ICLR 2023 Papers — Page 9
International Conference on Learning Representations · 1573 papers
Learning to Decompose Visual Features with Latent Textual Prompts
Feng Wang (Tsinghua University), Heng Ji (University of Illinois at Urbana-Champaign)
ClassificationTransformerPrompt EngineeringContrastive LearningImage
🎯 What it does: Proposes the DeFo (Decomposed Feature Prompting) model, which utilizes learnable text embedding sequences to prompt features in CLIP, achieving stronger classification performance through a linear classifier while maintaining visual-language alignment.
Learning to Estimate Shapley Values with Vision Transformers
Ian Connick Covert (University of Washington), Su-In Lee (University of Washington)
Explainability and InterpretabilityComputational EfficiencyTransformerImage
🎯 What it does: Train a dedicated ViT interpreter model to output approximate Shapley values for each image patch with a single forward pass.
Learning to Estimate Single-View Volumetric Flow Motions without 3D Supervision
Aleksandra Franz (Technical University of Munich), Nils Thuerey (Technical University of Munich)
GenerationData SynthesisDepth EstimationConvolutional Neural NetworkGenerative Adversarial NetworkOptical FlowVideo
🎯 What it does: Train a deep neural network for single-view video that can simultaneously estimate the 3D density field and velocity field of fluids without any 3D supervision data.
Learning to Extrapolate: A Transductive Approach
Aviv Netanyahu (Massachusetts Institute of Technology), Pulkit Agrawal (Massachusetts Institute of Technology)
Robotic IntelligencePoint Cloud
🎯 What it does: This paper proposes a transductive method aimed at addressing the issue of extrapolation at test points outside the training distribution in machine learning systems. By using transductive reparameterization, the extrapolation problem is transformed into a combinatorial generalization problem, thereby enabling effective utilization of hyperparameterized function approximators.
Learning to Generate Columns with Application to Vertex Coloring
Yuan Sun (La Trobe University), Jake Weiner (RMIT University)
OptimizationGraph Neural NetworkGraph
🎯 What it does: A machine learning-based column generation method is proposed to solve the vertex coloring problem of graphs.
Learning to Grow Pretrained Models for Efficient Transformer Training
Peihao Wang (University of Texas at Austin), Yoon Kim (Massachusetts Institute of Technology)
TransformerSupervised Fine-TuningImageText
🎯 What it does: Learning a linear mapping (LiGO) to map small model parameters to a large model, achieving rapid scaling and training of Transformers.
Learning to Induce Causal Structure
Nan Rosemary Ke (DeepMind), Danilo Jimenez Rezende
TransformerSupervised Fine-TuningGraph
🎯 What it does: A causal structure induction method based on supervised learning, CSIvA, is proposed, which directly maps observational and interventional data to causal graph structures.
Learning to Jointly Share and Prune Weights for Grounding Based Vision and Language Models
Shangqian Gao (University of Pittsburgh), Hongxia Jin (Samsung Research America)
RecognitionObject DetectionCompressionTransformerVision Language ModelMultimodality
🎯 What it does: This study investigates a unified framework that simultaneously learns weight sharing and pruning in visual-language models to reduce model parameters while maintaining performance.
Learning to Linearize Deep Neural Networks for Secure and Efficient Private Inference
Souvik Kundu (Intel Labs), Peter Anthony Beerel
OptimizationSafty and PrivacyComputational EfficiencyKnowledge DistillationConvolutional Neural NetworkImage
🎯 What it does: By automatically allocating and positioning ReLU, this paper optimizes private inference networks under a given ReLU budget, significantly reducing nonlinear computations, communication, and latency.
Learning to reason over visual objects
Shanka Subhra Mondal (Princeton University), Jonathan Cohen
Object DetectionSegmentationTransformerImage
🎯 What it does: A general visual reasoning model STSN based on object-centered encoding and Transformer inference is proposed to solve visual reasoning tasks such as Raven's Progressive Matrices.
Learning to Segment from Noisy Annotations: A Spatial Correction Approach
Jiachen Yao (Stony Brook University), Chao Chen (Stony Brook University)
SegmentationImageBiomedical DataMagnetic Resonance Imaging
🎯 What it does: A spatially correlated noise model based on Markov processes is proposed, and a label correction method is designed based on this model.
Learning to Solve Constraint Satisfaction Problems with Recurrent Transformer
Zhun Yang (Arizona State University), Joohyung Lee (Samsung Research)
TransformerImageText
🎯 What it does: The study uses a recursive Transformer to solve constraint satisfaction problems (CSP), particularly Sudoku and visual Sudoku, and implements sample-efficient and semi-supervised learning by directly embedding discrete constraints into the model using the STE method.
Learning topology-preserving data representations
Ilya Trofimov (Skolkovo Institute of Science and Technology), Serguei Barannikov (CNRS Université Paris Cité)
Representation LearningAuto EncoderImageBiomedical Data
🎯 What it does: This paper proposes a dimensionality reduction method that preserves topology using an autoencoder based on Representation Topology Divergence (RTD-AE).
Learning Uncertainty for Unknown Domains with Zero-Target-Assumption
Yu Yu (Stevens Institute of Technology), Jia Xu (Stevens Institute of Technology)
Domain AdaptationTransformerReinforcement LearningText
🎯 What it does: This paper proposes the Maximum Entropy Reward Reinforcement Learning framework (MERRL) to automatically select training samples without target domain information, enhancing the generalization performance of NLP models in unknown domains.
Learning Vortex Dynamics for Fluid Inference and Prediction
Yitong Deng (Dartmouth), Bo Zhu (Dartmouth)
VideoPhysics Related
🎯 What it does: This paper proposes a Differentiable Vortex Particle (DVP) method to infer and predict the implicit physical quantities and future evolution of two-dimensional inviscid turbulence from a single RGB video.
Learning What and Where: Disentangling Location and Identity Tracking Without Supervision
Manuel Traub (Neuro-Cognitive Modeling University of Tübingen), Martin V. Butz (Neuro-Cognitive Modeling University of Tübingen)
Object DetectionObject TrackingTransformerSupervised Fine-TuningVideo
🎯 What it does: A fully self-supervised object localization and identity decoupling network called Loci is proposed, which learns the 'what' and 'where' of objects through slot coding, and achieves object permanence and interaction through predictive coding, gated recursion, and multi-head attention.
Learning where and when to reason in neuro-symbolic inference
Cristina Cornelio (Samsung AI), Timothy Hospedales (Samsung AI)
Computational EfficiencyConvolutional Neural NetworkTransformerReinforcement LearningImage
🎯 What it does: A neural-symbolic integration pipeline (NASR) is proposed, consisting of a neural solver, a mask predictor, and a symbolic solver, capable of enforcing hard constraints during inference.
Learning with Auxiliary Activation for Memory-Efficient Training
Sunghyeon Woo (Seoul National University), Dongsuk Jeon (Seoul National University)
OptimizationComputational EfficiencyConvolutional Neural NetworkTransformerImageText
🎯 What it does: Proposes the Auxiliary Activation Learning algorithm, which uses auxiliary activations to replace the input activations in forward propagation, thereby reducing the activation data that needs to be stored during training and significantly lowering memory usage.
Learning with Logical Constraints but without Shortcut Satisfaction
Zenan Li (Nanjing University), Jian L\"{u}
ClassificationOptimizationImage
🎯 What it does: This paper proposes a new framework that seamlessly integrates logical constraints into deep neural networks, avoiding the shortcut satisfaction problem encountered in traditional methods.
Learning with Stochastic Orders
Carles Domingo-Enrich (New York University), Youssef Mroueh (IBM Research AI)
GenerationOptimizationGenerative Adversarial NetworkImagePoint Cloud
🎯 What it does: This paper proposes the use of convex/Choquet stochastic orders to learn high-dimensional probability distributions, presenting the Variational Dominance Criterion (VDC) and Choquet-Toland distance, and approximating convex functions using Input Convex Maxout Networks (ICMN) to form a trainable pseudo-distance. The framework is then applied to combinatorial optimization (second-order dominance constraints between benchmark and target combinations) and Generative Adversarial Networks (constraining the generated distribution's convex dominance based on a trained model), achieved through min-max optimization.
Learning without Prejudices: Continual Unbiased Learning via Benign and Malignant Forgetting
Myeongho Jeon (Seoul National University), Myungjoo Kang (Seoul National University)
ClassificationData-Centric LearningGenerative Adversarial NetworkContrastive LearningImage
🎯 What it does: A framework for continuous unbiased learning, LwP, is proposed, aiming to suppress malicious forgetting and promote beneficial forgetting while continuously learning tasks, in order to enhance the model's generalization ability under different bias distributions.
Learning Zero-Shot Cooperation with Humans, Assuming Humans Are Biased
Chao Yu (Tsinghua University), Yi Wu (Tsinghua University)
Reinforcement Learning
🎯 What it does: This paper proposes a Hidden-Utility Self-Play (HSP) framework that considers human preferences in zero-shot cooperation, enabling agents to be trained to cooperate smoothly with preference-based humans without human data.
Least-to-Most Prompting Enables Complex Reasoning in Large Language Models
Denny Zhou (Google Research), Ed H. Chi (Google Research)
TransformerLarge Language ModelPrompt EngineeringTextChain-of-Thought
🎯 What it does: A two-stage least-to-most prompting method is proposed, which first breaks down complex problems into a series of subproblems, and then sequentially uses the solved subproblems to complete subsequent subproblems, enabling large language models to reason on more difficult tasks that have not been seen in the prompt examples.
Leveraging Future Relationship Reasoning for Vehicle Trajectory Prediction
Daehee Park (Korea Advanced Institute of Science and Technology), Kuk-Jin Yoon (Korea Advanced Institute of Science and Technology)
Autonomous DrivingGraph Neural NetworkAuto EncoderTime Series
🎯 What it does: Proposes to use lane information to predict the future relationships between vehicles and generate multimodal trajectory predictions based on this;
Leveraging Importance Weights in Subset Selection
Gui Citovsky (Google Research), Yunjuan Wang (Johns Hopkins University)
ClassificationData-Centric LearningConvolutional Neural NetworkImage
🎯 What it does: The IWeS algorithm is proposed, which uses model entropy-based probabilities for importance sampling to select the most informative subset from large-scale datasets.
Leveraging Large Language Models for Multiple Choice Question Answering
Joshua Robinson (Brigham Young University), David Wingate (Brigham Young University)
TransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: By evaluating the multiple-choice answer symbol binding (MCSB) of large language models (LLMs) and replacing traditional cloze prompts (CP) with multiple-choice prompts (MCP), a systematic comparison of the performance of the two prompting methods on 20 multiple-choice question-answer datasets was conducted. It was found that models with high MCSB capability (such as Codex) showed significant improvement under MCP, nearly reaching or exceeding SOTA.
Leveraging Unlabeled Data to Track Memorization
Mahsa Forouzesh (École Polytechnique Fédérale de Lausanne), Patrick Thiran (École Polytechnique Fédérale de Lausanne)
ClassificationOptimizationData-Centric LearningConvolutional Neural NetworkImage
🎯 What it does: A metric that only uses unlabeled data is proposed - susceptibility to noisy labels, which is used to track the model's memory of noisy labels without the need for true labels.
LexMAE: Lexicon-Bottlenecked Pretraining for Large-Scale Retrieval
Tao Shen (Microsoft), Daxin Jiang (Microsoft)
RetrievalTransformerAuto EncoderContrastive LearningText
🎯 What it does: This paper proposes the LexMAE (Lexicon-Bottlenecked Masked AutoEncoder) model, which introduces a vocabulary importance bottleneck during the pre-training phase, allowing the masked language model to learn vocabulary weights that better meet retrieval needs, thus directly transferring to large-scale retrieval tasks.
LiftedCL: Lifting Contrastive Learning for Human-Centric Perception
Ziwei Chen (Southeast University), Wankou Yang (Southeast University)
Pose EstimationRepresentation LearningGenerative Adversarial NetworkContrastive LearningImage
🎯 What it does: We propose LiftedCL, which enhances 2D contrastive learning representations to 3D skeletons and implements adversarial learning to achieve unsupervised pre-training for human-centered tasks.
Light Sampling Field and BRDF Representation for Physically-based Neural Rendering
Jing Yang (Institute for Creative Technologies), Yajie Zhao (Institute for Creative Technologies)
GenerationData SynthesisNeural Radiance FieldImage
🎯 What it does: A light sampling field and layered BRDF model are proposed for neural physical rendering, achieving high-quality real-time rendering of facial skin.
LightGCL: Simple Yet Effective Graph Contrastive Learning for Recommendation
Xuheng Cai (University of Hong Kong), Xubin Ren (University of Hong Kong)
Recommendation SystemGraph Neural NetworkContrastive LearningGraph
🎯 What it does: A lightweight framework called LightGCL based on graph contrastive learning is proposed in the recommendation system, which enhances the global structure of the user-item interaction graph using singular value decomposition (SVD), and then performs contrastive learning with the original local information to improve recommendation performance under sparsity and popularity bias.
LilNetX: Lightweight Networks with EXtreme Model Compression and Structured Sparsification
Sharath Girish (University of Maryland), Abhinav Shrivastava (University of Maryland)
ClassificationCompressionConvolutional Neural NetworkImage
🎯 What it does: This paper proposes LilNetX, a unified end-to-end training framework that can simultaneously optimize model accuracy, disk storage size, and inference computation, achieving model compression and structured sparsity.
Limitless Stability for Graph Convolutional Networks
Christian Koke
Graph Neural NetworkGraph
🎯 What it does: This study investigates the stability and transferability of Graph Convolutional Networks (GCN), providing strict theoretical guarantees for node perturbations, edge weight perturbations, and structural (topological) perturbations, and proposes new filter spaces and resolution operator methods.
Linear Connectivity Reveals Generalization Strategies
Jeevesh Juneja (Delhi Technological University), Naomi Saphra (New York University)
ClassificationTransformerSupervised Fine-TuningText
🎯 What it does: This study investigates the linear mode connectivity in text classification tasks, finding that different fine-tuning results fall into different basins, which are associated with various generalization strategies.
Linear Convergence of Natural Policy Gradient Methods with Log-Linear Policies
Rui Yuan (Meta AI), Lin Xiao (Meta AI)
OptimizationReinforcement Learning
🎯 What it does: Under the infinite discount Markov decision process, the convergence properties of Natural Policy Gradient (NPG) and Q-NPG methods in the log-linear policy class are studied, proving that they achieve linear convergence in function approximation environments and providing upper bounds on sample complexity.
Linearly Mapping from Image to Text Space
Jack Merullo (Brown University), Ellie Pavlick (Brown University)
GenerationRetrievalTransformerLarge Language ModelPrompt EngineeringVision Language ModelImageTextMultimodality
🎯 What it does: This paper proposes a single-layer linear mapping (LiMBeR) that directly projects the representations of a frozen image encoder into soft prompts for a language model, enabling the text model to generate image descriptions and answer visual questions.
Link Prediction with Non-Contrastive Learning
William Shiao (University of California), Neil Shah (Snap Inc.)
Recommendation SystemRepresentation LearningGraph Neural NetworkContrastive LearningGraph
🎯 What it does: This paper focuses on non-contrastive learning methods in graph self-supervised learning, studying their performance in link prediction tasks, and proposes an improved non-contrastive framework called T-BGRL.
LipsFormer: Introducing Lipschitz Continuity to Vision Transformers
Xianbiao Qi (International Digital Economy Academy), Lei Zhang (International Digital Economy Academy)
ClassificationTransformerImage
🎯 What it does: A Transformer structure based on Lipschitz continuity, called LipsFormer, is proposed to fundamentally address the training instability issues of Transformers.
Liquid Structural State-Space Models
Ramin Hasani (Massachusetts Institute of Technology), Daniela Rus (Massachusetts Institute of Technology)
Recurrent Neural NetworkTransformerTime SeriesSequentialBenchmarkAudio
🎯 What it does: This paper presents Liquid-S4, a sequence modeling framework that combines Linear Liquid Time-Constant Networks (LTC) with Structured State Space Models (S4).
LMC: Fast Training of GNNs via Subgraph Sampling with Provable Convergence
Zhihao Shi (University of Science and Technology of China), Jie Wang (University of Science and Technology of China)
Graph Neural NetworkGraph
🎯 What it does: This paper proposes a subgraph sampling training method called LMC (Local Message Compensation), which achieves an unbiased estimate of the gradients of Graph Neural Networks (GNNs) by compensating for the discarded messages during forward and backward propagation, enabling scalable and efficient training on large-scale graphs.
LMSeg: Language-guided Multi-dataset Segmentation
Qiang Zhou (Alibaba Group), Fan Wang (Alibaba Group)
SegmentationTransformerVision Language ModelContrastive LearningImage
🎯 What it does: This paper proposes a language-guided multi-dataset segmentation framework called LMSeg, which supports semantic and panoptic segmentation. It utilizes text embeddings to unify the category representations of different datasets and achieves cross-dataset learning through category-guided decoding and dataset-aware enhancement.
Localized Randomized Smoothing for Collective Robustness Certification
Jan Schuchardt (Technical University of Munich), Stephan Günnemann (Technical University of Munich)
Object DetectionSegmentationAnomaly DetectionGraph Neural NetworkGaussian SplattingImageGraphStochastic Differential Equation
🎯 What it does: Proposes a Localized Randomized Smoothing method to prove the collective robustness of multi-output models under a single input.
Logical Entity Representation in Knowledge-Graphs for Differentiable Rule Learning
Chi Han (University of Illinois Urbana-Champaign), Heng Ji (University of Illinois Urbana-Champaign)
Knowledge DistillationRepresentation LearningGraph Neural NetworkGraphBiomedical Data
🎯 What it does: This study investigates a representation that encodes local subgraph information of entities into differentiable logical functions (LERP) and embeds it into probabilistic logic rule learning to enhance knowledge graph completion.
Logical Message Passing Networks with One-hop Inference on Atomic Formulas
Zihao Wang (Hong Kong University of Science and Technology), Simon See (NVIDIA)
Knowledge DistillationGraph Neural NetworkGraph
🎯 What it does: A logic message passing network (LMPNN) based on pre-trained knowledge graph embeddings is proposed to solve complex queries in EFO-1;
LogicDP: Creating Labels for Graph Data via Inductive Logic Programming
Yuan Yang (Georgia Institute of Technology), Ali Payani (Cisco)
ClassificationExplainability and InterpretabilityComputational EfficiencyGraph Neural NetworkReinforcement LearningGraph
🎯 What it does: This paper proposes LOGICDP, a data programming-based framework that generates training labels for graph data using automatically learned logical rules, thereby training graph reasoning models.
Long Range Language Modeling via Gated State Spaces
Harsh Mehta (Google Research), Behnam Neyshabur (Deepmind)
TransformerLarge Language ModelTextSequential
🎯 What it does: This paper proposes a gated state space-based long sequence autoregressive language model (GSS) and validates its fast training speed and competitive performance on long-range language modeling tasks through comparisons with Transformer and S4/DSS.
Long-Tailed Learning Requires Feature Learning
Thomas Laurent (Loyola Marymount University), Xavier Bresson (National University of Singapore)
ClassificationRepresentation LearningData-Centric LearningText
🎯 What it does: A long-tail distribution data model inspired by natural data (text, images) is proposed, and it is proven that learning features is a necessary condition for achieving good generalization.
Long-Tailed Partial Label Learning via Dynamic Rebalancing
Feng Hong (Shanghai Jiao Tong University), Yanfeng Wang (Shanghai Jiao Tong University)
ClassificationImage
🎯 What it does: This paper addresses the challenge of long-tail and partial label learning (LT-PLL) by proposing a dynamic rebalancing method called RECORDS, which is seamlessly integrated into the existing self-training PLL framework.
Loss Landscapes are All You Need: Neural Network Generalization Can Be Explained Without the Implicit Bias of Gradient Descent
Ping-yeh Chiang (University of Maryland), Tom Goldstein (University of Maryland)
OptimizationConvolutional Neural NetworkImage
🎯 What it does: Research shows that even without using gradient information, zero-order optimizers (such as guess-check, pattern search, and random greedy search) can achieve comparable or even better generalization performance than SGD in over-parameterized models; the 'volume hypothesis' has been experimentally validated - good solutions occupy a larger volume in the parameter space, making them naturally selected by random search.
Lossless Adaptation of Pretrained Vision Models For Robotic Manipulation
Mohit Sharma (Carnegie Mellon University), Yusuf Aytar (DeepMind)
Robotic IntelligenceTransformerImage
🎯 What it does: A method was designed and validated to enhance the performance of robotic manipulation tasks by inserting lightweight adapters at different layers of the model while keeping the original visual model functionality unchanged.
Lower Bounds on the Depth of Integral ReLU Neural Networks via Lattice Polytopes
Christian Alexander Haase (Freie Universität Berlin), Georg Loho (University of Twente)
🎯 What it does: It is proven that under integer weight conditions, the depth of ReLU neural networks strictly affects the set of representable functions; specifically, it is shown that for n=2^k, max_0{x_1,…,x_n} cannot be realized by a k-layer integer weight ReLU network.
LPT: Long-tailed Prompt Tuning for Image Classification
Bowen Dong (Harbin Institute of Technology), Wangmeng Zuo (Peng Cheng Laboratory)
ClassificationTransformerPrompt EngineeringImage
🎯 What it does: Proposes Long-tailed Prompt Tuning (LPT), which incorporates shared prompts and group-specific prompts into a frozen pre-trained ViT model, fine-tuning only a minimal number of parameters to adapt to long-tail classification tasks.
LS-IQ: Implicit Reward Regularization for Inverse Reinforcement Learning
Firas Al-Hafez (TU Darmstadt), Jan Peters (TU Darmstadt)
Reinforcement LearningSequential
🎯 What it does: A new inverse reinforcement learning algorithm, LS-IQ, is proposed, which improves learning stability through implicit reward regularization and effectively handles absorbing states.
M-L2O: Towards Generalizable Learning-to-Optimize by Test-Time Fast Self-Adaptation
Junjie Yang (Ohio State University), Yingbin Liang (Ohio State University)
OptimizationMeta LearningRecurrent Neural NetworkReinforcement LearningTabular
🎯 What it does: A novel learning optimizer M-L2O is proposed, which obtains an initialization point that can quickly converge with only a few adaptive steps during testing through meta-learning in the training phase.
MA-BERT: Towards Matrix Arithmetic-only BERT Inference by Eliminating Complex Non-Linear Functions
Neo Wei Ming (Agency for Science Technology and Research), Tao Luo (Agency for Science Technology and Research)
Computational EfficiencyKnowledge DistillationTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This paper proposes MA-BERT, which replaces all complex nonlinear functions (softmax, GELU, LayerNorm) in the Transformer model with implementations that only use matrix operations and ReLU to achieve more efficient inference.
Machine Unlearning of Federated Clusters
Chao Pan (University of Illinois), Olgica Milenkovic (University of Illinois)
Federated LearningSafty and PrivacyBiomedical Data
🎯 What it does: A machine unlearning framework for federated clustering is proposed, which includes Secure Compressed Multi-Set Aggregation (SCMA) and an efficient unlearning K-means++ initialization.
MACTA: A Multi-agent Reinforcement Learning Approach for Cache Timing Attacks and Detection
Jiaxun Cui (University of Texas at Austin), Yuandong Tian (Meta AI)
Adversarial AttackTransformerReinforcement LearningSequential
🎯 What it does: In this paper, the authors propose a multi-agent reinforcement learning framework called MACTA, which is used for the automatic learning and discovery of cache timing attack (CTA) strategies and corresponding detection strategies, and they have constructed a Gym environment, MA-AUTOCAT, that can realistically simulate CTAs.
MAESTRO: Open-Ended Environment Design for Multi-Agent Reinforcement Learning
Mikayel Samvelyan (Meta AI), Tim Rocktäschel (University College London)
Reinforcement Learning
🎯 What it does: Designed and implemented MAESTRO, a multi-agent unsupervised environment design method for two-player zero-sum games, which can automatically generate challenging tasks for both the environment and opponents during training, thereby enhancing the robustness of learners.
Make-A-Video: Text-to-Video Generation without Text-Video Data
Uriel Singer (Meta AI), Yaniv Taigman (Meta AI)
GenerationData SynthesisSuper ResolutionConvolutional Neural NetworkDiffusion modelVideoText
🎯 What it does: Extend the text-to-image generation model to a text-to-video generation framework called Make‑A‑Video.
Making Better Decision by Directly Planning in Continuous Control
Jinhua Zhu (University of Science and Technology of China), Houqiang Li (Microsoft Research AI4Science)
OptimizationRobotic IntelligenceReinforcement LearningSequential
🎯 What it does: A model-based reinforcement learning framework POMP is designed, which directly uses the learned continuous action space model as a planner, and proposes a Deep Differentiable Dynamic Programming (D3P) algorithm to optimize trajectories.
Making Substitute Models More Bayesian Can Enhance Transferability of Adversarial Examples
Qizhang Li (Harbin Institute of Technology), Hao Chen (University of California Davis)
Adversarial AttackConvolutional Neural NetworkTransformerGaussian SplattingImage
🎯 What it does: This paper proposes the use of a Bayesian model (adding probability distribution after given parameters) to enhance the transferability of adversarial attacks based on sub-models;
Malign Overfitting: Interpolation and Invariance are Fundamentally at Odds
Yoav Wald (Johns Hopkins University), Yair Carmon (Tel Aviv University)
ClassificationDomain AdaptationGaussian SplattingImageTabular
🎯 What it does: The study shows that in over-parameterized linear models, any interpolation learning rule with positive normalized margins cannot achieve invariance, and proposes a two-stage non-interpolation algorithm to implement a robust invariant classifier.
ManiSkill2: A Unified Benchmark for Generalizable Manipulation Skills
Jiayuan Gu (University of California San Diego), Hao Su (Tsinghua University)
Robotic IntelligenceReinforcement LearningPoint CloudBenchmark
🎯 What it does: Designed and implemented the ManiSkill2 benchmark, which includes 20 types of manipulation task families, over 2000 object models, 4 million demonstration frames, and provides a unified Gym interface, convertible action space, bi-directional coupling of soft and rigid body GPU MPM simulation, and an efficient asynchronous rendering/rendering server framework.
ManyDG: Many-domain Generalization for Healthcare Applications
Chaoqi Yang (University of Illinois Urbana-Champaign), Jimeng Sun (University of Illinois Urbana-Champaign)
Domain AdaptationRecommendation SystemBiomedical DataElectronic Health Records
🎯 What it does: The research addresses the multi-patient domain generalization problem in the medical field and proposes the ManyDG model.
MapTR: Structured Modeling and Learning for Online Vectorized HD Map Construction
Bencheng Liao (Huazhong University of Science and Technology), Chang Huang (Horizon Robotics)
Autonomous DrivingTransformerPoint Cloud
🎯 What it does: This paper proposes MapTR, a structured end-to-end online vectorization framework for high-definition map construction based on Transformer;
Markup-to-Image Diffusion Models with Scheduled Sampling
Yuntian Deng (Harvard University), Alexander M Rush
GenerationData SynthesisTransformerDiffusion modelImageText
🎯 What it does: This paper studies a novel markup-to-image generation task and proposes a fully data-driven rendering framework based on diffusion models.
MARS: Meta-learning as Score Matching in the Function Space
Krunoslav Lehman Pavasovic (ETH Zurich), Andreas Krause (ETH Zurich)
Meta LearningTransformerScore-based ModelBiomedical Data
🎯 What it does: A method for meta-learning called MARS is proposed, which estimates the score (gradient) of the data generation process in function space.
Martingale Posterior Neural Processes
Hyungi Lee (Korea Advanced Institute of Science and Technology), Juho Lee (Korea Advanced Institute of Science and Technology)
GenerationOptimizationImageTabular
🎯 What it does: This paper proposes a neural process model using Martingale Posterior (MPNP), which captures the uncertainty of functions by generating pseudo-context data and utilizing its uncertainty.
Masked Distillation with Receptive Tokens
Tao Huang (University of Sydney), Chang Xu (University of Sydney)
Object DetectionSegmentationKnowledge DistillationConvolutional Neural NetworkImage
🎯 What it does: A mask distillation method called MasKD based on learnable receptive tokens is proposed for feature distillation in dense prediction tasks such as object detection and semantic segmentation.
Masked Frequency Modeling for Self-Supervised Visual Pre-Training
Jiahao Xie (Nanyang Technological University), Chen Change Loy (Nanyang Technological University)
SegmentationRepresentation LearningConvolutional Neural NetworkTransformerImage
🎯 What it does: A self-supervised visual pre-training method (MFM) is proposed for low-pass/high-pass masking and predicting missing frequencies in the frequency domain.
Masked Image Modeling with Denoising Contrast
Kun Yi (Peking University), Xiaohu Qie
SegmentationRepresentation LearningTransformerContrastive LearningImage
🎯 What it does: A pure mask image modeling method called ConMIM is proposed, which achieves mask completion pre-training through denoising contrastive learning, avoiding the use of offline image tokenizers.
Masked Unsupervised Self-training for Label-free Image Classification
Junnan Li (Salesforce Research), Steven Hoi (Salesforce Research)
ClassificationTransformerContrastive LearningImage
🎯 What it does: The MASKED UNSUPERVISED SELF-TRAINING (MUST) method is proposed, which performs self-supervised fine-tuning on a pre-trained zero-shot classifier (such as CLIP) using unlabeled images to enhance classification performance.
Masked Vision and Language Modeling for Multi-modal Representation Learning
Gukyeong Kwon (AWS AI Labs), Stefano Soatto (AWS AI Labs)
RetrievalRepresentation LearningTransformerVision Language ModelContrastive LearningImageTextMultimodality
🎯 What it does: This paper proposes a joint masked visual-language modeling (MaskVLM) approach, which masks both images and text during pre-training and utilizes the complete information from the other modality to reconstruct the masked modality, thereby learning cross-modal alignment and semantic representation.
MaskFusion: Feature Augmentation for Click-Through Rate Prediction via Input-adaptive Mask Fusion
Chao Liao (Kuaishou Technology), Chengru Song (Kuaishou Technology)
Recommendation SystemTabular
🎯 What it does: The MaskFusion framework is proposed, which enhances CTR prediction performance through instance-adaptive masks for feature fusion at the DNN layer.
MaskViT: Masked Visual Pre-Training for Video Prediction
Agrim Gupta (Stanford University), Li Fei-Fei (Stanford University)
GenerationRobotic IntelligenceTransformerGenerative Adversarial NetworkVideo
🎯 What it does: This paper proposes MaskViT, a Transformer video prediction framework based on Masked Visual Modeling, which utilizes window attention for efficient encoding and employs iterative non-autoregressive decoding to generate future frames.
Mass-Editing Memory in a Transformer
Kevin Meng (Massachusetts Institute of Technology), David Bau (Northeastern University)
TransformerLarge Language ModelText
🎯 What it does: A large-scale memory editing method called MEMIT has been developed, which allows for direct modification of Transformer weights to batch update factual memories in LLMs without retraining.
Massively Scaling Heteroscedastic Classifiers
Mark Collier (Google AI), Effrosyni Kokiopoulou (Google AI)
ClassificationConvolutional Neural NetworkTransformerContrastive LearningImage
🎯 What it does: A novel heteroscedastic classifier HET-XL is proposed, which can decouple the additional parameter count from the number of classes in large-scale multi-class classification problems and automatically learn the temperature hyperparameter.
MAST: Masked Augmentation Subspace Training for Generalizable Self-Supervised Priors
Chen Huang (Apple Inc), Joshua M. Susskind
Object DetectionSegmentationRepresentation LearningContrastive LearningImage
🎯 What it does: A self-supervised learning framework named MAST is proposed, which splits the invariance corresponding to different data augmentations into several subspaces through learnable masks in a single feature space, thereby achieving general feature representation transferable to multiple tasks, and introduces uncertainty modeling to reduce similarity conflicts caused by strong augmentations.
Mastering the Game of No-Press Diplomacy via Human-Regularized Reinforcement Learning and Planning
Anton Bakhtin (Meta AI), Noam Brown (Meta AI)
Reinforcement LearningSequential
🎯 What it does: Developed and trained an AI for a non-verbal diplomacy game named Diplodocus, which combines human imitation with reinforcement learning using RL-DiL-piKL, and achieved first place in 200 games against humans.
Matching receptor to odorant with protein language and graph neural networks
Matej Hladiš (Université Côte d'Azur), Jérémie Topin (Université Côte d'Azur)
ClassificationDrug DiscoveryGraph Neural NetworkLarge Language ModelGraphBiomedical Data
🎯 What it does: This study investigates how to use protein language models and molecular graph neural networks to predict the activation of olfactory receptors (OR) by odor molecules, and to determine whether a molecule can activate a given OR through a binary classification approach.
Max-Margin Works while Large Margin Fails: Generalization without Uniform Convergence
Margalit Glasgow (Stanford University), Tengyu Ma (Stanford University)
ClassificationOptimization
🎯 What it does: In the context of over-parameterized linear classification and the two-layer neural network (XOR) problem, it is theoretically proven that when the model is near-max-margin, good generalization performance can still be achieved even in cases where uniform convergence (UC) fails.
Maximizing Communication Efficiency for Large-scale Training via 0/1 Adam
Yucheng Lu (Cornell University), Yuxiong He (Microsoft)
OptimizationTransformerLarge Language ModelImageText
🎯 What it does: A new Adam optimizer, 0/1 Adam, is proposed, which can achieve 1-bit gradient compression and local steps simultaneously in large-scale distributed training while maintaining the adaptive characteristics of Adam.
Maximizing Spatio-Temporal Entropy of Deep 3D CNNs for Efficient Video Recognition
Junyan Wang (University of New South Wales), Yang Song (University of New South Wales)
RecognitionOptimizationComputational EfficiencyNeural Architecture SearchConvolutional Neural NetworkVideo
🎯 What it does: Derive the space-time entropy score through the maximum entropy principle, and implement training-independent 3D CNN architecture search using analytical formulas;
MCAL: Minimum Cost Human-Machine Active Labeling
Hang Qiu (University of Southern California), Ramesh Govindan (University of Southern California)
ClassificationOptimizationConvolutional Neural NetworkImage
🎯 What it does: A minimal cost mixed human-machine active labeling framework (MCAL) is proposed, which jointly selects samples for human labeling and model training, and uses the trained classifier to automatically label the remaining samples, thereby minimizing the total labeling and training costs while satisfying the error threshold.
Measure the Predictive Heterogeneity
Jiashuo Liu (Tsinghua University), Peng Cui (Tsinghua University)
ClassificationDomain AdaptationOptimizationConvolutional Neural NetworkImageTabularAgriculture Related
🎯 What it does: The concept of Predictive Heterogeneity, which measures the predictive performance of machine learning models, is proposed, along with its theoretical definition, properties, and PAC estimation methods; the Information Maximization (IM) algorithm is used to partition data into subgroups, uncover heterogeneity, and enhance out-of-distribution (OOD) generalization.
Measuring axiomatic soundness of counterfactual image models
Miguel Monteiro (Imperial College London), Ben Glocker (Imperial College London)
GenerationData SynthesisAuto EncoderGenerative Adversarial NetworkImage
🎯 What it does: A framework based on Pearl's counterfactual axioms (composition, reversibility, effectiveness) is proposed to evaluate the soundness of image counterfactual models.
Measuring Forgetting of Memorized Training Examples
Matthew Jagielski (Google), Chiyuan Zhang (Google)
Safty and PrivacyConvolutional Neural NetworkTransformerImageTextAudio
🎯 What it does: The paper studies the forgetting behavior of deep learning models towards seen samples during the training process through the construction of privacy attack experiments and theoretical analysis, exploring its impact on privacy leakage.
MECTA: Memory-Economic Continual Test-Time Model Adaptation
Junyuan Hong (Michigan State University), Michael Spranger (Sony AI)
Domain AdaptationComputational EfficiencyConvolutional Neural NetworkSupervised Fine-TuningImage
🎯 What it does: A method called MECTA is proposed to reduce memory usage during gradient-based continual learning while maintaining or even improving out-of-distribution (OOD) recognition performance.
MEDFAIR: Benchmarking Fairness for Medical Imaging
Yongshuo Zong (University of Edinburgh), Timothy Hospedales (Samsung AI Centre)
Convolutional Neural NetworkImageMultimodalityBiomedical DataMagnetic Resonance ImagingComputed Tomography
🎯 What it does: A MEDFAIR benchmark was constructed to systematically evaluate fairness algorithms in medical imaging.
MEDICAL IMAGE UNDERSTANDING WITH PRETRAINED VISION LANGUAGE MODELS: A COMPREHENSIVE STUDY
Ziyuan Qin (West China Biomedical Big Data Center), Kang Li (West China Biomedical Big Data Center)
Object DetectionTransformerSupervised Fine-TuningPrompt EngineeringVision Language ModelImageBiomedical Data
🎯 What it does: This paper studies the transfer of pre-trained vision-language models (such as GLIP) from natural images to the medical imaging domain, enhancing object detection performance through the design or automatic generation of medical prompts.
Mega: Moving Average Equipped Gated Attention
Xuezhe Ma (Information Sciences Institute University of Southern California), Luke Zettlemoyer (Meta AI)
RetrievalOptimizationTransformerSupervised Fine-TuningImageText
🎯 What it does: MEGA is proposed—a Transformer replacement module that embeds Exponential Moving Average (EMA) into single-head gated attention, combining position awareness and scalability.
Memorization Capacity of Neural Networks with Conditional Computation
Erdem Koyuncu (University of Illinois Chicago)
OptimizationComputational Efficiency
🎯 What it does: This paper studies conditional computation in neural networks from the perspective of memory capacity, providing a general method to transform unconditional networks into conditional networks, and proving that under this framework, it is possible to perfectly memorize n input-output pairs with O(log n) operations.
Memorization-Dilation: Modeling Neural Collapse Under Noise
Duc Anh Nguyen (LMU Munich), Gitta Kutyniok (LMU Munich)
ClassificationOptimizationImage
🎯 What it does: The study investigates the phenomenon of Neural Collapse in neural networks under labeled noise environments, and introduces positive constraints and network expressiveness limitations in the layer-peeled model, constructing the Memorization-Dilation (MD) model to explain the enhancement of generalization by Label Smoothing (LS).
Memory Gym: Partially Observable Challenges to Memory-Based Agents
Marco Pleines (TU Dortmund University), Mike Preuss (LIACS Universiteit Leiden)
Recurrent Neural NetworkReinforcement LearningSequentialBenchmark
🎯 What it does: Proposed the Memory Gym benchmark, which includes three partially observable, memory-requiring 2D tasks (Mortar Mayhem, Mystery Path, Searing Spotlights) with adjustable difficulty;
MeshDiffusion: Score-based Generative 3D Mesh Modeling
Zhen Liu (Mila), Weiyang Liu (Max Planck Institute for Intelligent Systems)
GenerationData SynthesisConvolutional Neural NetworkDiffusion modelScore-based ModelPoint CloudMesh
🎯 What it does: This paper studies a 3D mesh generation method based on diffusion models called MeshDiffusion.
Meta Knowledge Condensation for Federated Learning
Ping Liu (Center for Frontier AI Research ASTAR), Joey Tianyi Zhou (Center for Frontier AI Research ASTAR)
Federated LearningKnowledge DistillationConvolutional Neural NetworkImage
🎯 What it does: The FedMK method is proposed, which extracts meta-knowledge on the client side and trains a global model on the server side to achieve efficient federated learning within limited communication rounds.
Meta Learning to Bridge Vision and Language Models for Multimodal Few-Shot Learning
Ivona Najdenkoska (University of Amsterdam), Marcel Worring (University of Amsterdam)
Meta LearningTransformerVision Language ModelImageTextMultimodality
🎯 What it does: A multi-modal few-shot learning framework based on meta-learning is proposed, which connects frozen large-scale visual and language models through a lightweight meta-mapper and quickly adapts to new tasks with a small number of samples.
Meta Temporal Point Processes
Wonho Bae (University of British Columbia), Gabriel L. Oliveira (Borealis AI)
Meta LearningTransformerTime SeriesSequential
🎯 What it does: A meta-learning framework is proposed that treats each event sequence as an independent task, using Neural Processes (NP) to model temporal point processes, and introduces latent variables and cross-attention modules based on this.
Meta-learning Adaptive Deep Kernel Gaussian Processes for Molecular Property Prediction
Wenlin Chen (University of Cambridge), José Miguel Hernández-Lobato (University of Cambridge)
Meta LearningDrug DiscoveryTabular
🎯 What it does: The ADKF-IFT framework is proposed for training deep kernel Gaussian processes (Deep Kernel GP) through meta-learning in few-shot molecular property prediction tasks, fitting the base kernel parameters at the task level via maximum likelihood.
Meta-Learning in Games
Keegan Harris (Carnegie Mellon University), Tuomas Sandholm (Carnegie Mellon University)
Meta LearningReinforcement Learning from Human FeedbackReinforcement LearningSequentialOrdinary Differential Equation
🎯 What it does: This paper studies meta-learning methods in multi-task game environments, providing theoretical convergence guarantees for zero-sum games, general-sum games, potential games, and Stackelberg games, and validates its acceleration effects in poker endgame experiments.