Add Batch 784217dd-8d09-474a-8724-fa53a2827331 data

.gitattributes

@@ -7298,3 +7298,67 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 2024/Cross-model[[:space:]]Control_[[:space:]]Improving[[:space:]]Multiple[[:space:]]Large[[:space:]]Language[[:space:]]Models[[:space:]]in[[:space:]]One-time[[:space:]]Training/f6a532be-1478-4cca-9b37-ca1dc0a9d16c_origin.pdf filter=lfs diff=lfs merge=lfs -text
 2024/Cross-video[[:space:]]Identity[[:space:]]Correlating[[:space:]]for[[:space:]]Person[[:space:]]Re-identification[[:space:]]Pre-training/2192f5f6-1f1c-4dba-a393-4f1cd8ace5da_origin.pdf filter=lfs diff=lfs merge=lfs -text
 2024/CryoBench_[[:space:]]Diverse[[:space:]]and[[:space:]]challenging[[:space:]]datasets[[:space:]]for[[:space:]]the[[:space:]]heterogeneity[[:space:]]problem[[:space:]]in[[:space:]]cryo-EM/5c8f8fd2-7cb8-48a0-ad7c-a85ba43575fa_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/CryoGEM_[[:space:]]Physics-Informed[[:space:]]Generative[[:space:]]Cryo-Electron[[:space:]]Microscopy/38602e45-e165-46e7-94f2-31f0a13239b8_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/CryoSPIN_[[:space:]]Improving[[:space:]]Ab-Initio[[:space:]]Cryo-EM[[:space:]]Reconstruction[[:space:]]with[[:space:]]Semi-Amortized[[:space:]]Pose[[:space:]]Inference/967fa365-aaa1-46bd-80f7-948df0680e03_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Cryptographic[[:space:]]Hardness[[:space:]]of[[:space:]]Score[[:space:]]Estimation/9e349f06-6d4c-4865-b127-6f5acd1df8c8_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Ctrl-X_[[:space:]]Controlling[[:space:]]Structure[[:space:]]and[[:space:]]Appearance[[:space:]]for[[:space:]]Text-To-Image[[:space:]]Generation[[:space:]]Without[[:space:]]Guidance/1041078b-a426-42ca-a128-147b63808e7e_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/CuMo_[[:space:]]Scaling[[:space:]]Multimodal[[:space:]]LLM[[:space:]]with[[:space:]]Co-Upcycled[[:space:]]Mixture-of-Experts/8e433d81-a0d1-4e53-9a22-fed61de1e57d_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/CultureLLM_[[:space:]]Incorporating[[:space:]]Cultural[[:space:]]Differences[[:space:]]into[[:space:]]Large[[:space:]]Language[[:space:]]Models/466f5c65-ec31-4ae6-afd0-26e62131d89f_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/CulturePark_[[:space:]]Boosting[[:space:]]Cross-cultural[[:space:]]Understanding[[:space:]]in[[:space:]]Large[[:space:]]Language[[:space:]]Models/e454ed12-d2ac-4594-9ae6-3ab30d0063bb_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Curriculum[[:space:]]Fine-tuning[[:space:]]of[[:space:]]Vision[[:space:]]Foundation[[:space:]]Model[[:space:]]for[[:space:]]Medical[[:space:]]Image[[:space:]]Classification[[:space:]]Under[[:space:]]Label[[:space:]]Noise/19e9ffb6-2b3d-4bfe-8b8b-16ee509b59b2_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Curvature[[:space:]]Clues_[[:space:]]Decoding[[:space:]]Deep[[:space:]]Learning[[:space:]]Privacy[[:space:]]with[[:space:]]Input[[:space:]]Loss[[:space:]]Curvature/a4c48c41-b991-4724-a6b8-4ecebc8e172c_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Customized[[:space:]]Multiple[[:space:]]Clustering[[:space:]]via[[:space:]]Multi-Modal[[:space:]]Subspace[[:space:]]Proxy[[:space:]]Learning/ebddc442-47a6-4fc3-9583-44dae924250c_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Customized[[:space:]]Subgraph[[:space:]]Selection[[:space:]]and[[:space:]]Encoding[[:space:]]for[[:space:]]Drug-drug[[:space:]]Interaction[[:space:]]Prediction/73b527a4-6fa3-44df-839b-ef2883a286f0_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Customizing[[:space:]]Language[[:space:]]Models[[:space:]]with[[:space:]]Instance-wise[[:space:]]LoRA[[:space:]]for[[:space:]]Sequential[[:space:]]Recommendation/fda7252a-f46c-492a-b1a2-6cc98b95132b_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/CycleNet_[[:space:]]Enhancing[[:space:]]Time[[:space:]]Series[[:space:]]Forecasting[[:space:]]through[[:space:]]Modeling[[:space:]]Periodic[[:space:]]Patterns/0ec04ea7-d3ae-4487-9c70-7c79da6a66c9_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/D-CPT[[:space:]]Law_[[:space:]]Domain-specific[[:space:]]Continual[[:space:]]Pre-Training[[:space:]]Scaling[[:space:]]Law[[:space:]]for[[:space:]]Large[[:space:]]Language[[:space:]]Models/dbe3729d-74af-42c5-a44a-a0e89a1c4c9a_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/D-LLM_[[:space:]]A[[:space:]]Token[[:space:]]Adaptive[[:space:]]Computing[[:space:]]Resource[[:space:]]Allocation[[:space:]]Strategy[[:space:]]for[[:space:]]Large[[:space:]]Language[[:space:]]Models/062b213c-a3f7-452d-99bc-c2f58659d609_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/D-MiSo_[[:space:]]Editing[[:space:]]Dynamic[[:space:]]3D[[:space:]]Scenes[[:space:]]using[[:space:]]Multi-Gaussians[[:space:]]Soup/9440fff5-9680-45b5-b6bf-39410db45369_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/D2R2_[[:space:]]Diffusion-based[[:space:]]Representation[[:space:]]with[[:space:]]Random[[:space:]]Distance[[:space:]]Matching[[:space:]]for[[:space:]]Tabular[[:space:]]Few-shot[[:space:]]Learning/c49cdd15-e7f6-42ca-a4df-d9af32962416_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/DA-Ada_[[:space:]]Learning[[:space:]]Domain-Aware[[:space:]]Adapter[[:space:]]for[[:space:]]Domain[[:space:]]Adaptive[[:space:]]Object[[:space:]]Detection/7b4cbeb2-cf95-4d16-86ff-ead0022e4565_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/DACO_[[:space:]]Towards[[:space:]]Application-Driven[[:space:]]and[[:space:]]Comprehensive[[:space:]]Data[[:space:]]Analysis[[:space:]]via[[:space:]]Code[[:space:]]Generation/23290edb-bd1a-4c14-b28d-8f7143113351_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/DAGER_[[:space:]]Exact[[:space:]]Gradient[[:space:]]Inversion[[:space:]]for[[:space:]]Large[[:space:]]Language[[:space:]]Models/da835ef8-87ac-4278-acba-3045bb817294_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/DALD_[[:space:]]Improving[[:space:]]Logits-based[[:space:]]Detector[[:space:]]without[[:space:]]Logits[[:space:]]from[[:space:]]Black-box[[:space:]]LLMs/66ea9ba0-b964-4af8-961a-82e6c7e1028a_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/DAPE_[[:space:]]Data-Adaptive[[:space:]]Positional[[:space:]]Encoding[[:space:]]for[[:space:]]Length[[:space:]]Extrapolation/3ae17c5f-65c8-45da-9c8c-bca73d5d54d1_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/DARG_[[:space:]]Dynamic[[:space:]]Evaluation[[:space:]]of[[:space:]]Large[[:space:]]Language[[:space:]]Models[[:space:]]via[[:space:]]Adaptive[[:space:]]Reasoning[[:space:]]Graph/371a2d4a-c020-4e7d-bae4-95a408ab5aa1_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/DARNet_[[:space:]]Dual[[:space:]]Attention[[:space:]]Refinement[[:space:]]Network[[:space:]]with[[:space:]]Spatiotemporal[[:space:]]Construction[[:space:]]for[[:space:]]Auditory[[:space:]]Attention[[:space:]]Detection/c8bd2b41-90d7-4423-a896-7a688e85e222_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/DART-Eval_[[:space:]]A[[:space:]]Comprehensive[[:space:]]DNA[[:space:]]Language[[:space:]]Model[[:space:]]Evaluation[[:space:]]Benchmark[[:space:]]on[[:space:]]Regulatory[[:space:]]DNA/7ee9bd4c-92ba-445e-9c7a-3fddf39655cf_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/DART-Math_[[:space:]]Difficulty-Aware[[:space:]]Rejection[[:space:]]Tuning[[:space:]]for[[:space:]]Mathematical[[:space:]]Problem-Solving/16eb38e8-f7d3-4989-9836-a2923351c9b2_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/DASH_[[:space:]]Warm-Starting[[:space:]]Neural[[:space:]]Network[[:space:]]Training[[:space:]]in[[:space:]]Stationary[[:space:]]Settings[[:space:]]without[[:space:]]Loss[[:space:]]of[[:space:]]Plasticity/75b1c7e7-50e0-424d-8571-27a8264ba247_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/DAT_[[:space:]]Improving[[:space:]]Adversarial[[:space:]]Robustness[[:space:]]via[[:space:]]Generative[[:space:]]Amplitude[[:space:]]Mix-up[[:space:]]in[[:space:]]Frequency[[:space:]]Domain/4f61c1bf-c1ce-451a-92c9-aac7465e9f9a_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/DC-Gaussian_[[:space:]]Improving[[:space:]]3D[[:space:]]Gaussian[[:space:]]Splatting[[:space:]]for[[:space:]]Reflective[[:space:]]Dash[[:space:]]Cam[[:space:]]Videos/718bebec-005e-47cc-a064-eeba3b3f2725_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/DCDepth_[[:space:]]Progressive[[:space:]]Monocular[[:space:]]Depth[[:space:]]Estimation[[:space:]]in[[:space:]]Discrete[[:space:]]Cosine[[:space:]]Domain/ae8d998a-4db6-4374-9ddb-daf71c2e86b2_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/DDGS-CT_[[:space:]]Direction-Disentangled[[:space:]]Gaussian[[:space:]]Splatting[[:space:]]for[[:space:]]Realistic[[:space:]]Volume[[:space:]]Rendering/57092fb4-667f-474d-b265-c9d2ea76355e_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/DDK_[[:space:]]Distilling[[:space:]]Domain[[:space:]]Knowledge[[:space:]]for[[:space:]]Efficient[[:space:]]Large[[:space:]]Language[[:space:]]Models/0b4cc55a-6c40-4faa-a5b5-5b35c3ed7006_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/DDN_[[:space:]]Dual-domain[[:space:]]Dynamic[[:space:]]Normalization[[:space:]]for[[:space:]]Non-stationary[[:space:]]Time[[:space:]]Series[[:space:]]Forecasting/2aefd221-d345-4ea5-b461-9bb5cc5e4430_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/DDR_[[:space:]]Exploiting[[:space:]]Deep[[:space:]]Degradation[[:space:]]Response[[:space:]]as[[:space:]]Flexible[[:space:]]Image[[:space:]]Descriptor/283c8885-ac93-4dce-9cd1-ba19acac2af0_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/DECO-Bench_[[:space:]]Unified[[:space:]]Benchmark[[:space:]]for[[:space:]]Decoupled[[:space:]]Task-Agnostic[[:space:]]Synthetic[[:space:]]Data[[:space:]]Release/3ac7bb8a-f7a5-4b40-8880-a489f2eaf677_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/DECRL_[[:space:]]A[[:space:]]Deep[[:space:]]Evolutionary[[:space:]]Clustering[[:space:]]Jointed[[:space:]]Temporal[[:space:]]Knowledge[[:space:]]Graph[[:space:]]Representation[[:space:]]Learning[[:space:]]Approach/4ed64841-fd1f-4283-ad1c-9d708a2e3f97_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/DEFT_[[:space:]]Efficient[[:space:]]Fine-tuning[[:space:]]of[[:space:]]Diffusion[[:space:]]Models[[:space:]]by[[:space:]]Learning[[:space:]]the[[:space:]]Generalised[[:space:]]$h$-transform/4459e056-4340-4b2e-83e5-79cd5b824cfe_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/DEL_[[:space:]]Discrete[[:space:]]Element[[:space:]]Learner[[:space:]]for[[:space:]]Learning[[:space:]]3D[[:space:]]Particle[[:space:]]Dynamics[[:space:]]with[[:space:]]Neural[[:space:]]Rendering/37793bd9-e21a-4cfc-bc46-b42b033d4cc0_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/DEPrune_[[:space:]]Depth-wise[[:space:]]Separable[[:space:]]Convolution[[:space:]]Pruning[[:space:]]for[[:space:]]Maximizing[[:space:]]GPU[[:space:]]Parallelism/8c9f69dd-3e3c-4ffb-8799-29ef667de431_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/DETAIL_[[:space:]]Task[[:space:]]DEmonsTration[[:space:]]Attribution[[:space:]]for[[:space:]]Interpretable[[:space:]]In-context[[:space:]]Learning/596d078a-7375-453b-ae0b-3ffbed5f7763_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/DEX_[[:space:]]Data[[:space:]]Channel[[:space:]]Extension[[:space:]]for[[:space:]]Efficient[[:space:]]CNN[[:space:]]Inference[[:space:]]on[[:space:]]Tiny[[:space:]]AI[[:space:]]Accelerators/2abfbbd2-08e6-4826-a9e2-fd1f9005bd4f_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/DF40_[[:space:]]Toward[[:space:]]Next-Generation[[:space:]]Deepfake[[:space:]]Detection/b11035d1-1c1c-410c-aa0a-f6cb51f7b656_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/DFA-GNN_[[:space:]]Forward[[:space:]]Learning[[:space:]]of[[:space:]]Graph[[:space:]]Neural[[:space:]]Networks[[:space:]]by[[:space:]]Direct[[:space:]]Feedback[[:space:]]Alignment/f19b169e-1e2c-45bf-953b-197f8eceac96_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/DG-SLAM_[[:space:]]Robust[[:space:]]Dynamic[[:space:]]Gaussian[[:space:]]Splatting[[:space:]]SLAM[[:space:]]with[[:space:]]Hybrid[[:space:]]Pose[[:space:]]Optimization/c4966176-f704-4670-ac93-8c4f2d39559e_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/DHA_[[:space:]]Learning[[:space:]]Decoupled-Head[[:space:]]Attention[[:space:]]from[[:space:]]Transformer[[:space:]]Checkpoints[[:space:]]via[[:space:]]Adaptive[[:space:]]Heads[[:space:]]Fusion/e2031ad2-7de5-4752-85fb-d2fdf1c96017_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/DI-MaskDINO_[[:space:]]A[[:space:]]Joint[[:space:]]Object[[:space:]]Detection[[:space:]]and[[:space:]]Instance[[:space:]]Segmentation[[:space:]]Model/43a9131a-3274-4d4c-b53f-678cd68b6d34_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/DINTR_[[:space:]]Tracking[[:space:]]via[[:space:]]Diffusion-based[[:space:]]Interpolation/7ac2b831-ef62-4876-8b71-23d43b9646d4_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/DISP-LLM_[[:space:]]Dimension-Independent[[:space:]]Structural[[:space:]]Pruning[[:space:]]for[[:space:]]Large[[:space:]]Language[[:space:]]Models/f4bd1eee-6588-498d-bc64-8c325e04cd1b_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/DMC-VB_[[:space:]]A[[:space:]]Benchmark[[:space:]]for[[:space:]]Representation[[:space:]]Learning[[:space:]]for[[:space:]]Control[[:space:]]with[[:space:]]Visual[[:space:]]Distractors/a212ad91-3aae-495c-8994-43ac073d72ea_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/DMNet_[[:space:]]Self-comparison[[:space:]]Driven[[:space:]]Model[[:space:]]for[[:space:]]Subject-independent[[:space:]]Seizure[[:space:]]Detection/96a57d97-d3ee-4eb9-8d77-89772d441f30_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/DMPlug_[[:space:]]A[[:space:]]Plug-in[[:space:]]Method[[:space:]]for[[:space:]]Solving[[:space:]]Inverse[[:space:]]Problems[[:space:]]with[[:space:]]Diffusion[[:space:]]Models/5816c8d2-e875-4ffe-b834-80847c587bd7_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/DMesh_[[:space:]]A[[:space:]]Differentiable[[:space:]]Mesh[[:space:]]Representation/5a5ad934-deed-4148-adbd-e75d87fb31bd_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/DN-4DGS_[[:space:]]Denoised[[:space:]]Deformable[[:space:]]Network[[:space:]]with[[:space:]]Temporal-Spatial[[:space:]]Aggregation[[:space:]]for[[:space:]]Dynamic[[:space:]]Scene[[:space:]]Rendering/cb75c6ac-a52f-4632-b821-5d8cc4147816_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/DOFEN_[[:space:]]Deep[[:space:]]Oblivious[[:space:]]Forest[[:space:]]ENsemble/96403268-052f-404e-8b49-1dee0347fbb6_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/DOGS_[[:space:]]Distributed-Oriented[[:space:]]Gaussian[[:space:]]Splatting[[:space:]]for[[:space:]]Large-Scale[[:space:]]3D[[:space:]]Reconstruction[[:space:]]Via[[:space:]]Gaussian[[:space:]]Consensus/2f6ad897-5010-47f8-899e-df1b10d944b8_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/DOPPLER_[[:space:]]Differentially[[:space:]]Private[[:space:]]Optimizers[[:space:]]with[[:space:]]Low-pass[[:space:]]Filter[[:space:]]for[[:space:]]Privacy[[:space:]]Noise[[:space:]]Reduction/fdc2e989-0f80-48fe-9502-aed9b4bb6a83_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/DPIC_[[:space:]]Decoupling[[:space:]]Prompt[[:space:]]and[[:space:]]Intrinsic[[:space:]]Characteristics[[:space:]]for[[:space:]]LLM[[:space:]]Generated[[:space:]]Text[[:space:]]Detection/42ff8850-4af1-48f3-82b0-9f3dd6b38cdb_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/DRACO_[[:space:]]A[[:space:]]Denoising-Reconstruction[[:space:]]Autoencoder[[:space:]]for[[:space:]]Cryo-EM/ca4a9ad8-fcad-4fb5-a22d-13c57585c711_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/DRIP_[[:space:]]Unleashing[[:space:]]Diffusion[[:space:]]Priors[[:space:]]for[[:space:]]Joint[[:space:]]Foreground[[:space:]]and[[:space:]]Alpha[[:space:]]Prediction[[:space:]]in[[:space:]]Image[[:space:]]Matting/89e4eb2d-1434-4492-80cc-951ba030a50c_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/DTGB_[[:space:]]A[[:space:]]Comprehensive[[:space:]]Benchmark[[:space:]]for[[:space:]]Dynamic[[:space:]]Text-Attributed[[:space:]]Graphs/9f2f6514-b86d-49ca-a3c6-094b30b94373_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/DU-Shapley_[[:space:]]A[[:space:]]Shapley[[:space:]]Value[[:space:]]Proxy[[:space:]]for[[:space:]]Efficient[[:space:]]Dataset[[:space:]]Valuation/6c241ad9-e559-4bd3-951d-f0299d39d50e_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/DapperFL_[[:space:]]Domain[[:space:]]Adaptive[[:space:]]Federated[[:space:]]Learning[[:space:]]with[[:space:]]Model[[:space:]]Fusion[[:space:]]Pruning[[:space:]]for[[:space:]]Edge[[:space:]]Devices/657735f6-98d8-4d17-befa-177d264ceaf6_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/DarkSAM_[[:space:]]Fooling[[:space:]]Segment[[:space:]]Anything[[:space:]]Model[[:space:]]to[[:space:]]Segment[[:space:]]Nothing/87d550d6-7557-4946-a8ea-b9a2e8e8c546_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Data[[:space:]]Acquisition[[:space:]]via[[:space:]]Experimental[[:space:]]Design[[:space:]]for[[:space:]]Data[[:space:]]Markets/cc6b2d70-05ec-4a1b-9e3c-1a896077c695_origin.pdf filter=lfs diff=lfs merge=lfs -text

2024/CryoGEM_ Physics-Informed Generative Cryo-Electron Microscopy/38602e45-e165-46e7-94f2-31f0a13239b8_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1b9cea923b66e38dd002219f70d04479a6aa44ebe92647446600b5789921ba4b
+size 15840859

2024/CryoGEM_ Physics-Informed Generative Cryo-Electron Microscopy/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:fd17234bbdd8c12a39650e7ef41308bf4aa1ba4ae15bc3c2b7c6d008371297c3
+size 1756554

2024/CryoSPIN_ Improving Ab-Initio Cryo-EM Reconstruction with Semi-Amortized Pose Inference/967fa365-aaa1-46bd-80f7-948df0680e03_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d2e99cbae902412528b65b7e9e6cb4b17eba2c211c9820f43b3c6e266c20b99d
+size 11580795

2024/CryoSPIN_ Improving Ab-Initio Cryo-EM Reconstruction with Semi-Amortized Pose Inference/full.md

@@ -0,0 +1,643 @@
+# CryoSPIN: Improving Ab-Initio Cryo-EM Reconstruction with Semi-Amortized Pose Inference
+
+Shayan Shekarforoush1,2
+
+shayan@cs.toronto.edu
+
+David B. Lindell<sup>1,2</sup>
+
+lindell@cs.toronto.edu
+
+Marcus A. Brubaker<sup>1,2,3,4</sup>
+
+mab@eecs.yorku.ca
+
+David J. Fleet<sup>1, 2, 4</sup>
+
+fleet@cs.toronto.edu
+
+<sup>1</sup>University of Toronto
+
+$^{2}$ Vector Institute
+
+$^{3}$ York University
+
+4Google DeepMind
+
+# Abstract
+
+Cryo-EM is an increasingly popular method for determining the atomic resolution 3D structure of macromolecular complexes (eg, proteins) from noisy 2D images captured by an electron microscope. The computational task is to reconstruct the 3D density of the particle, along with 3D pose of the particle in each 2D image, for which the posterior pose distribution is highly multi-modal. Recent developments in cryo-EM have focused on deep learning for which amortized inference has been used to predict pose. Here, we address key problems with this approach, and propose a new semi-amortized method, cryoSPIN, in which reconstruction begins with amortized inference and then switches to a form of auto-decoding to refine poses locally using stochastic gradient descent. Through evaluation on synthetic datasets, we demonstrate that cryoSPIN is able to handle multi-modal pose distributions during the amortized inference stage, while the later, more flexible stage of direct pose optimization yields faster and more accurate convergence of poses compared to baselines. On experimental data, we show that cryoSPIN outperforms the state-of-the-art cryoAI in speed and reconstruction quality. Project Webpage.
+
+# 1 Introduction
+
+Single particle electron cryo-microscopy (cryo-EM) has gained popularity among structural biologists as a powerful experimental method for determining the 3D structure of macromolecular complexes, such as proteins and viruses, unlocking our understanding of biological function at the scale of the cell. Thanks to pioneering advances in hardware and data processing techniques, cryo-EM has enabled reconstruction of challenging structures at atomic or near-atomic resolution [1].
+
+During a cryo-EM experiment,  $10^{4} - 10^{7}$  particle images, each containing an instance of the target bio-molecule, are acquired using a transmission electron microscope. From that particle stack, the goal is to reconstruct the unknown 3D density map [2]. This ab-initio reconstruction task presents some challenges. First, the 3D pose (orientation and position) of the particle in each image is unknown, and must be inferred along with the 3D structure. Second, low electron exposures are used to limit radiation damage to the particles (e.g., see Fig. 1). But this reduces the signal-to-noise ratio (SNR), obscuring high-resolution image detail and hindering pose and structure estimation. Third, bio-molecules are typically non-rigid and exhibit structural variations within a sample. Hence, for such heterogeneous datasets, it is crucial to account for the conformational variability in order to obtain high-resolution reconstruction [3-6].
+
+![](images/afaab8b03742740a7cf7a2acb9a89006d8d70668f5bf59b42b9e60ac453dcb8e.jpg)
+
+![](images/702d7fbd07feff3b1e2a867e9c27ca7498ba59c5a8896cef8c1bbdedeb75197f.jpg)  
+Figure 1: CryoSPIN consists of two stages: (i) an auto-encoding stage where an image encoder equipped with multiple heads maps the input image to the pose candidate set  $\{\phi_1, \dots, \phi_M\}$ , followed by computing projections by slicing through the volume decoder in Fourier space based on the pose set. These projections are compared with the input image and the one with the minimum error is used. (ii) An auto-decoding stage where pose parameters in axis-angle representation are stored for all images. The same volume decoder is used to obtain projections, and the reconstruction loss is computed for a single projection.
+
+The development of methods in cryo-EM has recently focused on deep learning. CryoDRGN [5] proposed an image encoder-volume decoder architecture to model continuous heterogeneity, but the poses were assumed to be known. CryoDRGNv2 [7] introduced hierarchical pose search, comprising grid search followed by a form of branch-and-bound (BnB) optimization, akin to cryoSPARC [8]. More recently, cryoPoseNet [9] and cryoAI [10] introduced amortized pose inference, to avoid the expense of orientation matching for each particle image. They employ a convolutional neural network (CNN) to map an input image to a pose estimate. Nevertheless, although efficient, amortized inference may not accurately represent multi-modal posterior pose distributions, which occur frequently early in reconstruction before the structure is estimated well. As such, compared to previous methods like cryoSPARC [8] and cryoDRGNv2 [7] that consider many pose candidates, amortized methods sometimes fail to identify the correct mode. Moreover, amortization confines the pose search to a global parametric function (the encoder) which can lead to slow convergence.
+
+Here, we introduce a new approach, called cryoSPIN, to ab-initio homogeneous reconstruction that is able to handle multi-modal pose distributions with a tailored encoder, and accelerates pose optimization with semi-amortization [11]. First, we perform amortized pose inference through a shared multi-choice encoder that maps each input image to multiple pose estimates (Fig. 1, top). In contrast to cryoPoseNet [9] and cryoAI [10], which use pose encoders that produce one or two estimates, we attach multiple pose predictor heads to a shared CNN feature extractor to predict multiple plausible poses for each image. By design, our multi-output encoder is able to account for pose uncertainty and encourage exploration of pose space during the initial stages of reconstruction. For the 3D decoder, unlike cryoDRGN and cryoAI that use MLPs with computationally expensive feed-forward implicit networks, we adopt an explicit parameterization to further accelerate the reconstruction. To train the encoder-decoder architecture, we apply a "winner-takes-all" loss in which the 3D decoder is queried to obtain a 3D-to-2D projection for each predicted pose. Inspired by the loss function in Multi-Choice Learning [12-14], we select the projection with the lowest image reconstruction error to determine the loss.
+
+During the course of optimization, as higher resolution details are resolved, the pose posterior becomes predominantly uni-modal. Thus, the pose search can be narrowed down to a neighborhood containing the most likely mode. In this stage, we propose to transition from auto-encoding to auto-decoding (Fig. 1, bottom). In particular, for each image, we choose the pose with the lowest reconstruction loss, iteratively refine it with stochastic gradient descent (SGD), and continue pose optimization
+
+and reconstruction in an alternating fashion. This direct, per-image optimization procedure achieves faster convergence to more accurate poses than amortized inference alone, which relies on potentially sub-optimal predictions of the encoder as a parametric function.
+
+We evaluate cryoSPIN in comparison to state-of-the-art methods, cryoSPARC [8], cryoAI [10], and cryoDRGN [7] on both synthetic and real experimental datasets. Using synthetic data, we find that semi-amortized inference noticeably accelerates the convergence of poses, and yields 3D reconstructions at similar or higher resolutions than baseline methods. We also show that our multi-choice encoder empirically accounts for multiple modes in the pose posterior. Finally, we apply cryoSPIN to an experimental dataset and obtain reconstructions with higher resolution than those from cryoAI, while matching the resolution of cryoSPARC reconstructions.
+
+To summarize, we make following contributions.
+
+- Building upon cryoAI, we develop a new encoder based on a multi-head architecture to return multiple plausible candidates to further mitigate pose ambiguity.  
+- We train the encoder coupled with an explicit volume decoder using a winner-take-all loss, which penalizes the pose hypothesis with least reconstruction error.  
+- Beyond the architecture and objective, we introduce semi-amortized pose inference, which begins with feed-forward amortized pose inference, followed by direct, per-particle pose optimization.  
+- In ab-initio reconstruction, our semi-amortized method, cryoSPIN, is faster than amortized method of cryoAI and achieves higher-resolution on synthetic and experimental datasets.
+
+# 2 Related work
+
+Cryo-EM reconstruction. Methods for cryo-EM reconstruction can be categorized as either homogeneous or heterogeneous. Homogeneous techniques [8-10] assume a rigid structure while heterogeneous ones [5, 7, 6, 15] allow for conformational variation. We focus on homogeneous reconstruction, but our optimization framework could be extended to heterogeneous data as well.
+
+Early reconstruction techniques rely on common-lines [16-18] or projection mapping [19, 20] to select optimal poses. Other works [21, 22] frame the reconstruction problem in the context of maximum a posteriori (MAP) estimation, and jointly reconstruct pose and structure via expectation maximization (EM). We compare our approach to cryoSPARC [8], a state-of-the-art method that uses stochastic gradient descent and a branch-and-bound search [23] for ab initio reconstruction and pose estimation. Like these methods, our auto-decoding stage directly optimizes pose of every image.
+
+More recently, amortized inference techniques have been proposed for pose estimation [24, 9, 10, 15]. These techniques avoid explicit per-particle pose optimization; instead, they train an auto-encoder or variational one [25] to associate each particle image with a predicted pose [9]. One challenge is that the auto-encoders can become stuck in local optima during training [9]. To address this issue, cryoAI [10] produces two pose estimates per image coupled with a symmetrized loss function that penalizes the best one. We build on this concept by adopting a multi-head neural architecture as the encoder to output multiple plausible pose candidates and avoid local optima.
+
+Multi-choice learning (MCL). Inspired by scenarios where a set of hypotheses needs to be generated to account for uncertainty in the prediction task, MCL [12] was introduced in a supervised setup to learn multiple structured-outputs with SSVMs [26]. Their motivating question was: can we learn to produce a set of plausible hypotheses? To address this, they define an "oracle" loss in which only the most accurate output pays the penalty. This loss is minimized even if there is only a single accurate prediction in the set. The early follow-up work [27, 13] uses the same loss to learn a deep CNN ensemble composed of  $M$  heads with a shared backbone network. Importantly, they show that the ensemble-mean loss hurts prediction diversity across different heads, while training with the "oracle" loss yields specialized heads. Variations have since been proposed to mitigate hypothesis collapse or overconfidence issues in MCL by modifying the loss or applying learnable probabilistic scoring schemes [28-30, 14]. MCL has been used to mitigate the ambiguity in several tasks including image segmentation [31], optical-flow estimation [32], trajectory forecasting [33], human pose and shape estimation [14, 34]. In our work, we use the "oracle" loss in context of auto-encoder which supervises the pose encoder indirectly through projections provided by the decoder.
+
+# 3 Problem Definition
+
+Image formation in cryo-EM is well-approximated using the weak-phase object model [35]. Under this model, the structure of interest is an unknown 3D density map,  $V: \mathbb{R}^3 \to \mathbb{R}_{\geq 0}$ , represented in some canonical coordinate frame. Cryo-EM images,  $\{I_i\}_{i=1}^N$ , are approximated as orthographic projections of the 3D map under some unknown coordinate transformation. We denote the unknown orientation by a rotation  $R_i \in SO(3)$  and the unknown position in terms of an in-plane translation  $t_i = (t_x, t_y) \in \mathbb{R}^2$ . Formally,
+
+$$
+I _ {i} (x, y) = \left[ g _ {i} \star \left(S _ {t _ {i}} \mathcal {P} _ {R _ {i}} V\right) \right] (x, y) + n (x, y), \tag {1}
+$$
+
+where  $\mathcal{P}_{R_i}(\cdot)$  is the linear operator computing the integral along the optical axis,  $z$ , over the input density map rotated by  $R_{i}$ , and  $S$  is the shift operator. The projection is convolved with the image-specific point-spread function (PSF),  $g_{i}$ , and corrupted by additive noise  $n$ . It is common to assume that  $n$  follows a zero-mean white (or colored) Gaussian distribution.
+
+By the Fourier slice theorem [36], the Fourier transform of a projection is equal to a central slice through the density map's 3D Fourier spectrum. That is,
+
+$$
+\hat {I} _ {i} \left(\omega_ {x}, \omega_ {y}\right) = \hat {g} _ {i} \hat {S} _ {t _ {i}} \left(\hat {\mathcal {P}} _ {R _ {i}} \hat {V}\right) \left(\omega_ {x}, \omega_ {y}\right) + \hat {n} \left(\omega_ {x}, \omega_ {y}\right), \tag {2}
+$$
+
+where  $\hat{I}$  and  $\hat{V}$  denote the 2D and 3D Fourier transforms of the image and the density map. The slice perpendicular to the projection is computed by  $(\hat{\mathcal{P}}_{R_i}\hat{V})$ . The translation by  $S_{t_i}$  becomes a phase shift operator  $\hat{S}_{t_i}$ , and convolution with  $g_{i}$  is equivalent to element-wise multiplication with,  $\hat{g}_i$ , the contrast transfer function (CTF). The noise  $\hat{n}$  remains zero-mean Gaussian.
+
+Under this model, given the structure  $\hat{V}$ , the negative log-likelihood of observing image  $\hat{I}_i$  with noise variance  $\sigma^2$  and pose  $(R_i,t_i)$  is
+
+$$
+\mathcal {L} = - \frac {1}{2 \sigma^ {2}} \sum_ {\omega_ {x}, \omega_ {y}} \left[ \hat {g} _ {i} \hat {S} _ {t _ {i}} \left(\hat {\mathcal {P}} _ {R _ {i}} \hat {V}\right) \left(\omega_ {x}, \omega_ {y}\right) - \hat {I} _ {i} \left(\omega_ {x}, \omega_ {y}\right) \right] ^ {2}. \tag {3}
+$$
+
+Ab-initio reconstruction methods [8-10, 7] solve jointly for the unknown structure  $\hat{V}$  and per-particle poses  $(R_{i},t_{i})$ . They often follow an Expectation-Maximization (EM) [37, 21] procedure in which the E-step aligns images with the structure yielding pose estimates  $(R_{i},t_{i})$ , and then in the M-step the volume  $\hat{V}$  is updated by minimizing the negative log-likelihood in Eq. 3. Since errors in pose estimates lead to blurry reconstructions, accurate pose estimates are crucial to finding high-resolution structures. As discussed above, poses are either optimized through search and projection matching [8, 7] or estimated by an encoder network [10, 9, 15].
+
+# 4 Methodology
+
+We introduce cryoSPIN, a two-staged approach to ab-initio cryo-EM reconstruction, combining auto-encoding and auto-decoding. We start with amortized inference (Fig. 1, auto-encoding) where a shared encoder equipped with multiple heads provide a set of pose guesses. Multiple guesses enable multimodal inference in the initial stages of reconstruction when the 3D structure is poorly resolved. As reconstruction improves, the pose posterior becomes less uncertain, at which point we switch to direct pose optimization (Fig. 1, auto-decoding). The former enables handling the pose uncertainty early in reconstruction while the latter circumvents sub-optimality of encoder network leading to arguably more accurate poses and faster convergence. We couple our pose estimation module with an explicit volumetric decoder representing the 3D structure in Hartley space [5]. Our explicit model enables faster evaluation of projections compared to multiple passes through implicit neural representations [38-40]. The following sections discuss these components in detail.
+
+# 4.1 Multi-choice encoder
+
+With randomly initialized or a low-resolution reconstruction, the pose posterior contains multiple modes. Also, due to high levels of noise in images and near-symmetries in biological structures, pose estimation inherently involves high uncertainty. As a result, there exist several plausible poses for each image, and naive optimization or search is prone to local minima.
+
+To account for uncertainty in pose estimation, we build upon cryoAI [10] and extend the encoder to return multiple candidate poses  $\phi$ . Formally, given the image  $I_{i} \in \mathbb{R}^{H \times W}$ ,  $M$  pairs of rotations and translations are computed as,
+
+$$
+\phi_ {i, j} = \left(R _ {i, j}, t _ {i, j}\right) = H _ {\theta_ {j}} \left(F _ {i}\right), \quad 1 \leq j \leq M \tag {4}
+$$
+
+where  $R_{i,j} \in SO(3)$  and  $t_{i,j} \in \mathbb{R}^2$ . The image-specific intermediate features,  $F_i = \mathrm{VGG}(I_i) \in \mathbb{R}^{C \times H \times W}$  are extracted by a shared convolutional backbone based on the VGG16 architecture [41], which are then supplied to  $M$  separate pose predictor heads,  $H_{\theta_j}, 1 \leq j \leq M$ , yielding poses  $\phi_{i,j}$ . The heads are composed of fully-connected layers with distinct learnable weights, while the backbone is shared across all heads. Using multiple heads facilitates pose exploration and reduces uncertainty in the pose estimation process. Moreover, as  $\theta_j$  are randomly initialized, different heads are free to specialize on subsets of pose space so that they collectively produce a set of likely poses  $\{\phi_{i,1}, \dots, \phi_{i,M}\}$ .
+
+Inspired by multi-choice learning [12-14], we optimize the encoder-decoder using a "winner-take-all" loss. For each image  $I_{i}$  and predicted pose  $\phi_{i,j}$ , the negative log-likelihood,  $\mathcal{L}_{i,j}$  in (Eq. 3), is computed, and the minimum is selected as the final loss for the corresponding image, i.e.,
+
+$$
+\mathcal {L} _ {i} = \min  _ {j} \mathcal {L} _ {i, j}. \tag {5}
+$$
+
+Minimizing this loss requires only one of the predicted poses to be accurate. Interestingly, we find that this approach produces heads that specialize to somewhat disjoint regions of pose space (see Supp. C), consistent with previous work showing that it improves diversity in prediction tasks [13, 27]. This 'divide-and-conquer' approach facilitates pose space exploration, which is crucial in order to avoid local minima during the early stages of reconstruction where uncertainty in the 3D structure is significant. CryoAI [10] can be viewed as a special case of this formulation; it assigns two poses to each image as a consequence of input augmentation, and selects the best one with a symmetrized loss. In contrast, our approach augments the output of the encoder network with multiple heads, each providing a pose estimate.
+
+# 4.2 Switching from auto-encoding to auto-decoding
+
+As optimization progresses and fine-grained details of the 3D structure are resolved, the posterior pose variance tends to decrease, with the posterior approaching a unimodal distribution. At this point, the gap between the amortized and variational posterior distributions is mainly determined by the error in the pose estimate (predicted mean), prioritizing accuracy over exploration. However, a feed-forward network, as a globally parameterized function of the input image, may be limited in the accuracy of its predictions, and hence the approximations inherent in amortized inference become significant sources of error in the refinement of the 3D structure. In prior work [11, 42-44], a related issue, called the amortization gap, is characterized by the KL-divergence between the true and predicted (amortized) variational posteriors.
+
+To address this issue, we adopt a semi-amortized inference scheme [11] comprising two stages. First, the encoder predicts a set of pose candidates using a multi-head architecture. In the second stage, rather than amortized inference, pose parameters are directly optimized for each image using stochastic gradient descent. To initialize poses for the  $i$ -th image, we choose the one with the lowest reconstruction loss from the set of candidates  $\{\phi_{i,1}, \ldots, \phi_{i,M}\}$ , namely  $\phi_i^* = \phi_{i,s}$  such that
+
+$$
+s = \arg \min  _ {j} \mathcal {L} _ {i, j}. \tag {6}
+$$
+
+Subsequently, the pose and structure are optimized by coordinate descent using the negative log-likelihood (Eq. 3) as the objective function. See Supp. A for more details on rotation optimization.
+
+# 4.3 Explicit representation as volume decoder
+
+Recent works [10, 5] use coordinate networks [39, 40, 38] to implicitly model the Fourier representation of the 3D structure. Instead, we couple the pose estimation module with an explicit parameterization (3D voxel array) of the structure in the Fourier domain. The explicit representation is less computationally expensive than an MLP to evaluate and update. Also, this choice is motivated by the fact the implicit decoder needs to be queried multiple times for each image with the multi-head encoder.
+
+We parameterize the volume using the Hartley representation [5]. The Fourier and Hartley transforms, respectively denoted as  $F(\omega)$  and  $H(\omega)$ , are related as
+
+$$
+H (\omega) = \mathcal {R} [ F (\omega) ] - \mathcal {I} [ F (\omega) ], \tag {7}
+$$
+
+where  $\omega$  denotes the frequency coordinate and  $\mathcal{R}$  and  $\mathcal{I}$  are the real and imaginary part, respectively. The Hartley representation is real-valued, and so more memory efficient to use than storing complex-valued Fourier coefficients. To account for high dynamic range of the Hartley coefficients, we assume the Hartley field is decomposed into mantissa,  $m(\omega)$  and exponent  $e(\omega)$  fields [10] as,
+
+$$
+H (\omega) = m (\omega) \times \exp (e (\omega)). \tag {8}
+$$
+
+This decomposition restricts the range of values for  $m(\omega)$  and  $e(\omega)$  and makes the reconstruction less sensitive to the initialization of the field.
+
+# 5 Experiments
+
+In what follows, we consider both synthetic and real datasets to empirically compare our semi-amortized ab-initio reconstruction method with state-of-the-art methods, cryoAI [10] and cryoDRGN [7] which use amortized inference for reconstruction, and cryoSPARC [8] which performs pose search separately for each particle image.
+
+Synthetic data. We generate synthetic datasets by simulating the image formation process formalized in Sec. 3 using atomic models deposited in the Protein Data Bank (PDB). We compute ground-truth density maps of size  $128^3$  for a heat shock protein (HSP) [45]  $(1.5\AA)$ , pre-catalytic spliceosome [46]  $(4.33\AA)$ , and SARS-CoV-2 spike protein [47]  $(2.13\AA)$ . Then,  $N = 50,000$  projections of size  $L = 128$  are generated by randomly rotating and projecting the density map along the canonical z-axis. To simplify the analysis and interpretation of results (e.g. Figs. 4, 5), we consider synthetic particles are centered, and therefore omit the estimation of 2D translation and allowing us to visualize the pose posterior more easily. Finally, a random CTF is applied in the Fourier space and Gaussian noise is added yielding  $\mathrm{SNR} = 0.1$ .
+
+Experimental data. As a widely adopted experimental benchmark, we use the 80S ribosome dataset (EMPIAR-10028 [48]), containing 105,247 particle images of linear box size  $L = 360$  with pixel size  $1.34\AA$ . Following cryoAI and cryoDRGN, we downsample the images to  $L = 128$  (3.76 Å) using cryoSPARC software [8], and randomly split the data into two halves and run the reconstruction methods independently on each (to enable Fourier Shell Correlation as a quantitative measure of the resolution of the 3D reconstruction). Unlike synthetic data, particles are not well-centered so it is required to estimate translation parameters as well as rotations. This is achieved by augmenting each head to return translation parameters as well.
+
+Implementation details. During auto-encoding, we use encoders with  $M = 7$  and  $M = 15$  heads for reconstruction on synthetic and real datasets, respectively, with Adam [49] to optimize encoder and decoder with learning rates 0.0001 and 0.05. Once switched to direct optimization (after 7 epochs for synthetic and 15 epochs for real data), we reduce the decoder learning rate to 0.02 and allocate a new optimizer for pose parameters with learning rate 0.05. To be consistent with cryoAI, we use a batch size of 64 and train for the same number of epochs (20 for synthetic, 30 for real data). We use the public cryoAI codebase, run cryoSPARC v4.4.0 [8] with default settings, and run cryoDRGN homogenous ab-initio job. Methods are implemented in Pytorch [50]. Experiments are run on a single NVIDIA A40 GPU.
+
+# 5.1 Results
+
+We first qualitatively compare final reconstructions of cryoSPIN with cryoAI [10], cryoDRGN [7], and cryoSPARC [8] on synthetic and real datasets (Fig. 2, left). We observe that cryoAI gets stuck in local minima when evaluated on experimental data (eg, 80S ribosome). In particular, we found that in the presence of unknown planar shift, cryoAI fails to estimate effective translation parameters, but after centering all particles via pre-processing, we could reproduce the cryoAI results. Hence, the provided results for cryoAI in Fig. 2 are obtained after this pre-processing step, whereas our method and others are given off-centered experimental images. Both cryoSPIN and cryoSPARC capture high-frequency details of the 3D structure on all datasets, whereas reconstructions by amortized
+
+![](images/c7adabbe0da0dbe4fc718106a107ec4b3b00a7106ddfb20187915281796f4214.jpg)  
+Figure 2: Qualitative and quantitative comparison of reconstructions obtained by our proposed semi-amortized method, cryoSPIN, with cryoAI [10], cryoDRGN [7] and cryoSPARC [8] We note that cryoAI often becomes stuck in local minima when particles are not centered, so for cryoAI we translate the input images to correct for the spatial offset. (Left) Final 3D reconstructions on three synthetic datasets and one experimental dataset (EMPIAR-10028) are depicted using ChimeraX [51]. (Right) FSC curves are visualized for quantitative comparison. The red dashed lines show the standard threshold levels of 0.5 and 0.143 to report the resolution (in Angstrom) for synthetic and real data, respectively. CryoSPIN achieves higher resolution on the Spliceosome and HSP datasets, and it is competitive with the state of the art on the Spike and EMPIAR-10028 datasets.
+
+methods, cryoAI and cryoDRGN, are inferior on the HSP and 80S datasets. In particular, on HSP, cryoAI gets stuck in local minima as it fails to handle high uncertainty in poses caused by symmetries in this structure.
+
+For quantitative comparison, we visualize the Fourier Shell Correlation (FSC) [52] between the reconstruction and the ground truth (Fig. 2, right). FSC is the gold-standard metric measuring the normalised cross-correlation coefficient between two 3D Fourier volumes along shells of increasing radius. Higher FSC implies more accurate reconstruction. On the Spliceosome, HSP and 80S experimental datasets, our reconstruction outperforms amortized methods of cryoAI and cryoDRGN. Also, our method outperforms cryoSPARC on Spliceosome, HSP and 80S, while achieving competitive FSC on Spike. We use the standard 0.5 and 0.143 cutoff thresholds to report the resolution for synthetic and real data, respectively. CryoSPIN achieves higher or competitive resolution compared to others on all datasets.
+
+![](images/bc9e6d962a3c32cfae8236398c122d99e966baed4ec596e7fc9128b2ae232ea8.jpg)  
+Figure 3: 3D resolution as a function of log time for different methods. These plots show the semi-amortized method is significantly faster than cryoAI.
+
+![](images/32c5c0b1f354e649f95098df22dc10fc46af150b96dc5c8619ebc724847abd9a.jpg)
+
+![](images/4adb6bbc1d0639d958ebaea10006007076e0a5809f0e801f040d83a498a3b5e1.jpg)
+
+![](images/34e5d9a2fea123a8f021a3872a8fb10323677668d21a0bb13072176a669c50a9.jpg)
+
+![](images/8fb34d0ba09d5cd54899e509abf13b4bc3531f7d5415434b3813df7dab91d8b1.jpg)
+
+![](images/dbdd933f149b758599da14fabb80f1962862c295d5e95311313ac85c1dde2c2f.jpg)
+
+![](images/66bf9400beecadc1d82bd3f764483179545b93ac7cf06746f5dac91e0ee8461b.jpg)
+
+![](images/e7b27ada05181acbe2291cadd4b4a0ae29a8be089c63d434a262142e4f9bdfa8.jpg)  
+Figure 4: Quantitative and qualitative comparison of fully- vs. semi-amortized methods in pose optimization for Spike (top) and Spliceosome (bottom) datasets. (Left) Mean geodesic distance between the predicted pose and ground-truth is visualized at different epochs. By switching from amortized inference to direct optimization, semi-amortized method (blue) enjoys accelerated pose convergence compared to amortized inference (red). (Right) To visualize pose inference, images depict the approximate log posterior for three particles (marginalized over in-plane rotations) as view-direction distribution over a HEALPix [53] uniform grid on a unit sphere  $S^2$ . After Gnomonic projection to 2D, we show the neighborhood of the mode of interest. Black dots depict starting points. Blue and red dots are pose estimates from fully- and semi-amortized methods.
+
+![](images/cf196aa3cd2294cd867d6fecace0c6ece0eabacd9135e04e5b7e67b9be4bda28.jpg)
+
+![](images/962dbfd29bb1619dc88dcffa71c6d6f388ddf2c7d6ee2a06da6fcc3204974fed.jpg)
+
+![](images/ab1fc5b18428a122fd52b671da241d81438efb8b86020d35e39598e8663ba3a9.jpg)
+
+In Fig. 3, we visualize resolution-time plots, showing that cryoSPIN gets to high-resolution reconstructions significantly faster than cryoAI. In fact, semi-amortization in cryoSPIN accelerates the improvement in the resolution and with our explicit structure decoder, we circumvent expensive MLP evaluations and train faster without any degradation to the final reconstruction quality as shown in Fig. 2. In Supp. B, through a detailed comparison with cryoAI, we show cryoSPIN is  $\sim 6x$  faster and uses  $\sim 5x$  less memory compared to cryoAI.
+
+Table 1: Mean/median errors of estimated rotations in units of degrees evaluated on synthetic datasets.  
+
+<table><tr><td>Method</td><td>HSP</td><td>Spike</td><td>Spliceosome</td></tr><tr><td>CryoSPARC [8]</td><td>6.23 / 1.05</td><td>1.61 / 1.51</td><td>1.41 / 1.36</td></tr><tr><td>CryoAI [10]</td><td>45.83 / 61.86</td><td>2.52 / 2.29</td><td>2.85 / 2.61</td></tr><tr><td>CryoSPIN (Ours)</td><td>3.27 / 0.97</td><td>1.54 / 0.90</td><td>0.68 / 0.61</td></tr></table>
+
+Finally, we report the mean and median errors in estimated poses on synthetic datasets in Table 1, showing that our method outperforms cryoSPARC and cryoAI. High errors by cryoAI on HSP dataset shows that it gets stuck in local minima and fails to accurately estimate poses.
+
+![](images/301a012b0e48e00a48975017894249617cb15a3429f090c912cea6583ae35e69.jpg)
+
+![](images/6a81e42c0e33be2e1efe164c447ed9af9eb1877ff92e368f8d53a0e304c794d5.jpg)
+
+![](images/e73cdd1acd2d2ec3dbcb8cda676526bd26cbb7da351f14bf7799a0f93cd979f0.jpg)
+
+![](images/51e0fa0008008a154b21ae85546edf72216bd5bb15c8fdcd44208866dbf4579e.jpg)
+
+![](images/a1ccdb2dd0cb3c2b1b2701f6272c1d9297935ba217a819ad64692426b3f568d1.jpg)
+
+![](images/d7b86ab7a68d03af337b9c1ecb7e7a7055d95fb41f225dc983fe4ef099b55bd0.jpg)
+
+![](images/0fd68bac095c86c69993820fbc0ffc360ab33975da92bf41a4e3da38bfc429ba.jpg)
+
+![](images/72cc34ff28f064d57941be247452f380d3cf5213343d0c4015558d44832fcd56.jpg)
+
+![](images/5d17314420c71608422a6ec3cc5101185f571f3259edd7c7b289424bd2334ebf.jpg)  
+Figure 5: Comparing performance of our multi-head pose encoder ( $M = 4$ ) with cryoAI pose encoder on the challenging HSP [45] dataset. (Left) The approximate log posterior of view direction is visualized on the unit sphere with highlighted areas showing modes of the distribution. CryoAI and multi-head encoders provide two and four pose estimates, respectively, which are marked with colored dots on the sphere (the order of poses is arbitrary). Below the sphere, the corresponding projections are illustrated. CryoAI fails to find the correct mode while our method is able to cover multiple modes. (Top, right) With our multi-head encoder, the reconstruction converges to a much higher resolution compared to cryoAI. (Bottom, right) Percentage of images assigned to each head is visualized as a bar plot confirming that all heads participate in pose estimation.
+
+![](images/0ca1b5127609c541de89cb0d3697588be489df75b386c4229774020c18f8241b.jpg)
+
+# 5.2 Semi-Amortized vs. Fully-Amortized
+
+To showcase the benefit of the auto-decoding stage, we compare our semi-amortized method with a fully-amortized baseline on spike and spliceosome datasets. Starting with auto-encoding, we branch the running reconstruction into two after 7 epochs: the first continues using the encoder while the second switches to direct pose optimization. We plot the average pose error with respect to the ground-truth at different epochs in Fig. 4 (left). When using the multi-head encoder, the pose with the least reconstruction error is selected among the candidates. Interestingly, as our method switches to direct pose optimization, the error in pose drops quickly, whereas the fully-amortized baseline exhibits slow convergence. This clearly shows the superiority of auto-decoding compared to auto-encoding during the later stages of optimization.
+
+Next, we analyze the evolution of pose estimates over the optimization landscape depicted as a heat map (Fig. 4, right). Qualitatively, poses obtained by direct optimization (blue dots) show stable convergence to the optimal point (highlighted area) while those inferred by the encoder (red dots) frequently oscillate. Indeed, the encoder in amortized inference is a globally parameterized function which might be too restrictive, yielding sub-optimal pose parameters. Therefore, poses inferred in an amortized fashion might fail to consistently converge to the optimal point. On other hand, direct optimization is intuitively more flexible as it is performed separately and locally for each image, exhibiting more stable convergence. See Supp. D for more examples.
+
+# 5.3 Multi-Modal Pose Posterior
+
+Lastly, we examine how well cryoSPIN multi-head encoder performs vs. cryoAI encoder in terms of handling the uncertainty in the pose on the challenging dataset of HSP [45]. To simplify the visualization, we run our method with  $M = 4$  heads in this experiment. We first inspect the behavior of encoders for two example cryo-EM images (Fig. 5, left, see Supp. E for more examples). In each row, the approximate posterior distribution over view direction is visualized for both cryoAI and
+
+multi-head encoders given the input image and reconstruction. Our multi-head encoder returns a set of plausible candidates, while cryoAI obtains two pose estimates by data augmentation. In both examples, the multi-head encoder identifies the correct mode while cryoAI selects the incorrect one. Note that the pose set predicted by the multi-head encoder contains other posterior modes as well. Intuitively, cryoAI encoder with two pose predictions cannot explore the pose space as much as our multi-head encoder. Thus, the pose posterior in cryoAI remains uncertain and multi-modal during reconstruction. This also contributes to slower convergence to higher-resolution reconstructions as depicted in the top left of Fig. 5. Finally, we visualize the percentage of images assigned to each head by our "winner-takes-all" loss. The relatively even spread of assignments to different heads shows that the multi-head encoder does not suffer hypotheses collapse [30], namely all heads actively participate in the pose prediction. In Supp. C, we investigate how each head specializes in pose prediction for non-overlapping regions of  $SO(3)$ .
+
+# 6 Conclusion, Limitations and Future Work
+
+In this paper, we introduce cryoSPIN, a new semi-amortized approach to ab initio cryo-EM reconstruction. We develop a new multi-head encoder that estimates a set of plausible candidate pose to handle the uncertainty. As the uncertainty is reduced, we transition to an auto-decoding stage where poses are iteratively optimized using SGD for each image. Our results show that our multi-head encoder is able to capture multiple modes of the pose distribution, and our flexible direct optimization enables accelerated convergence of poses and reconstructions. CryoSPIN outperforms cryoAI on experimental data and achieves competitive results with cryoSPARC.
+
+Limitations and future work. In this work, we assume that the 3D structure is rigid while many biomolecules are flexible in practice, exhibiting different conformations. Our results show that, even in this relatively constrained homogeneous case, there is a significant gap between amortized inference and our semi-amortized approach. We expect that the benefits of semi-amortized inference will also apply to such heterogeneous cases. In particular, our proposed multi-head encoder should be applicable to the estimation of the multimodal posterior over heterogeneity latent parameters as well as pose. Yet, there are significant challenges in heterogeneous reconstruction, such as extra uncertainty due to heterogeneity and conflation of latent spaces, making it outside the scope of our analysis on semi-amortized inference.
+
+The development of principled criteria for switching from amortized to auto-decoding is also an interesting direction for future research. Our experiments with a range of values show that after a sufficient number of epochs, the reconstruction error becomes insensitive to the switching time. Our intuition is that when switching too early, the pose posterior may have significant uncertainty, and the pose estimates may be too far from the correct basin of attraction to afford robust convergence. Therefore, it is better to switch late than early. However, a well-defined heuristic for the optimal switch time is lacking.
+
+Societal impact. Structure determination with Cryo-EM for macromolecules is one of the most exciting, fundamental areas in structural biology, and is already having with vast social impact. Cryo-EM was used to determine the structure of the SARS-COV2 spike protein, the discovery of its pre-fusion conformation, and the evaluation of medical counter-measures. It compliments advances in methods like alpha-fold for structure prediction, transforming our understanding of the process of life at the scale of the cell, and the design and development of new therapeutics.
+
+# Acknowledgments and Disclosure of Funding
+
+This research was supported in part by the Province of Ontario, the Government of Canada, through NSERC, CIFAR, and the Canada First Research Excellence Fund for the Vision: Science to Applications (VISTA) programme, and by companies sponsoring the Vector Institute.
+
+# References
+
+[1] Werner Kuhlbrandt. The resolution revolution. Science, 343(6178):1443-1444, 2014. 1  
+[2] Amit Singer and Fred J Sigworth. Computational methods for single-particle electron cryomicroscopy. Annual Review of Biomedical Data Science, 3:163-190, 2020. 1
+
+[3] Sjors HW Scheres, Haixiao Gao, Mikel Valle, Gabor T Herman, Paul PB Eggermont, Joachim Frank, and Jose-Maria Carazo. Disentangling conformational states of macromolecules in 3d-em through likelihood optimization. Nature Methods, 4(1):27-29, 2007. 1  
+[4] Roy R Lederman, Joakim Andén, and Amit Singer. Hyper-molecules: On the representation and recovery of dynamical structures, with application to flexible macro-molecular structures in cryo-em. arXiv preprint arXiv:1907.01589, 2019.  
+[5] Ellen D Zhong, Tristan Bepler, Bonnie Berger, and Joseph H Davis. CryoDRGN: reconstruction of heterogeneous cryo-EM structures using neural networks. Nature Methods, 18(2):176–185, 2021. 2, 3, 4, 5, 6  
+[6] Ali Punjani and David J Fleet. 3DFlex: Determining structure and motion of flexible proteins from cryo-EM. Nature Methods, 20(6):860–870, 2023. 1, 3  
+[7] Ellen D Zhong, Adam Lerer, Joseph H Davis, and Bonnie Berger. CryoDRGN2: Ab initio neural reconstruction of 3d protein structures from real cryo-EM images. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4066–4075, 2021. 2, 3, 4, 6, 7  
+[8] Ali Punjani, John L Rubinstein, David J Fleet, and Marcus A Brubaker. cryoSPARC: algorithms for rapid unsupervised cryo-EM structure determination. Nature Methods, 14(3):290–296, 2017. 2, 3, 4, 6, 7, 8  
+[9] Youssef SG Nashed, Frédéric Poitevin, Harshit Gupta, Geoffrey Woollard, Michael Kagan, Chun Hong Yoon, and Daniel Ratner. CryoPoseNet: End-to-end simultaneous learning of single-particle orientation and 3D map reconstruction from cryo-electron microscopy data. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4066-4076, 2021. 2, 3, 4  
+[10] Axel Levy, Frédéric Poitevin, Julien Martel, Youssef Nashed, Ariana Peck, Nina Miolane, Daniel Ratner, Mike Dunne, and Gordon Wetzstein. CryoAI: Amortized inference of poses for ab initio reconstruction of 3d molecular volumes from real cryo-EM images. In European Conference on Computer Vision, pages 540–557. Springer, 2022. 2, 3, 4, 5, 6, 7, 8, 18  
+[11] Yoon Kim, Sam Wiseman, Andrew Miller, David Sontag, and Alexander Rush. Semi-amortized variational autoencoders. In International Conference on Machine Learning, pages 2678-2687. PMLR, 2018. 2, 5  
+[12] Abner Guzman-Rivera, Dhruv Batra, and Pushmeet Kohli. Multiple choice learning: Learning to produce multiple structured outputs. Advances in Neural Information Processing Systems, 25, 2012. 2, 3, 5  
+[13] Stefan Lee, Senthil Purushwalkam Shiva Prakash, Michael Cogswell, Viresh Ranjan, David Crandall, and Dhruv Batra. Stochastic multiple choice learning for training diverse deep ensembles. Advances in Neural Information Processing Systems, 29, 2016. 3, 5, 16  
+[14] Christian Rupprecht, Iro Laina, Robert DiPietro, Maximilian Baust, Federico Tombari, Nassir Navab, and Gregory D Hager. Learning in an uncertain world: Representing ambiguity through multiple hypotheses. In Proceedings of the IEEE International Conference on Computer Vision, pages 3591–3600, 2017. 2, 3, 5  
+[15] Axel Levy, Gordon Wetzstein, Julien NP Martel, Frederic Poitevin, and Ellen Zhong. Amortized inference for heterogeneous reconstruction in cryo-EM. Advances in Neural Information Processing Systems, 35:13038–13049, 2022. 3, 4  
+[16] Amit Singer, Ronald R Coifman, Fred J Sigworth, David W Chester, and Yoel Shkolnisky. Detecting consistent common lines in cryo-EM by voting. Journal of Structural Biology, 169(3):312–322, 2010. 3  
+[17] Ido Greenberg and Yoel Shkolnisky. Common lines modeling for reference free ab-initio reconstruction in cryo-EM. Journal of Structural Biology, 200(2):106-117, 2017.  
+[18] Gabi Pragier and Yoel Shkolnisky. A common lines approach for ab initio modeling of cyclically symmetric molecules. Inverse Problems, 35(12):124005, 2019. 3
+
+[19] Pawel A Penczek, Robert A Grassucci, and Joachim Frank. The ribosome at improved resolution: new techniques for merging and orientation refinement in 3D cryo-electron microscopy of biological particles. Ultramicroscopy, 53(3):251-270, 1994. 3  
+[20] Timothy S Baker and R Holland Cheng. A model-based approach for determining orientations of biological macromolecules imaged by cryoelectron microscopy. Journal of Structural Biology, 116(1):120-130, 1996. 3  
+[21] Sjors HW Scheres. A Bayesian view on cryo-EM structure determination. Journal of Molecular Biology, 415(2):406–418, 2012. 3, 4  
+[22] Sjors HW Scheres. RELION: implementation of a Bayesian approach to cryo-EM structure determination. Journal of Structural Biology, 180(3):519–530, 2012. 3  
+[23] Eugene L Lawler and David E Wood. Branch-and-bound methods: A survey. Operations Research, 14(4):699-719, 1966. 3  
+[24] Dan Rosenbaum, Marta Garnelo, Michal Zielinski, Charlie Beattie, Ellen Clancy, Andrea Huber, Pushmeet Kohli, Andrew W Senior, John Jumper, Carl Doersch, et al. Inferring a continuous distribution of atom coordinates from cryo-EM images using VAEs. arXiv preprint arXiv:2106.14108, 2021. 3  
+[25] Diederik P Kingma and Max Welling. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 3  
+[26] Ioannis Tsochantaridis, Thorsten Joachims, Thomas Hofmann, Yasemin Altun, and Yoram Singer. Large margin methods for structured and interdependent output variables. Journal of Machine Learning Research, 6(9), 2005. 3  
+[27] Stefan Lee, Senthil Purushwalkam, Michael Cogswell, David Crandall, and Dhruv Batra. Why m heads are better than one: Training a diverse ensemble of deep networks. arXiv preprint arXiv:1511.06314, 2015. 3, 5, 16  
+[28] Kimin Lee, Changho Hwang, KyoungSoo Park, and Jinwoo Shin. Confident multiple choice learning. In International Conference on Machine Learning, pages 2014–2023. PMLR, 2017. 3  
+[29] Kai Tian, Yi Xu, Shuigeng Zhou, and Jihong Guan. Versatile multiple choice learning and its application to vision computing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6349-6357, 2019.  
+[30] Victor Letzelter, Mathieu Fontaine, Mickael Chen, Patrick Pérez, Slim Essid, and Gael Richard. Resilient multiple choice learning: A learned scoring scheme with application to audio scene analysis. Advances in Neural Information Processing Systems, 36, 2024. 3, 10  
+[31] Simon Kohl, Bernardino Romero-Paredes, Clemens Meyer, Jeffrey De Fauw, Joseph R Ledsam, Klaus Maier-Hein, SM Eslami, Danilo Jimenez Rezende, and Olaf Ronneberger. A probabilistic U-Net for segmentation of ambiguous images. Advances in Neural Information Processing Systems, 31, 2018. 3  
+[32] Eddy Ilg, Ozgun Cicek, Silvio Galesso, Aaron Klein, Osama Makansi, Frank Hutter, and Thomas Brox. Uncertainty estimates and multi-hypotheses networks for optical flow. In European Conference on Computer Vision, pages 652-667, 2018. 3  
+[33] Ye Yuan and Kris Kitani. Diverse trajectory forecasting with determinantal point processes. arXiv preprint arXiv:1907.04967, 2019. 3  
+[34] Benjamin Biggs, David Novotny, Sebastien Ehrhardt, Hanbyul Joo, Ben Graham, and Andrea Vedaldi. 3D Multi-bodies: Fitting sets of plausible 3d human models to ambiguous image data. Advances in Neural Information Processing Systems, 33:20496-20507, 2020. 3  
+[35] Earl J Kirkland. Advanced computing in electron microscopy, volume 12. Springer, 1998. 4  
+[36] Ronald N Bracewell. Strip integration in radio astronomy. Australian Journal of Physics, 9(2):198-217, 1956. 4
+
+[37] Todd K Moon. The expectation-maximization algorithm. IEEE Signal processing magazine, 13(6):47-60, 1996. 4  
+[38] Vincent Sitzmann, Julien Martel, Alexander Bergman, David Lindell, and Gordon Wetzstein. Implicit neural representations with periodic activation functions. Advances in Neural Information Processing Systems, 33:7462-7473, 2020. 4, 5  
+[39] Matthew Tancik, Pratul Srinivasan, Ben Mildenhall, Sara Fridovich-Keil, Nithin Raghavan, Utkarsh Singhal, Ravi Ramamoorthi, Jonathan Barron, and Ren Ng. Fourier features let networks learn high frequency functions in low dimensional domains. Advances in Neural Information Processing Systems, 33:7537–7547, 2020. 5  
+[40] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99–106, 2021. 4, 5  
+[41] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 5  
+[42] Devon Hjelm, Russ R Salakhutdinov, Kyunghyun Cho, Nebojsa Jojic, Vince Calhoun, and Junyoung Chung. Iterative refinement of the approximate posterior for directed belief networks. Advances in Neural Information Processing Systems, 29, 2016. 5  
+[43] Rahul Krishnan, Dawen Liang, and Matthew Hoffman. On the challenges of learning with inference networks on sparse, high-dimensional data. In International Conference on Artificial Intelligence and Statistics, pages 143-151. PMLR, 2018.  
+[44] Joe Marino, Yisong Yue, and Stephan Mandt. Iterative amortized inference. In International Conference on Machine Learning, pages 3403-3412. PMLR, 2018. 5  
+[45] D Goodsell. PDB-101 Molecule of the Month: Hsp90. 2008. 6, 9  
+[46] Clemens Plaschka, Pei-Chun Lin, and Kiyoshi Nagai. Structure of a pre-catalytic spliceosome. Nature, 546(7660):617-621, 2017. 6  
+[47] Alexandra C Walls, Young-Jun Park, M Alejandra Tortorici, Abigail Wall, Andrew T McGuire, and David Veesler. Structure, function, and antigenicity of the SARS-CoV-2 spike glycoprotein. Cell, 181(2):281–292, 2020. 6  
+[48] Wilson Wong, Xiao-chen Bai, Alan Brown, Israel S Fernandez, Eric Hanssen, Melanie Condron, Yan Hong Tan, Jake Baum, and Sjors HW Scheres. Cryo-EM structure of the Plasmodium falciparum 80S ribosome bound to the anti-protozoan drug emetine. *Elife*, 3:e03080, 2014. 6  
+[49] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6, 15  
+[50] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in Neural Information Processing Systems, 32, 2019. 6  
+[51] Thomas D Goddard, Conrad C Huang, Elaine C Meng, Eric F Pettersen, Gregory S Couch, John H Morris, and Thomas E Ferrin. UCSF ChimeraX: Meeting modern challenges in visualization and analysis. Protein Science, 27(1):14-25, 2018. 7  
+[52] Marin Van Heel and Michael Schatz. Fourier shell correlation threshold criteria. Journal of Structural Biology, 151(3):250–262, 2005. 7  
+[53] Krzysztof M Gorski, Eric Hivon, Anthony J Banday, Benjamin D Wandelt, Frode K Hansen, Mstvos Reinecke, and Matthia Bartelmann. HEALPix: A framework for high-resolution discretization and fast analysis of data distributed on the sphere. The Astrophysical Journal, 622(2):759, 2005. 8, 16, 17
+
+[54] Adam Paszke, Sam Gross, Soumith Chintala, Gregory Chanan, Edward Yang, Zachary DeVito, Zeming Lin, Alban Desmaison, Luca Antiga, and Adam Lerer. Automatic differentiation in pytorch. 2017. 15  
+[55] Julian Panetta. Optimizing over SO(3). Technical report, University of California, Davis, 2018. 15  
+[56] F Sebastian Grassia. Practical parameterization of rotations using the exponential map. Journal of Graphics Tools, 3(3):29-48, 1998. 15
+
+# Supplementary Material
+
+# CryoSPIN: Improving Ab-Initio Cryo-EM Reconstruction with Semi-Amortized Pose Inference
+
+# A Pose Optimization
+
+For the auto-encoding stage, we follow cryoAI and design the convolutional backbone to output the six-dimensional representation commonly referred to as  $S2S2$ . To compute the rotation matrix from this representation, the 6D vector is split into two 3D vectors and normalized, denoted as  $v_{1}, v_{2} \in \mathbb{R}^{3}$ . We then compute the cross product between them,  $v_{3} = v_{1} \times v_{2}$ , yielding a new unit vector. Selecting  $v_{1}$  and  $v_{3}$  as the first and third columns of the target rotation matrix, we compute  $\tilde{v}_{2} = v_{3} \times v_{1}$ , which is another unit vector orthogonal to both  $v_{1}$  and  $v_{3}$ . Then,  $R = [v_{1}, \tilde{v}_{2}, v_{3}]$ .
+
+Once switched to direct optimization, we change parameterization to axis-angle representation. During auto-decoding, we alternate between five iterations of pose SGD updates and one iteration of volume update. To update poses, we keep the volume fixed and optimize for the negative log-likelihood (Eq. 3) with respect to the pose parameters. We define the new pose estimate based on the current one as follows:
+
+$$
+R _ {t + 1} = R _ {\delta} R _ {t} \tag {9}
+$$
+
+where  $R_{\delta}$  is an infinitesimal rotation matrix perturbing the current estimate. The perturbation matrix  $R_{\delta}$  is parameterized by axis-angle representation. By a single vector  $\omega \in \mathbb{R}^3$ , one can represent both the axis  $||\omega||$  and the angle  $0 < \frac{\omega}{||\omega||} < \pi$  for any given rotation. Using Rodrigues formula, the perturbation matrix  $R_{\delta}(\omega)$  can be parameterized as a function of  $\omega$ . To find the optimal  $\omega$ , one can initialize it with zero vector, and then use automatic differentiation [54] in pytorch to compute the gradient with respect to  $\omega$  and make updates using Adam [49]. However, a naive implementation of the function  $R_{\delta}(\omega)$  would lead to numerically unstable calculations of the partial derivative  $\frac{\partial R_{\delta}}{\partial\omega}$ . In fact, there is a singularity at the zero vector, and the partial derivative involves terms that are unstable around the origin. Formally, the derivative of  $i$ -th column of the rotation matrix  $R_{\delta}$  with respect to the vector  $\omega$  is [55, 56],
+
+$$
+\begin{array}{l} \frac {\partial R _ {\delta} ^ {(i)}}{\partial \boldsymbol {\omega}} = - \left(\mathbf {e} ^ {(i)} \otimes \boldsymbol {\omega} + [ \mathbf {e} ^ {(i)} ] _ {\times}\right) \frac {\sin (\| \boldsymbol {\omega} \|)}{\| \boldsymbol {\omega} \|} + [ (\boldsymbol {\omega} \cdot \mathbf {e} ^ {(i)}) I + \boldsymbol {\omega} \otimes \mathbf {e} ^ {(i)} ] \left(\frac {1 - \cos (\| \boldsymbol {\omega} \|)}{\| \boldsymbol {\omega} \| ^ {2}}\right) \\ + (\boldsymbol {\omega} \otimes \boldsymbol {\omega}) \left(\left(\boldsymbol {\omega} \cdot \mathbf {e} ^ {(i)}\right) \frac {2 \cos (\| \boldsymbol {\omega} \|) - 2 + \| \boldsymbol {\omega} \| \sin (\| \boldsymbol {\omega} \|)}{\| \boldsymbol {\omega} \| ^ {4}}\right) \\ + \left[ (\boldsymbol {\omega} \times \mathbf {e} ^ {(i)}) \otimes \boldsymbol {\omega} \right] \frac {\| \boldsymbol {\omega} \| \cos (\| \boldsymbol {\omega} \|) - \sin (\| \boldsymbol {\omega} \|)}{\| \boldsymbol {\omega} \| ^ {3}}. \\ \end{array}
+$$
+
+where  $\otimes$  and  $\times$  are tensor and cross products, respectively.  $\mathbf{e}^{(i)}$  is the  $i$ -th standard basis in 3D and  $[\mathbf{v}]_{\times}$  denotes the cross product matrix for the vector  $\mathbf{v}$ . In all four terms, there are scalars such as  $\frac{\sin(\|\boldsymbol{\omega}\|)}{\|\boldsymbol{\omega}\|}$  or  $\frac{1 - \cos(\|\boldsymbol{\omega}\|)}{\|\boldsymbol{\omega}\|^2}$  that evaluate to  $\frac{0}{0}$  at zero angle  $\boldsymbol{\omega} = 0$ . Similar to [55], for  $||\boldsymbol{\omega}|| \ll 1$ , we substitute these terms with their numerically robust Taylor expansion, for instance,
+
+$$
+\frac {\sin (\| \boldsymbol {\omega} \|)}{\| \boldsymbol {\omega} \|} = 1 - \frac {| | \boldsymbol {\omega} | | ^ {2}}{6} + O (| | \boldsymbol {\omega} | | ^ {4}),
+$$
+
+$$
+\frac {1 - \cos (\| \pmb {\omega} \|)}{\| \pmb {\omega} \| ^ {2}} = \frac {1}{2} - \frac {| | \pmb {\omega} | | ^ {2}}{2 4} + O (| | \pmb {\omega} | | ^ {4}).
+$$
+
+We implement a differentiable and numerically stable version of the function  $R_{\delta}(\omega)$  in pytorch and use it in our pose estimation module.
+
+# B Ablation Study on Decoder
+
+We perform an ablation study on the decoder, detailed in Table 2, as well as a comparison with baselines in terms of reconstruction time per epoch, GPU memory usage, and number of parameters. As our method consists of two stages, we report numbers for each stage separately. In the first
+
+Table 2: The reconstruction time per epoch, GPU memory, and number of parameters for different methods averaged across three runs. GPU memory is recorded separately for the encoding and decoding modules. Also, two numbers provided for semi-amortized method corresponds to auto-encoding and auto-decoding stages, respectively. In CryoAI-explicit, the implicit decoder of CryoAI is replaced with an explicit decoder. Since we store rotations in 3D representation during direct optimization, the number of parameters of pose module is  $N \times 3$  with  $N$  denoting number of images in millions.  
+
+<table><tr><td rowspan="2">Model</td><td rowspan="2">Time (s)</td><td colspan="2">GPU Mem. (GB)</td><td rowspan="2"># Params (M)</td></tr><tr><td>Encoding</td><td>Decoding</td></tr><tr><td>Semi-Amortized</td><td>99.76, 81.61</td><td>1.17, -</td><td>2.22, 0.35</td><td>13.01, 4.29 + N × 3</td></tr><tr><td>Fully-Amortized</td><td>99.75</td><td>1.17</td><td>2.22</td><td>13.01</td></tr><tr><td>CryoAI-explicit</td><td>142.40</td><td>2.32</td><td>0.75</td><td>9.05</td></tr><tr><td>CryoAI</td><td>622.39</td><td>2.78</td><td>14.05</td><td>5.09</td></tr></table>
+
+stage, the encoder has  $\mathrm{H} = 7$  heads, while in the second stage, it is replaced with a pose module with size depending on the number of particles ( $N$ ). To facilitate comparison, we include another baseline, cryoAI-explicit, in which the implicit decoder is replaced by an explicit one as in our method. The explicit and implicit decoders have 4.29 and 0.33 million parameters, respectively. Despite a larger decoder, our method is 6x faster and uses 5x less memory compared to cryoAI. Importantly, swapping the implicit decoder with an explicit one (cryoAI-explicit) significantly drops time and memory, indicating that the implicit decoder is a major computational and memory bottleneck. Yet, cryoAI-explicit uses 2x more GPU memory for encoding and is 1.5x slower than our method as it performs early input augmentation and runs the entire encoder twice per image. Our method, by augmenting the encoder head, saves memory and time during pose encoding. Finally, the amortized baseline, which uses an explicit decoder coupled with a multi-head encoder ( $\mathrm{H} = 7$ ), uses more memory in decoding (negligible vs implicit decoder) and runs slower than the direct optimization stage of the semi-amortized method.
+
+# C Specialization of Encoder Heads
+
+A natural question about the multi-head encoder is: how each head does take part in pose encoding process? To address this, using the synthetic datasets, we conduct an experiment with our multi-head architecture  $(M = 4)$  and visualize the performance of each head on different regions of  $SO(3)$  space. In particular, as before, we define a uniform grid on the unit sphere using HEALPix [53] and assign images to their corresponding cells based on the view-direction. Now, for all images end up in the same cell, we compute the average rotation error and visualize it separately for each head. As shown in Fig. 6, all heads actively participate in pose estimation and they are able to specialize in prediction of poses for images with certain view-direction. A similar result has been provided in prior work on MCL [27, 13], to show that minimizing the error made by the best prediction ("oracle" loss) encourages diversity in deep ensembles. In our problem, by optimizing a "winner-takes-all" loss, the whole burden of pose estimation is no longer on a single network but it gets divided between multiple heads as separate predictors.
+
+# D More on Semi-Amortized vs. Fully-Amortized
+
+To validate the advantages of direct pose optimization in our semi-amortized method, we further show more qualitative examples of paths taken by pose estimates over the optimization landscape during reconstruction in Fig. 7. For both methods, optimization start from the same point marked by black dot in the vicinity of the distribution mode. It will then continue in paths colored in blue and red for semi-amortized and full-amortized methods, respectively. We observe in all examples that iterative updates by stochastic gradient descent demonstrate a stable convergence toward the optima while poses obtained by amortized inference show unstable behavior around the mode.
+
+![](images/b64f934c7f58dcf9595a7c4c4a79702ab30a3128a38889944a762c54f66b65b7.jpg)  
+Figure 6: Average rotation error is visualized over the sphere for different heads of the multi-head pose encoder  $(M = 4)$ . The sphere is uniformly divided into cells using HEALPix [53]; images are assigned to cells based on ground-truth view-direction. For each cell, the average rotation error is visualized, showing diverse behavior of different heads across the space. Blue and red colors show low and high error regions, respectively. Error ranges from zero to 180 degrees.
+
+![](images/de3e9ac6f326b2ad780a58010018db673b53d58109402c491d927bdbe0d21da4.jpg)  
+Figure 7: We compare the behavior of fully-amortized and semi-amortized pose inference on four examples per dataset. Two rows correspond to Spike and Spliceosome datasets, respectively. Each plot shows the approximate log pose posterior, marginalized over in-plane rotations represented as a heat map on a uniform grid over the unit sphere  $S^2$ . Gnomonic projection to 2D is also applied, followed by zooming on the proximity of the mode of interest. Black cross is the starting point while blue dots and red dots show poses estimated by fully- and semi-amortized methods, respectively.
+
+# E More on Multi-Modal Pose Posterior
+
+Through more examples (Fig. 8), we demonstrate that cryoAI fails to handle ambiguity in pose estimation on HSP dataset. The visualization shows that pose estimates by cryoAI become stuck in incorrect modes whereas our pose encoder with multi-head architecture is able to return a pose candidate that captures the correct mode.
+
+![](images/e91ef2121a66270de5c6c71c20257a61eec018f07152f130ecfb52c54f0f23fe.jpg)  
+Figure 8: The approximate log posterior of view-direction visualized on the unit sphere with highlighted areas showing modes of the distribution. CryoAI [10] and our multi-head encoders provide two and four pose estimates, respectively, which are marked with colored dots on the sphere (the order of poses is arbitrary). The corresponding projections are also illustrated. CryoAI cannot identify the correct mode of pose distribution.
+
+# NeurIPS Paper Checklist
+
+The checklist is designed to encourage best practices for responsible machine learning research, addressing issues of reproducibility, transparency, research ethics, and societal impact. Do not remove the checklist: The papers not including the checklist will be desk rejected. The checklist should follow the references and follow the (optional) supplemental material. The checklist does NOT count towards the page limit.
+
+Please read the checklist guidelines carefully for information on how to answer these questions. For each question in the checklist:
+
+- You should answer [Yes], [No], or [NA].  
+- [NA] means either that the question is Not Applicable for that particular paper or the relevant information is Not Available.  
+- Please provide a short (1–2 sentence) justification right after your answer (even for NA).
+
+The checklist answers are an integral part of your paper submission. They are visible to the reviewers, area chairs, senior area chairs, and ethics reviewers. You will be asked to also include it (after eventual revisions) with the final version of your paper, and its final version will be published with the paper.
+
+The reviewers of your paper will be asked to use the checklist as one of the factors in their evaluation. While "[Yes]" is generally preferable to "[No]", it is perfectly acceptable to answer "[No]" provided a proper justification is given (e.g., "error bars are not reported because it would be too computationally expensive" or "we were unable to find the license for the dataset we used"). In general, answering "[No]" or "[NA]" is not grounds for rejection. While the questions are phrased in a binary way, we acknowledge that the true answer is often more nuanced, so please just use your best judgment and write a justification to elaborate. All supporting evidence can appear either in the main paper or the supplemental material, provided in appendix. If you answer [Yes] to a question, in the justification please point to the section(s) where related material for the question can be found.
+
+IMPORTANT, please:
+
+- Delete this instruction block, but keep the section heading "NeurIPS paper checklist",  
+- Keep the checklist subsection headings, questions/answers and guidelines below.  
+- Do not modify the questions and only use the provided macros for your answers.
+
+# 1. Claims
+
+Question: Do the main claims made in the abstract and introduction accurately reflect the paper's contributions and scope?
+
+Answer: [Yes]
+
+Justification: We clearly state the claims in the abstract, discuss them through the introduction section, and summarize them in a list at the end of introduction section.
+
+Guidelines:
+
+- The answer NA means that the abstract and introduction do not include the claims made in the paper.  
+- The abstract and/or introduction should clearly state the claims made, including the contributions made in the paper and important assumptions and limitations. A No or NA answer to this question will not be perceived well by the reviewers.  
+- The claims made should match theoretical and experimental results, and reflect how much the results can be expected to generalize to other settings.  
+- It is fine to include aspirational goals as motivation as long as it is clear that these goals are not attained by the paper.
+
+# 2. Limitations
+
+Question: Does the paper discuss the limitations of the work performed by the authors?
+
+Answer: [Yes]
+
+Justification: We discuss limitations in the conclusion section. We acknowledge that ours assumptions about homogeneity and centered images could be limiting in some experimental setups. However, our framework could be extended to account for heterogeneity and translation parameters.
+
+# Guidelines:
+
+- The answer NA means that the paper has no limitation while the answer No means that the paper has limitations, but those are not discussed in the paper.  
+- The authors are encouraged to create a separate "Limitations" section in their paper.  
+- The paper should point out any strong assumptions and how robust the results are to violations of these assumptions (e.g., independence assumptions, noiseless settings, model well-specification, asymptotic approximations only holding locally). The authors should reflect on how these assumptions might be violated in practice and what the implications would be.  
+- The authors should reflect on the scope of the claims made, e.g., if the approach was only tested on a few datasets or with a few runs. In general, empirical results often depend on implicit assumptions, which should be articulated.  
+- The authors should reflect on the factors that influence the performance of the approach. For example, a facial recognition algorithm may perform poorly when image resolution is low or images are taken in low lighting. Or a speech-to-text system might not be used reliably to provide closed captions for online lectures because it fails to handle technical jargon.  
+- The authors should discuss the computational efficiency of the proposed algorithms and how they scale with dataset size.  
+- If applicable, the authors should discuss possible limitations of their approach to address problems of privacy and fairness.  
+- While the authors might fear that complete honesty about limitations might be used by reviewers as grounds for rejection, a worse outcome might be that reviewers discover limitations that aren't acknowledged in the paper. The authors should use their best judgment and recognize that individual actions in favor of transparency play an important role in developing norms that preserve the integrity of the community. Reviewers will be specifically instructed to not penalize honesty concerning limitations.
+
+# 3. Theory Assumptions and Proofs
+
+Question: For each theoretical result, does the paper provide the full set of assumptions and a complete (and correct) proof?
+
+Answer: [NA]
+
+Justification: We provide no theoretical results.
+
+# Guidelines:
+
+- The answer NA means that the paper does not include theoretical results.  
+- All the theorems, formulas, and proofs in the paper should be numbered and cross-referenced.  
+- All assumptions should be clearly stated or referenced in the statement of any theorems.  
+- The proofs can either appear in the main paper or the supplemental material, but if they appear in the supplemental material, the authors are encouraged to provide a short proof sketch to provide intuition.  
+- Inversely, any informal proof provided in the core of the paper should be complemented by formal proofs provided in appendix or supplemental material.  
+- Theorems and Lemmas that the proof relies upon should be properly referenced.
+
+# 4. Experimental Result Reproducibility
+
+Question: Does the paper fully disclose all the information needed to reproduce the main experimental results of the paper to the extent that it affects the main claims and/or conclusions of the paper (regardless of whether the code and data are provided or not)?
+
+Answer: [Yes]
+
+Justification: In a subsection called implementation details, we included all details needed to reproduce reconstruction results of our method, cryoAI and cryoSPARC. Steps to generate synthetic datasets and preprocessing steps to prepare experimental data are also provided.
+
+# Guidelines:
+
+- The answer NA means that the paper does not include experiments.  
+- If the paper includes experiments, a No answer to this question will not be perceived well by the reviewers: Making the paper reproducible is important, regardless of whether the code and data are provided or not.  
+- If the contribution is a dataset and/or model, the authors should describe the steps taken to make their results reproducible or verifiable.  
+- Depending on the contribution, reproducibility can be accomplished in various ways. For example, if the contribution is a novel architecture, describing the architecture fully might suffice, or if the contribution is a specific model and empirical evaluation, it may be necessary to either make it possible for others to replicate the model with the same dataset, or provide access to the model. In general, releasing code and data is often one good way to accomplish this, but reproducibility can also be provided via detailed instructions for how to replicate the results, access to a hosted model (e.g., in the case of a large language model), releasing of a model checkpoint, or other means that are appropriate to the research performed.  
+- While NeurIPS does not require releasing code, the conference does require all submissions to provide some reasonable avenue for reproducibility, which may depend on the nature of the contribution. For example  
+(a) If the contribution is primarily a new algorithm, the paper should make it clear how to reproduce that algorithm.  
+(b) If the contribution is primarily a new model architecture, the paper should describe the architecture clearly and fully.  
+(c) If the contribution is a new model (e.g., a large language model), then there should either be a way to access this model for reproducing the results or a way to reproduce the model (e.g., with an open-source dataset or instructions for how to construct the dataset).  
+(d) We recognize that reproducibility may be tricky in some cases, in which case authors are welcome to describe the particular way they provide for reproducibility. In the case of closed-source models, it may be that access to the model is limited in some way (e.g., to registered users), but it should be possible for other researchers to have some path to reproducing or verifying the results.
+
+# 5. Open access to data and code
+
+Question: Does the paper provide open access to the data and code, with sufficient instructions to faithfully reproduce the main experimental results, as described in supplemental material?
+
+Answer: [Yes]
+
+Justification:
+
+Guidelines:
+
+- The answer NA means that paper does not include experiments requiring code.  
+- Please see the NeurIPS code and data submission guidelines (https://nips.cc/public/guides/CodeSubmissionPolicy) for more details.  
+- While we encourage the release of code and data, we understand that this might not be possible, so "No" is an acceptable answer. Papers cannot be rejected simply for not including code, unless this is central to the contribution (e.g., for a new open-source benchmark).  
+- The instructions should contain the exact command and environment needed to run to reproduce the results. See the NeurIPS code and data submission guidelines (https://nips.cc/public/guides/CodeSubmissionPolicy) for more details.  
+- The authors should provide instructions on data access and preparation, including how to access the raw data, preprocessed data, intermediate data, and generated data, etc.
+
+- The authors should provide scripts to reproduce all experimental results for the new proposed method and baselines. If only a subset of experiments are reproducible, they should state which ones are omitted from the script and why.  
+- At submission time, to preserve anonymity, the authors should release anonymized versions (if applicable).  
+- Providing as much information as possible in supplemental material (appended to the paper) is recommended, but including URLs to data and code is permitted.
+
+# 6. Experimental Setting/Details
+
+Question: Does the paper specify all the training and test details (e.g., data splits, hyperparameters, how they were chosen, type of optimizer, etc.) necessary to understand the results?
+
+Answer: [Yes]
+
+Justification: We have discussed all the necessary information about experimental settings in implementation details.
+
+Guidelines:
+
+- The answer NA means that the paper does not include experiments.  
+- The experimental setting should be presented in the core of the paper to a level of detail that is necessary to appreciate the results and make sense of them.  
+- The full details can be provided either with the code, in appendix, or as supplemental material.
+
+# 7. Experiment Statistical Significance
+
+Question: Does the paper report error bars suitably and correctly defined or other appropriate information about the statistical significance of the experiments?
+
+Answer: [Yes]
+
+Justification: When reporting the resolution, we follow the standard to compute FSC by splitting data into two halves and make independent reconstructions and compare resulting half-maps.
+
+Guidelines:
+
+- The answer NA means that the paper does not include experiments.  
+- The authors should answer "Yes" if the results are accompanied by error bars, confidence intervals, or statistical significance tests, at least for the experiments that support the main claims of the paper.  
+- The factors of variability that the error bars are capturing should be clearly stated (for example, train/test split, initialization, random drawing of some parameter, or overall run with given experimental conditions).  
+- The method for calculating the error bars should be explained (closed form formula call to a library function, bootstrap, etc.)  
+- The assumptions made should be given (e.g., Normally distributed errors).  
+- It should be clear whether the error bar is the standard deviation or the standard error of the mean.  
+- It is OK to report 1-sigma error bars, but one should state it. The authors should preferably report a 2-sigma error bar than state that they have a  $96\%$  CI, if the hypothesis of Normality of errors is not verified.  
+- For asymmetric distributions, the authors should be careful not to show in tables or figures symmetric error bars that would yield results that are out of range (e.g. negative error rates).  
+- If error bars are reported in tables or plots, The authors should explain in the text how they were calculated and reference the corresponding figures or tables in the text.
+
+# 8. Experiments Compute Resources
+
+Question: For each experiment, does the paper provide sufficient information on the computer resources (type of compute workers, memory, time of execution) needed to reproduce the experiments?
+
+# Answer: [Yes]
+
+Justification: We have mentioned that we are using A40 Nvidia GPU to produc the results. Also, in Fig. 3 we compare the execution time of our method vs. cryoAI.
+
+# Guidelines:
+
+- The answer NA means that the paper does not include experiments.  
+- The paper should indicate the type of compute workers CPU or GPU, internal cluster, or cloud provider, including relevant memory and storage.  
+- The paper should provide the amount of compute required for each of the individual experimental runs as well as estimate the total compute.  
+- The paper should disclose whether the full research project required more compute than the experiments reported in the paper (e.g., preliminary or failed experiments that didn't make it into the paper).
+
+# 9. Code Of Ethics
+
+Question: Does the research conducted in the paper conform, in every respect, with the NeurIPS Code of Ethics https://neurips.cc/public/EthicsGuidelines?
+
+# Answer: [Yes]
+
+Justification: We conform with the NeurIPS Code of Ethics.
+
+# Guidelines:
+
+- The answer NA means that the authors have not reviewed the NeurIPS Code of Ethics.  
+- If the authors answer No, they should explain the special circumstances that require a deviation from the Code of Ethics.  
+- The authors should make sure to preserve anonymity (e.g., if there is a special consideration due to laws or regulations in their jurisdiction).
+
+# 10. Broader Impacts
+
+Question: Does the paper discuss both potential positive societal impacts and negative societal impacts of the work performed?
+
+# Answer: [Yes]
+
+Justification: Societal impacts of the work is discussed in the conclusion section.
+
+# Guidelines:
+
+- The answer NA means that there is no societal impact of the work performed.  
+- If the authors answer NA or No, they should explain why their work has no societal impact or why the paper does not address societal impact.  
+- Examples of negative societal impacts include potential malicious or unintended uses (e.g., disinformation, generating fake profiles, surveillance), fairness considerations (e.g., deployment of technologies that could make decisions that unfairly impact specific groups), privacy considerations, and security considerations.  
+- The conference expects that many papers will be foundational research and not tied to particular applications, let alone deployments. However, if there is a direct path to any negative applications, the authors should point it out. For example, it is legitimate to point out that an improvement in the quality of generative models could be used to generate deepfakes for disinformation. On the other hand, it is not needed to point out that a generic algorithm for optimizing neural networks could enable people to train models that generate Deepfakes faster.  
+- The authors should consider possible harms that could arise when the technology is being used as intended and functioning correctly, harms that could arise when the technology is being used as intended but gives incorrect results, and harms following from (intentional or unintentional) misuse of the technology.  
+- If there are negative societal impacts, the authors could also discuss possible mitigation strategies (e.g., gated release of models, providing defenses in addition to attacks, mechanisms for monitoring misuse, mechanisms to monitor how a system learns from feedback over time, improving the efficiency and accessibility of ML).
+
+# 11. Safeguards
+
+Question: Does the paper describe safeguards that have been put in place for responsible release of data or models that have a high risk for misuse (e.g., pretrained language models, image generators, or scraped datasets)?
+
+Answer: [NA]
+
+Justification:
+
+Guidelines:
+
+- The answer NA means that the paper poses no such risks.  
+- Released models that have a high risk for misuse or dual-use should be released with necessary safeguards to allow for controlled use of the model, for example by requiring that users adhere to usage guidelines or restrictions to access the model or implementing safety filters.  
+- Datasets that have been scraped from the Internet could pose safety risks. The authors should describe how they avoided releasing unsafe images.  
+- We recognize that providing effective safeguards is challenging, and many papers do not require this, but we encourage authors to take this into account and make a best faith effort.
+
+# 12. Licenses for existing assets
+
+Question: Are the creators or original owners of assets (e.g., code, data, models), used in the paper, properly credited and are the license and terms of use explicitly mentioned and properly respected?
+
+Answer: [Yes]
+
+Justification: For cryoAI, we are using the available code provided in github. For cryoSPARC, we have mentioned which version of the software is used. Synthetic datasets are generated based on atomic models on PDB which are free of all copyright restrictions and freely available for both non-commercial and commercial use. For experimental data, we use EMPIAR database which is freely and publicly available to the global community under the CC0 licence.
+
+# Guidelines:
+
+- The answer NA means that the paper does not use existing assets.  
+- The authors should cite the original paper that produced the code package or dataset.  
+- The authors should state which version of the asset is used and, if possible, include a URL.  
+- The name of the license (e.g., CC-BY 4.0) should be included for each asset.  
+- For scraped data from a particular source (e.g., website), the copyright and terms of service of that source should be provided.  
+- If assets are released, the license, copyright information, and terms of use in the package should be provided. For popular datasets, paperswithcode.com/datasets has curated licenses for some datasets. Their licensing guide can help determine the license of a dataset.  
+- For existing datasets that are re-packaged, both the original license and the license of the derived asset (if it has changed) should be provided.  
+- If this information is not available online, the authors are encouraged to reach out to the asset's creators.
+
+# 13. New Assets
+
+Question: Are new assets introduced in the paper well documented and is the documentation provided alongside the assets?
+
+Answer: [NA]
+
+Justification: We don't release any new assets.
+
+Guidelines:
+
+- The answer NA means that the paper does not release new assets.
+
+- Researchers should communicate the details of the dataset/code/model as part of their submissions via structured templates. This includes details about training, license, limitations, etc.  
+- The paper should discuss whether and how consent was obtained from people whose asset is used.  
+- At submission time, remember to anonymize your assets (if applicable). You can either create an anonymized URL or include an anonymized zip file.
+
+# 14. Crowdsourcing and Research with Human Subjects
+
+Question: For crowdsourcing experiments and research with human subjects, does the paper include the full text of instructions given to participants and screenshots, if applicable, as well as details about compensation (if any)?
+
+Answer: [NA]
+
+Justification: No crowdsourcing or research with human subject.
+
+Guidelines:
+
+- The answer NA means that the paper does not involve crowdsourcing nor research with human subjects.  
+- Including this information in the supplemental material is fine, but if the main contribution of the paper involves human subjects, then as much detail as possible should be included in the main paper.  
+- According to the NeurIPS Code of Ethics, workers involved in data collection, curation, or other labor should be paid at least the minimum wage in the country of the data collector.
+
+# 15. Institutional Review Board (IRB) Approvals or Equivalent for Research with Human Subjects
+
+Question: Does the paper describe potential risks incurred by study participants, whether such risks were disclosed to the subjects, and whether Institutional Review Board (IRB) approvals (or an equivalent approval/review based on the requirements of your country or institution) were obtained?
+
+Answer: [NA]
+
+Justification: No crowdsourcing or research with human subject.
+
+Guidelines:
+
+- The answer NA means that the paper does not involve crowdsourcing nor research with human subjects.  
+- Depending on the country in which research is conducted, IRB approval (or equivalent) may be required for any human subjects research. If you obtained IRB approval, you should clearly state this in the paper.  
+- We recognize that the procedures for this may vary significantly between institutions and locations, and we expect authors to adhere to the NeurIPS Code of Ethics and the guidelines for their institution.  
+- For initial submissions, do not include any information that would break anonymity (if applicable), such as the institution conducting the review.

2024/CryoSPIN_ Improving Ab-Initio Cryo-EM Reconstruction with Semi-Amortized Pose Inference/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a8ad9eca8fb7a14249cfb57ed73a3641f92aa7af0398f09bf140b8b71b5a90c2
+size 918246

2024/Cryptographic Hardness of Score Estimation/9e349f06-6d4c-4865-b127-6f5acd1df8c8_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:87203c4f1012d92651ce35ca67007cc965a18106c6e0fd2dbb04930d5a1f9404
+size 2959059

2024/Cryptographic Hardness of Score Estimation/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:348c0318c39647fad8c1bc591fc03fea9ec137394e6598cc19f99a7cac5d1266
+size 1169784

2024/Ctrl-X_ Controlling Structure and Appearance for Text-To-Image Generation Without Guidance/1041078b-a426-42ca-a128-147b63808e7e_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1db45ec10967a0c0b3d212cff1af9668b05d67ce871af95caf7283b9924cce65
+size 22683650

2024/Ctrl-X_ Controlling Structure and Appearance for Text-To-Image Generation Without Guidance/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:99f222d835e45d657d627ed6c579b09be1a2c38f361574ce8b0f1b17cdfeafa9
+size 2248814

2024/CuMo_ Scaling Multimodal LLM with Co-Upcycled Mixture-of-Experts/8e433d81-a0d1-4e53-9a22-fed61de1e57d_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a794287ae384bc35c5d569d3f53f5098646b89c1aebf4980355cf0add59eea92
+size 1184096

2024/CuMo_ Scaling Multimodal LLM with Co-Upcycled Mixture-of-Experts/full.md

@@ -0,0 +1,595 @@
+# CuMo: Scaling Multimodal LLM with Co-Upcycled Mixture-of-Experts
+
+Jiachen Li $^{1*}$  Xinyao Wang $^{2\dagger\dagger}$  Sijie Zhu $^{2}$  Chia-Wen Kuo $^{2}$  Lu Xu $^{2}$  Fan Chen $^{2}$  Jitesh Jain $^{1}$  Humphrey Shi $^{1\dagger}$  Longyin Wen $^{2}$ $^{1}$ SHI Labs @ Georgia Tech & UIUC  $^{2}$ ByteDance Inc., San Jose https://github.com/SHI-Labs/CuMo
+
+# Abstract
+
+Recent advancements in Multimodal Large Language Models (LLMs) have focused primarily on scaling by increasing text-image pair data and enhancing LLMs to improve performance on multimodal tasks. However, these scaling approaches are computationally expensive and overlook the significance of efficiently improving model capabilities from the vision side. Inspired by the successful applications of Mixture-of-Experts (MoE) in LLMs, which improves model scalability during training while keeping inference costs similar to those of smaller models, we propose CuMo, which incorporates Co-upcycled Top-K sparsely-gated Mixture-of-experts blocks into both the vision encoder and the MLP connector, thereby enhancing the multimodal LLMs with neglectable additional activated parameters during inference. CuMo first pre-trains the MLP blocks and then initializes each expert in the MoE block from the pre-trained MLP block during the visual instruction tuning stage, with auxiliary losses to ensure a balanced loading of experts. CuMo outperforms state-of-the-art multimodal LLMs across various VQA and visual-instruction-following benchmarks within each model size group, all while training exclusively on open-sourced datasets.
+
+# 1 Introduction
+
+The advent of GPT-4V [53] has sparked excitement within open-source communities to transform large language models (LLM) into multimodal LLMs. Recent multimodal LLMs [11, 46, 2] typically integrate pre-trained vision encoders with MLP connectors to LLMs, with visual instruction tuning data to fine-tune the pre-trained LLMs, enhancing their visual understanding capabilities. To further scale up multimodal LLMs, previous efforts [44, 45, 39, 51, 7, 42] primarily focus on training the model with a more extensive collection of text-image paired data and employing stronger LLMs, significantly increasing training efforts. On the vision side, recent work concentrates on leveraging multiple vision encoders [43, 18] to enrich visual content, employing larger vision encoders [9], and using advanced vision-language connectors [5] to improve performance on multimodal tasks. However, these techniques result in an increased number of additional parameters and generate extra visual tokens for LLMs to process, making it inefficient to scale.
+
+In terms of efficiently scaling up models, Mixture-of-Experts (MoE) has become the de-facto framework in modern large-scale neural networks, particularly in natural language processing (NLP). Most large language models (LLM) are built upon the transformer [64] architecture, wherein sparse
+
+![](images/35e5aa1b8c6fbcf557e2840900daee92cfdcda931040cba72f6316e61220ba25.jpg)  
+Figure 1: Left: Each MLP expert within the MoE block during the visual instruction tuning stage is initialized from the corresponding pre-trained MLP in CuMo. Right: CuMo outperforms strong open-sourced models such as Mini-Gemini and LLaVA-NeXT, as well as the private MM1 model.
+
+![](images/d9dad426b6476d0a278172c5a673c3f2e15eec663525b425541cac5c4d8aa411.jpg)
+
+MoE is used to replace the dense MLP block with the Top-K sparsely-gated MoE block [57]. Recent state-of-the-art open-sourced [28, 61] and private [55, 51] LLMs have predominantly adopted the sparse MoE architecture. These models are scaled up using the MoE design during training while maintaining relatively lower inference costs as only selected MLP experts are activated during the feed-forward process. Nevertheless, the development and optimization of MoE-based models have been largely tailored to LLMs, and the exploration of scaling multimodal LLMs with MoE, especially on the vision side, remains largely unexplored.
+
+Motivated by these observations, we introduce CuMo, which integrates Top-K sparsely-gated MoE blocks into the vision encoder and the MLP connector of multimodal LLMs. We also explore the associated training recipe and methodology for CuMo. Firstly, we pre-train the MLP connector and perform pre-finetuning to warm up the whole model without introducing the MoE architecture, which stabilizes the following visual instruction tuning stage with newly incorporated sparse MoE blocks. Then, we replace each MLP block with the sparse MoE block in the MLP connector and the vision encoder through co-upcycling. Each expert within the sparse MoE block is initialized from the corresponding MLP block after the pre-training and the pre-finetuning stages, as shown in Figure 1. Additionally, each MoE block contains a Top-K router trained from scratch to select experts during the visual instruction tuning stage with auxiliary losses on the router to maintain a balanced loading of experts. We conduct further comparisons between co-upcycled LLMs and pre-trained MoE-based LLMs. The results show that the pre-trained MoE-based LLMs significantly outperform the co-upcycled LLMs. As a result, the co-upcycling of LLMs is not included in CuMo. Our models are trained fully on open-sourced datasets that are converted to visual instruction following formats. Experimental results demonstrate that CuMo outperforms other state-of-the-art multimodal LLMs on various VQA and multimodal instruction-following benchmarks within the same model size group, as illustrated in Figure 1.
+
+Our contributions can be summarized as follows:
+
+- We introduce CuMo, which integrates co-upcycled sparsely-gated MoE layers into both the MLP connector and the vision encoder, enhancing the multimodal LLM with slightly additional activated parameters from the vision side.  
+- We outline the training methodology for CuMo, including a three-stage training process with auxiliary losses to stabilize training and ensure a balanced loading of experts.  
+- We train CuMo exclusively on open-sourced datasets and pre-trained models. It outperforms state-of-the-art open-sourced and private multimodal LLMs across multiple competitive benchmarks within each model size group.
+
+# 2 Related Works
+
+# 2.1 Multimodal LLM
+
+While the ultimate goal for multimodal models may be generative across various modalities [66, 3, 60], modern multimodal LLMs primarily focus on integrating additional modalities, such as vision, into LLMs. InstructBLIP [11] adopts Q-Former [37] to sample from visual tokens for LLM to feedforward and follow the instructions. Flamingo [1] and IDEFICS [23, 32] use shared decoder for visual-language understanding. Qwen-VL [2] uses three-stage training to convert QwenLM to QwenVL. LLaVA series [46, 44, 45] adopt visual instruction tuning that uses instruction-following data to convert LLM into multimodal LLM. ShareGPT4V [7] collects detailed image caption data from GPT4V to augment the LLaVA models. HoneyBee [5] investigates different designs of the MLP connector for better alignment. VILA [42] unfreezes the LLM during pre-training with interleaved image-text data. MoE-LLaVA [41] adopts the MoE design in small LLMs and reaches comparable performance to LLaVA with large LLMs. VCoder [26] adopts various vision adapters to enhance visual perception abilities. SPHINX [43, 18] adopts multiple visual encoders to enrich the visual features with scaled data and models. InternLM-Xcomposer [69, 12] is trained with interleaved text-image composition data and achieves state-of-the-art performance. InternVL [9] scales up the vision encoder to a 6B ViT model. MM1 [51] summarizes the essential steps towards building a strong multimodal LLM from a pre-trained LLM. Mini-Gemini [39] further collects guided generation into the pipeline.
+
+# 2.2 Mixture-of-Experts
+
+Mixture-of-Experts [24] is proposed to utilize a set of expert networks to address specific tasks by employing a gating network to determine the selection of these experts. Recently, it has gained popularity in the design of large language models [15]. The mainstream practice [57] is to replace the dense MLP layers with Top-K sparsely-gated mixture-of-experts (MoE) layers in the transformer [64]. MoE in Language Subsequent works [33, 16] have further scaled up MoE-based large language models with improved stability and load balancing of experts. The design of gating networks often involves selecting the top-k experts for each token [57, 33]. Various routing strategies have been explored, such as choosing top-k tokens by experts [71], one-to-one matching between experts and tokens [34]. Besides routing strategies, maintaining the load balance of experts is crucial for training MoE models. ST-MoE [73] adopts loading balancing loss and router-z loss to ensure a balanced distribution of the experts. Upcycling [31] proposes training sparse experts from dense checkpoints to stabilize training and lower the cost. Recent large language models like Gemini-Pro [55] and DBRX [61] are also based on the MoE design.
+
+MoE in Vision The success of MoE extends to the vision community, particularly following the popularity of vision transformers [13, 4, 72, 21, 20, 25, 36]. V-MoE [56] reaches comparable performance to dense ViT while only requiring half of the compute. LIMoE [52] replaces dense MLP layers with MoE layers in CLIP and observes improvements in zero-shot image classification. Residual MoE [65] incorporates residual design into MoE transformer and saves over  $30\%$  training cost. AdaMV-MoE [8] proposes an adaptive MoE framework for multi-task learning.
+
+# 3 Method
+
+In this section, we first review the sparse MoE block structure and the upcycling strategy utilized in previous studies. Subsequently, we describe how these sparsely-gated MoE blocks are integrated into each module of multimodal LLMs using co-upcycling strategies. Then, we introduce the three-stage training process and auxiliary loss functions employed to stabilize training and balance the loads of experts.
+
+# 3.1 Revisit Sparse MoE
+
+Sparse MoE Structure Previous mainstream practice [57] is to replace the dense MLP blocks with sparsely-gated mixture-of-experts blocks. Given input  $\mathbf{X} \in \mathbb{R}^{N \times C_{in}}$  and a MLP block,
+
+$$
+X _ {o u t} = \operatorname {M L P} (X) \in \mathbb {R} ^ {N \times C _ {o u t}} \tag {1}
+$$
+
+![](images/5af8be87179580bc01753121313134e830ac0b7d6bc30de8ae444417f3b2d276.jpg)  
+Figure 2: Architecture of CuMo. CuMo incorporates sparse Top-K MoE blocks into the CLIP vision encoder and vision-language MLP connector, thereby improving the multimodal LLM capabilities from the vision side. Skip connections are omitted for simplicity. Further implementation details are provided in Section 3.2.
+
+To scale up the model with multiple MLP blocks in parallel, a sparse MoE block includes a router network to select Top-K experts out of  $S$  total experts. This router network has a linear layer to compute the normalized weight matrix based on the inputs  $\mathbf{X}$  for voting, resulting in
+
+$$
+W = \operatorname {S o f t m a x} (\operatorname {L i n e a r} (X)) \in \mathbb {R} ^ {N \times S} \tag {2}
+$$
+
+The Top-K experts are selected for each token based on  $\mathbf{W}$ , and the re-normalized weights  $\mathbf{W}_{\mathbf{K}} \in \mathbb{R}^{N \times K}$  are computed using
+
+$$
+W _ {K} = \operatorname {S o f t m a x} (\operatorname {T o p K} (W)) \in \mathbb {R} ^ {N \times K} \tag {3}
+$$
+
+Each selected expert is represented by an MLP block, and the final output is obtained through a re-weighted sum
+
+$$
+X _ {o u t} = \sum_ {i} ^ {K} W _ {K} ^ {i} \circ \operatorname {M L P} _ {i} (X) \in \mathbb {R} ^ {N \times C _ {o u t}} \tag {4}
+$$
+
+the output  $\mathbf{X}_{out}$  maintains the same dimension as the output of a single dense MLP block.
+
+Sparse Upcycling Training MoE-based designs from scratch can be unstable and costly. Sparse Upcycling [31] addresses this challenge by initializing the experts in each MoE block from the corresponding MLP block in pre-trained dense checkpoints. This initialization approach provides a better starting point for training MoE-based models and reduces training costs compared to training from scratch.
+
+# 3.2 CuMo Architecture
+
+Sparse MoE in MLP Connector The MLP connector converts visual tokens into word embedding space, aligning dimensions between visual and text tokens. An effective architecture for the vision-language connector is an MLP block [44] that contains two linear layers. We start from a single MLP block and replace it with a Top-K sparse MoE block, incorporating a Top-K router and a set of experts for projecting visual tokens into word embedding space.
+
+Sparse MoE in Vision Encoder Vision encoders extract image features as sequences of visual tokens for reasoning in LLMs. CLIP [54] is one the most popular pre-trained vision encoders for multimodal LLM since it is pre-trained on large-scale image-text pairs, which makes it suitable for processing images for multimodal usage. The visual encoding part of CLIP is a ViT [13] model, which has consecutive MLP blocks in the transformer encoder. We substitute each MLP block with a Top-K sparse MoE block, retaining skip connections alongside MoE block outputs.
+
+![](images/f942c6834b811b85c66603d444c64a5f4e5453d27179c7cfe04a733a46c583da.jpg)  
+Figure 3: Training Stages of CuMo. The first stage involves pre-training the MLP for better alignment. Subsequently, the pre-finetuning stage trains all parameters as a warm-up before the next stage. Finally, the MLP experts within each MoE block are initialized from the weights of the corresponding MLP block, followed by training all parameters in the visual instruction tuning stage.
+
+Sparse MoE in LLM In terms of using MoE in LLM, we compare the co-upcycled LLM with pre-trained MoE-based LLM. We start from Mistral-7B and the upcycled Mistral-7B-MoE slightly outperforms Mistral-7B on certain benchmarks. However, considering the constrained knowledge base of upcycled experts from Mistral-7B, we compare it with the pre-trained Mixtral 8x7B with pre-trained experts of a diverse knowledge base. Experimental results reveal that pre-trained Mixtral 8x7B significantly outperforms Mistral-7B-MoE. As a result, LLM is not co-upcycled with CLIP and MLP connectors since it brings marginal improvements with great additional parameters.
+
+# 3.3 Training Recipe
+
+Co-Upcycling MoE blocks We start with training the added MoE blocks from scratch while the model is struggling to converge. Attempts to address this issue with lower learning rates perform worse compared to the baseline. As a result, we adopt a co-upcycling approach, initializing each module that integrates sparsely-gated MoE blocks with pre-trained MLPs to replace corresponding MLP blocks, as shown in Figure 1. This strategy consistently improves training stability and model performance.
+
+Three-Stage Training To further enhance training stability, we adopt a three-stage training strategy for CuMo models, as illustrated in Figure 3. In the first stage, we only pre-train the MLP connector, given that the vision encoder and LLM have already undergone pre-training on large-scale data. During the second pre-finetuning stage, we train all parameters using high-quality caption data to warm up the entire model before introducing MoE blocks in the subsequent stage. The third stage involves visual instruction finetuning, where the multimodal LLM is scaled up with upcycled MoE blocks and trained on visual instruction tuning data.
+
+Loss Function To maintain a load balance between experts in each MoE block, we adopt auxiliary losses based on the language modeling cross-entropy loss. The auxiliary losses comprise loading balance loss and router z-loss [73]. Hence, the total loss is
+
+$$
+L = L _ {c e} + \alpha_ {b} L _ {b} + \alpha_ {z} L _ {z} \tag {5}
+$$
+
+Here,  $L_{ce}$  represents the language modeling loss, which computes the cross-entropy of next-token predictions.  $\alpha_{b}$  and  $\alpha_{z}$  denote coefficients for loading balance loss  $L_{b}$  and router z-loss  $L_{z}$ , set to 0.1 and 0.01, respectively, across all experiments. These auxiliary losses, abbreviated as bzloss in Section 4, are individually applied to the MLP connector, vision encoder, and LLM for simplicity.
+
+# 4 Experiments
+
+We train the CuMo models on a mixture of open-sourced datasets, which are converted into the visual instruction tuning format. Then, we conduct comprehensive evaluations of the performance of CuMo models across various competitive VQA-based and instruction-following-based benchmarks. Additionally, we perform ablation studies on each module with upcycled MoE blocks with qualitative analysis of the results.
+
+<table><tr><td>Method</td><td>LLM</td><td>Act. (B)</td><td>SQA IMG</td><td>Text VQA</td><td>GQA</td><td>POPE</td><td>MME</td><td>MMB EN</td><td>CN</td><td>MM Vet</td><td>VQA v2</td><td>LLaVA Wild</td><td>SEED IMG</td><td>MMMU val</td></tr><tr><td colspan="15">7B to 13B Models</td></tr><tr><td>InstructBLIP [11]</td><td>Vicuna-7B</td><td>7.9</td><td>60.5</td><td>50.1</td><td>49.2</td><td>-</td><td>-</td><td>36.0</td><td>23.7</td><td>26.2</td><td>-</td><td>60.9</td><td>60.5</td><td>-</td></tr><tr><td>Qwen-VL-Chat [2]</td><td>Qwen-7B</td><td>-</td><td>68.2</td><td>61.5</td><td>57.5</td><td>-</td><td>1487.5</td><td>60.6</td><td>56.7</td><td>-</td><td>78.2</td><td>-</td><td>58.2</td><td>35.9</td></tr><tr><td>LLaVA-v1.5 [44]</td><td>Vicuna-7B</td><td>7.1</td><td>66.8</td><td>58.2</td><td>62.0</td><td>85.9</td><td>1510.7</td><td>64.3</td><td>58.3</td><td>30.5</td><td>78.5</td><td>63.4</td><td>66.1</td><td>-</td></tr><tr><td>VILA [42]</td><td>Vicuna-7B</td><td>7.1</td><td>68.2</td><td>64.4</td><td>62.3</td><td>85.5</td><td>1533.0</td><td>68.9</td><td>61.7</td><td>34.9</td><td>79.9</td><td>69.7</td><td>61.1</td><td>-</td></tr><tr><td>ShareGPT4V [7]</td><td>Vicuna-7B</td><td>7.1</td><td>68.4</td><td>-</td><td>-</td><td>-</td><td>1567.4</td><td>68.8</td><td>62.2</td><td>37.6</td><td>80.6</td><td>72.6</td><td>69.7</td><td>-</td></tr><tr><td>LLaMA-VID [38]</td><td>Vicuna-7B</td><td>-</td><td>68.3</td><td>-</td><td>64.3</td><td>86.0</td><td>1521.4</td><td>65.1</td><td>-</td><td>-</td><td>79.3</td><td>-</td><td>59.9</td><td>-</td></tr><tr><td>SPHINX-Intern2 [18]</td><td>InternLM2-7B</td><td>-</td><td>70.4</td><td>58.1</td><td>56.2</td><td>86.9</td><td>1260.4</td><td>57.9</td><td>-</td><td>36.5</td><td>75.5</td><td>57.6</td><td>68.8</td><td>-</td></tr><tr><td>LLaVA-NeXT [45]</td><td>Mistral-7B</td><td>7.6</td><td>72.8</td><td>65.7</td><td>64.8</td><td>86.7</td><td>1498</td><td>68.7</td><td>61.2</td><td>47.3</td><td>82.2</td><td>83.2</td><td>72.2</td><td>35.3</td></tr><tr><td>LLaVA-NeXT [45]</td><td>Vicuna-7B</td><td>7.1</td><td>70.1</td><td>64.9</td><td>64.2</td><td>86.5</td><td>1519</td><td>67.4</td><td>60.6</td><td>43.9</td><td>81.8</td><td>81.6</td><td>70.2</td><td>35.8</td></tr><tr><td>Mini-Gemini [39]</td><td>Vicuna-7B</td><td>7.3</td><td>65.2</td><td>-</td><td>-</td><td>-</td><td>1523</td><td>69.3</td><td>-</td><td>40.8</td><td>-</td><td>-</td><td>-</td><td>36.1</td></tr><tr><td>MM1 [51]</td><td>MM1-7B</td><td>-</td><td>72.6</td><td>72.8</td><td>-</td><td>86.6</td><td>1529.3</td><td>79.0</td><td>-</td><td>42.1</td><td>82.8</td><td>81.5</td><td>69.9</td><td>37.0</td></tr><tr><td>InstructBLIP [11]</td><td>Vicuna-13B</td><td>14.2</td><td>63.1</td><td>50.7</td><td>49.5</td><td>78.9</td><td>1212.8</td><td>-</td><td>-</td><td>25.6</td><td>-</td><td>58.2</td><td>63.1</td><td>-</td></tr><tr><td>LLaVA-v1.5 [44]</td><td>Vicuna-13B</td><td>13.4</td><td>71.6</td><td>61.3</td><td>63.3</td><td>85.9</td><td>1531.3</td><td>67.7</td><td>63.6</td><td>35.4</td><td>80.0</td><td>70.7</td><td>68.2</td><td>36.4</td></tr><tr><td>VILA [42]</td><td>Vicuna-13B</td><td>13.4</td><td>73.7</td><td>66.6</td><td>63.3</td><td>84.2</td><td>1570.1</td><td>70.3</td><td>64.3</td><td>38.8</td><td>80.8</td><td>73.0</td><td>62.8</td><td>-</td></tr><tr><td>InternVL-Chat [9]</td><td>Vicuna-13B</td><td>19</td><td>-</td><td>61.5</td><td>66.6</td><td>87.6</td><td>1586.4</td><td>-</td><td>-</td><td>-</td><td>81.2</td><td>-</td><td>-</td><td>-</td></tr><tr><td>LLaMA-VID [38]</td><td>Vicuna-13B</td><td>-</td><td>70.0</td><td>-</td><td>65.0</td><td>86.0</td><td>1542.3</td><td>66.6</td><td>-</td><td>-</td><td>80.0</td><td>-</td><td>62.3</td><td>-</td></tr><tr><td>SPHINX-Plus [18]</td><td>LLaMA2-13B</td><td>-</td><td>74.2</td><td>65.7</td><td>-</td><td>89.1</td><td>1457.7</td><td>71.0</td><td>-</td><td>47.9</td><td>-</td><td>71.7</td><td>74.8</td><td>-</td></tr><tr><td>Mini-Gemini[39]</td><td>Vicuna-13B</td><td>13.6</td><td>65.9</td><td>-</td><td>-</td><td>-</td><td>1565</td><td>68.5</td><td>-</td><td>46.0</td><td>-</td><td>-</td><td>-</td><td>38.1</td></tr><tr><td>LLaVA-NeXT [45]</td><td>Vicuna-13B</td><td>13.4</td><td>73.6</td><td>67.1</td><td>65.4</td><td>86.2</td><td>1575</td><td>70</td><td>64.4</td><td>48.4</td><td>82.8</td><td>87.3</td><td>71.9</td><td>36.2</td></tr><tr><td>CuMo</td><td>Mistral-7B</td><td>7.8</td><td>73.9</td><td>67.0</td><td>64.9</td><td>86.7</td><td>1548.6</td><td>73.0</td><td>66.6</td><td>51.0†</td><td>82.2</td><td>85.7†</td><td>72.1</td><td>39.1</td></tr><tr><td colspan="15">7B MoE Models</td></tr><tr><td>SPHINX-MoE [18]</td><td>Mixtral-8×7B</td><td>-</td><td>74.5</td><td>68.0</td><td>63.8</td><td>89.6</td><td>1485.3</td><td>71.3</td><td>-</td><td>40.9</td><td>81.1</td><td>70.2</td><td>73.0</td><td>31.1</td></tr><tr><td>MM1 [51]</td><td>MM1-7B-MoE</td><td>-</td><td>75.3</td><td>72.8</td><td>-</td><td>87.6</td><td>1629.0</td><td>79.7</td><td>-</td><td>47.0</td><td>83.4</td><td>82.0</td><td>70.4</td><td>40.9</td></tr><tr><td>Mini-Gemini [39]</td><td>Mixtral-8×7B</td><td>13.5</td><td>-</td><td>69.2</td><td>-</td><td>-</td><td>1639</td><td>75.6</td><td>-</td><td>45.8</td><td>-</td><td>-</td><td>-</td><td>41.8</td></tr><tr><td>CuMo</td><td>Mixtral-8×7B</td><td>13.5</td><td>77.9</td><td>66.0</td><td>63.8</td><td>85.7</td><td>1639.5</td><td>75.3</td><td>68.0</td><td>48.7†</td><td>81.8</td><td>84.7†</td><td>73.2</td><td>45.0</td></tr></table>
+
+Table 1: Comparisons between CuMo and other state-of-the-art multimodal LLMs on competitive benchmarks. These models are grouped by the size of the base LLM and bold indicates the best performance on a certain benchmark. Act.: activated parameters during inference. Numbers with  $\dagger$  are averaged by three inference runs of querying GPT API.
+
+# 4.1 Implementation Details
+
+Training Datasets During pre-training, we only utilize LLaVA-558K [46] to train the MLP connector for better alignment. In the subsequent pre-finetuning stage, detailed image caption data from ALLaVA [6] is employed to warm up all parameters of the multimodal LLM. For the final visual instruction tuning stage, a mixture of datasets including LLaVA-665K [44], ShareGPT4V [7], LAION-GPT-V [14], DocVQA [62], ChartQA [49], AI2D [29], InfoVQA [50], SynDog-EN [30], ALLaVA [6], and LIMA [70] is utilized to train the CuMo models with upcycled MoE blocks. The total data size for visual instruction tuning is approximately 1.65 million, and all training data are publicly accessible. The detailed breakdown of the training dataset is listed in Appendix A.
+
+Evaluation Benchmarks Evaluation of CuMo models primarily focuses on academic VQA-based datasets such as VQAv2 [19], GQA [22], Science-QA [48], and TextVQA [59], as well as instruction-following-based LMM benchmarks including POPE [40], MME [17], MMBench [47], SEED-Bench [35], LLaVA-Wild [46], and MM-Vet [67]. Additionally, the challenging MMMU [68] is evaluated to assess the visual reasoning abilities of the multimodal LLMs.
+
+Training Settings We employ the pre-trained CLIP ViT-L [54] as the vision encoder, a two-layer MLP as the vision-language connector, and Mistral-7B [27] as the LLM to establish the baseline model following LLaVA v1.5 [44]. We only use LLaVA-558K [44] as pre-training data and LLaVA-665K [44] as visual instruction tuning data to train the baseline model and make ablation studies for comparisons. The learning rate is set to 1e-3 for pre-training the MLP connector and reduced to 2e-5 for visual instruction tuning of both the MLP connector and CLIP. To further stabilize the visual instruction tuning process after scaling up with additional data, the learning rate is lowered to 2e-6 for all parameters of the CuMo models in the final results. More hyperparameters of the training process is listed in Appendix B.
+
+Evaluation Settings During evaluation, we adhere to the settings outlined in the LLaVA series [44], employing a greedy decoding strategy for all benchmarks. The data and questions are converted into visual instructions to prompt the multimodal LLMs. For benchmarks that utilize GPT API for evaluation, we adopt gpt-4-0613 for LLaVA-Wild [46].
+
+<table><tr><td>Method</td><td>LLM</td><td>PT</td><td>IT</td><td>SQA IMG</td><td>Text VQA</td><td>GQA</td><td>POPE</td><td>MME</td><td>MMBench EN</td><td>CN</td><td>MM Vet</td><td>VQA v2</td><td>LLvA Wild</td><td>SEED IMG</td></tr><tr><td>InstructBLIP [11]</td><td>Vicuna-7B</td><td>129M</td><td>1.2M</td><td>60.5</td><td>50.1</td><td>49.2</td><td>-</td><td>-</td><td>36.0</td><td>23.7</td><td>26.2</td><td>-</td><td>60.9</td><td>60.5</td></tr><tr><td>InstructBLIP [11]</td><td>Vicuna-13B</td><td>129M</td><td>1.2M</td><td>63.1</td><td>50.7</td><td>49.5</td><td>78.9</td><td>1212.8</td><td>-</td><td>-</td><td>25.6</td><td>-</td><td>58.2</td><td>63.1</td></tr><tr><td>IDEFICS-9B [23]</td><td>LLaMA-7B</td><td>353M</td><td>1M</td><td>-</td><td>25.9</td><td>38.4</td><td>-</td><td>-</td><td>48.2</td><td>25.2</td><td>-</td><td>50.9</td><td>-</td><td>-</td></tr><tr><td>IDEFICS-80B [23]</td><td>LLaMA-65B</td><td>353M</td><td>1M</td><td>-</td><td>30.9</td><td>45.2</td><td>-</td><td>-</td><td>54.5</td><td>38.1</td><td>-</td><td>60.0</td><td>-</td><td>-</td></tr><tr><td>Qwen-VL [2]</td><td>Qwen-7B</td><td>1.4B</td><td>50M</td><td>67.1</td><td>63.8</td><td>59.3</td><td>-</td><td>-</td><td>38.2</td><td>7.4</td><td>-</td><td>78.8</td><td>-</td><td>56.3</td></tr><tr><td>Qwen-VL-Chat [2]</td><td>Qwen-7B</td><td>1.4B</td><td>50M</td><td>68.2</td><td>61.5</td><td>57.5</td><td>-</td><td>1487.5</td><td>60.6</td><td>56.7</td><td>-</td><td>78.2</td><td>-</td><td>58.2</td></tr><tr><td>LLvA-v1.5 [44]</td><td>Vicuna-7B</td><td>558K</td><td>665K</td><td>66.8</td><td>58.2</td><td>62.0</td><td>85.9</td><td>1510.7</td><td>64.3</td><td>58.3</td><td>30.5</td><td>78.5</td><td>63.4</td><td>66.1</td></tr><tr><td>\(LLvA-v1.5^{\dagger}\)</td><td>Mistral-7B</td><td>558K</td><td>665K</td><td>72.8</td><td>57.6</td><td>60.0</td><td>86.3</td><td>1414.9</td><td>66.5</td><td>60.1</td><td>32.1</td><td>78.2</td><td>69.4</td><td>66.4</td></tr><tr><td>CuMo</td><td>Mistral-7B</td><td>558K</td><td>665K</td><td>71.7</td><td>59.3</td><td>63.2</td><td>87.1</td><td>1428.6</td><td>69.6</td><td>62.6</td><td>34.3</td><td>80.6</td><td>68.8</td><td>69.6</td></tr></table>
+
+Table 2: Comparisons between CuMo Mistral-7B and other multimodal LMM models with limited training data. The best performance are highlighted in **bold**. LLaVA-v1.5† with Mistral-7B is reproduced by us as a baseline model.
+
+# 4.2 Main Results
+
+Comparison with SoTA Multimodal LLMs In Table 1, we present a comparison of CuMo models with other state-of-the-art instruction-following-based multimodal LLMs. We categorize the models based on the size of the base LLMs, including 7B models, 13B models, and 7B MoE models. CuMo Mistral-7B outperforms other 7B-based state-of-the-art multimodal LLMs across multiple benchmarks. Moreover, the performance of the CuMo Mistral-7B model is comparable to many 13B-based multimodal LLMs. In the case of Mixtral-  $8\times 7\mathrm{B}$  models, CuMo achieves results on par with SPHINX-MoE, MM1, and Mini-Gemini. LLaMA-based LLMs [10, 63] are not utilized in our experiments due to license constraints.
+
+Comparison under limited training data To further evaluate the effectiveness of the co-upcycled MoE blocks, we train the vanilla CuMo mistral-7B under limited training data in Table 2. It shows that CuMo outperforms other 7B models and reaches comparable performance to LLaVA-v1.5 Vicuna-13B under the same training data.
+
+# 4.3 Ablation Study
+
+Upcycle MLP connector to MLP-MoE We initiate the ablation study by replacing the MLP connector with upcycled MLP-MoE, as depicted in Table 3(a). We start with a Top 2-in-4 router and train the MoE blocks from scratch, which leads to a clear performance drop on all benchmarks. Then, we adopt the upcycling strategy to initialize the MLP experts. We observe marginal improvements over the baseline, considering each expert comprises only two linear layers. Subsequently, the incorporation of bzloss to ensure a balanced loading of experts in the MLP-MoE yields noticeable enhancements on MMVet. However, employing a Top 2-in-8 router with upcycling and bzloss results in a slight performance decline, possibly due to the limited visual instruction tuning data to train robust and well-balanced eight experts.
+
+Empower CLIP with CLIP-MoE In Table 3(b), initially unfreezing CLIP based on MLP-MoE leads to noticeable improvements on TextVQA and MM Vet benchmarks. However, training with the added Top2-in-4 MoE blocks in CLIP from scratch proves unsuccessful, as the model fails to converge even with largely reduced learning rates. Consequently, adopting upcycled MoE blocks during the visual instruction tuning stage yields further enhancements on the TextVQA, MM Vet, and SEED benchmarks, as well as a more stable training process.
+
+Upcycle LLM vs Pre-trained LLM-MoE Upon replacing all MLP blocks with sparsely-gated MoE blocks in the visual part, we further investigate the utilization of the MoE architecture in the LLM. Starting from the Mistral-7B model, we first lower the learning rate to 2e-6 to set the baseline and the following experiments since a learning rate of 2e-5 induces training instabilities. Then, we upcycle each MLP block with a sparsely-gated MoE block, initializing the weight of each expert from the pre-trained MLP block. As demonstrated in Table 3(c), the upcycled Mistral-4×7B and 8×7B outperform the Mistral-7B model slightly except for TextVQA. However, considering that the upcycled experts significantly increase parameters without introducing new knowledge, we replace the upcycled Mistral 8×7B with Mixtral 8×7B [28]. In Mixtral 8×7B, all expert layers are pre-trained on large-scale language data, providing superior initialization and similar training stability compared to upcycling. The results indicate that CuMo Mixtral-8x7B outperforms its upcycled
+
+<table><tr><td>Method</td><td>SQA</td><td>\( VQAT^T \)</td><td>MMVet</td><td>SEED</td></tr><tr><td>Baseline on Mistral-7B</td><td>72.8</td><td>57.6</td><td>32.1</td><td>66.4</td></tr><tr><td>+ Top 2-in-4 &amp; Scratch</td><td>68.1</td><td>55.6</td><td>29.3</td><td>65.1</td></tr><tr><td>\(\rightleftharpoons\)Top 2-in-4 &amp; Upcycle</td><td>73.7</td><td>57.2</td><td>32.3</td><td>67.1</td></tr><tr><td>+bzloss</td><td>73.5</td><td>57.4</td><td>33.1</td><td>67.4</td></tr><tr><td>\(\rightleftharpoons\)Top 2-in-8 &amp; Upcycle</td><td>73.4</td><td>57.6</td><td>32.4</td><td>67.2</td></tr></table>
+
+(a) MLP-MoE  
+
+<table><tr><td>Method</td><td>SQA</td><td>VQAT</td><td>MM Vet</td><td>SEED</td></tr><tr><td>MLP-MoE</td><td>73.5</td><td>57.4</td><td>33.1</td><td>67.4</td></tr><tr><td>+ Unfreeze CLIP</td><td>72.0</td><td>58.9</td><td>34.7</td><td>69.0</td></tr><tr><td>+ Top 2-in-4 &amp; bzloss</td><td>72.8</td><td>59.7</td><td>35.4</td><td>69.8</td></tr><tr><td>\(\rightleftharpoons\)Top 2-in-8 &amp; bzloss</td><td>71.0</td><td>59.0</td><td>33.6</td><td>69.2</td></tr></table>
+
+(b) CLIP-MoE  
+
+<table><tr><td>Method</td><td>SQA</td><td>\( VQAT \)</td><td>MM Vet</td><td>SEED</td></tr><tr><td>MLP-MoE &amp; CLIP-MoE</td><td>72.8</td><td>59.7</td><td>35.4</td><td>69.8</td></tr><tr><td>+ lower lr to 2e-6</td><td>71.7</td><td>59.3</td><td>34.3</td><td>69.6</td></tr><tr><td>+ Mistral 4×7B &amp; Upcycle</td><td>72.8</td><td>57.0</td><td>35.2</td><td>69.9</td></tr><tr><td>= Mistral 8×7B &amp; Upcycle</td><td>73.2</td><td>56.4</td><td>35.7</td><td>70.5</td></tr><tr><td>= Mixtral 8×7B</td><td>74.2</td><td>60.6</td><td>40.0</td><td>72.6</td></tr></table>
+
+(c) LLM-MoE  
+
+<table><tr><td>1×</td><td>2×</td><td>3×</td><td>SQA</td><td>VQAT</td><td>MM Vet</td><td>SEED</td></tr><tr><td>✓</td><td>-</td><td>-</td><td>71.7</td><td>59.3</td><td>34.3</td><td>69.6</td></tr><tr><td>✓</td><td>✓</td><td>-</td><td>71.7</td><td>60.6</td><td>35.0</td><td>69.7</td></tr><tr><td>✓</td><td>-</td><td>✓</td><td>72.9</td><td>61.0</td><td>37.0</td><td>69.7</td></tr><tr><td>✓</td><td>✓</td><td>✓</td><td>72.2</td><td>60.5</td><td>36.9</td><td>70.1</td></tr></table>
+
+(d) Multi-resolution Feature  
+
+<table><tr><td>Method</td><td>SQA</td><td>VQAT</td><td>MM Vet</td><td>SEED</td></tr><tr><td>No PFT</td><td>71.7</td><td>59.3</td><td>34.3</td><td>69.6</td></tr><tr><td>+ ShareGPT4V</td><td>72.4</td><td>61.7</td><td>36.5</td><td>70.0</td></tr><tr><td>⇌ALLaVA</td><td>73.0</td><td>62.8</td><td>37.2</td><td>70.9</td></tr></table>
+
+(e) Pre-FineTuning Stage  
+
+<table><tr><td>Method</td><td>CLIP</td><td>MLP</td><td>LLM</td><td>Total</td></tr><tr><td>Mistral-7B</td><td>0.30</td><td>0.025</td><td>7.25</td><td>7.58</td></tr><tr><td>+ MLP-MoE</td><td>0.30</td><td>0.05</td><td>7.25</td><td>7.60</td></tr><tr><td>+ CLIP-MoE</td><td>0.50</td><td>0.05</td><td>7.25</td><td>7.80</td></tr><tr><td>⇌ Mixtral-8x7B</td><td>0.50</td><td>0.05</td><td>12.90</td><td>13.45</td></tr></table>
+
+(f) Activated billions of parameters
+
+Table 3: Ablation Studies during building CuMo. Each row represents a different configuration, with changes or additions marked using  $\rightleftharpoons$  and  $^+$  symbols, respectively. Settings highlighted with a light blue background are those adapted for final model in Table 1. For (b): all MoE blocks in CLIP are initialized with upcycling.
+
+counterparts significantly and is employed in the final models with bzloss to maintain a balanced loading of experts.
+
+Multi-Resolution Visual Features Incorporating multi-resolution inputs is crucial for enhancing the understanding of image content in multimodal LLMs. Following the approach outlined in  $S^2$  [58], we introduce multi-resolution inputs to CLIP and concatenate the feature maps channel-wise to maintain the total number of visual tokens consistent with low-resolution inputs. As illustrated in Table 3(d), an empirical combination of  $3 \times$  and  $1 \times$  reaches the best performance and we adopt this configuration for the final CuMo models.
+
+Pre-FineTuning Stage Previous ablation studies were conducted directly after the pre-training of the MLP connector, leading to observed training instabilities during visual instruction tuning. To address this, we introduce a pre-finetuning stage using high-quality image caption data, wherein all parameters are unfrozen. In Table 3(e), we leverage caption data from ALLaVA for this stage. Results indicate that ALLaVA data proves to be a superior option, providing fewer but higher-quality captions for training, ultimately leading to improved performance.
+
+Added Parameters In Table 3(f), we keep track of the added activated parameters during inference of CuMo. It shows that adding MoE blocks in the vision side upon MLP connector and CLIP only brings 0.22B extra parameters compared to the baseline model on Mistral-7B. More details can be found in Appendix C.
+
+# 4.4 More Analysis
+
+Expert Distribution As shown in Figure 4, we visualize the expert distributions in the MoE block from selected layers in CLIP during inference. The dataset used for analysis is the test set of the MME benchmark. The distribution indicates that the selected experts during infer
+
+![](images/6742270c178d524f5008be6c1453d7a93ee85b14a85512e65e9f661c6062181d.jpg)  
+Figure 4: Expert distributions of MoE blocks in CLIP. We select layers from CLIP and summarize the activated experts during inference.
+
+![](images/462bd3a2930c7cfbd26e13c569f4383deae22e2eea0dc6ea94b04cce32e903af.jpg)  
+can you describe what groceries are presented in this fridge  
+Figure 5: Dialogues between the user and multimodal LLMs on challenging images. We highlight the correct answers and hallucinations from the responses of the multimodal LLMs.
+
+LLaVA-Yi-348 The refrigerator in the image is well-stocked with a variety of groceries. Here's a list of what I can see: on the top shelf, there are several items including what appears to be a head of lettuce, a bunch of green onions, a container of strawberries, and a few blue plastic containers that could be used for food storage or serving; the middle shelf has a few bottles, which might contain milk or juice, and a carton of eggs; the bottom shelf is filled with fresh produce. There are bell peppers in different colors (red, yellow, and green), a bunch of spring onions, and a head of lettuce; the drawer at the bottom contains more fresh produce, including what looks like a bunch of green onions and possibly some other leafy greens. The refrigerator is well-organized and appears to be stocked with fresh and healthy food options.  
+The refrigerator is filled with a variety of groceries, including fruits, vegetables, dairy products, and beverages. There are several bottles, likely containing milk or other dairy products, and a carton of eggs. The fridge also contains a selection of fresh produce, such as broccoli, peppers, and lettuce. Additionally, there are some fruits, including apples and oranges. The presence of these items suggests that the person who owns the fridge prefers a balanced diet with a mix of healthy foods and beverages.  
+CuMo-Mistral-7B The groceries presented in the fridge include a variety of fresh produce such as bell peppers, lettuce, broccoli, and cucumbers. There are also eggs,milk,and cheese,which are dairy products.Additionally,there are cherry tomatoes and a jar of red sauce,which could be used for cooking or as a condiment. The presence of these items suggests that the fridge is well-stocked with a balance of fresh vegetables,dairy,and condiments, indicating a household that values a healthy and varied diet.  
+LLaVA-Yi-34B The image you've provided is from the animated movie "Coco." It features a variety of characters, both human and fantastical, which are not real people but animated characters. The central figure is a young boy named Miguel, who is holding a guitar. Surrounding him are other characters, including his family members and other characters from the film. The characters are depicted in a stylized, animated manner, and the image is still from the movie.  
+MiniGemini-Yi-34B In the image, there are a total of 12 people. This includes the boy in the foreground, the two skeletons, and the nine characters seated around the table. It's important to note that the skeletons are representations of people and are included in the count.  
+CuMo-Mistral-7B In the image, there are a total of 14 characters. This includes 12 characters seated around the table and 2 characters standing on the table. The characters are a mix of human and skeletal figures, with the skeleton figures being part of the Day of the Dead celebration.  
+This poster is for the movie "Avengers: Age of Ultron," which is a 2015 American superhero film based on the Marvel Comics superhero team the Avengers. The film is the sequel to "The Avengers" (2012) and the 11th film in the Marvel Cinematic Universe (MCU). It was directed by Joss Whedon and produced by Marvel Studios. The poster features several of the main characters from the film, including Iron Man, Thor, Captain America, the Hulk, Black Widow, and Hawkeye, all in their superhero costumes. The setting appears to be a cityscapes with a sense of destruction, suggesting a battle or conflict. The title "Avengers: Age of Ultron" is prominently displayed at the top of the poster, and the Marvel logo is visible at the bottom. The film's plot revolves around the Avengers attempting to stop the villain Ultron from enacting his plans for human extinction. The movie is known for its action-packed sequences, special effects, and ensemble cast of superheroes. It was a critical and commercial success, contributing to the ongoing popularity of the Marvel Cinematic Universe.  
+The movie poster is for "The Avengers" a 2012 American superhero film produced by Marvel Studios and distributed by Walt Disney Pictures. It is the sixth film in the Marvel Cinematic Universe (MCU). The film features an ensemble cast including Robert Downey Jr., Chris Evans, Mark Ruffalo, Chris Hensworth, Scarlett Johansson, Jeremy Renner, Tom Hildeston, and Samuel L Jackson. The plot revolves around Nick Fury, director of the peacekeeping organization S.H.I.E.L.D., recruiting a team of superheroes to help save the world from Loki, the adoptive brother of Thor, who is leading an alien invasion. The poster is designed to showcase the main characters and the action-packed nature of the film, with the characters in their iconic costumes and the cityscapes in the background suggesting the scale of the conflict. The release date is prominently displayed, indicating the film's availability in theaters. The poster also highlights the film's availability in various formats, including 3D, which was a significant selling point at the time of its release.  
+The poster is for the movie "The Avengers," which is part of the Marvel Cinematic Universe. It features a team of superheroes, each with unique abilities and costumes, coming together to face a significant threat. The central figure, with the iconic red and gold suit, is likely the leader of the group. The presence of the Hulk and Thor suggests that the film includes a mix of science fiction and fantasy elements. The cityscpe and the explosion in the background imply that the movie's plot involves a large-scale conflict that impacts the urban environment. The release date and the mention of "3D and IMAX 3D" indicate that the movie was available in various formats, catering to different viewing preferences. The poster is designed to convey the epic scale and the high stakes of the film, as well as the diverse team of heroes that viewers can expect to see on screen.
+
+ence are evenly spread across layers, providing further evidence of the effectiveness of the auxiliary losses in maintaining load balance.
+
+Dialogue Comparisons Presented in Figure 5, we contrast the responses from CuMo-Mistral-7B, LLaVA-Yi-34B, and MiniGemini-Yi-34B under challenging content understanding cases. It demonstrates that CuMo-Mistral-7B can effectively follow instructions and provide mostly correct answers to challenging questions derived from complex scenes. However, CuMo also exhibits instances of hallucinations, such as responding with "2 characters standing on the table", highlighting the need for further investigation to mitigate hallucinations and improve reliability of CuMo.
+
+Limitations The main limitation of CuMo is that, similarly to other large language models, it can generate hallucinated responses. This may constrain its potentials in real-world multimodal applications like used as a chatbot. Future works, such as Reinforcement Learning with Human Feedback (RLHF) and Retrieval Augmented Generation (RAG), can be undertaken to mitigate these hallucinations and improve the model's reliability.
+
+# 5 Conclusion
+
+In this study, we introduce the sparse mixture-of-experts design into multimodal LLMs from the vision side. Specifically, we replace each MLP block with a Top-K sparse MoE block in the MLP connector and the vision encoder. To enhance training stability, we employ a three-stage training approach, incorporating upcycled MoE blocks during the visual instruction tuning stage, along with auxiliary bzloss to maintain a balanced loading of experts. All CuMo models are trained and evaluated on fully open-sourced datasets and benchmarks. Through extensive experiments and ablation studies, we validate the effectiveness of the upcycled MoE blocks in each module. CuMo outperforms state-of-the-art models across multiple competitive benchmarks within the same group of model sizes.
+
+Acknowledgments We extend our gratitude to Chunyuan Li, Lei Chen, and Haibin Lin for their insightful and valuable discussions throughout this project. Li, Jain, Shi are in part supported by National Science Foundation CAREER Award #2427478, and by National Science Foundation and the Institute of Education Sciences, U.S. Department of Education under Award #2229873 - National AI Institute for Exceptional Education, Beckman Institute and ECE Department at UIUC, and Georgia Institute of Technology.
+
+# References
+
+[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in neural information processing systems, 35:23716-23736, 2022. 3  
+[2] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv preprint arXiv:2308.12966, 2023. 1, 3, 6, 7  
+[3] Fan Bao, Shen Nie, Kaiwen Xue, Chongxuan Li, Shi Pu, Yaole Wang, Gang Yue, Yue Cao, Hang Su, and Jun Zhu. One transformer fits all distributions in multi-modal diffusion at scale. In International Conference on Machine Learning (ICML), pages 1692-1717. PMLR, 2023. 3  
+[4] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision. Springer, 2020. 3  
+[5] Junbum Cha, Wooyoung Kang, Jonghwan Mun, and Byungseok Roh. Honeybee: Locality-enhanced projector for multimodal llm. arXiv preprint arXiv:2312.06742, 2023. 1, 3  
+[6] Guiming Hardy Chen, Shunian Chen, Ruifei Zhang, Junying Chen, Xiangbo Wu, Zhiyi Zhang, Zhihong Chen, Jianquan Li, Xiang Wan, and Benyou Wang. Allava: Harnessing gpt4v-synthesized data for a lite vision-language model. arXiv:2402.11684, 2024. 6  
+[7] Lin Chen, Jinsong Li, Xiaoyi Dong, Pan Zhang, Conghui He, Jiaqi Wang, Feng Zhao, and Dahua Lin. Sharegpt4v: Improving large multi-modal models with better captions, 2023. 1, 3, 6  
+[8] Tianlong Chen, Xuxi Chen, Xianzhi Du, Abdullah Rashwan, Fan Yang, Huizhong Chen, Zhangyang Wang, and Yeqing Li. Adamv-moe: Adaptive multi-task vision mixture-of-experts. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 17346-17357, October 2023. 3  
+[9] Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, Bin Li, Ping Luo, Tong Lu, Yu Qiao, and Jifeng Dai. Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. arXiv preprint arXiv:2312.14238, 2023. 1, 3, 6  
+[10] WL Chiang, Z Li, Z Lin, Y Sheng, Z Wu, H Zhang, L Zheng, S Zhuang, Y Zhuang, JE Gonzalez, et al. Vicuna: An open-source chatbot impressing gpt-4 with  $90\%$  * chatgpt quality, mar. 2023. 7  
+[11] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven Hoi. Instructclip: Towards general-purpose vision-language models with instruction tuning. arXiv preprint arXiv:2305.06500, 2023. 1, 3, 6, 7
+
+[12] Xiaoyi Dong, Pan Zhang, Yuhang Zang, Yuhang Cao, Bin Wang, Linke Ouyang, Songyang Zhang, Haodong Duan, Wenwei Zhang, Yining Li, et al. Internlm-xcomposer2-4khd: A pioneering large vision-language model handling resolutions from 336 pixels to 4k hd. arXiv preprint arXiv:2404.06512, 2024. 3  
+[13] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 3, 4  
+[14] LAION eV. Laion/gpt4v-dataset  $\cdot$  datasets at hugging face. 6  
+[15] William Fedus, Jeff Dean, and Barret Zoph. A review of sparse expert models in deep learning. arXiv preprint arXiv:2209.01667, 2022.3  
+[16] William Fedus, Barret Zoph, and Noam Shazeer. Switch transformers: Scaling to trillion parameter models with simple and efficient sparsity, 2022. 3  
+[17] Chaoyou Fu, Peixian Chen, Yunhang Shen, Yulei Qin, Mengdan Zhang, Xu Lin, Zhenyu Qiu, Wei Lin, Jinrui Yang, Xiawu Zheng, et al. Mme: A comprehensive evaluation benchmark for multimodal large language models. arXiv:2306.13394, 2023. 6  
+[18] Peng Gao, Renrui Zhang, Chris Liu, Longtian Qiu, Siyuan Huang, Weifeng Lin, Shitian Zhao, Shijie Geng, Ziyi Lin, Peng Jin, Kaipeng Zhang, Wenqi Shao, Chao Xu, Conghui He, Junjun He, Hao Shao, Pan Lu, Hongsheng Li, and Yu Qiao. *Sphinx-x: Scaling data and parameters for a family of multi-modal large language models.* ArXiv, abs/2402.05935, 2024. 1, 3, 6  
+[19] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Batra, and Devi Parikh. Making the v in vqa matter: Elevating the role of image understanding in visual question answering. In CVPR, 2017. 6  
+[20] Ali Hassani and Humphrey Shi. Dilated neighborhood attention transformer. arXiv preprint arXiv:2209.15001, 2022.3  
+[21] Ali Hassani, Steven Walton, Jiachen Li, Shen Li, and Humphrey Shi. Neighborhood attention transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 3  
+[22] Drew A Hudson and Christopher D Manning. Gqa: A new dataset for real-world visual reasoning and compositional question answering. In CVPR, 2019. 6  
+[23] IDEFICS. Introducing idefics: An open reproduction of state-of-the-art visual language model. https://huggingface.co/blog/idefics, 2023. 3, 7  
+[24] Robert A Jacobs, Michael I Jordan, Steven J Nowlan, and Geoffrey E Hinton. Adaptive mixtures of local experts. Neural computation, 3(1):79-87, 1991. 3  
+[25] Jitesh Jain, Jiachen Li, Mang Tik Chiu, Ali Hassani, Nikita Orlov, and Humphrey Shi. Oneformer: One transformer to rule universal image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2989-2998, 2023. 3  
+[26] Jitesh Jain, Jianwei Yang, and Humphrey Shi. Vcoder: Versatile vision encoders for multimodal large language models. arXiv preprint arXiv:2312.14233, 2023. 3  
+[27] Albert Q Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, et al. Mistral 7b. arXiv preprint arXiv:2310.06825, 2023. 6  
+[28] Albert Q Jiang, Alexandre Sablayrolles, Antoine Roux, Arthur Mensch, Blanche Savary, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Emma Bou Hanna, Florian Bressand, et al. Mixtral of experts. arXiv:2401.04088, 2024. 2, 7  
+[29] Aniruddha Kembhavi, Mike Salvato, Eric Kolve, Minjoon Seo, Hannaneh Hajishirzi, and Ali Farhadi. A diagram is worth a dozen images. In ECCV, 2016. 6  
+[30] Geewook Kim, Teakgyu Hong, Moonbin Yim, Jeong Yeon Nam, Jinyoung Park, Jinyeong Yim, Wonseok Hwang, Sangdoo Yun, Dongyoon Han, and Seunghyun Park. Ocr-free document understanding transformer. In European Conference on Computer Vision, pages 498-517. Springer, 2022. 6  
+[31] Aran Komatsuzaki, Joan Puigcerver, James Lee-Thorp, Carlos Riquelme Ruiz, Basil Mustafa, Joshua Ainslie, Yi Tay, Mostafa Dehghani, and Neil Houlsby. Sparse upcycling: Training mixture-of-experts from dense checkpoints, 2023. 3, 4
+
+[32] Hugo Laurençon, Léo Tronchon, Matthieu Cord, and Victor Sanh. What matters when building vision-language models? arXiv preprint arXiv:2405.02246, 2024. 3  
+[33] Dmitry Lepikhin, HyoukJoong Lee, Yuanzhong Xu, Dehao Chen, Orhan First, Yanping Huang, Maxim Krikun, Noam M. Shazeer, and Z. Chen. Gshard: Scaling giant models with conditional computation and automatic sharding. ArXiv, abs/2006.16668, 2020. 3  
+[34] Mike Lewis, Shruti Bhosale, Tim Dettmers, Naman Goyal, and Luke Zettlemoyer. Base layers: Simplifying training of large, sparse models, 2021. 3  
+[35] Bohao Li, Rui Wang, Guangzhi Wang, Yuying Ge, Yixiao Ge, and Ying Shan. Seed-bench: Benchmarking multimodal llms with generative comprehension. arXiv:2307.16125, 2023. 6  
+[36] Jiachen Li, Vidit Goel, Marianna Ohanyan, Shant Navasardyan, Yunchao Wei, and Humphrey Shi. Vformer: End-to-end video matting with transformer. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 6678-6687, 2024. 3  
+[37] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597, 2023. 3  
+[38] Yanwei Li, Chengyao Wang, and Jiaya Jia. Llama-vid: An image is worth 2 tokens in large language models. arXiv:2311.17043, 2023. 6  
+[39] Yanwei Li, Yuechen Zhang, Chengyao Wang, Zhisheng Zhong, Yixin Chen, Ruihang Chu, Shaoteng Liu, and Jiaya Jia. Mini-gemini: Mining the potential of multi-modality vision language models, 2024. 1, 3, 6  
+[40] Yifan Li, Yifan Du, Kun Zhou, Jinpeng Wang, Wayne Xin Zhao, and Ji-Rong Wen. Evaluating object hallucination in large vision-language models. arXiv:2305.10355, 2023. 6  
+[41] Bin Lin, Zhenyu Tang, Yang Ye, Jiaxi Cui, Bin Zhu, Peng Jin, Junwu Zhang, Munan Ning, and Li Yuan. Moe-llava: Mixture of experts for large vision-language models. arXiv preprint arXiv:2401.15947, 2024. 3  
+[42] Ji Lin, Hongxu Yin, Wei Ping, Yao Lu, Pavlo Molchanov, Andrew Tao, Huizi Mao, Jan Kautz, Mohammad Shoeybi, and Song Han. Vila: On pre-training for visual language models. arXiv preprint arXiv:2312.07533, 2023. 1, 3, 6  
+[43] Ziyi Lin, Chris Liu, Renrui Zhang, Peng Gao, Longtian Qiu, Han Xiao, Han Qiu, Chen Lin, Wenqi Shao, Keqin Chen, Jiaming Han, Siyuan Huang, Yichi Zhang, Xuming He, Hongsheng Li, and Yu Qiao. *Sphinxx: The joint mixing of weights, tasks, and visual embeddings for multi-modal large language models*, 2023. 1, 3  
+[44] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. arXiv preprint arXiv:2310.03744, 2023. 1, 3, 4, 6, 7  
+[45] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava-last: Improved reasoning,OCR, and world knowledge, January 2024. 1, 3, 6  
+[46] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning, 2023. 1, 3, 6  
+[47] Yuan Liu, Haodong Duan, Yuanhan Zhang, Bo Li, Songyang Zhang, Wangbo Zhao, Yike Yuan, Jiaqi Wang, Conghui He, Ziwei Liu, et al. Mmbench: Is your multi-modal model an all-around player? arXiv:2307.06281, 2023. 6  
+[48] Pan Lu, Swaroop Mishra, Tanglin Xia, Liang Qiu, Kai-Wei Chang, Song-Chun Zhu, Oyvind Tafjord, Peter Clark, and Ashwin Kalyan. Learn to explain: Multimodal reasoning via thought chains for science question answering. In NeurIPS, 2022. 6  
+[49] Ahmed Masry, Do Xuan Long, Jia Qing Tan, Shafiq Joty, and Enamul Hoque. Chartqa: A benchmark for question answering about charts with visual and logical reasoning. arXiv:2203.10244, 2022. 6  
+[50] Minesh Mathew, Viraj Bagal, Ruben Tito, Dimosthenis Karatzas, Ernest Valveny, and CV Jawahar. Infographicvqa. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1697-1706, 2022. 6  
+[51] Brandon McKinzie, Zhe Gan, Jean-Philippe Fauconnier, Sam Dodge, Bowen Zhang, Philipp Dufter, Dhruti Shah, Xianzhi Du, Futang Peng, Floris Weers, Anton Belyi, Haotian Zhang, Karanjeet Singh, Doug Kang, Ankur Jain, Hongyu He, Max Schwarzer, Tom Gunter, Xiang Kong, Aonan Zhang, Jianyu Wang, Chong Wang, Nan Du, Tao Lei, Sam Wiseman, Guoli Yin, Mark Lee, Zirui Wang, Ruoming Pang, Peter Grasch, Alexander Toshev, and Yinfei Yang. Mm1: Methods, analysis, insights from multimodal llm pre-training, 2024. 1, 2, 3, 6
+
+[52] Basil Mustafa, Carlos Riquelme, Joan Puigcerver, Rodolphe Jenatton, and Neil Houlsby. Multimodal contrastive learning with limoe: the language-image mixture of experts, 2022. 3  
+[53] OpenAI. Gpt-4v(ision) system card. 2023. 1  
+[54] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 4, 6  
+[55] Machel Reid, Nikolay Savinov, Denis Teptyashin, Dmitry Lepikhin, Timothy Lillicrap, Jeffrey Dean, and et al. Oriol Vinyals. Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context, 2024. 2, 3  
+[56] Carlos Riquelme, Joan Puigcerver, Basil Mustafa, Maxim Neumann, Rodolphe Jenatton, André Susano Pinto, Daniel Keysers, and Neil Houlsby. Scaling vision with sparse mixture of experts, 2021. 3  
+[57] Noam Shazeer, Azalia Mirhoseini, Krzysztof Maziarz, Andy Davis, Quoc Le, Geoffrey Hinton, and Jeff Dean. Outrageously large neural networks: The sparsely-gated mixture-of-experts layer, 2017. 2, 3  
+[58] Baifeng Shi, Ziyang Wu, Maolin Mao, Xin Wang, and Trevor Darrell. When do we not need larger vision models? arXiv preprint arXiv:2403.13043, 2024. 8  
+[59] Amanpreet Singh, Vivek Natarajan, Meet Shah, Yu Jiang, Xinlei Chen, Dhruv Batra, Devi Parikh, and Marcus Rohrbach. Towards vqa models that can read. In CVPR, 2019. 6  
+[60] Zineng Tang, Ziyi Yang, Chenguang Zhu, Michael Zeng, and Mohit Bansal. Any-to-any generation via composable diffusion. Advances in Neural Information Processing Systems, 36, 2024. 3  
+[61] The Mosaic Research Team. Introducing dbrx: A new state-of-the-art open llm, March 2024. 2, 3  
+[62] Rubén Tito, Dimosthenis Karatzas, and Ernest Valveny. Document collection visual question answering. In ICDAR 2021, 2021. 6  
+[63] Hugo Touvron and et al. Louis Martin. Llama 2: Open foundation and fine-tuned chat models, 2023. 7  
+[64] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 2017. 1, 3  
+[65] Lemeng Wu, Mengchen Liu, Yinpeng Chen, Dongdong Chen, Xiyang Dai, and Lu Yuan. Residual mixture of experts, 2022. 3  
+[66] Xingqian Xu, Zhangyang Wang, Gong Zhang, Kai Wang, and Humphrey Shi. Versatile diffusion: Text, images and variations all in one diffusion model. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7754-7765, 2023. 3  
+[67] Weihao Yu, Zhengyuan Yang, Linjie Li, Jianfeng Wang, Kevin Lin, Zicheng Liu, Xinchao Wang, and Lijuan Wang. Mm-vet: Evaluating large multimodal models for integrated capabilities. arXiv:2308.02490, 2023. 6  
+[68] Xiang Yue, Yuansheng Ni, Kai Zhang, Tianyu Zheng, Ruoqi Liu, Ge Zhang, Samuel Stevens, Dongfu Jiang, Weiming Ren, Yuxuan Sun, Cong Wei, Botao Yu, Ruibin Yuan, Renliang Sun, Ming Yin, Boyuan Zheng, Zhenzhu Yang, Yibo Liu, Wenhao Huang, Huan Sun, Yu Su, and Wenhu Chen. Mmmu: A massive multi-discipline multimodal understanding and reasoning benchmark for expert agi. In CVPR, 2024. 6  
+[69] Pan Zhang, Xiaoyi Dong Bin Wang, Yuhang Cao, Chao Xu, Linke Ouyang, Zhiyuan Zhao, Shuangrui Ding, Songyang Zhang, Haodong Duan, Hang Yan, et al. Internlm-xcomposer: A vision-language large model for advanced text-image comprehension and composition. arXiv preprint arXiv:2309.15112, 2023. 3  
+[70] Chunting Zhou, Pengfei Liu, Puxin Xu, Srinivasan Iyer, Jiao Sun, Yuning Mao, Xuezhe Ma, Avia Efrat, Ping Yu, Lili Yu, et al. Lima: Less is more for alignment. Advances in Neural Information Processing Systems, 36, 2024. 6  
+[71] Yanqi Zhou, Tao Lei, Hanxiao Liu, Nan Du, Yanping Huang, Vincent Zhao, Andrew Dai, Zhifeng Chen, Quoc Le, and James Laudon. Mixture-of-experts with expert choice routing, 2022. 3  
+[72] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. In International Conference on Learning Representations, 2020. 3  
+[73] Barret Zoph, Irwan Bello, Sameer Kumar, Nan Du, Yanping Huang, Jeff Dean, Noam Shazeer, and William Fedus. St-moe: Designing stable and transferable sparse expert models, 2022. 3, 5
+
+# Appendix
+
+The supplementary material elaborates on further aspects of our work concerning the experimental setups and dataset usage. In Appendix A, we provide details on the datasets used for the visual instruction tuning stage and how we converted the mixture of datasets into the visual instruction following formats. In Appendix B, we present the hyperparameters used for the three-stage trainings. In Appendix E, we include additional examples of dialogues between the user and our CuMo models.
+
+# A Dataset Details
+
+As outlined in Table 4, we provide detailed information on the datasets utilized for the three-stage training process mentioned in Section 3.3. All data are converted into the instruction-following format for training. For the Syndog-EN and DVQA datasets, we didn't use the entire training set as we observed that a large portion of synthetic data negatively impacts the zero-shot performance of the multimodal LLMs.
+
+<table><tr><td>Dataset</td><td>Size</td></tr><tr><td colspan="2">Pre-Training</td></tr><tr><td>LCS-558K</td><td>558K</td></tr><tr><td colspan="2">Pre-Finetuning</td></tr><tr><td>ALLaVA-Caption</td><td>708K</td></tr><tr><td colspan="2">Visual Instruction Tuning</td></tr><tr><td>LLaVA-665K</td><td>665K</td></tr><tr><td>ShareGPT4V</td><td>102K</td></tr><tr><td>LAION-GPT-V</td><td>11K</td></tr><tr><td>DocVQA</td><td>10K</td></tr><tr><td>SynDog-EN</td><td>50K</td></tr><tr><td>ChartQA</td><td>4K</td></tr><tr><td>DVQA</td><td>50K</td></tr><tr><td>AI2D</td><td>2K</td></tr><tr><td>InfoVQA</td><td>4K</td></tr><tr><td>ALLaVA</td><td>708K</td></tr><tr><td>LIMA</td><td>1K</td></tr><tr><td>ALLaVA-Text</td><td>143K</td></tr></table>
+
+Table 4: List of datasets used for three training stages.  
+
+<table><tr><td>Model</td><td>Vision Encoder</td><td>ImageNet Acc.</td><td>Res.</td><td>Params.</td><td>TextVQA</td><td>MM Vet</td><td>SEED</td></tr><tr><td>LLaVA-v1.5</td><td>CLIP-ViT-L</td><td>76.6</td><td>336</td><td>0.30B</td><td>57.6</td><td>32.1</td><td>66.4</td></tr><tr><td>CuMo</td><td>CLIP-ViT-L</td><td>76.6</td><td>336</td><td>0.50B</td><td>59.3</td><td>34.3</td><td>69.6</td></tr><tr><td>LLaVA-v1.5</td><td>SigLIP-SO400M</td><td>83.2</td><td>384</td><td>0.43B</td><td>58.1</td><td>32.5</td><td>67.5</td></tr><tr><td>CuMo</td><td>SigLIP-SO400M</td><td>83.2</td><td>384</td><td>0.72B</td><td>59.4</td><td>34.1</td><td>69.8</td></tr></table>
+
+Table 5: CuMo under different vision encoders.
+
+# B Experimental Setup Details
+
+Table 6 provides an overview of the main hyperparameters used during the three-stage training process. For the final results presented in Table 1, the model was trained using  $32 \times \mathrm{A}100$  GPUs with a total batch size of 256 and a learning rate of 4e-6. All ablation studies were conducted with a total batch size of 128 and learning rates of 2e-5 and 2e-6, as detailed in Section 4.3.
+
+<table><tr><td>Hyperparameter</td><td>PT</td><td>PFT</td><td>VIT</td></tr><tr><td>Learning rate</td><td>1e-3</td><td>2e-6</td><td>4e-6</td></tr><tr><td>LR schedule</td><td>Cosine</td><td>Cosine</td><td>Cosine</td></tr><tr><td>Batchsize per GPU</td><td>32</td><td>8</td><td>8</td></tr><tr><td>GPUs</td><td>8×A100</td><td>16×A100</td><td>32×A100</td></tr><tr><td>Zero</td><td>Zero2</td><td>Zero3</td><td>Zero3-offload</td></tr><tr><td>Optimizer</td><td>AdamW</td><td>AdamW</td><td>AdamW</td></tr><tr><td>MLP</td><td>Open</td><td>Open</td><td>Open</td></tr><tr><td>CLIP</td><td>Freeze</td><td>Open</td><td>Open</td></tr><tr><td>LLM</td><td>Freeze</td><td>Open</td><td>Open</td></tr><tr><td>MoE blocks</td><td>-</td><td>-</td><td>✓</td></tr><tr><td>Max token</td><td>2048</td><td>4096</td><td>4096</td></tr></table>
+
+# C More Vision Encoders
+
+Table 5 shows CuMo with different vision encoders. Here we use the pre-trained SigLIP-SO400M as the vision encoder and add MoE to the vision encoder. SigLIP-SO400M has a much better performance on ImageNet zero-shot classification than CLIP-ViT-L (83.2 vs 76.6). The added MoE can still make improvements to this stronger vision encoder but the average improvement shrinks compared to CLIP-ViT-L. However, the training data here is limited to LLaVA-665K for quick verification, which may not show the full potential of the model if training with more data.
+
+Table 6: Hyperparameters used in three-stage training on Mistral-7B. PT: Pre-Training stage. PFT: Pre-FineTuning stage. VIT: Visual Instruction tuning stage.  
+
+<table><tr><td>CuMo</td><td>CLIP</td><td>MLP</td><td>LLM</td><td>Total</td></tr><tr><td>Mistral-7B</td><td>0.30B</td><td>0.025B</td><td>7.25B</td><td>7.58B</td></tr><tr><td>⇒ Activation Params</td><td>0.30B</td><td>0.025B</td><td>7.25B</td><td>7.58B</td></tr><tr><td>+ Top 2-in-4 MLP-MoE</td><td>0.30B</td><td>0.10B</td><td>7.25B</td><td>7.65B</td></tr><tr><td>⇒ Activation Params</td><td>0.30B</td><td>0.05B</td><td>7.25B</td><td>7.60B</td></tr><tr><td>+ Top 2-in-4 CLIP-MoE</td><td>0.91B</td><td>0.10B</td><td>7.25B</td><td>8.26B</td></tr><tr><td>⇒ Activation Params</td><td>0.50B</td><td>0.05B</td><td>7.25B</td><td>7.80B</td></tr><tr><td>⇒ Mixtral-8x7B</td><td>0.91B</td><td>0.10B</td><td>46.70B</td><td>47.71B</td></tr><tr><td>⇒ Activation Params</td><td>0.50B</td><td>0.05B</td><td>12.90B</td><td>13.45B</td></tr></table>
+
+Table 7: Change of model parameters of CuMo. The 7.80B and 13.45B activation parameters corresponds to Act. of CuMo in Table 1.
+
+# D Model Parameters
+
+We include Table 7 to illustrate the evolution of parameters in the CuMo model throughout its construction process. The LLM constitutes a significant proportion of the total parameters, underscoring the potential for further scaling up the vision encoders to bolster the strength of multimodal LLMs.
+
+# E More Dialogues
+
+We add more dialogues between the questions from the users and the response from CuMo-7B in Figure 6.
+
+# F Border Impact
+
+The potential border impact of CuMo, as an open-sourced multimodal LLM, relies on two aspects: the development of artificial general intelligence (AGI) and the open-source community.
+
+<table><tr><td></td><td>What car is it in this image
+The car in the image is a Porsche Taycan, which is an electric vehicle produced by the German automobile manufacturer Porsche. The Taycan is known for its distinctive design elements that include the sloping roofline, the shape of the headlights, and the specific style of the wheels. The model name &quot;Taycan&quot; is visible on the front of the car, confirming its identity. This particular model is likely one of the higher-end versions, as indicated by the sporty design and the presence of what appears to be performance-oriented features.</td></tr><tr><td></td><td>What animals are in this image
+The animals in this image are a cat and a dog. The cat is a domestic short-haired feline, identifiable by its tabby pattern and the shape of its ears and eyes. The dog is a Border Collie, recognizable by its distinctive black and white coat, medium-length fur, and the shape of its ears and eyes. Both animals are common domestic pets and are known for their companionship with humans.</td></tr><tr><td></td><td>Can you tell me what happened in this image
+The image captures a moment where three individuals are standing beside a large piece of equipment, which appears to be a server or some form of industrial hardware. The presence of the yellow warning label suggests that the equipment is either new, being serviced, or requires caution when operating or moving. The setting seems to be a professional environment, possibly a data center or a technical facility, given the nature of the equipment and the clean, organized space. The individuals are dressed in casual to smart-casual attire, which might indicate a work-related context. The exact event or occasion is not clear from the image alone, but it could be a delivery, installation, or inspection of the equipment. The image does not provide any explicit information about the specific actions or events that took place prior to or after this moment.</td></tr><tr><td></td><td>How many characters in this image
+There are seven characters in this image.</td></tr><tr><td></td><td>How many calories estimated in this bowl
+Estimating the exact number of calories in a dish from a photograph is not possible without specific measurements and ingredient quantities. However, based on the visible ingredients and typical calorie counts for each, one could make an educated guess. A bowl of noodles with vegetables and a protein like chicken typically ranges from 300 to 500 calories. The addition of peanuts and sesame seeds would add a few dozen calories each. The sauce, which appears to be a light dressing, would contribute minimal calories. Therefore, a rough estimate might place this dish in the range of 400 to 600 calories, but this is a very rough estimate and could vary significantly based on the actual ingredients and portion sizes.</td></tr></table>
+
+Figure 6: More dialogues of CuMo-7B. We highlight the correct answers and hallucinations from the responses of CuMo.
+
+AGI: The development of AGI is a core topic, with the central component being the LLM. A multimodal LLM enhances its abilities to understand content such as images and videos, making it a more versatile agent and a positive contributor towards AGI.
+
+Open-sourced MLLM: Open-sourced weights and code can accelerate the development of MLLMs. However, this also has potential negative impacts on society. Therefore, we have added a non-society license to constrain the usage of our model.
+
+# NeurIPS Paper Checklist
+
+# 1. Claims
+
+Question: Do the main claims made in the abstract and introduction accurately reflect the paper's contributions and scope?
+
+Answer: [Yes]
+
+Justification: We summarize our contributions and scope in the abstract and introduction with the main experimental results to support our claims.
+
+Guidelines:
+
+- The answer NA means that the abstract and introduction do not include the claims made in the paper.  
+- The abstract and/or introduction should clearly state the claims made, including the contributions made in the paper and important assumptions and limitations. A No or NA answer to this question will not be perceived well by the reviewers.  
+- The claims made should match theoretical and experimental results, and reflect how much the results can be expected to generalize to other settings.  
+- It is fine to include aspirational goals as motivation as long as it is clear that these goals are not attained by the paper.
+
+# 2. Limitations
+
+Question: Does the paper discuss the limitations of the work performed by the authors?
+
+Answer: [Yes]
+
+Justification: We add section 4.4 to discuss the limitations of our CuMo model.
+
+Guidelines:
+
+- The answer NA means that the paper has no limitation while the answer No means that the paper has limitations, but those are not discussed in the paper.  
+- The authors are encouraged to create a separate "Limitations" section in their paper.  
+- The paper should point out any strong assumptions and how robust the results are to violations of these assumptions (e.g., independence assumptions, noiseless settings, model well-specification, asymptotic approximations only holding locally). The authors should reflect on how these assumptions might be violated in practice and what the implications would be.  
+- The authors should reflect on the scope of the claims made, e.g., if the approach was only tested on a few datasets or with a few runs. In general, empirical results often depend on implicit assumptions, which should be articulated.  
+- The authors should reflect on the factors that influence the performance of the approach. For example, a facial recognition algorithm may perform poorly when image resolution is low or images are taken in low lighting. Or a speech-to-text system might not be used reliably to provide closed captions for online lectures because it fails to handle technical jargon.  
+- The authors should discuss the computational efficiency of the proposed algorithms and how they scale with dataset size.  
+- If applicable, the authors should discuss possible limitations of their approach to address problems of privacy and fairness.  
+- While the authors might fear that complete honesty about limitations might be used by reviewers as grounds for rejection, a worse outcome might be that reviewers discover limitations that aren't acknowledged in the paper. The authors should use their best judgment and recognize that individual actions in favor of transparency play an important role in developing norms that preserve the integrity of the community. Reviewers will be specifically instructed to not penalize honesty concerning limitations.
+
+# 3. Theory Assumptions and Proofs
+
+Question: For each theoretical result, does the paper provide the full set of assumptions and a complete (and correct) proof?
+
+Answer: [NA]
+
+Justification: This paper is application oriented and we conduct extensive experiments to validate our assumptions.
+
+# Guidelines:
+
+- The answer NA means that the paper does not include theoretical results.  
+- All the theorems, formulas, and proofs in the paper should be numbered and cross-referenced.  
+- All assumptions should be clearly stated or referenced in the statement of any theorems.  
+- The proofs can either appear in the main paper or the supplemental material, but if they appear in the supplemental material, the authors are encouraged to provide a short proof sketch to provide intuition.  
+- Inversely, any informal proof provided in the core of the paper should be complemented by formal proofs provided in appendix or supplemental material.  
+- Theorems and Lemmas that the proof relies upon should be properly referenced.
+
+# 4. Experimental Result Reproducibility
+
+Question: Does the paper fully disclose all the information needed to reproduce the main experimental results of the paper to the extent that it affects the main claims and/or conclusions of the paper (regardless of whether the code and data are provided or not)?
+
+# Answer: [Yes]
+
+Justification: We provide all experimental details including the datasets, hyperparameters, and training devices that we used.
+
+# Guidelines:
+
+- The answer NA means that the paper does not include experiments.  
+- If the paper includes experiments, a No answer to this question will not be perceived well by the reviewers: Making the paper reproducible is important, regardless of whether the code and data are provided or not.  
+- If the contribution is a dataset and/or model, the authors should describe the steps taken to make their results reproducible or verifiable.  
+- Depending on the contribution, reproducibility can be accomplished in various ways. For example, if the contribution is a novel architecture, describing the architecture fully might suffice, or if the contribution is a specific model and empirical evaluation, it may be necessary to either make it possible for others to replicate the model with the same dataset, or provide access to the model. In general, releasing code and data is often one good way to accomplish this, but reproducibility can also be provided via detailed instructions for how to replicate the results, access to a hosted model (e.g., in the case of a large language model), releasing of a model checkpoint, or other means that are appropriate to the research performed.  
+- While NeurIPS does not require releasing code, the conference does require all submissions to provide some reasonable avenue for reproducibility, which may depend on the nature of the contribution. For example  
+(a) If the contribution is primarily a new algorithm, the paper should make it clear how to reproduce that algorithm.  
+(b) If the contribution is primarily a new model architecture, the paper should describe the architecture clearly and fully.  
+(c) If the contribution is a new model (e.g., a large language model), then there should either be a way to access this model for reproducing the results or a way to reproduce the model (e.g., with an open-source dataset or instructions for how to construct the dataset).  
+(d) We recognize that reproducibility may be tricky in some cases, in which case authors are welcome to describe the particular way they provide for reproducibility. In the case of closed-source models, it may be that access to the model is limited in some way (e.g., to registered users), but it should be possible for other researchers to have some path to reproducing or verifying the results.
+
+# 5. Open access to data and code
+
+Question: Does the paper provide open access to the data and code, with sufficient instructions to faithfully reproduce the main experimental results, as described in supplemental material?
+
+Answer: [Yes]
+
+Justification: We provide all the open-sourced datasets and codes that we used for training and evaluation.
+
+Guidelines:
+
+- The answer NA means that paper does not include experiments requiring code.  
+- Please see the NeurIPS code and data submission guidelines (https://nips.cc/public/guides/CodeSubmissionPolicy) for more details.  
+- While we encourage the release of code and data, we understand that this might not be possible, so "No" is an acceptable answer. Papers cannot be rejected simply for not including code, unless this is central to the contribution (e.g., for a new open-source benchmark).  
+- The instructions should contain the exact command and environment needed to run to reproduce the results. See the NeurIPS code and data submission guidelines (https://nips.cc/public/guides/CodeSubmissionPolicy) for more details.  
+- The authors should provide instructions on data access and preparation, including how to access the raw data, preprocessed data, intermediate data, and generated data, etc.  
+- The authors should provide scripts to reproduce all experimental results for the new proposed method and baselines. If only a subset of experiments are reproducible, they should state which ones are omitted from the script and why.  
+- At submission time, to preserve anonymity, the authors should release anonymized versions (if applicable).  
+- Providing as much information as possible in supplemental material (appended to the paper) is recommended, but including URLs to data and code is permitted.
+
+# 6. Experimental Setting/Details
+
+Question: Does the paper specify all the training and test details (e.g., data splits, hyperparameters, how they were chosen, type of optimizer, etc.) necessary to understand the results?
+
+Answer: [Yes]
+
+Justification: We provide all the training and inference details for reproducing and understanding our results.
+
+Guidelines:
+
+- The answer NA means that the paper does not include experiments.  
+- The experimental setting should be presented in the core of the paper to a level of detail that is necessary to appreciate the results and make sense of them.  
+- The full details can be provided either with the code, in appendix, or as supplemental material.
+
+# 7. Experiment Statistical Significance
+
+Question: Does the paper report error bars suitably and correctly defined or other appropriate information about the statistical significance of the experiments?
+
+Answer: [Yes]
+
+Justification: For results with statistical variance like querying GPT API, we averaged the results by three times to reduce the variance.
+
+Guidelines:
+
+- The answer NA means that the paper does not include experiments.  
+- The authors should answer "Yes" if the results are accompanied by error bars, confidence intervals, or statistical significance tests, at least for the experiments that support the main claims of the paper.
+
+- The factors of variability that the error bars are capturing should be clearly stated (for example, train/test split, initialization, random drawing of some parameter, or overall run with given experimental conditions).  
+- The method for calculating the error bars should be explained (closed form formula, call to a library function, bootstrap, etc.)  
+- The assumptions made should be given (e.g., Normally distributed errors).  
+- It should be clear whether the error bar is the standard deviation or the standard error of the mean.  
+- It is OK to report 1-sigma error bars, but one should state it. The authors should preferably report a 2-sigma error bar than state that they have a  $96\%$  CI, if the hypothesis of Normality of errors is not verified.  
+- For asymmetric distributions, the authors should be careful not to show in tables or figures symmetric error bars that would yield results that are out of range (e.g. negative error rates).  
+- If error bars are reported in tables or plots, The authors should explain in the text how they were calculated and reference the corresponding figures or tables in the text.
+
+# 8. Experiments Compute Resources
+
+Question: For each experiment, does the paper provide sufficient information on the computer resources (type of compute workers, memory, time of execution) needed to reproduce the experiments?
+
+Answer: [Yes]
+
+Justification: We provide the details of the compute resources settings to reproduce our results.
+
+Guidelines:
+
+- The answer NA means that the paper does not include experiments.  
+- The paper should indicate the type of compute workers CPU or GPU, internal cluster, or cloud provider, including relevant memory and storage.  
+- The paper should provide the amount of compute required for each of the individual experimental runs as well as estimate the total compute.  
+- The paper should disclose whether the full research project required more compute than the experiments reported in the paper (e.g., preliminary or failed experiments that didn't make it into the paper).
+
+# 9. Code Of Ethics
+
+Question: Does the research conducted in the paper conform, in every respect, with the NeurIPS Code of Ethics https://neurips.cc/public/EthicsGuidelines?
+
+Answer: [Yes]
+
+Justification: We strictly follow the code of ethics during the project.
+
+Guidelines:
+
+- The answer NA means that the authors have not reviewed the NeurIPS Code of Ethics.  
+- If the authors answer No, they should explain the special circumstances that require a deviation from the Code of Ethics.  
+- The authors should make sure to preserve anonymity (e.g., if there is a special consideration due to laws or regulations in their jurisdiction).
+
+# 10. Broader Impacts
+
+Question: Does the paper discuss both potential positive societal impacts and negative societal impacts of the work performed?
+
+Answer: [Yes]
+
+Justification: We add the discussions of border impact in Appendix F.
+
+Guidelines:
+
+- The answer NA means that there is no societal impact of the work performed.
+
+- If the authors answer NA or No, they should explain why their work has no societal impact or why the paper does not address societal impact.  
+- Examples of negative societal impacts include potential malicious or unintended uses (e.g., disinformation, generating fake profiles, surveillance), fairness considerations (e.g., deployment of technologies that could make decisions that unfairly impact specific groups), privacy considerations, and security considerations.  
+- The conference expects that many papers will be foundational research and not tied to particular applications, let alone deployments. However, if there is a direct path to any negative applications, the authors should point it out. For example, it is legitimate to point out that an improvement in the quality of generative models could be used to generate deepfakes for disinformation. On the other hand, it is not needed to point out that a generic algorithm for optimizing neural networks could enable people to train models that generate Deepfakes faster.  
+- The authors should consider possible harms that could arise when the technology is being used as intended and functioning correctly, harms that could arise when the technology is being used as intended but gives incorrect results, and harms following from (intentional or unintentional) misuse of the technology.  
+- If there are negative societal impacts, the authors could also discuss possible mitigation strategies (e.g., gated release of models, providing defenses in addition to attacks, mechanisms for monitoring misuse, mechanisms to monitor how a system learns from feedback over time, improving the efficiency and accessibility of ML).
+
+# 11. Safeguards
+
+Question: Does the paper describe safeguards that have been put in place for responsible release of data or models that have a high risk for misuse (e.g., pretrained language models, image generators, or scraped datasets)?
+
+Answer: [Yes]
+
+Justification: We provide usage guidelines for users to adhere when accessing to our model.
+
+Guidelines:
+
+- The answer NA means that the paper poses no such risks.  
+- Released models that have a high risk for misuse or dual-use should be released with necessary safeguards to allow for controlled use of the model, for example by requiring that users adhere to usage guidelines or restrictions to access the model or implementing safety filters.  
+- Datasets that have been scraped from the Internet could pose safety risks. The authors should describe how they avoided releasing unsafe images.  
+- We recognize that providing effective safeguards is challenging, and many papers do not require this, but we encourage authors to take this into account and make a best faith effort.
+
+# 12. Licenses for existing assets
+
+Question: Are the creators or original owners of assets (e.g., code, data, models), used in the paper, properly credited and are the license and terms of use explicitly mentioned and properly respected?
+
+Answer: [Yes]
+
+Justification: We strictly follow the original licenses of existing assets including datasets and codes.
+
+Guidelines:
+
+- The answer NA means that the paper does not use existing assets.  
+- The authors should cite the original paper that produced the code package or dataset.  
+- The authors should state which version of the asset is used and, if possible, include a URI.  
+- The name of the license (e.g., CC-BY 4.0) should be included for each asset.  
+- For scraped data from a particular source (e.g., website), the copyright and terms of service of that source should be provided.
+
+- If assets are released, the license, copyright information, and terms of use in the package should be provided. For popular datasets, paperswithcode.com/datasets has curated licenses for some datasets. Their licensing guide can help determine the license of a dataset.  
+- For existing datasets that are re-packaged, both the original license and the license of the derived asset (if it has changed) should be provided.  
+- If this information is not available online, the authors are encouraged to reach out to the asset's creators.
+
+# 13. New Assets
+
+Question: Are new assets introduced in the paper well documented and is the documentation provided alongside the assets?
+
+Answer: [Yes]
+
+Justification: Our models and codes are well documented for reproduction and protected by licenses.
+
+Guidelines:
+
+- The answer NA means that the paper does not release new assets.  
+- Researchers should communicate the details of the dataset/code/model as part of their submissions via structured templates. This includes details about training, license, limitations, etc.  
+- The paper should discuss whether and how consent was obtained from people whose asset is used.  
+- At submission time, remember to anonymize your assets (if applicable). You can either create an anonymized URL or include an anonymized zip file.
+
+# 14. Crowdsourcing and Research with Human Subjects
+
+Question: For crowdsourcing experiments and research with human subjects, does the paper include the full text of instructions given to participants and screenshots, if applicable, as well as details about compensation (if any)?
+
+Answer: [NA]
+
+Justification: Our paper does not include crowdsourcing or human subjects.
+
+Guidelines:
+
+- The answer NA means that the paper does not involve crowdsourcing nor research with human subjects.  
+- Including this information in the supplemental material is fine, but if the main contribution of the paper involves human subjects, then as much detail as possible should be included in the main paper.  
+- According to the NeurIPS Code of Ethics, workers involved in data collection, curation, or other labor should be paid at least the minimum wage in the country of the data collector.
+
+# 15. Institutional Review Board (IRB) Approvals or Equivalent for Research with Human Subjects
+
+Question: Does the paper describe potential risks incurred by study participants, whether such risks were disclosed to the subjects, and whether Institutional Review Board (IRB) approvals (or an equivalent approval/review based on the requirements of your country or institution) were obtained?
+
+Answer: [NA]
+
+Justification: Our paper does not include crowdsourcing or human subjects.
+
+Guidelines:
+
+- The answer NA means that the paper does not involve crowdsourcing nor research with human subjects.  
+- Depending on the country in which research is conducted, IRB approval (or equivalent) may be required for any human subjects research. If you obtained IRB approval, you should clearly state this in the paper.
+
+- We recognize that the procedures for this may vary significantly between institutions and locations, and we expect authors to adhere to the NeurIPS Code of Ethics and the guidelines for their institution.  
+- For initial submissions, do not include any information that would break anonymity (if applicable), such as the institution conducting the review.

2024/CuMo_ Scaling Multimodal LLM with Co-Upcycled Mixture-of-Experts/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:dfdf39c0322d8d4807c00130add8e735771a9ed17764be117c223781cdb0a20a
+size 859248

2024/CultureLLM_ Incorporating Cultural Differences into Large Language Models/466f5c65-ec31-4ae6-afd0-26e62131d89f_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2573759ca1e60f46b608e62cbeebea23e339e486e73b7d5011e7d882e604ff4d
+size 1225189

2024/CultureLLM_ Incorporating Cultural Differences into Large Language Models/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:98f62add92cd80f24d37b90973a7f46b0c48ac000ee6b0a6190040ef961f1e56
+size 2225416

2024/CulturePark_ Boosting Cross-cultural Understanding in Large Language Models/e454ed12-d2ac-4594-9ae6-3ab30d0063bb_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:669d690dfdc9f7f5d003752b6ddeb2ec8f03ffdff79fdd5ebb68ac6cc6dbf1eb
+size 1777094

2024/CulturePark_ Boosting Cross-cultural Understanding in Large Language Models/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:de6f532d801e5a009c14022e7b8e37231a9b4b6ff642c65f011f806bd719ec88
+size 1787449

2024/Curriculum Fine-tuning of Vision Foundation Model for Medical Image Classification Under Label Noise/19e9ffb6-2b3d-4bfe-8b8b-16ee509b59b2_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b74dfe200b9d5e93417e13aee8ee43836ea1e96c6a21e24a780e3b9af5ae0bcd
+size 1281031

2024/Curriculum Fine-tuning of Vision Foundation Model for Medical Image Classification Under Label Noise/full.md

@@ -0,0 +1,525 @@
+# Curriculum Fine-tuning of Vision Foundation Model for Medical Image Classification Under Label Noise
+
+Yeonguk Yu Minhwan Ko Sungho Shin Kangmin Kim Kyoobin Lee*
+
+Gwangju Institute of Science and Technology
+
+{yeon_guk, mhko1998, hogili89, kgmin156}@gm.gist.ac.kr kyoobinlee@gist.ac.kr
+
+# Abstract
+
+Deep neural networks have demonstrated remarkable performance in various vision tasks, but their success heavily depends on the quality of the training data. Noisy labels are a critical issue in medical datasets and can significantly degrade model performance. Previous clean sample selection methods have not utilized the well pre-trained features of vision foundation models (VFMs) and assumed that training begins from scratch. In this paper, we propose CUFIT, a curriculum fine-tuning paradigm of VFMs for medical image classification under label noise. Our method is motivated by the fact that linear probing of VFMs is relatively unaffected by noisy samples, as it does not update the feature extractor of the VFM, thus robustly classifying the training samples. Subsequently, curriculum fine-tuning of two adapters is conducted, starting with clean sample selection from the linear probing phase. Our experimental results demonstrate that CUFIT outperforms previous methods across various medical image benchmarks. Specifically, our method surpasses previous baselines by  $5.0\%$ ,  $2.1\%$ ,  $4.6\%$ , and  $5.8\%$  at a  $40\%$  noise rate on the HAM10000, APTOS-2019, BloodMnist, and OrgancMnist datasets, respectively. Furthermore, we provide extensive analyses to demonstrate the impact of our method on noisy label detection. For instance, our method shows higher label precision and recall compared to previous approaches. Our work highlights the potential of leveraging VFMs in medical image classification under challenging conditions of noisy labels.
+
+# 1 Introduction
+
+Deep neural networks have demonstrated remarkable performance across various tasks, including classification, detection, and segmentation [20, 16, 19, 57]. In medical imaging, these neural networks leverage large amounts of labeled data to train models capable of accurately detecting or classifying medical conditions from images such as dermatoscopes, X-rays, MRIs, and CT scans. However, in practical settings, data often contain noisy labels and it is well established that neural networks perform well only when the quality of training data is sufficiently high [42, 29, 5]. Noisy labels occur when the data annotations—the labels assigned to training images—are incorrect or inconsistent. This issue is particularly problematic in medical imaging, where annotating images is more complex compared to natural images [53, 25]. Consequently, improving the robustness of neural networks against noisy labels is a crucial area of research, directly affecting the effectiveness and reliability of medical imaging technologies.
+
+A large number of algorithms have been developed to address the issue of performance degradation caused by noisy samples [42]. In particular, clean sample selection methods, such as MentorNet [24], Co-teaching [17], Co-teaching+ [56], JoCor [47], and CoDis [50], have demonstrated superior performance without requiring modifications to the model architecture or training loss. The core
+
+![](images/3101adea8d267551dc71691cef8e8e61febf0cbf4838ecf702626a4fbb8f871a.jpg)  
+(a) Linear probing
+
+![](images/fe8b88419c2f60145a6050bab3bd3cbd93c856f97ea32d57680f3d9ad9eab90d.jpg)  
+(b) Using adapters
+
+![](images/e502209d4b87825231ace83b6eb0ac0e5b8203e0438097579b49e2e09e5eb190.jpg)  
+(c) Performance comparison  
+Figure 1: Illustration of linear probing (a) and adapter usage (b). Specifically, the weights of the foundation model are frozen, while the fully connected layer or adapter weights (shown in orange) are updated during the training phase. In (c), a performance comparison using a simulated noisy dataset (HAM10000) is presented. It demonstrates that linear probing is more robust to noisy labels compared to the adapter, whereas the adapter outperforms linear probing when there are no noisy labels.
+
+principle behind these methods is that small-loss samples are likely to be clean, as they are easier to classify and the model memorizes them faster. Additionally, using two different but homogeneous networks to select small-loss samples for each other is more stable than relying on a single model for sample selection. These methods have shown outstanding performance using traditional neural network architectures based on convolutional layers. However, there are practical issues with these methods in two aspects: (i) it cannot be guaranteed that these methods will perform equally well with transformer-based architectures, which have recently gained significant attention, and (ii) their assumption that training starts from scratch is impractical, as it prevents the use of rich features from pre-trained models, which could be beneficial for filtering noisy labels.
+
+Recently, large-scale vision foundation models (VFMs) with transformer-based architectures [13], such as CLIP [39], MAE [18], SAM [28], and DINOv2 [37], have gained attention for their performance and applicability across various tasks. The self-supervised training of VFMs on large-scale datasets enhances their robustness against various image corruptions and improves their generalization capabilities [10, 48, 38, 40]. The inherent robustness and rich features of VFMs can be beneficial for detecting noisy labels. For instance, linear probing of VFMs is relatively unaffected by noisy samples since it does not modify the VFM's feature extractor, preventing the memorization of noisy data, as shown in Figure 1. However, linear probing does not fully leverage the VFM's capabilities when there is a domain gap between the pretraining task of the VFM and the target task (e.g., pretraining on natural images versus medical image classification) [48]. To address this issue, some researchers have proposed using trainable fine-tuning adapters for VFMs [22, 23, 48, 9]. Yet, these adapters might degrade performance by memorizing noisy labels due to their trainable parameters involved in feature extraction. Therefore, we state our research question as follows: How can we use the power of pre-trained vision foundation models for medical image classification in the presence of noisy labels?
+
+In this paper, we introduce a Curriculum Fine-Tuning paradigm for vision foundation models in medical image classification under noisy labels, called CUFIT. CUFIT is a curriculum-learning framework designed to fine-tune VFMs with noisy medical datasets. The framework consists of three training modules: the Linear Probing Module (LPM), the Intermediate Adapter Module (IAM), and the Last Adapter Module (LAM). During the training stage, clean samples are selected based on an agreement criterion, where a sample is selected if its annotation matches the module's prediction. Specifically, the LPM is trained using all available samples, as linear probing is robust against noisy labels. Subsequently, the IAM is trained with the samples selected by the LPM, and the LAM is trained with the samples selected by the IAM. This inter-module curriculum training (i.e.,  $\mathrm{LPM} \rightarrow \mathrm{IAM} \rightarrow \mathrm{LAM}$ ) is beneficial for increasing the number of clean samples available to train the LAM, considering that the LPM only selects a limited number of samples due to performance degradation caused by the domain gap. Consequently, CUFIT leverages the LAM for final predictions,
+
+offering strong fine-tuning performance for medical image classification in the presence of noisy labels by utilizing the strengths of both linear probing and adapters, as illustrated in Figure 1-c. In summary, the main contributions of this paper are as follows:
+
+- We introduce CUFIT, a simple yet effective fine-tuning paradigm for medical image classification using VFMs in the presence of noisy labels. This method leverages the robustness of linear probing and the generalization capability of fine-tuning adapters to handle noisy datasets during training stage.  
+- We conduct various experiments demonstrating that CUFIT significantly improves the robustness of VFMs against noisy labels in medical datasets. We show that CUFIT outperforms previous methods across several medical image benchmarks.  
+- We provide extensive analyses to enhance the understanding of our fine-tuning paradigm. Additionally, we validate our framework with various VFMs and adapter configurations.
+
+# 2 Related Work
+
+Vision foundation models. Vision transformers (ViTs) embed 2D images into 1D tokens and model their global correlations using the self-attention mechanism [13, 35, 43]. ViTs are known to be effective when large datasets are used, and the concept of vision foundation models is introduced. Several studies have developed pre-trained vision transformers based on self-supervised learning. For instance, contrastive language-image pre-training (CLIP) provides high-quality visual representations through contrastive learning with a large amount of image-text pairs [39]. Additionally, masked autoencoder (MAE) offers high-capacity models that generalize well through self-supervised learning with masked auto-encoding, where the task is to reconstruct token patches from the given masked tokens [18]. Moreover, knowledge distillation with no labels (DINOv1) [7] proposed a teacher-student framework for self-training without annotations, resulting in a well-generalized ViT. More recently, DINOv2 introduced a self-training framework that combines masked autoencoding and teacher-student training based on carefully curated datasets [37].
+
+Since large models and self-training in VFs provide strong generalization capabilities for various tasks, parameter-efficient fine-tuning has gained attention. Parameter-efficient fine-tuning (PEFT) aims to adapt foundation models to new tasks by training only a few adapter parameters while keeping the model itself frozen. Notably, there is research that proposes low-rank adaptation (LoRA), which introduces trainable rank decomposition matrices into each layer of the transformer architecture in large language models [22]. For the VFs, visual prompt tuning [23] proposed appending prompts to the input sequence of each transformer block, achieving excellent fine-tuning performance with minimal parameters. Similarly, Adaptformer [9] introduces a novel MLP block to replace the original one in transformer blocks, allowing for the use of both original and few trainable parameters. More recently, Wei et al. proposed Rein, which aims to adapt VFs for semantic segmentation with domain generalization capabilities [48]. Also, Dutt et al. investigate PEFT algorithms across both convolutional and large transformer-based networks for medical image classification, demonstrating the effectiveness of PEFT, particularly in the low-data regimes common in medical imaging [14]. In this paper, we focus on fine-tuning VFs for image classification in the presence of noisy labels using adapters. Rather than introducing a new adapter, we utilize an existing adapter within our training paradigm.
+
+Learning with noisy label. Deep neural networks have demonstrated remarkable performance on large-scale datasets. However, it is well-known that neural networks can easily memorize noisy labeled samples, leading to degraded performance. Several studies have been conducted to explore robust supervised learning in the presence of noisy labels. These studies can be categorized into five approaches [42]: (i) robust architectures, (ii) robust regularization, (iii) robust loss functions [15, 33, 46, 59], (iv) loss adjustment [26, 31, 32], and (v) sample selection [24, 36, 17, 56, 50]. In this paper, we categorize our method as a sample selection method, which selects samples with clean labels from a noisy training dataset. While previous sample selection methods typically consider training from scratch, we focus on training starting from a pre-trained model, which is known to be more robust to noisy labels [21]. Additionally, research has explored using CLIP to enable robust training by leveraging its text-image matching capability on noisy datasets [49].
+
+Various methods have been proposed for clean sample selection from noisy datasets. For example, MentorNet introduced the use of a teacher network to guide the student network to focus on clean labels [24]. Similarly, Decoupling proposed updating two networks by using only the samples with differing predictions between them [36]. Co-teaching also trains two networks simultaneously, updating them based on sample recommendations from each other [17]. Co-teaching+ [56] improved upon Co-teaching by introducing the "update by disagreement" strategy, where only the samples with differing predictions between the two networks are used. More recently, Xia et al. proposed CoDis, an extension of Co-teaching+ that employs an "update by discrepancy" strategy, selecting samples with high-discrepancy prediction probabilities between the two networks to utilize more samples [50]. These methods are based on the assumption that clean samples can be identified using certain criteria, and that network collaboration is more stable than self-selection, which may lead to error accumulation. In this paper, we develop our method based on same assumption, but we assume that we start the learning process from the pre-trained VFM.
+
+# 3 Problem Setup
+
+We consider a  $k$ -class classification task using a neural network. Let  $\mathcal{X} \in \mathbb{R}^d$  denote the input space and  $\mathcal{Y} \in \mathbb{R}^k$  represent the ground-truth label space. In a typical classification task, the neural network is trained to align the input space with the label space. To this end, a training dataset  $D = \{(x_i, \hat{y}_i)\}_{i=1}^n$  is used for supervised learning with cross-entropy loss. In practice, a sample  $(x_i, \hat{y}_i)$  is considered as a noisy labeled sample when human-annotated label  $\hat{y}_i$  does not match the true label  $y_i$ . The objective of this paper is to develop a fine-tuning approach for VFMs that is robust to noise and capable of performing accurately on noisy datasets.
+
+Given a pre-trained VFM with parameters  $\theta_{\mathrm{VFM}}$ , consisting of a sequence of layers (e.g., attention blocks in ViT [13])  $L_{1}, L_{2}, \ldots, L_{M}$ , where  $M$  is the depth of  $\theta_{\mathrm{VFM}}$ , the learning objective for a classification problem can be formulated as:
+
+$$
+\min  _ {\theta_ {t}} \sum_ {i = 1} ^ {n} \mathcal {L} _ {c e} \left(p \left(x _ {i} \mid \theta_ {\mathrm {V F M}}, \theta_ {t}\right), \hat {y} _ {i}\right), \tag {1}
+$$
+
+where  $\theta_{t}$  and  $\mathcal{L}_{ce}$  represent the parameters targeted for updating and the cross-entropy loss, respectively. Here,  $p(\cdot |\theta)$  refers to the prediction for a given input using parameters  $\theta$ . We refer to the training process as linear probing when  $\theta_{t}$  is limited to the parameters of the linear layer  $\theta_{l}$ . Additionally, we refer to the training process as full-tuning when  $\theta_{t}$  includes  $\theta_{\mathrm{VFM}}$ , and as adapter tuning when it includes adapter parameters  $\theta_{a}$ , which are not part of  $\theta_{\mathrm{VFM}}$ .
+
+# 4 Method
+
+In this section, we begin by describing the adapter method for fine-tuning the VFM in Section 4.1. Following this, in Section 4.2, we introduce our method, CUFIT, which utilizes three modules: a linear layer and two adapters, to combat noisy labels. The key idea behind CUFIT is to leverage the well-pre-trained features of the VFM without updating the feature extractor when handling corrupted samples. Subsequently, the adapters are trained using the samples selected in a curriculum-based training manner, as shown in Figure 2 (i.e., linear probing  $\rightarrow$  intermediate adapter  $\rightarrow$  last adapter). This approach helps increase the number of selected samples by reducing the domain gap between the pretraining task and the medical image task. It is important to note that our framework does not train the modules sequentially (i.e., where one module starts training only after another finishes); instead, it trains the modules simultaneously on the current batch, similar to multi-task training.
+
+# 4.1 Learning with Adapter
+
+We consider various adapters for fine-tuning VFMs on medical image datasets. In particular, adapters like Visual Prompt Tuning (VPT [23]), AdaptFormer [9], Low Rank Adaptation (LoRA [22]) and Rein [48] can be used. These methods have been shown to be efficient for various image and video tasks, even compared to full model training [9, 48]. Typically, when an adapter is used for fine-tuning, the parameters of the VFM are frozen and not included in the optimization process.
+
+In this section, we briefly introduce how an adapter works. Note that our goal is not to propose a novel adapter but rather to present a training paradigm that can be applied to various adapters. For
+
+![](images/95ecfc2225c99cc29f55d181122e1b7a1318f79cf57984dfafe3e3c343426d4c.jpg)  
+Figure 2: Illustration of our proposed training framework, CUFIT, which consists of a pre-trained VFM and three distinct modules: (a) the linear probing module (LPM), (b) the intermediate adapter module (IAM), and (c) the last adapter module (LAM). During the training stage, the LPM selects clean samples for the IAM based on the agreement criterion, and the IAM selects clean samples for the LPM. During the inference stage, only the LAM is used for prediction.
+
+vision transformers (ViTs), the output of the attention block for the given input patches is calculated as follows:
+
+$$
+x _ {l} ^ {\prime} = \operatorname {A t t e n t i o n} (Q, K, V) = \operatorname {S o f t m a x} \left(\frac {Q K ^ {T}}{\sqrt {d}} V\right) + x _ {l - 1}, \tag {2}
+$$
+
+where  $x_{l-1}$  is the output token of the previous block. Here,  $Q$ ,  $K$ , and  $V$  refer to the query, key, and value vectors, respectively, which are derived from linear projection and LayerNorm [6] applied to  $x_{l-1}$ . The final output of the block,  $x_l$ , is then computed using LayerNorm and an MLP. Without using an adapter, this process is formulated as:
+
+$$
+x _ {l} = \operatorname {M L P} \left(\operatorname {L N} \left(x _ {l} ^ {\prime}\right)\right) + x _ {l} ^ {\prime}, \tag {3}
+$$
+
+where  $x_{l}$  is the output token of the  $l$ -th block. When an adapter is used, the Eq 3 is replaced by the following:
+
+$$
+x _ {l} = \operatorname {M L P} \left(\operatorname {L N} \left(x _ {l} ^ {\prime}\right)\right) + x _ {l} ^ {\prime} + \operatorname {A d a p t} \left(x _ {l - 1}; \theta_ {l - 1} ^ {a}\right), \tag {4}
+$$
+
+where  $\mathrm{Adapt}(\cdot ;\theta_l^a)$  refers to the adapter function for the  $l$  -th layer, parameterized by  $\theta_l^a$  . We consider this process to be an arbitrary function, as various adapters can be used.In the last block, the [CLS] token is passed to the following linear layer for final image classification.
+
+# 4.2 Curriculum training of three different modules
+
+We consider a pre-trained VFM,  $\theta_{\mathrm{VFM}}$ , with a single linear layer parameterized by  $\theta_{\mathrm{LPM}} \in \mathbb{R}^{c \times k}$ , an intermediate adapter module parameterized by  $\theta_{\mathrm{IAM}}$ , and a last adapter module parameterized by  $\theta_{\mathrm{LAM}}$ , where  $c$  refers to the dimension of the class token (e.g., 384 dimensions for the ViT-small architecture). Then, we propose a curriculum training framework for these three modules, in which the LPM is trained with all samples from the given batch, while the adapter modules are trained with filtered samples selected by their corresponding module using the agreement criterion. The agreement criterion refers to a method where a sample is considered clean if the module's prediction matches the sample's annotation. The idea behind this criterion is that a robust classifier will correctly predict the sample under the assumption that clean labels are in the majority within a noisy class. Therefore, a sample is selected as clean if it meets the agreement criterion (e.g., a "dog" image with a "dog" annotation). Thus, we build the curriculum training framework based on the robustness of the LPM against noisy labels using the agreement criterion.
+
+In particular, the linear probing module (LPM) is trained as follows:
+
+$$
+\min  _ {\theta_ {\mathrm {L P M}}} \sum_ {i = 1} ^ {n} \mathcal {L} _ {c e} \left(p \left(x _ {i} \mid \theta_ {\mathrm {V F M}}, \theta_ {\mathrm {L P M}}\right), \hat {y} _ {i}\right), \tag {5}
+$$
+
+which directly represents supervised learning using the given images and corresponding labels. Here,  $p(x_{i}|\theta_{\mathrm{VFM}},\theta_{\mathrm{LPM}})$  refers to the output of the network using  $\theta_{\mathrm{VFM}}$  and  $\theta_{\mathrm{LPM}}$  for the given image  $x_{i}$ . During the training stage, the intermediate adapter module (IAM) is trained as follows:
+
+$$
+\min  _ {\theta_ {\mathrm {I A M}}} \sum_ {i = 1} ^ {n} \mathcal {L} _ {c e} \left(p \left(x _ {i} \mid \theta_ {\mathrm {V F M}}, \theta_ {\mathrm {I A M}}\right), \hat {y} _ {i}\right) \mathbb {1} \left\{\arg \max  p \left(x _ {i} \mid \theta_ {\mathrm {V F M}}, \theta_ {\mathrm {L P M}}\right) = \hat {y} _ {i} \right\}, \tag {6}
+$$
+
+where  $\mathbb{1}\{\cdot\}$  is the indicator function. This simple modification using the indicator function ensures that the adapter module is trained only on selected samples chosen by the linear layer. Finally, the last adapter module (LAM) is trained as follows:
+
+$$
+\min  _ {\theta_ {\mathrm {L A M}}} \sum_ {i = 1} ^ {n} \mathcal {L} _ {c e} \left(p \left(x _ {i} \mid \theta_ {\mathrm {V F M}}, \theta_ {\mathrm {L A M}}\right), \hat {y} _ {i}\right) \mathbb {1} \left\{\arg \max  p \left(x _ {i} \mid \theta_ {\mathrm {V F M}}, \theta_ {\mathrm {I A M}}\right) = \hat {y} _ {i} \right\}. \tag {7}
+$$
+
+The Eq 7 is equivalent to Eq 6, but it uses LAM and IAM instead of IAM and LPM, respectively. This simple yet effective sample selection strategy is well-suited for fine-tuning the VFM on noisy image datasets. Notably, it does not require any hyperparameters like the estimated noise rate, which are commonly needed in previous works [17, 47, 50], where they assume the noise rate is known in order to select small-loss samples (e.g., selecting  $60\%$  of samples in a batch for a known noise rate of  $40\%$ ). After training is completed, only the last adapter module is used to predict the given test image.
+
+# 5 Experiments
+
+# 5.1 Settings
+
+Datasets. We evaluate our approach on four simulated noisy label medical multi-class image classification benchmarks: HAM10000 [44], APTOS-2019 [11], BloodMnist [3], and OrgancMnist [52]. Additionally, we conduct an evaluation on a real-world noisy label benchmark, Kaggle-EyePACS [2]. In particular, the detail of datasets are as follows:
+
+- HAM10000 [44]: This dataset contains 10,015 dermatoscopic images for skin lesion classification, with each image classified into one of seven possible disease categories. We use all the images for training, and the evaluation is conducted using the 1,512 test images provided by the ISIC 2018 challenge [1].  
+- APTOS-2019 [11]: This dataset consists of 3,662 retina images taken with fundus photography under various imaging conditions. Each image is rated for the severity of diabetic retinopathy (DR) on a scale from 0 to 4. We use 2,930 images for training and 366 images for evaluation.  
+- BloodMnist [3]: This dataset contains 17,092 images of individual cells, with each image annotated as one of eight possible cell types. We use 11,959 images for training and 3,421 images for evaluation.  
+- OrgancMnist [52]: This dataset includes 23,538 images that are center-sliced from the Hounsfield-Unit of 3D images in a coronal view. Each image is labeled as one of eleven body organs. We use 12,975 images for training and 8,216 images for evaluation.  
+- Kaggle-EyePACS [2]: This Kaggle competition dataset provides 35,126 retina images categorized into five DR severity grades for training, which are known to contain noisy labels [25]. Specifically, some DR category labels (e.g., mild DR labeled as moderate DR) are noisy, and some images considered normal may actually contain retinal diseases such as glaucoma or drusen, which are not included in the classification categories. It is estimated that there is approximately a  $30\% -40\%$  label error in this dataset [25, 45]. We use the original 35,126 training images and their annotations for training, and all images from APTOS-2019 for evaluation. Additionally, we use the FGADR [60] dataset for further evaluation.
+
+Baselines. We compare the performance of CUFIT with basic training paradigms: full training, linear probing, and fine-tuning with Rein [48]. Additionally, we evaluate our approach against other training-based methods, including Co-teaching [17], JoCor [47], and CoDis [50]. Like ours, these methods do not modify the training loss or architecture. Specifically, Co-teaching trains two networks simultaneously, with each network selecting small-loss samples from its peer's predictions to guide
+
+<table><tr><td rowspan="2">Dataset</td><td rowspan="2">Noise rate</td><td colspan="7">Method</td></tr><tr><td>Full-training</td><td>Linear probing</td><td>Rein</td><td>Co-teaching</td><td>JoCor</td><td>CoDis</td><td>CUFIT</td></tr><tr><td rowspan="5">HAM10000</td><td>0.1</td><td>66.5</td><td>75.6</td><td>78.6</td><td>81.5</td><td>81.1</td><td>81.9</td><td>82.6</td></tr><tr><td>0.2</td><td>62.6</td><td>75.3</td><td>72.1</td><td>79.1</td><td>79.4</td><td>80.1</td><td>81.5</td></tr><tr><td>0.4</td><td>56.1</td><td>71.0</td><td>54.9</td><td>74.3</td><td>73.9</td><td>74.1</td><td>79.1</td></tr><tr><td>0.6</td><td>59.9</td><td>61.9</td><td>37.8</td><td>67.3</td><td>67.1</td><td>66.1</td><td>70.1</td></tr><tr><td>Mean</td><td>61.3</td><td>71.0</td><td>60.8</td><td>75.5</td><td>75.4</td><td>75.5</td><td>78.3</td></tr><tr><td rowspan="5">APTOS-2019</td><td>0.1</td><td>66.8</td><td>79.2</td><td>82.5</td><td>82.8</td><td>84.8</td><td>83.2</td><td>84.2</td></tr><tr><td>0.2</td><td>65.9</td><td>79.4</td><td>78.7</td><td>81.2</td><td>83.1</td><td>82.0</td><td>84.2</td></tr><tr><td>0.4</td><td>69.9</td><td>79.5</td><td>77.2</td><td>79.5</td><td>76.0</td><td>79.5</td><td>81.6</td></tr><tr><td>0.6</td><td>48.2</td><td>66.9</td><td>42.0</td><td>72.9</td><td>74.2</td><td>75.7</td><td>76.3</td></tr><tr><td>Mean</td><td>62.7</td><td>76.3</td><td>68.9</td><td>79.1</td><td>79.5</td><td>80.1</td><td>81.6</td></tr><tr><td rowspan="5">BloodMnist</td><td>0.1</td><td>95.4</td><td>97.2</td><td>95.9</td><td>98.6</td><td>98.5</td><td>98.5</td><td>99.0</td></tr><tr><td>0.2</td><td>93.9</td><td>96.7</td><td>89.0</td><td>97.6</td><td>97.3</td><td>97.2</td><td>98.8</td></tr><tr><td>0.4</td><td>91.8</td><td>95.8</td><td>69.3</td><td>93.7</td><td>93.0</td><td>93.5</td><td>98.3</td></tr><tr><td>0.6</td><td>87.9</td><td>90.3</td><td>45.6</td><td>88.7</td><td>87.3</td><td>88.0</td><td>98.2</td></tr><tr><td>Mean</td><td>92.3</td><td>95.0</td><td>75.0</td><td>94.7</td><td>94.0</td><td>94.3</td><td>98.6</td></tr><tr><td rowspan="5">OrgancMnist</td><td>0.1</td><td>85.3</td><td>83.3</td><td>87.4</td><td>92.1</td><td>92.1</td><td>92.1</td><td>93.7</td></tr><tr><td>0.2</td><td>79.9</td><td>82.9</td><td>82.0</td><td>90.9</td><td>91.9</td><td>90.7</td><td>93.6</td></tr><tr><td>0.4</td><td>72.1</td><td>79.9</td><td>63.8</td><td>85.8</td><td>85.3</td><td>85.8</td><td>91.6</td></tr><tr><td>0.6</td><td>64.5</td><td>72.2</td><td>43.1</td><td>82.8</td><td>82.6</td><td>81.9</td><td>87.4</td></tr><tr><td>Mean</td><td>75.5</td><td>79.6</td><td>69.1</td><td>87.9</td><td>88.0</td><td>87.6</td><td>91.6</td></tr></table>
+
+Table 1: Average test accuracy (%) on four simulated noisy datasets with different noise levels. The test accuracy is averaged over the last ten epochs. The best and second-best results in each case are highlighted in bold and underline, respectively.
+
+<table><tr><td rowspan="2">Testset</td><td colspan="7">Method</td></tr><tr><td>Full-training</td><td>Linear probing</td><td>Rein</td><td>Co-teaching</td><td>JoCor</td><td>CoDis</td><td>CUFIT</td></tr><tr><td>APTOS-2019</td><td>34.2</td><td>65.4</td><td>69.1</td><td>70.9</td><td>69.3</td><td>69.2</td><td>69.8</td></tr><tr><td>FGADR</td><td>14.3</td><td>46.4</td><td>48.8</td><td>44.9</td><td>53.1</td><td>53.0</td><td>53.7</td></tr><tr><td>Total</td><td>27.5</td><td>59.0</td><td>62.3</td><td>62.2</td><td>63.9</td><td>63.8</td><td>64.4</td></tr></table>
+
+Table 2: Average test accuracy (%) on real-world noisy datasets (Kaggle-EyePACS for training). After the training is done, we evaluate the model on two datasets: APTOS-2019 and FGADR. The best result and second-best result in each case are highlighted in bold and underline, respectively.
+
+learning. JoCor extends this idea by incorporating co-regularization to maximize agreement between the two networks. CoDis further refines this process by selecting samples that not only have small losses but also show high divergence between the two networks. It is important to note that we do not compare our proposed framework with state-of-the-art methods that modify the training loss (e.g., semi-supervised learning) or model architecture [51, 8]. In the experiments, we apply these methods to VFMs with adapters, as they do not require specific model architectures, and VFMs with adapters outperform the linear probing of VFMs (i.e., DINOv2 with the Rein adapter is used as the default setting for training with Co-teaching, JoCor, and CoDis for a fair comparison).
+
+Implementation details. For the experiments, we use DINOv2 [37] with the ViT-small [13] backbone as our basic vision foundation model. Additionally, we use Rein [48] as the fine-tuning adapter, originally proposed for domain-generalized semantic segmentation of VFMs. In our setup, we utilize the class token from the block for classification, rather than the patch tokens from multiple blocks.
+
+We use the PyTorch [4] codebase for our experiments. BloodMnist and OrgancMnist datasets are sourced from MedMnist [54, 55]. We use the ViT-small architecture and the Adam optimizer [27]. All training runs for 100 epochs with a batch size of 32. The initial learning rate is 0.001, which decays by a factor of 10 at epochs 50, 75, and 90. For full-parameter training, however, we start with an initial learning rate of 0.0001. For the simulated noisy label benchmarks, we generate symmetric noise [17] for evaluation, with noise rates set at  $10\%$ ,  $20\%$ ,  $40\%$ , and  $60\%$ .
+
+# 5.2 Simulated noisy medical image classification benchmark
+
+First, we evaluate our framework on simulated noisy label benchmarks using four medical datasets. The average classification test accuracy for each dataset is provided in Table 1. Our framework consistently outperforms previous baselines, demonstrating its effectiveness under noisy labels by leveraging the pre-trained features of DINOv2 and the Rein adapter. Notably, our framework proves
+
+![](images/5d8c038c6f3335ec64d33aa35ec190b5d9c157e62e4773b58fd574acb6984ae8.jpg)  
+Figure 3: Illustration of label precision (a,d), label recall (b,e), and test accuracy (c,f) vs. epoch. The first row is for HAM10000 with  $40\%$  noise rate, and the second row is for APTOS-2019 with  $40\%$  noise rate.
+
+to be more effective as the noise rate increases. For example, CUFIT achieves  $0.85\%$  relatively higher accuracy than CoDis on HAM10000 with a  $10\%$  noise rate, while the improvement rises to  $3.7\%$  at a  $60\%$  noise rate. This result indicates that the pre-trained features of the VFM are particularly useful for handling noisy labels in the given datasets.
+
+# 5.3 Real-world noisy medical image classification benchmark
+
+We train DINOv2 with Rein on the real-world benchmark, using the Kaggle-EyePACS dataset for training and the APTOS-2019 and FGADR datasets for testing. Given the highly imbalanced training set (e.g., approximately  $73\%$  of the samples are labeled as the normal class), we use weighted cross-entropy loss to train the model. Since previous sample selection methods require a noise rate hyperparameter, we employed the noise estimation method from [34], following Co-teaching [17].
+
+In Table 2, we report the classification accuracy on the APTOS-2019 and FGADR datasets, as well as the overall accuracy across both datasets. Our method outperforms other baselines on the FGADR dataset and the combined dataset, while Co-teaching achieves the highest accuracy on the APTOS-2019 dataset. We believe this discrepancy is due to the distribution of normal class samples—approximately  $50\%$  in APTOS-2019 and about  $5\%$  in FGADR. Co-teaching performs well in classifying the normal class, whereas our method excels at classifying diseased samples. For example, our method achieves  $53.9\%$  macro-average test accuracy, while Co-teaching achieves  $48.5\%$  on the combined test set.
+
+# 6 Discussion
+
+# 6.1 How does CUFIT works?
+
+So far, we have demonstrated through empirical results that our framework significantly improves the robustness of VFM fine-tuning against noisy labels. However, we have not yet discussed why our framework is effective in learning with noisy labels. In Figure 3, we present label precision, label recall, and test accuracy over the number of epochs to illustrate how our framework functions. In principle, higher label precision indicates fewer noisy samples in the selected data, while higher label recall indicates fewer clean samples in the unselected data. We have following three observations:
+
+![](images/352593a1b48781d07003757ebda2bc9e7f243987cad335e57a153d9d15cf2207.jpg)  
+(a) HAM10000 with  $40\%$  noise rate
+
+![](images/3852cd54f2a695a490abccbbef62c4ff4bd59d5f495ca5457f2c1042461ff011.jpg)  
+(b) APTOS-2019 with  $40\%$  noise rate  
+Figure 4: Test accuracy of our method with various VFMs (DINOv1 [7], MAE [18], DINOv2 [37]) and adapters (VPT [23], AdaptFormer [9], Rein [48]). We use HAM10000 and APTOS-2019 with  $40\%$  noise rate for training.
+
+<table><tr><td rowspan="2">Dataset</td><td rowspan="2">Noise Rate</td><td colspan="2">Full-training</td><td colspan="2">Linear probing</td><td colspan="2">CoDis</td><td colspan="3">CUFIT</td></tr><tr><td>ResNet</td><td>DINOv2</td><td>ResNet</td><td>DINOv2</td><td>ResNet</td><td>DINOv2</td><td>ResNet</td><td>ResNet+rein</td><td>DINOv2</td></tr><tr><td rowspan="2">HAM10000</td><td>0.2</td><td>73.1</td><td>66.5</td><td>71.1</td><td>75.6</td><td>74.9</td><td>80.1</td><td>77.7</td><td>79.9</td><td>82.6</td></tr><tr><td>0.4</td><td>59.6</td><td>62.6</td><td>67.8</td><td>75.3</td><td>72.4</td><td>74.1</td><td>73.8</td><td>75.4</td><td>81.5</td></tr><tr><td rowspan="2">APTOS-2019</td><td>0.2</td><td>80.4</td><td>66.8</td><td>80.3</td><td>79.2</td><td>80.3</td><td>82.0</td><td>82.4</td><td>82.2</td><td>84.2</td></tr><tr><td>0.4</td><td>64.7</td><td>65.9</td><td>73.4</td><td>79.4</td><td>78.1</td><td>79.5</td><td>80.5</td><td>82.2</td><td>84.2</td></tr></table>
+
+Table 3: Average test accuracy on simulated noisy datasets (HAM10000 and APTOS-2019) using the ResNet50 architecture. Test accuracy is averaged over the last ten epochs.
+
+- We find that CoDis exhibits lower label precision during training compared to LPM and IAM, suggesting that previous sample selection methods fail to effectively utilize the pre-trained features of VFMs, leading to lower test accuracy. These methods train the network on all training data without sample selection during the early stages of training (e.g., epochs 1 to 10 in their default setting). However, this approach may harm the feature extraction capability of VFMs and result in degraded performance.  
+- LPM consistently achieves the highest label precision across epochs but has the lowest label recall, indicating that it effectively prevents the memorization of noisy samples. However, because the feature extractor remains unchanged, its overall accuracy is limited, thus selecting only a small number of clean samples.  
+- IAM, on the other hand, achieves similar label precision but higher label recall by leveraging the adapter module, which contributes to the improved test accuracy of LAM. This suggests that by adapting the feature extractor through training the adapter on a few certain clean samples, IAM can be a more accurate module. It can then provide more clean samples to LAM, resulting in better overall performance.
+
+# 6.2 Performance comparison across various VFMs and adapters
+
+To validate the performance of CUFIT in various settings, we present experimental results using three VFMs and adapters in Figure 4. We use the same experimental setup (i.e., 100 epochs with the Adam optimizer) to train the network across all backbones and adapters. Specifically, we utilize four backbones, including DINOv1 [7], MAE [18], DINOv2 [37], and BioMedCLIP [58] and four adapters, including VPT [23], AdaptFormer [9], LoRA [22], and Rein [48]. BioMedClip is a CLIP-like model trained with the PMC-15M dataset, which contains 15 million biomedical image-text pairs collected from 4.4 million scientific articles. Our results demonstrate that our framework consistently helps build a robust classifier across different VFMs and adapters. For example, our framework achieves better performance compared to both adapter-based methods and linear probing. Additionally, we observe that linear probing consistently outperforms the adapter method in all cases, indicating that the performance of adapters can be degraded by noisy labels across various adapters.
+
+<table><tr><td rowspan="2">Dataset</td><td rowspan="2">Noise rate</td><td colspan="7">Method</td></tr><tr><td>Full-training</td><td>Linear probing</td><td>Rein</td><td>Co-teaching</td><td>JoCor</td><td>CoDis</td><td>CUFIT</td></tr><tr><td>CIFAR10</td><td rowspan="2">0.8</td><td>25.9</td><td>79.0</td><td>24.8</td><td>78.2</td><td>75.7</td><td>76.3</td><td>83.9</td></tr><tr><td>CIFAR100</td><td>6.3</td><td>59.6</td><td>25.6</td><td>66.7</td><td>64.2</td><td>63.7</td><td>73.8</td></tr><tr><td>ANIMAL10N</td><td>0.08*</td><td>74.5</td><td>89.1</td><td>88.0</td><td>92.2</td><td>91.9</td><td>91.7</td><td>92.3</td></tr></table>
+
+Table 4: Average test accuracy on the natural image dataset with simulated noisy labels (CIFAR, symmetric noise at  $80\%$ ) and real-world noisy labels (ANIMAL10N [41], which has an estimated noise ratio of  $8\%$ ). The test accuracy is averaged over the last ten epochs. We use DINOv2 with Rein adapter for the experiment. Bold values the best result.
+
+# 6.3 Performance on CNNs with adapters
+
+We designed CUFIT for VFMs due to their strong pre-trained feature extraction capabilities, enabled by self-supervised training on large datasets. However, CNN-based architectures like ResNet [20] also utilize pre-trained weights instead of starting from scratch. Therefore, we validate CUFIT on ImageNet [12] pre-trained ResNet50, with and without the Rein adapter modified for ResNet. The experimental results are shown in Table 3. We observe that our method outperforms other training paradigms when using the ImageNet pre-trained ResNet architecture. Additionally, the Rein adapter for ResNet improves performance, demonstrating that using fewer trainable parameters with an adapter, compared to full training, is beneficial for combating noisy labels. Finally, we show that the more representative pre-trained features of DINOv2 outperform the ImageNet pre-trained features across all training methods.
+
+# 6.4 Performance on noisy natural image classification benchmark
+
+Since our framework is easily applicable not only to medical image classification but also to natural image classification, we present experimental results on the CIFAR [30] simulated noisy classification benchmarks and ANIMAL10N [41] real-world noisy classification benchmark in Table 4. We validate our framework under an extremely high noise rate setting (80%) for CIFAR benchmark, as it is intuitive that our framework performs well under low noise rates due to the feature extraction capabilities of VFM. As shown in Table 4, our framework outperforms other sample selection methods in natural image classification benchmarks as well. This demonstrates the effectiveness of our framework, highlighting that using well pre-trained VFMs is beneficial for detecting noisy labels in natural images, as expected.
+
+# 7 Conclusion
+
+This paper presents a curriculum fine-tuning paradigm called CUFIT, designed to robustly fine-tune vision foundation models (VFMs) for medical image classification. Our framework is based on the insight that linear probing of VFs is robust to noisy labels, as it does not modify the feature extraction process. Building on this, CUFIT consists of three training modules: the linear probing module (LPM), the intermediate adapter module (IAM), and the last adapter module (LAM). These modules are trained simultaneously, with each selecting clean samples for the next module. Specifically, while the LPM is trained on all samples, the LPM and IAM select clean samples for the IAM and LAM, respectively (i.e., LPM  $\rightarrow$  IAM  $\rightarrow$  LAM). Experiments demonstrate that CUFIT significantly improves the performance of VFs in the presence of noisy labels for medical image classification. Additionally, we provide extensive analyses to enhance the understanding of CUFIT. We hope our insights inspire future research to further explore the robustness of vision foundation models when learning with noisy labels for various medical imaging tasks.
+
+# Acknowledgments
+
+This work was partly supported by Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government(MSIT) (No. RS-2022-II0951, Development of Uncertainty-Aware Agents Learning by Asking Questions,  $90\%$  ) and Institute of Civil Military Technology Cooperation funded by the Defense Acquisition Program Administration and Ministry of Trade, Industry and Energy of Korean government under grant No. 22-CM-GU-08,  $10\%$ ).
+
+# References
+
+[1] Isic 2018 challenge. https://challenge.isic-archive.com/landing/2018/.  
+[2] Kaggle diabetic retinopathy detection competition. https://www.kaggle.com/c/ diabetic-retinopathy-detection.  
+[3] Andrea Acevedo, Anna Merino González, Edwin Santiago Alférez Baquero, Ángel Molina Borras, Laura Boldú Nebot, and José Rodellar Benedé. A dataset of microscopic peripheral blood cell images for development of automatic recognition systems. Data in brief, 30(article 105474), 2020.  
+[4] Jason Ansel, Edward Yang, Horace He, Natalia Gimelshein, Animesh Jain, Michael Voznesensky, Bin Bao, Peter Bell, David Berard, Evgeni Burovski, Geeta Chauhan, Anjali Chourdia, Will Constable, Alban Desmaison, Zachary DeVito, Elias Ellison, Will Feng, Jiong Gong, Michael Gschwind, Brian Hirsh, Sherlock Huang, Kshitteej Kalambarkar, Laurent Kirsch, Michael Lazos, Mario Lezcano, Yanbo Liang, Jason Liang, Yinghai Lu, CK Luk, Bert Maher, Yunjie Pan, Christian Puhrsch, Matthias Reso, Mark Saroufim, Marcos Yukio Siraichi, Helen Suk, Michael Suo, Phil Tillet, Eikan Wang, Xiaodong Wang, William Wen, Shunting Zhang, Xu Zhao, Keren Zhou, Richard Zou, Ajit Mathews, Gregory Chanan, Peng Wu, and Soumith Chintala. PyTorch 2: Faster Machine Learning Through Dynamic Python Bytecode Transformation and Graph Compilation. In 29th ACM International Conference on Architectural Support for Programming Languages and Operating Systems, Volume 2 (ASPLOS '24). ACM, April 2024.  
+[5] Devansh Arpit, Stanisław Jastrzebski, Nicolas Ballas, David Krueger, Emmanuel Bengio, Maxinder S Kanwal, Tegan Maharaj, Asja Fischer, Aaron Courville, Yoshua Bengio, et al. A closer look at memorization in deep networks. In International conference on machine learning, pages 233-242. PMLR, 2017.  
+[6] Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. Layer normalization. arXiv preprint arXiv:1607.06450, 2016.  
+[7] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021.  
+[8] Bingzhi Chen, Zhanhao Ye, Yishu Liu, Zheng Zhang, Jiahui Pan, Biqing Zeng, and Guangming Lu. Combating medical label noise via robust semi-supervised contrastive learning. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 562-572. Springer, 2023.  
+[9] Shoufa Chen, Chongjian Ge, Zhan Tong, Jiangliu Wang, Yibing Song, Jue Wang, and Ping Luo. Adaptformer: Adapting vision transformers for scalable visual recognition. Advances in Neural Information Processing Systems, 35:16664-16678, 2022.  
+[10] Xi Chen, Lianghua Huang, Yu Liu, Yujun Shen, Deli Zhao, and Hengshuang Zhao. Anydoor: Zero-shot object-level image customization. arXiv preprint arXiv:2307.09481, 2023.  
+[11] Omar Dekhil, Ahmed Naglah, Mohamed Shaban, Mohammed Ghazal, Fatma Taher, and Ayman Elbaz. Deep learning based method for computer aided diagnosis of diabetic retinopathy. In 2019 IEEE International Conference on Imaging Systems and Techniques (IST), pages 1-4. IEEE, 2019.  
+[12] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248–255. IEEE, 2009.  
+[13] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2021.  
+[14] Raman Dutt, Linus Ericsson, Pedro Sanchez, Sotirios A Tsaftaris, and Timothy Hospedales. Parameter-efficient fine-tuning for medical image analysis: The missed opportunity. In Medical Imaging with Deep Learning.  
+[15] Aritra Ghosh, Himanshu Kumar, and P Shanti Sastry. Robust loss functions under label noise for deep neural networks. In Proceedings of the AAAI conference on artificial intelligence, volume 31, 2017.  
+[16] Ross Girshick. Fast r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 1440-1448, 2015.
+
+[17] Bo Han, Quanming Yao, Xingrui Yu, Gang Niu, Miao Xu, Weihua Hu, Ivor Tsang, and Masashi Sugiyama. Co-teaching: Robust training of deep neural networks with extremely noisy labels. Advances in neural information processing systems, 31, 2018.  
+[18] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000-16009, 2022.  
+[19] Kaiming He, Georgia Gkioxari, Piotr Dolkar, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017.  
+[20] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016.  
+[21] Dan Hendrycks, Kimin Lee, and Mantas Mazeika. Using pre-training can improve model robustness and uncertainty. In International conference on machine learning, pages 2712-2721. PMLR, 2019.  
+[22] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. In International Conference on Learning Representations, 2022.  
+[23] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In European Conference on Computer Vision, pages 709-727. Springer, 2022.  
+[24] Lu Jiang, Zhengyuan Zhou, Thomas Leung, Li-Jia Li, and Li Fei-Fei. Mentornet: Learning data-driven curriculum for very deep neural networks on corrupted labels. In International conference on machine learning, pages 2304–2313. PMLR, 2018.  
+[25] Lie Ju, Xin Wang, Lin Wang, Dwarikanath Mahapatra, Xin Zhao, Quan Zhou, Tongliang Liu, and Zongyuan Ge. Improving medical images classification with label noise using dual-uncertainty estimation. IEEE transactions on medical imaging, 41(6):1533–1546, 2022.  
+[26] Taehyeon Kim, Jongwoo Ko, JinHwan Choi, Se-Young Yun, et al. Fine samples for learning with noisy labels. Advances in Neural Information Processing Systems, 34:24137-24149, 2021.  
+[27] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. CoRR, abs/1412.6980, 2014.  
+[28] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015-4026, 2023.  
+[29] Jonathan Krause, Benjamin Sapp, Andrew Howard, Howard Zhou, Alexander Toshev, Tom Duerig, James Philbin, and Li Fei-Fei. The unreasonable effectiveness of noisy data for fine-grained recognition. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part III 14, pages 301-320. Springer, 2016.  
+[30] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009.  
+[31] Junnan Li, Richard Socher, and Steven CH Hoi. Dividemix: Learning with noisy labels as semi-supervised learning. arXiv preprint arXiv:2002.07394, 2020.  
+[32] Junnan Li, Caiming Xiong, and Steven CH Hoi. Learning from noisy data with robust representation learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9485-9494, 2021.  
+[33] Sheng Liu, Jonathan Niles-Weed, Narges Razavian, and Carlos Fernandez-Granda. Early-learning regularization prevents memorization of noisy labels. Advances in neural information processing systems, 33:20331-20342, 2020.  
+[34] Tongliang Liu and Dacheng Tao. Classification with noisy labels by importance reweighting. IEEE Transactions on pattern analysis and machine intelligence, 38(3):447-461, 2015.  
+[35] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021.
+
+[36] Eran Malach and Shai Shalev-Shwartz. Decoupling" when to update" from" how to update". Advances in neural information processing systems, 30, 2017.  
+[37] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023.  
+[38] Yu Qiao, Chaoning Zhang, Taegoo Kang, Donghun Kim, Shehbaz Tariq, Chenshuang Zhang, and Choong Seon Hong. Robustness of sam: Segment anything under corruptions and beyond. arXiv preprint arXiv:2306.07713, 2023.  
+[39] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021.  
+[40] Poulami Sinhamahapatra, Franziska Schwaiger, Shirsha Bose, Huiyu Wang, Karsten Roscher, and Stephan Guennemann. Finding dino: A plug-and-play framework for unsupervised detection of out-of-distribution objects using prototypes. arXiv preprint arXiv:2404.07664, 2024.  
+[41] Hwanjun Song, Minseok Kim, and Jae-Gil Lee. SELFFIE: Refurbishing unclean samples for robust deep learning. In Kamalika Chaudhuri and Ruslan Salakhutdinov, editors, Proceedings of the 36th International Conference on Machine Learning, volume 97 of Proceedings of Machine Learning Research, pages 5907-5915. PMLR, 09-15 Jun 2019.  
+[42] Hwanjun Song, Minseok Kim, Dongmin Park, Yooju Shin, and Jae-Gil Lee. Learning from noisy labels with deep neural networks: A survey. IEEE transactions on neural networks and learning systems, 2022.  
+[43] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In International conference on machine learning, pages 10347-10357. PMLR, 2021.  
+[44] Philipp Tschandl, Cliff Rosendahl, and Harald Kittler. The ham10000 dataset, a large collection of multi-source dermatoscopic images of common pigmented skin lesions. Scientific data, 5(1):1-9, 2018.  
+[45] Xin Wang, Lie Ju, Xin Zhao, and Zongyuan Ge. Retinal abnormalities recognition using regional multitask learning. In Medical Image Computing and Computer Assisted Intervention-MICCAI 2019: 22nd International Conference, Shenzhen, China, October 13-17, 2019, Proceedings, Part I 22, pages 30-38. Springer, 2019.  
+[46] Yisen Wang, Xingjun Ma, Zaiyi Chen, Yuan Luo, Jinfeng Yi, and James Bailey. Symmetric cross entropy for robust learning with noisy labels. In Proceedings of the IEEE/CVF international conference on computer vision, pages 322-330, 2019.  
+[47] Hongxin Wei, Lei Feng, Xiangyu Chen, and Bo An. Combating noisy labels by agreement: A joint training method with co-regularization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13726-13735, 2020.  
+[48] Zhixiang Wei, Lin Chen, Yi Jin, Xiaoxiao Ma, Tianle Liu, Pengyang Ling, Ben Wang, Huaian Chen, and Jinjin Zheng. Stronger, fewer, & superior: Harnessing vision foundation models for domain generalized semantic segmentation. arXiv preprint arXiv:2312.04265, 2023.  
+[49] Cheng-En Wu, Yu Tian, Haichao Yu, Heng Wang, Pedro Morgado, Yu Hen Hu, and Linjie Yang. Why is prompt tuning for vision-language models robust to noisy labels? In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15488–15497, 2023.  
+[50] Xiaobo Xia, Bo Han, Yibing Zhan, Jun Yu, Mingming Gong, Chen Gong, and Tongliang Liu. Combating noisy labels with sample selection by mining high-discrepancy examples. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1833–1843, 2023.  
+[51] Xiaohan Xing, Zhen Chen, Zhifan Gao, and Yixuan Yuan. Gradient and feature conformity-steered medical image classification with noisy labels. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 75-84. Springer, 2023.  
+[52] Xuanang Xu, Fugen Zhou, Bo Liu, Dongshan Fu, and Xiangzhi Bai. Efficient multiple organ localization in ct image using 3d region proposal network. IEEE transactions on medical imaging, 38(8):1885-1898, 2019.
+
+[53] Cheng Xue, Lequan Yu, Pengfei Chen, Qi Dou, and Pheng-Ann Heng. Robust medical image classification from noisy labeled data with global and local representation guided co-training. IEEE transactions on medical imaging, 41(6):1371-1382, 2022.  
+[54] Jiancheng Yang, Rui Shi, and Bingbing Ni. Medmnist classification decathlon: A lightweight automl benchmark for medical image analysis. In IEEE 18th International Symposium on Biomedical Imaging (ISBI), pages 191-195, 2021.  
+[55] Jiancheng Yang, Rui Shi, Donglai Wei, Zequan Liu, Lin Zhao, Bilian Ke, Hanspeter Pfister, and Bingbing Ni. Medmnist v2-a large-scale lightweight benchmark for 2d and 3d biomedical image classification. Scientific Data, 10(1):41, 2023.  
+[56] Xingrui Yu, Bo Han, Jiangchao Yao, Gang Niu, Ivor Tsang, and Masashi Sugiyama. How does disagreement help generalization against label corruption? In International conference on machine learning, pages 7164-7173. PMLR, 2019.  
+[57] Li Yuan, Yunpeng Chen, Tao Wang, Weihao Yu, Yujun Shi, Zi-Hang Jiang, Francis EH Tay, Jiashi Feng, and Shuicheng Yan. Tokens-to-token vit: Training vision transformers from scratch onImagenet. In Proceedings of the IEEE/CVF international conference on computer vision, pages 558-567, 2021.  
+[58] Sheng Zhang, Yanbo Xu, Naoto Usuyama, Jaspreet Bagga, Robert Tinn, Sam Preston, Rajesh Rao, Mu Wei, Naveen Valluri, Cliff Wong, Matthew Lungren, Tristan Naumann, and Hoifung Poon. Large-scale domain-specific pretraining for biomedical vision-language processing, 2023.  
+[59] Zhilu Zhang and Mert Sabuncu. Generalized cross entropy loss for training deep neural networks with noisy labels. Advances in neural information processing systems, 31, 2018.  
+[60] Yi Zhou, Boyang Wang, Lei Huang, Shanshan Cui, and Ling Shao. A benchmark for studying diabetic retinopathy: segmentation, grading, and transferability. IEEE Transactions on Medical Imaging, 40(3):818-828, 2020.
+
+# NeurIPS Paper Checklist
+
+# 1. Claims
+
+Question: Do the main claims made in the abstract and introduction accurately reflect the paper's contributions and scope?
+
+Answer: [Yes]
+
+Justification: We provide the main claims in the introduction and validate it in the experiments and discussion sections.
+
+Guidelines:
+
+- The answer NA means that the abstract and introduction do not include the claims made in the paper.  
+- The abstract and/or introduction should clearly state the claims made, including the contributions made in the paper and important assumptions and limitations. A No or NA answer to this question will not be perceived well by the reviewers.  
+- The claims made should match theoretical and experimental results, and reflect how much the results can be expected to generalize to other settings.  
+- It is fine to include aspirational goals as motivation as long as it is clear that these goals are not attained by the paper.
+
+# 2. Limitations
+
+Question: Does the paper discuss the limitations of the work performed by the authors?
+
+Answer: [Yes]
+
+Justification: We discuss the limitation in the conclusion.
+
+Guidelines:
+
+- The answer NA means that the paper has no limitation while the answer No means that the paper has limitations, but those are not discussed in the paper.  
+- The authors are encouraged to create a separate "Limitations" section in their paper.  
+- The paper should point out any strong assumptions and how robust the results are to violations of these assumptions (e.g., independence assumptions, noiseless settings, model well-specification, asymptotic approximations only holding locally). The authors should reflect on how these assumptions might be violated in practice and what the implications would be.  
+- The authors should reflect on the scope of the claims made, e.g., if the approach was only tested on a few datasets or with a few runs. In general, empirical results often depend on implicit assumptions, which should be articulated.  
+- The authors should reflect on the factors that influence the performance of the approach. For example, a facial recognition algorithm may perform poorly when image resolution is low or images are taken in low lighting. Or a speech-to-text system might not be used reliably to provide closed captions for online lectures because it fails to handle technical jargon.  
+- The authors should discuss the computational efficiency of the proposed algorithms and how they scale with dataset size.  
+- If applicable, the authors should discuss possible limitations of their approach to address problems of privacy and fairness.  
+- While the authors might fear that complete honesty about limitations might be used by reviewers as grounds for rejection, a worse outcome might be that reviewers discover limitations that aren't acknowledged in the paper. The authors should use their best judgment and recognize that individual actions in favor of transparency play an important role in developing norms that preserve the integrity of the community. Reviewers will be specifically instructed to not penalize honesty concerning limitations.
+
+# 3. Theory Assumptions and Proofs
+
+Question: For each theoretical result, does the paper provide the full set of assumptions and a complete (and correct) proof?
+
+Answer: [NA]
+
+Justification: We do not provide the theoretical result in the paper.
+
+# Guidelines:
+
+- The answer NA means that the paper does not include theoretical results.  
+- All the theorems, formulas, and proofs in the paper should be numbered and cross-referenced.  
+- All assumptions should be clearly stated or referenced in the statement of any theorems.  
+- The proofs can either appear in the main paper or the supplemental material, but if they appear in the supplemental material, the authors are encouraged to provide a short proof sketch to provide intuition.  
+- Inversely, any informal proof provided in the core of the paper should be complemented by formal proofs provided in appendix or supplemental material.  
+- Theorems and Lemmas that the proof relies upon should be properly referenced.
+
+# 4. Experimental Result Reproducibility
+
+Question: Does the paper fully disclose all the information needed to reproduce the main experimental results of the paper to the extent that it affects the main claims and/or conclusions of the paper (regardless of whether the code and data are provided or not)?
+
+# Answer: [Yes]
+
+Justification: We provide the experimental setting of our paper. Also, we provide source code for experiments in supplementary.
+
+# Guidelines:
+
+- The answer NA means that the paper does not include experiments.  
+- If the paper includes experiments, a No answer to this question will not be perceived well by the reviewers: Making the paper reproducible is important, regardless of whether the code and data are provided or not.  
+- If the contribution is a dataset and/or model, the authors should describe the steps taken to make their results reproducible or verifiable.  
+- Depending on the contribution, reproducibility can be accomplished in various ways. For example, if the contribution is a novel architecture, describing the architecture fully might suffice, or if the contribution is a specific model and empirical evaluation, it may be necessary to either make it possible for others to replicate the model with the same dataset, or provide access to the model. In general, releasing code and data is often one good way to accomplish this, but reproducibility can also be provided via detailed instructions for how to replicate the results, access to a hosted model (e.g., in the case of a large language model), releasing of a model checkpoint, or other means that are appropriate to the research performed.  
+- While NeurIPS does not require releasing code, the conference does require all submissions to provide some reasonable avenue for reproducibility, which may depend on the nature of the contribution. For example  
+(a) If the contribution is primarily a new algorithm, the paper should make it clear how to reproduce that algorithm.  
+(b) If the contribution is primarily a new model architecture, the paper should describe the architecture clearly and fully.  
+(c) If the contribution is a new model (e.g., a large language model), then there should either be a way to access this model for reproducing the results or a way to reproduce the model (e.g., with an open-source dataset or instructions for how to construct the dataset).  
+(d) We recognize that reproducibility may be tricky in some cases, in which case authors are welcome to describe the particular way they provide for reproducibility. In the case of closed-source models, it may be that access to the model is limited in some way (e.g., to registered users), but it should be possible for other researchers to have some path to reproducing or verifying the results.
+
+# 5. Open access to data and code
+
+Question: Does the paper provide open access to the data and code, with sufficient instructions to faithfully reproduce the main experimental results, as described in supplemental material?
+
+# Answer: [Yes]
+
+Justification: Datasets used in this paper are available publicly. The code will be open-sourced upon publication.
+
+# Guidelines:
+
+- The answer NA means that paper does not include experiments requiring code.  
+- Please see the NeurIPS code and data submission guidelines (https://nips.cc/public/guides/CodeSubmissionPolicy) for more details.  
+- While we encourage the release of code and data, we understand that this might not be possible, so "No" is an acceptable answer. Papers cannot be rejected simply for not including code, unless this is central to the contribution (e.g., for a new open-source benchmark).  
+- The instructions should contain the exact command and environment needed to run to reproduce the results. See the NeurIPS code and data submission guidelines (https://nips.cc/public/guides/CodeSubmissionPolicy) for more details.  
+- The authors should provide instructions on data access and preparation, including how to access the raw data, preprocessed data, intermediate data, and generated data, etc.  
+- The authors should provide scripts to reproduce all experimental results for the new proposed method and baselines. If only a subset of experiments are reproducible, they should state which ones are omitted from the script and why.  
+- At submission time, to preserve anonymity, the authors should release anonymized versions (if applicable).  
+- Providing as much information as possible in supplemental material (appended to the paper) is recommended, but including URLs to data and code is permitted.
+
+# 6. Experimental Setting/Details
+
+Question: Does the paper specify all the training and test details (e.g., data splits, hyperparameters, how they were chosen, type of optimizer, etc.) necessary to understand the results?
+
+# Answer: [Yes]
+
+Justification: We give the training and test details in the Experiment section.
+
+# Guidelines:
+
+- The answer NA means that the paper does not include experiments.  
+- The experimental setting should be presented in the core of the paper to a level of detail that is necessary to appreciate the results and make sense of them.  
+- The full details can be provided either with the code, in appendix, or as supplemental material.
+
+# 7. Experiment Statistical Significance
+
+Question: Does the paper report error bars suitably and correctly defined or other appropriate information about the statistical significance of the experiments?
+
+# Answer: [No]
+
+Justification: We do not report the statistical significance.
+
+# Guidelines:
+
+- The answer NA means that the paper does not include experiments.  
+- The authors should answer "Yes" if the results are accompanied by error bars, confidence intervals, or statistical significance tests, at least for the experiments that support the main claims of the paper.  
+- The factors of variability that the error bars are capturing should be clearly stated (for example, train/test split, initialization, random drawing of some parameter, or overall run with given experimental conditions).  
+- The method for calculating the error bars should be explained (closed form formula, call to a library function, bootstrap, etc.)  
+- The assumptions made should be given (e.g., Normally distributed errors).
+
+- It should be clear whether the error bar is the standard deviation or the standard error of the mean.  
+- It is OK to report 1-sigma error bars, but one should state it. The authors should preferably report a 2-sigma error bar than state that they have a  $96\%$  CI, if the hypothesis of Normality of errors is not verified.  
+- For asymmetric distributions, the authors should be careful not to show in tables or figures symmetric error bars that would yield results that are out of range (e.g. negative error rates).  
+- If error bars are reported in tables or plots, The authors should explain in the text how they were calculated and reference the corresponding figures or tables in the text.
+
+# 8. Experiments Compute Resources
+
+Question: For each experiment, does the paper provide sufficient information on the computer resources (type of compute workers, memory, time of execution) needed to reproduce the experiments?
+
+Answer: [No]
+
+Justification: Our experiments are relatively free from computational resources since it updates adapters of the model using small batch-size.
+
+Guidelines:
+
+- The answer NA means that the paper does not include experiments.  
+- The paper should indicate the type of compute workers CPU or GPU, internal cluster, or cloud provider, including relevant memory and storage.  
+- The paper should provide the amount of compute required for each of the individual experimental runs as well as estimate the total compute.  
+- The paper should disclose whether the full research project required more compute than the experiments reported in the paper (e.g., preliminary or failed experiments that didn't make it into the paper).
+
+# 9. Code Of Ethics
+
+Question: Does the research conducted in the paper conform, in every respect, with the NeurIPS Code of Ethics https://neurips.cc/public/EthicsGuidelines?
+
+Answer: [Yes]
+
+Justification: We have reviewed the Code of Ethics.
+
+Guidelines:
+
+- The answer NA means that the authors have not reviewed the NeurIPS Code of Ethics.  
+- If the authors answer No, they should explain the special circumstances that require a deviation from the Code of Ethics.  
+- The authors should make sure to preserve anonymity (e.g., if there is a special consideration due to laws or regulations in their jurisdiction).
+
+# 10. Broader Impacts
+
+Question: Does the paper discuss both potential positive societal impacts and negative societal impacts of the work performed?
+
+Answer: [NA]
+
+Justification:
+
+Guidelines:
+
+- The answer NA means that there is no societal impact of the work performed.  
+- If the authors answer NA or No, they should explain why their work has no societal impact or why the paper does not address societal impact.  
+- Examples of negative societal impacts include potential malicious or unintended uses (e.g., disinformation, generating fake profiles, surveillance), fairness considerations (e.g., deployment of technologies that could make decisions that unfairly impact specific groups), privacy considerations, and security considerations.
+
+- The conference expects that many papers will be foundational research and not tied to particular applications, let alone deployments. However, if there is a direct path to any negative applications, the authors should point it out. For example, it is legitimate to point out that an improvement in the quality of generative models could be used to generate deepfakes for disinformation. On the other hand, it is not needed to point out that a generic algorithm for optimizing neural networks could enable people to train models that generate Deepfakes faster.  
+- The authors should consider possible harms that could arise when the technology is being used as intended and functioning correctly, harms that could arise when the technology is being used as intended but gives incorrect results, and harms following from (intentional or unintentional) misuse of the technology.  
+- If there are negative societal impacts, the authors could also discuss possible mitigation strategies (e.g., gated release of models, providing defenses in addition to attacks, mechanisms for monitoring misuse, mechanisms to monitor how a system learns from feedback over time, improving the efficiency and accessibility of ML).
+
+# 11. Safeguards
+
+Question: Does the paper describe safeguards that have been put in place for responsible release of data or models that have a high risk for misuse (e.g., pretrained language models, image generators, or scraped datasets)?
+
+Answer: [NA]
+
+Justification:
+
+Guidelines:
+
+- The answer NA means that the paper poses no such risks.  
+- Released models that have a high risk for misuse or dual-use should be released with necessary safeguards to allow for controlled use of the model, for example by requiring that users adhere to usage guidelines or restrictions to access the model or implementing safety filters.  
+- Datasets that have been scraped from the Internet could pose safety risks. The authors should describe how they avoided releasing unsafe images.  
+- We recognize that providing effective safeguards is challenging, and many papers do not require this, but we encourage authors to take this into account and make a best faith effort.
+
+# 12. Licenses for existing assets
+
+Question: Are the creators or original owners of assets (e.g., code, data, models), used in the paper, properly credited and are the license and terms of use explicitly mentioned and properly respected?
+
+Answer: [Yes]
+
+Justification: We have used publicly opened code and data.
+
+Guidelines:
+
+- The answer NA means that the paper does not use existing assets.  
+- The authors should cite the original paper that produced the code package or dataset.  
+- The authors should state which version of the asset is used and, if possible, include a URL.  
+- The name of the license (e.g., CC-BY 4.0) should be included for each asset.  
+- For scraped data from a particular source (e.g., website), the copyright and terms of service of that source should be provided.  
+- If assets are released, the license, copyright information, and terms of use in the package should be provided. For popular datasets, paperswithcode.com/datasets has curated licenses for some datasets. Their licensing guide can help determine the license of a dataset.  
+- For existing datasets that are re-packaged, both the original license and the license of the derived asset (if it has changed) should be provided.
+
+- If this information is not available online, the authors are encouraged to reach out to the asset's creators.
+
+# 13. New Assets
+
+Question: Are new assets introduced in the paper well documented and is the documentation provided alongside the assets?
+
+Answer: [NA]
+
+Justification:
+
+Guidelines:
+
+- The answer NA means that the paper does not release new assets.  
+- Researchers should communicate the details of the dataset/code/model as part of their submissions via structured templates. This includes details about training, license, limitations, etc.  
+- The paper should discuss whether and how consent was obtained from people whose asset is used.  
+- At submission time, remember to anonymize your assets (if applicable). You can either create an anonymized URL or include an anonymized zip file.
+
+# 14. Crowdsourcing and Research with Human Subjects
+
+Question: For crowdsourcing experiments and research with human subjects, does the paper include the full text of instructions given to participants and screenshots, if applicable, as well as details about compensation (if any)?
+
+Answer: [NA]
+
+Justification:
+
+Guidelines:
+
+- The answer NA means that the paper does not involve crowdsourcing nor research with human subjects.  
+- Including this information in the supplemental material is fine, but if the main contribution of the paper involves human subjects, then as much detail as possible should be included in the main paper.  
+- According to the NeurIPS Code of Ethics, workers involved in data collection, curation, or other labor should be paid at least the minimum wage in the country of the data collector.
+
+# 15. Institutional Review Board (IRB) Approvals or Equivalent for Research with Human Subjects
+
+Question: Does the paper describe potential risks incurred by study participants, whether such risks were disclosed to the subjects, and whether Institutional Review Board (IRB) approvals (or an equivalent approval/review based on the requirements of your country or institution) were obtained?
+
+Answer: [NA]
+
+Justification:
+
+Guidelines:
+
+- The answer NA means that the paper does not involve crowdsourcing nor research with human subjects.  
+- Depending on the country in which research is conducted, IRB approval (or equivalent) may be required for any human subjects research. If you obtained IRB approval, you should clearly state this in the paper.  
+- We recognize that the procedures for this may vary significantly between institutions and locations, and we expect authors to adhere to the NeurIPS Code of Ethics and the guidelines for their institution.  
+- For initial submissions, do not include any information that would break anonymity (if applicable), such as the institution conducting the review.

2024/Curriculum Fine-tuning of Vision Foundation Model for Medical Image Classification Under Label Noise/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a1ea76e043c58e43d8f7bc69b95f35e6c212f6f671c8ca4d57514a99e3104be6
+size 459084