ICLR 2023 Papers — Page 16
International Conference on Learning Representations · 1573 papers
Uniform-in-time propagation of chaos for the mean-field gradient Langevin dynamics
Taiji Suzuki (University of Tokyo), Denny Wu (University of Toronto)
TabularStochastic Differential Equation
🎯 What it does: This paper studies the mean field Langevin dynamics of two-layer neural networks, proving the existence of uniformly weakly propagating chaos in time and providing a quantification upper bound with a discretization error of O(1/N).
UniKGQA: Unified Retrieval and Reasoning for Solving Multi-hop Question Answering Over Knowledge Graph
Jinhao Jiang (Renmin University of China), Ji-Rong Wen (Renmin University of China)
RetrievalGraph Neural NetworkLarge Language ModelContrastive LearningTextGraphBenchmark
🎯 What it does: This paper proposes a unified model called UniKGQA, which combines retrieval and reasoning in two stages of knowledge graph question answering (KGQA). It achieves the sharing and transfer of retrieval and reasoning through a unified semantic matching and information propagation module.
UniMax: Fairer and More Effective Language Sampling for Large-Scale Multilingual Pretraining
Hyung Won Chung (Google Research), Noah Constant (Google Research)
TransformerLarge Language ModelTextBenchmark
🎯 What it does: This paper proposes a new multilingual pre-training sampling strategy called UNIMAX, aimed at balancing the data distribution of different languages, reducing the excessive repetition of low-resource languages, and enhancing the model's generalization ability.
Universal Few-shot Learning of Dense Prediction Tasks with Visual Token Matching
Donggyun Kim (Korea Advanced Institute of Science and Technology), Seunghoon Hong (Korea Advanced Institute of Science and Technology)
SegmentationMeta LearningTransformerImage
🎯 What it does: A general few-shot dense prediction framework VTM is proposed, which can learn any dense prediction task with only a small amount of labeled data.
Universal Vision-Language Dense Retrieval: Learning A Unified Representation Space for Multi-Modal Retrieval
Zhenghao Liu (Northeastern University), Ge Yu (Tsinghua University)
RetrievalVision Language ModelContrastive LearningImageTextMultimodality
🎯 What it does: A unified multimodal retrieval framework, UniVL-DR, has been constructed to simultaneously handle text, images, and queries in the same embedding space, enabling one-time retrieval and fusion of multimodal results.
Unmasking the Lottery Ticket Hypothesis: What's Encoded in a Winning Ticket's Mask?
Mansheej Paul (Stanford), Gintare Karolina Dziugaite (Google Research)
Convolutional Neural NetworkSupervised Fine-TuningImage
🎯 What it does: This paper studies and explains the mechanism of Iterative Magnitude Pruning (IMP) in the process of finding sparse trainable sub-networks (winning tickets), revealing the roles of masks, SGD robustness, Hessian spectrum, and retraining through the geometry of the loss landscape.
Unsupervised 3D Object Learning through Neuron Activity aware Plasticity
Beomseok Kang (Georgia Institute of Technology), Saibal Mukhopadhyay (Georgia Institute of Technology)
ClassificationRecognitionPoint Cloud
🎯 What it does: Proposes an unsupervised 3D object classification framework: utilizes a multi-layer MLP encoder combined with WTA and cross-suppression for feature extraction, and employs the Neuron Activity Aware Hebbian (NeAW) learning rule for unsupervised training on the encoder, followed by a supervised fully connected classifier for final recognition.
Unsupervised Learning for Combinatorial Optimization Needs Meta Learning
Haoyu Peter Wang (Georgia Institute of Technology), Pan Li (Purdue University)
OptimizationMeta LearningGraph Neural NetworkGraph
🎯 What it does: This paper proposes Meta-EGN, an unsupervised learning framework based on MAML, designed to learn model initializations that can be quickly fine-tuned for combinatorial optimization problems, thereby achieving high-quality solutions for individual instances.
Unsupervised Manifold Alignment with Joint Multidimensional Scaling
Dexiong Chen (ETH Zurich), Karsten Borgwardt (ETH Zurich)
Domain AdaptationOptimizationRepresentation LearningProtein Structure PredictionMultimodalityBiomedical Data
🎯 What it does: For datasets without correspondence between two domains, a Joint MDS method is proposed to achieve unsupervised manifold alignment and low-dimensional embedding.
Unsupervised Meta-learning via Few-shot Pseudo-supervised Contrastive Learning
Huiwon Jang (Korea Advanced Institute of Science and Technology), Jinwoo Shin (Korea Advanced Institute of Science and Technology)
ClassificationMeta LearningContrastive LearningImage
🎯 What it does: An unsupervised meta-learning framework PsCo is proposed, which utilizes momentum networks and momentum queues to online construct few-shot tasks, and trains the model through pseudo-supervised contrastive learning to achieve few-shot classification.
Unsupervised Model Selection for Time Series Anomaly Detection
Mononito Goswami (Carnegie Mellon University), Andrey Kan (Amazon Research)
Anomaly DetectionRecurrent Neural NetworkTime Series
🎯 What it does: This paper studies the unsupervised model selection problem for anomaly detection on unlabeled time series data, proposing a model selection framework based on proxy metrics and performing robust ranking aggregation.
Unsupervised Semantic Segmentation with Self-supervised Object-centric Representations
Andrii Zadaianchuk (Max-Planck Institute for Intelligent Systems), Thomas Brox (University of Freiburg)
SegmentationTransformerContrastive LearningImage
🎯 What it does: An unsupervised semantic segmentation framework COMUS is proposed, which utilizes self-supervised features and unsupervised saliency detection for object discovery, and enhances multi-object segmentation through iterative self-training.
Unsupervised visualization of image datasets using contrastive learning
Niklas Böhm, Dmitry Kobak (University of Tuebingen)
ClassificationRepresentation LearningConvolutional Neural NetworkContrastive LearningImage
🎯 What it does: A two-dimensional visualization method based on contrastive learning, t-SimCNE, is proposed for unsupervised mapping of image data to a two-dimensional space.
Unveiling the sampling density in non-uniform geometric graphs
Raffaele Paolino (Ludwig Maximilians University Munich), Ron Levie (Technion Israel Institute of Technology)
ClassificationExplainability and InterpretabilityRepresentation LearningGraph Neural NetworkGraph
🎯 What it does: A non-uniform geometric graph (NuG) model is proposed, along with a self-supervised method for estimating sampling density from the graph structure, utilizing this density to correct the graph shift operator (GSO) to enhance graph learning performance.
User-Interactive Offline Reinforcement Learning
Phillip Swazinna (Siemens and Technical University of Munich), Thomas Runkler (Siemens and Technical University of Munich)
Reinforcement LearningTabularBenchmark
🎯 What it does: Designed and implemented LION, an offline reinforcement learning algorithm that allows real-time control of the proximity between the policy and the original policy by adjusting the hyperparameter λ after deployment, enabling user interactive adjustment.
Using Both Demonstrations and Language Instructions to Efficiently Learn Robotic Tasks
Albert Yu (University of Texas at Austin), Ray Mooney
Robotic IntelligenceConvolutional Neural NetworkContrastive LearningVideoTextMultimodality
🎯 What it does: By combining demonstration videos and natural language instructions, a single multi-task robot controller is trained, enabling the robot to complete new tasks with only one demonstration and a single sentence instruction.
Using Language to Extend to Unseen Domains
Lisa Dunlap (University of California Berkeley), Anna Rohrbach (University of California Berkeley)
Domain AdaptationTransformerVision Language ModelContrastive LearningImageMultimodality
🎯 What it does: This paper proposes a language description-based domain expansion method called LADS, which utilizes the multimodal embedding space of CLIP to map training domain image features to unseen domains through language guidance, thereby improving cross-domain performance without using target domain images.
VA-DepthNet: A Variational Approach to Single Image Depth Prediction
Ce Liu (ETH Zurich), Luc Van Gool (KU Leuven)
Depth EstimationTransformerImage
🎯 What it does: Proposes VA-DepthNet, a monocular depth prediction network that utilizes first-order variational constraints.
Valid P-Value for Deep Learning-driven Salient Region
Miwa Daiki, Ichiro Takeuchi (Nagoya University and RIKEN)
Object DetectionSegmentationConvolutional Neural NetworkAuto EncoderImageBiomedical DataComputed Tomography
🎯 What it does: Using Conditional Selective Inference to provide testable p-values for significant regions generated by deep learning, thereby quantifying the reliability of significant regions.
Value Memory Graph: A Graph-Structured World Model for Offline Reinforcement Learning
Deyao Zhu (King Abdullah University of Science and Technology), Mohamed Elhoseiny
Graph Neural NetworkReinforcement LearningContrastive LearningWorld ModelGraphBenchmark
🎯 What it does: This paper proposes the Value Memory Graph (VMG), a world model that aggregates offline RL data into a directed graph after mapping it to a metric space. It employs value iteration and Dijkstra multi-step search control for agents on this graph, achieving efficient learning for sparse rewards and long-horizon tasks.
Variance Reduction is an Antidote to Byzantines: Better Rates, Weaker Assumptions and Communication Compression as a Cherry on the Top
Eduard Gorbunov (Mohamed bin Zayed University of Artificial Intelligence), Gauthier Gidel (Mila)
OptimizationFederated LearningTabular
🎯 What it does: A new distributed optimization algorithm, Byz-VR-MARINA, is designed and implemented. This algorithm achieves robust training under Byzantine attack scenarios through variance reduction and communication compression, and provides theoretical convergence guarantees for non-convex and Polyak-Łojasiewicz (PŁ) objective functions.
Variance-Aware Sparse Linear Bandits
Yan Dai (Tsinghua University), Simon Shaolei Du
OptimizationReinforcement Learning
🎯 What it does: This paper studies a variance-adaptive algorithm for sparse linear bandits and proposes a general framework VASLB, which can transform any variance-aware linear bandit algorithm into a sparse version with theoretical guarantees.
Variational Information Pursuit for Interpretable Predictions
Aditya Chattopadhyay (Johns Hopkins University), Rene Vidal (University of Pennsylvania)
ClassificationExplainability and InterpretabilityComputational EfficiencyConvolutional Neural NetworkReinforcement LearningImageTextBiomedical Data
🎯 What it does: A variational information tracking (V-IP) method is proposed, which directly learns the query selection strategy and classifier using deep networks, achieving interpretable predictions without relying on generative models.
Variational Latent Branching Model for Off-Policy Evaluation
Qitong Gao (North Carolina State University), Miroslav Pajic (Duke University)
Recurrent Neural NetworkReinforcement LearningAuto EncoderSequentialBenchmark
🎯 What it does: A Variational Latent Branch Model (VLBM) is designed and implemented to learn MDP transitions from limited coverage offline trajectories and perform offline policy evaluation.
Verifying the Union of Manifolds Hypothesis for Image Data
Bradley CA Brown, Gabriel Loaiza-Ganem (Layer 6 AI)
ClassificationGenerationAuto EncoderImage
🎯 What it does: This paper experimentally verifies that image data better fits the 'multi-manifold union' hypothesis, proving that the support set of the data is discrete and that different subsets have different intrinsic dimensions.
Versatile Neural Processes for Learning Implicit Neural Representations
Zongyu Guo (University of Science and Technology of China), Zhibo Chen (University of Science and Technology of China)
GenerationRepresentation LearningAuto EncoderImagePoint Cloud
🎯 What it does: The Versatile Neural Processes (VNP) framework is proposed for efficiently learning implicit neural representations (INR), allowing for the inference of an entire continuous function in a single forward pass with only partial observations.
Video Scene Graph Generation from Single-Frame Weak Supervision
Siqi Chen (Zhejiang University), Long Chen (Hong Kong University of Science and Technology)
GenerationKnowledge DistillationTransformerVideo
🎯 What it does: This paper proposes a single-frame weakly supervised video scene graph generation task and automatically generates localized scene graphs for each frame in the video through a Pseudo Label Assignment framework (PLA), enabling the training of fully supervised models without the need for complete video-level annotations.
View Synthesis with Sculpted Neural Points
Yiming Zuo (Princeton University), Jia Deng (Princeton University)
GenerationData SynthesisComputational EfficiencyConvolutional Neural NetworkNeural Radiance FieldPoint Cloud
🎯 What it does: This paper proposes a point cloud-based view synthesis method called Sculpted Neural Points (SNP), which utilizes initial point clouds generated by multi-view stereo (MVS) and achieves high-quality, fast view synthesis through global sculpting (point pruning and point augmentation) and differentiable rendering.
ViewCo: Discovering Text-Supervised Segmentation Masks via Multi-View Semantic Consistency
Pengzhen Ren (Sun Yat-sen University), Xiaodan Liang (MBZUAI)
SegmentationContrastive LearningImageText
🎯 What it does: A text-supervised semantic segmentation is performed by constructing a multi-view consistency learning framework (ViewCo) to enhance the consistency of segmentation across different view images and cross-modal semantic alignment.
VIP: Towards Universal Visual Reward and Representation via Value-Implicit Pre-Training
Yecheng Jason Ma (Meta AI), Amy Zhang (Meta AI)
Representation LearningRobotic IntelligenceConvolutional Neural NetworkReinforcement LearningContrastive LearningVideo
🎯 What it does: This paper proposes a visual reward and representation learning framework based on Value Implicit Pre-training (VIP), which can self-supervisedly learn a network from massive offline human videos that can generate dense rewards and serve as general visual features.
VIPeR: Provably Efficient Algorithm for Offline RL with Neural Function Approximation
Thanh Nguyen-Tang (Johns Hopkins University), Raman Arora (Johns Hopkins University)
Reinforcement LearningTabular
🎯 What it does: An offline reinforcement learning algorithm named VIPER is proposed, which achieves conservatism (pessimism) by adding random noise to the rewards and using the minimum of multiple models, thereby achieving theoretically provable effectiveness without explicitly constructing confidence intervals.
Vision Transformer Adapter for Dense Predictions
Zhe Chen (Nanjing University), Yu Qiao (Shanghai AI Laboratory)
Object DetectionSegmentationTransformerImageMultimodality
🎯 What it does: A pre-trained, task-agnostic adapter named ViT-Adapter is proposed, which transforms a standard Vision Transformer into a model suitable for dense prediction tasks (object detection, instance segmentation, semantic segmentation) through convolutional priors and multi-scale feature reconstruction.
Visual Classification via Description from Large Language Models
Sachit Menon (Columbia University), Carl Vondrick (Columbia University)
ClassificationExplainability and InterpretabilityTransformerLarge Language ModelVision Language ModelImageText
🎯 What it does: Automatically generate descriptive words for visual categories using large language models, and utilize Vision-Language models to match images with these descriptions for interpretable zero-shot classification;
Visual Imitation Learning with Patch Rewards
Minghuan Liu (Shanghai Jiaotong University), Zhongwen Xu (Sea AI Lab)
Explainability and InterpretabilityRobotic IntelligenceConvolutional Neural NetworkReinforcement LearningImage
🎯 What it does: Proposes the PatchAIL framework, which uses a patch-based discriminator to generate dense rewards for visual imitation learning, balancing efficiency and interpretability.
Visual Recognition with Deep Nearest Centroids
Wenguan Wang (Zhejiang University), Dongfang Liu (Zhejiang University)
ClassificationRecognitionSegmentationConvolutional Neural NetworkImage
🎯 What it does: A Deep Nearest Center (DNC) network is proposed, using a non-parametric nearest center decision instead of the traditional softmax classifier;
Visually-Augmented Language Modeling
Weizhi Wang (University of California), Furu Wei (Microsoft Research)
RetrievalTransformerLarge Language ModelVision Language ModelImageTextMultimodalityRetrieval-Augmented Generation
🎯 What it does: Introducing retrieval-based visual completion during the pre-training phase of the language model, dynamically retrieving and integrating corresponding image information for each text token, thereby achieving visually enhanced autoregressive language modeling.
VoGE: A Differentiable Volume Renderer using Gaussian Ellipsoids for Analysis-by-Synthesis
Angtian Wang (Johns Hopkins University), Alan Yuille (Johns Hopkins University)
Pose EstimationImage
🎯 What it does: Using a differentiable voxel renderer VoGE, the geometry of objects is represented by 3D Gaussian ellipsoids to achieve volume rendering;
Voint Cloud: Multi-View Point Cloud Representation for 3D Understanding
Abdullah Hamdi (King Abdullah University of Science and Technology), Bernard Ghanem (King Abdullah University of Science and Technology)
ClassificationSegmentationRetrievalGraph Neural NetworkPoint Cloud
🎯 What it does: This paper proposes a multi-view point cloud representation, where each 3D point is described by a set of features from multiple views, and designs VointNet to learn feature aggregation in this space.
Volumetric Optimal Transportation by Fast Fourier Transform
Na Lei (Dalian University of Technology), David Gu (Stony Brook University)
OptimizationComputational EfficiencyBiomedical Data
🎯 What it does: The three-dimensional optimal transport problem is solved using FFT, achieving efficient computation on GPU through the linearization of the Monge-Ampère equation and mean coefficient approximation.
Voxurf: Voxel-based Efficient and Accurate Neural Surface Reconstruction
Tong Wu (Shanghai AI Laboratory), Dahua Lin (Max Planck Institute for Informatics)
RestorationGenerationComputational EfficiencyNeural Radiance FieldPoint CloudMeshBenchmark
🎯 What it does: A voxel-based neural surface reconstruction method called Voxurf is proposed, capable of completing high-quality surface reconstruction in just a few minutes.
Warping the Space: Weight Space Rotation for Class-Incremental Few-Shot Learning
Do-Yeon Kim (Korea Advanced Institute of Science and Technology), Jaekyun Moon (Korea Advanced Institute of Science and Technology)
ClassificationRepresentation LearningConvolutional Neural NetworkImage
🎯 What it does: A weight space rotation (WaRP) reparameterization method is proposed for class-incremental few-shot learning, which can effectively learn new categories while preserving old knowledge.
Wasserstein Auto-encoded MDPs: Formal Verification of Efficiently Distilled RL Policies with Many-sided Guarantees
Florent Delgrange (Vrije Universiteit Brussel), Guillermo Perez
Knowledge DistillationReinforcement LearningAuto EncoderSequential
🎯 What it does: This paper proposes a framework based on Wasserstein Autoencoders (WAE-MDP) for distilling deep reinforcement learning (DRL) policies into discrete verifiable subspace models, and ensures the formal verification of the original policy by approximating parallel execution that guarantees the formalized Bisimulation property.
wav2tok: Deep Sequence Tokenizer for Audio Retrieval
Adhiraj Banerjee (Indian Institute of Technology Kanpur), Vipul Arora (Indian Institute of Technology Kanpur)
RetrievalRecurrent Neural NetworkContrastive LearningAudio
🎯 What it does: A deep learning-based audio sequence tokenizer called wav2tok is proposed, which can learn semantic discrete tokens from similar audio pairs for efficient retrieval.
Weakly Supervised Explainable Phrasal Reasoning with Neural Fuzzy Logic
Zijun Wu (University of Alberta), Lili Mou (University of Alberta)
Explainability and InterpretabilityTransformerLarge Language ModelText
🎯 What it does: We propose a weakly supervised interpretable phrase reasoning model EPR to accomplish the natural language inference task; the model first detects phrases, aligns corresponding phrases, predicts phrase-level logical relationships, and then derives sentence-level labels through neural fuzzy logic.
Weakly Supervised Knowledge Transfer with Probabilistic Logical Reasoning for Object Detection
Martijn Oldenhof (KU Leuven), Edward De Brouwer (KU Leuven)
Object DetectionTransformerImage
🎯 What it does: Proposes the ProbKT framework, which utilizes probabilistic logic reasoning to achieve weakly supervised knowledge transfer for object detection.
Weakly-supervised HOI Detection via Prior-guided Bi-level Representation Learning
Bo Wan (KU Leuven), Xuming He (ShanghaiTech University)
Object DetectionRepresentation LearningConvolutional Neural NetworkContrastive LearningImage
🎯 What it does: A weakly supervised human-object interaction (HOI) detection framework is proposed, utilizing the CLIP pre-trained model to construct a dual-layer knowledge base, injecting prior knowledge at both the image level and alignment level, and suppressing erroneous human-object associations through a self-learning correlation classifier, ultimately achieving end-to-end weakly supervised HOI detection.
Weighted Clock Logic Point Process
Ruixuan Yan (Rensselaer Polytechnic Institute), Anak Agung Julius
GenerationOptimizationExplainability and InterpretabilityTabularTime SeriesElectronic Health Records
🎯 What it does: A neural-symbolic framework called Clock Logic Neural Network (CLNN) is proposed, which explains the generation mechanism of multivariate event streams and predicts event occurrences by learning weighted clock logic formulas (wCL).
Weighted Ensemble Self-Supervised Learning
Yangjun Ruan (University of Toronto), Joshua V. Dillon (Google Research)
Representation LearningContrastive LearningImage
🎯 What it does: An efficient method is proposed for multi-head ensemble of the projection head and codebook of self-supervised learning models, applied only during the training phase, and a data-dependent weighted cross-entropy loss is designed to enhance representation learning performance.
What Can we Learn From The Selective Prediction And Uncertainty Estimation Performance Of 523 Imagenet Classifiers?
Ido Galil (Technion), Ran El-Yaniv (Deci.AI)
ClassificationKnowledge DistillationTransformerImage
🎯 What it does: A systematic evaluation of 523 ImageNet pre-trained models was conducted regarding their performance in uncertainty estimation (selective prediction, calibration, etc.), and influencing factors were explored.
What Do Self-Supervised Vision Transformers Learn?
Namuk Park (Genentech), Sangdoo Yun (Naver Corporation)
Object DetectionRepresentation LearningTransformerContrastive LearningImage
🎯 What it does: This paper compares two mainstream methods in self-supervised learning—Contrastive Learning (CL) and Masked Image Modeling (MIM)—in terms of their behavior and effects during Vision Transformer (ViT) pre-training, analyzing their self-attention patterns, representation transformations, and important layer positions. It demonstrates that the two methods are complementary and proposes a simple linear combination objective that can simultaneously leverage the advantages of both.
What Is Missing in IRM Training and Evaluation? Challenges and Solutions
Yihua Zhang (Michigan State University), Sijia Liu (IBM Research)
Domain AdaptationOptimizationImage
🎯 What it does: This paper re-examines the training and evaluation of Invariant Risk Minimization (IRM), revealing performance misjudgments caused by large-batch training and single-test environment evaluation, and proposes a new method called BLOC-IRM, which involves small-batch training, multi-environment evaluation, and consensus-constrained bi-level optimization.
What learning algorithm is in-context learning? Investigations with linear models
Ekin Akyürek (Massachusetts Institute of Technology), Denny Zhou (Google Research)
OptimizationRepresentation LearningTransformerTabular
🎯 What it does: This study investigates the context learning mechanism of Transformer in linear regression tasks, proving that it can implicitly implement gradient descent and closed-form ridge regression, and can exhibit algorithm phase transitions under different capacities and noise levels.
What Makes Convolutional Models Great on Long Sequence Modeling?
Yuhong Li (University of Illinois Urbana-Champaign), Debadeepta Dey (Microsoft Research)
Convolutional Neural NetworkTransformerImageTextAudio
🎯 What it does: A long sequence modeling framework SGConv based on global convolution is proposed, and it is integrated as a general module into Transformer, ConvNet, and visual models.
What shapes the loss landscape of self supervised learning?
Liu Ziyin (University of Tokyo), Hidenori Tanaka (NTT Research)
OptimizationRepresentation LearningTransformerContrastive LearningImage
🎯 What it does: Analyzes the loss landscape of self-supervised learning (SSL) from a theoretical perspective, revealing the fundamental causes of dimensional collapse and providing control methods.
When and Why Vision-Language Models Behave like Bags-Of-Words, and What to Do About It?
Mert Yuksekgonul (Stanford University), James Zou (Stanford University)
RetrievalTransformerVision Language ModelContrastive LearningImageTextMultimodalityBenchmark
🎯 What it does: This paper proposes the ARO (Attribution, Relation, Order) benchmark for systematically evaluating the performance of visual language models in understanding attributes, relations, and order. It finds that existing models often perform like a bag-of-words on these tasks; the use of a negative sample strategy in contrastive learning (composition-aware hard negatives) significantly enhances the model's compositionality and order sensitivity.
When Data Geometry Meets Deep Function: Generalizing Offline Reinforcement Learning
Jianxiong Li (Institute for Artificial Intelligence Industry Research), Ya-Qin Zhang (Institute for Artificial Intelligence Industry Research)
Reinforcement LearningTabularBenchmark
🎯 What it does: The DOGE algorithm is proposed, which combines the geometric information of the dataset with the approximation error characteristics of the deep Q function in offline reinforcement learning. It utilizes a state-conditioned distance function as a policy constraint to achieve reasonable utilization of the OOB region.
When Source-Free Domain Adaptation Meets Learning with Noisy Labels
Li Yi (University of Western Ontario), Boyu Wang (University of Western Ontario)
Domain AdaptationConvolutional Neural NetworkSupervised Fine-TuningImage
🎯 What it does: This paper analyzes the source-free domain adaptation (SFDA) problem from the perspective of learning with noisy labels (LLN), proving that the noise in SFDA is unbounded. By introducing early training regularization (ELR), it effectively suppresses the memorization of noisy labels through the early training phenomenon (ETP), thereby improving classification performance in the target domain.
When to Make and Break Commitments?
Alihan Hüyük (University of Cambridge), Mihaela van der Schaar (University of Cambridge)
OptimizationReinforcement LearningTabularBiomedical Data
🎯 What it does: This paper proposes a new Optimal Commitment Problem (OCP) aimed at helping decision-makers determine when to break commitments to avoid future costs, and potentially shift to other commitments or cease commitments altogether.
Where to Begin? On the Impact of Pre-Training and Initialization in Federated Learning
John Nguyen (Meta AI), Michael Rabbat (Meta AI)
OptimizationFederated LearningTransformerSupervised Fine-TuningImageText
🎯 What it does: This paper systematically studies the role of pre-trained models in federated learning, comparing the performance of random initialization and pre-trained initialization across different federated optimization algorithms, and quantifying how pre-training reduces the impact of data and system heterogeneity.
Where to Diffuse, How to Diffuse, and How to Get Back: Automated Learning for Multivariate Diffusions
Raghav Singhal (New York University), Rajesh Ranganath (New York University)
GenerationData SynthesisDiffusion modelImageStochastic Differential Equation
🎯 What it does: A multivariate diffusion model (MDM) is proposed along with an automated training algorithm (AMDT) that can learn the inference process without manually deriving the ELBO or transition kernels, thereby improving image generation quality.
Which Layer is Learning Faster? A Systematic Exploration of Layer-wise Convergence Rate for Deep Neural Networks
Yixiong Chen (Chinese University of Hong Kong), Zongwei Zhou (Johns Hopkins University)
Convolutional Neural NetworkImage
🎯 What it does: Explores and verifies the phenomenon that shallow layers converge faster than deep layers in deep neural networks (layer convergence bias)
Why (and When) does Local SGD Generalize Better than SGD?
Xinran Gu (Tsinghua University), Sanjeev Arora (Princeton University)
OptimizationImageStochastic Differential Equation
🎯 What it does: This paper theoretically analyzes the long-term generalization behavior of Local SGD under a small learning rate regime by deriving the Slow SDE model.
Why adversarial training can hurt robust accuracy
Jacob Clarysse (ETH Zurich), Fanny Yang (ETH Zurich)
ClassificationAdversarial AttackImage
🎯 What it does: This paper studies why adversarial training reduces robust accuracy under low sample sizes in the context of perceptible and direct attacks (i.e., attacks focused on weakening classification information), providing theoretical proof and experimental validation.
WikiWhy: Answering and Explaining Cause-and-Effect Questions
Matthew Ho (University of California), William Yang Wang (University of California)
GenerationRetrievalExplainability and InterpretabilityTransformerLarge Language ModelPrompt EngineeringTextRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: A WIKIWHY dataset was constructed, collecting 9,406 Wikipedia-based 'why' question-answer and explanation triples;
Win: Weight-Decay-Integrated Nesterov Acceleration for Adaptive Gradient Algorithms
Pan Zhou (Sea AI Lab), Shuicheng YAN
OptimizationTransformerImageText
🎯 What it does: A general acceleration framework named Weight‑Decay‑Integrated Nesterov Acceleration (WIN) is proposed to enhance the convergence speed and final performance of adaptive gradient optimizers such as AdamW, Adam, LAMB, and SGD.
WiNeRT: Towards Neural Ray Tracing for Wireless Channel Modelling and Differentiable Simulations
Tribhuvanesh Orekondy (Qualcomm AI Research), Arash Behboodi (Qualcomm AI Research)
Neural Radiance FieldMesh
🎯 What it does: A differentiable neural ray tracing network was trained to replace traditional wireless ray tracers, providing wireless channel characteristics in indoor environments.
Winning Both the Accuracy of Floating Point Activation and the Simplicity of Integer Arithmetic
Yulhwa Kim (Seoul National University), jae-joon kim
Computational EfficiencyConvolutional Neural NetworkImageText
🎯 What it does: A method for matrix multiplication that replaces FP operations with integer arithmetic while maintaining floating-point activation accuracy, utilizing weight quantization and pre-alignment to achieve integer addition calculations.
Words are all you need? Language as an approximation for human similarity judgments
Raja Marjieh (Princeton University), Nori Jacoby (Max Planck Institute for Empirical Aesthetics)
RetrievalRecommendation SystemTransformerLarge Language ModelImageVideoMultimodalityAudio
🎯 What it does: The performance of various methods based on pre-trained deep networks, language models, and term frequency analysis in approximating human similarity judgments was evaluated and compared. An O(N) similarity approximation method based on language and an adaptive label collection process called STEP-Tag were proposed. A large-scale human similarity dataset and practical guidelines were also released.
Write and Paint: Generative Vision-Language Models are Unified Modal Learners
Shizhe Diao (Hong Kong University of Science and Technology), Jiawei Wang (Shanghai Jiao Tong University)
GenerationData SynthesisTransformerVision Language ModelImageTextMultimodality
🎯 What it does: This paper proposes a unified prefix multimodal pre-training framework, constructing a general visual-language foundation model DAVINCI that can simultaneously learn 'writing' (image → text) and 'painting' (text → image).
Your Contrastive Learning Is Secretly Doing Stochastic Neighbor Embedding
Tianyang Hu (Huawei Noah's Ark Lab), Weiran Huang (Shanghai Jiao Tong University)
Domain AdaptationRepresentation LearningContrastive LearningImageMultimodality
🎯 What it does: This paper reinterprets self-supervised contrastive learning (SSCL) from the perspective of SNE (Stochastic Neighbor Embedding) and considers SSCL as a special form of SNE; based on this, it reveals the principles of alignment and uniformity in SSCL, implicit bias, and distance preservation, among other theories; further, it proposes improvement methods inspired by SNE (weighted positive samples, t-SimCLR / t-MoCo) and validates their effectiveness through various datasets.
Zero-Shot Image Restoration Using Denoising Diffusion Null-Space Model
Yinhuai Wang (Peking University), Jian Zhang (Peking University)
RestorationSuper ResolutionDiffusion modelImage
🎯 What it does: A zero-shot image restoration framework DDNM and its improved version DDNM+ are proposed, utilizing a pre-trained denoising diffusion model and null space decomposition, capable of solving any linear inverse problem (such as super-resolution, deblurring, colorization, compressed sensing, inpainting, etc.) without training any networks, and supporting noise recovery.
Zeroth-Order Optimization with Trajectory-Informed Derivative Estimation
Yao Shu (National University of Singapore), Bryan Kian Hsiang Low (National University of Singapore)
OptimizationAdversarial AttackTabular
🎯 What it does: This paper proposes a zero-order optimization method ZORD based on trajectory information, which uses the queried optimization trajectories to estimate gradients and achieves multi-step gradient descent through dynamic virtual updates, significantly reducing the number of queries.
ZiCo: Zero-shot NAS via inverse Coefficient of Variation on Gradients
Guihong Li (University of Texas at Austin), Radu Marculescu (University of Texas at Austin)
OptimizationNeural Architecture SearchImage
🎯 What it does: A new zero-shot NAS proxy, ZiCo, is proposed, which predicts network performance using the inverse coefficient of gradient mean and variance to directly search for the optimal network.