness of our strategies decisions compared to their alternatives.\n", "date": "2023-06-21", "categories": []}
{"id": "2404.14361", "title": "Better Synthetic Data by Retrieving and Transforming Existing Datasets", "abstract": "  Despite recent advances in large language models, building dependable and\ndeployable NLP models typically requires abundant, high-quality training data.\nHowever, task-specific data is not available for many use cases, and manually\ncurating task-specific data is labor-intensive. Recent work has studied\nprompt-driven synthetic data generation using large language models, but these\ngenerated datasets tend to lack complexity and diversity. To address these\nlimitations, we introduce a method, DataTune, to make better use of existing,\npublicly available datasets to improve automatic dataset generation. DataTune\nperforms dataset transformation, enabling the repurposing of publicly available\ndatasets into a format that is directly aligned with the specific requirements\nof target tasks. On a diverse set of language-based tasks from the BIG-Bench\nbenchmark, we find that finetuning language models via DataTune improves over a\nfew-shot prompting baseline by 49% and improves over existing methods that use\nsynthetic or retrieved training data by 34%. We find that dataset\ntransformation significantly increases the diversity and difficulty of\ngenerated data on many tasks. We integrate DataTune into an open-source\nrepository to make this method accessible to the community:\nhttps://github.com/neulab/prompt2model.\n", "date": "2024-04-22", "categories": []}
{"id": "2404.17732", "title": "Generative Dataset Distillation: Balancing Global Structure and Local\n  Details", "abstract": "  In this paper, we propose a new dataset distillation method that considers\nbalancing global structure and local details when distilling the information\nfrom a large dataset into a generative model. Dataset distillation has been\nproposed to reduce the size of the required dataset when training models. The\nconventional dataset distillation methods face the problem of long redeployment\ntime and poor cross-architecture performance. Moreover, previous methods\nfocused too much on the high-level semantic attributes between the synthetic\ndataset and the original dataset while ignoring the local features such as\ntexture and shape. Based on the above understanding, we propose a new method\nfor distilling the original image dataset into a generative model. Our method\ninvolves using a conditional generative adversarial network to generate the\ndistilled dataset. Subsequently, we ensure balancing global structure and local\ndetails in the distillation process, continuously optimizing the generator for\nmore information-dense dataset generation.\n", "date": "2024-04-26", "categories": []}
{"id": "2404.18881", "title": "Human-in-the-Loop Synthetic Text Data Inspection with Provenance\n  Tracking", "abstract": "  Data augmentation techniques apply transformations to existing texts to\ngenerate additional data. The transformations may produce low-quality texts,\nwhere the meaning of the text is changed and the text may even be mangled\nbeyond human comprehension. Analyzing the synthetically generated texts and\ntheir corresponding labels is slow and demanding. To winnow out texts with\nincorrect labels, we develop INSPECTOR, a human-in-the-loop data inspection\ntechnique. INSPECTOR combines the strengths of provenance tracking techniques\nwith assistive labeling. INSPECTOR allows users to group related texts by their\ntransformation provenance, i.e., the transformations applied to the original\ntext, or feature provenance, the linguistic features of the original text. For\nassistive labeling, INSPECTOR computes metrics that approximate data quality,\nand allows users to compare the corresponding label of each text against the\npredictions of a large language model. In a user study, INSPECTOR increases the\nnumber of texts with correct labels identified by 3X on a sentiment analysis\ntask and by 4X on a hate speech detection task. The participants found grouping\nthe synthetically generated texts by their common transformation to be the most\nuseful technique. Surprisingly, grouping texts by common linguistic features\nwas perceived to be unhelpful. Contrary to prior work, our study finds that no\nsingle technique obviates the need for human inspection effort. This validates\nthe design of INSPECTOR which combines both analysis of data provenance and\nassistive labeling to reduce human inspection effort.\n", "date": "2024-04-29", "categories": []}
{"id": "2404.16886", "title": "Review of Data-centric Time Series Analysis from Sample, Feature, and\n  Period", "abstract": "  Data is essential to performing time series analysis utilizing machine\nlearning approaches, whether for classic models or today's large language\nmodels. A good time-series dataset is advantageous for the model's accuracy,\nrobustness, and convergence, as well as task outcomes and costs. The emergence\nof data-centric AI represents a shift in the landscape from model refinement to\nprioritizing data quality. Even though time-series data processing methods\nfrequently come up in a wide range of research fields, it hasn't been well\ninvestigated as a specific topic. To fill the gap, in this paper, we\nsystematically review different data-centric methods in time series analysis,\ncovering a wide range of research topics. Based on the time-series data\ncharacteristics at sample, feature, and period, we propose a taxonomy for the\nreviewed data selection methods. In addition to discussing and summarizing\ntheir characteristics, benefits, and drawbacks targeting time-series data, we\nalso introduce the challenges and opportunities by proposing recommendations,\nopen problems, and possible research topics.\n", "date": "2024-04-24", "categories": []}
{"id": "2308.00994", "title": "SYNAuG: Exploiting Synthetic Data for Data Imbalance Problems", "abstract": "  Data imbalance in training data often leads to biased predictions from\ntrained models, which in turn causes ethical and social issues. A\nstraightforward solution is to carefully curate training data, but given the\nenormous scale of modern neural networks, this is prohibitively labor-intensive\nand thus impractical. Inspired by recent developments in generative models,\nthis paper explores the potential of synthetic data to address the data\nimbalance problem. To be specific, our method, dubbed SYNAuG, leverages\nsynthetic data to equalize the unbalanced distribution of training data. Our\nexperiments demonstrate that, although a domain gap between real and synthetic\ndata exists, training with SYNAuG followed by fine-tuning with a few real\nsamples allows to achieve impressive performance on diverse tasks with\ndifferent data imbalance issues, surpassing existing task-specific methods for\nthe same purpose.\n", "date": "2023-08-02", "categories": []}
{"id": "2301.13389", "title": "Differentially Private Kernel Inducing Points using features from\n  ScatterNets (DP-KIP-ScatterNet) for Privacy Preserving Data Distillation", "abstract": "  Data distillation aims to generate a small data set that closely mimics the\nperformance of a given learning algorithm on the original data set. The\ndistilled dataset is hence useful to simplify the training process thanks to\nits small data size. However, distilled data samples are not necessarily\nprivacy-preserving, even if they are generally humanly indiscernible. To\naddress this limitation, we introduce differentially private kernel inducing\npoints (DP-KIP) for privacy-preserving data distillation. Unlike our original\nintention to simply apply DP-SGD to the framework of KIP, we find that KIP\nusing infinitely-wide convolutional neural tangent kernels (conv-NTKs) performs\nbetter compared to KIP using fully-connected NTKs. However, KIP with conv-NTKs,\ndue to its convolutional and pooling operations, introduces an unbearable\ncomputational complexity, requiring hundreds of V100 GPUs in parallel to train,\nwhich is impractical and more importantly, such computational resources are\ninaccessible to many. To overcome this issue, we propose an alternative that\ndoes not require pre-training (to avoid a privacy loss) and can well capture\ncomplex information on images, as those features from conv-NKTs do, while the\ncomputational cost is manageable by a single V100 GPU. To this end, we propose\nDP-KIP-ScatterNet, which uses the wavelet features from Scattering networks\n(ScatterNet) instead of those from conv-NTKs, to perform DP-KIP at a reasonable\ncomputational cost. We implement DP-KIP-ScatterNet in -- computationally\nefficient -- JAX and test on several popular image datasets to show its\nefficacy and its superior performance compared to state-of-the art methods in\nimage data distillation with differential privacy guarantees.\n", "date": "2023-01-31", "categories": []}
{"id": "2404.14006", "title": "Distilled Datamodel with Reverse Gradient Matching", "abstract": "  The proliferation of large-scale AI models trained on extensive datasets has\nrevolutionized machine learning. With these models taking on increasingly\ncentral roles in various applications, the need to understand their behavior\nand enhance interpretability has become paramount. To investigate the impact of\nchanges in training data on a pre-trained model, a common approach is\nleave-one-out retraining. This entails systematically altering the training\ndataset by removing specific samples to observe resulting changes within the\nmodel. However, retraining the model for each altered dataset presents a\nsignificant computational challenge, given the need to perform this operation\nfor every dataset variation. In this paper, we introduce an efficient framework\nfor assessing data impact, comprising offline training and online evaluation\nstages. During the offline training phase, we approximate the influence of\ntraining data on the target model through a distilled synset, formulated as a\nreversed gradient matching problem. For online evaluation, we expedite the\nleave-one-out process using the synset, which is then utilized to compute the\nattribution matrix based on the evaluation objective. Experimental evaluations,\nincluding training data attribution and assessments of data quality,\ndemonstrate that our proposed method achieves comparable model behavior\nevaluation while significantly speeding up the process compared to the direct\nretraining method.\n", "date": "2024-04-22", "categories": []}
{"id": "2305.02942", "title": "Incentivising the federation: gradient-based metrics for data selection\n  and valuation in private decentralised training", "abstract": "  Obtaining high-quality data for collaborative training of machine learning\nmodels can be a challenging task due to A) regulatory concerns and B) a lack of\ndata owner incentives to participate. The first issue can be addressed through\nthe combination of distributed machine learning techniques (e.g. federated\nlearning) and privacy enhancing technologies (PET), such as the differentially\nprivate (DP) model training. The second challenge can be addressed by rewarding\nthe participants for giving access to data which is beneficial to the training\nmodel, which is of particular importance in federated settings, where the data\nis unevenly distributed. However, DP noise can adversely affect the\nunderrepresented and the atypical (yet often informative) data samples, making\nit difficult to assess their usefulness. In this work, we investigate how to\nleverage gradient information to permit the participants of private training\nsettings to select the data most beneficial for the jointly trained model. We\nassess two such methods, namely variance of gradients (VoG) and the privacy\nloss-input susceptibility score (PLIS). We show that these techniques can\nprovide the federated clients with tools for principled data selection even in\nstricter privacy settings.\n", "date": "2023-05-04", "categories": []}
{"id": "2312.00362", "title": "Dancing with Still Images: Video Distillation via Static-Dynamic\n  Disentanglement", "abstract": "  Recently, dataset distillation has paved the way towards efficient machine\nlearning, especially for image datasets. However, the distillation for videos,\ncharacterized by an exclusive temporal dimension, remains an underexplored\ndomain. In this work, we provide the first systematic study of video\ndistillation and introduce a taxonomy to categorize temporal compression. Our\ninvestigation reveals that the temporal information is usually not well learned\nduring distillation, and the temporal dimension of synthetic data contributes\nlittle. The observations motivate our unified framework of disentangling the\ndynamic and static information in the videos. It first distills the videos into\nstill images as static memory and then compensates the dynamic and motion\ninformation with a learnable dynamic memory block. Our method achieves\nstate-of-the-art on video datasets at different scales, with a notably smaller\nmemory storage budget. Our code is available at\nhttps://github.com/yuz1wan/video_distillation.\n", "date": "2023-12-01", "categories": []}
{"id": "2403.06075", "title": "Multisize Dataset Condensation", "abstract": "  While dataset condensation effectively enhances training efficiency, its\napplication in on-device scenarios brings unique challenges. 1) Due to the\nfluctuating computational resources of these devices, there's a demand for a\nflexible dataset size that diverges from a predefined size. 2) The limited\ncomputational power on devices often prevents additional condensation\noperations. These two challenges connect to the \"subset degradation problem\" in\ntraditional dataset condensation: a subset from a larger condensed dataset is\noften unrepresentative compared to directly condensing the whole dataset to\nthat smaller size. In this paper, we propose Multisize Dataset Condensation\n(MDC) by compressing N condensation processes into a single condensation\nprocess to obtain datasets with multiple sizes. Specifically, we introduce an\n\"adaptive subset loss\" on top of the basic condensation loss to mitigate the\n\"subset degradation problem\". Our MDC method offers several benefits: 1) No\nadditional condensation process is required; 2) reduced storage requirement by\nreusing condensed images. Experiments validate our findings on networks\nincluding ConvNet, ResNet and DenseNet, and datasets including SVHN, CIFAR-10,\nCIFAR-100 and ImageNet. For example, we achieved 5.22%-6.40% average accuracy\ngains on condensing CIFAR-10 to ten images per class. Code is available at:\nhttps://github.com/he-y/Multisize-Dataset-Condensation.\n", "date": "2024-03-10", "categories": []}
{"id": "2404.08820", "title": "Single-image driven 3d viewpoint training data augmentation for\n  effective wine label recognition", "abstract": "  Confronting the critical challenge of insufficient training data in the field\nof complex image recognition, this paper introduces a novel 3D viewpoint\naugmentation technique specifically tailored for wine label recognition. This\nmethod enhances deep learning model performance by generating visually\nrealistic training samples from a single real-world wine label image,\novercoming the challenges posed by the intricate combinations of text and\nlogos. Classical Generative Adversarial Network (GAN) methods fall short in\nsynthesizing such intricate content combination. Our proposed solution\nleverages time-tested computer vision and image processing strategies to expand\nour training dataset, thereby broadening the range of training samples for deep\nlearning applications. This innovative approach to data augmentation\ncircumvents the constraints of limited training resources. Using the augmented\ntraining images through batch-all triplet metric learning on a Vision\nTransformer (ViT) architecture, we can get the most discriminative embedding\nfeatures for every wine label, enabling us to perform one-shot recognition of\nexisting wine labels in the training classes or future newly collected wine\nlabels unavailable in the training. Experimental results show a significant\nincrease in recognition accuracy over conventional 2D data augmentation\ntechniques.\n", "date": "2024-04-12", "categories": []}
{"id": "2404.09161", "title": "Coreset Selection for Object Detection", "abstract": "  Coreset selection is a method for selecting a small, representative subset of\nan entire dataset. It has been primarily researched in image classification,\nassuming there is only one object per image. However, coreset selection for\nobject detection is more challenging as an image can contain multiple objects.\nAs a result, much research has yet to be done on this topic. Therefore, we\nintroduce a new approach, Coreset Selection for Object Detection (CSOD). CSOD\ngenerates imagewise and classwise representative feature vectors for multiple\nobjects of the same class within each image. Subsequently, we adopt submodular\noptimization for considering both representativeness and diversity and utilize\nthe representative vectors in the submodular optimization process to select a\nsubset. When we evaluated CSOD on the Pascal VOC dataset, CSOD outperformed\nrandom selection by +6.4%p in AP$_{50}$ when selecting 200 images.\n", "date": "2024-04-14", "categories": []}
{"id": "2404.09703", "title": "AI Competitions and Benchmarks: Dataset Development", "abstract": "  Machine learning is now used in many applications thanks to its ability to\npredict, generate, or discover patterns from large quantities of data. However,\nthe process of collecting and transforming data for practical use is intricate.\nEven in today's digital era, where substantial data is generated daily, it is\nuncommon for it to be readily usable; most often, it necessitates meticulous\nmanual data preparation. The haste in developing new models can frequently\nresult in various shortcomings, potentially posing risks when deployed in\nreal-world scenarios (eg social discrimination, critical failures), leading to\nthe failure or substantial escalation of costs in AI-based projects. This\nchapter provides a comprehensive overview of established methodological tools,\nenriched by our practical experience, in the development of datasets for\nmachine learning. Initially, we develop the tasks involved in dataset\ndevelopment and offer insights into their effective management (including\nrequirements, design, implementation, evaluation, distribution, and\nmaintenance). Then, we provide more details about the implementation process\nwhich includes data collection, transformation, and quality evaluation.\nFinally, we address practical considerations regarding dataset distribution and\nmaintenance.\n", "date": "2024-04-15", "categories": []}
{"id": "2310.06511", "title": "Self-Supervised Dataset Distillation for Transfer Learning", "abstract": "  Dataset distillation methods have achieved remarkable success in distilling a\nlarge dataset into a small set of representative samples. However, they are not\ndesigned to produce a distilled dataset that can be effectively used for\nfacilitating self-supervised pre-training. To this end, we propose a novel\nproblem of distilling an unlabeled dataset into a set of small synthetic\nsamples for efficient self-supervised learning (SSL). We first prove that a\ngradient of synthetic samples with respect to a SSL objective in naive bilevel\noptimization is \\textit{biased} due to the randomness originating from data\naugmentations or masking. To address this issue, we propose to minimize the\nmean squared error (MSE) between a model's representations of the synthetic\nexamples and their corresponding learnable target feature representations for\nthe inner objective, which does not introduce any randomness. Our primary\nmotivation is that the model obtained by the proposed inner optimization can\nmimic the \\textit{self-supervised target model}. To achieve this, we also\nintroduce the MSE between representations of the inner model and the\nself-supervised target model on the original full dataset for outer\noptimization. Lastly, assuming that a feature extractor is fixed, we only\noptimize a linear head on top of the feature extractor, which allows us to\nreduce the computational cost and obtain a closed-form solution of the head\nwith kernel ridge regression. We empirically validate the effectiveness of our\nmethod on various applications involving transfer learning.\n", "date": "2023-10-10", "categories": []}
{"id": "2404.07976", "title": "Self-supervised Dataset Distillation: A Good Compression Is All You Need", "abstract": "  Dataset distillation aims to compress information from a large-scale original\ndataset to a new compact dataset while striving to preserve the utmost degree\nof the original data informational essence. Previous studies have predominantly\nconcentrated on aligning the intermediate statistics between the original and\ndistilled data, such as weight trajectory, features, gradient, BatchNorm, etc.\nIn this work, we consider addressing this task through the new lens of model\ninformativeness in the compression stage on the original dataset pretraining.\nWe observe that with the prior state-of-the-art SRe$^2$L, as model sizes\nincrease, it becomes increasingly challenging for supervised pretrained models\nto recover learned information during data synthesis, as the channel-wise mean\nand variance inside the model are flatting and less informative. We further\nnotice that larger variances in BN statistics from self-supervised models\nenable larger loss signals to update the recovered data by gradients, enjoying\nmore informativeness during synthesis. Building on this observation, we\nintroduce SC-DD, a simple yet effective Self-supervised Compression framework\nfor Dataset Distillation that facilitates diverse information compression and\nrecovery compared to traditional supervised learning schemes, further reaps the\npotential of large pretrained models with enhanced capabilities. Extensive\nexperiments are conducted on CIFAR-100, Tiny-ImageNet and ImageNet-1K datasets\nto demonstrate the superiority of our proposed approach. The proposed SC-DD\noutperforms all previous state-of-the-art supervised dataset distillation\nmethods when employing larger models, such as SRe$^2$L, MTT, TESLA, DC, CAFE,\netc., by large margins under the same recovery and post-training budgets. Code\nis available at https://github.com/VILA-Lab/SRe2L/tree/main/SCDD/.\n", "date": "2024-04-11", "categories": []}
{"id": "2404.07177", "title": "Scaling Laws for Data Filtering -- Data Curation cannot be Compute\n  Agnostic", "abstract": "  Vision-language models (VLMs) are trained for thousands of GPU hours on\ncarefully curated web datasets. In recent times, data curation has gained\nprominence with several works developing strategies to retain 'high-quality'\nsubsets of 'raw' scraped data. For instance, the LAION public dataset retained\nonly 10% of the total crawled data. However, these strategies are typically\ndeveloped agnostic of the available compute for training. In this paper, we\nfirst demonstrate that making filtering decisions independent of training\ncompute is often suboptimal: the limited high-quality data rapidly loses its\nutility when repeated, eventually requiring the inclusion of 'unseen' but\n'lower-quality' data. To address this quality-quantity tradeoff\n($\\texttt{QQT}$), we introduce neural scaling laws that account for the\nnon-homogeneous nature of web data, an angle ignored in existing literature.\nOur scaling laws (i) characterize the $\\textit{differing}$ 'utility' of various\nquality subsets of web data; (ii) account for how utility diminishes for a data\npoint at its 'nth' repetition; and (iii) formulate the mutual interaction of\nvarious data pools when combined, enabling the estimation of model performance\non a combination of multiple data pools without ever jointly training on them.\nOur key message is that data curation $\\textit{cannot}$ be agnostic of the\ntotal compute that a model will be trained for. Our scaling laws allow us to\ncurate the best possible pool for achieving top performance on Datacomp at\nvarious compute budgets, carving out a pareto-frontier for data curation. Code\nis available at https://github.com/locuslab/scaling_laws_data_filtering.\n", "date": "2024-04-10", "categories": []}
{"id": "2303.16854", "title": "AnnoLLM: Making Large Language Models to Be Better Crowdsourced\n  Annotators", "abstract": "  Many natural language processing (NLP) tasks rely on labeled data to train\nmachine learning models with high performance. However, data annotation is\ntime-consuming and expensive, especially when the task involves a large amount\nof data or requires specialized domains. Recently, GPT-3.5 series models have\ndemonstrated remarkable few-shot and zero-shot ability across various NLP\ntasks. In this paper, we first claim that large language models (LLMs), such as\nGPT-3.5, can serve as an excellent crowdsourced annotator when provided with\nsufficient guidance and demonstrated examples. Accordingly, we propose AnnoLLM,\nan annotation system powered by LLMs, which adopts a two-step approach,\nexplain-then-annotate. Concretely, we first prompt LLMs to provide explanations\nfor why the specific ground truth answer/label was assigned for a given\nexample. Then, we construct the few-shot chain-of-thought prompt with the\nself-generated explanation and employ it to annotate the unlabeled data with\nLLMs. Our experiment results on three tasks, including user input and keyword\nrelevance assessment, BoolQ, and WiC, demonstrate that AnnoLLM surpasses or\nperforms on par with crowdsourced annotators. Furthermore, we build the first\nconversation-based information retrieval dataset employing AnnoLLM. This\ndataset is designed to facilitate the development of retrieval models capable\nof retrieving pertinent documents for conversational text. Human evaluation has\nvalidated the dataset's high quality.\n", "date": "2023-03-29", "categories": []}
{"id": "2310.20550", "title": "CapsFusion: Rethinking Image-Text Data at Scale", "abstract": "  Large multimodal models demonstrate remarkable generalist ability to perform\ndiverse multimodal tasks in a zero-shot manner. Large-scale web-based\nimage-text pairs contribute fundamentally to this success, but suffer from\nexcessive noise. Recent studies use alternative captions synthesized by\ncaptioning models and have achieved notable benchmark performance. However, our\nexperiments reveal significant Scalability Deficiency and World Knowledge Loss\nissues in models trained with synthetic captions, which have been largely\nobscured by their initial benchmark success. Upon closer examination, we\nidentify the root cause as the overly-simplified language structure and lack of\nknowledge details in existing synthetic captions. To provide higher-quality and\nmore scalable multimodal pretraining data, we propose CapsFusion, an advanced\nframework that leverages large language models to consolidate and refine\ninformation from both web-based image-text pairs and synthetic captions.\nExtensive experiments show that CapsFusion captions exhibit remarkable\nall-round superiority over existing captions in terms of model performance\n(e.g., 18.8 and 18.3 improvements in CIDEr score on COCO and NoCaps), sample\nefficiency (requiring 11-16 times less computation than baselines), world\nknowledge depth, and scalability. These effectiveness, efficiency and\nscalability advantages position CapsFusion as a promising candidate for future\nscaling of LMM training.\n", "date": "2023-10-31", "categories": []}
{"id": "2404.00264", "title": "DiLM: Distilling Dataset into Language Model for Text-level Dataset\n  Distillation", "abstract": "  Dataset distillation aims to compress a training dataset by creating a small\nnumber of informative synthetic samples such that neural networks trained on\nthem perform as well as those trained on the original training dataset. Current\ntext dataset distillation methods create each synthetic sample as a sequence of\nword embeddings instead of a text to apply gradient-based optimization;\nhowever, such embedding-level distilled datasets cannot be used for training\nother models whose word embedding weights are different from the model used for\ndistillation. To address this issue, we propose a novel text dataset\ndistillation approach, called Distilling dataset into Language Model (DiLM),\nwhich trains a language model to generate informative synthetic training\nsamples as text data, instead of directly optimizing synthetic samples. We\nevaluated DiLM on various text classification datasets and showed that\ndistilled synthetic datasets from DiLM outperform those from current coreset\nselection methods. DiLM achieved remarkable generalization performance in\ntraining different types of models and in-context learning of large language\nmodels. Our code will be available at https://github.com/arumaekawa/DiLM.\n", "date": "2024-03-30", "categories": []}
{"id": "2404.00563", "title": "Exploiting Inter-sample and Inter-feature Relations in Dataset\n  Distillation", "abstract": "  Dataset distillation has emerged as a promising approach in deep learning,\nenabling efficient training with small synthetic datasets derived from larger\nreal ones. Particularly, distribution matching-based distillation methods\nattract attention thanks to its effectiveness and low computational cost.\nHowever, these methods face two primary limitations: the dispersed feature\ndistribution within the same class in synthetic datasets, reducing class\ndiscrimination, and an exclusive focus on mean feature consistency, lacking\nprecision and comprehensiveness. To address these challenges, we introduce two\nnovel constraints: a class centralization constraint and a covariance matching\nconstraint. The class centralization constraint aims to enhance class\ndiscrimination by more closely clustering samples within classes. The\ncovariance matching constraint seeks to achieve more accurate feature\ndistribution matching between real and synthetic datasets through local feature\ncovariance matrices, particularly beneficial when sample sizes are much smaller\nthan the number of features. Experiments demonstrate notable improvements with\nthese constraints, yielding performance boosts of up to 6.6% on CIFAR10, 2.9%\non SVHN, 2.5% on CIFAR100, and 2.5% on TinyImageNet, compared to the\nstate-of-the-art relevant methods. In addition, our method maintains robust\nperformance in cross-architecture settings, with a maximum performance drop of\n1.7% on four architectures. Code is available at\nhttps://github.com/VincenDen/IID.\n", "date": "2024-03-31", "categories": []}
{"id": "2403.20012", "title": "Colorful Cutout: Enhancing Image Data Augmentation with Curriculum\n  Learning", "abstract": "  Data augmentation is one of the regularization strategies for the training of\ndeep learning models, which enhances generalizability and prevents overfitting,\nleading to performance improvement. Although researchers have proposed various\ndata augmentation techniques, they often lack consideration for the difficulty\nof augmented data. Recently, another line of research suggests incorporating\nthe concept of curriculum learning with data augmentation in the field of\nnatural language processing. In this study, we adopt curriculum data\naugmentation for image data augmentation and propose colorful cutout, which\ngradually increases the noise and difficulty introduced in the augmented image.\nOur experimental results highlight the possibility of curriculum data\naugmentation for image data. We publicly released our source code to improve\nthe reproducibility of our study.\n", "date": "2024-03-29", "categories": []}
{"id": "2403.19600", "title": "Enhance Image Classification via Inter-Class Image Mixup with Diffusion\n  Model", "abstract": "  Text-to-image (T2I) generative models have recently emerged as a powerful\ntool, enabling the creation of photo-realistic images and giving rise to a\nmultitude of applications. However, the effective integration of T2I models\ninto fundamental image classification tasks remains an open question. A\nprevalent strategy to bolster image classification performance is through\naugmenting the training set with synthetic images generated by T2I models. In\nthis study, we scrutinize the shortcomings of both current generative and\nconventional data augmentation techniques. Our analysis reveals that these\nmethods struggle to produce images that are both faithful (in terms of\nforeground objects) and diverse (in terms of background contexts) for\ndomain-specific concepts. To tackle this challenge, we introduce an innovative\ninter-class data augmentation method known as Diff-Mix\n(https://github.com/Zhicaiwww/Diff-Mix), which enriches the dataset by\nperforming image translations between classes. Our empirical results\ndemonstrate that Diff-Mix achieves a better balance between faithfulness and\ndiversity, leading to a marked improvement in performance across diverse image\nclassification scenarios, including few-shot, conventional, and long-tail\nclassifications for domain-specific datasets.\n", "date": "2024-03-28", "categories": []}
{"id": "2307.09136", "title": "The Effects of Mixed Sample Data Augmentation are Class Dependent", "abstract": "  Mixed Sample Data Augmentation (MSDA) techniques, such as Mixup, CutMix, and\nPuzzleMix, have been widely acknowledged for enhancing performance in a variety\nof tasks. A previous study reported the class dependency of traditional data\naugmentation (DA), where certain classes benefit disproportionately compared to\nothers. This paper reveals a class dependent effect of MSDA, where some classes\nexperience improved performance while others experience degraded performance.\nThis research addresses the issue of class dependency in MSDA and proposes an\nalgorithm to mitigate it. The approach involves training on a mixture of MSDA\nand non-MSDA data, which not only mitigates the negative impact on the affected\nclasses, but also improves overall accuracy. Furthermore, we provide in-depth\nanalysis and discussion of why MSDA introduced class dependencies and which\nclasses are most likely to have them.\n", "date": "2023-07-18", "categories": []}
{"id": "2403.17130", "title": "Exploring the potential of prototype-based soft-labels data distillation\n  for imbalanced data classification", "abstract": "  Dataset distillation aims at synthesizing a dataset by a small number of\nartificially generated data items, which, when used as training data, reproduce\nor approximate a machine learning (ML) model as if it were trained on the\nentire original dataset. Consequently, data distillation methods are usually\ntied to a specific ML algorithm. While recent literature deals mainly with\ndistillation of large collections of images in the context of neural network\nmodels, tabular data distillation is much less represented and mainly focused\non a theoretical perspective. The current paper explores the potential of a\nsimple distillation technique previously proposed in the context of\nLess-than-one shot learning. The main goal is to push further the performance\nof prototype-based soft-labels distillation in terms of classification\naccuracy, by integrating optimization steps in the distillation process. The\nanalysis is performed on real-world data sets with various degrees of\nimbalance. Experimental studies trace the capability of the method to distill\nthe data, but also the opportunity to act as an augmentation method, i.e. to\ngenerate new data that is able to increase model accuracy when used in\nconjunction with - as opposed to instead of - the original data.\n", "date": "2024-03-25", "categories": []}
{"id": "2311.11202", "title": "Unmasking and Improving Data Credibility: A Study with Datasets for\n  Training Harmless Language Models", "abstract": "  Language models have shown promise in various tasks but can be affected by\nundesired data during training, fine-tuning, or alignment. For example, if some\nunsafe conversations are wrongly annotated as safe ones, the model fine-tuned\non these samples may be harmful. Therefore, the correctness of annotations,\ni.e., the credibility of the dataset, is important. This study focuses on the\ncredibility of real-world datasets, including the popular benchmarks Jigsaw\nCivil Comments, Anthropic Harmless & Red Team, PKU BeaverTails & SafeRLHF, that\ncan be used for training a harmless language model. Given the cost and\ndifficulty of cleaning these datasets by humans, we introduce a systematic\nframework for evaluating the credibility of datasets, identifying label errors,\nand evaluating the influence of noisy labels in the curated language data,\nspecifically focusing on unsafe comments and conversation classification. With\nthe framework, we find and fix an average of 6.16% label errors in 11 datasets\nconstructed from the above benchmarks. The data credibility and downstream\nlearning performance can be remarkably improved by directly fixing label\nerrors, indicating the significance of cleaning existing real-world datasets.\nWe provide an open-source tool, Docta, for data cleaning at\nhttps://github.com/Docta-ai/docta.\n", "date": "2023-11-19", "categories": []}
{"id": "2311.13614", "title": "HalluciDoctor: Mitigating Hallucinatory Toxicity in Visual Instruction\n  Data", "abstract": "  Multi-modal Large Language Models (MLLMs) tuned on machine-generated\ninstruction-following data have demonstrated remarkable performance in various\nmulti-modal understanding and generation tasks. However, the hallucinations\ninherent in machine-generated data, which could lead to hallucinatory outputs\nin MLLMs, remain under-explored. This work aims to investigate various\nhallucinations (i.e., object, relation, attribute hallucinations) and mitigate\nthose hallucinatory toxicities in large-scale machine-generated visual\ninstruction datasets. Drawing on the human ability to identify factual errors,\nwe present a novel hallucination detection and elimination framework,\nHalluciDoctor, based on the cross-checking paradigm. We use our framework to\nidentify and eliminate hallucinations in the training data automatically.\nInterestingly, HalluciDoctor also indicates that spurious correlations arising\nfrom long-tail object co-occurrences contribute to hallucinations. Based on\nthat, we execute counterfactual visual instruction expansion to balance data\ndistribution, thereby enhancing MLLMs' resistance to hallucinations.\nComprehensive experiments on hallucination evaluation benchmarks show that our\nmethod successfully mitigates 44.6% hallucinations relatively and maintains\ncompetitive performance compared to LLaVA. The data and code for this paper are\npublicly available. \\url{https://github.com/Yuqifan1117/HalluciDoctor}.\n", "date": "2023-11-22", "categories": []}
{"id": "2311.15529", "title": "Efficient Dataset Distillation via Minimax Diffusion", "abstract": "  Dataset distillation reduces the storage and computational consumption of\ntraining a network by generating a small surrogate dataset that encapsulates\nrich information of the original large-scale one. However, previous\ndistillation methods heavily rely on the sample-wise iterative optimization\nscheme. As the images-per-class (IPC) setting or image resolution grows larger,\nthe necessary computation will demand overwhelming time and resources. In this\nwork, we intend to incorporate generative diffusion techniques for computing\nthe surrogate dataset. Observing that key factors for constructing an effective\nsurrogate dataset are representativeness and diversity, we design additional\nminimax criteria in the generative training to enhance these facets for the\ngenerated images of diffusion models. We present a theoretical model of the\nprocess as hierarchical diffusion control demonstrating the flexibility of the\ndiffusion process to target these criteria without jeopardizing the\nfaithfulness of the sample to the desired distribution. The proposed method\nachieves state-of-the-art validation performance while demanding much less\ncomputational resources. Under the 100-IPC setting on ImageWoof, our method\nrequires less than one-twentieth the distillation time of previous methods, yet\nyields even better performance. Source code and generated data are available in\nhttps://github.com/vimar-gu/MinimaxDiffusion.\n", "date": "2023-11-27", "categories": []}
{"id": "2403.14137", "title": "SynerMix: Synergistic Mixup Solution for Enhanced Intra-Class Cohesion\n  and Inter-Class Separability in Image Classification", "abstract": "  To address the issues of MixUp and its variants (e.g., Manifold MixUp) in\nimage classification tasks-namely, their neglect of mixing within the same\nclass (intra-class mixup) and their inadequacy in enhancing intra-class\ncohesion through their mixing operations-we propose a novel mixup method named\nSynerMix-Intra and, building upon this, introduce a synergistic mixup solution\nnamed SynerMix. SynerMix-Intra specifically targets intra-class mixup to\nbolster intra-class cohesion, a feature not addressed by current mixup methods.\nFor each mini-batch, it leverages feature representations of unaugmented\noriginal images from each class to generate a synthesized feature\nrepresentation through random linear interpolation. All synthesized\nrepresentations are then fed into the classification and loss layers to\ncalculate an average classification loss that significantly enhances\nintra-class cohesion. Furthermore, SynerMix combines SynerMix-Intra with an\nexisting mixup approach (e.g., MixUp, Manifold MixUp), which primarily focuses\non inter-class mixup and has the benefit of enhancing inter-class separability.\nIn doing so, it integrates both inter- and intra-class mixup in a balanced way\nwhile concurrently improving intra-class cohesion and inter-class separability.\nExperimental results on six datasets show that SynerMix achieves a 0.1% to\n3.43% higher accuracy than the best of either MixUp or SynerMix-Intra alone,\naveraging a 1.16% gain. It also surpasses the top-performer of either Manifold\nMixUp or SynerMix-Intra by 0.12% to 5.16%, with an average gain of 1.11%. Given\nthat SynerMix is model-agnostic, it holds significant potential for application\nin other domains where mixup methods have shown promise, such as speech and\ntext classification. Our code is publicly available at:\nhttps://github.com/wxitxy/synermix.git.\n", "date": "2024-03-21", "categories": []}
{"id": "2403.15938", "title": "LlamBERT: Large-scale low-cost data annotation in NLP", "abstract": "  Large Language Models (LLMs), such as GPT-4 and Llama 2, show remarkable\nproficiency in a wide range of natural language processing (NLP) tasks. Despite\ntheir effectiveness, the high costs associated with their use pose a challenge.\nWe present LlamBERT, a hybrid approach that leverages LLMs to annotate a small\nsubset of large, unlabeled databases and uses the results for fine-tuning\ntransformer encoders like BERT and RoBERTa. This strategy is evaluated on two\ndiverse datasets: the IMDb review dataset and the UMLS Meta-Thesaurus. Our\nresults indicate that the LlamBERT approach slightly compromises on accuracy\nwhile offering much greater cost-effectiveness.\n", "date": "2024-03-23", "categories": []}
{"id": "2403.16028", "title": "Exploring the Impact of Dataset Bias on Dataset Distillation", "abstract": "  Dataset Distillation (DD) is a promising technique to synthesize a smaller\ndataset that preserves essential information from the original dataset. This\nsynthetic dataset can serve as a substitute for the original large-scale one,\nand help alleviate the training workload. However, current DD methods typically\noperate under the assumption that the dataset is unbiased, overlooking\npotential bias issues within the dataset itself. To fill in this blank, we\nsystematically investigate the influence of dataset bias on DD. To the best of\nour knowledge, this is the first exploration in the DD domain. Given that there\nare no suitable biased datasets for DD, we first construct two biased datasets,\nCMNIST-DD and CCIFAR10-DD, to establish a foundation for subsequent analysis.\nThen we utilize existing DD methods to generate synthetic datasets on CMNIST-DD\nand CCIFAR10-DD, and evaluate their performance following the standard process.\nExperiments demonstrate that biases present in the original dataset\nsignificantly impact the performance of the synthetic dataset in most cases,\nwhich highlights the necessity of identifying and mitigating biases in the\noriginal datasets during DD. Finally, we reformulate DD within the context of a\nbiased dataset. Our code along with biased datasets are available at\nhttps://github.com/yaolu-zjut/Biased-DD.\n", "date": "2024-03-24", "categories": []}
{"id": "2403.15194", "title": "Your Image is My Video: Reshaping the Receptive Field via Image-To-Video\n  Differentiable AutoAugmentation and Fusion", "abstract": "  The landscape of deep learning research is moving towards innovative\nstrategies to harness the true potential of data. Traditionally, emphasis has\nbeen on scaling model architectures, resulting in large and complex neural\nnetworks, which can be difficult to train with limited computational resources.\nHowever, independently of the model size, data quality (i.e. amount and\nvariability) is still a major factor that affects model generalization. In this\nwork, we propose a novel technique to exploit available data through the use of\nautomatic data augmentation for the tasks of image classification and semantic\nsegmentation. We introduce the first Differentiable Augmentation Search method\n(DAS) to generate variations of images that can be processed as videos.\nCompared to previous approaches, DAS is extremely fast and flexible, allowing\nthe search on very large search spaces in less than a GPU day. Our intuition is\nthat the increased receptive field in the temporal dimension provided by DAS\ncould lead to benefits also to the spatial receptive field. More specifically,\nwe leverage DAS to guide the reshaping of the spatial receptive field by\nselecting task-dependant transformations. As a result, compared to standard\naugmentation alternatives, we improve in terms of accuracy on ImageNet,\nCifar10, Cifar100, Tiny-ImageNet, Pascal-VOC-2012 and CityScapes datasets when\nplugging-in our DAS over different light-weight video backbones.\n", "date": "2024-03-22", "categories": []}
{"id": "2310.10402", "title": "Real-Fake: Effective Training Data Synthesis Through Distribution\n  Matching", "abstract": "  Synthetic training data has gained prominence in numerous learning tasks and\nscenarios, offering advantages such as dataset augmentation, generalization\nevaluation, and privacy preservation. Despite these benefits, the efficiency of\nsynthetic data generated by current methodologies remains inferior when\ntraining advanced deep models exclusively, limiting its practical utility. To\naddress this challenge, we analyze the principles underlying training data\nsynthesis for supervised learning and elucidate a principled theoretical\nframework from the distribution-matching perspective that explicates the\nmechanisms governing synthesis efficacy. Through extensive experiments, we\ndemonstrate the effectiveness of our synthetic data across diverse image\nclassification tasks, both as a replacement for and augmentation to real\ndatasets, while also benefits such as out-of-distribution generalization,\nprivacy preservation, and scalability. Specifically, we achieve 70.9% top1\nclassification accuracy on ImageNet1K when training solely with synthetic data\nequivalent to 1 X the original real data size, which increases to 76.0% when\nscaling up to 10 X synthetic data.\n", "date": "2023-10-16", "categories": []}
{"id": "2403.13233", "title": "Technical Report: Competition Solution For BetterMixture", "abstract": "  In the era of flourishing large-scale models, the challenge of selecting and\noptimizing datasets from the vast and complex sea of data, to enhance the\nperformance of large language models within the constraints of limited\ncomputational resources, has become paramount. This paper details our solution\nfor the BetterMixture challenge, which focuses on the fine-tuning data mixing\nfor large language models. Our approach, which secured third place,\nincorporates data deduplication, low-level and high-level quality filtering,\nand diversity selection. The foundation of our solution is Ke-Data-Juicer, an\nextension of Data-Juicer, demonstrating its robust capabilities in handling and\noptimizing data for large language models.\n", "date": "2024-03-20", "categories": []}
{"id": "2403.13469", "title": "Progressive trajectory matching for medical dataset distillation", "abstract": "  It is essential but challenging to share medical image datasets due to\nprivacy issues, which prohibit building foundation models and knowledge\ntransfer. In this paper, we propose a novel dataset distillation method to\ncondense the original medical image datasets into a synthetic one that\npreserves useful information for building an analysis model without accessing\nthe original datasets. Existing methods tackle only natural images by randomly\nmatching parts of the training trajectories of the model parameters trained by\nthe whole real datasets. However, through extensive experiments on medical\nimage datasets, the training process is extremely unstable and achieves\ninferior distillation results. To solve these barriers, we propose to design a\nnovel progressive trajectory matching strategy to improve the training\nstability for medical image dataset distillation. Additionally, it is observed\nthat improved stability prevents the synthetic dataset diversity and final\nperformance improvements. Therefore, we propose a dynamic overlap mitigation\nmodule that improves the synthetic dataset diversity by dynamically eliminating\nthe overlap across different images and retraining parts of the synthetic\nimages for better convergence. Finally, we propose a new medical image dataset\ndistillation benchmark of various modalities and configurations to promote fair\nevaluations. It is validated that our proposed method achieves 8.33%\nimprovement over previous state-of-the-art methods on average, and 11.7%\nimprovement when ipc=2 (i.e., image per class is 2). Codes and benchmarks will\nbe released.\n", "date": "2024-03-20", "categories": []}
{"id": "2307.03132", "title": "T-MARS: Improving Visual Representations by Circumventing Text Feature\n  Learning", "abstract": "  Large web-sourced multimodal datasets have powered a slew of new methods for\nlearning general-purpose visual representations, advancing the state of the art\nin computer vision and revolutionizing zero- and few-shot recognition. One\ncrucial decision facing practitioners is how, if at all, to curate these\never-larger datasets. For example, the creators of the LAION-5B dataset chose\nto retain only image-caption pairs whose CLIP similarity score exceeded a\ndesignated threshold. In this paper, we propose a new state-of-the-art data\nfiltering approach motivated by our observation that nearly 40% of LAION's\nimages contain text that overlaps significantly with the caption. Intuitively,\nsuch data could be wasteful as it incentivizes models to perform optical\ncharacter recognition rather than learning visual features. However, naively\nremoving all such data could also be wasteful, as it throws away images that\ncontain visual features (in addition to overlapping text). Our simple and\nscalable approach, T-MARS (Text Masking and Re-Scoring), filters out only those\npairs where the text dominates the remaining visual features -- by first\nmasking out the text and then filtering out those with a low CLIP similarity\nscore of the masked image. Experimentally, T-MARS outperforms the top-ranked\nmethod on the \"medium scale\" of DataComp (a data filtering benchmark) by a\nmargin of 6.5% on ImageNet and 4.7% on VTAB. Additionally, our systematic\nevaluation on various data pool sizes from 2M to 64M shows that the accuracy\ngains enjoyed by T-MARS linearly increase as data and compute are scaled\nexponentially. Code is available at https://github.com/locuslab/T-MARS.\n", "date": "2023-07-06", "categories": []}
{"id": "2312.03526", "title": "On the Diversity and Realism of Distilled Dataset: An Efficient Dataset\n  Distillation Paradigm", "abstract": "  Contemporary machine learning requires training large neural networks on\nmassive datasets and thus faces the challenges of high computational demands.\nDataset distillation, as a recent emerging strategy, aims to compress\nreal-world datasets for efficient training. However, this line of research\ncurrently struggle with large-scale and high-resolution datasets, hindering its\npracticality and feasibility. To this end, we re-examine the existing dataset\ndistillation methods and identify three properties required for large-scale\nreal-world applications, namely, realism, diversity, and efficiency. As a\nremedy, we propose RDED, a novel computationally-efficient yet effective data\ndistillation paradigm, to enable both diversity and realism of the distilled\ndata. Extensive empirical results over various neural architectures and\ndatasets demonstrate the advancement of RDED: we can distill the full\nImageNet-1K to a small dataset comprising 10 images per class within 7 minutes,\nachieving a notable 42% top-1 accuracy with ResNet-18 on a single RTX-4090 GPU\n(while the SOTA only achieves 21% but requires 6 hours).\n", "date": "2023-12-06", "categories": []}
{"id": "2403.12429", "title": "TransformMix: Learning Transformation and Mixing Strategies from Data", "abstract": "  Data augmentation improves the generalization power of deep learning models\nby synthesizing more training samples. Sample-mixing is a popular data\naugmentation approach that creates additional data by combining existing\nsamples. Recent sample-mixing methods, like Mixup and Cutmix, adopt simple\nmixing operations to blend multiple inputs. Although such a heuristic approach\nshows certain performance gains in some computer vision tasks, it mixes the\nimages blindly and does not adapt to different datasets automatically. A mixing\nstrategy that is effective for a particular dataset does not often generalize\nwell to other datasets. If not properly configured, the methods may create\nmisleading mixed images, which jeopardize the effectiveness of sample-mixing\naugmentations. In this work, we propose an automated approach, TransformMix, to\nlearn better transformation and mixing augmentation strategies from data. In\nparticular, TransformMix applies learned transformations and mixing masks to\ncreate compelling mixed images that contain correct and important information\nfor the target tasks. We demonstrate the effectiveness of TransformMix on\nmultiple datasets in transfer learning, classification, object detection, and\nknowledge distillation settings. Experimental results show that our method\nachieves better performance as well as efficiency when compared with strong\nsample-mixing baselines.\n", "date": "2024-03-19", "categories": []}
{"id": "2403.12776", "title": "Automated Data Curation for Robust Language Model Fine-Tuning", "abstract": "  Large Language Models have become the de facto approach to\nsequence-to-sequence text generation tasks, but for specialized tasks/domains,\na pretrained LLM lacks specific capabilities to produce accurate or\nwell-formatted responses. Supervised fine-tuning specializes a LLM by training\nit on dataset of example prompts with target responses, but real-world data\ntends to be noisy. While many fine-tuning algorithms exist, here we consider a\n\\emph{data-centric AI} perspective on LLM fine-tuning, studying how to\n\\emph{systematically} curate the training dataset to improve the LLM produced\nvia \\emph{any} fine-tuning algorithm.\n  We introduce an automated data curation pipeline CLEAR (Confidence-based LLM\nEvaluation And Rectification) for instruction tuning datasets, that can be used\nwith any LLM and fine-tuning procedure. CLEAR estimates which training data is\nlow-quality and either filters or corrects it. Automatically identifying which\ndata to filter or correct is done via LLM-derived confidence estimates, to\nensure only confident modifications to the dataset. Unlike existing data\ncuration techniques, CLEAR is a comprehensive framework that can improve a\ndataset (and trained model outputs) without additional fine-tuning\ncomputations. We don't assume access to a stronger LLM than the model being\nfine-tuned (e.g.\\ relying on GPT-4 when fine-tuning GPT-3.5), to see whether\nCLEAR can meaningfully improve the capabilities of any LLM. Experiments reveal\nthat CLEAR consistently improves the performance of fine-tuned models across\nmany datasets and models (like GPT-3.5 and Llama2).\n", "date": "2024-03-19", "categories": []}
{"id": "2403.12803", "title": "DreamDA: Generative Data Augmentation with Diffusion Models", "abstract": "  The acquisition of large-scale, high-quality data is a resource-intensive and\ntime-consuming endeavor. Compared to conventional Data Augmentation (DA)\ntechniques (e.g. cropping and rotation), exploiting prevailing diffusion models\nfor data generation has received scant attention in classification tasks.\nExisting generative DA methods either inadequately bridge the domain gap\nbetween real-world and synthesized images, or inherently suffer from a lack of\ndiversity. To solve these issues, this paper proposes a new\nclassification-oriented framework DreamDA, which enables data synthesis and\nlabel generation by way of diffusion models. DreamDA generates diverse samples\nthat adhere to the original data distribution by considering training images in\nthe original data as seeds and perturbing their reverse diffusion process. In\naddition, since the labels of the generated data may not align with the labels\nof their corresponding seed images, we introduce a self-training paradigm for\ngenerating pseudo labels and training classifiers using the synthesized data.\nExtensive experiments across four tasks and five datasets demonstrate\nconsistent improvements over strong baselines, revealing the efficacy of\nDreamDA in synthesizing high-quality and diverse images with accurate labels.\nOur code will be available at https://github.com/yunxiangfu2001/DreamDA.\n", "date": "2024-03-19", "categories": []}
{"id": "2310.05773", "title": "Towards Lossless Dataset Distillation via Difficulty-Aligned Trajectory\n  Matching", "abstract": "  The ultimate goal of Dataset Distillation is to synthesize a small synthetic\ndataset such that a model trained on this synthetic set will perform equally\nwell as a model trained on the full, real dataset. Until now, no method of\nDataset Distillation has reached this completely lossless goal, in part due to\nthe fact that previous methods only remain effective when the total number of\nsynthetic samples is extremely small. Since only so much information can be\ncontained in such a small number of samples, it seems that to achieve truly\nloss dataset distillation, we must develop a distillation method that remains\neffective as the size of the synthetic dataset grows. In this work, we present\nsuch an algorithm and elucidate why existing methods fail to generate larger,\nhigh-quality synthetic sets. Current state-of-the-art methods rely on\ntrajectory-matching, or optimizing the synthetic data to induce similar\nlong-term training dynamics as the real data. We empirically find that the\ntraining stage of the trajectories we choose to match (i.e., early or late)\ngreatly affects the effectiveness of the distilled dataset. Specifically, early\ntrajectories (where the teacher network learns easy patterns) work well for a\nlow-cardinality synthetic set since there are fewer examples wherein to\ndistribute the necessary information. Conversely, late trajectories (where the\nteacher network learns hard patterns) provide better signals for larger\nsynthetic sets since there are now enough samples to represent the necessary\ncomplex patterns. Based on our findings, we propose to align the difficulty of\nthe generated patterns with the size of the synthetic dataset. In doing so, we\nsuccessfully scale trajectory matching-based methods to larger synthetic\ndatasets, achieving lossless dataset distillation for the very first time. Code\nand distilled datasets are available at https://gzyaftermath.github.io/DATM.\n", "date": "2023-10-09", "categories": []}
{"id": "2311.17950", "title": "Generalized Large-Scale Data Condensation via Various Backbone and\n  Statistical Matching", "abstract": "  The lightweight \"local-match-global\" matching introduced by SRe2L\nsuccessfully creates a distilled dataset with comprehensive information on the\nfull 224x224 ImageNet-1k. However, this one-sided approach is limited to a\nparticular backbone, layer, and statistics, which limits the improvement of the\ngeneralization of a distilled dataset. We suggest that sufficient and various\n\"local-match-global\" matching are more precise and effective than a single one\nand has the ability to create a distilled dataset with richer information and\nbetter generalization. We call this perspective \"generalized matching\" and\npropose Generalized Various Backbone and Statistical Matching (G-VBSM) in this\nwork, which aims to create a synthetic dataset with densities, ensuring\nconsistency with the complete dataset across various backbones, layers, and\nstatistics. As experimentally demonstrated, G-VBSM is the first algorithm to\nobtain strong performance across both small-scale and large-scale datasets.\nSpecifically, G-VBSM achieves a performance of 38.7% on CIFAR-100 with\n128-width ConvNet, 47.6% on Tiny-ImageNet with ResNet18, and 31.4% on the full\n224x224 ImageNet-1k with ResNet18, under images per class (IPC) 10, 50, and 10,\nrespectively. These results surpass all SOTA methods by margins of 3.9%, 6.5%,\nand 10.1%, respectively.\n", "date": "2023-11-29", "categories": []}
{"id": "2402.19282", "title": "WanJuan-CC: A Safe and High-Quality Open-sourced English Webtext Dataset", "abstract": "  This paper presents WanJuan-CC, a safe and high-quality open-sourced English\nwebtext dataset derived from Common Crawl data. The study addresses the\nchallenges of constructing large-scale pre-training datasets for language\nmodels, which require vast amounts of high-quality data. A comprehensive\nprocess was designed to handle Common Crawl data, including extraction,\nheuristic rule filtering, fuzzy deduplication, content safety filtering, and\ndata quality filtering. From approximately 68 billion original English\ndocuments, we obtained 2.22T Tokens of safe data and selected 1.0T Tokens of\nhigh-quality data as part of WanJuan-CC. We have open-sourced 100B Tokens from\nthis dataset. The paper also provides statistical information related to data\nquality, enabling users to select appropriate data according to their needs. To\nevaluate the quality and utility of the dataset, we trained 1B-parameter and\n3B-parameter models using WanJuan-CC and another dataset, RefinedWeb. Results\nshow that WanJuan-CC performs better on validation datasets and downstream\ntasks.\n", "date": "2024-02-29", "categories": []}
{"id": "2403.12040", "title": "Distilling Datasets Into Less Than One Image", "abstract": "  Dataset distillation aims to compress a dataset into a much smaller one so\nthat a model trained on the distilled dataset achieves high accuracy. Current\nmethods frame this as maximizing the distilled classification accuracy for a\nbudget of K distilled images-per-class, where K is a positive integer. In this\npaper, we push the boundaries of dataset distillation, compressing the dataset\ninto less than an image-per-class. It is important to realize that the\nmeaningful quantity is not the number of distilled images-per-class but the\nnumber of distilled pixels-per-dataset. We therefore, propose Poster Dataset\nDistillation (PoDD), a new approach that distills the entire original dataset\ninto a single poster. The poster approach motivates new technical solutions for\ncreating training images and learnable labels. Our method can achieve\ncomparable or better performance with less than an image-per-class compared to\nexisting methods that use one image-per-class. Specifically, our method\nestablishes a new state-of-the-art performance on CIFAR-10, CIFAR-100, and\nCUB200 using as little as 0.3 images-per-class.\n", "date": "2024-03-18", "categories": []}
{"id": "2302.09195", "title": "Data-Efficient Contrastive Self-supervised Learning: Most Beneficial\n  Examples for Supervised Learning Contribute the Least", "abstract": "  Self-supervised learning (SSL) learns high-quality representations from large\npools of unlabeled training data. As datasets grow larger, it becomes crucial\nto identify the examples that contribute the most to learning such\nrepresentations. This enables efficient SSL by reducing the volume of data\nrequired. Nevertheless, quantifying the value of examples for SSL has remained\nan open question. In this work, we address this problem for the first time, by\nproving that examples that contribute the most to contrastive SSL are those\nthat have the most similar augmentations to other examples, in expectation. We\nprovide rigorous guarantees for the generalization performance of contrastive\nlearning on such subsets. Through extensive experiments, we show that we can\nsafely exclude 20% of examples from CIFAR100 and 40% from STL10 and\nTinyImageNet, without affecting downstream task performance. In general,\nsubsets selected by our method outperform random subsets by over 3% across\nthese datasets. Interestingly, we also discover the subsets that contribute the\nmost to contrastive learning are those that contribute the least to supervised\nlearning. Code available at\nhttps://github.com/bigml-cs-ucla/sas-data-efficient-contrastive-learning.\n", "date": "2023-02-18", "categories": []}
{"id": "2310.00902", "title": "DataInf: Efficiently Estimating Data Influence in LoRA-tuned LLMs and\n  Diffusion Models", "abstract": "  Quantifying the impact of training data points is crucial for understanding\nthe outputs of machine learning models and for improving the transparency of\nthe AI pipeline. The influence function is a principled and popular data\nattribution method, but its computational cost often makes it challenging to\nuse. This issue becomes more pronounced in the setting of large language models\nand text-to-image models. In this work, we propose DataInf, an efficient\ninfluence approximation method that is practical for large-scale generative AI\nmodels. Leveraging an easy-to-compute closed-form expression, DataInf\noutperforms existing influence computation algorithms in terms of computational\nand memory efficiency. Our theoretical analysis shows that DataInf is\nparticularly well-suited for parameter-efficient fine-tuning techniques such as\nLoRA. Through systematic empirical evaluations, we show that DataInf accurately\napproximates influence scores and is orders of magnitude faster than existing\nmethods. In applications to RoBERTa-large, Llama-2-13B-chat, and\nstable-diffusion-v1.5 models, DataInf effectively identifies the most\ninfluential fine-tuning examples better than other approximate influence\nscores. Moreover, it can help to identify which data points are mislabeled.\n", "date": "2023-10-02", "categories": []}
{"id": "2310.04780", "title": "IPMix: Label-Preserving Data Augmentation Method for Training Robust\n  Classifiers", "abstract": "  Data augmentation has been proven effective for training high-accuracy\nconvolutional neural network classifiers by preventing overfitting. However,\nbuilding deep neural networks in real-world scenarios requires not only high\naccuracy on clean data but also robustness when data distributions shift. While\nprior methods have proposed that there is a trade-off between accuracy and\nrobustness, we propose IPMix, a simple data augmentation approach to improve\nrobustness without hurting clean accuracy. IPMix integrates three levels of\ndata augmentation (image-level, patch-level, and pixel-level) into a coherent\nand label-preserving technique to increase the diversity of training data with\nlimited computational overhead. To further improve the robustness, IPMix\nintroduces structural complexity at different levels to generate more diverse\nimages and adopts the random mixing method for multi-scale information fusion.\nExperiments demonstrate that IPMix outperforms state-of-the-art corruption\nrobustness on CIFAR-C and ImageNet-C. In addition, we show that IPMix also\nsignificantly improves the other safety measures, including robustness to\nadversarial perturbations, calibration, prediction consistency, and anomaly\ndetection, achieving state-of-the-art or comparable results on several\nbenchmarks, including ImageNet-R, ImageNet-A, and ImageNet-O.\n", "date": "2023-10-07", "categories": []}
{"id": "2211.14611", "title": "The Principles of Data-Centric AI (DCAI)", "abstract": "  Data is a crucial infrastructure to how artificial intelligence (AI) systems\nlearn. However, these systems to date have been largely model-centric, putting\na premium on the model at the expense of the data quality. Data quality issues\nbeset the performance of AI systems, particularly in downstream deployments and\nin real-world applications. Data-centric AI (DCAI) as an emerging concept\nbrings data, its quality and its dynamism to the forefront in considerations of\nAI systems through an iterative and systematic approach. As one of the first\noverviews, this article brings together data-centric perspectives and concepts\nto outline the foundations of DCAI. It specifically formulates six guiding\nprinciples for researchers and practitioners and gives direction for future\nadvancement of DCAI.\n", "date": "2022-11-26", "categories": []}
{"id": "2401.04578", "title": "Effective pruning of web-scale datasets based on complexity of concept\n  clusters", "abstract": "  Utilizing massive web-scale datasets has led to unprecedented performance\ngains in machine learning models, but also imposes outlandish compute\nrequirements for their training. In order to improve training and data\nefficiency, we here push the limits of pruning large-scale multimodal datasets\nfor training CLIP-style models. Today's most effective pruning method on\nImageNet clusters data samples into separate concepts according to their\nembedding and prunes away the most prototypical samples. We scale this approach\nto LAION and improve it by noting that the pruning rate should be\nconcept-specific and adapted to the complexity of the concept. Using a simple\nand intuitive complexity measure, we are able to reduce the training cost to a\nquarter of regular training. By filtering from the LAION dataset, we find that\ntraining on a smaller set of high-quality data can lead to higher performance\nwith significantly lower training costs. More specifically, we are able to\noutperform the LAION-trained OpenCLIP-ViT-B32 model on ImageNet zero-shot\naccuracy by 1.1p.p. while only using 27.7% of the data and training compute.\nDespite a strong reduction in training cost, we also see improvements on\nImageNet dist. shifts, retrieval tasks and VTAB. On the DataComp Medium\nbenchmark, we achieve a new state-of-the-art\nImagehttps://info.arxiv.org/help/prep#commentsNet zero-shot accuracy and a\ncompetitive average zero-shot accuracy on 38 evaluation tasks.\n", "date": "2024-01-09", "categories": []}
{"id": "2403.07142", "title": "One Category One Prompt: Dataset Distillation using Diffusion Models", "abstract": "  The extensive amounts of data required for training deep neural networks pose\nsignificant challenges on storage and transmission fronts. Dataset distillation\nhas emerged as a promising technique to condense the information of massive\ndatasets into a much smaller yet representative set of synthetic samples.\nHowever, traditional dataset distillation approaches often struggle to scale\neffectively with high-resolution images and more complex architectures due to\nthe limitations in bi-level optimization. Recently, several works have proposed\nexploiting knowledge distillation with decoupled optimization schemes to scale\nup dataset distillation. Although these methods effectively address the\nscalability issue, they rely on extensive image augmentations requiring the\nstorage of soft labels for augmented images. In this paper, we introduce\nDataset Distillation using Diffusion Models (D3M) as a novel paradigm for\ndataset distillation, leveraging recent advancements in generative\ntext-to-image foundation models. Our approach utilizes textual inversion, a\ntechnique for fine-tuning text-to-image generative models, to create concise\nand informative representations for large datasets. By employing these learned\ntext prompts, we can efficiently store and infer new samples for introducing\ndata variability within a fixed memory budget. We show the effectiveness of our\nmethod through extensive experiments across various computer vision benchmark\ndatasets with different memory budgets.\n", "date": "2024-03-11", "categories": []}
{"id": "2310.02110", "title": "Sieve: Multimodal Dataset Pruning Using Image Captioning Models", "abstract": "  Vision-Language Models (VLMs) are pretrained on large, diverse, and noisy\nweb-crawled datasets. This underscores the critical need for dataset pruning,\nas the quality of these datasets is strongly correlated with the performance of\nVLMs on downstream tasks. Using CLIPScore from a pretrained model to only train\nmodels using highly-aligned samples is one of the most successful methods for\npruning. We argue that this approach suffers from multiple limitations\nincluding: false positives and negatives due to CLIP's pretraining on noisy\nlabels. We propose a pruning signal, Sieve, that employs synthetic captions\ngenerated by image-captioning models pretrained on small, diverse, and\nwell-aligned image-text pairs to evaluate the alignment of noisy image-text\npairs. To bridge the gap between the limited diversity of generated captions\nand the high diversity of alternative text (alt-text), we estimate the semantic\ntextual similarity in the embedding space of a language model pretrained on\nunlabeled text corpus. Using DataComp, a multimodal dataset filtering\nbenchmark, when evaluating on 38 downstream tasks, our pruning approach,\nsurpasses CLIPScore by 2.6\\% and 1.7\\% on medium and large scale respectively.\nIn addition, on retrieval tasks, Sieve leads to a significant improvement of\n2.7% and 4.5% on medium and large scale respectively.\n", "date": "2023-10-03", "categories": []}
{"id": "2403.06139", "title": "Fine-grainedly Synthesize Streaming Data Based On Large Language Models\n  With Graph Structure Understanding For Data Sparsity", "abstract": "  Due to the sparsity of user data, sentiment analysis on user reviews in\ne-commerce platforms often suffers from poor performance, especially when faced\nwith extremely sparse user data or long-tail labels. Recently, the emergence of\nLLMs has introduced new solutions to such problems by leveraging graph\nstructures to generate supplementary user profiles. However, previous\napproaches have not fully utilized the graph understanding capabilities of LLMs\nand have struggled to adapt to complex streaming data environments. In this\nwork, we propose a fine-grained streaming data synthesis framework that\ncategorizes sparse users into three categories: Mid-tail, Long-tail, and\nExtreme. Specifically, we design LLMs to comprehensively understand three key\ngraph elements in streaming data, including Local-global Graph Understanding,\nSecond-Order Relationship Extraction, and Product Attribute Understanding,\nwhich enables the generation of high-quality synthetic data to effectively\naddress sparsity across different categories. Experimental results on three\nreal datasets demonstrate significant performance improvements, with\nsynthesized data contributing to MSE reductions of 45.85%, 3.16%, and 62.21%,\nrespectively.\n", "date": "2024-03-10", "categories": []}
{"id": "2401.04305", "title": "Advancing Deep Active Learning & Data Subset Selection: Unifying\n  Principles with Information-Theory Intuitions", "abstract": "  At its core, this thesis aims to enhance the practicality of deep learning by\nimproving the label and training efficiency of deep learning models. To this\nend, we investigate data subset selection techniques, specifically active\nlearning and active sampling, grounded in information-theoretic principles.\nActive learning improves label efficiency, while active sampling enhances\ntraining efficiency. Supervised deep learning models often require extensive\ntraining with labeled data. Label acquisition can be expensive and\ntime-consuming, and training large models is resource-intensive, hindering the\nadoption outside academic research and \"big tech.\" Existing methods for data\nsubset selection in deep learning often rely on heuristics or lack a principled\ninformation-theoretic foundation. In contrast, this thesis examines several\nobjectives for data subset selection and their applications within deep\nlearning, striving for a more principled approach inspired by information\ntheory. We begin by disentangling epistemic and aleatoric uncertainty in single\nforward-pass deep neural networks, which provides helpful intuitions and\ninsights into different forms of uncertainty and their relevance for data\nsubset selection. We then propose and investigate various approaches for active\nlearning and data subset selection in (Bayesian) deep learning. Finally, we\nrelate various existing and proposed approaches to approximations of\ninformation quantities in weight or prediction space. Underpinning this work is\na principled and practical notation for information-theoretic quantities that\nincludes both random variables and observed outcomes. This thesis demonstrates\nthe benefits of working from a unified perspective and highlights the potential\nimpact of our contributions to the practical application of deep learning.\n", "date": "2024-01-09", "categories": []}
{"id": "2402.01943", "title": "Precedence-Constrained Winter Value for Effective Graph Data Valuation", "abstract": "  Data valuation is essential for quantifying data's worth, aiding in assessing\ndata quality and determining fair compensation. While existing data valuation\nmethods have proven effective in evaluating the value of Euclidean data, they\nface limitations when applied to the increasingly popular graph-structured\ndata. Particularly, graph data valuation introduces unique challenges,\nprimarily stemming from the intricate dependencies among nodes and the\nexponential growth in value estimation costs. To address the challenging\nproblem of graph data valuation, we put forth an innovative solution,\nPrecedence-Constrained Winter (PC-Winter) Value, to account for the complex\ngraph structure. Furthermore, we develop a variety of strategies to address the\ncomputational challenges and enable efficient approximation of PC-Winter.\nExtensive experiments demonstrate the effectiveness of PC-Winter across diverse\ndatasets and tasks.\n", "date": "2024-02-02", "categories": []}
{"id": "2403.05174", "title": "VTruST: Controllable value function based subset selection for\n  Data-Centric Trustworthy AI", "abstract": "  Trustworthy AI is crucial to the widespread adoption of AI in high-stakes\napplications with fairness, robustness, and accuracy being some of the key\ntrustworthiness metrics. In this work, we propose a controllable framework for\ndata-centric trustworthy AI (DCTAI)- VTruST, that allows users to control the\ntrade-offs between the different trustworthiness metrics of the constructed\ntraining datasets. A key challenge in implementing an efficient DCTAI framework\nis to design an online value-function-based training data subset selection\nalgorithm. We pose the training data valuation and subset selection problem as\nan online sparse approximation formulation. We propose a novel online version\nof the Orthogonal Matching Pursuit (OMP) algorithm for solving this problem.\nExperimental results show that VTruST outperforms the state-of-the-art\nbaselines on social, image, and scientific datasets. We also show that the data\nvalues generated by VTruST can provide effective data-centric explanations for\ndifferent trustworthiness metrics.\n", "date": "2024-03-08", "categories": []}
{"id": "2403.04120", "title": "A data-centric approach to class-specific bias in image data\n  augmentation", "abstract": "  Data augmentation (DA) enhances model generalization in computer vision but\nmay introduce biases, impacting class accuracy unevenly. Our study extends this\ninquiry, examining DA's class-specific bias across various datasets, including\nthose distinct from ImageNet, through random cropping. We evaluated this\nphenomenon with ResNet50, EfficientNetV2S, and SWIN ViT, discovering that while\nresidual models showed similar bias effects, Vision Transformers exhibited\ngreater robustness or altered dynamics. This suggests a nuanced approach to\nmodel selection, emphasizing bias mitigation. We also refined a \"data\naugmentation robustness scouting\" method to manage DA-induced biases more\nefficiently, reducing computational demands significantly (training 112 models\ninstead of 1860; a reduction of factor 16.2) while still capturing essential\nbias trends.\n", "date": "2024-03-07", "categories": []}
{"id": "2403.04551", "title": "Dissecting Sample Hardness: A Fine-Grained Analysis of Hardness\n  Characterization Methods for Data-Centric AI", "abstract": "  Characterizing samples that are difficult to learn from is crucial to\ndeveloping highly performant ML models. This has led to numerous Hardness\nCharacterization Methods (HCMs) that aim to identify \"hard\" samples. However,\nthere is a lack of consensus regarding the definition and evaluation of\n\"hardness\". Unfortunately, current HCMs have only been evaluated on specific\ntypes of hardness and often only qualitatively or with respect to downstream\nperformance, overlooking the fundamental quantitative identification task. We\naddress this gap by presenting a fine-grained taxonomy of hardness types.\nAdditionally, we propose the Hardness Characterization Analysis Toolkit\n(H-CAT), which supports comprehensive and quantitative benchmarking of HCMs\nacross the hardness taxonomy and can easily be extended to new HCMs, hardness\ntypes, and datasets. We use H-CAT to evaluate 13 different HCMs across 8\nhardness types. This comprehensive evaluation encompassing over 14K setups\nuncovers strengths and weaknesses of different HCMs, leading to practical tips\nto guide HCM selection and future development. Our findings highlight the need\nfor more comprehensive HCM evaluation, while we hope our hardness taxonomy and\ntoolkit will advance the principled evaluation and uptake of data-centric AI\nmethods.\n", "date": "2024-03-07", "categories": []}
{"id": "2403.02677", "title": "Finetuned Multimodal Language Models Are High-Quality Image-Text Data\n  Filters", "abstract": "  We propose a novel framework for filtering image-text data by leveraging\nfine-tuned Multimodal Language Models (MLMs). Our approach outperforms\npredominant filtering methods (e.g., CLIPScore) via integrating the recent\nadvances in MLMs. We design four distinct yet complementary metrics to\nholistically measure the quality of image-text data. A new pipeline is\nestablished to construct high-quality instruction data for fine-tuning MLMs as\ndata filters. Comparing with CLIPScore, our MLM filters produce more precise\nand comprehensive scores that directly improve the quality of filtered data and\nboost the performance of pre-trained models. We achieve significant\nimprovements over CLIPScore on popular foundation models (i.e., CLIP and BLIP2)\nand various downstream tasks. Our MLM filter can generalize to different models\nand tasks, and be used as a drop-in replacement for CLIPScore. An additional\nablation study is provided to verify our design choices for the MLM filter.\n", "date": "2024-03-05", "categories": []}
{"id": "2211.05764", "title": "DC-Check: A Data-Centric AI checklist to guide the development of\n  reliable machine learning systems", "abstract": "  While there have been a number of remarkable breakthroughs in machine\nlearning (ML), much of the focus has been placed on model development. However,\nto truly realize the potential of machine learning in real-world settings,\nadditional aspects must be considered across the ML pipeline. Data-centric AI\nis emerging as a unifying paradigm that could enable such reliable end-to-end\npipelines. However, this remains a nascent area with no standardized framework\nto guide practitioners to the necessary data-centric considerations or to\ncommunicate the design of data-centric driven ML systems. To address this gap,\nwe propose DC-Check, an actionable checklist-style framework to elicit\ndata-centric considerations at different stages of the ML pipeline: Data,\nTraining, Testing, and Deployment. This data-centric lens on development aims\nto promote thoughtfulness and transparency prior to system development.\nAdditionally, we highlight specific data-centric AI challenges and research\nopportunities. DC-Check is aimed at both practitioners and researchers to guide\nday-to-day development. As such, to easily engage with and use DC-Check and\nassociated resources, we provide a DC-Check companion website\n(https://www.vanderschaar-lab.com/dc-check/). The website will also serve as an\nupdated resource as methods and tooling evolve over time.\n", "date": "2022-11-09", "categories": []}
{"id": "2401.15422", "title": "A Survey on Data Augmentation in Large Model Era", "abstract": "  Large models, encompassing large language and diffusion models, have shown\nexceptional promise in approximating human-level intelligence, garnering\nsignificant interest from both academic and industrial spheres. However, the\ntraining of these large models necessitates vast quantities of high-quality\ndata, and with continuous updates to these models, the existing reservoir of\nhigh-quality data may soon be depleted. This challenge has catalyzed a surge in\nresearch focused on data augmentation methods. Leveraging large models, these\ndata augmentation techniques have outperformed traditional approaches. This\npaper offers an exhaustive review of large model-driven data augmentation\nmethods, adopting a comprehensive perspective. We begin by establishing a\nclassification of relevant studies into three main categories: image\naugmentation, text augmentation, and paired data augmentation. Following this,\nwe delve into various data post-processing techniques pertinent to large\nmodel-based data augmentation. Our discussion then expands to encompass the\narray of applications for these data augmentation methods within natural\nlanguage processing, computer vision, and audio signal processing. We proceed\nto evaluate the successes and limitations of large model-based data\naugmentation across different scenarios. Concluding our review, we highlight\nprospective challenges and avenues for future exploration in the field of data\naugmentation. Our objective is to furnish researchers with critical insights,\nultimately contributing to the advancement of more sophisticated large models.\nWe consistently maintain the related open-source materials at:\nhttps://github.com/MLGroup-JLU/LLM-data-aug-survey.\n", "date": "2024-01-27", "categories": []}
{"id": "2401.16193", "title": "Contributing Dimension Structure of Deep Feature for Coreset Selection", "abstract": "  Coreset selection seeks to choose a subset of crucial training samples for\nefficient learning. It has gained traction in deep learning, particularly with\nthe surge in training dataset sizes. Sample selection hinges on two main\naspects: a sample's representation in enhancing performance and the role of\nsample diversity in averting overfitting. Existing methods typically measure\nboth the representation and diversity of data based on similarity metrics, such\nas L2-norm. They have capably tackled representation via distribution matching\nguided by the similarities of features, gradients, or other information between\ndata. However, the results of effectively diverse sample selection are mired in\nsub-optimality. This is because the similarity metrics usually simply aggregate\ndimension similarities without acknowledging disparities among the dimensions\nthat significantly contribute to the final similarity. As a result, they fall\nshort of adequately capturing diversity. To address this, we propose a\nfeature-based diversity constraint, compelling the chosen subset to exhibit\nmaximum diversity. Our key lies in the introduction of a novel Contributing\nDimension Structure (CDS) metric. Different from similarity metrics that\nmeasure the overall similarity of high-dimensional features, our CDS metric\nconsiders not only the reduction of redundancy in feature dimensions, but also\nthe difference between dimensions that contribute significantly to the final\nsimilarity. We reveal that existing methods tend to favor samples with similar\nCDS, leading to a reduced variety of CDS types within the coreset and\nsubsequently hindering model performance. In response, we enhance the\nperformance of five classical selection methods by integrating the CDS\nconstraint. Our experiments on three datasets demonstrate the general\neffectiveness of the proposed method in boosting existing methods.\n", "date": "2024-01-29", "categories": []}
{"id": "2403.00999", "title": "Distributional Dataset Distillation with Subtask Decomposition", "abstract": "  What does a neural network learn when training from a task-specific dataset?\nSynthesizing this knowledge is the central idea behind Dataset Distillation,\nwhich recent work has shown can be used to compress large datasets into a small\nset of input-label pairs ($\\textit{prototypes}$) that capture essential aspects\nof the original dataset. In this paper, we make the key observation that\nexisting methods distilling into explicit prototypes are very often suboptimal,\nincurring in unexpected storage cost from distilled labels. In response, we\npropose $\\textit{Distributional Dataset Distillation}$ (D3), which encodes the\ndata using minimal sufficient per-class statistics and paired with a decoder,\nwe distill dataset into a compact distributional representation that is more\nmemory-efficient compared to prototype-based methods. To scale up the process\nof learning these representations, we propose $\\textit{Federated\ndistillation}$, which decomposes the dataset into subsets, distills them in\nparallel using sub-task experts and then re-aggregates them. We thoroughly\nevaluate our algorithm on a three-dimensional metric and show that our method\nachieves state-of-the-art results on TinyImageNet and ImageNet-1K.\nSpecifically, we outperform the prior art by $6.9\\%$ on ImageNet-1K under the\nstorage budget of 2 images per class.\n", "date": "2024-03-01", "categories": []}
{"id": "2403.01832", "title": "Model-Based Data-Centric AI: Bridging the Divide Between Academic Ideals\n  and Industrial Pragmatism", "abstract": "  This paper delves into the contrasting roles of data within academic and\nindustrial spheres, highlighting the divergence between Data-Centric AI and\nModel-Agnostic AI approaches. We argue that while Data-Centric AI focuses on\nthe primacy of high-quality data for model performance, Model-Agnostic AI\nprioritizes algorithmic flexibility, often at the expense of data quality\nconsiderations. This distinction reveals that academic standards for data\nquality frequently do not meet the rigorous demands of industrial applications,\nleading to potential pitfalls in deploying academic models in real-world\nsettings. Through a comprehensive analysis, we address these disparities,\npresenting both the challenges they pose and strategies for bridging the gap.\nFurthermore, we propose a novel paradigm: Model-Based Data-Centric AI, which\naims to reconcile these differences by integrating model considerations into\ndata optimization processes. This approach underscores the necessity for\nevolving data requirements that are sensitive to the nuances of both academic\nresearch and industrial deployment. By exploring these discrepancies, we aim to\nfoster a more nuanced understanding of data's role in AI development and\nencourage a convergence of academic and industrial standards to enhance AI's\nreal-world applicability.\n", "date": "2024-03-04", "categories": []}
{"id": "2310.11046", "title": "Fast Graph Condensation with Structure-based Neural Tangent Kernel", "abstract": "  The rapid development of Internet technology has given rise to a vast amount\nof graph-structured data. Graph Neural Networks (GNNs), as an effective method\nfor various graph mining tasks, incurs substantial computational resource costs\nwhen dealing with large-scale graph data. A data-centric manner solution is\nproposed to condense the large graph dataset into a smaller one without\nsacrificing the predictive performance of GNNs. However, existing efforts\ncondense graph-structured data through a computational intensive bi-level\noptimization architecture also suffer from massive computation costs. In this\npaper, we propose reforming the graph condensation problem as a Kernel Ridge\nRegression (KRR) task instead of iteratively training GNNs in the inner loop of\nbi-level optimization. More specifically, We propose a novel dataset\ncondensation framework (GC-SNTK) for graph-structured data, where a\nStructure-based Neural Tangent Kernel (SNTK) is developed to capture the\ntopology of graph and serves as the kernel function in KRR paradigm.\nComprehensive experiments demonstrate the effectiveness of our proposed model\nin accelerating graph condensation while maintaining high prediction\nperformance. The source code is available on\nhttps://github.com/WANGLin0126/GCSNTK.\n", "date": "2023-10-17", "categories": []}
{"id": "2311.08675", "title": "Refined Coreset Selection: Towards Minimal Coreset Size under Model\n  Performance Constraints", "abstract": "  Coreset selection is powerful in reducing computational costs and\naccelerating data processing for deep learning algorithms. It strives to\nidentify a small subset from large-scale data, so that training only on the\nsubset practically performs on par with full data. Practitioners regularly\ndesire to identify the smallest possible coreset in realistic scenes while\nmaintaining comparable model performance, to minimize costs and maximize\nacceleration. Motivated by this desideratum, for the first time, we pose the\nproblem of refined coreset selection, in which the minimal coreset size under\nmodel performance constraints is explored. Moreover, to address this problem,\nwe propose an innovative method, which maintains optimization priority order\nover the model performance and coreset size, and efficiently optimizes them in\nthe coreset selection procedure. Theoretically, we provide the convergence\nguarantee of the proposed method. Empirically, extensive experiments confirm\nits superiority compared with previous strategies, often yielding better model\nperformance with smaller coreset sizes.\n", "date": "2023-11-15", "categories": []}
{"id": "2402.17599", "title": "DAGnosis: Localized Identification of Data Inconsistencies using\n  Structures", "abstract": "  Identification and appropriate handling of inconsistencies in data at\ndeployment time is crucial to reliably use machine learning models. While\nrecent data-centric methods are able to identify such inconsistencies with\nrespect to the training set, they suffer from two key limitations: (1)\nsuboptimality in settings where features exhibit statistical independencies,\ndue to their usage of compressive representations and (2) lack of localization\nto pin-point why a sample might be flagged as inconsistent, which is important\nto guide future data collection. We solve these two fundamental limitations\nusing directed acyclic graphs (DAGs) to encode the training set's features\nprobability distribution and independencies as a structure. Our method, called\nDAGnosis, leverages these structural interactions to bring valuable and\ninsightful data-centric conclusions. DAGnosis unlocks the localization of the\ncauses of inconsistencies on a DAG, an aspect overlooked by previous\napproaches. Moreover, we show empirically that leveraging these interactions\n(1) leads to more accurate conclusions in detecting inconsistencies, as well as\n(2) provides more detailed insights into why some samples are flagged.\n", "date": "2024-02-26", "categories": []}
{"id": "2402.17327", "title": "Data-Efficient Learning via Clustering-Based Sensitivity Sampling:\n  Foundation Models and Beyond", "abstract": "  We study the data selection problem, whose aim is to select a small\nrepresentative subset of data that can be used to efficiently train a machine\nlearning model. We present a new data selection approach based on $k$-means\nclustering and sensitivity sampling. Assuming access to an embedding\nrepresentation of the data with respect to which the model loss is H\\\"older\ncontinuous, our approach provably allows selecting a set of ``typical'' $k +\n1/\\varepsilon^2$ elements whose average loss corresponds to the average loss of\nthe whole dataset, up to a multiplicative $(1\\pm\\varepsilon)$ factor and an\nadditive $\\varepsilon \\lambda \\Phi_k$, where $\\Phi_k$ represents the $k$-means\ncost for the input embeddings and $\\lambda$ is the H\\\"older constant.\n  We furthermore demonstrate the performance and scalability of our approach on\nfine-tuning foundation models and show that it outperforms state-of-the-art\nmethods. We also show how it can be applied on linear regression, leading to a\nnew sampling strategy that surprisingly matches the performances of leverage\nscore sampling, while being conceptually simpler and more scalable.\n", "date": "2024-02-27", "categories": []}
{"id": "2312.15927", "title": "M3D: Dataset Condensation by Minimizing Maximum Mean Discrepancy", "abstract": "  Training state-of-the-art (SOTA) deep models often requires extensive data,\nresulting in substantial training and storage costs. To address these\nchallenges, dataset condensation has been developed to learn a small synthetic\nset that preserves essential information from the original large-scale dataset.\nNowadays, optimization-oriented methods have been the primary method in the\nfield of dataset condensation for achieving SOTA results. However, the bi-level\noptimization process hinders the practical application of such methods to\nrealistic and larger datasets. To enhance condensation efficiency, previous\nworks proposed Distribution-Matching (DM) as an alternative, which\nsignificantly reduces the condensation cost. Nonetheless, current DM-based\nmethods still yield less comparable results to SOTA optimization-oriented\nmethods. In this paper, we argue that existing DM-based methods overlook the\nhigher-order alignment of the distributions, which may lead to sub-optimal\nmatching results. Inspired by this, we present a novel DM-based method named\nM3D for dataset condensation by Minimizing the Maximum Mean Discrepancy between\nfeature representations of the synthetic and real images. By embedding their\ndistributions in a reproducing kernel Hilbert space, we align all orders of\nmoments of the distributions of real and synthetic images, resulting in a more\ngeneralized condensed set. Notably, our method even surpasses the SOTA\noptimization-oriented method IDC on the high-resolution ImageNet dataset.\nExtensive analysis is conducted to verify the effectiveness of the proposed\nmethod. Source codes are available at https://github.com/Hansong-Zhang/M3D.\n", "date": "2023-12-26", "categories": []}
{"id": "2402.14888", "title": "Efficient data selection employing Semantic Similarity-based Graph\n  Structures for model training", "abstract": "  Recent developments in natural language processing (NLP) have highlighted the\nneed for substantial amounts of data for models to capture textual information\naccurately. This raises concerns regarding the computational resources and time\nrequired for training such models. This paper introduces Semantics for data\nSAliency in Model performance Estimation (SeSaME). It is an efficient data\nsampling mechanism solely based on textual information without passing the data\nthrough a compute-heavy model or other intensive pre-processing\ntransformations. The application of this approach is demonstrated in the use\ncase of low-resource automated speech recognition (ASR) models, which\nexcessively rely on text-to-speech (TTS) calls when using augmented data.\nSeSaME learns to categorize new incoming data points into speech recognition\ndifficulty buckets by employing semantic similarity-based graph structures and\ndiscrete ASR information from homophilous neighbourhoods through message\npassing. The results indicate reliable projections of ASR performance, with a\n93% accuracy increase when using the proposed method compared to random\npredictions, bringing non-trivial information on the impact of textual\nrepresentations in speech models. Furthermore, a series of experiments show\nboth the benefits and challenges of using the ASR information on incoming data\nto fine-tune the model. We report a 7% drop in validation loss compared to\nrandom sampling, 7% WER drop with non-local aggregation when evaluating against\na highly difficult dataset, and 1.8% WER drop with local aggregation and high\nsemantic similarity between datasets.\n", "date": "2024-02-22", "categories": []}
{"id": "2402.13482", "title": "Retrieval-Augmented Data Augmentation for Low-Resource Domain Tasks", "abstract": "  Despite large successes of recent language models on diverse tasks, they\nsuffer from severe performance degeneration in low-resource settings with\nlimited training data available. Many existing works tackle this problem by\ngenerating synthetic data from the training data and then training models on\nthem, recently using Large Language Models (LLMs). However, in low-resource\nsettings, the amount of seed data samples to use for data augmentation is very\nsmall, which makes generated samples suboptimal and less diverse. To tackle\nthis challenge, we propose a novel method that augments training data by\nincorporating a wealth of examples from other datasets, along with the given\ntraining data. Specifically, we first retrieve the relevant instances from\nother datasets, such as their input-output pairs or contexts, based on their\nsimilarities with the given seed data, and then prompt LLMs to generate new\nsamples with the contextual information within and across the original and\nretrieved samples. This approach can ensure that the generated data is not only\nrelevant but also more diverse than what could be achieved using the limited\nseed data alone. We validate our proposed Retrieval-Augmented Data Augmentation\n(RADA) framework on multiple datasets under low-resource settings of training\nand test-time data augmentation scenarios, on which it outperforms existing\nLLM-powered data augmentation baselines.\n", "date": "2024-02-21", "categories": []}
{"id": "2402.13007", "title": "Improve Cross-Architecture Generalization on Dataset Distillation", "abstract": "  Dataset distillation, a pragmatic approach in machine learning, aims to\ncreate a smaller synthetic dataset from a larger existing dataset. However,\nexisting distillation methods primarily adopt a model-based paradigm, where the\nsynthetic dataset inherits model-specific biases, limiting its generalizability\nto alternative models. In response to this constraint, we propose a novel\nmethodology termed \"model pool\". This approach involves selecting models from a\ndiverse model pool based on a specific probability distribution during the data\ndistillation process. Additionally, we integrate our model pool with the\nestablished knowledge distillation approach and apply knowledge distillation to\nthe test process of the distilled dataset. Our experimental results validate\nthe effectiveness of the model pool approach across a range of existing models\nwhile testing, demonstrating superior performance compared to existing\nmethodologies.\n", "date": "2024-02-20", "categories": []}
{"id": "2402.12010", "title": "Training Green AI Models Using Elite Samples", "abstract": "  The substantial increase in AI model training has considerable environmental\nimplications, mandating more energy-efficient and sustainable AI practices. On\nthe one hand, data-centric approaches show great potential towards training\nenergy-efficient AI models. On the other hand, instance selection methods\ndemonstrate the capability of training AI models with minimised training sets\nand negligible performance degradation. Despite the growing interest in both\ntopics, the impact of data-centric training set selection on energy efficiency\nremains to date unexplored. This paper presents an evolutionary-based sampling\nframework aimed at (i) identifying elite training samples tailored for datasets\nand model pairs, (ii) comparing model performance and energy efficiency gains\nagainst typical model training practice, and (iii) investigating the\nfeasibility of this framework for fostering sustainable model training\npractices. To evaluate the proposed framework, we conducted an empirical\nexperiment including 8 commonly used AI classification models and 25 publicly\navailable datasets. The results showcase that by considering 10% elite training\nsamples, the models' performance can show a 50% improvement and remarkable\nenergy savings of 98% compared to the common training practice.\n", "date": "2024-02-19", "categories": []}
{"id": "2206.13508", "title": "Data Augmentation techniques in time series domain: A survey and\n  taxonomy", "abstract": "  With the latest advances in Deep Learning-based generative models, it has not\ntaken long to take advantage of their remarkable performance in the area of\ntime series. Deep neural networks used to work with time series heavily depend\non the size and consistency of the datasets used in training. These features\nare not usually abundant in the real world, where they are usually limited and\noften have constraints that must be guaranteed. Therefore, an effective way to\nincrease the amount of data is by using Data Augmentation techniques, either by\nadding noise or permutations and by generating new synthetic data. This work\nsystematically reviews the current state-of-the-art in the area to provide an\noverview of all available algorithms and proposes a taxonomy of the most\nrelevant research. The efficiency of the different variants will be evaluated\nas a central part of the process, as well as the different metrics to evaluate\nthe performance and the main problems concerning each model will be analysed.\nThe ultimate aim of this study is to provide a summary of the evolution and\nperformance of areas that produce better results to guide future researchers in\nthis field.\n", "date": "2022-06-25", "categories": []}
{"id": "2402.10423", "title": "Connect the dots: Dataset Condensation, Differential Privacy, and\n  Adversarial Uncertainty", "abstract": "  Our work focuses on understanding the underpinning mechanism of dataset\ncondensation by drawing connections with ($\\epsilon$, $\\delta$)-differential\nprivacy where the optimal noise, $\\epsilon$, is chosen by adversarial\nuncertainty \\cite{Grining2017}. We can answer the question about the inner\nworkings of the dataset condensation procedure. Previous work \\cite{dong2022}\nproved the link between dataset condensation (DC) and ($\\epsilon$,\n$\\delta$)-differential privacy. However, it is unclear from existing works on\nablating DC to obtain a lower-bound estimate of $\\epsilon$ that will suffice\nfor creating high-fidelity synthetic data. We suggest that adversarial\nuncertainty is the most appropriate method to achieve an optimal noise level,\n$\\epsilon$. As part of the internal dynamics of dataset condensation, we adopt\na satisfactory scheme for noise estimation that guarantees high-fidelity data\nwhile providing privacy.\n", "date": "2024-02-16", "categories": []}
{"id": "2309.07822", "title": "CATfOOD: Counterfactual Augmented Training for Improving Out-of-Domain\n  Performance and Calibration", "abstract": "  In recent years, large language models (LLMs) have shown remarkable\ncapabilities at scale, particularly at generating text conditioned on a prompt.\nIn our work, we investigate the use of LLMs to augment training data of small\nlanguage models~(SLMs) with automatically generated counterfactual~(CF)\ninstances -- i.e. minimally altered inputs -- in order to improve\nout-of-domain~(OOD) performance of SLMs in the extractive question\nanswering~(QA) setup. We show that, across various LLM generators, such data\naugmentation consistently enhances OOD performance and improves model\ncalibration for both confidence-based and rationale-augmented calibrator\nmodels. Furthermore, these performance improvements correlate with higher\ndiversity of CF instances in terms of their surface form and semantic content.\nFinally, we show that CF augmented models which are easier to calibrate also\nexhibit much lower entropy when assigning importance, indicating that\nrationale-augmented calibrators prefer concise explanations.\n", "date": "2023-09-14", "categories": []}
{"id": "2402.08209", "title": "Thresholding Data Shapley for Data Cleansing Using Multi-Armed Bandits", "abstract": "  Data cleansing aims to improve model performance by removing a set of harmful\ninstances from the training dataset. Data Shapley is a common theoretically\nguaranteed method to evaluate the contribution of each instance to model\nperformance; however, it requires training on all subsets of the training data,\nwhich is computationally expensive. In this paper, we propose an\niterativemethod to fast identify a subset of instances with low data Shapley\nvalues by using the thresholding bandit algorithm. We provide a theoretical\nguarantee that the proposed method can accurately select harmful instances if a\nsufficiently large number of iterations is conducted. Empirical evaluation\nusing various models and datasets demonstrated that the proposed method\nefficiently improved the computational speed while maintaining the model\nperformance.\n", "date": "2024-02-13", "categories": []}
{"id": "2306.13092", "title": "Squeeze, Recover and Relabel: Dataset Condensation at ImageNet Scale\n  From A New Perspective", "abstract": "  We present a new dataset condensation framework termed Squeeze, Recover and\nRelabel (SRe$^2$L) that decouples the bilevel optimization of model and\nsynthetic data during training, to handle varying scales of datasets, model\narchitectures and image resolutions for efficient dataset condensation. The\nproposed method demonstrates flexibility across diverse dataset scales and\nexhibits multiple advantages in terms of arbitrary resolutions of synthesized\nimages, low training cost and memory consumption with high-resolution\nsynthesis, and the ability to scale up to arbitrary evaluation network\narchitectures. Extensive experiments are conducted on Tiny-ImageNet and full\nImageNet-1K datasets. Under 50 IPC, our approach achieves the highest 42.5% and\n60.8% validation accuracy on Tiny-ImageNet and ImageNet-1K, outperforming all\nprevious state-of-the-art methods by margins of 14.5% and 32.9%, respectively.\nOur approach also surpasses MTT in terms of speed by approximately 52$\\times$\n(ConvNet-4) and 16$\\times$ (ResNet-18) faster with less memory consumption of\n11.6$\\times$ and 6.4$\\times$ during data synthesis. Our code and condensed\ndatasets of 50, 200 IPC with 4K recovery budget are available at\nhttps://github.com/VILA-Lab/SRe2L.\n", "date": "2023-06-22", "categories": []}
{"id": "2402.06811", "title": "Discipline and Label: A WEIRD Genealogy and Social Theory of Data\n  Annotation", "abstract": "  Data annotation remains the sine qua non of machine learning and AI. Recent\nempirical work on data annotation has begun to highlight the importance of\nrater diversity for fairness, model performance, and new lines of research have\nbegun to examine the working conditions for data annotation workers, the\nimpacts and role of annotator subjectivity on labels, and the potential\npsychological harms from aspects of annotation work. This paper outlines a\ncritical genealogy of data annotation; starting with its psychological and\nperceptual aspects. We draw on similarities with critiques of the rise of\ncomputerized lab-based psychological experiments in the 1970's which question\nwhether these experiments permit the generalization of results beyond the\nlaboratory settings within which these results are typically obtained. Do data\nannotations permit the generalization of results beyond the settings, or\nlocations, in which they were obtained? Psychology is overly reliant on\nparticipants from Western, Educated, Industrialized, Rich, and Democratic\nsocieties (WEIRD). Many of the people who work as data annotation platform\nworkers, however, are not from WEIRD countries; most data annotation workers\nare based in Global South countries. Social categorizations and classifications\nfrom WEIRD countries are imposed on non-WEIRD annotators through instructions\nand tasks, and through them, on data, which is then used to train or evaluate\nAI models in WEIRD countries. We synthesize evidence from several recent lines\nof research and argue that data annotation is a form of automated social\ncategorization that risks entrenching outdated and static social categories\nthat are in reality dynamic and changing. We propose a framework for\nunderstanding the interplay of the global social conditions of data annotation\nwith the subjective phenomenological experience of data annotation work.\n", "date": "2024-02-09", "categories": []}
{"id": "2402.06971", "title": "In-Context Data Distillation with TabPFN", "abstract": "  Foundation models have revolutionized tasks in computer vision and natural\nlanguage processing. However, in the realm of tabular data, tree-based models\nlike XGBoost continue to dominate. TabPFN, a transformer model tailored for\ntabular data, mirrors recent foundation models in its exceptional in-context\nlearning capability, being competitive with XGBoost's performance without the\nneed for task-specific training or hyperparameter tuning. Despite its promise,\nTabPFN's applicability is hindered by its data size constraint, limiting its\nuse in real-world scenarios. To address this, we present in-context data\ndistillation (ICD), a novel methodology that effectively eliminates these\nconstraints by optimizing TabPFN's context. ICD efficiently enables TabPFN to\nhandle significantly larger datasets with a fixed memory budget, improving\nTabPFN's quadratic memory complexity but at the cost of a linear number of\ntuning steps. Notably, TabPFN, enhanced with ICD, demonstrates very strong\nperformance against established tree-based models and modern deep learning\nmethods on 48 large tabular datasets from OpenML.\n", "date": "2024-02-10", "categories": []}
{"id": "2402.05962", "title": "EXGC: Bridging Efficiency and Explainability in Graph Condensation", "abstract": "  Graph representation learning on vast datasets, like web data, has made\nsignificant strides. However, the associated computational and storage\noverheads raise concerns. In sight of this, Graph condensation (GCond) has been\nintroduced to distill these large real datasets into a more concise yet\ninformation-rich synthetic graph. Despite acceleration efforts, existing GCond\nmethods mainly grapple with efficiency, especially on expansive web data\ngraphs. Hence, in this work, we pinpoint two major inefficiencies of current\nparadigms: (1) the concurrent updating of a vast parameter set, and (2)\npronounced parameter redundancy. To counteract these two limitations\ncorrespondingly, we first (1) employ the Mean-Field variational approximation\nfor convergence acceleration, and then (2) propose the objective of Gradient\nInformation Bottleneck (GDIB) to prune redundancy. By incorporating the leading\nexplanation techniques (e.g., GNNExplainer and GSAT) to instantiate the GDIB,\nour EXGC, the Efficient and eXplainable Graph Condensation method is proposed,\nwhich can markedly boost efficiency and inject explainability. Our extensive\nevaluations across eight datasets underscore EXGC's superiority and relevance.\nCode is available at https://github.com/MangoKiller/EXGC.\n", "date": "2024-02-05", "categories": []}
{"id": "2402.02318", "title": "Diversity Measurement and Subset Selection for Instruction Tuning\n  Datasets", "abstract": "  We aim to select data subsets for the fine-tuning of large language models to\nmore effectively follow instructions. Prior work has emphasized the importance\nof diversity in dataset curation but relied on heuristics such as the number of\ntasks. In this paper, we use determinantal point processes to capture the\ndiversity and quality of instruction tuning datasets for subset selection. We\npropose to measure dataset diversity with log determinant distance that is the\ndistance between the dataset of interest and a maximally diverse reference\ndataset. Our experiments demonstrate that the proposed diversity measure in the\nnormalized weight gradient space is correlated with downstream\ninstruction-following performance. Consequently, it can be used to inform when\ndata selection is the most helpful and to analyze dataset curation strategies.\nWe demonstrate the utility of our approach on various instruction tuning\ndatasets.\n", "date": "2024-02-04", "categories": []}
{"id": "2402.02055", "title": "Variance Alignment Score: A Simple But Tough-to-Beat Data Selection\n  Method for Multimodal Contrastive Learning", "abstract": "  In recent years, data selection has emerged as a core issue for large-scale\nvisual-language model pretraining, especially on noisy web-curated datasets.\nOne widely adopted strategy assigns quality scores such as CLIP similarity for\neach sample and retains the data pairs with the highest scores. However, these\napproaches are agnostic of data distribution and always fail to select the most\ninformative samples. To solve this problem, we propose a simple yet\ntheoretically principled metric named Variance Alignment Score (VAS), which has\nthe form $\\langle \\Sigma_{\\text{test}}, \\Sigma_i\\rangle$. Here,\n$\\Sigma_{\\text{test}}$ represents the target (cross-)covariance matrix we aim\nto align, potentially based on prior knowledge, while $\\Sigma_i$ denotes the\ntensor product of single or multi-modal representations for the $i$-th sample.\nWe further design a new data selection method that maximizes the total VAS. We\nprovide theoretical analysis in a simplified setting to demonstrate the\ntheoretical advantage of VAS over random or other existing data selection.\nExperimentally, applying VAS and CLIP scores together can outperform baselines\nby a margin of $1.3\\%$ average on 38 evaluation sets for noisy dataset DataComp\nand $2.5\\%$ on VTAB for high-quality dataset CC12M. Additionally, our ablation\nstudy also shows visual features are better than text for calculating VAS, and\nthe related classical experimental design methods may fail under this context.\n", "date": "2024-02-03", "categories": []}
{"id": "2401.17736", "title": "Leveraging Human-Machine Interactions for Computer Vision Dataset\n  Quality Enhancement", "abstract": "  Large-scale datasets for single-label multi-class classification, such as\n\\emph{ImageNet-1k}, have been instrumental in advancing deep learning and\ncomputer vision. However, a critical and often understudied aspect is the\ncomprehensive quality assessment of these datasets, especially regarding\npotential multi-label annotation errors. In this paper, we introduce a\nlightweight, user-friendly, and scalable framework that synergizes human and\nmachine intelligence for efficient dataset validation and quality enhancement.\nWe term this novel framework \\emph{Multilabelfy}. Central to Multilabelfy is an\nadaptable web-based platform that systematically guides annotators through the\nre-evaluation process, effectively leveraging human-machine interactions to\nenhance dataset quality. By using Multilabelfy on the ImageNetV2 dataset, we\nfound that approximately $47.88\\%$ of the images contained at least two labels,\nunderscoring the need for more rigorous assessments of such influential\ndatasets. Furthermore, our analysis showed a negative correlation between the\nnumber of potential labels per image and model top-1 accuracy, illuminating a\ncrucial factor in model evaluation and selection. Our open-source framework,\nMultilabelfy, offers a convenient, lightweight solution for dataset\nenhancement, emphasizing multi-label proportions. This study tackles major\nchallenges in dataset integrity and provides key insights into model\nperformance evaluation. Moreover, it underscores the advantages of integrating\nhuman expertise with machine capabilities to produce more robust models and\ntrustworthy data development. The source code for Multilabelfy will be\navailable at https://github.com/esla/Multilabelfy.\n  \\keywords{Computer Vision \\and Dataset Quality Enhancement \\and Dataset\nValidation \\and Human-Computer Interaction \\and Multi-label Annotation.}\n", "date": "2024-01-31", "categories": []}
{"id": "2401.15863", "title": "Importance-Aware Adaptive Dataset Distillation", "abstract": "  Herein, we propose a novel dataset distillation method for constructing small\ninformative datasets that preserve the information of the large original\ndatasets. The development of deep learning models is enabled by the\navailability of large-scale datasets. Despite unprecedented success,\nlarge-scale datasets considerably increase the storage and transmission costs,\nresulting in a cumbersome model training process. Moreover, using raw data for\ntraining raises privacy and copyright concerns. To address these issues, a new\ntask named dataset distillation has been introduced, aiming to synthesize a\ncompact dataset that retains the essential information from the large original\ndataset. State-of-the-art (SOTA) dataset distillation methods have been\nproposed by matching gradients or network parameters obtained during training\non real and synthetic datasets. The contribution of different network\nparameters to the distillation process varies, and uniformly treating them\nleads to degraded distillation performance. Based on this observation, we\npropose an importance-aware adaptive dataset distillation (IADD) method that\ncan improve distillation performance by automatically assigning importance\nweights to different network parameters during distillation, thereby\nsynthesizing more robust distilled datasets. IADD demonstrates superior\nperformance over other SOTA dataset distillation methods based on parameter\nmatching on multiple benchmark datasets and outperforms them in terms of\ncross-architecture generalization. In addition, the analysis of self-adaptive\nweights demonstrates the effectiveness of IADD. Furthermore, the effectiveness\nof IADD is validated in a real-world medical application such as COVID-19\ndetection.\n", "date": "2024-01-29", "categories": []}
{"id": "2211.04625", "title": "Soft Augmentation for Image Classification", "abstract": "  Modern neural networks are over-parameterized and thus rely on strong\nregularization such as data augmentation and weight decay to reduce overfitting\nand improve generalization. The dominant form of data augmentation applies\ninvariant transforms, where the learning target of a sample is invariant to the\ntransform applied to that sample. We draw inspiration from human visual\nclassification studies and propose generalizing augmentation with invariant\ntransforms to soft augmentation where the learning target softens non-linearly\nas a function of the degree of the transform applied to the sample: e.g., more\naggressive image crop augmentations produce less confident learning targets. We\ndemonstrate that soft targets allow for more aggressive data augmentation,\noffer more robust performance boosts, work with other augmentation policies,\nand interestingly, produce better calibrated models (since they are trained to\nbe less confident on aggressively cropped/occluded examples). Combined with\nexisting aggressive augmentation strategies, soft target 1) doubles the top-1\naccuracy boost across Cifar-10, Cifar-100, ImageNet-1K, and ImageNet-V2, 2)\nimproves model occlusion performance by up to $4\\times$, and 3) halves the\nexpected calibration error (ECE). Finally, we show that soft augmentation\ngeneralizes to self-supervised classification tasks. Code available at\nhttps://github.com/youngleox/soft_augmentation\n", "date": "2022-11-09", "categories": []}
{"id": "2401.12225", "title": "Multimodal Data Curation via Object Detection and Filter Ensembles", "abstract": "  We propose an approach for curating multimodal data that we used for our\nentry in the 2023 DataComp competition filtering track. Our technique combines\nobject detection and weak supervision-based ensembling. In the first of two\nsteps in our approach, we employ an out-of-the-box zero-shot object detection\nmodel to extract granular information and produce a variety of filter designs.\nIn the second step, we employ weak supervision to ensemble filtering rules.\nThis approach results in a 4% performance improvement when compared to the\nbest-performing baseline, producing the top-ranking position in the small scale\ntrack at the time of writing. Furthermore, in the medium scale track, we\nachieve a noteworthy 4.2% improvement over the baseline by simply ensembling\nexisting baselines with weak supervision.\n", "date": "2024-01-05", "categories": []}
{"id": "2401.12926", "title": "DsDm: Model-Aware Dataset Selection with Datamodels", "abstract": "  When selecting data for training large-scale models, standard practice is to\nfilter for examples that match human notions of data quality. Such filtering\nyields qualitatively clean datapoints that intuitively should improve model\nbehavior. However, in practice the opposite can often happen: we find that\nselecting according to similarity with \"high quality\" data sources may not\nincrease (and can even hurt) performance compared to randomly selecting data.\n  To develop better methods for selecting data, we start by framing dataset\nselection as an optimization problem that we can directly solve for: given\ntarget tasks, a learning algorithm, and candidate data, select the subset that\nmaximizes model performance. This framework thus avoids handpicked notions of\ndata quality, and instead models explicitly how the learning process uses train\ndatapoints to predict on the target tasks. Our resulting method greatly\nimproves language model (LM) performance on both pre-specified tasks and\npreviously unseen tasks. Specifically, choosing target tasks representative of\nstandard LM problems and evaluating on diverse held-out benchmarks, our\nselected datasets provide a 2x compute multiplier over baseline methods.\n", "date": "2024-01-23", "categories": []}
{"id": "2311.14212", "title": "Annotation Sensitivity: Training Data Collection Methods Affect Model\n  Performance", "abstract": "  When training data are collected from human annotators, the design of the\nannotation instrument, the instructions given to annotators, the\ncharacteristics of the annotators, and their interactions can impact training\ndata. This study demonstrates that design choices made when creating an\nannotation instrument also impact the models trained on the resulting\nannotations. We introduce the term annotation sensitivity to refer to the\nimpact of annotation data collection methods on the annotations themselves and\non downstream model performance and predictions. We collect annotations of hate\nspeech and offensive language in five experimental conditions of an annotation\ninstrument, randomly assigning annotators to conditions. We then fine-tune BERT\nmodels on each of the five resulting datasets and evaluate model performance on\na holdout portion of each condition. We find considerable differences between\nthe conditions for 1) the share of hate speech/offensive language annotations,\n2) model performance, 3) model predictions, and 4) model learning curves. Our\nresults emphasize the crucial role played by the annotation instrument which\nhas received little attention in the machine learning literature. We call for\nadditional research into how and why the instrument impacts the annotations to\ninform the development of best practices in instrument design.\n", "date": "2023-11-23", "categories": []}
{"id": "2312.11413", "title": "DeRDaVa: Deletion-Robust Data Valuation for Machine Learning", "abstract": "  Data valuation is concerned with determining a fair valuation of data from\ndata sources to compensate them or to identify training examples that are the\nmost or least useful for predictions. With the rising interest in personal data\nownership and data protection regulations, model owners will likely have to\nfulfil more data deletion requests. This raises issues that have not been\naddressed by existing works: Are the data valuation scores still fair with\ndeletions? Must the scores be expensively recomputed? The answer is no. To\navoid recomputations, we propose using our data valuation framework DeRDaVa\nupfront for valuing each data source's contribution to preserving robust model\nperformance after anticipated data deletions. DeRDaVa can be efficiently\napproximated and will assign higher values to data that are more useful or less\nlikely to be deleted. We further generalize DeRDaVa to Risk-DeRDaVa to cater to\nrisk-averse/seeking model owners who are concerned with the worst/best-cases\nmodel utility. We also empirically demonstrate the practicality of our\nsolutions.\n", "date": "2023-12-18", "categories": []}
{"id": "2312.13141", "title": "Augment on Manifold: Mixup Regularization with UMAP", "abstract": "  Data augmentation techniques play an important role in enhancing the\nperformance of deep learning models. Despite their proven benefits in computer\nvision tasks, their application in the other domains remains limited. This\npaper proposes a Mixup regularization scheme, referred to as UMAP Mixup,\ndesigned for ``on-manifold\" automated data augmentation for deep learning\npredictive models. The proposed approach ensures that the Mixup operations\nresult in synthesized samples that lie on the data manifold of the features and\nlabels by utilizing a dimensionality reduction technique known as uniform\nmanifold approximation and projection. Evaluations across diverse regression\ntasks show that UMAP Mixup is competitive with or outperforms other Mixup\nvariants, show promise for its potential as an effective tool for enhancing the\ngeneralization performance of deep learning models.\n", "date": "2023-12-20", "categories": []}
{"id": "2401.11103", "title": "Efficient Data Shapley for Weighted Nearest Neighbor Algorithms", "abstract": "  This work aims to address an open problem in data valuation literature\nconcerning the efficient computation of Data Shapley for weighted $K$ nearest\nneighbor algorithm (WKNN-Shapley). By considering the accuracy of hard-label\nKNN with discretized weights as the utility function, we reframe the\ncomputation of WKNN-Shapley into a counting problem and introduce a\nquadratic-time algorithm, presenting a notable improvement from $O(N^K)$, the\nbest result from existing literature. We develop a deterministic approximation\nalgorithm that further improves computational efficiency while maintaining the\nkey fairness properties of the Shapley value. Through extensive experiments, we\ndemonstrate WKNN-Shapley's computational efficiency and its superior\nperformance in discerning data quality compared to its unweighted counterpart.\n", "date": "2024-01-20", "categories": []}
{"id": "2212.11854", "title": "Data-Centric Artificial Intelligence", "abstract": "  Data-centric artificial intelligence (data-centric AI) represents an emerging\nparadigm emphasizing that the systematic design and engineering of data is\nessential for building effective and efficient AI-based systems. The objective\nof this article is to introduce practitioners and researchers from the field of\nInformation Systems (IS) to data-centric AI. We define relevant terms, provide\nkey characteristics to contrast the data-centric paradigm to the model-centric\none, and introduce a framework for data-centric AI. We distinguish data-centric\nAI from related concepts and discuss its longer-term implications for the IS\ncommunity.\n", "date": "2022-12-22", "categories": []}
{"id": "2401.07540", "title": "Study Features via Exploring Distribution Structure", "abstract": "  In this paper, we present a novel framework for data redundancy measurement\nbased on probabilistic modeling of datasets, and a new criterion for redundancy\ndetection that is resilient to noise. We also develop new methods for data\nredundancy reduction using both deterministic and stochastic optimization\ntechniques. Our framework is flexible and can handle different types of\nfeatures, and our experiments on benchmark datasets demonstrate the\neffectiveness of our methods. We provide a new perspective on feature\nselection, and propose effective and robust approaches for both supervised and\nunsupervised learning problems.\n", "date": "2024-01-15", "categories": []}
{"id": "2305.03711", "title": "Medical records condensation: a roadmap towards healthcare data\n  democratisation", "abstract": "  The prevalence of artificial intelligence (AI) has envisioned an era of\nhealthcare democratisation that promises every stakeholder a new and better way\nof life. However, the advancement of clinical AI research is significantly\nhurdled by the dearth of data democratisation in healthcare. To truly\ndemocratise data for AI studies, challenges are two-fold: 1. the sensitive\ninformation in clinical data should be anonymised appropriately, and 2.\nAI-oriented clinical knowledge should flow freely across organisations. This\npaper considers a recent deep-learning advent, dataset condensation (DC), as a\nstone that kills two birds in democratising healthcare data. The condensed data\nafter DC, which can be viewed as statistical metadata, abstracts original\nclinical records and irreversibly conceals sensitive information at individual\nlevels; nevertheless, it still preserves adequate knowledge for learning deep\nneural networks (DNNs). More favourably, the compressed volumes and the\naccelerated model learnings of condensed data portray a more efficient clinical\nknowledge sharing and flowing system, as necessitated by data democratisation.\nWe underline DC's prospects for democratising clinical data, specifically\nelectrical healthcare records (EHRs), for AI research through experimental\nresults and analysis across three healthcare datasets of varying data types.\n", "date": "2023-05-05", "categories": []}
{"id": "2401.01867", "title": "Dataset Difficulty and the Role of Inductive Bias", "abstract": "  Motivated by the goals of dataset pruning and defect identification, a\ngrowing body of methods have been developed to score individual examples within\na dataset. These methods, which we call \"example difficulty scores\", are\ntypically used to rank or categorize examples, but the consistency of rankings\nbetween different training runs, scoring methods, and model architectures is\ngenerally unknown. To determine how example rankings vary due to these random\nand controlled effects, we systematically compare different formulations of\nscores over a range of runs and model architectures. We find that scores\nlargely share the following traits: they are noisy over individual runs of a\nmodel, strongly correlated with a single notion of difficulty, and reveal\nexamples that range from being highly sensitive to insensitive to the inductive\nbiases of certain model architectures. Drawing from statistical genetics, we\ndevelop a simple method for fingerprinting model architectures using a few\nsensitive examples. These findings guide practitioners in maximizing the\nconsistency of their scores (e.g. by choosing appropriate scoring methods,\nnumber of runs, and subsets of examples), and establishes comprehensive\nbaselines for evaluating scores in the future.\n", "date": "2024-01-03", "categories": []}
{"id": "2310.06594", "title": "On the Evaluation and Refinement of Vision-Language Instruction Tuning\n  Datasets", "abstract": "  There is an emerging line of research on multimodal instruction tuning, and a\nline of benchmarks has been proposed for evaluating these models recently.\nInstead of evaluating the models directly, in this paper, we try to evaluate\nthe Vision-Language Instruction-Tuning (VLIT) datasets. Also, we seek the way\nof building a dataset for developing an all-powerful VLIT model, which we\nbelieve could also be of utility for establishing a grounded protocol for\nbenchmarking VLIT models. For effective evaluation of VLIT datasets that\nremains an open question, we propose a tune-cross-evaluation paradigm: tuning\non one dataset and evaluating on the others in turn. For each single\ntune-evaluation experiment set, we define the Meta Quality (MQ) as the mean\nscore obtained by a set of caption metrics including BLEU, METEOR, and ROUGE-L\nto quantify the quality of a certain dataset or a sample. On this basis, to\nevaluate the comprehensiveness of a dataset, we develop the Dataset Quality\n(DQ) covering all tune-evaluation sets. To lay the foundation for building a\ncomprehensive dataset and developing an all-powerful model for practical\napplications, we define the Sample Quality (SQ) to quantify the all-sided\nquality of each sample. Extensive experiments validate the rationality of the\nproposed evaluation paradigm. Based on the holistic evaluation, we build a new\ndataset, REVO-LION (REfining VisiOn-Language InstructiOn tuNing), by collecting\nsamples with higher SQ from each dataset. Remarkably, even with only half of\nthe complete data, the model trained on REVO-LION can achieve the performance\ncomparable to simply adding all VLIT datasets up. Furthermore, REVO-LION not\nonly facilitates the development of a powerful model but also incorporates an\nevaluation set, which is designed to serve as a convenient benchmark for future\nresearch in the field.\n", "date": "2023-10-10", "categories": []}
{"id": "2312.16627", "title": "MIM4DD: Mutual Information Maximization for Dataset Distillation", "abstract": "  Dataset distillation (DD) aims to synthesize a small dataset whose test\nperformance is comparable to a full dataset using the same model.\nState-of-the-art (SoTA) methods optimize synthetic datasets primarily by\nmatching heuristic indicators extracted from two networks: one from real data\nand one from synthetic data (see Fig.1, Left), such as gradients and training\ntrajectories. DD is essentially a compression problem that emphasizes\nmaximizing the preservation of information contained in the data. We argue that\nwell-defined metrics which measure the amount of shared information between\nvariables in information theory are necessary for success measurement but are\nnever considered by previous works. Thus, we introduce mutual information (MI)\nas the metric to quantify the shared information between the synthetic and the\nreal datasets, and devise MIM4DD numerically maximizing the MI via a newly\ndesigned optimizable objective within a contrastive learning framework to\nupdate the synthetic dataset. Specifically, we designate the samples in\ndifferent datasets that share the same labels as positive pairs and vice versa\nnegative pairs. Then we respectively pull and push those samples in positive\nand negative pairs into contrastive space via minimizing NCE loss. As a result,\nthe targeted MI can be transformed into a lower bound represented by feature\nmaps of samples, which is numerically feasible. Experiment results show that\nMIM4DD can be implemented as an add-on module to existing SoTA DD methods.\n", "date": "2023-12-27", "categories": []}
{"id": "2301.05603", "title": "A Comprehensive Survey of Dataset Distillation", "abstract": "  Deep learning technology has developed unprecedentedly in the last decade and\nhas become the primary choice in many application domains. This progress is\nmainly attributed to a systematic collaboration in which rapidly growing\ncomputing resources encourage advanced algorithms to deal with massive data.\nHowever, it has gradually become challenging to handle the unlimited growth of\ndata with limited computing power. To this end, diverse approaches are proposed\nto improve data processing efficiency. Dataset distillation, a dataset\nreduction method, addresses this problem by synthesizing a small typical\ndataset from substantial data and has attracted much attention from the deep\nlearning community. Existing dataset distillation methods can be taxonomized\ninto meta-learning and data matching frameworks according to whether they\nexplicitly mimic the performance of target data. Although dataset distillation\nhas shown surprising performance in compressing datasets, there are still\nseveral limitations such as distilling high-resolution data or data with\ncomplex label spaces. This paper provides a holistic understanding of dataset\ndistillation from multiple aspects, including distillation frameworks and\nalgorithms, factorized dataset distillation, performance comparison, and\napplications. Finally, we discuss challenges and promising directions to\nfurther promote future studies on dataset distillation.\n", "date": "2023-01-13", "categories": []}
{"id": "2305.00054", "title": "LAVA: Data Valuation without Pre-Specified Learning Algorithms", "abstract": "  Traditionally, data valuation (DV) is posed as a problem of equitably\nsplitting the validation performance of a learning algorithm among the training\ndata. As a result, the calculated data values depend on many design choices of\nthe underlying learning algorithm. However, this dependence is undesirable for\nmany DV use cases, such as setting priorities over different data sources in a\ndata acquisition process and informing pricing mechanisms in a data\nmarketplace. In these scenarios, data needs to be valued before the actual\nanalysis and the choice of the learning algorithm is still undetermined then.\nAnother side-effect of the dependence is that to assess the value of individual\npoints, one needs to re-run the learning algorithm with and without a point,\nwhich incurs a large computation burden. This work leapfrogs over the current\nlimits of data valuation methods by introducing a new framework that can value\ntraining data in a way that is oblivious to the downstream learning algorithm.\nOur main results are as follows. (1) We develop a proxy for the validation\nperformance associated with a training set based on a non-conventional\nclass-wise Wasserstein distance between training and validation sets. We show\nthat the distance characterizes the upper bound of the validation performance\nfor any given model under certain Lipschitz conditions. (2) We develop a novel\nmethod to value individual data based on the sensitivity analysis of the\nclass-wise Wasserstein distance. Importantly, these values can be directly\nobtained for free from the output of off-the-shelf optimization solvers when\ncomputing the distance. (3) We evaluate our new data valuation framework over\nvarious use cases related to detecting low-quality data and show that,\nsurprisingly, the learning-agnostic feature of our framework enables a\nsignificant improvement over SOTA performance while being orders of magnitude\nfaster.\n", "date": "2023-04-28", "categories": []}
{"id": "2309.02033", "title": "Data-Juicer: A One-Stop Data Processing System for Large Language Models", "abstract": "  The immense evolution in Large Language Models (LLMs) has underscored the\nimportance of massive, heterogeneous, and high-quality data. A data recipe is a\nmixture of data from different sources for training LLMs, which plays a vital\nrole in LLMs' performance. Existing open-source tools for LLM data processing\nare mostly tailored for specific data recipes. To continuously uncover the\npotential of LLMs, incorporate data from new sources, and improve LLMs'\nperformance, we build a new system named Data-Juicer, with which we can\nefficiently generate diverse data recipes, explore different possibilities in\nforming data mixtures, and evaluate their effects on model performance.\nDifferent from traditional data-analytics pipelines, Data-Juicer faces some\nunique challenges. Firstly, the possible data sources for forming data recipes\nare truly heterogeneous and massive with various qualities. Secondly, it is\nextremely expensive to precisely evaluate data recipes' impact on LLMs'\nperformance. Thirdly, the end users of Data-Juicer, model developers, need\nsufficient flexibility to configure and evaluate different data recipes.\n  Data-Juicer features a fine-grained abstraction of pipelines for constructing\ndata recipes, with over 50 built-in operators for easy composition and\nextension. By incorporating visualization and auto-evaluation capabilities,\nData-Juicer enables a timely feedback loop for both LLM pre-training and\nfine-tuning. Further, Data-Juicer is optimized and integrated with ecosystems\nfor LLM training, evaluation, and distributed computing. The data recipes\nderived with Data-Juicer gain notable improvements on state-of-the-art LLMs, by\nup to 7.45% increase in averaged score across 16 LLM benchmarks and 17.5%\nhigher win rate in pair-wise GPT-4 evaluations. Our system, data recipes, and\ntutorials are released, calling for broader data-centric research on training\nand understanding LLMs.\n", "date": "2023-09-05", "categories": []}
{"id": "2312.06022", "title": "Exploiting Representation Bias for Data Distillation in Abstractive Text\n  Summarization", "abstract": "  Abstractive text summarization is surging with the number of training samples\nto cater to the needs of the deep learning models. These models tend to exploit\nthe training data representations to attain superior performance by improving\nthe quantitative element of the resultant summary. However, increasing the size\nof the training set may not always be the ideal solution to maximize the\nperformance, and therefore, a need to revisit the quality of training samples\nand the learning protocol of deep learning models is a must. In this paper, we\naim to discretize the vector space of the abstractive text summarization models\nto understand the characteristics learned between the input embedding space and\nthe models' encoder space. We show that deep models fail to capture the\ndiversity of the input space. Further, the distribution of data points on the\nencoder space indicates that an unchecked increase in the training samples does\nnot add value; rather, a tear-down of data samples is highly needed to make the\nmodels focus on variability and faithfulness. We employ clustering techniques\nto learn the diversity of a model's sample space and how data points are mapped\nfrom the embedding space to the encoder space and vice versa. Further, we\ndevise a metric to filter out redundant data points to make the model more\nrobust and less data hungry. We benchmark our proposed method using\nquantitative metrics, such as Rouge, and qualitative metrics, such as\nBERTScore, FEQA and Pyramid score. We also quantify the reasons that inhibit\nthe models from learning the diversity from the varied input samples.\n", "date": "2023-12-10", "categories": []}
{"id": "2312.10602", "title": "A Weighted K-Center Algorithm for Data Subset Selection", "abstract": "  The success of deep learning hinges on enormous data and large models, which\nrequire labor-intensive annotations and heavy computation costs. Subset\nselection is a fundamental problem that can play a key role in identifying\nsmaller portions of the training data, which can then be used to produce\nsimilar models as the ones trained with full data. Two prior methods are shown\nto achieve impressive results: (1) margin sampling that focuses on selecting\npoints with high uncertainty, and (2) core-sets or clustering methods such as\nk-center for informative and diverse subsets. We are not aware of any work that\ncombines these methods in a principled manner. To this end, we develop a novel\nand efficient factor 3-approximation algorithm to compute subsets based on the\nweighted sum of both k-center and uncertainty sampling objective functions. To\nhandle large datasets, we show a parallel algorithm to run on multiple machines\nwith approximation guarantees. The proposed algorithm achieves similar or\nbetter performance compared to other strong baselines on vision datasets such\nas CIFAR-10, CIFAR-100, and ImageNet.\n", "date": "2023-12-17", "categories": []}
{"id": "2312.08912", "title": "Dataset Distillation via Adversarial Prediction Matching", "abstract": "  Dataset distillation is the technique of synthesizing smaller condensed\ndatasets from large original datasets while retaining necessary information to\npersist the effect. In this paper, we approach the dataset distillation problem\nfrom a novel perspective: we regard minimizing the prediction discrepancy on\nthe real data distribution between models, which are respectively trained on\nthe large original dataset and on the small distilled dataset, as a conduit for\ncondensing information from the raw data into the distilled version. An\nadversarial framework is proposed to solve the problem efficiently. In contrast\nto existing distillation methods involving nested optimization or long-range\ngradient unrolling, our approach hinges on single-level optimization. This\nensures the memory efficiency of our method and provides a flexible tradeoff\nbetween time and memory budgets, allowing us to distil ImageNet-1K using a\nminimum of only 6.5GB of GPU memory. Under the optimal tradeoff strategy, it\nrequires only 2.5$\\times$ less memory and 5$\\times$ less runtime compared to\nthe state-of-the-art. Empirically, our method can produce synthetic datasets\njust 10% the size of the original, yet achieve, on average, 94% of the test\naccuracy of models trained on the full original datasets including ImageNet-1K,\nsignificantly surpassing state-of-the-art. Additionally, extensive tests reveal\nthat our distilled datasets excel in cross-architecture generalization\ncapabilities.\n", "date": "2023-12-14", "categories": []}
{"id": "2312.08066", "title": "A Novel Metric for Measuring Data Quality in Classification Applications\n  (extended version)", "abstract": "  Data quality is a key element for building and optimizing good learning\nmodels. Despite many attempts to characterize data quality, there is still a\nneed for rigorous formalization and an efficient measure of the quality from\navailable observations. Indeed, without a clear understanding of the training\nand testing processes, it is hard to evaluate the intrinsic performance of a\nmodel. Besides, tools allowing to measure data quality specific to machine\nlearning are still lacking. In this paper, we introduce and explain a novel\nmetric to measure data quality. This metric is based on the correlated\nevolution between the classification performance and the deterioration of data.\nThe proposed method has the major advantage of being model-independent.\nFurthermore, we provide an interpretation of each criterion and examples of\nassessment levels. We confirm the utility of the proposed metric with intensive\nnumerical experiments and detail some illustrative cases with controlled and\ninterpretable qualities.\n", "date": "2023-12-13", "categories": []}
{"id": "2312.04911", "title": "Collinear datasets augmentation using Procrustes validation sets", "abstract": "  In this paper, we propose a new method for the augmentation of numeric and\nmixed datasets. The method generates additional data points by utilizing\ncross-validation resampling and latent variable modeling. It is particularly\nefficient for datasets with moderate to high degrees of collinearity, as it\ndirectly utilizes this property for generation. The method is simple, fast, and\nhas very few parameters, which, as shown in the paper, do not require specific\ntuning. It has been tested on several real datasets; here, we report detailed\nresults for two cases, prediction of protein in minced meat based on near\ninfrared spectra (fully numeric data with high degree of collinearity) and\ndiscrimination of patients referred for coronary angiography (mixed data, with\nboth numeric and categorical variables, and moderate collinearity). In both\ncases, artificial neural networks were employed for developing the regression\nand the discrimination models. The results show a clear improvement in the\nperformance of the models; thus for the prediction of meat protein, fitting the\nmodel to the augmented data resulted in a reduction in the root mean squared\nerror computed for the independent test set by 1.5 to 3 times.\n", "date": "2023-12-08", "categories": []}
{"id": "2312.03437", "title": "Data-Centric Digital Agriculture: A Perspective", "abstract": "  In response to the increasing global demand for food, feed, fiber, and fuel,\ndigital agriculture is rapidly evolving to meet these demands while reducing\nenvironmental impact. This evolution involves incorporating data science,\nmachine learning, sensor technologies, robotics, and new management strategies\nto establish a more sustainable agricultural framework. So far, machine\nlearning research in digital agriculture has predominantly focused on\nmodel-centric approaches, focusing on model design and evaluation. These\nefforts aim to optimize model accuracy and efficiency, often treating data as a\nstatic benchmark. Despite the availability of agricultural data and\nmethodological advancements, a saturation point has been reached, with many\nestablished machine learning methods achieving comparable levels of accuracy\nand facing similar limitations. To fully realize the potential of digital\nagriculture, it is crucial to have a comprehensive understanding of the role of\ndata in the field and to adopt data-centric machine learning. This involves\ndeveloping strategies to acquire and curate valuable data and implementing\neffective learning and evaluation strategies that utilize the intrinsic value\nof data. This approach has the potential to create accurate, generalizable, and\nadaptable machine learning methods that effectively and sustainably address\nagricultural tasks such as yield prediction, weed detection, and early disease\nidentification\n", "date": "2023-12-06", "categories": []}
{"id": "2312.02200", "title": "An Empirical Study of Automated Mislabel Detection in Real World Vision\n  Datasets", "abstract": "  Major advancements in computer vision can primarily be attributed to the use\nof labeled datasets. However, acquiring labels for datasets often results in\nerrors which can harm model performance. Recent works have proposed methods to\nautomatically identify mislabeled images, but developing strategies to\neffectively implement them in real world datasets has been sparsely explored.\nTowards improved data-centric methods for cleaning real world vision datasets,\nwe first conduct more than 200 experiments carefully benchmarking recently\ndeveloped automated mislabel detection methods on multiple datasets under a\nvariety of synthetic and real noise settings with varying noise levels. We\ncompare these methods to a Simple and Efficient Mislabel Detector (SEMD) that\nwe craft, and find that SEMD performs similarly to or outperforms prior\nmislabel detection approaches. We then apply SEMD to multiple real world\ncomputer vision datasets and test how dataset size, mislabel removal strategy,\nand mislabel removal amount further affect model performance after retraining\non the cleaned data. With careful design of the approach, we find that mislabel\nremoval leads per-class performance improvements of up to 8% of a retrained\nclassifier in smaller data regimes.\n", "date": "2023-12-02", "categories": []}
{"id": "2312.01537", "title": "Unlocking the Potential of Federated Learning: The Symphony of Dataset\n  Distillation via Deep Generative Latents", "abstract": "  Data heterogeneity presents significant challenges for federated learning\n(FL). Recently, dataset distillation techniques have been introduced, and\nperformed at the client level, to attempt to mitigate some of these challenges.\nIn this paper, we propose a highly efficient FL dataset distillation framework\non the server side, significantly reducing both the computational and\ncommunication demands on local devices while enhancing the clients' privacy.\nUnlike previous strategies that perform dataset distillation on local devices\nand upload synthetic data to the server, our technique enables the server to\nleverage prior knowledge from pre-trained deep generative models to synthesize\nessential data representations from a heterogeneous model architecture. This\nprocess allows local devices to train smaller surrogate models while enabling\nthe training of a larger global model on the server, effectively minimizing\nresource utilization. We substantiate our claim with a theoretical analysis,\ndemonstrating the asymptotic resemblance of the process to the hypothetical\nideal of completely centralized training on a heterogeneous dataset. Empirical\nevidence from our comprehensive experiments indicates our method's superiority,\ndelivering an accuracy enhancement of up to 40% over non-dataset-distillation\ntechniques in highly heterogeneous FL contexts, and surpassing existing\ndataset-distillation methods by 18%. In addition to the high accuracy, our\nframework converges faster than the baselines because rather than the server\ntrains on several sets of heterogeneous data distributions, it trains on a\nmulti-modal distribution. Our code is available at\nhttps://github.com/FedDG23/FedDG-main.git\n", "date": "2023-12-03", "categories": []}
{"id": "2304.10253", "title": "Image retrieval outperforms diffusion models on data augmentation", "abstract": "  Many approaches have been proposed to use diffusion models to augment\ntraining datasets for downstream tasks, such as classification. However,\ndiffusion models are themselves trained on large datasets, often with noisy\nannotations, and it remains an open question to which extent these models\ncontribute to downstream classification performance. In particular, it remains\nunclear if they generalize enough to improve over directly using the additional\ndata of their pre-training process for augmentation. We systematically evaluate\na range of existing methods to generate images from diffusion models and study\nnew extensions to assess their benefit for data augmentation. Personalizing\ndiffusion models towards the target data outperforms simpler prompting\nstrategies. However, using the pre-training data of the diffusion model alone,\nvia a simple nearest-neighbor retrieval procedure, leads to even stronger\ndownstream performance. Our study explores the potential of diffusion models in\ngenerating new training data, and surprisingly finds that these sophisticated\nmodels are not yet able to beat a simple and strong image retrieval baseline on\nsimple downstream vision tasks.\n", "date": "2023-04-20", "categories": []}
{"id": "2311.17967", "title": "Discovering Galaxy Features via Dataset Distillation", "abstract": "  In many applications, Neural Nets (NNs) have classification performance on\npar or even exceeding human capacity. Moreover, it is likely that NNs leverage\nunderlying features that might differ from those humans perceive to classify.\nCan we \"reverse-engineer\" pertinent features to enhance our scientific\nunderstanding? Here, we apply this idea to the notoriously difficult task of\ngalaxy classification: NNs have reached high performance for this task, but\nwhat does a neural net (NN) \"see\" when it classifies galaxies? Are there\nmorphological features that the human eye might overlook that could help with\nthe task and provide new insights? Can we visualize tracers of early evolution,\nor additionally incorporated spectral data? We present a novel way to summarize\nand visualize galaxy morphology through the lens of neural networks, leveraging\nDataset Distillation, a recent deep-learning methodology with the primary\nobjective to distill knowledge from a large dataset and condense it into a\ncompact synthetic dataset, such that a model trained on this synthetic dataset\nachieves performance comparable to a model trained on the full dataset. We\ncurate a class-balanced, medium-size high-confidence version of the Galaxy Zoo\n2 dataset, and proceed with dataset distillation from our accurate\nNN-classifier to create synthesized prototypical images of galaxy morphological\nfeatures, demonstrating its effectiveness. Of independent interest, we\nintroduce a self-adaptive version of the state-of-the-art Matching Trajectory\nalgorithm to automate the distillation process, and show enhanced performance\non computer vision benchmarks.\n", "date": "2023-11-29", "categories": []}
{"id": "2311.18064", "title": "GELDA: A generative language annotation framework to reveal visual\n  biases in datasets", "abstract": "  Bias analysis is a crucial step in the process of creating fair datasets for\ntraining and evaluating computer vision models. The bottleneck in dataset\nanalysis is annotation, which typically requires: (1) specifying a list of\nattributes relevant to the dataset domain, and (2) classifying each\nimage-attribute pair. While the second step has made rapid progress in\nautomation, the first has remained human-centered, requiring an experimenter to\ncompile lists of in-domain attributes. However, an experimenter may have\nlimited foresight leading to annotation \"blind spots,\" which in turn can lead\nto flawed downstream dataset analyses. To combat this, we propose GELDA, a\nnearly automatic framework that leverages large generative language models\n(LLMs) to propose and label various attributes for a domain. GELDA takes a\nuser-defined domain caption (e.g., \"a photo of a bird,\" \"a photo of a living\nroom\") and uses an LLM to hierarchically generate attributes. In addition,\nGELDA uses the LLM to decide which of a set of vision-language models (VLMs) to\nuse to classify each attribute in images. Results on real datasets show that\nGELDA can generate accurate and diverse visual attribute suggestions, and\nuncover biases such as confounding between class labels and background\nfeatures. Results on synthetic datasets demonstrate that GELDA can be used to\nevaluate the biases of text-to-image diffusion models and generative\nadversarial networks. Overall, we show that while GELDA is not accurate enough\nto replace human annotators, it can serve as a complementary tool to help\nhumans analyze datasets in a cheap, low-effort, and flexible manner.\n", "date": "2023-11-29", "categories": []}
{"id": "2311.06965", "title": "Anchor Data Augmentation", "abstract": "  We propose a novel algorithm for data augmentation in nonlinear\nover-parametrized regression. Our data augmentation algorithm borrows from the\nliterature on causality and extends the recently proposed Anchor regression\n(AR) method for data augmentation, which is in contrast to the current\nstate-of-the-art domain-agnostic solutions that rely on the Mixup literature.\nOur Anchor Data Augmentation (ADA) uses several replicas of the modified\nsamples in AR to provide more training examples, leading to more robust\nregression predictions. We apply ADA to linear and nonlinear regression\nproblems using neural networks. ADA is competitive with state-of-the-art\nC-Mixup solutions.\n", "date": "2023-11-12", "categories": []}
{"id": "2311.16298", "title": "Influence Scores at Scale for Efficient Language Data Sampling", "abstract": "  Modern ML systems ingest data aggregated from diverse sources, such as\nsynthetic, human-annotated, and live customer traffic. Understanding\n\\textit{which} examples are important to the performance of a learning\nalgorithm is crucial for efficient model training. Recently, a growing body of\nliterature has given rise to various \"influence scores,\" which use training\nartifacts such as model confidence or checkpointed gradients to identify\nimportant subsets of data. However, these methods have primarily been developed\nin computer vision settings, and it remains unclear how well they generalize to\nlanguage-based tasks using pretrained models.\n  In this paper, we explore the applicability of influence scores in language\nclassification tasks. We evaluate a diverse subset of these scores on the SNLI\ndataset by quantifying accuracy changes in response to pruning training data\nthrough random and influence-score-based sampling. We then stress-test one of\nthe scores -- \"variance of gradients\" (VoG) from Agarwal et al. (2022) -- in an\nNLU model stack that was exposed to dynamic user speech patterns in a voice\nassistant type of setting. Our experiments demonstrate that in many cases,\nencoder-based language models can be finetuned on roughly 50% of the original\ndata without degradation in performance metrics. Along the way, we summarize\nlessons learned from applying out-of-the-box implementations of influence\nscores, quantify the effects of noisy and class-imbalanced data, and offer\nrecommendations on score-based sampling for better accuracy and training\nefficiency.\n", "date": "2023-11-27", "categories": []}
{"id": "2311.16302", "title": "Comprehensive Benchmarking of Entropy and Margin Based Scoring Metrics\n  for Data Selection", "abstract": "  While data selection methods have been studied extensively in active\nlearning, data pruning, and data augmentation settings, there is little\nevidence for the efficacy of these methods in industry scale settings,\nparticularly in low-resource languages. Our work presents ways of assessing\nprospective training examples in those settings for their \"usefulness\" or\n\"difficulty\". We also demonstrate how these measures can be used in selecting\nimportant examples for training supervised machine learning models. We\nprimarily experiment with entropy and Error L2-Norm (EL2N) scores. We use these\nmetrics to curate high quality datasets from a large pool of \\textit{Weak\nSignal Labeled} data, which assigns no-defect high confidence hypotheses during\ninference as ground truth labels. We then conduct training data augmentation\nexperiments using these de-identified datasets and demonstrate that score-based\nselection can result in a 2% decrease in semantic error rate and 4%-7% decrease\nin domain classification error rate when compared to the baseline technique of\nrandom selection.\n", "date": "2023-11-27", "categories": []}
{"id": "2311.16508", "title": "RandMSAugment: A Mixed-Sample Augmentation for Limited-Data Scenarios", "abstract": "  The high costs of annotating large datasets suggests a need for effectively\ntraining CNNs with limited data, and data augmentation is a promising\ndirection. We study foundational augmentation techniques, including Mixed\nSample Data Augmentations (MSDAs) and a no-parameter variant of RandAugment\ntermed Preset-RandAugment, in the fully supervised scenario. We observe that\nPreset-RandAugment excels in limited-data contexts while MSDAs are moderately\neffective. We show that low-level feature transforms play a pivotal role in\nthis performance difference, postulate a new property of augmentations related\nto their data efficiency, and propose new ways to measure the diversity and\nrealism of augmentations. Building on these insights, we introduce a novel\naugmentation technique called RandMSAugment that integrates complementary\nstrengths of existing methods. RandMSAugment significantly outperforms the\ncompetition on CIFAR-100, STL-10, and Tiny-Imagenet. With very small training\nsets (4, 25, 100 samples/class), RandMSAugment achieves compelling performance\ngains between 4.1% and 6.75%. Even with more training data (500 samples/class)\nwe improve performance by 1.03% to 2.47%. RandMSAugment does not require\nhyperparameter tuning, extra validation data, or cumbersome optimizations.\n", "date": "2023-11-25", "categories": []}
{"id": "2304.04258", "title": "A Note on \"Efficient Task-Specific Data Valuation for Nearest Neighbor\n  Algorithms\"", "abstract": "  Data valuation is a growing research field that studies the influence of\nindividual data points for machine learning (ML) models. Data Shapley, inspired\nby cooperative game theory and economics, is an effective method for data\nvaluation. However, it is well-known that the Shapley value (SV) can be\ncomputationally expensive. Fortunately, Jia et al. (2019) showed that for\nK-Nearest Neighbors (KNN) models, the computation of Data Shapley is\nsurprisingly simple and efficient.\n  In this note, we revisit the work of Jia et al. (2019) and propose a more\nnatural and interpretable utility function that better reflects the performance\nof KNN models. We derive the corresponding calculation procedure for the Data\nShapley of KNN classifiers/regressors with the new utility functions. Our new\napproach, dubbed soft-label KNN-SV, achieves the same time complexity as the\noriginal method. We further provide an efficient approximation algorithm for\nsoft-label KNN-SV based on locality sensitive hashing (LSH). Our experimental\nresults demonstrate that Soft-label KNN-SV outperforms the original method on\nmost datasets in the task of mislabeled data detection, making it a better\nbaseline for future work on data valuation.\n", "date": "2023-04-09", "categories": []}
{"id": "2308.15709", "title": "Threshold KNN-Shapley: A Linear-Time and Privacy-Friendly Approach to\n  Data Valuation", "abstract": "  Data valuation aims to quantify the usefulness of individual data sources in\ntraining machine learning (ML) models, and is a critical aspect of data-centric\nML research. However, data valuation faces significant yet frequently\noverlooked privacy challenges despite its importance. This paper studies these\nchallenges with a focus on KNN-Shapley, one of the most practical data\nvaluation methods nowadays. We first emphasize the inherent privacy risks of\nKNN-Shapley, and demonstrate the significant technical difficulties in adapting\nKNN-Shapley to accommodate differential privacy (DP). To overcome these\nchallenges, we introduce TKNN-Shapley, a refined variant of KNN-Shapley that is\nprivacy-friendly, allowing for straightforward modifications to incorporate DP\nguarantee (DP-TKNN-Shapley). We show that DP-TKNN-Shapley has several\nadvantages and offers a superior privacy-utility tradeoff compared to naively\nprivatized KNN-Shapley in discerning data quality. Moreover, even non-private\nTKNN-Shapley achieves comparable performance as KNN-Shapley. Overall, our\nfindings suggest that TKNN-Shapley is a promising alternative to KNN-Shapley,\nparticularly for real-world applications involving sensitive data.\n", "date": "2023-08-30", "categories": []}
{"id": "2310.10541", "title": "AST: Effective Dataset Distillation through Alignment with Smooth and\n  High-Quality Expert Trajectories", "abstract": "  Training large AI models typically requires large-scale datasets in the\nmachine learning process, making training and parameter-tuning process both\ntime-consuming and costly. Some researchers address this problem by carefully\nsynthesizing a very small number of highly representative and informative\nsamples from real-world datasets. This approach, known as Dataset Distillation\n(DD), proposes a perspective for data-efficient learning. Despite recent\nprogress in this field, the performance of existing methods still cannot meet\nexpectations, and distilled datasets cannot effectively replace original\ndatasets. In this paper, unlike previous methods that focus solely on improving\nthe effectiveness of student distillation, we recognize and leverage the\nimportant mutual influence between expert and student models. We observed that\nthe smoothness of expert trajectories has a significant impact on subsequent\nstudent parameter alignment. Based on this, we propose an effective DD\nframework named AST, standing for Alignment with Smooth and high-quality expert\nTrajectories. We devise the integration of clipping loss and gradient penalty\nto regulate the rate of parameter changes in expert trajectory generation. To\nfurther refine the student parameter alignment with expert trajectory, we put\nforward representative initialization for the synthetic dataset and balanced\ninner-loop loss in response to the sensitivity exhibited towards randomly\ninitialized variables during distillation. We also propose two enhancement\nstrategies, namely intermediate matching loss and weight perturbation, to\nmitigate the potential occurrence of cumulative errors. We conduct extensive\nexperiments on datasets of different scales, sizes, and resolutions. The\nresults demonstrate that the proposed method significantly outperforms prior\nmethods.\n", "date": "2023-10-16", "categories": []}
{"id": "2311.15134", "title": "SwiftLearn: A Data-Efficient Training Method of Deep Learning Models\n  using Importance Sampling", "abstract": "  In this paper, we present SwiftLearn, a data-efficient approach to accelerate\ntraining of deep learning models using a subset of data samples selected during\nthe warm-up stages of training. This subset is selected based on an importance\ncriteria measured over the entire dataset during warm-up stages, aiming to\npreserve the model performance with fewer examples during the rest of training.\nThe importance measure we propose could be updated during training every once\nin a while, to make sure that all of the data samples have a chance to return\nto the training loop if they show a higher importance. The model architecture\nis unchanged but since the number of data samples controls the number of\nforward and backward passes during training, we can reduce the training time by\nreducing the number of training samples used in each epoch of training.\nExperimental results on a variety of CV and NLP models during both pretraining\nand finetuning show that the model performance could be preserved while\nachieving a significant speed-up during training. More specifically, BERT\nfinetuning on GLUE benchmark shows that almost 90% of the data can be dropped\nachieving an end-to-end average speedup of 3.36x while keeping the average\naccuracy drop less than 0.92%.\n", "date": "2023-11-25", "categories": []}
{"id": "2311.15547", "title": "Dataset Distillation in Latent Space", "abstract": "  Dataset distillation (DD) is a newly emerging research area aiming at\nalleviating the heavy computational load in training models on large datasets.\nIt tries to distill a large dataset into a small and condensed one so that\nmodels trained on the distilled dataset can perform comparably with those\ntrained on the full dataset when performing downstream tasks. Among the\nprevious works in this area, there are three key problems that hinder the\nperformance and availability of the existing DD methods: high time complexity,\nhigh space complexity, and low info-compactness. In this work, we\nsimultaneously attempt to settle these three problems by moving the DD\nprocesses from conventionally used pixel space to latent space. Encoded by a\npretrained generic autoencoder, latent codes in the latent space are naturally\ninfo-compact representations of the original images in much smaller sizes.\nAfter transferring three mainstream DD algorithms to latent space, we\nsignificantly reduce time and space consumption while achieving similar\nperformance, allowing us to distill high-resolution datasets or target at\ngreater data ratio that previous methods have failed. Besides, within the same\nstorage budget, we can also quantitatively deliver more latent codes than\npixel-level images, which further boosts the performance of our methods.\n", "date": "2023-11-27", "categories": []}
{"id": "2311.15772", "title": "Attend Who is Weak: Enhancing Graph Condensation via Cross-Free\n  Adversarial Training", "abstract": "  In this paper, we study the \\textit{graph condensation} problem by\ncompressing the large, complex graph into a concise, synthetic representation\nthat preserves the most essential and discriminative information of structure\nand features. We seminally propose the concept of Shock Absorber (a type of\nperturbation) that enhances the robustness and stability of the original graphs\nagainst changes in an adversarial training fashion. Concretely, (I) we forcibly\nmatch the gradients between pre-selected graph neural networks (GNNs) trained\non a synthetic, simplified graph and the original training graph at regularly\nspaced intervals. (II) Before each update synthetic graph point, a Shock\nAbsorber serves as a gradient attacker to maximize the distance between the\nsynthetic dataset and the original graph by selectively perturbing the parts\nthat are underrepresented or insufficiently informative. We iteratively repeat\nthe above two processes (I and II) in an adversarial training fashion to\nmaintain the highly-informative context without losing correlation with the\noriginal dataset. More importantly, our shock absorber and the synthesized\ngraph parallelly share the backward process in a free training manner. Compared\nto the original adversarial training, it introduces almost no additional time\noverhead.\n  We validate our framework across 8 datasets (3 graph and 5 node\nclassification datasets) and achieve prominent results: for example, on Cora,\nCiteseer and Ogbn-Arxiv, we can gain nearly 1.13% to 5.03% improvements compare\nwith SOTA models. Moreover, our algorithm adds only about 0.2% to 2.2%\nadditional time overhead over Flicker, Citeseer and Ogbn-Arxiv. Compared to the\ngeneral adversarial training, our approach improves time efficiency by nearly\n4-fold.\n", "date": "2023-11-27", "categories": []}
{"id": "2311.13712", "title": "Data Acquisition: A New Frontier in Data-centric AI", "abstract": "  As Machine Learning (ML) systems continue to grow, the demand for relevant\nand comprehensive datasets becomes imperative. There is limited study on the\nchallenges of data acquisition due to ad-hoc processes and lack of consistent\nmethodologies. We first present an investigation of current data marketplaces,\nrevealing lack of platforms offering detailed information about datasets,\ntransparent pricing, standardized data formats. With the objective of inciting\nparticipation from the data-centric AI community, we then introduce the DAM\nchallenge, a benchmark to model the interaction between the data providers and\nacquirers. The benchmark was released as a part of DataPerf. Our evaluation of\nthe submitted strategies underlines the need for effective data acquisition\nstrategies in ML.\n", "date": "2023-11-22", "categories": []}
{"id": "2311.12800", "title": "Understanding Data Augmentation from a Robustness Perspective", "abstract": "  In the realm of visual recognition, data augmentation stands out as a pivotal\ntechnique to amplify model robustness. Yet, a considerable number of existing\nmethodologies lean heavily on heuristic foundations, rendering their intrinsic\nmechanisms ambiguous. This manuscript takes both a theoretical and empirical\napproach to understanding the phenomenon. Theoretically, we frame the discourse\naround data augmentation within game theory's constructs. Venturing deeper, our\nempirical evaluations dissect the intricate mechanisms of emblematic data\naugmentation strategies, illuminating that these techniques primarily stimulate\nmid- and high-order game interactions. Beyond the foundational exploration, our\nexperiments span multiple datasets and diverse augmentation techniques,\nunderscoring the universal applicability of our findings. Recognizing the vast\narray of robustness metrics with intricate correlations, we unveil a\nstreamlined proxy. This proxy not only simplifies robustness assessment but\nalso offers invaluable insights, shedding light on the inherent dynamics of\nmodel game interactions and their relation to overarching system robustness.\nThese insights provide a novel lens through which we can re-evaluate model\nsafety and robustness in visual recognition tasks.\n", "date": "2023-09-07", "categories": []}
{"id": "2305.10429", "title": "DoReMi: Optimizing Data Mixtures Speeds Up Language Model Pretraining", "abstract": "  The mixture proportions of pretraining data domains (e.g., Wikipedia, books,\nweb text) greatly affect language model (LM) performance. In this paper, we\npropose Domain Reweighting with Minimax Optimization (DoReMi), which first\ntrains a small proxy model using group distributionally robust optimization\n(Group DRO) over domains to produce domain weights (mixture proportions)\nwithout knowledge of downstream tasks. We then resample a dataset with these\ndomain weights and train a larger, full-sized model. In our experiments, we use\nDoReMi on a 280M-parameter proxy model to set the domain weights for training\nan 8B-parameter model (30x larger) more efficiently. On The Pile, DoReMi\nimproves perplexity across all domains, even when it downweights a domain.\nDoReMi improves average few-shot downstream accuracy by 6.5% points over a\nbaseline model trained using The Pile's default domain weights and reaches the\nbaseline accuracy with 2.6x fewer training steps. On the GLaM dataset, DoReMi,\nwhich has no knowledge of downstream tasks, even matches the performance of\nusing domain weights tuned on downstream tasks.\n", "date": "2023-05-17", "categories": []}
{"id": "2311.12537", "title": "Oasis: Data Curation and Assessment System for Pretraining of Large\n  Language Models", "abstract": "  Data is one of the most critical elements in building a large language model.\nHowever, existing systems either fail to customize a corpus curation pipeline\nor neglect to leverage comprehensive corpus assessment for iterative\noptimization of the curation. To this end, we present a pretraining corpus\ncuration and assessment platform called Oasis -- a one-stop system for data\nquality improvement and quantification with user-friendly interactive\ninterfaces. Specifically, the interactive modular rule filter module can devise\ncustomized rules according to explicit feedback. The debiased neural filter\nmodule builds the quality classification dataset in a negative-centric manner\nto remove the undesired bias. The adaptive document deduplication module could\nexecute large-scale deduplication with limited memory resources. These three\nparts constitute the customized data curation module. And in the holistic data\nassessment module, a corpus can be assessed in local and global views, with\nthree evaluation means including human, GPT-4, and heuristic metrics. We\nexhibit a complete process to use Oasis for the curation and assessment of\npretraining data. In addition, an 800GB bilingual corpus curated by Oasis is\npublicly released.\n", "date": "2023-11-21", "categories": []}
{"id": "2302.03169", "title": "Data Selection for Language Models via Importance Resampling", "abstract": "  Selecting a suitable pretraining dataset is crucial for both general-domain\n(e.g., GPT-3) and domain-specific (e.g., Codex) language models (LMs). We\nformalize this problem as selecting a subset of a large raw unlabeled dataset\nto match a desired target distribution given unlabeled target samples. Due to\nthe scale and dimensionality of the raw text data, existing methods use simple\nheuristics or require human experts to manually curate data. Instead, we extend\nthe classic importance resampling approach used in low-dimensions for LM data\nselection. We propose Data Selection with Importance Resampling (DSIR), an\nefficient and scalable framework that estimates importance weights in a reduced\nfeature space for tractability and selects data with importance resampling\naccording to these weights. We instantiate the DSIR framework with hashed\nn-gram features for efficiency, enabling the selection of 100M documents from\nthe full Pile dataset in 4.5 hours. To measure whether hashed n-gram features\npreserve the aspects of the data that are relevant to the target, we define KL\nreduction, a data metric that measures the proximity between the selected\npretraining data and the target on some feature space. Across 8 data selection\nmethods (including expert selection), KL reduction on hashed n-gram features\nhighly correlates with average downstream accuracy (r=0.82). When selecting\ndata for continued pretraining on a specific domain, DSIR performs comparably\nto expert curation across 8 target distributions. When pretraining\ngeneral-domain models (target is Wikipedia and books), DSIR improves over\nrandom selection and heuristic filtering baselines by 2-2.5% on the GLUE\nbenchmark. Code is available at https://github.com/p-lambda/dsir.\n", "date": "2023-02-06", "categories": []}
{"id": "2310.08782", "title": "Selectivity Drives Productivity: Efficient Dataset Pruning for Enhanced\n  Transfer Learning", "abstract": "  Massive data is often considered essential for deep learning applications,\nbut it also incurs significant computational and infrastructural costs.\nTherefore, dataset pruning (DP) has emerged as an effective way to improve data\nefficiency by identifying and removing redundant training samples without\nsacrificing performance. In this work, we aim to address the problem of DP for\ntransfer learning, i.e., how to prune a source dataset for improved pretraining\nefficiency and lossless finetuning accuracy on downstream target tasks. To our\nbest knowledge, the problem of DP for transfer learning remains open, as\nprevious studies have primarily addressed DP and transfer learning as separate\nproblems. By contrast, we establish a unified viewpoint to integrate DP with\ntransfer learning and find that existing DP methods are not suitable for the\ntransfer learning paradigm. We then propose two new DP methods, label mapping\nand feature mapping, for supervised and self-supervised pretraining settings\nrespectively, by revisiting the DP problem through the lens of source-target\ndomain mapping. Furthermore, we demonstrate the effectiveness of our approach\non numerous transfer learning tasks. We show that source data classes can be\npruned by up to 40% ~ 80% without sacrificing downstream performance, resulting\nin a significant 2 ~ 5 times speed-up during the pretraining stage. Besides,\nour proposal exhibits broad applicability and can improve other computationally\nintensive transfer learning techniques, such as adversarial pretraining. Codes\nare available at https://github.com/OPTML-Group/DP4TL.\n", "date": "2023-10-13", "categories": []}
{"id": "2311.08819", "title": "Frequency Domain-based Dataset Distillation", "abstract": "  This paper presents FreD, a novel parameterization method for dataset\ndistillation, which utilizes the frequency domain to distill a small-sized\nsynthetic dataset from a large-sized original dataset. Unlike conventional\napproaches that focus on the spatial domain, FreD employs frequency-based\ntransforms to optimize the frequency representations of each data instance. By\nleveraging the concentration of spatial domain information on specific\nfrequency components, FreD intelligently selects a subset of frequency\ndimensions for optimization, leading to a significant reduction in the required\nbudget for synthesizing an instance. Through the selection of frequency\ndimensions based on the explained variance, FreD demonstrates both theoretical\nand empirical evidence of its ability to operate efficiently within a limited\nbudget, while better preserving the information of the original dataset\ncompared to conventional parameterization methods. Furthermore, based on the\northogonal compatibility of FreD with existing methods, we confirm that FreD\nconsistently improves the performances of existing distillation methods over\nthe evaluation scenarios with different benchmark datasets. We release the code\nat https://github.com/sdh0818/FreD.\n", "date": "2023-11-15", "categories": []}
{"id": "2311.07025", "title": "Embarassingly Simple Dataset Distillation", "abstract": "  Dataset distillation extracts a small set of synthetic training samples from\na large dataset with the goal of achieving competitive performance on test data\nwhen trained on this sample. In this work, we tackle dataset distillation at\nits core by treating it directly as a bilevel optimization problem.\nRe-examining the foundational back-propagation through time method, we study\nthe pronounced variance in the gradients, computational burden, and long-term\ndependencies. We introduce an improved method: Random Truncated Backpropagation\nThrough Time (RaT-BPTT) to address them. RaT-BPTT incorporates a truncation\ncoupled with a random window, effectively stabilizing the gradients and\nspeeding up the optimization while covering long dependencies. This allows us\nto establish new state-of-the-art for a variety of standard dataset benchmarks.\nA deeper dive into the nature of distilled data unveils pronounced\nintercorrelation. In particular, subsets of distilled datasets tend to exhibit\nmuch worse performance than directly distilled smaller datasets of the same\nsize. Leveraging RaT-BPTT, we devise a boosting mechanism that generates\ndistilled datasets that contain subsets with near optimal performance across\ndifferent data budgets.\n", "date": "2023-11-13", "categories": []}
{"id": "2302.06755", "title": "Dataset Distillation with Convexified Implicit Gradients", "abstract": "  We propose a new dataset distillation algorithm using reparameterization and\nconvexification of implicit gradients (RCIG), that substantially improves the\nstate-of-the-art. To this end, we first formulate dataset distillation as a\nbi-level optimization problem. Then, we show how implicit gradients can be\neffectively used to compute meta-gradient updates. We further equip the\nalgorithm with a convexified approximation that corresponds to learning on top\nof a frozen finite-width neural tangent kernel. Finally, we improve bias in\nimplicit gradients by parameterizing the neural network to enable analytical\ncomputation of final-layer parameters given the body parameters. RCIG\nestablishes the new state-of-the-art on a diverse series of dataset\ndistillation tasks. Notably, with one image per class, on resized ImageNet,\nRCIG sees on average a 108\\% improvement over the previous state-of-the-art\ndistillation algorithm. Similarly, we observed a 66\\% gain over SOTA on\nTiny-ImageNet and 37\\% on CIFAR-100.\n", "date": "2023-02-13", "categories": []}
{"id": "2311.05265", "title": "Don't Waste a Single Annotation: Improving Single-Label Classifiers\n  Through Soft Labels", "abstract": "  In this paper, we address the limitations of the common data annotation and\ntraining methods for objective single-label classification tasks. Typically,\nwhen annotating such tasks annotators are only asked to provide a single label\nfor each sample and annotator disagreement is discarded when a final hard label\nis decided through majority voting. We challenge this traditional approach,\nacknowledging that determining the appropriate label can be difficult due to\nthe ambiguity and lack of context in the data samples. Rather than discarding\nthe information from such ambiguous annotations, our soft label method makes\nuse of them for training. Our findings indicate that additional annotator\ninformation, such as confidence, secondary label and disagreement, can be used\nto effectively generate soft labels. Training classifiers with these soft\nlabels then leads to improved performance and calibration on the hard label\ntest set.\n", "date": "2023-11-09", "categories": []}
{"id": "2311.05346", "title": "Accelerated Shapley Value Approximation for Data Evaluation", "abstract": "  Data valuation has found various applications in machine learning, such as\ndata filtering, efficient learning and incentives for data sharing. The most\npopular current approach to data valuation is the Shapley value. While popular\nfor its various applications, Shapley value is computationally expensive even\nto approximate, as it requires repeated iterations of training models on\ndifferent subsets of data. In this paper we show that the Shapley value of data\npoints can be approximated more efficiently by leveraging the structural\nproperties of machine learning problems. We derive convergence guarantees on\nthe accuracy of the approximate Shapley value for different learning settings\nincluding Stochastic Gradient Descent with convex and non-convex loss\nfunctions. Our analysis suggests that in fact models trained on small subsets\nare more important in the context of data valuation. Based on this idea, we\ndescribe $\\delta$-Shapley -- a strategy of only using small subsets for the\napproximation. Experiments show that this approach preserves approximate value\nand rank of data, while achieving speedup of up to 9.9x. In pre-trained\nnetworks the approach is found to bring more efficiency in terms of accurate\nevaluation using small subsets.\n", "date": "2023-11-09", "categories": []}
{"id": "2311.05538", "title": "Embedding Space Interpolation Beyond Mini-Batch, Beyond Pairs and Beyond\n  Examples", "abstract": "  Mixup refers to interpolation-based data augmentation, originally motivated\nas a way to go beyond empirical risk minimization (ERM). Its extensions mostly\nfocus on the definition of interpolation and the space (input or feature) where\nit takes place, while the augmentation process itself is less studied. In most\nmethods, the number of generated examples is limited to the mini-batch size and\nthe number of examples being interpolated is limited to two (pairs), in the\ninput space.\n  We make progress in this direction by introducing MultiMix, which generates\nan arbitrarily large number of interpolated examples beyond the mini-batch size\nand interpolates the entire mini-batch in the embedding space. Effectively, we\nsample on the entire convex hull of the mini-batch rather than along linear\nsegments between pairs of examples.\n  On sequence data, we further extend to Dense MultiMix. We densely interpolate\nfeatures and target labels at each spatial location and also apply the loss\ndensely. To mitigate the lack of dense labels, we inherit labels from examples\nand weight interpolation factors by attention as a measure of confidence.\n  Overall, we increase the number of loss terms per mini-batch by orders of\nmagnitude at little additional cost. This is only possible because of\ninterpolating in the embedding space. We empirically show that our solutions\nyield significant improvement over state-of-the-art mixup methods on four\ndifferent benchmarks, despite interpolation being only linear. By analyzing the\nembedding space, we show that the classes are more tightly clustered and\nuniformly spread over the embedding space, thereby explaining the improved\nbehavior.\n", "date": "2023-11-09", "categories": []}
{"id": "2311.04345", "title": "A Taxonomy of Rater Disagreements: Surveying Challenges & Opportunities\n  from the Perspective of Annotating Online Toxicity", "abstract": "  Toxicity is an increasingly common and severe issue in online spaces.\nConsequently, a rich line of machine learning research over the past decade has\nfocused on computationally detecting and mitigating online toxicity. These\nefforts crucially rely on human-annotated datasets that identify toxic content\nof various kinds in social media texts. However, such annotations historically\nyield low inter-rater agreement, which was often dealt with by taking the\nmajority vote or other such approaches to arrive at a single ground truth\nlabel. Recent research has pointed out the importance of accounting for the\nsubjective nature of this task when building and utilizing these datasets, and\nthis has triggered work on analyzing and better understanding rater\ndisagreements, and how they could be effectively incorporated into the machine\nlearning developmental pipeline. While these efforts are filling an important\ngap, there is a lack of a broader framework about the root causes of rater\ndisagreement, and therefore, we situate this work within that broader\nlandscape. In this survey paper, we analyze a broad set of literature on the\nreasons behind rater disagreements focusing on online toxicity, and propose a\ndetailed taxonomy for the same. Further, we summarize and discuss the potential\nsolutions targeting each reason for disagreement. We also discuss several open\nissues, which could promote the future development of online toxicity research.\n", "date": "2023-11-07", "categories": []}
{"id": "2212.05404", "title": "Cap2Aug: Caption guided Image to Image data Augmentation", "abstract": "  Visual recognition in a low-data regime is challenging and often prone to\noverfitting. To mitigate this issue, several data augmentation strategies have\nbeen proposed. However, standard transformations, e.g., rotation, cropping, and\nflipping provide limited semantic variations. To this end, we propose Cap2Aug,\nan image-to-image diffusion model-based data augmentation strategy using image\ncaptions as text prompts. We generate captions from the limited training images\nand using these captions edit the training images using an image-to-image\nstable diffusion model to generate semantically meaningful augmentations. This\nstrategy generates augmented versions of images similar to the training images\nyet provides semantic diversity across the samples. We show that the variations\nwithin the class can be captured by the captions and then translated to\ngenerate diverse samples using the image-to-image diffusion model guided by the\ncaptions. However, naive learning on synthetic images is not adequate due to\nthe domain gap between real and synthetic images. Thus, we employ a maximum\nmean discrepancy (MMD) loss to align the synthetic images to the real images\nfor minimizing the domain gap. We evaluate our method on few-shot and long-tail\nclassification tasks and obtain performance improvements over state-of-the-art,\nespecially in the low-data regimes.\n", "date": "2022-12-11", "categories": []}
{"id": "2308.10544", "title": "Towards Accelerated Model Training via Bayesian Data Selection", "abstract": "  Mislabeled, duplicated, or biased data in real-world scenarios can lead to\nprolonged training and even hinder model convergence. Traditional solutions\nprioritizing easy or hard samples lack the flexibility to handle such a variety\nsimultaneously. Recent work has proposed a more reasonable data selection\nprinciple by examining the data's impact on the model's generalization loss.\nHowever, its practical adoption relies on less principled approximations and\nadditional holdout data. This work solves these problems by leveraging a\nlightweight Bayesian treatment and incorporating off-the-shelf zero-shot\npredictors built on large-scale pre-trained models. The resulting algorithm is\nefficient and easy to implement. We perform extensive empirical studies on\nchallenging benchmarks with considerable data noise and imbalance in the online\nbatch selection scenario, and observe superior training efficiency over\ncompetitive baselines. Notably, on the challenging WebVision benchmark, our\nmethod can achieve similar predictive performance with significantly fewer\ntraining iterations than leading data selection methods.\n", "date": "2023-08-21", "categories": []}
{"id": "2302.06960", "title": "Data pruning and neural scaling laws: fundamental limitations of\n  score-based algorithms", "abstract": "  Data pruning algorithms are commonly used to reduce the memory and\ncomputational cost of the optimization process. Recent empirical results reveal\nthat random data pruning remains a strong baseline and outperforms most\nexisting data pruning methods in the high compression regime, i.e., where a\nfraction of $30\\%$ or less of the data is kept. This regime has recently\nattracted a lot of interest as a result of the role of data pruning in\nimproving the so-called neural scaling laws; in [Sorscher et al.], the authors\nshowed the need for high-quality data pruning algorithms in order to beat the\nsample power law.\n  In this work, we focus on score-based data pruning algorithms and show\ntheoretically and empirically why such algorithms fail in the high compression\nregime. We demonstrate ``No Free Lunch\" theorems for data pruning and present\ncalibration protocols that enhance the performance of existing pruning\nalgorithms in this high compression regime using randomization.\n", "date": "2023-02-14", "categories": []}
{"id": "2309.17425", "title": "Data Filtering Networks", "abstract": "  Large training sets have become a cornerstone of machine learning and are the\nfoundation for recent advances in language modeling and multimodal learning.\nWhile data curation for pre-training is often still ad-hoc, one common paradigm\nis to first collect a massive pool of data from the Web and then filter this\ncandidate pool down to an actual training set via various heuristics. In this\nwork, we study the problem of learning a data filtering network (DFN) for this\nsecond step of filtering a large uncurated dataset. Our key finding is that the\nquality of a network for filtering is distinct from its performance on\ndownstream tasks: for instance, a model that performs well on ImageNet can\nyield worse training sets than a model with low ImageNet accuracy that is\ntrained on a small amount of high-quality data. Based on our insights, we\nconstruct new data filtering networks that induce state-of-the-art image-text\ndatasets. Specifically, our best performing dataset DFN-5B enables us to train\nstate-of-the-art CLIP models for their compute budgets: among other\nimprovements on a variety of tasks, a ViT-H trained on our dataset achieves\n84.4% zero-shot transfer accuracy on ImageNet, out-performing models trained on\nother datasets such as LAION-2B, DataComp-1B, or OpenAI's WIT. In order to\nfacilitate further research in dataset design, we also release a new 2 billion\nexample dataset DFN-2B and show that high performance data filtering networks\ncan be trained from scratch using only publicly available data.\n", "date": "2023-09-29", "categories": []}
{"id": "2311.02709", "title": "Benchmarking a Benchmark: How Reliable is MS-COCO?", "abstract": "  Benchmark datasets are used to profile and compare algorithms across a\nvariety of tasks, ranging from image classification to segmentation, and also\nplay a large role in image pretraining algorithms. Emphasis is placed on\nresults with little regard to the actual content within the dataset. It is\nimportant to question what kind of information is being learned from these\ndatasets and what are the nuances and biases within them. In the following\nwork, Sama-COCO, a re-annotation of MS-COCO, is used to discover potential\nbiases by leveraging a shape analysis pipeline. A model is trained and\nevaluated on both datasets to examine the impact of different annotation\nconditions. Results demonstrate that annotation styles are important and that\nannotation pipelines should closely consider the task of interest. The dataset\nis made publicly available at https://www.sama.com/sama-coco-dataset/ .\n", "date": "2023-11-05", "categories": []}
{"id": "2311.01570", "title": "Sequential Subset Matching for Dataset Distillation", "abstract": "  Dataset distillation is a newly emerging task that synthesizes a small-size\ndataset used in training deep neural networks (DNNs) for reducing data storage\nand model training costs. The synthetic datasets are expected to capture the\nessence of the knowledge contained in real-world datasets such that the former\nyields a similar performance as the latter. Recent advancements in distillation\nmethods have produced notable improvements in generating synthetic datasets.\nHowever, current state-of-the-art methods treat the entire synthetic dataset as\na unified entity and optimize each synthetic instance equally. This static\noptimization approach may lead to performance degradation in dataset\ndistillation. Specifically, we argue that static optimization can give rise to\na coupling issue within the synthetic data, particularly when a larger amount\nof synthetic data is being optimized. This coupling issue, in turn, leads to\nthe failure of the distilled dataset to extract the high-level features learned\nby the deep neural network (DNN) in the latter epochs.\n  In this study, we propose a new dataset distillation strategy called\nSequential Subset Matching (SeqMatch), which tackles this problem by adaptively\noptimizing the synthetic data to encourage sequential acquisition of knowledge\nduring dataset distillation. Our analysis indicates that SeqMatch effectively\naddresses the coupling issue by sequentially generating the synthetic\ninstances, thereby enhancing its performance significantly. Our proposed\nSeqMatch outperforms state-of-the-art methods in various datasets, including\nSVNH, CIFAR-10, CIFAR-100, and Tiny ImageNet. Our code is available at\nhttps://github.com/shqii1j/seqmatch.\n", "date": "2023-11-02", "categories": []}
{"id": "2311.01002", "title": "Robust Data Pruning under Label Noise via Maximizing Re-labeling\n  Accuracy", "abstract": "  Data pruning, which aims to downsize a large training set into a small\ninformative subset, is crucial for reducing the enormous computational costs of\nmodern deep learning. Though large-scale data collections invariably contain\nannotation noise and numerous robust learning methods have been developed, data\npruning for the noise-robust learning scenario has received little attention.\nWith state-of-the-art Re-labeling methods that self-correct erroneous labels\nwhile training, it is challenging to identify which subset induces the most\naccurate re-labeling of erroneous labels in the entire training set. In this\npaper, we formalize the problem of data pruning with re-labeling. We first show\nthat the likelihood of a training example being correctly re-labeled is\nproportional to the prediction confidence of its neighborhood in the subset.\nTherefore, we propose a novel data pruning algorithm, Prune4Rel, that finds a\nsubset maximizing the total neighborhood confidence of all training examples,\nthereby maximizing the re-labeling accuracy and generalization performance.\nExtensive experiments on four real and one synthetic noisy datasets show that\n\\algname{} outperforms the baselines with Re-labeling models by up to 9.1% as\nwell as those with a standard model by up to 21.6%.\n", "date": "2023-11-02", "categories": []}
{"id": "2211.10586", "title": "Scaling Up Dataset Distillation to ImageNet-1K with Constant Memory", "abstract": "  Dataset Distillation is a newly emerging area that aims to distill large\ndatasets into much smaller and highly informative synthetic ones to accelerate\ntraining and reduce storage. Among various dataset distillation methods,\ntrajectory-matching-based methods (MTT) have achieved SOTA performance in many\ntasks, e.g., on CIFAR-10/100. However, due to exorbitant memory consumption\nwhen unrolling optimization through SGD steps, MTT fails to scale to\nlarge-scale datasets such as ImageNet-1K. Can we scale this SOTA method to\nImageNet-1K and does its effectiveness on CIFAR transfer to ImageNet-1K? To\nanswer these questions, we first propose a procedure to exactly compute the\nunrolled gradient with constant memory complexity, which allows us to scale MTT\nto ImageNet-1K seamlessly with ~6x reduction in memory footprint. We further\ndiscover that it is challenging for MTT to handle datasets with a large number\nof classes, and propose a novel soft label assignment that drastically improves\nits convergence. The resulting algorithm sets new SOTA on ImageNet-1K: we can\nscale up to 50 IPCs (Image Per Class) on ImageNet-1K on a single GPU (all\nprevious methods can only scale to 2 IPCs on ImageNet-1K), leading to the best\naccuracy (only 5.9% accuracy drop against full dataset training) while\nutilizing only 4.2% of the number of data points - an 18.2% absolute gain over\nprior SOTA. Our code is available at https://github.com/justincui03/tesla\n", "date": "2022-11-19", "categories": []}
{"id": "2310.20111", "title": "Making Large Language Models Better Data Creators", "abstract": "  Although large language models (LLMs) have advanced the state-of-the-art in\nNLP significantly, deploying them for downstream applications is still\nchallenging due to cost, responsiveness, control, or concerns around privacy\nand security. As such, trainable models are still the preferred option in some\ncases. However, these models still require human-labeled data for optimal\nperformance, which is expensive and time-consuming to obtain. In order to\naddress this issue, several techniques to reduce human effort involve labeling\nor generating data using LLMs. Although these methods are effective for certain\napplications, in practice they encounter difficulties in real-world scenarios.\nLabeling data requires careful data selection, while generating data\nnecessitates task-specific prompt engineering. In this paper, we propose a\nunified data creation pipeline that requires only a single formatting example,\nand which is applicable to a broad range of tasks, including traditionally\nproblematic ones with semantically devoid label spaces. In our experiments we\ndemonstrate that instruction-following LLMs are highly cost-effective data\ncreators, and that models trained with these data exhibit performance better\nthan those trained with human-labeled data (by up to 17.5%) on\nout-of-distribution evaluation, while maintaining comparable performance on\nin-distribution tasks. These results have important implications for the\nrobustness of NLP systems deployed in the real-world.\n", "date": "2023-10-31", "categories": []}
{"id": "2305.16289", "title": "Diversify Your Vision Datasets with Automatic Diffusion-Based\n  Augmentation", "abstract": "  Many fine-grained classification tasks, like rare animal identification, have\nlimited training data and consequently classifiers trained on these datasets\noften fail to generalize to variations in the domain like changes in weather or\nlocation. As such, we explore how natural language descriptions of the domains\nseen in training data can be used with large vision models trained on diverse\npretraining datasets to generate useful variations of the training data. We\nintroduce ALIA (Automated Language-guided Image Augmentation), a method which\nutilizes large vision and language models to automatically generate natural\nlanguage descriptions of a dataset's domains and augment the training data via\nlanguage-guided image editing. To maintain data integrity, a model trained on\nthe original dataset filters out minimal image edits and those which corrupt\nclass-relevant information. The resulting dataset is visually consistent with\nthe original training data and offers significantly enhanced diversity. We show\nthat ALIA is able to surpasses traditional data augmentation and text-to-image\ngenerated data on fine-grained classification tasks, including cases of domain\ngeneralization and contextual bias. Code is available at\nhttps://github.com/lisadunlap/ALIA.\n", "date": "2023-05-25", "categories": []}
{"id": "2310.18970", "title": "TRIAGE: Characterizing and auditing training data for improved\n  regression", "abstract": "  Data quality is crucial for robust machine learning algorithms, with the\nrecent interest in data-centric AI emphasizing the importance of training data\ncharacterization. However, current data characterization methods are largely\nfocused on classification settings, with regression settings largely\nunderstudied. To address this, we introduce TRIAGE, a novel data\ncharacterization framework tailored to regression tasks and compatible with a\nbroad class of regressors. TRIAGE utilizes conformal predictive distributions\nto provide a model-agnostic scoring method, the TRIAGE score. We operationalize\nthe score to analyze individual samples' training dynamics and characterize\nsamples as under-, over-, or well-estimated by the model. We show that TRIAGE's\ncharacterization is consistent and highlight its utility to improve performance\nvia data sculpting/filtering, in multiple regression settings. Additionally,\nbeyond sample level, we show TRIAGE enables new approaches to dataset selection\nand feature acquisition. Overall, TRIAGE highlights the value unlocked by data\ncharacterization in real-world regression applications\n", "date": "2023-10-29", "categories": []}
{"id": "2310.17852", "title": "Function Space Bayesian Pseudocoreset for Bayesian Neural Networks", "abstract": "  A Bayesian pseudocoreset is a compact synthetic dataset summarizing essential\ninformation of a large-scale dataset and thus can be used as a proxy dataset\nfor scalable Bayesian inference. Typically, a Bayesian pseudocoreset is\nconstructed by minimizing a divergence measure between the posterior\nconditioning on the pseudocoreset and the posterior conditioning on the full\ndataset. However, evaluating the divergence can be challenging, particularly\nfor the models like deep neural networks having high-dimensional parameters. In\nthis paper, we propose a novel Bayesian pseudocoreset construction method that\noperates on a function space. Unlike previous methods, which construct and\nmatch the coreset and full data posteriors in the space of model parameters\n(weights), our method constructs variational approximations to the coreset\nposterior on a function space and matches it to the full data posterior in the\nfunction space. By working directly on the function space, our method could\nbypass several challenges that may arise when working on a weight space,\nincluding limited scalability and multi-modality issue. Through various\nexperiments, we demonstrate that the Bayesian pseudocoresets constructed from\nour method enjoys enhanced uncertainty quantification and better robustness\nacross various model architectures.\n", "date": "2023-10-27", "categories": []}
{"id": "2307.10350", "title": "Improving Multimodal Datasets with Image Captioning", "abstract": "  Massive web datasets play a key role in the success of large vision-language\nmodels like CLIP and Flamingo. However, the raw web data is noisy, and existing\nfiltering methods to reduce noise often come at the expense of data diversity.\nOur work focuses on caption quality as one major source of noise, and studies\nhow generated captions can increase the utility of web-scraped datapoints with\nnondescript text. Through exploring different mixing strategies for raw and\ngenerated captions, we outperform the best filtering method proposed by the\nDataComp benchmark by 2% on ImageNet and 4% on average across 38 tasks, given a\ncandidate pool of 128M image-text pairs. Our best approach is also 2x better at\nFlickr and MS-COCO retrieval. We then analyze what makes synthetic captions an\neffective source of text supervision. In experimenting with different image\ncaptioning models, we also demonstrate that the performance of a model on\nstandard image captioning benchmarks (e.g., NoCaps CIDEr) is not a reliable\nindicator of the utility of the captions it generates for multimodal training.\nFinally, our experiments with using generated captions at DataComp's large\nscale (1.28B image-text pairs) offer insights into the limitations of synthetic\ntext, as well as the importance of image curation with increasing training data\nquantity. The synthetic captions used in our experiments are now available on\nHuggingFace.\n", "date": "2023-07-19", "categories": []}
{"id": "2310.16981", "title": "Reimagining Synthetic Tabular Data Generation through Data-Centric AI: A\n  Comprehensive Benchmark", "abstract": "  Synthetic data serves as an alternative in training machine learning models,\nparticularly when real-world data is limited or inaccessible. However, ensuring\nthat synthetic data mirrors the complex nuances of real-world data is a\nchallenging task. This paper addresses this issue by exploring the potential of\nintegrating data-centric AI techniques which profile the data to guide the\nsynthetic data generation process. Moreover, we shed light on the often ignored\nconsequences of neglecting these data profiles during synthetic data generation\n-- despite seemingly high statistical fidelity. Subsequently, we propose a\nnovel framework to evaluate the integration of data profiles to guide the\ncreation of more representative synthetic data. In an empirical study, we\nevaluate the performance of five state-of-the-art models for tabular data\ngeneration on eleven distinct tabular datasets. The findings offer critical\ninsights into the successes and limitations of current synthetic data\ngeneration techniques. Finally, we provide practical recommendations for\nintegrating data-centric insights into the synthetic data generation process,\nwith a specific focus on classification performance, model selection, and\nfeature selection. This study aims to reevaluate conventional approaches to\nsynthetic data generation and promote the application of data-centric AI\ntechniques in improving the quality and effectiveness of synthetic data.\n", "date": "2023-10-25", "categories": []}
{"id": "2310.14664", "title": "Data Pruning via Moving-one-Sample-out", "abstract": "  In this paper, we propose a novel data-pruning approach called\nmoving-one-sample-out (MoSo), which aims to identify and remove the least\ninformative samples from the training set. The core insight behind MoSo is to\ndetermine the importance of each sample by assessing its impact on the optimal\nempirical risk. This is achieved by measuring the extent to which the empirical\nrisk changes when a particular sample is excluded from the training set.\nInstead of using the computationally expensive leaving-one-out-retraining\nprocedure, we propose an efficient first-order approximator that only requires\ngradient information from different training stages. The key idea behind our\napproximation is that samples with gradients that are consistently aligned with\nthe average gradient of the training set are more informative and should\nreceive higher scores, which could be intuitively understood as follows: if the\ngradient from a specific sample is consistent with the average gradient vector,\nit implies that optimizing the network using the sample will yield a similar\neffect on all remaining samples. Experimental results demonstrate that MoSo\neffectively mitigates severe performance degradation at high pruning ratios and\nachieves satisfactory performance across various settings.\n", "date": "2023-10-23", "categories": []}
{"id": "2310.16499", "title": "Data Optimization in Deep Learning: A Survey", "abstract": "  Large-scale, high-quality data are considered an essential factor for the\nsuccessful application of many deep learning techniques. Meanwhile, numerous\nreal-world deep learning tasks still have to contend with the lack of\nsufficient amounts of high-quality data. Additionally, issues such as model\nrobustness, fairness, and trustworthiness are also closely related to training\ndata. Consequently, a huge number of studies in the existing literature have\nfocused on the data aspect in deep learning tasks. Some typical data\noptimization techniques include data augmentation, logit perturbation, sample\nweighting, and data condensation. These techniques usually come from different\ndeep learning divisions and their theoretical inspirations or heuristic\nmotivations may seem unrelated to each other. This study aims to organize a\nwide range of existing data optimization methodologies for deep learning from\nthe previous literature, and makes the effort to construct a comprehensive\ntaxonomy for them. The constructed taxonomy considers the diversity of split\ndimensions, and deep sub-taxonomies are constructed for each dimension. On the\nbasis of the taxonomy, connections among the extensive data optimization\nmethods for deep learning are built in terms of four aspects. We probe into\nrendering several promising and interesting future directions. The constructed\ntaxonomy and the revealed connections will enlighten the better understanding\nof existing methods and the design of novel data optimization techniques.\nFurthermore, our aspiration for this survey is to promote data optimization as\nan independent subdivision of deep learning. A curated, up-to-date list of\nresources related to data optimization in deep learning is available at\n\\url{https://github.com/YaoRujing/Data-Optimization}.\n", "date": "2023-10-25", "categories": []}
{"id": "2306.02664", "title": "Structure-free Graph Condensation: From Large-scale Graphs to Condensed\n  Graph-free Data", "abstract": "  Graph condensation, which reduces the size of a large-scale graph by\nsynthesizing a small-scale condensed graph as its substitution, has immediate\nbenefits for various graph learning tasks. However, existing graph condensation\nmethods rely on the joint optimization of nodes and structures in the condensed\ngraph, and overlook critical issues in effectiveness and generalization\nability. In this paper, we advocate a new Structure-Free Graph Condensation\nparadigm, named SFGC, to distill a large-scale graph into a small-scale graph\nnode set without explicit graph structures, i.e., graph-free data. Our idea is\nto implicitly encode topology structure information into the node attributes in\nthe synthesized graph-free data, whose topology is reduced to an identity\nmatrix. Specifically, SFGC contains two collaborative components: (1) a\ntraining trajectory meta-matching scheme for effectively synthesizing\nsmall-scale graph-free data; (2) a graph neural feature score metric for\ndynamically evaluating the quality of the condensed data. Through training\ntrajectory meta-matching, SFGC aligns the long-term GNN learning behaviors\nbetween the large-scale graph and the condensed small-scale graph-free data,\nensuring comprehensive and compact transfer of informative knowledge to the\ngraph-free data. Afterward, the underlying condensed graph-free data would be\ndynamically evaluated with the graph neural feature score, which is a\nclosed-form metric for ensuring the excellent expressiveness of the condensed\ngraph-free data. Extensive experiments verify the superiority of SFGC across\ndifferent condensation ratios.\n", "date": "2023-06-05", "categories": []}
{"id": "2307.13899", "title": "Regularizing Neural Networks with Meta-Learning Generative Models", "abstract": "  This paper investigates methods for improving generative data augmentation\nfor deep learning. Generative data augmentation leverages the synthetic samples\nproduced by generative models as an additional dataset for classification with\nsmall dataset settings. A key challenge of generative data augmentation is that\nthe synthetic data contain uninformative samples that degrade accuracy. This is\nbecause the synthetic samples do not perfectly represent class categories in\nreal data and uniform sampling does not necessarily provide useful samples for\ntasks. In this paper, we present a novel strategy for generative data\naugmentation called meta generative regularization (MGR). To avoid the\ndegradation of generative data augmentation, MGR utilizes synthetic samples in\nthe regularization term for feature extractors instead of in the loss function,\ne.g., cross-entropy. These synthetic samples are dynamically determined to\nminimize the validation losses through meta-learning. We observed that MGR can\navoid the performance degradation of na\\\"ive generative data augmentation and\nboost the baselines. Experiments on six datasets showed that MGR is effective\nparticularly when datasets are smaller and stably outperforms baselines.\n", "date": "2023-07-26", "categories": []}
{"id": "2310.14019", "title": "You Only Condense Once: Two Rules for Pruning Condensed Datasets", "abstract": "  Dataset condensation is a crucial tool for enhancing training efficiency by\nreducing the size of the training dataset, particularly in on-device scenarios.\nHowever, these scenarios have two significant challenges: 1) the varying\ncomputational resources available on the devices require a dataset size\ndifferent from the pre-defined condensed dataset, and 2) the limited\ncomputational resources often preclude the possibility of conducting additional\ncondensation processes. We introduce You Only Condense Once (YOCO) to overcome\nthese limitations. On top of one condensed dataset, YOCO produces smaller\ncondensed datasets with two embarrassingly simple dataset pruning rules: Low\nLBPE Score and Balanced Construction. YOCO offers two key advantages: 1) it can\nflexibly resize the dataset to fit varying computational constraints, and 2) it\neliminates the need for extra condensation processes, which can be\ncomputationally prohibitive. Experiments validate our findings on networks\nincluding ConvNet, ResNet and DenseNet, and datasets including CIFAR-10,\nCIFAR-100 and ImageNet. For example, our YOCO surpassed various dataset\ncondensation and dataset pruning methods on CIFAR-10 with ten Images Per Class\n(IPC), achieving 6.98-8.89% and 6.31-23.92% accuracy gains, respectively. The\ncode is available at: https://github.com/he-y/you-only-condense-once.\n", "date": "2023-10-21", "categories": []}
{"id": "2310.14413", "title": "Data Augmentation: a Combined Inductive-Deductive Approach featuring\n  Answer Set Programming", "abstract": "  Although the availability of a large amount of data is usually given for\ngranted, there are relevant scenarios where this is not the case; for instance,\nin the biomedical/healthcare domain, some applications require to build huge\ndatasets of proper images, but the acquisition of such images is often hard for\ndifferent reasons (e.g., accessibility, costs, pathology-related variability),\nthus causing limited and usually imbalanced datasets. Hence, the need for\nsynthesizing photo-realistic images via advanced Data Augmentation techniques\nis crucial. In this paper we propose a hybrid inductive-deductive approach to\nthe problem; in particular, starting from a limited set of real labeled images,\nthe proposed framework makes use of logic programs for declaratively specifying\nthe structure of new images, that is guaranteed to comply with both a set of\nconstraints coming from the domain knowledge and some specific desiderata. The\nresulting labeled images undergo a dedicated process based on Deep Learning in\ncharge of creating photo-realistic images that comply with the generated label.\n", "date": "2023-10-22", "categories": []}
{"id": "2310.14581", "title": "Leveraging Image-Text Similarity and Caption Modification for the\n  DataComp Challenge: Filtering Track and BYOD Track", "abstract": "  Large web crawl datasets have already played an important role in learning\nmultimodal features with high generalization capabilities. However, there are\nstill very limited studies investigating the details or improvements of data\ndesign. Recently, a DataComp challenge has been designed to propose the best\ntraining data with the fixed models. This paper presents our solution to both\nfiltering track and BYOD track of the DataComp challenge. Our solution adopts\nlarge multimodal models CLIP and BLIP-2 to filter and modify web crawl data,\nand utilize external datasets along with a bag of tricks to improve the data\nquality. Experiments show our solution significantly outperforms DataComp\nbaselines (filtering track: 6.6% improvement, BYOD track: 48.5% improvement).\n", "date": "2023-10-23", "categories": []}
{"id": "2310.15052", "title": "DREAM+: Efficient Dataset Distillation by Bidirectional Representative\n  Matching", "abstract": "  Dataset distillation plays a crucial role in creating compact datasets with\nsimilar training performance compared with original large-scale ones. This is\nessential for addressing the challenges of data storage and training costs.\nPrevalent methods facilitate knowledge transfer by matching the gradients,\nembedding distributions, or training trajectories of synthetic images with\nthose of the sampled original images. Although there are various matching\nobjectives, currently the strategy for selecting original images is limited to\nnaive random sampling. We argue that random sampling overlooks the evenness of\nthe selected sample distribution, which may result in noisy or biased matching\ntargets. Besides, the sample diversity is also not constrained by random\nsampling. Additionally, current methods predominantly focus on\nsingle-dimensional matching, where information is not fully utilized. To\naddress these challenges, we propose a novel matching strategy called Dataset\nDistillation by Bidirectional REpresentAtive Matching (DREAM+), which selects\nrepresentative original images for bidirectional matching. DREAM+ is applicable\nto a variety of mainstream dataset distillation frameworks and significantly\nreduces the number of distillation iterations by more than 15 times without\naffecting performance. Given sufficient training time, DREAM+ can further\nimprove the performance and achieve state-of-the-art results. We have released\nthe code at github.com/NUS-HPC-AI-Lab/DREAM+.\n", "date": "2023-10-23", "categories": []}
{"id": "2303.04947", "title": "InfoBatch: Lossless Training Speed Up by Unbiased Dynamic Data Pruning", "abstract": "  Data pruning aims to obtain lossless performances with less overall cost. A\ncommon approach is to filter out samples that make less contribution to the\ntraining. This could lead to gradient expectation bias compared to the original\ndata. To solve this problem, we propose \\textbf{InfoBatch}, a novel framework\naiming to achieve lossless training acceleration by unbiased dynamic data\npruning. Specifically, InfoBatch randomly prunes a portion of less informative\nsamples based on the loss distribution and rescales the gradients of the\nremaining samples to approximate the original gradient. As a plug-and-play and\narchitecture-agnostic framework, InfoBatch consistently obtains lossless\ntraining results on classification, semantic segmentation, vision pertaining,\nand instruction fine-tuning tasks. On CIFAR10/100, ImageNet-1K, and ADE20K,\nInfoBatch losslessly saves 40\\% overall cost. For pertaining MAE and diffusion\nmodel, InfoBatch can respectively save 24.8\\% and 27\\% cost. For LLaMA\ninstruction fine-tuning, InfoBatch is also able to save 20\\% cost and is\ncompatible with coreset selection methods. The code is publicly available at\n\\href{https://github.com/henryqin1997/InfoBatch}{github.com/NUS-HPC-AI-Lab/InfoBatch}.\n", "date": "2023-03-08", "categories": []}
{"id": "2304.14108", "title": "DataComp: In search of the next generation of multimodal datasets", "abstract": "  Multimodal datasets are a critical component in recent breakthroughs such as\nStable Diffusion and GPT-4, yet their design does not receive the same research\nattention as model architectures or training algorithms. To address this\nshortcoming in the ML ecosystem, we introduce DataComp, a testbed for dataset\nexperiments centered around a new candidate pool of 12.8 billion image-text\npairs from Common Crawl. Participants in our benchmark design new filtering\ntechniques or curate new data sources and then evaluate their new dataset by\nrunning our standardized CLIP training code and testing the resulting model on\n38 downstream test sets. Our benchmark consists of multiple compute scales\nspanning four orders of magnitude, which enables the study of scaling trends\nand makes the benchmark accessible to researchers with varying resources. Our\nbaseline experiments show that the DataComp workflow leads to better training\nsets. In particular, our best baseline, DataComp-1B, enables training a CLIP\nViT-L/14 from scratch to 79.2% zero-shot accuracy on ImageNet, outperforming\nOpenAI's CLIP ViT-L/14 by 3.7 percentage points while using the same training\nprocedure and compute. We release DataComp and all accompanying code at\nwww.datacomp.ai.\n", "date": "2023-04-27", "categories": []}
{"id": "2310.13671", "title": "Let's Synthesize Step by Step: Iterative Dataset Synthesis with Large\n  Language Models by Extrapolating Errors from Small Models", "abstract": "  *Data Synthesis* is a promising way to train a small model with very little\nlabeled data. One approach for data synthesis is to leverage the rich knowledge\nfrom large language models to synthesize pseudo training examples for small\nmodels, making it possible to achieve both data and compute efficiency at the\nsame time. However, a key challenge in data synthesis is that the synthesized\ndataset often suffers from a large distributional discrepancy from the *real\ntask* data distribution. Thus, in this paper, we propose *Synthesis Step by\nStep* (**S3**), a data synthesis framework that shrinks this distribution gap\nby iteratively extrapolating the errors made by a small model trained on the\nsynthesized dataset on a small real-world validation dataset using a large\nlanguage model. Extensive experiments on multiple NLP tasks show that our\napproach improves the performance of a small model by reducing the gap between\nthe synthetic dataset and the real data, resulting in significant improvement\ncompared to several baselines: 9.48% improvement compared to ZeroGen and 2.73%\ncompared to GoldGen, and at most 15.17% improvement compared to the small model\ntrained on human-annotated data.\n", "date": "2023-10-20", "categories": []}
{"id": "2310.12168", "title": "RK-core: An Established Methodology for Exploring the Hierarchical\n  Structure within Datasets", "abstract": "  Recently, the field of machine learning has undergone a transition from\nmodel-centric to data-centric. The advancements in diverse learning tasks have\nbeen propelled by the accumulation of more extensive datasets, subsequently\nfacilitating the training of larger models on these datasets. However, these\ndatasets remain relatively under-explored. To this end, we introduce a\npioneering approach known as RK-core, to empower gaining a deeper understanding\nof the intricate hierarchical structure within datasets. Across several\nbenchmark datasets, we find that samples with low coreness values appear less\nrepresentative of their respective categories, and conversely, those with high\ncoreness values exhibit greater representativeness. Correspondingly, samples\nwith high coreness values make a more substantial contribution to the\nperformance in comparison to those with low coreness values. Building upon\nthis, we further employ RK-core to analyze the hierarchical structure of\nsamples with different coreset selection methods. Remarkably, we find that a\nhigh-quality coreset should exhibit hierarchical diversity instead of solely\nopting for representative samples. The code is available at\nhttps://github.com/yaolu-zjut/Kcore.\n", "date": "2023-10-10", "categories": []}
{"id": "2302.13959", "title": "Make Every Example Count: On the Stability and Utility of Self-Influence\n  for Learning from Noisy NLP Datasets", "abstract": "  Increasingly larger datasets have become a standard ingredient to advancing\nthe state-of-the-art in NLP. However, data quality might have already become\nthe bottleneck to unlock further gains. Given the diversity and the sizes of\nmodern datasets, standard data filtering is not straight-forward to apply,\nbecause of the multifacetedness of the harmful data and elusiveness of\nfiltering rules that would generalize across multiple tasks. We study the\nfitness of task-agnostic self-influence scores of training examples for data\ncleaning, analyze their efficacy in capturing naturally occurring outliers, and\ninvestigate to what extent self-influence based data cleaning can improve\ndownstream performance in machine translation, question answering and text\nclassification, building up on recent approaches to self-influence calculation\nand automated curriculum learning.\n", "date": "2023-02-27", "categories": []}
{"id": "2310.10955", "title": "A State-Vector Framework for Dataset Effects", "abstract": "  The impressive success of recent deep neural network (DNN)-based systems is\nsignificantly influenced by the high-quality datasets used in training.\nHowever, the effects of the datasets, especially how they interact with each\nother, remain underexplored. We propose a state-vector framework to enable\nrigorous studies in this direction. This framework uses idealized probing test\nresults as the bases of a vector space. This framework allows us to quantify\nthe effects of both standalone and interacting datasets. We show that the\nsignificant effects of some commonly-used language understanding datasets are\ncharacteristic and are concentrated on a few linguistic dimensions.\nAdditionally, we observe some ``spill-over'' effects: the datasets could impact\nthe models along dimensions that may seem unrelated to the intended tasks. Our\nstate-vector framework paves the way for a systematic understanding of the\ndataset effects, a crucial component in responsible and robust model\ndevelopment.\n", "date": "2023-10-17", "categories": []}
{"id": "2212.13381", "title": "MixupE: Understanding and Improving Mixup from Directional Derivative\n  Perspective", "abstract": "  Mixup is a popular data augmentation technique for training deep neural\nnetworks where additional samples are generated by linearly interpolating pairs\nof inputs and their labels. This technique is known to improve the\ngeneralization performance in many learning paradigms and applications. In this\nwork, we first analyze Mixup and show that it implicitly regularizes infinitely\nmany directional derivatives of all orders. Based on this new insight, we\npropose an improved version of Mixup, theoretically justified to deliver better\ngeneralization performance than the vanilla Mixup. To demonstrate the\neffectiveness of the proposed method, we conduct experiments across various\ndomains such as images, tabular data, speech, and graphs. Our results show that\nthe proposed method improves Mixup across multiple datasets using a variety of\narchitectures, for instance, exhibiting an improvement over Mixup by 0.8% in\nImageNet top-1 accuracy.\n", "date": "2022-12-27", "categories": []}
{"id": "2310.08139", "title": "DualAug: Exploiting Additional Heavy Augmentation with OOD Data\n  Rejection", "abstract": "  Data augmentation is a dominant method for reducing model overfitting and\nimproving generalization. Most existing data augmentation methods tend to find\na compromise in augmenting the data, \\textit{i.e.}, increasing the amplitude of\naugmentation carefully to avoid degrading some data too much and doing harm to\nthe model performance. We delve into the relationship between data augmentation\nand model performance, revealing that the performance drop with heavy\naugmentation comes from the presence of out-of-distribution (OOD) data.\nNonetheless, as the same data transformation has different effects for\ndifferent training samples, even for heavy augmentation, there remains part of\nin-distribution data which is beneficial to model training. Based on the\nobservation, we propose a novel data augmentation method, named\n\\textbf{DualAug}, to keep the augmentation in distribution as much as possible\nat a reasonable time and computational cost. We design a data mixing strategy\nto fuse augmented data from both the basic- and the heavy-augmentation\nbranches. Extensive experiments on supervised image classification benchmarks\nshow that DualAug improve various automated data augmentation method. Moreover,\nthe experiments on semi-supervised learning and contrastive self-supervised\nlearning demonstrate that our DualAug can also improve related method. Code is\navailable at\n\\href{https://github.com/shuguang99/DualAug}{https://github.com/shuguang99/DualAug}.\n", "date": "2023-10-12", "categories": []}
{"id": "2310.09983", "title": "Farzi Data: Autoregressive Data Distillation", "abstract": "  We study data distillation for auto-regressive machine learning tasks, where\nthe input and output have a strict left-to-right causal structure. More\nspecifically, we propose Farzi, which summarizes an event sequence dataset into\na small number of synthetic sequences -- Farzi Data -- which are optimized to\nmaintain (if not improve) model performance compared to training on the full\ndataset. Under the hood, Farzi conducts memory-efficient data distillation by\n(i) deriving efficient reverse-mode differentiation of the Adam optimizer by\nleveraging Hessian-Vector Products; and (ii) factorizing the high-dimensional\ndiscrete event-space into a latent-space which provably promotes implicit\nregularization. Empirically, for sequential recommendation and language\nmodeling tasks, we are able to achieve 98-120% of downstream full-data\nperformance when training state-of-the-art models on Farzi Data of size as\nlittle as 0.1% of the original dataset. Notably, being able to train better\nmodels with significantly less data sheds light on the design of future large\nauto-regressive models, and opens up new opportunities to further scale up\nmodel and data sizes.\n", "date": "2023-10-15", "categories": []}
{"id": "2207.10062", "title": "DataPerf: Benchmarks for Data-Centric AI Development", "abstract": "  Machine learning research has long focused on models rather than datasets,\nand prominent datasets are used for common ML tasks without regard to the\nbreadth, difficulty, and faithfulness of the underlying problems. Neglecting\nthe fundamental importance of data has given rise to inaccuracy, bias, and\nfragility in real-world applications, and research is hindered by saturation\nacross existing dataset benchmarks. In response, we present DataPerf, a\ncommunity-led benchmark suite for evaluating ML datasets and data-centric\nalgorithms. We aim to foster innovation in data-centric AI through competition,\ncomparability, and reproducibility. We enable the ML community to iterate on\ndatasets, instead of just architectures, and we provide an open, online\nplatform with multiple rounds of challenges to support this iterative\ndevelopment. The first iteration of DataPerf contains five benchmarks covering\na wide spectrum of data-centric techniques, tasks, and modalities in vision,\nspeech, acquisition, debugging, and diffusion prompting, and we support hosting\nnew contributed benchmarks from the community. The benchmarks, online\nevaluation platform, and baseline implementations are open source, and the\nMLCommons Association will maintain DataPerf to ensure long-term benefits to\nacademia and industry.\n", "date": "2022-07-20", "categories": []}
{"id": "2306.10577", "title": "OpenDataVal: a Unified Benchmark for Data Valuation", "abstract": "  Assessing the quality and impact of individual data points is critical for\nimproving model performance and mitigating undesirable biases within the\ntraining dataset. Several data valuation algorithms have been proposed to\nquantify data quality, however, there lacks a systemic and standardized\nbenchmarking system for data valuation. In this paper, we introduce\nOpenDataVal, an easy-to-use and unified benchmark framework that empowers\nresearchers and practitioners to apply and compare various data valuation\nalgorithms. OpenDataVal provides an integrated environment that includes (i) a\ndiverse collection of image, natural language, and tabular datasets, (ii)\nimplementations of eleven different state-of-the-art data valuation algorithms,\nand (iii) a prediction model API that can import any models in scikit-learn.\nFurthermore, we propose four downstream machine learning tasks for evaluating\nthe quality of data values. We perform benchmarking analysis using OpenDataVal,\nquantifying and comparing the efficacy of state-of-the-art data valuation\napproaches. We find that no single algorithm performs uniformly best across all\ntasks, and an appropriate algorithm should be employed for a user's downstream\ntask. OpenDataVal is publicly available at https://opendataval.github.io with\ncomprehensive documentation. Furthermore, we provide a leaderboard where\nresearchers can evaluate the effectiveness of their own data valuation\nalgorithms.\n", "date": "2023-06-18", "categories": []}
{"id": "2310.09192", "title": "Does Graph Distillation See Like Vision Dataset Counterpart?", "abstract": "  Training on large-scale graphs has achieved remarkable results in graph\nrepresentation learning, but its cost and storage have attracted increasing\nconcerns. Existing graph condensation methods primarily focus on optimizing the\nfeature matrices of condensed graphs while overlooking the impact of the\nstructure information from the original graphs. To investigate the impact of\nthe structure information, we conduct analysis from the spectral domain and\nempirically identify substantial Laplacian Energy Distribution (LED) shifts in\nprevious works. Such shifts lead to poor performance in cross-architecture\ngeneralization and specific tasks, including anomaly detection and link\nprediction. In this paper, we propose a novel Structure-broadcasting Graph\nDataset Distillation (SGDD) scheme for broadcasting the original structure\ninformation to the generation of the synthetic one, which explicitly prevents\noverlooking the original structure information. Theoretically, the synthetic\ngraphs by SGDD are expected to have smaller LED shifts than previous works,\nleading to superior performance in both cross-architecture settings and\nspecific tasks. We validate the proposed SGDD across 9 datasets and achieve\nstate-of-the-art results on all of them: for example, on the YelpChi dataset,\nour approach maintains 98.6% test accuracy of training on the original graph\ndataset with 1,000 times saving on the scale of the graph. Moreover, we\nempirically evaluate there exist 17.6% ~ 31.4% reductions in LED shift crossing\n9 datasets. Extensive experiments and analysis verify the effectiveness and\nnecessity of the proposed designs. The code is available in the GitHub\nrepository: https://github.com/RingBDStack/SGDD.\n", "date": "2023-10-13", "categories": []}
{"id": "2310.07931", "title": "D2 Pruning: Message Passing for Balancing Diversity and Difficulty in\n  Data Pruning", "abstract": "  Analytical theories suggest that higher-quality data can lead to lower test\nerrors in models trained on a fixed data budget. Moreover, a model can be\ntrained on a lower compute budget without compromising performance if a dataset\ncan be stripped of its redundancies. Coreset selection (or data pruning) seeks\nto select a subset of the training data so as to maximize the performance of\nmodels trained on this subset, also referred to as coreset. There are two\ndominant approaches: (1) geometry-based data selection for maximizing data\ndiversity in the coreset, and (2) functions that assign difficulty scores to\nsamples based on training dynamics. Optimizing for data diversity leads to a\ncoreset that is biased towards easier samples, whereas, selection by difficulty\nranking omits easy samples that are necessary for the training of deep learning\nmodels. This demonstrates that data diversity and importance scores are two\ncomplementary factors that need to be jointly considered during coreset\nselection. We represent a dataset as an undirected graph and propose a novel\npruning algorithm, D2 Pruning, that uses forward and reverse message passing\nover this dataset graph for coreset selection. D2 Pruning updates the\ndifficulty scores of each example by incorporating the difficulty of its\nneighboring examples in the dataset graph. Then, these updated difficulty\nscores direct a graph-based sampling method to select a coreset that\nencapsulates both diverse and difficult regions of the dataset space. We\nevaluate supervised and self-supervised versions of our method on various\nvision and language datasets. Results show that D2 Pruning improves coreset\nselection over previous state-of-the-art methods for up to 70% pruning rates.\nAdditionally, we find that using D2 Pruning for filtering large multimodal\ndatasets leads to increased diversity in the dataset and improved\ngeneralization of pretrained models.\n", "date": "2023-10-11", "categories": []}
{"id": "2303.04707", "title": "DiM: Distilling Dataset into Generative Model", "abstract": "  Dataset distillation reduces the network training cost by synthesizing small\nand informative datasets from large-scale ones. Despite the success of the\nrecent dataset distillation algorithms, three drawbacks still limit their wider\napplication: i). the synthetic images perform poorly on large architectures;\nii). they need to be re-optimized when the distillation ratio changes; iii).\nthe limited diversity restricts the performance when the distillation ratio is\nlarge. In this paper, we propose a novel distillation scheme to\n\\textbf{D}istill information of large train sets \\textbf{i}nto generative\n\\textbf{M}odels, named DiM. Specifically, DiM learns to use a generative model\nto store the information of the target dataset. During the distillation phase,\nwe minimize the differences in logits predicted by a models pool between real\nand generated images. At the deployment stage, the generative model synthesizes\nvarious training samples from random noises on the fly. Due to the simple yet\neffective designs, the trained DiM can be directly applied to different\ndistillation ratios and large architectures without extra cost. We validate the\nproposed DiM across 4 datasets and achieve state-of-the-art results on all of\nthem. To the best of our knowledge, we are the first to achieve higher accuracy\non complex architectures than simple ones, such as 75.1\\% with ResNet-18 and\n72.6\\% with ConvNet-3 on ten images per class of CIFAR-10. Besides, DiM\noutperforms previous methods with 10\\% $\\sim$ 22\\% when images per class are 1\nand 10 on the SVHN dataset.\n", "date": "2023-03-08", "categories": []}
{"id": "2310.06982", "title": "Data Distillation Can Be Like Vodka: Distilling More Times For Better\n  Quality", "abstract": "  Dataset distillation aims to minimize the time and memory needed for training\ndeep networks on large datasets, by creating a small set of synthetic images\nthat has a similar generalization performance to that of the full dataset.\nHowever, current dataset distillation techniques fall short, showing a notable\nperformance gap when compared to training on the original data. In this work,\nwe are the first to argue that using just one synthetic subset for distillation\nwill not yield optimal generalization performance. This is because the training\ndynamics of deep networks drastically change during the training. Hence,\nmultiple synthetic subsets are required to capture the training dynamics at\ndifferent phases of training. To address this issue, we propose Progressive\nDataset Distillation (PDD). PDD synthesizes multiple small sets of synthetic\nimages, each conditioned on the previous sets, and trains the model on the\ncumulative union of these subsets without requiring additional training time.\nOur extensive experiments show that PDD can effectively improve the performance\nof existing dataset distillation methods by up to 4.3%. In addition, our method\nfor the first time enable generating considerably larger synthetic datasets.\n", "date": "2023-10-10", "categories": []}
{"id": "2211.13976", "title": "Expanding Small-Scale Datasets with Guided Imagination", "abstract": "  The power of DNNs relies heavily on the quantity and quality of training\ndata. However, collecting and annotating data on a large scale is often\nexpensive and time-consuming. To address this issue, we explore a new task,\ntermed dataset expansion, aimed at expanding a ready-to-use small dataset by\nautomatically creating new labeled samples. To this end, we present a Guided\nImagination Framework (GIF) that leverages cutting-edge generative models like\nDALL-E2 and Stable Diffusion (SD) to \"imagine\" and create informative new data\nfrom the input seed data. Specifically, GIF conducts data imagination by\noptimizing the latent features of the seed data in the semantically meaningful\nspace of the prior model, resulting in the creation of photo-realistic images\nwith new content. To guide the imagination towards creating informative samples\nfor model training, we introduce two key criteria, i.e., class-maintained\ninformation boosting and sample diversity promotion. These criteria are\nverified to be essential for effective dataset expansion: GIF-SD obtains 13.5%\nhigher model accuracy on natural image datasets than unguided expansion with\nSD. With these essential criteria, GIF successfully expands small datasets in\nvarious scenarios, boosting model accuracy by 36.9% on average over six natural\nimage datasets and by 13.5% on average over three medical datasets. The source\ncode is available at https://github.com/Vanint/DatasetExpansion.\n", "date": "2022-11-25", "categories": []}
{"id": "2308.06160", "title": "DatasetDM: Synthesizing Data with Perception Annotations Using Diffusion\n  Models", "abstract": "  Current deep networks are very data-hungry and benefit from training on\nlargescale datasets, which are often time-consuming to collect and annotate. By\ncontrast, synthetic data can be generated infinitely using generative models\nsuch as DALL-E and diffusion models, with minimal effort and cost. In this\npaper, we present DatasetDM, a generic dataset generation model that can\nproduce diverse synthetic images and the corresponding high-quality perception\nannotations (e.g., segmentation masks, and depth). Our method builds upon the\npre-trained diffusion model and extends text-guided image synthesis to\nperception data generation. We show that the rich latent code of the diffusion\nmodel can be effectively decoded as accurate perception annotations using a\ndecoder module. Training the decoder only needs less than 1% (around 100\nimages) manually labeled images, enabling the generation of an infinitely large\nannotated dataset. Then these synthetic data can be used for training various\nperception models for downstream tasks. To showcase the power of the proposed\napproach, we generate datasets with rich dense pixel-wise labels for a wide\nrange of downstream tasks, including semantic segmentation, instance\nsegmentation, and depth estimation. Notably, it achieves 1) state-of-the-art\nresults on semantic segmentation and instance segmentation; 2) significantly\nmore robust on domain generalization than using the real data alone; and\nstate-of-the-art results in zero-shot segmentation setting; and 3) flexibility\nfor efficient application and novel task composition (e.g., image editing). The\nproject website and code can be found at\nhttps://weijiawu.github.io/DatasetDM_page/ and\nhttps://github.com/showlab/DatasetDM, respectively\n", "date": "2023-08-11", "categories": []}
{"id": "2301.07014", "title": "Dataset Distillation: A Comprehensive Review", "abstract": "  Recent success of deep learning is largely attributed to the sheer amount of\ndata used for training deep neural networks.Despite the unprecedented success,\nthe massive data, unfortunately, significantly increases the burden on storage\nand transmission and further gives rise to a cumbersome model training process.\nBesides, relying on the raw data for training \\emph{per se} yields concerns\nabout privacy and copyright. To alleviate these shortcomings, dataset\ndistillation~(DD), also known as dataset condensation (DC), was introduced and\nhas recently attracted much research attention in the community. Given an\noriginal dataset, DD aims to derive a much smaller dataset containing synthetic\nsamples, based on which the trained models yield performance comparable with\nthose trained on the original dataset. In this paper, we give a comprehensive\nreview and summary of recent advances in DD and its application. We first\nintroduce the task formally and propose an overall algorithmic framework\nfollowed by all existing DD methods. Next, we provide a systematic taxonomy of\ncurrent methodologies in this area, and discuss their theoretical\ninterconnections. We also present current challenges in DD through extensive\nexperiments and envision possible directions for future works.\n", "date": "2023-01-17", "categories": []}
{"id": "2210.11717", "title": "A Survey of Dataset Refinement for Problems in Computer Vision Datasets", "abstract": "  Large-scale datasets have played a crucial role in the advancement of\ncomputer vision. However, they often suffer from problems such as class\nimbalance, noisy labels, dataset bias, or high resource costs, which can\ninhibit model performance and reduce trustworthiness. With the advocacy of\ndata-centric research, various data-centric solutions have been proposed to\nsolve the dataset problems mentioned above. They improve the quality of\ndatasets by re-organizing them, which we call dataset refinement. In this\nsurvey, we provide a comprehensive and structured overview of recent advances\nin dataset refinement for problematic computer vision datasets. Firstly, we\nsummarize and analyze the various problems encountered in large-scale computer\nvision datasets. Then, we classify the dataset refinement algorithms into three\ncategories based on the refinement process: data sampling, data subset\nselection, and active learning. In addition, we organize these dataset\nrefinement methods according to the addressed data problems and provide a\nsystematic comparative description. We point out that these three types of\ndataset refinement have distinct advantages and disadvantages for dataset\nproblems, which informs the choice of the data-centric method appropriate to a\nparticular research objective. Finally, we summarize the current literature and\npropose potential future research topics.\n", "date": "2022-10-21", "categories": []}
{"id": "2310.03295", "title": "Can pre-trained models assist in dataset distillation?", "abstract": "  Dataset Distillation (DD) is a prominent technique that encapsulates\nknowledge from a large-scale original dataset into a small synthetic dataset\nfor efficient training. Meanwhile, Pre-trained Models (PTMs) function as\nknowledge repositories, containing extensive information from the original\ndataset. This naturally raises a question: Can PTMs effectively transfer\nknowledge to synthetic datasets, guiding DD accurately? To this end, we conduct\npreliminary experiments, confirming the contribution of PTMs to DD. Afterwards,\nwe systematically study different options in PTMs, including initialization\nparameters, model architecture, training epoch and domain knowledge, revealing\nthat: 1) Increasing model diversity enhances the performance of synthetic\ndatasets; 2) Sub-optimal models can also assist in DD and outperform\nwell-trained ones in certain cases; 3) Domain-specific PTMs are not mandatory\nfor DD, but a reasonable domain match is crucial. Finally, by selecting optimal\noptions, we significantly improve the cross-architecture generalization over\nbaseline DD methods. We hope our work will facilitate researchers to develop\nbetter DD techniques. Our code is available at\nhttps://github.com/yaolu-zjut/DDInterpreter.\n", "date": "2023-10-05", "categories": []}
{"id": "2309.04747", "title": "When to Learn What: Model-Adaptive Data Augmentation Curriculum", "abstract": "  Data augmentation (DA) is widely used to improve the generalization of neural\nnetworks by enforcing the invariances and symmetries to pre-defined\ntransformations applied to input data. However, a fixed augmentation policy may\nhave different effects on each sample in different training stages but existing\napproaches cannot adjust the policy to be adaptive to each sample and the\ntraining model. In this paper, we propose Model Adaptive Data Augmentation\n(MADAug) that jointly trains an augmentation policy network to teach the model\nwhen to learn what. Unlike previous work, MADAug selects augmentation operators\nfor each input image by a model-adaptive policy varying between training\nstages, producing a data augmentation curriculum optimized for better\ngeneralization. In MADAug, we train the policy through a bi-level optimization\nscheme, which aims to minimize a validation-set loss of a model trained using\nthe policy-produced data augmentations. We conduct an extensive evaluation of\nMADAug on multiple image classification tasks and network architectures with\nthorough comparisons to existing DA approaches. MADAug outperforms or is on par\nwith other baselines and exhibits better fairness: it brings improvement to all\nclasses and more to the difficult ones. Moreover, MADAug learned policy shows\nbetter performance when transferred to fine-grained datasets. In addition, the\nauto-optimized policy in MADAug gradually introduces increasing perturbations\nand naturally forms an easy-to-hard curriculum.\n", "date": "2023-09-09", "categories": []}
{"id": "2309.15954", "title": "The Devil is in the Details: A Deep Dive into the Rabbit Hole of Data\n  Filtering", "abstract": "  The quality of pre-training data plays a critical role in the performance of\nfoundation models. Popular foundation models often design their own recipe for\ndata filtering, which makes it hard to analyze and compare different data\nfiltering approaches. DataComp is a new benchmark dedicated to evaluating\ndifferent methods for data filtering. This paper describes our learning and\nsolution when participating in the DataComp challenge. Our filtering strategy\nincludes three stages: single-modality filtering, cross-modality filtering, and\ndata distribution alignment. We integrate existing methods and propose new\nsolutions, such as computing CLIP score on horizontally flipped images to\nmitigate the interference of scene text, using vision and language models to\nretrieve training samples for target downstream tasks, rebalancing the data\ndistribution to improve the efficiency of allocating the computational budget,\netc. We slice and dice our design choices, provide in-depth analysis, and\ndiscuss open questions. Our approach outperforms the best method from the\nDataComp paper by over 4% on the average performance of 38 tasks and by over 2%\non ImageNet.\n", "date": "2023-09-27", "categories": []}
{"id": "2301.04272", "title": "Data Distillation: A Survey", "abstract": "  The popularity of deep learning has led to the curation of a vast number of\nmassive and multifarious datasets. Despite having close-to-human performance on\nindividual tasks, training parameter-hungry models on large datasets poses\nmulti-faceted problems such as (a) high model-training time; (b) slow research\niteration; and (c) poor eco-sustainability. As an alternative, data\ndistillation approaches aim to synthesize terse data summaries, which can serve\nas effective drop-in replacements of the original dataset for scenarios like\nmodel training, inference, architecture search, etc. In this survey, we present\na formal framework for data distillation, along with providing a detailed\ntaxonomy of existing approaches. Additionally, we cover data distillation\napproaches for different data modalities, namely images, graphs, and user-item\ninteractions (recommender systems), while also identifying current challenges\nand future research directions.\n", "date": "2023-01-11", "categories": []}
{"id": "2309.13770", "title": "Devil in the Number: Towards Robust Multi-modality Data Filter", "abstract": "  In order to appropriately filter multi-modality data sets on a web-scale, it\nbecomes crucial to employ suitable filtering methods to boost performance and\nreduce training costs. For instance, LAION papers employs the CLIP score filter\nto select data with CLIP scores surpassing a certain threshold. On the other\nhand, T-MARS achieves high-quality data filtering by detecting and masking text\nwithin images and then filtering by CLIP score. Through analyzing the dataset,\nwe observe a significant proportion of redundant information, such as numbers,\npresent in the textual content. Our experiments on a subset of the data unveil\nthe profound impact of these redundant elements on the CLIP scores. A logical\napproach would involve reevaluating the CLIP scores after eliminating these\ninfluences. Experimentally, our text-based CLIP filter outperforms the\ntop-ranked method on the ``small scale\" of DataComp (a data filtering\nbenchmark) on ImageNet distribution shifts, achieving a 3.6% performance\nimprovement. The results also demonstrate that our proposed text-masked filter\noutperforms the original CLIP score filter when selecting the top 40% of the\ndata. The impact of numbers on CLIP and their handling provide valuable\ninsights for improving the effectiveness of CLIP training, including language\nrewrite techniques.\n", "date": "2023-09-24", "categories": []}
{"id": "2303.08983", "title": "Reinforce Data, Multiply Impact: Improved Model Accuracy and Robustness\n  with Dataset Reinforcement", "abstract": "  We propose Dataset Reinforcement, a strategy to improve a dataset once such\nthat the accuracy of any model architecture trained on the reinforced dataset\nis improved at no additional training cost for users. We propose a Dataset\nReinforcement strategy based on data augmentation and knowledge distillation.\nOur generic strategy is designed based on extensive analysis across CNN- and\ntransformer-based models and performing large-scale study of distillation with\nstate-of-the-art models with various data augmentations. We create a reinforced\nversion of the ImageNet training dataset, called ImageNet+, as well as\nreinforced datasets CIFAR-100+, Flowers-102+, and Food-101+. Models trained\nwith ImageNet+ are more accurate, robust, and calibrated, and transfer well to\ndownstream tasks (e.g., segmentation and detection). As an example, the\naccuracy of ResNet-50 improves by 1.7% on the ImageNet validation set, 3.5% on\nImageNetV2, and 10.0% on ImageNet-R. Expected Calibration Error (ECE) on the\nImageNet validation set is also reduced by 9.9%. Using this backbone with\nMask-RCNN for object detection on MS-COCO, the mean average precision improves\nby 0.8%. We reach similar gains for MobileNets, ViTs, and Swin-Transformers.\nFor MobileNetV3 and Swin-Tiny, we observe significant improvements on\nImageNet-R/A/C of up to 20% improved robustness. Models pretrained on ImageNet+\nand fine-tuned on CIFAR-100+, Flowers-102+, and Food-101+, reach up to 3.4%\nimproved accuracy. The code, datasets, and pretrained models are available at\nhttps://github.com/apple/ml-dr.\n", "date": "2023-03-15", "categories": []}
{"id": "2309.10979", "title": "Towards Data-centric Graph Machine Learning: Review and Outlook", "abstract": "  Data-centric AI, with its primary focus on the collection, management, and\nutilization of data to drive AI models and applications, has attracted\nincreasing attention in recent years. In this article, we conduct an in-depth\nand comprehensive review, offering a forward-looking outlook on the current\nefforts in data-centric AI pertaining to graph data-the fundamental data\nstructure for representing and capturing intricate dependencies among massive\nand diverse real-life entities. We introduce a systematic framework,\nData-centric Graph Machine Learning (DC-GML), that encompasses all stages of\nthe graph data lifecycle, including graph data collection, exploration,\nimprovement, exploitation, and maintenance. A thorough taxonomy of each stage\nis presented to answer three critical graph-centric questions: (1) how to\nenhance graph data availability and quality; (2) how to learn from graph data\nwith limited-availability and low-quality; (3) how to build graph MLOps systems\nfrom the graph data-centric view. Lastly, we pinpoint the future prospects of\nthe DC-GML domain, providing insights to navigate its advancements and\napplications.\n", "date": "2023-09-20", "categories": []}
{"id": "2206.09191", "title": "Gender Artifacts in Visual Datasets", "abstract": "  Gender biases are known to exist within large-scale visual datasets and can\nbe reflected or even amplified in downstream models. Many prior works have\nproposed methods for mitigating gender biases, often by attempting to remove\ngender expression information from images. To understand the feasibility and\npracticality of these approaches, we investigate what $\\textit{gender\nartifacts}$ exist within large-scale visual datasets. We define a\n$\\textit{gender artifact}$ as a visual cue that is correlated with gender,\nfocusing specifically on those cues that are learnable by a modern image\nclassifier and have an interpretable human corollary. Through our analyses, we\nfind that gender artifacts are ubiquitous in the COCO and OpenImages datasets,\noccurring everywhere from low-level information (e.g., the mean value of the\ncolor channels) to the higher-level composition of the image (e.g., pose and\nlocation of people). Given the prevalence of gender artifacts, we claim that\nattempts to remove gender artifacts from such datasets are largely infeasible.\nInstead, the responsibility lies with researchers and practitioners to be aware\nthat the distribution of images within datasets is highly gendered and hence\ndevelop methods which are robust to these distributional shifts across groups.\n", "date": "2022-06-18", "categories": []}
{"id": "2209.15031", "title": "Automatic Data Augmentation via Invariance-Constrained Learning", "abstract": "  Underlying data structures, such as symmetries or invariances to\ntransformations, are often exploited to improve the solution of learning tasks.\nHowever, embedding these properties in models or learning algorithms can be\nchallenging and computationally intensive. Data augmentation, on the other\nhand, induces these symmetries during training by applying multiple\ntransformations to the input data. Despite its ubiquity, its effectiveness\ndepends on the choices of which transformations to apply, when to do so, and\nhow often. In fact, there is both empirical and theoretical evidence that the\nindiscriminate use of data augmentation can introduce biases that outweigh its\nbenefits. This work tackles these issues by automatically adapting the data\naugmentation while solving the learning task. To do so, it formulates data\naugmentation as an invariance-constrained learning problem and leverages Monte\nCarlo Markov Chain (MCMC) sampling to solve it. The result is a practical\nalgorithm that not only does away with a priori searches for augmentation\ndistributions, but also dynamically controls if and when data augmentation is\napplied. Our experiments illustrate the performance of this method, which\nachieves state-of-the-art results in automatic data augmentation benchmarks for\nCIFAR datasets. Furthermore, this approach can be used to gather insights on\nthe actual symmetries underlying a learning task.\n", "date": "2022-09-29", "categories": []}
{"id": "2307.12463", "title": "Rethinking Data Distillation: Do Not Overlook Calibration", "abstract": "  Neural networks trained on distilled data often produce over-confident output\nand require correction by calibration methods. Existing calibration methods\nsuch as temperature scaling and mixup work well for networks trained on\noriginal large-scale data. However, we find that these methods fail to\ncalibrate networks trained on data distilled from large source datasets. In\nthis paper, we show that distilled data lead to networks that are not\ncalibratable due to (i) a more concentrated distribution of the maximum logits\nand (ii) the loss of information that is semantically meaningful but unrelated\nto classification tasks. To address this problem, we propose Masked Temperature\nScaling (MTS) and Masked Distillation Training (MDT) which mitigate the\nlimitations of distilled data and achieve better calibration results while\nmaintaining the efficiency of dataset distillation.\n", "date": "2023-07-24", "categories": []}
{"id": "2309.07698", "title": "Dataset Condensation via Generative Model", "abstract": "  Dataset condensation aims to condense a large dataset with a lot of training\nsamples into a small set. Previous methods usually condense the dataset into\nthe pixels format. However, it suffers from slow optimization speed and large\nnumber of parameters to be optimized. When increasing image resolutions and\nclasses, the number of learnable parameters grows accordingly, prohibiting\ncondensation methods from scaling up to large datasets with diverse classes.\nMoreover, the relations among condensed samples have been neglected and hence\nthe feature distribution of condensed samples is often not diverse. To solve\nthese problems, we propose to condense the dataset into another format, a\ngenerative model. Such a novel format allows for the condensation of large\ndatasets because the size of the generative model remains relatively stable as\nthe number of classes or image resolution increases. Furthermore, an\nintra-class and an inter-class loss are proposed to model the relation of\ncondensed samples. Intra-class loss aims to create more diverse samples for\neach class by pushing each sample away from the others of the same class.\nMeanwhile, inter-class loss increases the discriminability of samples by\nwidening the gap between the centers of different classes. Extensive\ncomparisons with state-of-the-art methods and our ablation studies confirm the\neffectiveness of our method and its individual component. To our best\nknowledge, we are the first to successfully conduct condensation on\nImageNet-1k.\n", "date": "2023-09-14", "categories": []}
{"id": "2305.10293", "title": "Infinite Class Mixup", "abstract": "  Mixup is a widely adopted strategy for training deep networks, where\nadditional samples are augmented by interpolating inputs and labels of training\npairs. Mixup has shown to improve classification performance, network\ncalibration, and out-of-distribution generalisation. While effective, a\ncornerstone of Mixup, namely that networks learn linear behaviour patterns\nbetween classes, is only indirectly enforced since the output interpolation is\nperformed at the probability level. This paper seeks to address this limitation\nby mixing the classifiers directly instead of mixing the labels for each mixed\npair. We propose to define the target of each augmented sample as a uniquely\nnew classifier, whose parameters are a linear interpolation of the classifier\nvectors of the input pair. The space of all possible classifiers is continuous\nand spans all interpolations between classifier pairs. To make optimisation\ntractable, we propose a dual-contrastive Infinite Class Mixup loss, where we\ncontrast the classifier of a mixed pair to both the classifiers and the\npredicted outputs of other mixed pairs in a batch. Infinite Class Mixup is\ngeneric in nature and applies to many variants of Mixup. Empirically, we show\nthat it outperforms standard Mixup and variants such as RegMixup and Remix on\nbalanced, long-tailed, and data-constrained benchmarks, highlighting its broad\napplicability.\n", "date": "2023-05-17", "categories": []}
{"id": "2302.14416", "title": "DREAM: Efficient Dataset Distillation by Representative Matching", "abstract": "  Dataset distillation aims to synthesize small datasets with little\ninformation loss from original large-scale ones for reducing storage and\ntraining costs. Recent state-of-the-art methods mainly constrain the sample\nsynthesis process by matching synthetic images and the original ones regarding\ngradients, embedding distributions, or training trajectories. Although there\nare various matching objectives, currently the strategy for selecting original\nimages is limited to naive random sampling.\n  We argue that random sampling overlooks the evenness of the selected sample\ndistribution, which may result in noisy or biased matching targets.\n  Besides, the sample diversity is also not constrained by random sampling.\nThese factors together lead to optimization instability in the distilling\nprocess and degrade the training efficiency. Accordingly, we propose a novel\nmatching strategy named as \\textbf{D}ataset distillation by\n\\textbf{RE}present\\textbf{A}tive \\textbf{M}atching (DREAM), where only\nrepresentative original images are selected for matching. DREAM is able to be\neasily plugged into popular dataset distillation frameworks and reduce the\ndistilling iterations by more than 8 times without performance drop. Given\nsufficient training time, DREAM further provides significant improvements and\nachieves state-of-the-art performances.\n", "date": "2023-02-28", "categories": []}
{"id": "2305.01975", "title": "A Survey on Dataset Distillation: Approaches, Applications and Future\n  Directions", "abstract": "  Dataset distillation is attracting more attention in machine learning as\ntraining sets continue to grow and the cost of training state-of-the-art models\nbecomes increasingly high. By synthesizing datasets with high information\ndensity, dataset distillation offers a range of potential applications,\nincluding support for continual learning, neural architecture search, and\nprivacy protection. Despite recent advances, we lack a holistic understanding\nof the approaches and applications. Our survey aims to bridge this gap by first\nproposing a taxonomy of dataset distillation, characterizing existing\napproaches, and then systematically reviewing the data modalities, and related\napplications. In addition, we summarize the challenges and discuss future\ndirections for this field of research.\n", "date": "2023-05-03", "categories": []}
{"id": "2308.12284", "title": "D4: Improving LLM Pretraining via Document De-Duplication and\n  Diversification", "abstract": "  Over recent years, an increasing amount of compute and data has been poured\ninto training large language models (LLMs), usually by doing one-pass learning\non as many tokens as possible randomly selected from large-scale web corpora.\nWhile training on ever-larger portions of the internet leads to consistent\nperformance improvements, the size of these improvements diminishes with scale,\nand there has been little work exploring the effect of data selection on\npre-training and downstream performance beyond simple de-duplication methods\nsuch as MinHash. Here, we show that careful data selection (on top of\nde-duplicated data) via pre-trained model embeddings can speed up training (20%\nefficiency gains) and improves average downstream accuracy on 16 NLP tasks (up\nto 2%) at the 6.7B model scale. Furthermore, we show that repeating data\nintelligently consistently outperforms baseline training (while repeating\nrandom data performs worse than baseline training). Our results indicate that\nclever data selection can significantly improve LLM pre-training, calls into\nquestion the common practice of training for a single epoch on as much data as\npossible, and demonstrates a path to keep improving our models past the limits\nof randomly sampling web data.\n", "date": "2023-08-23", "categories": []}
{"id": "2209.14609", "title": "Dataset Distillation Using Parameter Pruning", "abstract": "  In this study, we propose a novel dataset distillation method based on\nparameter pruning. The proposed method can synthesize more robust distilled\ndatasets and improve distillation performance by pruning difficult-to-match\nparameters during the distillation process. Experimental results on two\nbenchmark datasets show the superiority of the proposed method.\n", "date": "2022-09-29", "categories": []}
{"id": "2308.10524", "title": "Dataset Quantization", "abstract": "  State-of-the-art deep neural networks are trained with large amounts\n(millions or even billions) of data. The expensive computation and memory costs\nmake it difficult to train them on limited hardware resources, especially for\nrecent popular large language models (LLM) and computer vision models (CV).\nRecent popular dataset distillation methods are thus developed, aiming to\nreduce the number of training samples via synthesizing small-scale datasets via\ngradient matching. However, as the gradient calculation is coupled with the\nspecific network architecture, the synthesized dataset is biased and performs\npoorly when used for training unseen architectures. To address these\nlimitations, we present dataset quantization (DQ), a new framework to compress\nlarge-scale datasets into small subsets which can be used for training any\nneural network architectures. Extensive experiments demonstrate that DQ is able\nto generate condensed small datasets for training unseen network architectures\nwith state-of-the-art compression ratios for lossless model training. To the\nbest of our knowledge, DQ is the first method that can successfully distill\nlarge-scale datasets such as ImageNet-1k with a state-of-the-art compression\nratio. Notably, with 60% data from ImageNet and 20% data from Alpaca's\ninstruction tuning data, the models can be trained with negligible or no\nperformance drop for both vision tasks (including classification, semantic\nsegmentation, and object detection) as well as language tasks (including\ninstruction tuning tasks such as BBH and DROP).\n", "date": "2023-08-21", "categories": []}
{"id": "2308.08656", "title": "Flickr Africa: Examining Geo-Diversity in Large-Scale, Human-Centric\n  Visual Data", "abstract": "  Biases in large-scale image datasets are known to influence the performance\nof computer vision models as a function of geographic context. To investigate\nthe limitations of standard Internet data collection methods in low- and\nmiddle-income countries, we analyze human-centric image geo-diversity on a\nmassive scale using geotagged Flickr images associated with each nation in\nAfrica. We report the quantity and content of available data with comparisons\nto population-matched nations in Europe as well as the distribution of data\naccording to fine-grained intra-national wealth estimates. Temporal analyses\nare performed at two-year intervals to expose emerging data trends.\nFurthermore, we present findings for an ``othering'' phenomenon as evidenced by\na substantial number of images from Africa being taken by non-local\nphotographers. The results of our study suggest that further work is required\nto capture image data representative of African people and their environments\nand, ultimately, to improve the applicability of computer vision models in a\nglobal context.\n", "date": "2023-08-16", "categories": []}
{"id": "2305.00136", "title": "Optimizing the AI Development Process by Providing the Best Support\n  Environment", "abstract": "  The purpose of this study is to investigate the development process for\nArtificial inelegance (AI) and machine learning (ML) applications in order to\nprovide the best support environment. The main stages of ML are problem\nunderstanding, data management, model building, model deployment and\nmaintenance. This project focuses on investigating the data management stage of\nML development and its obstacles as it is the most important stage of machine\nlearning development because the accuracy of the end model is relying on the\nkind of data fed into the model. The biggest obstacle found on this stage was\nthe lack of sufficient data for model learning, especially in the fields where\ndata is confidential. This project aimed to build and develop a framework for\nresearchers and developers that can help solve the lack of sufficient data\nduring data management stage. The framework utilizes several data augmentation\ntechniques that can be used to generate new data from the original dataset\nwhich can improve the overall performance of the ML applications by increasing\nthe quantity and quality of available data to feed the model with the best\npossible data. The framework was built using python language to perform data\naugmentation using deep learning advancements.\n", "date": "2023-04-29", "categories": []}
{"id": "2308.02804", "title": "MiAMix: Enhancing Image Classification through a Multi-stage Augmented\n  Mixed Sample Data Augmentation Method", "abstract": "  Despite substantial progress in the field of deep learning, overfitting\npersists as a critical challenge, and data augmentation has emerged as a\nparticularly promising approach due to its capacity to enhance model\ngeneralization in various computer vision tasks. While various strategies have\nbeen proposed, Mixed Sample Data Augmentation (MSDA) has shown great potential\nfor enhancing model performance and generalization. We introduce a novel mixup\nmethod called MiAMix, which stands for Multi-stage Augmented Mixup. MiAMix\nintegrates image augmentation into the mixup framework, utilizes multiple\ndiversified mixing methods concurrently, and improves the mixing method by\nrandomly selecting mixing mask augmentation methods. Recent methods utilize\nsaliency information and the MiAMix is designed for computational efficiency as\nwell, reducing additional overhead and offering easy integration into existing\ntraining pipelines. We comprehensively evaluate MiaMix using four image\nbenchmarks and pitting it against current state-of-the-art mixed sample data\naugmentation techniques to demonstrate that MIAMix improves performance without\nheavy computational overhead.\n", "date": "2023-08-05", "categories": []}
{"id": "2308.06451", "title": "Semantic Equivariant Mixup", "abstract": "  Mixup is a well-established data augmentation technique, which can extend the\ntraining distribution and regularize the neural networks by creating ''mixed''\nsamples based on the label-equivariance assumption, i.e., a proportional mixup\nof the input data results in the corresponding labels being mixed in the same\nproportion. However, previous mixup variants may fail to exploit the\nlabel-independent information in mixed samples during training, which usually\ncontains richer semantic information. To further release the power of mixup, we\nfirst improve the previous label-equivariance assumption by the\nsemantic-equivariance assumption, which states that the proportional mixup of\nthe input data should lead to the corresponding representation being mixed in\nthe same proportion. Then a generic mixup regularization at the representation\nlevel is proposed, which can further regularize the model with the semantic\ninformation in mixed samples. At a high level, the proposed semantic\nequivariant mixup (sem) encourages the structure of the input data to be\npreserved in the representation space, i.e., the change of input will result in\nthe obtained representation information changing in the same way. Different\nfrom previous mixup variants, which tend to over-focus on the label-related\ninformation, the proposed method aims to preserve richer semantic information\nin the input with semantic-equivariance assumption, thereby improving the\nrobustness of the model against distribution shifts. We conduct extensive\nempirical studies and qualitative analyzes to demonstrate the effectiveness of\nour proposed method. The code of the manuscript is in the supplement.\n", "date": "2023-08-12", "categories": []}
{"id": "2212.04875", "title": "Expeditious Saliency-guided Mix-up through Random Gradient Thresholding", "abstract": "  Mix-up training approaches have proven to be effective in improving the\ngeneralization ability of Deep Neural Networks. Over the years, the research\ncommunity expands mix-up methods into two directions, with extensive efforts to\nimprove saliency-guided procedures but minimal focus on the arbitrary path,\nleaving the randomization domain unexplored. In this paper, inspired by the\nsuperior qualities of each direction over one another, we introduce a novel\nmethod that lies at the junction of the two routes. By combining the best\nelements of randomness and saliency utilization, our method balances speed,\nsimplicity, and accuracy. We name our method R-Mix following the concept of\n\"Random Mix-up\". We demonstrate its effectiveness in generalization, weakly\nsupervised object localization, calibration, and robustness to adversarial\nattacks. Finally, in order to address the question of whether there exists a\nbetter decision protocol, we train a Reinforcement Learning agent that decides\nthe mix-up policies based on the classifier's performance, reducing dependency\non human-designed objectives and hyperparameter tuning. Extensive experiments\nfurther show that the agent is capable of performing at the cutting-edge level,\nlaying the foundation for a fully automatic mix-up. Our code is released at\n[https://github.com/minhlong94/Random-Mixup].\n", "date": "2022-12-09", "categories": []}
{"id": "2308.04982", "title": "Exploring Multilingual Text Data Distillation", "abstract": "  With the rise of deep learning, large datasets and complex models have become\ncommon, requiring significant computing power. To address this, data\ndistillation has emerged as a technique to quickly train models with lower\nmemory and time requirements. However, data distillation on text-based datasets\nhasn't been explored much because of the challenges rising due to its discrete\nnature. Additionally, existing dataset distillation methods often struggle to\ngeneralize to new architectures. In the paper, we propose several data\ndistillation techniques for multilingual text classification datasets using\nlanguage-model-based learning methods. We conduct experiments to analyze their\nperformance in terms of classification strength, and cross-architecture\ngeneralization. Furthermore, we investigate the language-specific fairness of\nthe data summaries generated by these methods. Our approach builds upon\nexisting techniques, enhancing cross-architecture generalization in the text\ndata distillation domain.\n", "date": "2023-08-09", "categories": []}
{"id": "2212.01452", "title": "CLIP: Train Faster with Less Data", "abstract": "  Deep learning models require an enormous amount of data for training.\nHowever, recently there is a shift in machine learning from model-centric to\ndata-centric approaches. In data-centric approaches, the focus is to refine and\nimprove the quality of the data to improve the learning performance of the\nmodels rather than redesigning model architectures. In this paper, we propose\nCLIP i.e., Curriculum Learning with Iterative data Pruning. CLIP combines two\ndata-centric approaches i.e., curriculum learning and dataset pruning to\nimprove the model learning accuracy and convergence speed. The proposed scheme\napplies loss-aware dataset pruning to iteratively remove the least significant\nsamples and progressively reduces the size of the effective dataset in the\ncurriculum learning training. Extensive experiments performed on crowd density\nestimation models validate the notion behind combining the two approaches by\nreducing the convergence time and improving generalization. To our knowledge,\nthe idea of data pruning as an embedded process in curriculum learning is\nnovel.\n", "date": "2022-12-02", "categories": []}
{"id": "2307.14430", "title": "Skill-it! A Data-Driven Skills Framework for Understanding and Training\n  Language Models", "abstract": "  The quality of training data impacts the performance of pre-trained large\nlanguage models (LMs). Given a fixed budget of tokens, we study how to best\nselect data that leads to good downstream model performance across tasks. We\ndevelop a new framework based on a simple hypothesis: just as humans acquire\ninterdependent skills in a deliberate order, language models also follow a\nnatural order when learning a set of skills from their training data. If such\nan order exists, it can be utilized for improved understanding of LMs and for\ndata-efficient training. Using this intuition, our framework formalizes the\nnotion of a skill and of an ordered set of skills in terms of the associated\ndata. First, using both synthetic and real data, we demonstrate that these\nordered skill sets exist, and that their existence enables more advanced skills\nto be learned with less data when we train on their prerequisite skills.\nSecond, using our proposed framework, we introduce an online data sampling\nalgorithm, Skill-It, over mixtures of skills for both continual pre-training\nand fine-tuning regimes, where the objective is to efficiently learn multiple\nskills in the former and an individual skill in the latter. On the LEGO\nsynthetic in the continual pre-training setting, Skill-It obtains 36.5 points\nhigher accuracy than random sampling. On the Natural Instructions dataset in\nthe fine-tuning setting, Skill-It reduces the validation loss on the target\nskill by 13.6% versus training on data associated with the target skill itself.\nWe apply our skills framework on the recent RedPajama dataset to continually\npre-train a 3B-parameter LM, achieving higher accuracy on the LM Evaluation\nHarness with 1B tokens than the baseline approach of sampling uniformly over\ndata sources with 3B tokens.\n", "date": "2023-07-26", "categories": []}
{"id": "2306.10473", "title": "2D-Shapley: A Framework for Fragmented Data Valuation", "abstract": "  Data valuation -- quantifying the contribution of individual data sources to\ncertain predictive behaviors of a model -- is of great importance to enhancing\nthe transparency of machine learning and designing incentive systems for data\nsharing. Existing work has focused on evaluating data sources with the shared\nfeature or sample space. How to valuate fragmented data sources of which each\nonly contains partial features and samples remains an open question. We start\nby presenting a method to calculate the counterfactual of removing a fragment\nfrom the aggregated data matrix. Based on the counterfactual calculation, we\nfurther propose 2D-Shapley, a theoretical framework for fragmented data\nvaluation that uniquely satisfies some appealing axioms in the fragmented data\ncontext. 2D-Shapley empowers a range of new use cases, such as selecting useful\ndata fragments, providing interpretation for sample-wise data values, and\nfine-grained data issue diagnosis.\n", "date": "2023-06-18", "categories": []}
{"id": "2307.11823", "title": "HybridAugment++: Unified Frequency Spectra Perturbations for Model\n  Robustness", "abstract": "  Convolutional Neural Networks (CNN) are known to exhibit poor generalization\nperformance under distribution shifts. Their generalization have been studied\nextensively, and one line of work approaches the problem from a\nfrequency-centric perspective. These studies highlight the fact that humans and\nCNNs might focus on different frequency components of an image. First, inspired\nby these observations, we propose a simple yet effective data augmentation\nmethod HybridAugment that reduces the reliance of CNNs on high-frequency\ncomponents, and thus improves their robustness while keeping their clean\naccuracy high. Second, we propose HybridAugment++, which is a hierarchical\naugmentation method that attempts to unify various frequency-spectrum\naugmentations. HybridAugment++ builds on HybridAugment, and also reduces the\nreliance of CNNs on the amplitude component of images, and promotes phase\ninformation instead. This unification results in competitive to or better than\nstate-of-the-art results on clean accuracy (CIFAR-10/100 and ImageNet),\ncorruption benchmarks (ImageNet-C, CIFAR-10-C and CIFAR-100-C), adversarial\nrobustness on CIFAR-10 and out-of-distribution detection on various datasets.\nHybridAugment and HybridAugment++ are implemented in a few lines of code, does\nnot require extra data, ensemble models or additional networks.\n", "date": "2023-07-21", "categories": []}
{"id": "2307.11375", "title": "LatentAugment: Data Augmentation via Guided Manipulation of GAN's Latent\n  Space", "abstract": "  Data Augmentation (DA) is a technique to increase the quantity and diversity\nof the training data, and by that alleviate overfitting and improve\ngeneralisation. However, standard DA produces synthetic data for augmentation\nwith limited diversity. Generative Adversarial Networks (GANs) may unlock\nadditional information in a dataset by generating synthetic samples having the\nappearance of real images. However, these models struggle to simultaneously\naddress three key requirements: fidelity and high-quality samples; diversity\nand mode coverage; and fast sampling. Indeed, GANs generate high-quality\nsamples rapidly, but have poor mode coverage, limiting their adoption in DA\napplications. We propose LatentAugment, a DA strategy that overcomes the low\ndiversity of GANs, opening up for use in DA applications. Without external\nsupervision, LatentAugment modifies latent vectors and moves them into latent\nspace regions to maximise the synthetic images' diversity and fidelity. It is\nalso agnostic to the dataset and the downstream task. A wide set of experiments\nshows that LatentAugment improves the generalisation of a deep model\ntranslating from MRI-to-CT beating both standard DA as well GAN-based sampling.\nMoreover, still in comparison with GAN-based sampling, LatentAugment synthetic\nsamples show superior mode coverage and diversity. Code is available at:\nhttps://github.com/ltronchin/LatentAugment.\n", "date": "2023-07-21", "categories": []}
{"id": "2210.08363", "title": "Data-Efficient Augmentation for Training Neural Networks", "abstract": "  Data augmentation is essential to achieve state-of-the-art performance in\nmany deep learning applications. However, the most effective augmentation\ntechniques become computationally prohibitive for even medium-sized datasets.\nTo address this, we propose a rigorous technique to select subsets of data\npoints that when augmented, closely capture the training dynamics of full data\naugmentation. We first show that data augmentation, modeled as additive\nperturbations, improves learning and generalization by relatively enlarging and\nperturbing the smaller singular values of the network Jacobian, while\npreserving its prominent directions. This prevents overfitting and enhances\nlearning the harder to learn information. Then, we propose a framework to\niteratively extract small subsets of training data that when augmented, closely\ncapture the alignment of the fully augmented Jacobian with labels/residuals. We\nprove that stochastic gradient descent applied to the augmented subsets found\nby our approach has similar training dynamics to that of fully augmented data.\nOur experiments demonstrate that our method achieves 6.3x speedup on CIFAR10\nand 2.2x speedup on SVHN, and outperforms the baselines by up to 10% across\nvarious subset sizes. Similarly, on TinyImageNet and ImageNet, our method beats\nthe baselines by up to 8%, while achieving up to 3.3x speedup across various\nsubset sizes. Finally, training on and augmenting 50% subsets using our method\non a version of CIFAR10 corrupted with label noise even outperforms using the\nfull dataset. Our code is available at:\nhttps://github.com/tianyu139/data-efficient-augmentation\n", "date": "2022-10-15", "categories": []}
{"id": "2307.09742", "title": "Improved Distribution Matching for Dataset Condensation", "abstract": "  Dataset Condensation aims to condense a large dataset into a smaller one\nwhile maintaining its ability to train a well-performing model, thus reducing\nthe storage cost and training effort in deep learning applications. However,\nconventional dataset condensation methods are optimization-oriented and\ncondense the dataset by performing gradient or parameter matching during model\noptimization, which is computationally intensive even on small datasets and\nmodels. In this paper, we propose a novel dataset condensation method based on\ndistribution matching, which is more efficient and promising. Specifically, we\nidentify two important shortcomings of naive distribution matching (i.e.,\nimbalanced feature numbers and unvalidated embeddings for distance computation)\nand address them with three novel techniques (i.e., partitioning and expansion\naugmentation, efficient and enriched model sampling, and class-aware\ndistribution regularization). Our simple yet effective method outperforms most\nprevious optimization-oriented methods with much fewer computational resources,\nthereby scaling data condensation to larger datasets and models. Extensive\nexperiments demonstrate the effectiveness of our method. Codes are available at\nhttps://github.com/uitrbn/IDM\n", "date": "2023-07-19", "categories": []}
{"id": "2307.07119", "title": "DataAssist: A Machine Learning Approach to Data Cleaning and Preparation", "abstract": "  Current automated machine learning (ML) tools are model-centric, focusing on\nmodel selection and parameter optimization. However, the majority of the time\nin data analysis is devoted to data cleaning and wrangling, for which limited\ntools are available. Here we present DataAssist, an automated data preparation\nand cleaning platform that enhances dataset quality using ML-informed methods.\nWe show that DataAssist provides a pipeline for exploratory data analysis and\ndata cleaning, including generating visualization for user-selected variables,\nunifying data annotation, suggesting anomaly removal, and preprocessing data.\nThe exported dataset can be readily integrated with other autoML tools or\nuser-specified model for downstream analysis. Our data-centric tool is\napplicable to a variety of fields, including economics, business, and\nforecasting applications saving over 50% time of the time spent on data\ncleansing and preparation.\n", "date": "2023-07-14", "categories": []}
{"id": "2307.08086", "title": "Dataset Distillation Meets Provable Subset Selection", "abstract": "  Deep learning has grown tremendously over recent years, yielding\nstate-of-the-art results in various fields. However, training such models\nrequires huge amounts of data, increasing the computational time and cost. To\naddress this, dataset distillation was proposed to compress a large training\ndataset into a smaller synthetic one that retains its performance -- this is\nusually done by (1) uniformly initializing a synthetic set and (2) iteratively\nupdating/learning this set according to a predefined loss by uniformly sampling\ninstances from the full data. In this paper, we improve both phases of dataset\ndistillation: (1) we present a provable, sampling-based approach for\ninitializing the distilled set by identifying important and removing redundant\npoints in the data, and (2) we further merge the idea of data subset selection\nwith dataset distillation, by training the distilled set on ``important''\nsampled points during the training procedure instead of randomly sampling the\nnext batch. To do so, we define the notion of importance based on the relative\ncontribution of instances with respect to two different loss functions, i.e.,\none for the initialization phase (a kernel fitting function for kernel ridge\nregression and $K$-means based loss function for any other distillation\nmethod), and the relative cross-entropy loss (or any other predefined loss)\nfunction for the training phase. Finally, we provide experimental results\nshowing how our method can latch on to existing dataset distillation techniques\nand improve their performance.\n", "date": "2023-07-16", "categories": []}
{"id": "2307.02460", "title": "Performance Scaling via Optimal Transport: Enabling Data Selection from\n  Partially Revealed Sources", "abstract": "  Traditionally, data selection has been studied in settings where all samples\nfrom prospective sources are fully revealed to a machine learning developer.\nHowever, in practical data exchange scenarios, data providers often reveal only\na limited subset of samples before an acquisition decision is made. Recently,\nthere have been efforts to fit scaling laws that predict model performance at\nany size and data source composition using the limited available samples.\nHowever, these scaling functions are black-box, computationally expensive to\nfit, highly susceptible to overfitting, or/and difficult to optimize for data\nselection. This paper proposes a framework called <projektor>, which predicts\nmodel performance and supports data selection decisions based on partial\nsamples of prospective data sources. Our approach distinguishes itself from\nexisting work by introducing a novel *two-stage* performance inference process.\nIn the first stage, we leverage the Optimal Transport distance to predict the\nmodel's performance for any data mixture ratio within the range of disclosed\ndata sizes. In the second stage, we extrapolate the performance to larger\nundisclosed data sizes based on a novel parameter-free mapping technique\ninspired by neural scaling laws. We further derive an efficient gradient-based\nmethod to select data sources based on the projected model performance.\nEvaluation over a diverse range of applications demonstrates that <projektor>\nsignificantly improves existing performance scaling approaches in terms of both\nthe accuracy of performance inference and the computation costs associated with\nconstructing the performance predictor. Also, <projektor> outperforms by a wide\nmargin in data selection effectiveness compared to a range of other\noff-the-shelf solutions.\n", "date": "2023-07-05", "categories": []}
{"id": "2306.15538", "title": "DataCI: A Platform for Data-Centric AI on Streaming Data", "abstract": "  We introduce DataCI, a comprehensive open-source platform designed\nspecifically for data-centric AI in dynamic streaming data settings. DataCI\nprovides 1) an infrastructure with rich APIs for seamless streaming dataset\nmanagement, data-centric pipeline development and evaluation on streaming\nscenarios, 2) an carefully designed versioning control function to track the\npipeline lineage, and 3) an intuitive graphical interface for a better\ninteractive user experience. Preliminary studies and demonstrations attest to\nthe easy-to-use and effectiveness of DataCI, highlighting its potential to\nrevolutionize the practice of data-centric AI in streaming data contexts.\n", "date": "2023-06-27", "categories": []}
{"id": "2306.16612", "title": "GuidedMixup: An Efficient Mixup Strategy Guided by Saliency Maps", "abstract": "  Data augmentation is now an essential part of the image training process, as\nit effectively prevents overfitting and makes the model more robust against\nnoisy datasets. Recent mixing augmentation strategies have advanced to generate\nthe mixup mask that can enrich the saliency information, which is a supervisory\nsignal. However, these methods incur a significant computational burden to\noptimize the mixup mask. From this motivation, we propose a novel\nsaliency-aware mixup method, GuidedMixup, which aims to retain the salient\nregions in mixup images with low computational overhead. We develop an\nefficient pairing algorithm that pursues to minimize the conflict of salient\nregions of paired images and achieve rich saliency in mixup images. Moreover,\nGuidedMixup controls the mixup ratio for each pixel to better preserve the\nsalient region by interpolating two paired images smoothly. The experiments on\nseveral datasets demonstrate that GuidedMixup provides a good trade-off between\naugmentation overhead and generalization performance on classification\ndatasets. In addition, our method shows good performance in experiments with\ncorrupted or reduced datasets.\n", "date": "2023-06-29", "categories": []}
{"id": "2306.15392", "title": "Assessing Dataset Quality Through Decision Tree Characteristics in\n  Autoencoder-Processed Spaces", "abstract": "  In this paper, we delve into the critical aspect of dataset quality\nassessment in machine learning classification tasks. Leveraging a variety of\nnine distinct datasets, each crafted for classification tasks with varying\ncomplexity levels, we illustrate the profound impact of dataset quality on\nmodel training and performance. We further introduce two additional datasets\ndesigned to represent specific data conditions - one maximizing entropy and the\nother demonstrating high redundancy. Our findings underscore the importance of\nappropriate feature selection, adequate data volume, and data quality in\nachieving high-performing machine learning models. To aid researchers and\npractitioners, we propose a comprehensive framework for dataset quality\nassessment, which can help evaluate if the dataset at hand is sufficient and of\nthe required quality for specific tasks. This research offers valuable insights\ninto data assessment practices, contributing to the development of more\naccurate and robust machine learning models.\n", "date": "2023-06-27", "categories": []}
{"id": "2301.01228", "title": "DMOps: Data Management Operation and Recipes", "abstract": "  Data-centric AI has shed light on the significance of data within the machine\nlearning (ML) pipeline. Recognizing its significance, academia, industry, and\ngovernment departments have suggested various NLP data research initiatives.\nWhile the ability to utilize existing data is essential, the ability to build a\ndataset has become more critical than ever, especially in the industry. In\nconsideration of this trend, we propose a \"Data Management Operations and\nRecipes\" to guide the industry in optimizing the building of datasets for NLP\nproducts. This paper presents the concept of DMOps which is derived from\nreal-world experiences with NLP data management and aims to streamline data\noperations by offering a baseline.\n", "date": "2023-01-02", "categories": []}
{"id": "2306.14035", "title": "Thinking Like an Annotator: Generation of Dataset Labeling Instructions", "abstract": "  Large-scale datasets are essential to modern day deep learning. Advocates\nargue that understanding these methods requires dataset transparency (e.g.\n\"dataset curation, motivation, composition, collection process, etc...\").\nHowever, almost no one has suggested the release of the detailed definitions\nand visual category examples provided to annotators - information critical to\nunderstanding the structure of the annotations present in each dataset. These\nlabels are at the heart of public datasets, yet few datasets include the\ninstructions that were used to generate them. We introduce a new task, Labeling\nInstruction Generation, to address missing publicly available labeling\ninstructions. In Labeling Instruction Generation, we take a reasonably\nannotated dataset and: 1) generate a set of examples that are visually\nrepresentative of each category in the dataset; 2) provide a text label that\ncorresponds to each of the examples. We introduce a framework that requires no\nmodel training to solve this task and includes a newly created rapid retrieval\nsystem that leverages a large, pre-trained vision and language model. This\nframework acts as a proxy to human annotators that can help to both generate a\nfinal labeling instruction set and evaluate its quality. Our framework\ngenerates multiple diverse visual and text representations of dataset\ncategories. The optimized instruction set outperforms our strongest baseline\nacross 5 folds by 7.06 mAP for NuImages and 12.9 mAP for COCO.\n", "date": "2023-06-24", "categories": []}
{"id": "2306.14113", "title": "Exploring Data Redundancy in Real-world Image Classification through\n  Data Selection", "abstract": "  Deep learning models often require large amounts of data for training,\nleading to increased costs. It is particularly challenging in medical imaging,\ni.e., gathering distributed data for centralized training, and meanwhile,\nobtaining quality labels remains a tedious job. Many methods have been proposed\nto address this issue in various training paradigms, e.g., continual learning,\nactive learning, and federated learning, which indeed demonstrate certain forms\nof the data valuation process. However, existing methods are either overly\nintuitive or limited to common clean/toy datasets in the experiments. In this\nwork, we present two data valuation metrics based on Synaptic Intelligence and\ngradient norms, respectively, to study the redundancy in real-world image data.\nNovel online and offline data selection algorithms are then proposed via\nclustering and grouping based on the examined data values. Our online approach\neffectively evaluates data utilizing layerwise model parameter updates and\ngradients in each epoch and can accelerate model training with fewer epochs and\na subset (e.g., 19%-59%) of data while maintaining equivalent levels of\naccuracy in a variety of datasets. It also extends to the offline coreset\nconstruction, producing subsets of only 18%-30% of the original. The codes for\nthe proposed adaptive data selection and coreset computation are available\n(https://github.com/ZhenyuTANG2023/data_selection).\n", "date": "2023-06-25", "categories": []}
{"id": "2306.14377", "title": "Synthetic Alone: Exploring the Dark Side of Synthetic Data for\n  Grammatical Error Correction", "abstract": "  Data-centric AI approach aims to enhance the model performance without\nmodifying the model and has been shown to impact model performance positively.\nWhile recent attention has been given to data-centric AI based on synthetic\ndata, due to its potential for performance improvement, data-centric AI has\nlong been exclusively validated using real-world data and publicly available\nbenchmark datasets. In respect of this, data-centric AI still highly depends on\nreal-world data, and the verification of models using synthetic data has not\nyet been thoroughly carried out. Given the challenges above, we ask the\nquestion: Does data quality control (noise injection and balanced data), a\ndata-centric AI methodology acclaimed to have a positive impact, exhibit the\nsame positive impact in models trained solely with synthetic data? To address\nthis question, we conducted comparative analyses between models trained on\nsynthetic and real-world data based on grammatical error correction (GEC) task.\nOur experimental results reveal that the data quality control method has a\npositive impact on models trained with real-world data, as previously reported\nin existing studies, while a negative impact is observed in models trained\nsolely on synthetic data.\n", "date": "2023-06-26", "categories": []}
{"id": "2306.10165", "title": "Data Selection for Fine-tuning Large Language Models Using Transferred\n  Shapley Values", "abstract": "  Although Shapley values have been shown to be highly effective for\nidentifying harmful training instances, dataset size and model complexity\nconstraints limit the ability to apply Shapley-based data valuation to\nfine-tuning large pre-trained language models. To address this, we propose\nTS-DShapley, an algorithm that reduces computational cost of Shapley-based data\nvaluation through: 1) an efficient sampling-based method that aggregates\nShapley values computed from subsets for valuation of the entire training set,\nand 2) a value transfer method that leverages value information extracted from\na simple classifier trained using representations from the target language\nmodel. Our experiments applying TS-DShapley to select data for fine-tuning\nBERT-based language models on benchmark natural language understanding (NLU)\ndatasets show that TS-DShapley outperforms existing data selection methods.\nFurther, TS-DShapley can filter fine-tuning data to increase language model\nperformance compared to training with the full fine-tuning dataset.\n", "date": "2023-06-16", "categories": []}
{"id": "2303.10158", "title": "Data-centric Artificial Intelligence: A Survey", "abstract": "  Artificial Intelligence (AI) is making a profound impact in almost every\ndomain. A vital enabler of its great success is the availability of abundant\nand high-quality data for building machine learning models. Recently, the role\nof data in AI has been significantly magnified, giving rise to the emerging\nconcept of data-centric AI. The attention of researchers and practitioners has\ngradually shifted from advancing model design to enhancing the quality and\nquantity of the data. In this survey, we discuss the necessity of data-centric\nAI, followed by a holistic view of three general data-centric goals (training\ndata development, inference data development, and data maintenance) and the\nrepresentative methods. We also organize the existing literature from\nautomation and collaboration perspectives, discuss the challenges, and tabulate\nthe benchmarks for various tasks. We believe this is the first comprehensive\nsurvey that provides a global view of a spectrum of tasks across various stages\nof the data lifecycle. We hope it can help the readers efficiently grasp a\nbroad picture of this field, and equip them with the techniques and further\nresearch ideas to systematically engineer data for building AI systems. A\ncompanion list of data-centric AI resources will be regularly updated on\nhttps://github.com/daochenzha/data-centric-AI\n", "date": "2023-03-17", "categories": []}
{"id": "2305.19344", "title": "infoVerse: A Universal Framework for Dataset Characterization with\n  Multidimensional Meta-information", "abstract": "  The success of NLP systems often relies on the availability of large,\nhigh-quality datasets. However, not all samples in these datasets are equally\nvaluable for learning, as some may be redundant or noisy. Several methods for\ncharacterizing datasets based on model-driven meta-information (e.g., model's\nconfidence) have been developed, but the relationship and complementary effects\nof these methods have received less attention. In this paper, we introduce\ninfoVerse, a universal framework for dataset characterization, which provides a\nnew feature space that effectively captures multidimensional characteristics of\ndatasets by incorporating various model-driven meta-information. infoVerse\nreveals distinctive regions of the dataset that are not apparent in the\noriginal semantic space, hence guiding users (or models) in identifying which\nsamples to focus on for exploration, assessment, or annotation. Additionally,\nwe propose a novel sampling method on infoVerse to select a set of data points\nthat maximizes informativeness. In three real-world applications (data pruning,\nactive learning, and data annotation), the samples chosen on infoVerse space\nconsistently outperform strong baselines in all applications. Our code and demo\nare publicly available.\n", "date": "2023-05-30", "categories": []}
{"id": "2306.06866", "title": "Generating Synthetic Datasets by Interpolating along Generalized\n  Geodesics", "abstract": "  Data for pretraining machine learning models often consists of collections of\nheterogeneous datasets. Although training on their union is reasonable in\nagnostic settings, it might be suboptimal when the target domain -- where the\nmodel will ultimately be used -- is known in advance. In that case, one would\nideally pretrain only on the dataset(s) most similar to the target one. Instead\nof limiting this choice to those datasets already present in the pretraining\ncollection, here we explore extending this search to all datasets that can be\nsynthesized as `combinations' of them. We define such combinations as\nmulti-dataset interpolations, formalized through the notion of generalized\ngeodesics from optimal transport (OT) theory. We compute these geodesics using\na recent notion of distance between labeled datasets, and derive alternative\ninterpolation schemes based on it: using either barycentric projections or\noptimal transport maps, the latter computed using recent neural OT methods.\nThese methods are scalable, efficient, and -- notably -- can be used to\ninterpolate even between datasets with distinct and unrelated label sets.\nThrough various experiments in transfer learning in computer vision, we\ndemonstrate this is a promising new approach for targeted on-demand dataset\nsynthesis.\n", "date": "2023-06-12", "categories": []}
{"id": "2306.07171", "title": "Shapley Value on Probabilistic Classifiers", "abstract": "  Data valuation has become an increasingly significant discipline in data\nscience due to the economic value of data. In the context of machine learning\n(ML), data valuation methods aim to equitably measure the contribution of each\ndata point to the utility of an ML model. One prevalent method is Shapley\nvalue, which helps identify data points that are beneficial or detrimental to\nan ML model. However, traditional Shapley-based data valuation methods may not\neffectively distinguish between beneficial and detrimental training data points\nfor probabilistic classifiers. In this paper, we propose Probabilistic Shapley\n(P-Shapley) value by constructing a probability-wise utility function that\nleverages the predicted class probabilities of probabilistic classifiers rather\nthan binarized prediction results in the traditional Shapley value. We also\noffer several activation functions for confidence calibration to effectively\nquantify the marginal contribution of each data point to the probabilistic\nclassifiers. Extensive experiments on four real-world datasets demonstrate the\neffectiveness of our proposed P-Shapley value in evaluating the importance of\ndata for building a high-usability and trustworthy ML model.\n", "date": "2023-06-12", "categories": []}
{"id": "2306.04349", "title": "GPT Self-Supervision for a Better Data Annotator", "abstract": "  The task of annotating data into concise summaries poses a significant\nchallenge across various domains, frequently requiring the allocation of\nsignificant time and specialized knowledge by human experts. Despite existing\nefforts to use large language models for annotation tasks, significant problems\nsuch as limited applicability to unlabeled data, the absence of self-supervised\nmethods, and the lack of focus on complex structured data still persist. In\nthis work, we propose a GPT self-supervision annotation method, which embodies\na generating-recovering paradigm that leverages the one-shot learning\ncapabilities of the Generative Pretrained Transformer (GPT). The proposed\napproach comprises a one-shot tuning phase followed by a generation phase. In\nthe one-shot tuning phase, we sample a data from the support set as part of the\nprompt for GPT to generate a textual summary, which is then used to recover the\noriginal data. The alignment score between the recovered and original data\nserves as a self-supervision navigator to refine the process. In the generation\nstage, the optimally selected one-shot sample serves as a template in the\nprompt and is applied to generating summaries from challenging datasets. The\nannotation performance is evaluated by tuning several human feedback reward\nnetworks and by calculating alignment scores between original and recovered\ndata at both sentence and structure levels. Our self-supervised annotation\nmethod consistently achieves competitive scores, convincingly demonstrating its\nrobust strength in various data-to-summary annotation tasks.\n", "date": "2023-06-07", "categories": []}
{"id": "2306.04240", "title": "T-ADAF: Adaptive Data Augmentation Framework for Image Classification\n  Network based on Tensor T-product Operator", "abstract": "  Image classification is one of the most fundamental tasks in Computer Vision.\nIn practical applications, the datasets are usually not as abundant as those in\nthe laboratory and simulation, which is always called as Data Hungry. How to\nextract the information of data more completely and effectively is very\nimportant. Therefore, an Adaptive Data Augmentation Framework based on the\ntensor T-product Operator is proposed in this paper, to triple one image data\nto be trained and gain the result from all these three images together with\nonly less than 0.1% increase in the number of parameters. At the same time,\nthis framework serves the functions of column image embedding and global\nfeature intersection, enabling the model to obtain information in not only\nspatial but frequency domain, and thus improving the prediction accuracy of the\nmodel. Numerical experiments have been designed for several models, and the\nresults demonstrate the effectiveness of this adaptive framework. Numerical\nexperiments show that our data augmentation framework can improve the\nperformance of original neural network model by 2%, which provides competitive\nresults to state-of-the-art methods.\n", "date": "2023-06-07", "categories": []}
{"id": "2306.02543", "title": "Addressing Budget Allocation and Revenue Allocation in Data Market\n  Environments Using an Adaptive Sampling Algorithm", "abstract": "  High-quality machine learning models are dependent on access to high-quality\ntraining data. When the data are not already available, it is tedious and\ncostly to obtain them. Data markets help with identifying valuable training\ndata: model consumers pay to train a model, the market uses that budget to\nidentify data and train the model (the budget allocation problem), and finally\nthe market compensates data providers according to their data contribution\n(revenue allocation problem). For example, a bank could pay the data market to\naccess data from other financial institutions to train a fraud detection model.\nCompensating data contributors requires understanding data's contribution to\nthe model; recent efforts to solve this revenue allocation problem based on the\nShapley value are inefficient to lead to practical data markets.\n  In this paper, we introduce a new algorithm to solve budget allocation and\nrevenue allocation problems simultaneously in linear time. The new algorithm\nemploys an adaptive sampling process that selects data from those providers who\nare contributing the most to the model. Better data means that the algorithm\naccesses those providers more often, and more frequent accesses corresponds to\nhigher compensation. Furthermore, the algorithm can be deployed in both\ncentralized and federated scenarios, boosting its applicability. We provide\ntheoretical guarantees for the algorithm that show the budget is used\nefficiently and the properties of revenue allocation are similar to Shapley's.\nFinally, we conduct an empirical evaluation to show the performance of the\nalgorithm in practical scenarios and when compared to other baselines. Overall,\nwe believe that the new algorithm paves the way for the implementation of\npractical data markets.\n", "date": "2023-06-05", "categories": []}
{"id": "2306.01244", "title": "Towards Sustainable Learning: Coresets for Data-efficient Deep Learning", "abstract": "  To improve the efficiency and sustainability of learning deep models, we\npropose CREST, the first scalable framework with rigorous theoretical\nguarantees to identify the most valuable examples for training non-convex\nmodels, particularly deep networks. To guarantee convergence to a stationary\npoint of a non-convex function, CREST models the non-convex loss as a series of\nquadratic functions and extracts a coreset for each quadratic sub-region. In\naddition, to ensure faster convergence of stochastic gradient methods such as\n(mini-batch) SGD, CREST iteratively extracts multiple mini-batch coresets from\nlarger random subsets of training data, to ensure nearly-unbiased gradients\nwith small variances. Finally, to further improve scalability and efficiency,\nCREST identifies and excludes the examples that are learned from the coreset\nselection pipeline. Our extensive experiments on several deep networks trained\non vision and NLP datasets, including CIFAR-10, CIFAR-100, TinyImageNet, and\nSNLI, confirm that CREST speeds up training deep networks on very large\ndatasets, by 1.7x to 2.5x with minimum loss in the performance. By analyzing\nthe learning difficulty of the subsets selected by CREST, we show that deep\nmodels benefit the most by learning from subsets of increasing difficulty\nlevels.\n", "date": "2023-06-02", "categories": []}
{"id": "2304.07718", "title": "Data-OOB: Out-of-bag Estimate as a Simple and Efficient Data Value", "abstract": "  Data valuation is a powerful framework for providing statistical insights\ninto which data are beneficial or detrimental to model training. Many\nShapley-based data valuation methods have shown promising results in various\ndownstream tasks, however, they are well known to be computationally\nchallenging as it requires training a large number of models. As a result, it\nhas been recognized as infeasible to apply to large datasets. To address this\nissue, we propose Data-OOB, a new data valuation method for a bagging model\nthat utilizes the out-of-bag estimate. The proposed method is computationally\nefficient and can scale to millions of data by reusing trained weak learners.\nSpecifically, Data-OOB takes less than 2.25 hours on a single CPU processor\nwhen there are $10^6$ samples to evaluate and the input dimension is 100.\nFurthermore, Data-OOB has solid theoretical interpretations in that it\nidentifies the same important data point as the infinitesimal jackknife\ninfluence function when two different points are compared. We conduct\ncomprehensive experiments using 12 classification datasets, each with thousands\nof sample sizes. We demonstrate that the proposed method significantly\noutperforms existing state-of-the-art data valuation methods in identifying\nmislabeled data and finding a set of helpful (or harmful) data points,\nhighlighting the potential for applying data values in real-world applications.\n", "date": "2023-04-16", "categories": []}
{"id": "2305.19617", "title": "MSMix:An Interpolation-Based Text Data Augmentation Method Manifold Swap\n  Mixup", "abstract": "  To solve the problem of poor performance of deep neural network models due to\ninsufficient data, a simple yet effective interpolation-based data augmentation\nmethod is proposed: MSMix (Manifold Swap Mixup). This method feeds two\ndifferent samples to the same deep neural network model, and then randomly\nselect a specific layer and partially replace hidden features at that layer of\none of the samples by the counterpart of the other. The mixed hidden features\nare fed to the model and go through the rest of the network. Two different\nselection strategies are also proposed to obtain richer hidden representation.\nExperiments are conducted on three Chinese intention recognition datasets, and\nthe results show that the MSMix method achieves better results than other\nmethods in both full-sample and small-sample configurations.\n", "date": "2023-05-31", "categories": []}
{"id": "2210.06828", "title": "Rethinking Annotation: Can Language Learners Contribute?", "abstract": "  Researchers have traditionally recruited native speakers to provide\nannotations for widely used benchmark datasets. However, there are languages\nfor which recruiting native speakers can be difficult, and it would help to\nfind learners of those languages to annotate the data. In this paper, we\ninvestigate whether language learners can contribute annotations to benchmark\ndatasets. In a carefully controlled annotation experiment, we recruit 36\nlanguage learners, provide two types of additional resources (dictionaries and\nmachine-translated sentences), and perform mini-tests to measure their language\nproficiency. We target three languages, English, Korean, and Indonesian, and\nthe four NLP tasks of sentiment analysis, natural language inference, named\nentity recognition, and machine reading comprehension. We find that language\nlearners, especially those with intermediate or advanced levels of language\nproficiency, are able to provide fairly accurate labels with the help of\nadditional resources. Moreover, we show that data annotation improves learners'\nlanguage proficiency in terms of vocabulary and grammar. One implication of our\nfindings is that broadening the annotation task to include language learners\ncan open up the opportunity to build benchmark datasets for languages for which\nit is difficult to recruit native speakers.\n", "date": "2022-10-13", "categories": []}
{"id": "2305.03355", "title": "A Comprehensive Study on Dataset Distillation: Performance, Privacy,\n  Robustness and Fairness", "abstract": "  The aim of dataset distillation is to encode the rich features of an original\ndataset into a tiny dataset. It is a promising approach to accelerate neural\nnetwork training and related studies. Different approaches have been proposed\nto improve the informativeness and generalization performance of distilled\nimages. However, no work has comprehensively analyzed this technique from a\nsecurity perspective and there is a lack of systematic understanding of\npotential risks. In this work, we conduct extensive experiments to evaluate\ncurrent state-of-the-art dataset distillation methods. We successfully use\nmembership inference attacks to show that privacy risks still remain. Our work\nalso demonstrates that dataset distillation can cause varying degrees of impact\non model robustness and amplify model unfairness across classes when making\npredictions. This work offers a large-scale benchmarking framework for dataset\ndistillation evaluation.\n", "date": "2023-05-05", "categories": []}
{"id": "2305.17476", "title": "Toward Understanding Generative Data Augmentation", "abstract": "  Generative data augmentation, which scales datasets by obtaining fake labeled\nexamples from a trained conditional generative model, boosts classification\nperformance in various learning tasks including (semi-)supervised learning,\nfew-shot learning, and adversarially robust learning. However, little work has\ntheoretically investigated the effect of generative data augmentation. To fill\nthis gap, we establish a general stability bound in this not independently and\nidentically distributed (non-i.i.d.) setting, where the learned distribution is\ndependent on the original train set and generally not the same as the true\ndistribution. Our theoretical result includes the divergence between the\nlearned distribution and the true distribution. It shows that generative data\naugmentation can enjoy a faster learning rate when the order of divergence term\nis $o(\\max\\left( \\log(m)\\beta_m, 1 / \\sqrt{m})\\right)$, where $m$ is the train\nset size and $\\beta_m$ is the corresponding stability constant. We further\nspecify the learning setup to the Gaussian mixture model and generative\nadversarial nets. We prove that in both cases, though generative data\naugmentation does not enjoy a faster learning rate, it can improve the learning\nguarantees at a constant level when the train set is small, which is\nsignificant when the awful overfitting occurs. Simulation results on the\nGaussian mixture model and empirical results on generative adversarial nets\nsupport our theoretical conclusions. Our code is available at\nhttps://github.com/ML-GSAI/Understanding-GDA.\n", "date": "2023-05-27", "categories": []}
{"id": "2303.00716", "title": "Aligning benchmark datasets for table structure recognition", "abstract": "  Benchmark datasets for table structure recognition (TSR) must be carefully\nprocessed to ensure they are annotated consistently. However, even if a\ndataset's annotations are self-consistent, there may be significant\ninconsistency across datasets, which can harm the performance of models trained\nand evaluated on them. In this work, we show that aligning these\nbenchmarks$\\unicode{x2014}$removing both errors and inconsistency between\nthem$\\unicode{x2014}$improves model performance significantly. We demonstrate\nthis through a data-centric approach where we adopt one model architecture, the\nTable Transformer (TATR), that we hold fixed throughout. Baseline exact match\naccuracy for TATR evaluated on the ICDAR-2013 benchmark is 65% when trained on\nPubTables-1M, 42% when trained on FinTabNet, and 69% combined. After reducing\nannotation mistakes and inter-dataset inconsistency, performance of TATR\nevaluated on ICDAR-2013 increases substantially to 75% when trained on\nPubTables-1M, 65% when trained on FinTabNet, and 81% combined. We show through\nablations over the modification steps that canonicalization of the table\nannotations has a significantly positive effect on performance, while other\nchoices balance necessary trade-offs that arise when deciding a benchmark\ndataset's final composition. Overall we believe our work has significant\nimplications for benchmark design for TSR and potentially other tasks as well.\nDataset processing and training code will be released at\nhttps://github.com/microsoft/table-transformer.\n", "date": "2023-03-01", "categories": []}
{"id": "2305.13520", "title": "Tied-Augment: Controlling Representation Similarity Improves Data\n  Augmentation", "abstract": "  Data augmentation methods have played an important role in the recent advance\nof deep learning models, and have become an indispensable component of\nstate-of-the-art models in semi-supervised, self-supervised, and supervised\ntraining for vision. Despite incurring no additional latency at test time, data\naugmentation often requires more epochs of training to be effective. For\nexample, even the simple flips-and-crops augmentation requires training for\nmore than 5 epochs to improve performance, whereas RandAugment requires more\nthan 90 epochs. We propose a general framework called Tied-Augment, which\nimproves the efficacy of data augmentation in a wide range of applications by\nadding a simple term to the loss that can control the similarity of\nrepresentations under distortions. Tied-Augment can improve state-of-the-art\nmethods from data augmentation (e.g. RandAugment, mixup), optimization (e.g.\nSAM), and semi-supervised learning (e.g. FixMatch). For example,\nTied-RandAugment can outperform RandAugment by 2.0% on ImageNet. Notably, using\nTied-Augment, data augmentation can be made to improve generalization even when\ntraining for a few epochs and when fine-tuning. We open source our code at\nhttps://github.com/ekurtulus/tied-augment/tree/main.\n", "date": "2023-05-22", "categories": []}
{"id": "2305.14113", "title": "On the Size and Approximation Error of Distilled Sets", "abstract": "  Dataset Distillation is the task of synthesizing small datasets from large\nones while still retaining comparable predictive accuracy to the original\nuncompressed dataset. Despite significant empirical progress in recent years,\nthere is little understanding of the theoretical limitations/guarantees of\ndataset distillation, specifically, what excess risk is achieved by\ndistillation compared to the original dataset, and how large are distilled\ndatasets? In this work, we take a theoretical view on kernel ridge regression\n(KRR) based methods of dataset distillation such as Kernel Inducing Points. By\ntransforming ridge regression in random Fourier features (RFF) space, we\nprovide the first proof of the existence of small (size) distilled datasets and\ntheir corresponding excess risk for shift-invariant kernels. We prove that a\nsmall set of instances exists in the original input space such that its\nsolution in the RFF space coincides with the solution of the original data. We\nfurther show that a KRR solution can be generated using this distilled set of\ninstances which gives an approximation towards the KRR solution optimized on\nthe full input data. The size of this set is linear in the dimension of the RFF\nspace of the input set or alternatively near linear in the number of effective\ndegrees of freedom, which is a function of the kernel, number of datapoints,\nand the regularization parameter $\\lambda$. The error bound of this distilled\nset is also a function of $\\lambda$. We verify our bounds analytically and\nempirically.\n", "date": "2023-05-23", "categories": []}
{"id": "2305.14115", "title": "RLBoost: Boosting Supervised Models using Deep Reinforcement Learning", "abstract": "  Data quality or data evaluation is sometimes a task as important as\ncollecting a large volume of data when it comes to generating accurate\nartificial intelligence models. In fact, being able to evaluate the data can\nlead to a larger database that is better suited to a particular problem because\nwe have the ability to filter out data obtained automatically of dubious\nquality. In this paper we present RLBoost, an algorithm that uses deep\nreinforcement learning strategies to evaluate a particular dataset and obtain a\nmodel capable of estimating the quality of any new data in order to improve the\nfinal predictive quality of a supervised learning model. This solution has the\nadvantage that of being agnostic regarding the supervised model used and,\nthrough multi-attention strategies, takes into account the data in its context\nand not only individually. The results of the article show that this model\nobtains better and more stable results than other state-of-the-art algorithms\nsuch as LOO, DataShapley or DVRL.\n", "date": "2023-05-23", "categories": []}
{"id": "2210.16835", "title": "Variance reduced Shapley value estimation for trustworthy data valuation", "abstract": "  Data valuation, especially quantifying data value in algorithmic prediction\nand decision-making, is a fundamental problem in data trading scenarios. The\nmost widely used method is to define the data Shapley and approximate it by\nmeans of the permutation sampling algorithm. To make up for the large\nestimation variance of the permutation sampling that hinders the development of\nthe data marketplace, we propose a more robust data valuation method using\nstratified sampling, named variance reduced data Shapley (VRDS for short). We\ntheoretically show how to stratify, how many samples are taken at each stratum,\nand the sample complexity analysis of VRDS. Finally, the effectiveness of VRDS\nis illustrated in different types of datasets and data removal applications.\n", "date": "2022-10-30", "categories": []}
{"id": "2305.12811", "title": "Label Smarter, Not Harder: CleverLabel for Faster Annotation of\n  Ambiguous Image Classification with Higher Quality", "abstract": "  High-quality data is crucial for the success of machine learning, but\nlabeling large datasets is often a time-consuming and costly process. While\nsemi-supervised learning can help mitigate the need for labeled data, label\nquality remains an open issue due to ambiguity and disagreement among\nannotators. Thus, we use proposal-guided annotations as one option which leads\nto more consistency between annotators. However, proposing a label increases\nthe probability of the annotators deciding in favor of this specific label.\nThis introduces a bias which we can simulate and remove. We propose a new\nmethod CleverLabel for Cost-effective LabEling using Validated proposal-guidEd\nannotations and Repaired LABELs. CleverLabel can reduce labeling costs by up to\n30.0%, while achieving a relative improvement in Kullback-Leibler divergence of\nup to 29.8% compared to the previous state-of-the-art on a multi-domain\nreal-world image classification benchmark. CleverLabel offers a novel solution\nto the challenge of efficiently labeling large datasets while also improving\nthe label quality.\n", "date": "2023-05-22", "categories": []}
{"id": "2207.06658", "title": "Universal Adaptive Data Augmentation", "abstract": "  Existing automatic data augmentation (DA) methods either ignore updating DA's\nparameters according to the target model's state during training or adopt\nupdate strategies that are not effective enough. In this work, we design a\nnovel data augmentation strategy called \"Universal Adaptive Data Augmentation\"\n(UADA). Different from existing methods, UADA would adaptively update DA's\nparameters according to the target model's gradient information during\ntraining: given a pre-defined set of DA operations, we randomly decide types\nand magnitudes of DA operations for every data batch during training, and\nadaptively update DA's parameters along the gradient direction of the loss\nconcerning DA's parameters. In this way, UADA can increase the training loss of\nthe target networks, and the target networks would learn features from harder\nsamples to improve the generalization. Moreover, UADA is very general and can\nbe utilized in numerous tasks, e.g., image classification, semantic\nsegmentation and object detection. Extensive experiments with various models\nare conducted on CIFAR-10, CIFAR-100, ImageNet, tiny-ImageNet, Cityscapes, and\nVOC07+12 to prove the significant performance improvements brought by UADA.\n", "date": "2022-07-14", "categories": []}
{"id": "2305.01649", "title": "Generalizing Dataset Distillation via Deep Generative Prior", "abstract": "  Dataset Distillation aims to distill an entire dataset's knowledge into a few\nsynthetic images. The idea is to synthesize a small number of synthetic data\npoints that, when given to a learning algorithm as training data, result in a\nmodel approximating one trained on the original data. Despite recent progress\nin the field, existing dataset distillation methods fail to generalize to new\narchitectures and scale to high-resolution datasets. To overcome the above\nissues, we propose to use the learned prior from pre-trained deep generative\nmodels to synthesize the distilled data. To achieve this, we present a new\noptimization algorithm that distills a large number of images into a few\nintermediate feature vectors in the generative model's latent space. Our method\naugments existing techniques, significantly improving cross-architecture\ngeneralization in all settings.\n", "date": "2023-05-02", "categories": []}
{"id": "2305.02668", "title": "LatentAugment: Dynamically Optimized Latent Probabilities of Data\n  Augmentation", "abstract": "  Although data augmentation is a powerful technique for improving the\nperformance of image classification tasks, it is difficult to identify the best\naugmentation policy. The optimal augmentation policy, which is the latent\nvariable, cannot be directly observed. To address this problem, this study\nproposes $\\textit{LatentAugment}$, which estimates the latent probability of\noptimal augmentation. The proposed method is appealing in that it can\ndynamically optimize the augmentation strategies for each input and model\nparameter in learning iterations. Theoretical analysis shows that LatentAugment\nis a general model that includes other augmentation methods as special cases,\nand it is simple and computationally efficient in comparison with existing\naugmentation methods. Experimental results show that the proposed LatentAugment\nhas higher test accuracy than previous augmentation methods on the CIFAR-10,\nCIFAR-100, SVHN, and ImageNet datasets.\n", "date": "2023-05-04", "categories": []}
{"id": "2305.01660", "title": "Data valuation: The partial ordinal Shapley value for machine learning", "abstract": "  Data valuation using Shapley value has emerged as a prevalent research domain\nin machine learning applications. However, it is a challenge to address the\nrole of order in data cooperation as most research lacks such discussion. To\ntackle this problem, this paper studies the definition of the partial ordinal\nShapley value by group theory in abstract algebra. Besides, since the\ncalculation of the partial ordinal Shapley value requires exponential time,\nthis paper also gives three algorithms for approximating the results. The\nTruncated Monte Carlo algorithm is derived from the classic Shapley value\napproximation algorithm. The Classification Monte Carlo algorithm and the\nClassification Truncated Monte Carlo algorithm are based on the fact that the\ndata points in the same class provide similar information, then we can\naccelerate the calculation by leaving out some data points in each class.\n", "date": "2023-05-02", "categories": []}
{"id": "2304.04554", "title": "Use the Detection Transformer as a Data Augmenter", "abstract": "  Detection Transformer (DETR) is a Transformer architecture based object\ndetection model. In this paper, we demonstrate that it can also be used as a\ndata augmenter. We term our approach as DETR assisted CutMix, or DeMix for\nshort. DeMix builds on CutMix, a simple yet highly effective data augmentation\ntechnique that has gained popularity in recent years. CutMix improves model\nperformance by cutting and pasting a patch from one image onto another,\nyielding a new image. The corresponding label for this new example is specified\nas the weighted average of the original labels, where the weight is\nproportional to the area of the patch. CutMix selects a random patch to be cut.\nIn contrast, DeMix elaborately selects a semantically rich patch, located by a\npre-trained DETR. The label of the new image is specified in the same way as in\nCutMix. Experimental results on benchmark datasets for image classification\ndemonstrate that DeMix significantly outperforms prior art data augmentation\nmethods including CutMix. Oue code is available at\nhttps://github.com/ZJLAB-AMMI/DeMix.\n", "date": "2023-04-10", "categories": []}
{"id": "2304.13636", "title": "AutoCure: Automated Tabular Data Curation Technique for ML Pipelines", "abstract": "  Machine learning algorithms have become increasingly prevalent in multiple\ndomains, such as autonomous driving, healthcare, and finance. In such domains,\ndata preparation remains a significant challenge in developing accurate models,\nrequiring significant expertise and time investment to search the huge search\nspace of well-suited data curation and transformation tools. To address this\nchallenge, we present AutoCure, a novel and configuration-free data curation\npipeline that improves the quality of tabular data. Unlike traditional data\ncuration methods, AutoCure synthetically enhances the density of the clean data\nfraction through an adaptive ensemble-based error detection method and a data\naugmentation module. In practice, AutoCure can be integrated with open source\ntools, e.g., Auto-sklearn, H2O, and TPOT, to promote the democratization of\nmachine learning. As a proof of concept, we provide a comparative evaluation of\nAutoCure against 28 combinations of traditional data curation tools,\ndemonstrating superior performance and predictive accuracy without user\nintervention. Our evaluation shows that AutoCure is an effective approach to\nautomating data preparation and improving the accuracy of machine learning\nmodels.\n", "date": "2023-04-26", "categories": []}
{"id": "2304.10036", "title": "Visual DNA: Representing and Comparing Images using Distributions of\n  Neuron Activations", "abstract": "  Selecting appropriate datasets is critical in modern computer vision.\nHowever, no general-purpose tools exist to evaluate the extent to which two\ndatasets differ. For this, we propose representing images - and by extension\ndatasets - using Distributions of Neuron Activations (DNAs). DNAs fit\ndistributions, such as histograms or Gaussians, to activations of neurons in a\npre-trained feature extractor through which we pass the image(s) to represent.\nThis extractor is frozen for all datasets, and we rely on its generally\nexpressive power in feature space. By comparing two DNAs, we can evaluate the\nextent to which two datasets differ with granular control over the comparison\nattributes of interest, providing the ability to customise the way distances\nare measured to suit the requirements of the task at hand. Furthermore, DNAs\nare compact, representing datasets of any size with less than 15 megabytes. We\ndemonstrate the value of DNAs by evaluating their applicability on several\ntasks, including conditional dataset comparison, synthetic image evaluation,\nand transfer learning, and across diverse datasets, ranging from synthetic cat\nimages to celebrity faces and urban driving scenes.\n", "date": "2023-04-20", "categories": []}
{"id": "2304.08989", "title": "Incremental Image Labeling via Iterative Refinement", "abstract": "  Data quality is critical for multimedia tasks, while various types of\nsystematic flaws are found in image benchmark datasets, as discussed in recent\nwork. In particular, the existence of the semantic gap problem leads to a\nmany-to-many mapping between the information extracted from an image and its\nlinguistic description. This unavoidable bias further leads to poor performance\non current computer vision tasks. To address this issue, we introduce a\nKnowledge Representation (KR)-based methodology to provide guidelines driving\nthe labeling process, thereby indirectly introducing intended semantics in ML\nmodels. Specifically, an iterative refinement-based annotation method is\nproposed to optimize data labeling by organizing objects in a classification\nhierarchy according to their visual properties, ensuring that they are aligned\nwith their linguistic descriptions. Preliminary results verify the\neffectiveness of the proposed method.\n", "date": "2023-04-18", "categories": []}
{"id": "2212.06152", "title": "Accelerating Dataset Distillation via Model Augmentation", "abstract": "  Dataset Distillation (DD), a newly emerging field, aims at generating much\nsmaller but efficient synthetic training datasets from large ones. Existing DD\nmethods based on gradient matching achieve leading performance; however, they\nare extremely computationally intensive as they require continuously optimizing\na dataset among thousands of randomly initialized models. In this paper, we\nassume that training the synthetic data with diverse models leads to better\ngeneralization performance. Thus we propose two model augmentation techniques,\ni.e. using early-stage models and parameter perturbation to learn an\ninformative synthetic set with significantly reduced training cost. Extensive\nexperiments demonstrate that our method achieves up to 20x speedup and\ncomparable performance on par with state-of-the-art methods.\n", "date": "2022-12-12", "categories": []}
{"id": "2304.08442", "title": "The MiniPile Challenge for Data-Efficient Language Models", "abstract": "  The ever-growing diversity of pre-training text corpora has equipped language\nmodels with generalization capabilities across various downstream tasks.\nHowever, such diverse datasets are often too large for academic budgets; hence,\nmost research on Transformer architectures, training procedures, optimizers,\netc. gets conducted on smaller, homogeneous datasets. To this end, we present\nThe MiniPile Challenge, where one pre-trains a language model on a diverse text\ncorpus containing at most 1M documents. MiniPile is a 6GB subset of the\ndeduplicated 825GB The Pile corpus. To curate MiniPile, we perform a simple,\nthree-step data filtering process: we (1) infer embeddings for all documents of\nthe Pile, (2) cluster the embedding space using $k$-means, and (3) filter out\nlow-quality clusters. To verify MiniPile's suitability for language model\npre-training, we use it to pre-train a BERT and T5 model, yielding a\nperformance drop of only $1.9\\%$/$2.5\\%$ on the GLUE and SNI benchmarks\ncompared to the original pre-trained checkpoints trained on $2.6$x/$745$x the\namount of data. MiniPile is available at\nhttps://huggingface.co/datasets/JeanKaddour/minipile.\n", "date": "2023-04-17", "categories": []}
{"id": "2304.01224", "title": "Optimizing Data Shapley Interaction Calculation from O(2^n) to O(t n^2)\n  for KNN models", "abstract": "  With the rapid growth of data availability and usage, quantifying the added\nvalue of each training data point has become a crucial process in the field of\nartificial intelligence. The Shapley values have been recognized as an\neffective method for data valuation, enabling efficient training set\nsummarization, acquisition, and outlier removal. In this paper, we introduce\n\"STI-KNN\", an innovative algorithm that calculates the exact pair-interaction\nShapley values for KNN models in O(t n^2) time, which is a significant\nimprovement over the O(2^n)$ time complexity of baseline methods. By using\nSTI-KNN, we can efficiently and accurately evaluate the value of individual\ndata points, leading to improved training outcomes and ultimately enhancing the\neffectiveness of artificial intelligence applications.\n", "date": "2023-04-02", "categories": []}
{"id": "2301.04819", "title": "Data-centric AI: Perspectives and Challenges", "abstract": "  The role of data in building AI systems has recently been significantly\nmagnified by the emerging concept of data-centric AI (DCAI), which advocates a\nfundamental shift from model advancements to ensuring data quality and\nreliability. Although our community has continuously invested efforts into\nenhancing data in different aspects, they are often isolated initiatives on\nspecific tasks. To facilitate the collective initiative in our community and\npush forward DCAI, we draw a big picture and bring together three general\nmissions: training data development, inference data development, and data\nmaintenance. We provide a top-level discussion on representative DCAI tasks and\nshare perspectives. Finally, we list open challenges. More resources are\nsummarized at https://github.com/daochenzha/data-centric-AI\n", "date": "2023-01-12", "categories": []}
{"id": "2210.06441", "title": "How Much Data Are Augmentations Worth? An Investigation into Scaling\n  Laws, Invariance, and Implicit Regularization", "abstract": "  Despite the clear performance benefits of data augmentations, little is known\nabout why they are so effective. In this paper, we disentangle several key\nmechanisms through which data augmentations operate. Establishing an exchange\nrate between augmented and additional real data, we find that in\nout-of-distribution testing scenarios, augmentations which yield samples that\nare diverse, but inconsistent with the data distribution can be even more\nvaluable than additional training data. Moreover, we find that data\naugmentations which encourage invariances can be more valuable than invariance\nalone, especially on small and medium sized training sets. Following this\nobservation, we show that augmentations induce additional stochasticity during\ntraining, effectively flattening the loss landscape.\n", "date": "2022-10-12", "categories": []}
{"id": "2303.16186", "title": "Large-scale Training Data Search for Object Re-identification", "abstract": "  We consider a scenario where we have access to the target domain, but cannot\nafford on-the-fly training data annotation, and instead would like to construct\nan alternative training set from a large-scale data pool such that a\ncompetitive model can be obtained. We propose a search and pruning (SnP)\nsolution to this training data search problem, tailored to object\nre-identification (re-ID), an application aiming to match the same object\ncaptured by different cameras. Specifically, the search stage identifies and\nmerges clusters of source identities which exhibit similar distributions with\nthe target domain. The second stage, subject to a budget, then selects\nidentities and their images from the Stage I output, to control the size of the\nresulting training set for efficient training. The two steps provide us with\ntraining sets 80\\% smaller than the source pool while achieving a similar or\neven higher re-ID accuracy. These training sets are also shown to be superior\nto a few existing search methods such as random sampling and greedy sampling\nunder the same budget on training data size. If we release the budget, training\nsets resulting from the first stage alone allow even higher re-ID accuracy. We\nprovide interesting discussions on the specificity of our method to the re-ID\nproblem and particularly its role in bridging the re-ID domain gap. The code is\navailable at https://github.com/yorkeyao/SnP.\n", "date": "2023-03-28", "categories": []}
{"id": "2211.11004", "title": "Minimizing the Accumulated Trajectory Error to Improve Dataset\n  Distillation", "abstract": "  Model-based deep learning has achieved astounding successes due in part to\nthe availability of large-scale real-world data. However, processing such\nmassive amounts of data comes at a considerable cost in terms of computations,\nstorage, training and the search for good neural architectures. Dataset\ndistillation has thus recently come to the fore. This paradigm involves\ndistilling information from large real-world datasets into tiny and compact\nsynthetic datasets such that processing the latter ideally yields similar\nperformances as the former. State-of-the-art methods primarily rely on learning\nthe synthetic dataset by matching the gradients obtained during training\nbetween the real and synthetic data. However, these gradient-matching methods\nsuffer from the so-called accumulated trajectory error caused by the\ndiscrepancy between the distillation and subsequent evaluation. To mitigate the\nadverse impact of this accumulated trajectory error, we propose a novel\napproach that encourages the optimization algorithm to seek a flat trajectory.\nWe show that the weights trained on synthetic data are robust against the\naccumulated errors perturbations with the regularization towards the flat\ntrajectory. Our method, called Flat Trajectory Distillation (FTD), is shown to\nboost the performance of gradient-matching methods by up to 4.7% on a subset of\nimages of the ImageNet dataset with higher resolution images. We also validate\nthe effectiveness and generalizability of our method with datasets of different\nresolutions and demonstrate its applicability to neural architecture search.\nCode is available at https://github.com/AngusDujw/FTD-distillation.\n", "date": "2022-11-20", "categories": []}
{"id": "2303.13251", "title": "A Bag-of-Prototypes Representation for Dataset-Level Applications", "abstract": "  This work investigates dataset vectorization for two dataset-level tasks:\nassessing training set suitability and test set difficulty. The former measures\nhow suitable a training set is for a target domain, while the latter studies\nhow challenging a test set is for a learned model. Central to the two tasks is\nmeasuring the underlying relationship between datasets. This needs a desirable\ndataset vectorization scheme, which should preserve as much discriminative\ndataset information as possible so that the distance between the resulting\ndataset vectors can reflect dataset-to-dataset similarity. To this end, we\npropose a bag-of-prototypes (BoP) dataset representation that extends the\nimage-level bag consisting of patch descriptors to dataset-level bag consisting\nof semantic prototypes. Specifically, we develop a codebook consisting of K\nprototypes clustered from a reference dataset. Given a dataset to be encoded,\nwe quantize each of its image features to a certain prototype in the codebook\nand obtain a K-dimensional histogram. Without assuming access to dataset\nlabels, the BoP representation provides a rich characterization of the dataset\nsemantic distribution. Furthermore, BoP representations cooperate well with\nJensen-Shannon divergence for measuring dataset-to-dataset similarity. Although\nvery simple, BoP consistently shows its advantage over existing representations\non a series of benchmarks for two dataset-level tasks.\n", "date": "2023-03-23", "categories": []}
{"id": "2303.09540", "title": "SemDeDup: Data-efficient learning at web-scale through semantic\n  deduplication", "abstract": "  Progress in machine learning has been driven in large part by massive\nincreases in data. However, large web-scale datasets such as LAION are largely\nuncurated beyond searches for exact duplicates, potentially leaving much\nredundancy. Here, we introduce SemDeDup, a method which leverages embeddings\nfrom pre-trained models to identify and remove semantic duplicates: data pairs\nwhich are semantically similar, but not exactly identical. Removing semantic\nduplicates preserves performance and speeds up learning. Analyzing a subset of\nLAION, we show that SemDeDup can remove 50% of the data with minimal\nperformance loss, effectively halving training time. Moreover, performance\nincreases out of distribution. Also, analyzing language models trained on C4, a\npartially curated dataset, we show that SemDeDup improves over prior approaches\nwhile providing efficiency gains. SemDeDup provides an example of how simple\nways of leveraging quality embeddings can be used to make models learn faster\nwith less data.\n", "date": "2023-03-16", "categories": []}
{"id": "2212.12977", "title": "SMMix: Self-Motivated Image Mixing for Vision Transformers", "abstract": "  CutMix is a vital augmentation strategy that determines the performance and\ngeneralization ability of vision transformers (ViTs). However, the\ninconsistency between the mixed images and the corresponding labels harms its\nefficacy. Existing CutMix variants tackle this problem by generating more\nconsistent mixed images or more precise mixed labels, but inevitably introduce\nheavy training overhead or require extra information, undermining ease of use.\nTo this end, we propose an novel and effective Self-Motivated image Mixing\nmethod (SMMix), which motivates both image and label enhancement by the model\nunder training itself. Specifically, we propose a max-min attention region\nmixing approach that enriches the attention-focused objects in the mixed\nimages. Then, we introduce a fine-grained label assignment technique that\nco-trains the output tokens of mixed images with fine-grained supervision.\nMoreover, we devise a novel feature consistency constraint to align features\nfrom mixed and unmixed images. Due to the subtle designs of the self-motivated\nparadigm, our SMMix is significant in its smaller training overhead and better\nperformance than other CutMix variants. In particular, SMMix improves the\naccuracy of DeiT-T/S/B, CaiT-XXS-24/36, and PVT-T/S/M/L by more than +1% on\nImageNet-1k. The generalization capability of our method is also demonstrated\non downstream tasks and out-of-distribution datasets. Our project is\nanonymously available at https://github.com/ChenMnZ/SMMix.\n", "date": "2022-12-26", "categories": []}
{"id": "2303.02095", "title": "Data-Efficient Training of CNNs and Transformers with Coresets: A\n  Stability Perspective", "abstract": "  Coreset selection is among the most effective ways to reduce the training\ntime of CNNs, however, only limited is known on how the resultant models will\nbehave under variations of the coreset size, and choice of datasets and models.\nMoreover, given the recent paradigm shift towards transformer-based models, it\nis still an open question how coreset selection would impact their performance.\nThere are several similar intriguing questions that need to be answered for a\nwide acceptance of coreset selection methods, and this paper attempts to answer\nsome of these. We present a systematic benchmarking setup and perform a\nrigorous comparison of different coreset selection methods on CNNs and\ntransformers. Our investigation reveals that under certain circumstances,\nrandom selection of subsets is more robust and stable when compared with the\nSOTA selection methods. We demonstrate that the conventional concept of uniform\nsubset sampling across the various classes of the data is not the appropriate\nchoice. Rather samples should be adaptively chosen based on the complexity of\nthe data distribution for each class. Transformers are generally pretrained on\nlarge datasets, and we show that for certain target datasets, it helps to keep\ntheir performance stable at even very small coreset sizes. We further show that\nwhen no pretraining is done or when the pretrained transformer models are used\nwith non-natural images (e.g. medical data), CNNs tend to generalize better\nthan transformers at even very small coreset sizes. Lastly, we demonstrate that\nin the absence of the right pretraining, CNNs are better at learning the\nsemantic coherence between spatially distant objects within an image, and these\ntend to outperform transformers at almost all choices of the coreset size.\n", "date": "2023-03-03", "categories": []}
{"id": "2303.05498", "title": "Mark My Words: Dangers of Watermarked Images in ImageNet", "abstract": "  The utilization of pre-trained networks, especially those trained on\nImageNet, has become a common practice in Computer Vision. However, prior\nresearch has indicated that a significant number of images in the ImageNet\ndataset contain watermarks, making pre-trained networks susceptible to learning\nartifacts such as watermark patterns within their latent spaces. In this paper,\nwe aim to assess the extent to which popular pre-trained architectures display\nsuch behavior and to determine which classes are most affected. Additionally,\nwe examine the impact of watermarks on the extracted features. Contrary to the\npopular belief that the Chinese logographic watermarks impact the \"carton\"\nclass only, our analysis reveals that a variety of ImageNet classes, such as\n\"monitor\", \"broom\", \"apron\" and \"safe\" rely on spurious correlations. Finally,\nwe propose a simple approach to mitigate this issue in fine-tuned networks by\nignoring the encodings from the feature-extractor layer of ImageNet pre-trained\nnetworks that are most susceptible to watermark imprints.\n", "date": "2023-03-09", "categories": []}
{"id": "2303.04449", "title": "Loss-Curvature Matching for Dataset Selection and Condensation", "abstract": "  Training neural networks on a large dataset requires substantial\ncomputational costs. Dataset reduction selects or synthesizes data instances\nbased on the large dataset, while minimizing the degradation in generalization\nperformance from the full dataset. Existing methods utilize the neural network\nduring the dataset reduction procedure, so the model parameter becomes\nimportant factor in preserving the performance after reduction. By depending\nupon the importance of parameters, this paper introduces a new reduction\nobjective, coined LCMat, which Matches the Loss Curvatures of the original\ndataset and reduced dataset over the model parameter space, more than the\nparameter point. This new objective induces a better adaptation of the reduced\ndataset on the perturbed parameter region than the exact point matching.\nParticularly, we identify the worst case of the loss curvature gap from the\nlocal parameter region, and we derive the implementable upper bound of such\nworst-case with theoretical analyses. Our experiments on both coreset selection\nand condensation benchmarks illustrate that LCMat shows better generalization\nperformances than existing baselines.\n", "date": "2023-03-08", "categories": []}
{"id": "2301.00930", "title": "Data Valuation Without Training of a Model", "abstract": "  Many recent works on understanding deep learning try to quantify how much\nindividual data instances influence the optimization and generalization of a\nmodel. Such attempts reveal characteristics and importance of individual\ninstances, which may provide useful information in diagnosing and improving\ndeep learning. However, most of the existing works on data valuation require\nactual training of a model, which often demands high-computational cost. In\nthis paper, we provide a training-free data valuation score, called\ncomplexity-gap score, which is a data-centric score to quantify the influence\nof individual instances in generalization of two-layer overparameterized neural\nnetworks. The proposed score can quantify irregularity of the instances and\nmeasure how much each data instance contributes in the total movement of the\nnetwork parameters during training. We theoretically analyze and empirically\ndemonstrate the effectiveness of the complexity-gap score in finding `irregular\nor mislabeled' data instances, and also provide applications of the score in\nanalyzing datasets and diagnosing training dynamics. Our code is publicly\navailable at https://github.com/JJchy/CG_score\n", "date": "2023-01-03", "categories": []}
{"id": "2210.15809", "title": "Coverage-centric Coreset Selection for High Pruning Rates", "abstract": "  One-shot coreset selection aims to select a representative subset of the\ntraining data, given a pruning rate, that can later be used to train future\nmodels while retaining high accuracy. State-of-the-art coreset selection\nmethods pick the highest importance examples based on an importance metric and\nare found to perform well at low pruning rates. However, at high pruning rates,\nthey suffer from a catastrophic accuracy drop, performing worse than even\nrandom sampling. This paper explores the reasons behind this accuracy drop both\ntheoretically and empirically. We first propose a novel metric to measure the\ncoverage of a dataset on a specific distribution by extending the classical\ngeometric set cover problem to a distribution cover problem. This metric helps\nexplain why coresets selected by SOTA methods at high pruning rates perform\npoorly compared to random sampling because of worse data coverage. We then\npropose a novel one-shot coreset selection method, Coverage-centric Coreset\nSelection (CCS), that jointly considers overall data coverage upon a\ndistribution as well as the importance of each example. We evaluate CCS on five\ndatasets and show that, at high pruning rates (e.g., 90%), it achieves\nsignificantly better accuracy than previous SOTA methods (e.g., at least 19.56%\nhigher on CIFAR10) as well as random selection (e.g., 7.04% higher on CIFAR10)\nand comparable accuracy at low pruning rates. We make our code publicly\navailable at\nhttps://github.com/haizhongzheng/Coverage-centric-coreset-selection.\n", "date": "2022-10-28", "categories": []}
{"id": "2303.02081", "title": "Unproportional mosaicing", "abstract": "  Data shift is a gap between data distribution used for training and data\ndistribution encountered in the real-world. Data augmentations help narrow the\ngap by generating new data samples, increasing data variability, and data space\ncoverage. We present a new data augmentation: Unproportional mosaicing\n(Unprop). Our augmentation randomly splits an image into various-sized blocks\nand swaps its content (pixels) while maintaining block sizes. Our method\nachieves a lower error rate when combined with other state-of-the-art\naugmentations.\n", "date": "2023-03-03", "categories": []}
{"id": "2302.11188", "title": "What Are Effective Labels for Augmented Data? Improving Calibration and\n  Robustness with AutoLabel", "abstract": "  A wide breadth of research has devised data augmentation approaches that can\nimprove both accuracy and generalization performance for neural networks.\nHowever, augmented data can end up being far from the clean training data and\nwhat is the appropriate label is less clear. Despite this, most existing work\nsimply uses one-hot labels for augmented data. In this paper, we show re-using\none-hot labels for highly distorted data might run the risk of adding noise and\ndegrading accuracy and calibration. To mitigate this, we propose a generic\nmethod AutoLabel to automatically learn the confidence in the labels for\naugmented data, based on the transformation distance between the clean\ndistribution and augmented distribution. AutoLabel is built on label smoothing\nand is guided by the calibration-performance over a hold-out validation set. We\nsuccessfully apply AutoLabel to three different data augmentation techniques:\nthe state-of-the-art RandAug, AugMix, and adversarial training. Experiments on\nCIFAR-10, CIFAR-100 and ImageNet show that AutoLabel significantly improves\nexisting data augmentation techniques over models' calibration and accuracy,\nespecially under distributional shift.\n", "date": "2023-02-22", "categories": []}
{"id": "2302.11431", "title": "A Note on \"Towards Efficient Data Valuation Based on the Shapley Value''", "abstract": "  The Shapley value (SV) has emerged as a promising method for data valuation.\nHowever, computing or estimating the SV is often computationally expensive. To\novercome this challenge, Jia et al. (2019) propose an advanced SV estimation\nalgorithm called ``Group Testing-based SV estimator'' which achieves favorable\nasymptotic sample complexity. In this technical note, we present several\nimprovements in the analysis and design choices of this SV estimator. Moreover,\nwe point out that the Group Testing-based SV estimator does not fully reuse the\ncollected samples. Our analysis and insights contribute to a better\nunderstanding of the challenges in developing efficient SV estimation\nalgorithms for data valuation.\n", "date": "2023-02-22", "categories": []}
{"id": "2302.10679", "title": "Evaluating the effect of data augmentation and BALD heuristics on\n  distillation of Semantic-KITTI dataset", "abstract": "  Active Learning (AL) has remained relatively unexplored for LiDAR perception\ntasks in autonomous driving datasets. In this study we evaluate Bayesian active\nlearning methods applied to the task of dataset distillation or core subset\nselection (subset with near equivalent performance as full dataset). We also\nstudy the effect of application of data augmentation (DA) within Bayesian AL\nbased dataset distillation. We perform these experiments on the full\nSemantic-KITTI dataset. We extend our study over our existing work only on\n1/4th of the same dataset. Addition of DA and BALD have a negative impact over\nthe labeling efficiency and thus the capacity to distill datasets. We\ndemonstrate key issues in designing a functional AL framework and finally\nconclude with a review of challenges in real world active learning.\n", "date": "2023-02-21", "categories": []}
{"id": "2302.08682", "title": "Random Padding Data Augmentation", "abstract": "  The convolutional neural network (CNN) learns the same object in different\npositions in images, which can improve the recognition accuracy of the model.\nAn implication of this is that CNN may know where the object is. The usefulness\nof the features' spatial information in CNNs has not been well investigated. In\nthis paper, we found that the model's learning of features' position\ninformation hindered the learning of the features' relationship. Therefore, we\nintroduced Random Padding, a new type of padding method for training CNNs that\nimpairs the architecture's capacity to learn position information by adding\nzero-padding randomly to half of the border of feature maps. Random Padding is\nparameter-free, simple to construct, and compatible with the majority of\nCNN-based recognition models. This technique is also complementary to data\naugmentations such as random cropping, rotation, flipping and erasing, and\nconsistently improves the performance of image classification over strong\nbaselines.\n", "date": "2023-02-17", "categories": []}
{"id": "2212.05129", "title": "Measuring Data", "abstract": "  We identify the task of measuring data to quantitatively characterize the\ncomposition of machine learning data and datasets. Similar to an object's\nheight, width, and volume, data measurements quantify different attributes of\ndata along common dimensions that support comparison. Several lines of research\nhave proposed what we refer to as measurements, with differing terminology; we\nbring some of this work together, particularly in fields of computer vision and\nlanguage, and build from it to motivate measuring data as a critical component\nof responsible AI development. Measuring data aids in systematically building\nand analyzing machine learning (ML) data towards specific goals and gaining\nbetter control of what modern ML systems will learn. We conclude with a\ndiscussion of the many avenues of future work, the limitations of data\nmeasurements, and how to leverage these measurement approaches in research and\npractice.\n", "date": "2022-12-09", "categories": []}
{"id": "2302.01374", "title": "Neural Network Architecture for Database Augmentation Using Shared\n  Features", "abstract": "  The popularity of learning from data with machine learning and neural\nnetworks has lead to the creation of many new datasets for almost every problem\ndomain. However, even within a single domain, these datasets are often\ncollected with disparate features, sampled from different sub-populations, and\nrecorded at different time points. Even with the plethora of individual\ndatasets, large data science projects can be difficult as it is often not\ntrivial to merge these smaller datasets. Inherent challenges in some domains\nsuch as medicine also makes it very difficult to create large single source\ndatasets or multi-source datasets with identical features. Instead of trying to\nmerge these non-matching datasets directly, we propose a neural network\narchitecture that can provide data augmentation using features common between\nthese datasets. Our results show that this style of data augmentation can work\nfor both image and tabular data.\n", "date": "2023-02-02", "categories": []}
{"id": "2208.05516", "title": "Quality Not Quantity: On the Interaction between Dataset Design and\n  Robustness of CLIP", "abstract": "  Web-crawled datasets have enabled remarkable generalization capabilities in\nrecent image-text models such as CLIP (Contrastive Language-Image pre-training)\nor Flamingo, but little is known about the dataset creation processes. In this\nwork, we introduce a testbed of six publicly available data sources - YFCC,\nLAION, Conceptual Captions, WIT, RedCaps, Shutterstock - to investigate how\npre-training distributions induce robustness in CLIP. We find that the\nperformance of the pre-training data varies substantially across distribution\nshifts, with no single data source dominating. Moreover, we systematically\nstudy the interactions between these data sources and find that combining\nmultiple sources does not necessarily yield better models, but rather dilutes\nthe robustness of the best individual data source. We complement our empirical\nfindings with theoretical insights from a simple setting, where combining the\ntraining data also results in diluted robustness. In addition, our theoretical\nmodel provides a candidate explanation for the success of the CLIP-based data\nfiltering technique recently employed in the LAION dataset. Overall our results\ndemonstrate that simply gathering a large amount of data from the web is not\nthe most effective way to build a pre-training dataset for robust\ngeneralization, necessitating further study into dataset design. Code is\navailable at https://github.com/mlfoundations/clip_quality_not_quantity.\n", "date": "2022-08-10", "categories": []}
{"id": "2301.12052", "title": "Leveraging Importance Weights in Subset Selection", "abstract": "  We present a subset selection algorithm designed to work with arbitrary model\nfamilies in a practical batch setting. In such a setting, an algorithm can\nsample examples one at a time but, in order to limit overhead costs, is only\nable to update its state (i.e. further train model weights) once a large enough\nbatch of examples is selected. Our algorithm, IWeS, selects examples by\nimportance sampling where the sampling probability assigned to each example is\nbased on the entropy of models trained on previously selected batches. IWeS\nadmits significant performance improvement compared to other subset selection\nalgorithms for seven publicly available datasets. Additionally, it is\ncompetitive in an active learning setting, where the label information is not\navailable at selection time. We also provide an initial theoretical analysis to\nsupport our importance weighting approach, proving generalization and sampling\nrate bounds.\n", "date": "2023-01-28", "categories": []}
{"id": "2212.14229", "title": "Condensed Representation of Machine Learning Data", "abstract": "  Training of a Machine Learning model requires sufficient data. The\nsufficiency of the data is not always about the quantity, but about the\nrelevancy and reduced redundancy. Data-generating processes create massive\namounts of data. When used raw, such big data is causing much computational\nresource utilization. Instead of using the raw data, a proper Condensed\nRepresentation can be used instead. Combining K-means, a well-known clustering\nmethod, with some correction and refinement facilities a novel Condensed\nRepresentation method for Machine Learning applications is introduced. To\npresent the novel method meaningfully and visually, synthetically generated\ndata is employed. It has been shown that by using the condensed representation,\ninstead of the raw data, acceptably accurate model training is possible.\n", "date": "2022-12-29", "categories": []}
{"id": "2209.14603", "title": "Dataset Distillation for Medical Dataset Sharing", "abstract": "  Sharing medical datasets between hospitals is challenging because of the\nprivacy-protection problem and the massive cost of transmitting and storing\nmany high-resolution medical images. However, dataset distillation can\nsynthesize a small dataset such that models trained on it achieve comparable\nperformance with the original large dataset, which shows potential for solving\nthe existing medical sharing problems. Hence, this paper proposes a novel\ndataset distillation-based method for medical dataset sharing. Experimental\nresults on a COVID-19 chest X-ray image dataset show that our method can\nachieve high detection performance even using scarce anonymized chest X-ray\nimages.\n", "date": "2022-09-29", "categories": []}
{"id": "2212.10054", "title": "VoronoiPatches: Evaluating A New Data Augmentation Method", "abstract": "  Overfitting is a problem in Convolutional Neural Networks (CNN) that causes\npoor generalization of models on unseen data. To remediate this problem, many\nnew and diverse data augmentation methods (DA) have been proposed to supplement\nor generate more training data, and thereby increase its quality. In this work,\nwe propose a new data augmentation algorithm: VoronoiPatches (VP). We primarily\nutilize non-linear recombination of information within an image, fragmenting\nand occluding small information patches. Unlike other DA methods, VP uses small\nconvex polygon-shaped patches in a random layout to transport information\naround within an image. Sudden transitions created between patches and the\noriginal image can, optionally, be smoothed. In our experiments, VP\noutperformed current DA methods regarding model variance and overfitting\ntendencies. We demonstrate data augmentation utilizing non-linear\nre-combination of information within images, and non-orthogonal shapes and\nstructures improves CNN model robustness on unseen data.\n", "date": "2022-12-20", "categories": []}
{"id": "2212.11468", "title": "IPProtect: protecting the intellectual property of visual datasets\n  during data valuation", "abstract": "  Data trading is essential to accelerate the development of data-driven\nmachine learning pipelines. The central problem in data trading is to estimate\nthe utility of a seller's dataset with respect to a given buyer's machine\nlearning task, also known as data valuation. Typically, data valuation requires\none or more participants to share their raw dataset with others, leading to\npotential risks of intellectual property (IP) violations. In this paper, we\ntackle the novel task of preemptively protecting the IP of datasets that need\nto be shared during data valuation. First, we identify and formalize two kinds\nof novel IP risks in visual datasets: data-item (image) IP and statistical\n(dataset) IP. Then, we propose a novel algorithm to convert the raw dataset\ninto a sanitized version, that provides resistance to IP violations, while at\nthe same time allowing accurate data valuation. The key idea is to limit the\ntransfer of information from the raw dataset to the sanitized dataset, thereby\nprotecting against potential intellectual property violations. Next, we analyze\nour method for the likely existence of a solution and immunity against\nreconstruction attacks. Finally, we conduct extensive experiments on three\ncomputer vision datasets demonstrating the advantages of our method in\ncomparison to other baselines.\n", "date": "2022-12-22", "categories": []}
{"id": "2212.10553", "title": "RangeAugment: Efficient Online Augmentation with Range Learning", "abstract": "  State-of-the-art automatic augmentation methods (e.g., AutoAugment and\nRandAugment) for visual recognition tasks diversify training data using a large\nset of augmentation operations. The range of magnitudes of many augmentation\noperations (e.g., brightness and contrast) is continuous. Therefore, to make\nsearch computationally tractable, these methods use fixed and manually-defined\nmagnitude ranges for each operation, which may lead to sub-optimal policies. To\nanswer the open question on the importance of magnitude ranges for each\naugmentation operation, we introduce RangeAugment that allows us to efficiently\nlearn the range of magnitudes for individual as well as composite augmentation\noperations. RangeAugment uses an auxiliary loss based on image similarity as a\nmeasure to control the range of magnitudes of augmentation operations. As a\nresult, RangeAugment has a single scalar parameter for search, image\nsimilarity, which we simply optimize via linear search. RangeAugment integrates\nseamlessly with any model and learns model- and task-specific augmentation\npolicies. With extensive experiments on the ImageNet dataset across different\nnetworks, we show that RangeAugment achieves competitive performance to\nstate-of-the-art automatic augmentation methods with 4-5 times fewer\naugmentation operations. Experimental results on semantic segmentation, object\ndetection, foundation models, and knowledge distillation further shows\nRangeAugment's effectiveness.\n", "date": "2022-12-20", "categories": []}
{"id": "2212.04059", "title": "MixBoost: Improving the Robustness of Deep Neural Networks by Boosting\n  Data Augmentation", "abstract": "  As more and more artificial intelligence (AI) technologies move from the\nlaboratory to real-world applications, the open-set and robustness challenges\nbrought by data from the real world have received increasing attention. Data\naugmentation is a widely used method to improve model performance, and some\nrecent works have also confirmed its positive effect on the robustness of AI\nmodels. However, most of the existing data augmentation methods are heuristic,\nlacking the exploration of their internal mechanisms. We apply the explainable\nartificial intelligence (XAI) method, explore the internal mechanisms of\npopular data augmentation methods, analyze the relationship between game\ninteractions and some widely used robustness metrics, and propose a new proxy\nfor model robustness in the open-set environment. Based on the analysis of the\ninternal mechanisms, we develop a mask-based boosting method for data\naugmentation that comprehensively improves several robustness measures of AI\nmodels and beats state-of-the-art data augmentation approaches. Experiments\nshow that our method can be widely applied to many popular data augmentation\nmethods. Different from the adversarial training, our boosting method not only\nsignificantly improves the robustness of models, but also improves the accuracy\nof test sets. Our code is available at\n\\url{https://github.com/Anonymous_for_submission}.\n", "date": "2022-12-08", "categories": []}
{"id": "2211.15846", "title": "LUMix: Improving Mixup by Better Modelling Label Uncertainty", "abstract": "  Modern deep networks can be better generalized when trained with noisy\nsamples and regularization techniques. Mixup and CutMix have been proven to be\neffective for data augmentation to help avoid overfitting. Previous Mixup-based\nmethods linearly combine images and labels to generate additional training\ndata. However, this is problematic if the object does not occupy the whole\nimage as we demonstrate in Figure 1. Correctly assigning the label weights is\nhard even for human beings and there is no clear criterion to measure it. To\ntackle this problem, in this paper, we propose LUMix, which models such\nuncertainty by adding label perturbation during training. LUMix is simple as it\ncan be implemented in just a few lines of code and can be universally applied\nto any deep networks \\eg CNNs and Vision Transformers, with minimal\ncomputational cost. Extensive experiments show that our LUMix can consistently\nboost the performance for networks with a wide range of diversity and capacity\non ImageNet, \\eg $+0.7\\%$ for a small model DeiT-S and $+0.6\\%$ for a large\nvariant XCiT-L. We also demonstrate that LUMix can lead to better robustness\nwhen evaluated on ImageNet-O and ImageNet-A. The source code can be found\n\\href{https://github.com/kevin-ssy/LUMix}{here}\n", "date": "2022-11-29", "categories": []}
{"id": "2206.02916", "title": "Remember the Past: Distilling Datasets into Addressable Memories for\n  Neural Networks", "abstract": "  We propose an algorithm that compresses the critical information of a large\ndataset into compact addressable memories. These memories can then be recalled\nto quickly re-train a neural network and recover the performance (instead of\nstoring and re-training on the full original dataset). Building upon the\ndataset distillation framework, we make a key observation that a shared common\nrepresentation allows for more efficient and effective distillation.\nConcretely, we learn a set of bases (aka ``memories'') which are shared between\nclasses and combined through learned flexible addressing functions to generate\na diverse set of training examples. This leads to several benefits: 1) the size\nof compressed data does not necessarily grow linearly with the number of\nclasses; 2) an overall higher compression rate with more effective distillation\nis achieved; and 3) more generalized queries are allowed beyond recalling the\noriginal classes. We demonstrate state-of-the-art results on the dataset\ndistillation task across six benchmarks, including up to 16.5% and 9.7% in\nretained accuracy improvement when distilling CIFAR10 and CIFAR100\nrespectively. We then leverage our framework to perform continual learning,\nachieving state-of-the-art results on four benchmarks, with 23.2% accuracy\nimprovement on MANY. The code is released on our project webpage\nhttps://github.com/princetonvisualai/RememberThePast-DatasetDistillation.\n", "date": "2022-06-06", "categories": []}
{"id": "2211.10944", "title": "Feature Weaken: Vicinal Data Augmentation for Classification", "abstract": "  Deep learning usually relies on training large-scale data samples to achieve\nbetter performance. However, over-fitting based on training data always remains\na problem. Scholars have proposed various strategies, such as feature dropping\nand feature mixing, to improve the generalization continuously. For the same\npurpose, we subversively propose a novel training method, Feature Weaken, which\ncan be regarded as a data augmentation method. Feature Weaken constructs the\nvicinal data distribution with the same cosine similarity for model training by\nweakening features of the original samples. In especially, Feature Weaken\nchanges the spatial distribution of samples, adjusts sample boundaries, and\nreduces the gradient optimization value of back-propagation. This work can not\nonly improve the classification performance and generalization of the model,\nbut also stabilize the model training and accelerate the model convergence. We\nconduct extensive experiments on classical deep convolution neural models with\nfive common image classification datasets and the Bert model with four common\ntext classification datasets. Compared with the classical models or the\ngeneralization improvement methods, such as Dropout, Mixup, Cutout, and CutMix,\nFeature Weaken shows good compatibility and performance. We also use\nadversarial samples to perform the robustness experiments, and the results show\nthat Feature Weaken is effective in improving the robustness of the model.\n", "date": "2022-11-20", "categories": []}
{"id": "2211.09859", "title": "Data-Centric Debugging: mitigating model failures via targeted data\n  collection", "abstract": "  Deep neural networks can be unreliable in the real world when the training\nset does not adequately cover all the settings where they are deployed.\nFocusing on image classification, we consider the setting where we have an\nerror distribution $\\mathcal{E}$ representing a deployment scenario where the\nmodel fails. We have access to a small set of samples $\\mathcal{E}_{sample}$\nfrom $\\mathcal{E}$ and it can be expensive to obtain additional samples. In the\ntraditional model development framework, mitigating failures of the model in\n$\\mathcal{E}$ can be challenging and is often done in an ad hoc manner. In this\npaper, we propose a general methodology for model debugging that can\nsystemically improve model performance on $\\mathcal{E}$ while maintaining its\nperformance on the original test set. Our key assumption is that we have access\nto a large pool of weakly (noisily) labeled data $\\mathcal{F}$. However,\nnaively adding $\\mathcal{F}$ to the training would hurt model performance due\nto the large extent of label noise. Our Data-Centric Debugging (DCD) framework\ncarefully creates a debug-train set by selecting images from $\\mathcal{F}$ that\nare perceptually similar to the images in $\\mathcal{E}_{sample}$. To do this,\nwe use the $\\ell_2$ distance in the feature space (penultimate layer\nactivations) of various models including ResNet, Robust ResNet and DINO where\nwe observe DINO ViTs are significantly better at discovering similar images\ncompared to Resnets. Compared to LPIPS, we find that our method reduces compute\nand storage requirements by 99.58\\%. Compared to the baselines that maintain\nmodel performance on the test set, we achieve significantly (+9.45\\%) improved\nresults on the debug-heldout sets.\n", "date": "2022-11-17", "categories": []}
{"id": "2211.07859", "title": "Local Magnification for Data and Feature Augmentation", "abstract": "  In recent years, many data augmentation techniques have been proposed to\nincrease the diversity of input data and reduce the risk of overfitting on deep\nneural networks. In this work, we propose an easy-to-implement and model-free\ndata augmentation method called Local Magnification (LOMA). Different from\nother geometric data augmentation methods that perform global transformations\non images, LOMA generates additional training data by randomly magnifying a\nlocal area of the image. This local magnification results in geometric changes\nthat significantly broaden the range of augmentations while maintaining the\nrecognizability of objects. Moreover, we extend the idea of LOMA and random\ncropping to the feature space to augment the feature map, which further boosts\nthe classification accuracy considerably. Experiments show that our proposed\nLOMA, though straightforward, can be combined with standard data augmentation\nto significantly improve the performance on image classification and object\ndetection. And further combination with our feature augmentation techniques,\ntermed LOMA_IF&FO, can continue to strengthen the model and outperform advanced\nintensity transformation methods for data augmentation.\n", "date": "2022-11-15", "categories": []}
{"id": "2211.07959", "title": "The Lean Data Scientist: Recent Advances towards Overcoming the Data\n  Bottleneck", "abstract": "  Machine learning (ML) is revolutionizing the world, affecting almost every\nfield of science and industry. Recent algorithms (in particular, deep networks)\nare increasingly data-hungry, requiring large datasets for training. Thus, the\ndominant paradigm in ML today involves constructing large, task-specific\ndatasets.\n  However, obtaining quality datasets of such magnitude proves to be a\ndifficult challenge. A variety of methods have been proposed to address this\ndata bottleneck problem, but they are scattered across different areas, and it\nis hard for a practitioner to keep up with the latest developments. In this\nwork, we propose a taxonomy of these methods. Our goal is twofold: (1) We wish\nto raise the community's awareness of the methods that already exist and\nencourage more efficient use of resources, and (2) we hope that such a taxonomy\nwill contribute to our understanding of the problem, inspiring novel ideas and\nstrategies to replace current annotation-heavy approaches.\n", "date": "2022-11-15", "categories": []}
{"id": "2211.06800", "title": "CS-Shapley: Class-wise Shapley Values for Data Valuation in\n  Classification", "abstract": "  Data valuation, or the valuation of individual datum contributions, has seen\ngrowing interest in machine learning due to its demonstrable efficacy for tasks\nsuch as noisy label detection. In particular, due to the desirable axiomatic\nproperties, several Shapley value approximation methods have been proposed. In\nthese methods, the value function is typically defined as the predictive\naccuracy over the entire development set. However, this limits the ability to\ndifferentiate between training instances that are helpful or harmful to their\nown classes. Intuitively, instances that harm their own classes may be noisy or\nmislabeled and should receive a lower valuation than helpful instances. In this\nwork, we propose CS-Shapley, a Shapley value with a new value function that\ndiscriminates between training instances' in-class and out-of-class\ncontributions. Our theoretical analysis shows the proposed value function is\n(essentially) the unique function that satisfies two desirable properties for\nevaluating data values in classification. Further, our experiments on two\nbenchmark evaluation tasks (data removal and noisy label detection) and four\nclassifiers demonstrate the effectiveness of CS-Shapley over existing methods.\nLastly, we evaluate the \"transferability\" of data values estimated from one\nclassifier to others, and our results suggest Shapley-based data valuation is\ntransferable for application across different models.\n", "date": "2022-11-13", "categories": []}
{"id": "2207.06214", "title": "Is one annotation enough? A data-centric image classification benchmark\n  for noisy and ambiguous label estimation", "abstract": "  High-quality data is necessary for modern machine learning. However, the\nacquisition of such data is difficult due to noisy and ambiguous annotations of\nhumans. The aggregation of such annotations to determine the label of an image\nleads to a lower data quality. We propose a data-centric image classification\nbenchmark with ten real-world datasets and multiple annotations per image to\nallow researchers to investigate and quantify the impact of such data quality\nissues. With the benchmark we can study the impact of annotation costs and\n(semi-)supervised methods on the data quality for image classification by\napplying a novel methodology to a range of different algorithms and diverse\ndatasets. Our benchmark uses a two-phase approach via a data label improvement\nmethod in the first phase and a fixed evaluation model in the second phase.\nThereby, we give a measure for the relation between the input labeling effort\nand the performance of (semi-)supervised algorithms to enable a deeper insight\ninto how labels should be created for effective model training. Across\nthousands of experiments, we show that one annotation is not enough and that\nthe inclusion of multiple annotations allows for a better approximation of the\nreal underlying class distribution. We identify that hard labels can not\ncapture the ambiguity of the data and this might lead to the common issue of\noverconfident models. Based on the presented datasets, benchmarked methods, and\nanalysis, we create multiple research opportunities for the future directed at\nthe improvement of label noise estimation approaches, data annotation schemes,\nrealistic (semi-)supervised learning, or more reliable image collection.\n", "date": "2022-07-13", "categories": []}
{"id": "2211.01847", "title": "Seeing the Unseen: Errors and Bias in Visual Datasets", "abstract": "  From face recognition in smartphones to automatic routing on self-driving\ncars, machine vision algorithms lie in the core of these features. These\nsystems solve image based tasks by identifying and understanding objects,\nsubsequently making decisions from these information. However, errors in\ndatasets are usually induced or even magnified in algorithms, at times\nresulting in issues such as recognising black people as gorillas and\nmisrepresenting ethnicities in search results. This paper tracks the errors in\ndatasets and their impacts, revealing that a flawed dataset could be a result\nof limited categories, incomprehensive sourcing and poor classification.\n", "date": "2022-11-03", "categories": []}
{"id": "2211.00824", "title": "Adversarial Auto-Augment with Label Preservation: A Representation\n  Learning Principle Guided Approach", "abstract": "  Data augmentation is a critical contributing factor to the success of deep\nlearning but heavily relies on prior domain knowledge which is not always\navailable. Recent works on automatic data augmentation learn a policy to form a\nsequence of augmentation operations, which are still pre-defined and restricted\nto limited options. In this paper, we show that a prior-free autonomous data\naugmentation's objective can be derived from a representation learning\nprinciple that aims to preserve the minimum sufficient information of the\nlabels. Given an example, the objective aims at creating a distant \"hard\npositive example\" as the augmentation, while still preserving the original\nlabel. We then propose a practical surrogate to the objective that can be\noptimized efficiently and integrated seamlessly into existing methods for a\nbroad class of machine learning tasks, e.g., supervised, semi-supervised, and\nnoisy-label learning. Unlike previous works, our method does not require\ntraining an extra generative model but instead leverages the intermediate layer\nrepresentations of the end-task model for generating data augmentations. In\nexperiments, we show that our method consistently brings non-trivial\nimprovements to the three aforementioned learning tasks from both efficiency\nand final performance, either or not combined with strong pre-defined\naugmentations, e.g., on medical images when domain knowledge is unavailable and\nthe existing augmentation techniques perform poorly. Code is available at:\nhttps://github.com/kai-wen-yang/LPA3}{https://github.com/kai-wen-yang/LPA3.\n", "date": "2022-11-02", "categories": []}
{"id": "2211.00113", "title": "SAGE: Saliency-Guided Mixup with Optimal Rearrangements", "abstract": "  Data augmentation is a key element for training accurate models by reducing\noverfitting and improving generalization. For image classification, the most\npopular data augmentation techniques range from simple photometric and\ngeometrical transformations, to more complex methods that use visual saliency\nto craft new training examples. As augmentation methods get more complex, their\nability to increase the test accuracy improves, yet, such methods become\ncumbersome, inefficient and lead to poor out-of-domain generalization, as we\nshow in this paper. This motivates a new augmentation technique that allows for\nhigh accuracy gains while being simple, efficient (i.e., minimal computation\noverhead) and generalizable. To this end, we introduce Saliency-Guided Mixup\nwith Optimal Rearrangements (SAGE), which creates new training examples by\nrearranging and mixing image pairs using visual saliency as guidance. By\nexplicitly leveraging saliency, SAGE promotes discriminative foreground objects\nand produces informative new images useful for training. We demonstrate on\nCIFAR-10 and CIFAR-100 that SAGE achieves better or comparable performance to\nthe state of the art while being more efficient. Additionally, evaluations in\nthe out-of-distribution setting, and few-shot learning on mini-ImageNet, show\nthat SAGE achieves improved generalization performance without trading off\nrobustness.\n", "date": "2022-10-31", "categories": []}
{"id": "2210.16774", "title": "Dataset Distillation via Factorization", "abstract": "  In this paper, we study \\xw{dataset distillation (DD)}, from a novel\nperspective and introduce a \\emph{dataset factorization} approach, termed\n\\emph{HaBa}, which is a plug-and-play strategy portable to any existing DD\nbaseline. Unlike conventional DD approaches that aim to produce distilled and\nrepresentative samples, \\emph{HaBa} explores decomposing a dataset into two\ncomponents: data \\emph{Ha}llucination networks and \\emph{Ba}ses, where the\nlatter is fed into the former to reconstruct image samples. The flexible\ncombinations between bases and hallucination networks, therefore, equip the\ndistilled data with exponential informativeness gain, which largely increase\nthe representation capability of distilled datasets. To furthermore increase\nthe data efficiency of compression results, we further introduce a pair of\nadversarial contrastive constraints on the resultant hallucination networks and\nbases, which increase the diversity of generated images and inject more\ndiscriminant information into the factorization. Extensive comparisons and\nexperiments demonstrate that our method can yield significant improvement on\ndownstream classification tasks compared with previous state of the arts, while\nreducing the total number of compressed parameters by up to 65\\%. Moreover,\ndistilled datasets by our approach also achieve \\textasciitilde10\\% higher\naccuracy than baseline methods in cross-architecture generalization. Our code\nis available \\href{https://github.com/Huage001/DatasetFactorization}{here}.\n", "date": "2022-10-30", "categories": []}
{"id": "2210.15721", "title": "GraphMAD: Graph Mixup for Data Augmentation using Data-Driven Convex\n  Clustering", "abstract": "  We develop a novel data-driven nonlinear mixup mechanism for graph data\naugmentation and present different mixup functions for sample pairs and their\nlabels. Mixup is a data augmentation method to create new training data by\nlinearly interpolating between pairs of data samples and their labels. Mixup of\ngraph data is challenging since the interpolation between graphs of potentially\ndifferent sizes is an ill-posed operation. Hence, a promising approach for\ngraph mixup is to first project the graphs onto a common latent feature space\nand then explore linear and nonlinear mixup strategies in this latent space. In\nthis context, we propose to (i) project graphs onto the latent space of\ncontinuous random graph models known as graphons, (ii) leverage convex\nclustering in this latent space to generate nonlinear data-driven mixup\nfunctions, and (iii) investigate the use of different mixup functions for\nlabels and data samples. We evaluate our graph data augmentation performance on\nbenchmark datasets and demonstrate that nonlinear data-driven mixup functions\ncan significantly improve graph classification.\n", "date": "2022-10-27", "categories": []}
{"id": "2210.10951", "title": "Automatic Document Selection for Efficient Encoder Pretraining", "abstract": "  Building pretrained language models is considered expensive and\ndata-intensive, but must we increase dataset size to achieve better\nperformance? We propose an alternative to larger training sets by automatically\nidentifying smaller yet domain-representative subsets. We extend Cynical Data\nSelection, a statistical sentence scoring method that conditions on a\nrepresentative target domain corpus. As an example, we treat the OntoNotes\ncorpus as a target domain and pretrain a RoBERTa-like encoder from a cynically\nselected subset of the Pile. On both perplexity and across several downstream\ntasks in the target domain, it consistently outperforms random selection with\n20x less data, 3x fewer training iterations, and 2x less estimated cloud\ncompute cost, validating the recipe of automatic document selection for LM\npretraining.\n", "date": "2022-10-20", "categories": []}
{"id": "2206.00719", "title": "Dataset Distillation using Neural Feature Regression", "abstract": "  Dataset distillation aims to learn a small synthetic dataset that preserves\nmost of the information from the original dataset. Dataset distillation can be\nformulated as a bi-level meta-learning problem where the outer loop optimizes\nthe meta-dataset and the inner loop trains a model on the distilled data.\nMeta-gradient computation is one of the key challenges in this formulation, as\ndifferentiating through the inner loop learning procedure introduces\nsignificant computation and memory costs. In this paper, we address these\nchallenges using neural Feature Regression with Pooling (FRePo), achieving the\nstate-of-the-art performance with an order of magnitude less memory requirement\nand two orders of magnitude faster training than previous methods. The proposed\nalgorithm is analogous to truncated backpropagation through time with a pool of\nmodels to alleviate various types of overfitting in dataset distillation. FRePo\nsignificantly outperforms the previous methods on CIFAR100, Tiny ImageNet, and\nImageNet-1K. Furthermore, we show that high-quality distilled data can greatly\nimprove various downstream applications, such as continual learning and\nmembership inference defense. Please check out our webpage at\nhttps://sites.google.com/view/frepo.\n", "date": "2022-06-01", "categories": []}
{"id": "2208.10033", "title": "Evaluating and Crafting Datasets Effective for Deep Learning With Data\n  Maps", "abstract": "  Rapid development in deep learning model construction has prompted an\nincreased need for appropriate training data. The popularity of large datasets\n- sometimes known as \"big data\" - has diverted attention from assessing their\nquality. Training on large datasets often requires excessive system resources\nand an infeasible amount of time. Furthermore, the supervised machine learning\nprocess has yet to be fully automated: for supervised learning, large datasets\nrequire more time for manually labeling samples. We propose a method of\ncurating smaller datasets with comparable out-of-distribution model accuracy\nafter an initial training session using an appropriate distribution of samples\nclassified by how difficult it is for a model to learn from them.\n", "date": "2022-08-22", "categories": []}
{"id": "2210.11560", "title": "Finding Dataset Shortcuts with Grammar Induction", "abstract": "  Many NLP datasets have been found to contain shortcuts: simple decision rules\nthat achieve surprisingly high accuracy. However, it is difficult to discover\nshortcuts automatically. Prior work on automatic shortcut detection has focused\non enumerating features like unigrams or bigrams, which can find only low-level\nshortcuts, or relied on post-hoc model interpretability methods like saliency\nmaps, which reveal qualitative patterns without a clear statistical\ninterpretation. In this work, we propose to use probabilistic grammars to\ncharacterize and discover shortcuts in NLP datasets. Specifically, we use a\ncontext-free grammar to model patterns in sentence classification datasets and\nuse a synchronous context-free grammar to model datasets involving sentence\npairs. The resulting grammars reveal interesting shortcut features in a number\nof datasets, including both simple and high-level features, and automatically\nidentify groups of test examples on which conventional classifiers fail.\nFinally, we show that the features we discover can be used to generate\ndiagnostic contrast examples and incorporated into standard robust optimization\nmethods to improve worst-group accuracy.\n", "date": "2022-10-20", "categories": []}
{"id": "2210.12067", "title": "Efficient Dataset Distillation Using Random Feature Approximation", "abstract": "  Dataset distillation compresses large datasets into smaller synthetic\ncoresets which retain performance with the aim of reducing the storage and\ncomputational burden of processing the entire dataset. Today's best-performing\nalgorithm, \\textit{Kernel Inducing Points} (KIP), which makes use of the\ncorrespondence between infinite-width neural networks and kernel-ridge\nregression, is prohibitively slow due to the exact computation of the neural\ntangent kernel matrix, scaling $O(|S|^2)$, with $|S|$ being the coreset size.\nTo improve this, we propose a novel algorithm that uses a random feature\napproximation (RFA) of the Neural Network Gaussian Process (NNGP) kernel, which\nreduces the kernel matrix computation to $O(|S|)$. Our algorithm provides at\nleast a 100-fold speedup over KIP and can run on a single GPU. Our new method,\ntermed an RFA Distillation (RFAD), performs competitively with KIP and other\ndataset condensation algorithms in accuracy over a range of large-scale\ndatasets, both in kernel regression and finite-width network training. We\ndemonstrate the effectiveness of our approach on tasks involving model\ninterpretability and privacy preservation.\n", "date": "2022-10-21", "categories": []}
{"id": "2207.09639", "title": "DC-BENCH: Dataset Condensation Benchmark", "abstract": "  Dataset Condensation is a newly emerging technique aiming at learning a tiny\ndataset that captures the rich information encoded in the original dataset. As\nthe size of datasets contemporary machine learning models rely on becomes\nincreasingly large, condensation methods become a prominent direction for\naccelerating network training and reducing data storage. Despite numerous\nmethods have been proposed in this rapidly growing field, evaluating and\ncomparing different condensation methods is non-trivial and still remains an\nopen issue. The quality of condensed dataset are often shadowed by many\ncritical contributing factors to the end performance, such as data augmentation\nand model architectures. The lack of a systematic way to evaluate and compare\ncondensation methods not only hinders our understanding of existing techniques,\nbut also discourages practical usage of the synthesized datasets. This work\nprovides the first large-scale standardized benchmark on Dataset Condensation.\nIt consists of a suite of evaluations to comprehensively reflect the\ngenerability and effectiveness of condensation methods through the lens of\ntheir generated dataset. Leveraging this benchmark, we conduct a large-scale\nstudy of current condensation methods, and report many insightful findings that\nopen up new possibilities for future development. The benchmark library,\nincluding evaluators, baseline methods, and generated datasets, is open-sourced\nto facilitate future research and application.\n", "date": "2022-07-20", "categories": []}
{"id": "2209.09449", "title": "Data-Centric AI Paradigm Based on Application-Driven Fine-Grained\n  Dataset Design", "abstract": "  Deep learning has a wide range of applications in industrial scenario, but\nreducing false alarm (FA) remains a major difficulty. Optimizing network\narchitecture or network parameters is used to tackle this challenge in academic\ncircles, while ignoring the essential characteristics of data in application\nscenarios, which often results in increased FA in new scenarios. In this paper,\nwe propose a novel paradigm for fine-grained design of datasets, driven by\nindustrial applications. We flexibly select positive and negative sample sets\naccording to the essential features of the data and application requirements,\nand add the remaining samples to the training set as uncertainty classes. We\ncollect more than 10,000 mask-wearing recognition samples covering various\napplication scenarios as our experimental data. Compared with the traditional\ndata design methods, our method achieves better results and effectively reduces\nFA. We make all contributions available to the research community for broader\nuse. The contributions will be available at\nhttps://github.com/huh30/OpenDatasets.\n", "date": "2022-09-20", "categories": []}
{"id": "2210.07562", "title": "TokenMixup: Efficient Attention-guided Token-level Data Augmentation for\n  Transformers", "abstract": "  Mixup is a commonly adopted data augmentation technique for image\nclassification. Recent advances in mixup methods primarily focus on mixing\nbased on saliency. However, many saliency detectors require intense computation\nand are especially burdensome for parameter-heavy transformer models. To this\nend, we propose TokenMixup, an efficient attention-guided token-level data\naugmentation method that aims to maximize the saliency of a mixed set of\ntokens. TokenMixup provides x15 faster saliency-aware data augmentation\ncompared to gradient-based methods. Moreover, we introduce a variant of\nTokenMixup which mixes tokens within a single instance, thereby enabling\nmulti-scale feature augmentation. Experiments show that our methods\nsignificantly improve the baseline models' performance on CIFAR and\nImageNet-1K, while being more efficient than previous methods. We also reach\nstate-of-the-art performance on CIFAR-100 among from-scratch transformer\nmodels. Code is available at https://github.com/mlvlab/TokenMixup.\n", "date": "2022-10-14", "categories": []}
{"id": "2206.02626", "title": "Infinite Recommendation Networks: A Data-Centric Approach", "abstract": "  We leverage the Neural Tangent Kernel and its equivalence to training\ninfinitely-wide neural networks to devise $\\infty$-AE: an autoencoder with\ninfinitely-wide bottleneck layers. The outcome is a highly expressive yet\nsimplistic recommendation model with a single hyper-parameter and a closed-form\nsolution. Leveraging $\\infty$-AE's simplicity, we also develop Distill-CF for\nsynthesizing tiny, high-fidelity data summaries which distill the most\nimportant knowledge from the extremely large and sparse user-item interaction\nmatrix for efficient and accurate subsequent data-usage like model training,\ninference, architecture search, etc. This takes a data-centric approach to\nrecommendation, where we aim to improve the quality of logged user-feedback\ndata for subsequent modeling, independent of the learning algorithm. We\nparticularly utilize the concept of differentiable Gumbel-sampling to handle\nthe inherent data heterogeneity, sparsity, and semi-structuredness, while being\nscalable to datasets with hundreds of millions of user-item interactions. Both\nof our proposed approaches significantly outperform their respective\nstate-of-the-art and when used together, we observe 96-105% of $\\infty$-AE's\nperformance on the full dataset with as little as 0.1% of the original dataset\nsize, leading us to explore the counter-intuitive question: Is more data what\nyou need for better recommendation?\n", "date": "2022-06-03", "categories": []}
{"id": "2210.05775", "title": "C-Mixup: Improving Generalization in Regression", "abstract": "  Improving the generalization of deep networks is an important open challenge,\nparticularly in domains without plentiful data. The mixup algorithm improves\ngeneralization by linearly interpolating a pair of examples and their\ncorresponding labels. These interpolated examples augment the original training\nset. Mixup has shown promising results in various classification tasks, but\nsystematic analysis of mixup in regression remains underexplored. Using mixup\ndirectly on regression labels can result in arbitrarily incorrect labels. In\nthis paper, we propose a simple yet powerful algorithm, C-Mixup, to improve\ngeneralization on regression tasks. In contrast with vanilla mixup, which picks\ntraining examples for mixing with uniform probability, C-Mixup adjusts the\nsampling probability based on the similarity of the labels. Our theoretical\nanalysis confirms that C-Mixup with label similarity obtains a smaller mean\nsquare error in supervised regression and meta-regression than vanilla mixup\nand using feature similarity. Another benefit of C-Mixup is that it can improve\nout-of-distribution robustness, where the test distribution is different from\nthe training distribution. By selectively interpolating examples with similar\nlabels, it mitigates the effects of domain-associated information and yields\ndomain-invariant representations. We evaluate C-Mixup on eleven datasets,\nranging from tabular to video data. Compared to the best prior approach,\nC-Mixup achieves 6.56%, 4.76%, 5.82% improvements in in-distribution\ngeneralization, task generalization, and out-of-distribution robustness,\nrespectively. Code is released at https://github.com/huaxiuyao/C-Mixup.\n", "date": "2022-10-11", "categories": []}
{"id": "2210.06205", "title": "On Divergence Measures for Bayesian Pseudocoresets", "abstract": "  A Bayesian pseudocoreset is a small synthetic dataset for which the posterior\nover parameters approximates that of the original dataset. While promising, the\nscalability of Bayesian pseudocoresets is not yet validated in realistic\nproblems such as image classification with deep neural networks. On the other\nhand, dataset distillation methods similarly construct a small dataset such\nthat the optimization using the synthetic dataset converges to a solution with\nperformance competitive with optimization using full data. Although dataset\ndistillation has been empirically verified in large-scale settings, the\nframework is restricted to point estimates, and their adaptation to Bayesian\ninference has not been explored. This paper casts two representative dataset\ndistillation algorithms as approximations to methods for constructing\npseudocoresets by minimizing specific divergence measures: reverse KL\ndivergence and Wasserstein distance. Furthermore, we provide a unifying view of\nsuch divergence measures in Bayesian pseudocoreset construction. Finally, we\npropose a novel Bayesian pseudocoreset algorithm based on minimizing forward KL\ndivergence. Our empirical results demonstrate that the pseudocoresets\nconstructed from these methods reflect the true posterior even in\nhigh-dimensional Bayesian inference problems.\n", "date": "2022-10-12", "categories": []}
{"id": "2206.06544", "title": "A Survey of Automated Data Augmentation Algorithms for Deep\n  Learning-based Image Classification Tasks", "abstract": "  In recent years, one of the most popular techniques in the computer vision\ncommunity has been the deep learning technique. As a data-driven technique,\ndeep model requires enormous amounts of accurately labelled training data,\nwhich is often inaccessible in many real-world applications. A data-space\nsolution is Data Augmentation (DA), that can artificially generate new images\nout of original samples. Image augmentation strategies can vary by dataset, as\ndifferent data types might require different augmentations to facilitate model\ntraining. However, the design of DA policies has been largely decided by the\nhuman experts with domain knowledge, which is considered to be highly\nsubjective and error-prone. To mitigate such problem, a novel direction is to\nautomatically learn the image augmentation policies from the given dataset\nusing Automated Data Augmentation (AutoDA) techniques. The goal of AutoDA\nmodels is to find the optimal DA policies that can maximize the model\nperformance gains. This survey discusses the underlying reasons of the\nemergence of AutoDA technology from the perspective of image classification. We\nidentify three key components of a standard AutoDA model: a search space, a\nsearch algorithm and an evaluation function. Based on their architecture, we\nprovide a systematic taxonomy of existing image AutoDA approaches. This paper\npresents the major works in AutoDA field, discussing their pros and cons, and\nproposing several potential directions for future improvements.\n", "date": "2022-06-14", "categories": []}
{"id": "2210.01234", "title": "Optimizing Data Collection for Machine Learning", "abstract": "  Modern deep learning systems require huge data sets to achieve impressive\nperformance, but there is little guidance on how much or what kind of data to\ncollect. Over-collecting data incurs unnecessary present costs, while\nunder-collecting may incur future costs and delay workflows. We propose a new\nparadigm for modeling the data collection workflow as a formal optimal data\ncollection problem that allows designers to specify performance targets,\ncollection costs, a time horizon, and penalties for failing to meet the\ntargets. Additionally, this formulation generalizes to tasks requiring multiple\ndata sources, such as labeled and unlabeled data used in semi-supervised\nlearning. To solve our problem, we develop Learn-Optimize-Collect (LOC), which\nminimizes expected future collection costs. Finally, we numerically compare our\nframework to the conventional baseline of estimating data requirements by\nextrapolating from neural scaling laws. We significantly reduce the risks of\nfailing to meet desired performance targets on several classification,\nsegmentation, and detection tasks, while maintaining low total collection\ncosts.\n", "date": "2022-10-03", "categories": []}
{"id": "2210.00987", "title": "Data Budgeting for Machine Learning", "abstract": "  Data is the fuel powering AI and creates tremendous value for many domains.\nHowever, collecting datasets for AI is a time-consuming, expensive, and\ncomplicated endeavor. For practitioners, data investment remains to be a leap\nof faith in practice. In this work, we study the data budgeting problem and\nformulate it as two sub-problems: predicting (1) what is the saturating\nperformance if given enough data, and (2) how many data points are needed to\nreach near the saturating performance. Different from traditional\ndataset-independent methods like PowerLaw, we proposed a learning method to\nsolve data budgeting problems. To support and systematically evaluate the\nlearning-based method for data budgeting, we curate a large collection of 383\ntabular ML datasets, along with their data vs performance curves. Our empirical\nevaluation shows that it is possible to perform data budgeting given a small\npilot study dataset with as few as $50$ data points.\n", "date": "2022-10-03", "categories": []}
{"id": "2206.07137", "title": "Prioritized Training on Points that are Learnable, Worth Learning, and\n  Not Yet Learnt", "abstract": "  Training on web-scale data can take months. But most computation and time is\nwasted on redundant and noisy points that are already learnt or not learnable.\nTo accelerate training, we introduce Reducible Holdout Loss Selection\n(RHO-LOSS), a simple but principled technique which selects approximately those\npoints for training that most reduce the model's generalization loss. As a\nresult, RHO-LOSS mitigates the weaknesses of existing data selection methods:\ntechniques from the optimization literature typically select 'hard' (e.g. high\nloss) points, but such points are often noisy (not learnable) or less\ntask-relevant. Conversely, curriculum learning prioritizes 'easy' points, but\nsuch points need not be trained on once learned. In contrast, RHO-LOSS selects\npoints that are learnable, worth learning, and not yet learnt. RHO-LOSS trains\nin far fewer steps than prior art, improves accuracy, and speeds up training on\na wide range of datasets, hyperparameters, and architectures (MLPs, CNNs, and\nBERT). On the large web-scraped image dataset Clothing-1M, RHO-LOSS trains in\n18x fewer steps and reaches 2% higher final accuracy than uniform data\nshuffling.\n", "date": "2022-06-14", "categories": []}
{"id": "2209.10015", "title": "Metadata Archaeology: Unearthing Data Subsets by Leveraging Training\n  Dynamics", "abstract": "  Modern machine learning research relies on relatively few carefully curated\ndatasets. Even in these datasets, and typically in `untidy' or raw data,\npractitioners are faced with significant issues of data quality and diversity\nwhich can be prohibitively labor intensive to address. Existing methods for\ndealing with these challenges tend to make strong assumptions about the\nparticular issues at play, and often require a priori knowledge or metadata\nsuch as domain labels. Our work is orthogonal to these methods: we instead\nfocus on providing a unified and efficient framework for Metadata Archaeology\n-- uncovering and inferring metadata of examples in a dataset. We curate\ndifferent subsets of data that might exist in a dataset (e.g. mislabeled,\natypical, or out-of-distribution examples) using simple transformations, and\nleverage differences in learning dynamics between these probe suites to infer\nmetadata of interest. Our method is on par with far more sophisticated\nmitigation methods across different tasks: identifying and correcting\nmislabeled examples, classifying minority-group samples, prioritizing points\nrelevant for training and enabling scalable human auditing of relevant\nexamples.\n", "date": "2022-09-20", "categories": []}
{"id": "2207.08037", "title": "Data Representativeness in Accessibility Datasets: A Meta-Analysis", "abstract": "  As data-driven systems are increasingly deployed at scale, ethical concerns\nhave arisen around unfair and discriminatory outcomes for historically\nmarginalized groups that are underrepresented in training data. In response,\nwork around AI fairness and inclusion has called for datasets that are\nrepresentative of various demographic groups. In this paper, we contribute an\nanalysis of the representativeness of age, gender, and race & ethnicity in\naccessibility datasets - datasets sourced from people with disabilities and\nolder adults - that can potentially play an important role in mitigating bias\nfor inclusive AI-infused applications. We examine the current state of\nrepresentation within datasets sourced by people with disabilities by reviewing\npublicly-available information of 190 datasets, we call these accessibility\ndatasets. We find that accessibility datasets represent diverse ages, but have\ngender and race representation gaps. Additionally, we investigate how the\nsensitive and complex nature of demographic variables makes classification\ndifficult and inconsistent (e.g., gender, race & ethnicity), with the source of\nlabeling often unknown. By reflecting on the current challenges and\nopportunities for representation of disabled data contributors, we hope our\neffort expands the space of possibility for greater inclusion of marginalized\ncommunities in AI-infused systems.\n", "date": "2022-07-16", "categories": []}
{"id": "2209.05122", "title": "Data Augmentation by Selecting Mixed Classes Considering Distance\n  Between Classes", "abstract": "  Data augmentation is an essential technique for improving recognition\naccuracy in object recognition using deep learning. Methods that generate mixed\ndata from multiple data sets, such as mixup, can acquire new diversity that is\nnot included in the training data, and thus contribute significantly to\naccuracy improvement. However, since the data selected for mixing are randomly\nsampled throughout the training process, there are cases where appropriate\nclasses or data are not selected. In this study, we propose a data augmentation\nmethod that calculates the distance between classes based on class\nprobabilities and can select data from suitable classes to be mixed in the\ntraining process. Mixture data is dynamically adjusted according to the\ntraining trend of each class to facilitate training. The proposed method is\napplied in combination with conventional methods for generating mixed data.\nEvaluation experiments show that the proposed method improves recognition\nperformance on general and long-tailed image recognition datasets.\n", "date": "2022-09-12", "categories": []}
{"id": "2209.05297", "title": "DoubleMix: Simple Interpolation-Based Data Augmentation for Text\n  Classification", "abstract": "  This paper proposes a simple yet effective interpolation-based data\naugmentation approach termed DoubleMix, to improve the robustness of models in\ntext classification. DoubleMix first leverages a couple of simple augmentation\noperations to generate several perturbed samples for each training data, and\nthen uses the perturbed data and original data to carry out a two-step\ninterpolation in the hidden space of neural models. Concretely, it first mixes\nup the perturbed data to a synthetic sample and then mixes up the original data\nand the synthetic perturbed data. DoubleMix enhances models' robustness by\nlearning the \"shifted\" features in hidden space. On six text classification\nbenchmark datasets, our approach outperforms several popular text augmentation\nmethods including token-level, sentence-level, and hidden-level data\naugmentation techniques. Also, experiments in low-resource settings show our\napproach consistently improves models' performance when the training data is\nscarce. Extensive ablation studies and case studies confirm that each component\nof our approach contributes to the final performance and show that our approach\nexhibits superior performance on challenging counterexamples. Additionally,\nvisual analysis shows that text features generated by our approach are highly\ninterpretable. Our code for this paper can be found at\nhttps://github.com/declare-lab/DoubleMix.git.\n", "date": "2022-09-12", "categories": []}
{"id": "2206.07746", "title": "Condensing Graphs via One-Step Gradient Matching", "abstract": "  As training deep learning models on large dataset takes a lot of time and\nresources, it is desired to construct a small synthetic dataset with which we\ncan train deep learning models sufficiently. There are recent works that have\nexplored solutions on condensing image datasets through complex bi-level\noptimization. For instance, dataset condensation (DC) matches network gradients\nw.r.t. large-real data and small-synthetic data, where the network weights are\noptimized for multiple steps at each outer iteration. However, existing\napproaches have their inherent limitations: (1) they are not directly\napplicable to graphs where the data is discrete; and (2) the condensation\nprocess is computationally expensive due to the involved nested optimization.\nTo bridge the gap, we investigate efficient dataset condensation tailored for\ngraph datasets where we model the discrete graph structure as a probabilistic\nmodel. We further propose a one-step gradient matching scheme, which performs\ngradient matching for only one single step without training the network\nweights. Our theoretical analysis shows this strategy can generate synthetic\ngraphs that lead to lower classification loss on real graphs. Extensive\nexperiments on various graph datasets demonstrate the effectiveness and\nefficiency of the proposed method. In particular, we are able to reduce the\ndataset size by 90% while approximating up to 98% of the original performance\nand our method is significantly faster than multi-step gradient matching (e.g.\n15x in CIFAR10 for synthesizing 500 graphs). Code is available at\n\\url{https://github.com/amazon-research/DosCond}.\n", "date": "2022-06-15", "categories": []}
{"id": "2209.02609", "title": "Merged-GHCIDR: Geometrical Approach to Reduce Image Data", "abstract": "  The computational resources required to train a model have been increasing\nsince the inception of deep networks. Training neural networks on massive\ndatasets have become a challenging and time-consuming task. So, there arises a\nneed to reduce the dataset without compromising the accuracy. In this paper, we\npresent novel variations of an earlier approach called reduction through\nhomogeneous clustering for reducing dataset size. The proposed methods are\nbased on the idea of partitioning the dataset into homogeneous clusters and\nselecting images that contribute significantly to the accuracy. We propose two\nvariations: Geometrical Homogeneous Clustering for Image Data Reduction\n(GHCIDR) and Merged-GHCIDR upon the baseline algorithm - Reduction through\nHomogeneous Clustering (RHC) to achieve better accuracy and training time. The\nintuition behind GHCIDR involves selecting data points by cluster weights and\ngeometrical distribution of the training set. Merged-GHCIDR involves merging\nclusters having the same labels using complete linkage clustering. We used\nthree deep learning models- Fully Connected Networks (FCN), VGG1, and VGG16. We\nexperimented with the two variants on four datasets- MNIST, CIFAR10,\nFashion-MNIST, and Tiny-Imagenet. Merged-GHCIDR with the same percentage\nreduction as RHC showed an increase of 2.8%, 8.9%, 7.6% and 3.5% accuracy on\nMNIST, Fashion-MNIST, CIFAR10, and Tiny-Imagenet, respectively.\n", "date": "2022-09-06", "categories": []}
{"id": "2208.13079", "title": "Geometrical Homogeneous Clustering for Image Data Reduction", "abstract": "  In this paper, we present novel variations of an earlier approach called\nhomogeneous clustering algorithm for reducing dataset size. The intuition\nbehind the approaches proposed in this paper is to partition the dataset into\nhomogeneous clusters and select some images which contribute significantly to\nthe accuracy. Selected images are the proper subset of the training data and\nthus are human-readable. We propose four variations upon the baseline\nalgorithm-RHC. The intuition behind the first approach, RHCKON, is that the\nboundary points contribute significantly towards the representation of\nclusters. It involves selecting k farthest and one nearest neighbour of the\ncentroid of the clusters. In the following two approaches (KONCW and CWKC), we\nintroduce the concept of cluster weights. They are based on the fact that\nlarger clusters contribute more than smaller sized clusters. The final\nvariation is GHCIDR which selects points based on the geometrical aspect of\ndata distribution. We performed the experiments on two deep learning models-\nFully Connected Networks (FCN) and VGG1. We experimented with the four variants\non three datasets- MNIST, CIFAR10, and Fashion-MNIST. We found that GHCIDR gave\nthe best accuracy of 99.35%, 81.10%, and 91.66% and a training data reduction\nof 87.27%, 32.34%, and 76.80% on MNIST, CIFAR10, and Fashion-MNIST\nrespectively.\n", "date": "2022-08-27", "categories": []}
{"id": "2208.12354", "title": "Fundamentals of Task-Agnostic Data Valuation", "abstract": "  We study valuing the data of a data owner/seller for a data seeker/buyer.\nData valuation is often carried out for a specific task assuming a particular\nutility metric, such as test accuracy on a validation set, that may not exist\nin practice. In this work, we focus on task-agnostic data valuation without any\nvalidation requirements. The data buyer has access to a limited amount of data\n(which could be publicly available) and seeks more data samples from a data\nseller. We formulate the problem as estimating the differences in the\nstatistical properties of the data at the seller with respect to the baseline\ndata available at the buyer. We capture these statistical differences through\nsecond moment by measuring diversity and relevance of the seller's data for the\nbuyer; we estimate these measures through queries to the seller without\nrequesting raw data. We design the queries with the proposed approach so that\nthe seller is blind to the buyer's raw data and has no knowledge to fabricate\nresponses to queries to obtain a desired outcome of the diversity and relevance\ntrade-off.We will show through extensive experiments on real tabular and image\ndatasets that the proposed estimates capture the diversity and relevance of the\nseller's data for the buyer.\n", "date": "2022-08-25", "categories": []}
{"id": "2208.11532", "title": "A novel method for data augmentation: Nine Dot Moving Least Square\n  (ND-MLS)", "abstract": "  Data augmentation greatly increases the amount of data obtained based on\nlabeled data to save on expenses and labor for data collection and labeling. We\npresent a new approach for data augmentation called nine-dot MLS (ND-MLS). This\napproach is proposed based on the idea of image defor-mation. Images are\ndeformed based on control points, which are calculated by ND-MLS. The method\ncan generate over 2000 images for one exist-ing dataset in a short time. To\nverify this data augmentation method, extensive tests were performed covering 3\nmain tasks of computer vision, namely, classification, detection and\nsegmentation. The results show that 1) in classification, 10 images per\ncategory were used for training, and VGGNet can obtain 92% top-1 acc on the\nMNIST dataset of handwritten digits by ND-MLS. In the Omniglot dataset, the\nfew-shot accuracy usu-ally decreases with the increase in character categories.\nHowever, the ND-MLS method has stable performance and obtains 96.5 top-1 acc in\nRes-Net on 100 different handwritten character classification tasks; 2) in\nsegmentation, under the premise of only ten original images, DeepLab obtains\n93.5%, 85%, and 73.3% m_IOU(10) on the bottle, horse, and grass test datasets,\nrespectively, while the cat test dataset obtains 86.7% m_IOU(10) with the\nSegNet model; 3) with only 10 original images from each category in object\ndetection, YOLO v4 obtains 100% and 97.2% bottle and horse detection,\nrespectively, while the cat dataset obtains 93.6% with YOLO v3. In summary,\nND-MLS can perform well on classification, object detec-tion, and semantic\nsegmentation tasks by using only a few data.\n", "date": "2022-08-24", "categories": []}
{"id": "2208.11695", "title": "Bugs in the Data: How ImageNet Misrepresents Biodiversity", "abstract": "  ImageNet-1k is a dataset often used for benchmarking machine learning (ML)\nmodels and evaluating tasks such as image recognition and object detection.\nWild animals make up 27% of ImageNet-1k but, unlike classes representing people\nand objects, these data have not been closely scrutinized. In the current\npaper, we analyze the 13,450 images from 269 classes that represent wild\nanimals in the ImageNet-1k validation set, with the participation of expert\necologists. We find that many of the classes are ill-defined or overlapping,\nand that 12% of the images are incorrectly labeled, with some classes having\n>90% of images incorrect. We also find that both the wildlife-related labels\nand images included in ImageNet-1k present significant geographical and\ncultural biases, as well as ambiguities such as artificial animals, multiple\nspecies in the same image, or the presence of humans. Our findings highlight\nserious issues with the extensive use of this dataset for evaluating ML\nsystems, the use of such algorithms in wildlife-related tasks, and more broadly\nthe ways in which ML datasets are commonly created and curated.\n", "date": "2022-08-24", "categories": []}
{"id": "2208.10494", "title": "Dataset Condensation with Latent Space Knowledge Factorization and\n  Sharing", "abstract": "  In this paper, we introduce a novel approach for systematically solving\ndataset condensation problem in an efficient manner by exploiting the\nregularity in a given dataset. Instead of condensing the dataset directly in\nthe original input space, we assume a generative process of the dataset with a\nset of learnable codes defined in a compact latent space followed by a set of\ntiny decoders which maps them differently to the original input space. By\ncombining different codes and decoders interchangeably, we can dramatically\nincrease the number of synthetic examples with essentially the same parameter\ncount, because the latent space is much lower dimensional and since we can\nassume as many decoders as necessary to capture different styles represented in\nthe dataset with negligible cost. Such knowledge factorization allows efficient\nsharing of information between synthetic examples in a systematic way,\nproviding far better trade-off between compression ratio and quality of the\ngenerated examples. We experimentally show that our method achieves new\nstate-of-the-art records by significant margins on various benchmark datasets\nsuch as SVHN, CIFAR10, CIFAR100, and TinyImageNet.\n", "date": "2022-08-21", "categories": []}
{"id": "2207.08817", "title": "Research Trends and Applications of Data Augmentation Algorithms", "abstract": "  In the Machine Learning research community, there is a consensus regarding\nthe relationship between model complexity and the required amount of data and\ncomputation power. In real world applications, these computational requirements\nare not always available, motivating research on regularization methods. In\naddition, current and past research have shown that simpler classification\nalgorithms can reach state-of-the-art performance on computer vision tasks\ngiven a robust method to artificially augment the training dataset. Because of\nthis, data augmentation techniques became a popular research topic in recent\nyears. However, existing data augmentation methods are generally less\ntransferable than other regularization methods. In this paper we identify the\nmain areas of application of data augmentation algorithms, the types of\nalgorithms used, significant research trends, their progression over time and\nresearch gaps in data augmentation literature. To do this, the related\nliterature was collected through the Scopus database. Its analysis was done\nfollowing network science, text mining and exploratory analysis approaches. We\nexpect readers to understand the potential of data augmentation, as well as\nidentify future research directions and open questions within data augmentation\nresearch.\n", "date": "2022-07-18", "categories": []}
{"id": "2208.00311", "title": "Delving into Effective Gradient Matching for Dataset Condensation", "abstract": "  As deep learning models and datasets rapidly scale up, network training is\nextremely time-consuming and resource-costly. Instead of training on the entire\ndataset, learning with a small synthetic dataset becomes an efficient solution.\nExtensive research has been explored in the direction of dataset condensation,\namong which gradient matching achieves state-of-the-art performance. The\ngradient matching method directly targets the training dynamics by matching the\ngradient when training on the original and synthetic datasets. However, there\nare limited deep investigations into the principle and effectiveness of this\nmethod. In this work, we delve into the gradient matching method from a\ncomprehensive perspective and answer the critical questions of what, how, and\nwhere to match. We propose to match the multi-level gradients to involve both\nintra-class and inter-class gradient information. We demonstrate that the\ndistance function should focus on the angle, considering the magnitude\nsimultaneously to delay the overfitting. An overfitting-aware adaptive learning\nstep strategy is also proposed to trim unnecessary optimization steps for\nalgorithmic efficiency improvement. Ablation and comparison experiments\ndemonstrate that our proposed methodology shows superior accuracy, efficiency,\nand generalization compared to prior work.\n", "date": "2022-07-30", "categories": []}
{"id": "2207.13887", "title": "Adaptive Second Order Coresets for Data-efficient Machine Learning", "abstract": "  Training machine learning models on massive datasets incurs substantial\ncomputational costs. To alleviate such costs, there has been a sustained effort\nto develop data-efficient training methods that can carefully select subsets of\nthe training examples that generalize on par with the full training data.\nHowever, existing methods are limited in providing theoretical guarantees for\nthe quality of the models trained on the extracted subsets, and may perform\npoorly in practice. We propose AdaCore, a method that leverages the geometry of\nthe data to extract subsets of the training examples for efficient machine\nlearning. The key idea behind our method is to dynamically approximate the\ncurvature of the loss function via an exponentially-averaged estimate of the\nHessian to select weighted subsets (coresets) that provide a close\napproximation of the full gradient preconditioned with the Hessian. We prove\nrigorous guarantees for the convergence of various first and second-order\nmethods applied to the subsets chosen by AdaCore. Our extensive experiments\nshow that AdaCore extracts coresets with higher quality compared to baselines\nand speeds up training of convex and non-convex machine learning models, such\nas logistic regression and neural networks, by over 2.9x over the full data and\n4.5x over random subsets.\n", "date": "2022-07-28", "categories": []}
{"id": "2206.14868", "title": "Teach me how to Interpolate a Myriad of Embeddings", "abstract": "  Mixup refers to interpolation-based data augmentation, originally motivated\nas a way to go beyond empirical risk minimization (ERM). Yet, its extensions\nfocus on the definition of interpolation and the space where it takes place,\nwhile the augmentation itself is less studied: For a mini-batch of size $m$,\nmost methods interpolate between $m$ pairs with a single scalar interpolation\nfactor $\\lambda$.\n  In this work, we make progress in this direction by introducing MultiMix,\nwhich interpolates an arbitrary number $n$ of tuples, each of length $m$, with\none vector $\\lambda$ per tuple. On sequence data, we further extend to dense\ninterpolation and loss computation over all spatial positions. Overall, we\nincrease the number of tuples per mini-batch by orders of magnitude at little\nadditional cost. This is possible by interpolating at the very last layer\nbefore the classifier. Finally, to address inconsistencies due to linear target\ninterpolation, we introduce a self-distillation approach to generate and\ninterpolate synthetic targets.\n  We empirically show that our contributions result in significant improvement\nover state-of-the-art mixup methods on four benchmarks. By analyzing the\nembedding space, we observe that the classes are more tightly clustered and\nuniformly spread over the embedding space, thereby explaining the improved\nbehavior.\n", "date": "2022-06-29", "categories": []}
{"id": "2206.15015", "title": "Exploring Temporally Dynamic Data Augmentation for Video Recognition", "abstract": "  Data augmentation has recently emerged as an essential component of modern\ntraining recipes for visual recognition tasks. However, data augmentation for\nvideo recognition has been rarely explored despite its effectiveness. Few\nexisting augmentation recipes for video recognition naively extend the image\naugmentation methods by applying the same operations to the whole video frames.\nOur main idea is that the magnitude of augmentation operations for each frame\nneeds to be changed over time to capture the real-world video's temporal\nvariations. These variations should be generated as diverse as possible using\nfewer additional hyper-parameters during training. Through this motivation, we\npropose a simple yet effective video data augmentation framework, DynaAugment.\nThe magnitude of augmentation operations on each frame is changed by an\neffective mechanism, Fourier Sampling that parameterizes diverse, smooth, and\nrealistic temporal variations. DynaAugment also includes an extended search\nspace suitable for video for automatic data augmentation methods. DynaAugment\nexperimentally demonstrates that there are additional performance rooms to be\nimproved from static augmentations on diverse video models. Specifically, we\nshow the effectiveness of DynaAugment on various video datasets and tasks:\nlarge-scale video recognition (Kinetics-400 and Something-Something-v2),\nsmall-scale video recognition (UCF- 101 and HMDB-51), fine-grained video\nrecognition (Diving-48 and FineGym), video action segmentation on Breakfast,\nvideo action localization on THUMOS'14, and video object detection on MOT17Det.\nDynaAugment also enables video models to learn more generalized representation\nto improve the model robustness on the corrupted videos.\n", "date": "2022-06-30", "categories": []}
{"id": "2206.04374", "title": "Uncovering bias in the PlantVillage dataset", "abstract": "  We report our investigation on the use of the popular PlantVillage dataset\nfor training deep learning based plant disease detection models. We trained a\nmachine learning model using only 8 pixels from the PlantVillage image\nbackgrounds. The model achieved 49.0% accuracy on the held-out test set, well\nabove the random guessing accuracy of 2.6%. This result indicates that the\nPlantVillage dataset contains noise correlated with the labels and deep\nlearning models can easily exploit this bias to make predictions. Possible\napproaches to alleviate this problem are discussed.\n", "date": "2022-06-09", "categories": []}
{"id": "2206.02734", "title": "Global Mixup: Eliminating Ambiguity with Clustering", "abstract": "  Data augmentation with \\textbf{Mixup} has been proven an effective method to\nregularize the current deep neural networks. Mixup generates virtual samples\nand corresponding labels at once through linear interpolation. However, this\none-stage generation paradigm and the use of linear interpolation have the\nfollowing two defects: (1) The label of the generated sample is directly\ncombined from the labels of the original sample pairs without reasonable\njudgment, which makes the labels likely to be ambiguous. (2) linear combination\nsignificantly limits the sampling space for generating samples. To tackle these\nproblems, we propose a novel and effective augmentation method based on global\nclustering relationships named \\textbf{Global Mixup}. Specifically, we\ntransform the previous one-stage augmentation process into two-stage,\ndecoupling the process of generating virtual samples from the labeling. And for\nthe labels of the generated samples, relabeling is performed based on\nclustering by calculating the global relationships of the generated samples. In\naddition, we are no longer limited to linear relationships but generate more\nreliable virtual samples in a larger sampling space. Extensive experiments for\n\\textbf{CNN}, \\textbf{LSTM}, and \\textbf{BERT} on five tasks show that Global\nMixup significantly outperforms previous state-of-the-art baselines. Further\nexperiments also demonstrate the advantage of Global Mixup in low-resource\nscenarios.\n", "date": "2022-06-06", "categories": []}
{"id": "2206.00511", "title": "Differentially Private Shapley Values for Data Evaluation", "abstract": "  The Shapley value has been proposed as a solution to many applications in\nmachine learning, including for equitable valuation of data. Shapley values are\ncomputationally expensive and involve the entire dataset. The query for a\npoint's Shapley value can also compromise the statistical privacy of other data\npoints. We observe that in machine learning problems such as empirical risk\nminimization, and in many learning algorithms (such as those with uniform\nstability), a diminishing returns property holds, where marginal benefit per\ndata point decreases rapidly with data sample size. Based on this property, we\npropose a new stratified approximation method called the Layered Shapley\nAlgorithm. We prove that this method operates on small (O(\\polylog(n))) random\nsamples of data and small sized ($O(\\log n)$) coalitions to achieve the results\nwith guaranteed probabilistic accuracy, and can be modified to incorporate\ndifferential privacy. Experimental results show that the algorithm correctly\nidentifies high-value data points that improve validation accuracy, and that\nthe differentially private evaluations preserve approximate ranking of data.\n", "date": "2022-06-01", "categories": []}
{"id": "2606.21337", "title": "DataClaw0: Agentic Tailoring Multimodal Data from Raw Streams", "abstract": "Massive unstructured multimodal streams suffer from high \"data entropy,\" impeding both efficient human knowledge acquisition and high-quality AI post-training. Existing passive annotation paradigms, heavily reliant on heuristic rules or general VLMs, are costly, monotonous, and fail to unlock the deep procedural logic embedded in raw data. We elevate data processing to a learnable capability, proposing a paradigm shift towards Agentic Data Tailoring, which actively refining and structuring data to align with diverse user and downstream intents. To overcome the data scarcity bottleneck in training such high-order capabilities, we design a two-stage pipeline grounding generative semantic synthesis in deterministic Factual Anchors, yielding a large-scale dataset spanning five core physical and digital domains. Building upon this, $\\text{DataClaw}_0$-9B model synergizes Supervised Fine-Tuning (SFT) with Group Relative Policy Optimization (GRPO), achieving robust alignment with complex refinement and tailoring intents. To systematically quantify this capability, we construct $\\text{DataClaw}_0$-val, the first benchmark dedicated to data refinement. Crucially, we adopt downstream post-training as the ultimate validation touchstone. Evaluations on video generation, real-world VQA, and GUI navigation confirm that $\\text{DataClaw}_0$ delivers high-information-density tailored data, facilitating efficient model adaptation to new tasks under limited training data regimes. Project page: https://czjdsg.github.io/MakeAnyData", "date": "2026-06-19", "categories": []}
{"id": "2606.21705", "title": "Structural Assessment for Understanding and Guiding Dataset Distillation in Discrete Token Space", "abstract": "Dataset distillation (DD) has proven to reduce training cost while preserving accuracy. While promising, the factors that make one distilled dataset more effective than another remain poorly understood. In this work, we investigate this question through the lens of discrete visual tokenizers. Whereas many prior DD efforts emphasize matching global data distributions, we suggest that the effectiveness depends on which semantic concepts are captured and how they are composed. Discrete visual tokenizers provide a finite vocabulary that enables direct statistical analysis of such compositional structure. Through quantitative analysis of token-level statistics, we introduce the structural score to measure the adequacy of token compositions. We observe that distilled datasets with balanced token composition yield higher validation performance. On the other hand, divergence from the original data does not necessarily harm performance. We further show that samples with high structural scores in the discrete token space can effectively guide diffusion-based DD. Our findings highlight the importance of token composition in dataset effectiveness, offering a principled complement to distributional similarity considerations in DD.", "date": "2026-06-19", "categories": []}
{"id": "2606.21916", "title": "Data Pruning: Redundant, Problematic, and Interdependent Samples", "abstract": "The performance of deep learning models is affected by not only data quantity but also data quality. Data pruning is a process by which practitioners can reduce the size of a dataset by only keeping the most important training data points, thereby achieving similar test set performance. We empirically investigate two popular data pruning methods under noisy and noiseless conditions and show that these methods fail in the presence of significant label noise. We highlight that the success of data pruning is distinctly affected by three factors: redundancy in the dataset, the presence of problematic samples, and interdependence between samples. We perform a detailed investigation on commonly used benchmark classification datasets and neural network architectures. We find that our observations are consistent across data distributions and training protocols.", "date": "2026-06-20", "categories": []}
{"id": "2606.22975", "title": "TaLK: Text-attributed Graph Dataset Distillation via Coupling Language Model with Graph-Aware Kernel", "abstract": "Text-attributed graphs (TAGs) are widely used in many real-world domains, and learning on TAGs requires jointly modeling text semantics and graph structure. A standard approach for modeling TAGs is to combine a language model (LM) and a graph neural network (GNN), but joint training is computationally expensive and difficult to scale. Dataset distillation is a promising way to reduce training costs, but existing methods are not well suited to TAGs because they are typically designed for a single modality or still require repeatedly training expensive LM-GNN models on the full dataset during distillation. To address this, we propose TaLK, an effective dataset distillation method for TAGs that couples an LM with a graph-aware neural tangent kernel.This design enables efficient dataset distillation, avoiding repeated joint training on the full dataset while reflecting both textual and structural information for effective TAG learning.Experiments on multiple TAG benchmarks show that TaLK consistently outperforms existing baselines and achieves up to 97% of full-dataset performance with only 1% synthetic data.", "date": "2026-06-22", "categories": []}
{"id": "2606.23204", "title": "Unmasking LAION-5B: Age, Gender, Race, and Emotion Biases in Large-Scale Image Datasets", "abstract": "Large-scale image-text datasets, such as LAION-5B, are foundational to modern AI systems, yet their vast scale and uncurated nature raise significant concerns about demographic and stereotypical biases. This study presents a comprehensive analysis of the demographic composition and representational, stereotypical, and intersectional biases in LAION-2B-en and LAION-2B-multi, the two main components of the LAION-5B dataset. Using state-of-the-art models -- FairFace, DeepFace, and Emo-AffectNet -- we analyze faces detected in the dataset to identify biases across age, gender, race, and expressed emotion. Our findings reveal substantial overrepresentation of young adults (20--39), White individuals, and males, alongside consistent underrepresentation of minority racial groups and middle-aged or older women across both dataset components. We also observe stereotypical associations between demographic attributes and emotions, such as ``Anger'' being predominantly linked to males and ``Happiness'' to females, pointing to systemic imbalances in the data. The consistency of these patterns across two demographic models and both components of LAION-5B demonstrates that these biases are deeply embedded in one of the most widely-used training datasets. Given the scale at which LAION-5B is used to train generative models, these demographic imbalances could shape the behavior and outputs of numerous downstream AI systems.", "date": "2026-06-22", "categories": []}
{"id": "2606.25784", "title": "$S^{2}$-FracMix: Label-Preserving Self-Saliency Mixup Augmentation", "abstract": "Data augmentation is known to improve generalization of deep visual models. Recent methods favor mixup strategies that generate interpolated samples to improve model performance. However, these techniques not only incur significant computational overhead, they also lead to semantic disruption of augmentation data due to cross-sample mixing. We first propose Self-Saliency ($S^2$) Mixup, which constructs challenging yet label-consistent samples by extracting multi-scale salient patches and reinserting them into non-salient regions of the same image. This promotes scale-invariant feature learning while avoiding cross-sample interference. To further enhance model robustness, we introduce FracMix, a mixing scheme that injects self-similarity patterns into salient regions using adaptive ratios. Collectively, our unified framework, $S^{2}$-FracMix, enables simultaneous learning from fractal and non-fractal structures within a single image, yielding a targeted and structurally coherent augmentation strategy. We theoretically analyze the advantage of our technique, and empirically establish its superiority over the existing methods by achieving state-of-the-art performance in extensive evaluation with seven benchmarks across classification (coarse and fine-grained), robustness, calibration, object detection, and transfer learning tasks. Project page is available at \\href{https://fracmix-data-augmentation.github.io/}{fracmix-data-augmentation.github.io}", "date": "2026-06-24", "categories": []}
{"id": "2606.25996", "title": "Autodata: An agentic data scientist to create high quality synthetic data", "abstract": "We introduce Autodata, a general method that enables AI agents to act as data scientists who build high quality training and evaluation data. We show how to train (meta-optimize) such a data scientist agent, so that it learns to create even stronger data. We describe the overall formulation, and a specific practical implementation, Agentic Self-Instruct. We conduct experiments on computer science research tasks, legal reasoning tasks and reasoning with mathematical objects, where we obtain improved results compared to classical synthetic dataset creation methods. Further, meta-optimizing the data scientist agent itself delivers an even larger performance uplift. Agentic data creation provides a way to convert increased inference compute into higher quality model training. Overall, we believe this direction has the potential to change the way we build AI data.", "date": "2026-06-24", "categories": []}
{"id": "2606.28460", "title": "Counterfactual Residual Data Augmentation for Regression", "abstract": "Data-driven modeling in real-world regression tasks often suffers from limited training samples, high collection costs, and noisy observations. Inspired by the impact of data augmentation in vision and language, we propose a novel Counterfactual Residual Data Augmentation (CRDA) technique for tabular regression. Our key insight is that once a regressor has modeled the systematic component of the data, the remaining noise can be viewed as an invariant residual that remains stable under small perturbations of carefully selected features. We exploit this residual invariance to generate new, yet realistic, training samples, effectively expanding the dataset without requiring additional real data. Our method is model-agnostic and readily applicable to various types of regressors. In experiments across datasets from a variety of benchmark repositories, on average, CRDA reduces an MLP Regressor's MSE by 22.9% and an XGBoost Regressor's MSE by 6.4%. When compared to existing state-of-the-art data generators and augmentation techniques, CRDA consistently outperforms in MSE reduction. By adding principled counterfactual variations to the training data, our method offers a simple and efficient remedy for noise-prone, small-sample regression settings.", "date": "2026-06-26", "categories": []}
{"id": "2606.29464", "title": "Rank-Aware Hyperbolic Alignment for Vision-Language Dataset Distillation", "abstract": "Vision-language dataset distillation (VLDD) compresses a large image-text paired dataset into a small set of synthetic pairs that can efficiently train contrastive vision-language models under strict data and compute budgets. Most existing methods match expert trajectories or cross-modal statistics, yet still enforce full-dimensional alignment in a Euclidean embedding space. This is often overly restrictive due to rank-deficient image--text correlation, with shared semantics concentrated in a low-dimensional range and remaining variation spread across a weakly correlated residual subspace. LoRS relaxes alignment at the similarity level by low-rank factorization, but does not explicitly control dominant alignment capacity and structure in the representation space. We thus propose a rank-aware hyperbolic alignment (RAHA) that combines hierarchical geometry with explicit alignment-capacity control. RAHA lifts multimodal representations to hyperbolic space and optimizes distilled pairs with asymmetric objectives that enforce geodesic alignment in the shared range while regularizing the residual subspace to preserve modality-private diversity and improve transfer robustness. Experiments on benchmarks show that RAHA demonstrates competitive cross-modal retrieval and improved transfer indicators under fixed budgets.", "date": "2026-06-28", "categories": []}
{"id": "2606.29837", "title": "Robust Trajectory Distillation: Hybrid Reweighting Meets Teacher-Inspired Targets", "abstract": "Dataset distillation (DD) condenses large corpora into compact, information-rich subsets for efficient training and reuse. However, under noisy supervision, DD risks condensing corrupted associations together with useful signals, degrading robustness. Conventional noisy-label remedies (sample selection, loss weighting, label correction) tightly couple noise estimation with model optimization, often require clean anchors, and can amplify confirmation bias-assumptions that are misaligned with DD's goal of compact, plug-and-play supervision. We therefore propose a trajectory-based DD framework that jointly suppresses noise and preserves transferable knowledge without relabeling or clean subsets. It comprises two complementary components: Selective Guidance Reweighting (SGR), which fuses global forgetting patterns (second-split forgetting) with local neighborhood consistency into a progressive reweighting scheme that prioritizes clean supervision along the teacher trajectory; and Teacher-Inspired Auxiliary Targets (TIAT), which inject auxiliary residual guidance distilled from intermediate teacher dynamics to reinforce informative signals while remaining internally consistent. Together, SGR and TIAT produce distilled datasets with cleaner and richer representations under noisy supervision. The framework is robust, label-preserving, computationally lightweight, and broadly applicable, yielding consistent gains over state-of-the-art DD baselines across symmetric, asymmetric, and real-world noise.", "date": "2026-06-29", "categories": []}
{"id": "2606.30077", "title": "Online Data Selection for Instruction Tuning via Gaussian Processes", "abstract": "With Large Language Model (LLM) pre-training and fine-tuning shifting its focus from data volume to data quality, quality data selection has emerged as a critical research topic. Existing online data selection methods for LLM training are typically \"batch-constrained\", limiting optimization to local utility within random batches. To overcome this, we propose GAIA (Global Adaptive Instruction tuning via GAussian processes), a framework that formulates data valuation as a global estimation process. GAIA employs Gaussian Process regression to model continuous utility manifolds across the semantic space, utilizing an adaptive strategy fusion mechanism to dynamically prioritize high-utility samples. By casting the strategy-posterior update as an instance of the classical fixed-share Hedge framework for tracking the best expert, we inherit a dynamic-regret guarantee that characterizes GAIA's robustness under non-stationary quality scores during training. Empirical evaluations on three datasets demonstrate that GAIA significantly outperforms state-of-the-art baselines like \\greats, establishing our method as a scalable and robust solution for efficient instruction tuning.", "date": "2026-06-29", "categories": []}