ICML 2024 Papers — Page 8
International Conference on Machine Learning · 2610 papers
Efficient and Effective Time-Series Forecasting with Spiking Neural Networks
Changze Lv (Fudan University), Dongsheng Li (Microsoft Research Asia)
Spiking Neural NetworkTransformerTime Series
🎯 What it does: This paper proposes a spiking neural network framework for time series prediction, which transforms continuous sequences into event-driven spike sequences using temporal alignment and spike encoding.
Efficient Black-box Adversarial Attacks via Bayesian Optimization Guided by a Function Prior
Shuyu Cheng (Tsinghua University), Jun Zhu (Tsinghua University)
OptimizationAdversarial AttackTransformerImage
🎯 What it does: By incorporating the loss function of the surrogate model as a global function prior into the Gaussian process, we propose Prior-guided Bayesian Optimization (P-BO) for efficient black-box adversarial attacks.
Efficient Contextual Bandits with Uninformed Feedback Graphs
Mengxiao Zhang (University of Southern California), Paul Mineiro (Microsoft Research)
Recommendation SystemOptimizationGraph Neural NetworkReinforcement LearningGraphTabular
🎯 What it does: This paper proposes a novel contextual multi-armed bandit algorithm, SquareCB UG, which can simultaneously learn the loss and feedback graph in an 'unforeseen' feedback graph environment, achieving efficient learning and obtaining approximately optimal returns.
Efficient Contrastive Learning for Fast and Accurate Inference on Graphs
Teng Xiao (Penn State University), Vasant G Honavar (Penn State University)
ClassificationComputational EfficiencyRepresentation LearningGraph Neural NetworkContrastive LearningGraph
🎯 What it does: This paper proposes GraphECL, a cross-model contrastive learning framework that utilizes MLP as the encoder during inference, avoiding the message passing of traditional GNNs and significantly reducing inference latency.
Efficient Denoising Diffusion via Probabilistic Masking
WEIZHONG ZHANG, Kani Chen (Hong Kong University of Science and Technology)
RestorationGenerationData SynthesisOptimizationComputational EfficiencyDiffusion modelImageTime Series
🎯 What it does: By introducing a probabilistic masking module, the diffusion model and mask probabilities are jointly trained to automatically identify and eliminate redundant diffusion steps, significantly accelerating the sampling process.
Efficient Error Certification for Physics-Informed Neural Networks
Francisco Eiras (University of Oxford), M. Pawan Kumar (Google DeepMind)
OptimizationExplainability and InterpretabilityComputational EfficiencyTabularPhysics Related
🎯 What it does: A CROWN-based error certification framework ∂-CROWN is proposed, which can provide an upper bound on the residual error in the PINN continuous domain and further tighten it through a greedy input branching.
Efficient Exploration for LLMs
Vikranth Dwaracherla (Google DeepMind), Benjamin Van Roy (Stanford University)
Reinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: With limited human feedback, fine-tuning large language models through active exploration methods (especially double Thompson sampling) significantly improves the quality of responses to prompts.
Efficient Exploration in Average-Reward Constrained Reinforcement Learning: Achieving Near-Optimal Regret With Posterior Sampling
Danil Provodin (Eindhoven University of Technology), Mykola Pechenizkiy (University of Jyvaskyla)
Reinforcement Learning
🎯 What it does: This paper proposes a posterior sampling-based algorithm PSCONRL for efficient exploration and learning in infinite average reward constrained Markov decision processes (CMDP).
Efficient Low-Rank Matrix Estimation, Experimental Design, and Arm-Set-Dependent Low-Rank Bandits
Kyoungseok Jang (Università degli Studi di Milano), Kwang-Sung Jun (University of Arizona)
OptimizationReinforcement Learning from Human FeedbackTabular
🎯 What it does: A low-rank matrix estimation method called LowPopArt is proposed, along with a difficulty measure B_Q based on the covariance of the measurement distribution; subsequently, two low-rank bandit algorithms, LPA-ETC and LPA-ESTR, suitable for any arm set, are designed based on this estimator, achieving adaptive exploration of the arm set geometry and a better regret upper bound.
Efficient Mixture Learning in Black-Box Variational Inference
Alexandra Hotti (KTH Royal Institute of Technology), Jens Lagergren (KTH Royal Institute of Technology)
OptimizationComputational EfficiencyRepresentation LearningMixture of ExpertsAuto EncoderImage
🎯 What it does: Proposes the MISVAE architecture and two new MISELBO estimators S2A and S2S, enhancing the scalability and efficiency of black-box variational inference for mixture distributions.
Efficient Non-stationary Online Learning by Wavelets with Applications to Online Distribution Shift Adaptation
Yu-Yang Qian (Nanjing University), Zhi-Hua Zhou (Nanjing University)
Domain AdaptationOptimizationTabularTime Series
🎯 What it does: A non-stationary online learning framework based on a single-layer model is proposed, utilizing waveform detection and adaptive restart to track environmental drift; this framework is also applied to the online label shift problem, introducing new algorithms Wav-O and Wav-R.
Efficient Online Set-valued Classification with Bandit Feedback
Zhou Wang (Binghamton University), Xingye Qiao (Binghamton University)
ClassificationReinforcement LearningImage
🎯 What it does: This paper proposes a Bandit Class-specific Conformal Prediction (BCCP) in an online multi-class bandit feedback environment, generating prediction sets based on adaptive quantiles through unbiased label estimation.
Efficient PAC Learnability of Dynamical Systems Over Multilayer Networks
Zirou Qiu (University of Virginia), Anil Kumar Vullikanti
Graph
🎯 What it does: For threshold dynamic systems under multilayer networks, an efficient PAC learning algorithm is proposed, and upper and lower bounds for its sample complexity and Natarajan dimension are provided.
Efficient Pareto Manifold Learning with Low-Rank Structure
Weiyu Chen (Hong Kong University of Science and Technology), James Kwok
OptimizationRepresentation LearningImageTabular
🎯 What it does: A structure utilizing a main network combined with low-rank matrices is proposed to continuously approximate the Pareto front of multi-objective tasks, significantly reducing the number of parameters and enhancing shared feature learning.
Efficient Policy Evaluation with Offline Data Informed Behavior Policy Design
Shuze Liu (University of Virginia), Shangtong Zhang (University of Virginia)
OptimizationRobotic IntelligenceReinforcement LearningTabular
🎯 What it does: Utilize offline data to learn a closed-form behavior policy, enhancing the unbiased Monte Carlo evaluation variance of the target policy and reducing the required amount of online interactions.
Efficient Precision and Recall Metrics for Assessing Generative Models using Hubness-aware Sampling
Yuanbang Liang (Cardiff University), Yipeng Qin (Cardiff University)
GenerationComputational EfficiencyDiffusion modelGenerative Adversarial NetworkImage
🎯 What it does: An efficient precision and recall (eP&R) metric is proposed, which significantly reduces the computational cost of evaluating generative models through hubness-aware sampling.
Efficient Stochastic Approximation of Minimax Excess Risk Optimization
Lijun Zhang (Nanjing University), Yao Hu (Xiaohongshu Inc.)
OptimizationTabular
🎯 What it does: Developed an efficient stochastic approximation method that directly optimizes the minimization of maximum excess risk (MERO), providing nearly optimal convergence rates and algorithms for arbitrary stopping;
Efficient Value Iteration for s-rectangular Robust Markov Decision Processes
Navdeep Kumar (Technion), Shie Mannor (Nvidia)
OptimizationReinforcement Learning
🎯 What it does: This paper studies s-rectangular robust Markov decision processes (MDPs) and proposes an efficient value iteration method that can quickly compute the optimal robust Bellman operator in large state spaces.
Efficient World Models with Context-Aware Tokenization
Vincent Micheli (University of Geneva), François Fleuret (University of Geneva)
Computational EfficiencyTransformerReinforcement LearningAuto EncoderWorld ModelImageVideoBenchmark
🎯 What it does: This paper proposes a new scalable world model based on Transformer, ∆-IRIS, which utilizes a context-aware discrete autoencoder to encode only the random increments (delta) between time steps, and inserts continuous I-tokens into the dynamic model to reduce the number of tokens and enhance inference efficiency, thereby achieving faster and more efficient imagination-based reinforcement learning.
EfficientZero V2: Mastering Discrete and Continuous Control with Limited Data
Shengjie Wang (Tsinghua University), Yang Gao (Tsinghua University)
Robotic IntelligenceReinforcement LearningWorld ModelImageVideoBenchmark
🎯 What it does: EfficientZero V2 (EZ-V2) is proposed, a general, sample-efficient reinforcement learning framework aimed at discrete and continuous control, visual and low-dimensional input, and sparse and dense rewards.
EiG-Search: Generating Edge-Induced Subgraphs for GNN Explanation in Linear Time
Shengyao Lu (University of Alberta), Di Niu (University of Alberta)
Explainability and InterpretabilityComputational EfficiencyGraph Neural NetworkGraph
🎯 What it does: A training-independent, linear-time subgraph explanation method called EiG-Search is proposed, which directly ranks edges by importance and searches for the best subgraph to explain GNN predictions.
ELF: Encoding Speaker-Specific Latent Speech Feature for Speech Synthesis
Jungil Kong (SK Telecom), Sangjin Kim (SK Telecom)
GenerationData SynthesisTransformerFlow-based ModelAuto EncoderGenerative Adversarial NetworkAudio
🎯 What it does: Proposes the ELF method, which achieves zero-shot multi-speaker TTS without training by encoding and clustering speaker latent features.
ELTA: An Enhancer against Long-Tail for Aesthetics-oriented Models
Limin Liu (Beijing University of Posts and Telecommunications), Huadong Ma (Beijing University of Posts and Telecommunications)
ClassificationContrastive LearningImage
🎯 What it does: A long-tail enhancement framework ELTA is proposed for the task of image aesthetic assessment, addressing the issue of sample imbalance.
Eluder-based Regret for Stochastic Contextual MDPs
Orin Levy (Tel Aviv University), Yishay Mansour (Tel Aviv University)
OptimizationReinforcement Learning
🎯 What it does: The E-UCRL algorithm is proposed, achieving sublinear regret in the context of a random Contextual Markov Decision Process (CMDP) under the premise of relying solely on an offline regression oracle (least squares and log loss).
Embarrassingly Parallel GFlowNets
Tiago Silva, Diego Mesquita (Getulio Vargas Foundation)
OptimizationFederated LearningFlow-based ModelContrastive LearningTabularSequential
🎯 What it does: A GFlowNet sampling method called EP-GFlowNet has been developed for use in distributed/federated environments, which can aggregate locally trained GFlowNets from various clients into a global GFlowNet without sharing local rewards, thereby achieving efficient sampling of target posteriors or product distributions.
Embodied CoT Distillation From LLM To Off-the-shelf Agents
Wonje Choi (Sungkyunkwan University), Honguk Woo (Sungkyunkwan University)
Knowledge DistillationRobotic IntelligenceTransformerLarge Language ModelContrastive LearningTextChain-of-Thought
🎯 What it does: This paper proposes the DEDER framework, which distills the embodied reasoning capabilities of large language models into a two-layer strategy for small language models, enabling complex task planning on resource-constrained devices.
EMC$^2$: Efficient MCMC Negative Sampling for Contrastive Learning with Global Convergence
Chung-Yiu Yau (Chinese University of Hong Kong), Mingyi Hong (University of Minnesota)
OptimizationRepresentation LearningContrastive LearningImage
🎯 What it does: A negative sampling method based on MCMC (EMC 2) is proposed for efficiently optimizing global contrastive learning loss, achieving convergence under small batch training.
Emergence of In-Context Reinforcement Learning from Noise Distillation
Ilya Zisman (AIRI), Sergey Kolesnikov (Tinkoff)
Knowledge DistillationTransformerReinforcement LearningSequential
🎯 What it does: The ADε method is proposed, which generates learning history by injecting adjustable noise into the policy, allowing the Transformer to achieve unsupervised in-context reinforcement learning without the need for a large amount of single-task RL or optimal policies.
Emergent Equivariance in Deep Ensembles
Jan E Gerken, Pan Kessel (Genentech Roche)
ClassificationData-Centric LearningConvolutional Neural NetworkImage
🎯 What it does: The study investigates deep ensemble models under the condition of full data augmentation, proving that in the infinite width limit, their average predictions maintain equivariance to symmetric transformations of the input data (such as rotations, flips, etc.); this equivariance is 'emergent', meaning individual members do not need to exhibit equivariance.
Emergent Representations of Program Semantics in Language Models Trained on Programs
Charles Jin (Massachusetts Institute of Technology), Martin Rinard (Massachusetts Institute of Technology)
TransformerLarge Language ModelText
🎯 What it does: Train a Transformer language model on a synthetic program corpus based on Karel DSL, learning program semantics solely through the next token prediction task, and extracting intermediate abstract states of program execution using a detector.
Empowering Graph Invariance Learning with Deep Spurious Infomax
Tianjun Yao (Mohamed bin Zayed University of Artificial Intelligence), Kun Zhang (Mohamed bin Zayed University of Artificial Intelligence)
Drug DiscoveryGraph Neural NetworkGraph
🎯 What it does: This paper proposes a graph-invariant learning paradigm called EQuAD based on deep self-supervised information maximization, which can reliably extract pseudo-features and eliminate pseudo-correlations with labels.
Enabling Few-Shot Learning with PID Control: A Layer Adaptive Optimizer
Le Yu (Southeast University), Fir Dunkin (Southeast University)
OptimizationMeta LearningImage
🎯 What it does: A layer adaptive PID (LA-PID) optimizer is proposed under the MAML framework for inner-loop gradient updates in few-shot learning tasks.
Enabling Uncertainty Estimation in Iterative Neural Networks
Nikita Durasov (Ecole Polytechnique Federale de Lausanne), Pascal Fua (Ecole Polytechnique Federale de Lausanne)
Object DetectionOptimizationGraph Neural NetworkImageMesh
🎯 What it does: This study investigates a method for estimating uncertainty using the convergence speed of outputs from iterative neural networks, applying it to aerial image road detection and two-dimensional and three-dimensional aerodynamic performance prediction and optimization.
Encodings for Prediction-based Neural Architecture Search
Yash Akhauri (Cornell University), Mohamed S Abdelfattah
Neural Architecture SearchGraph Neural NetworkTransformerAuto EncoderGraphBenchmark
🎯 What it does: A systematic evaluation and classification of encoding methods for predictive NAS is conducted, proposing the FLAN predictor that integrates graph convolution, graph attention, and operation embedding, and designs a unified encoding to achieve transfer learning across search spaces, tasks, and datasets.
End-to-End Neuro-Symbolic Reinforcement Learning with Textual Explanations
Lirui Luo (Peking University), Qing Li (Peking University)
Explainability and InterpretabilityKnowledge DistillationConvolutional Neural NetworkLarge Language ModelReinforcement LearningImage
🎯 What it does: The INSIGHT framework is proposed, which can learn structured states from visual input in an end-to-end manner and utilize symbolic policies for reinforcement learning, while also generating natural language policy and decision explanations.
Energy-based Backdoor Defense without Task-Specific Samples and Model Retraining
Yudong Gao (China University of Petroleum), Huajie Shao (College of William and Mary)
Anomaly DetectionImageTextAudio
🎯 What it does: This paper proposes two energy statistics-based backdoor defense methods, EBBA and EBBA+, which can achieve backdoor model detection, trigger detection, and backdoor removal without task-specific samples and model retraining.
Energy-Efficient Gaussian Processes Using Low-Precision Arithmetic
Nicolas Alder (Hasso Plattner Institute), Ralf Herbrich (Hasso Plattner Institute)
OptimizationComputational EfficiencyTabular
🎯 What it does: This study investigates the impact of low-precision floating-point representation on the energy consumption of Gaussian process regression and proposes a method to significantly reduce energy consumption while maintaining predictive performance.
Energy-Guided Diffusion Sampling for Offline-to-Online Reinforcement Learning
Xu-Hui Liu (Nanjing University), Yang Yu (Nanjing University)
Reinforcement LearningDiffusion modelContrastive LearningTabularBenchmarkStochastic Differential Equation
🎯 What it does: Proposes the Energy-Guided Diffusion Sampling (EDIS) method, which utilizes diffusion models to generate samples guided by energy that match the online distribution, thereby alleviating the data distribution shift problem in offline-online reinforcement learning.
Enforcing Constraints in RNA Secondary Structure Predictions: A Post-Processing Framework Based on the Assignment Problem
Geewon Suh (Spidercore Inc.), Mingeun Kang (Spidercore Inc.)
OptimizationProtein Structure PredictionTransformerBiomedical Data
🎯 What it does: A post-processing framework based on the integer linear programming assignment problem is proposed to correct the RNA secondary structure predicted by machine learning models, ensuring that all hard constraints are satisfied.
Enhancing Adversarial Robustness in SNNs with Sparse Gradients
Yujia Liu (Peking University), Zhaofei Yu (Peking University)
Adversarial AttackSpiking Neural NetworkImage
🎯 What it does: Proposes training Spiking Neural Networks (SNN) through gradient sparse regularization to enhance their robustness against adversarial attacks.
Enhancing Class-Imbalanced Learning with Pre-Trained Guidance through Class-Conditional Knowledge Distillation
Lan Li (Nanjing University), De-Chuan Zhan (Nanjing University)
ClassificationKnowledge DistillationConvolutional Neural NetworkContrastive LearningImage
🎯 What it does: Utilizing large-scale pre-trained models (CLIP) to enhance the generalization ability of minority class features under imbalanced data through class conditional knowledge distillation (CCKD) and its enhanced version (ACCKD).
Enhancing Cross-Modal Fine-Tuning with Gradually Intermediate Modality Generation
Lincan Cai (Beijing Institute of Technology), Chengwei Zhu (Tencent)
Domain AdaptationKnowledge DistillationTransformerSupervised Fine-TuningMultimodalityTabularOrdinary Differential Equation
🎯 What it does: This paper proposes an end-to-end cross-modal fine-tuning method called PaRe, which gradually replaces patches after scoring the embeddings of the source and target modalities with a gating mechanism, generating intermediate modalities from easy to difficult, thereby bridging the modality gap and enhancing the stability and performance of cross-modal fine-tuning.
Enhancing Implicit Shape Generators Using Topological Regularizations
Liyan Chen (University of Texas at Austin), Qixing Huang (University of Texas at Austin)
GenerationData SynthesisPoint Cloud
🎯 What it does: This paper proposes the use of persistent graph generators and topological regularization to enhance the topological generalization ability of implicit shape generators.
Enhancing Size Generalization in Graph Neural Networks through Disentangled Representation Learning
Zheng Huang (Dartmouth), Yujun Yan (Dartmouth)
Representation LearningGraph Neural NetworkContrastive LearningGraph
🎯 What it does: This paper proposes the DISGEN framework, which enhances data through size-invariant and task-invariant transformations on graphs, and introduces a decoupling loss to separate size information from task information in graph representations, thereby improving the generalization ability of GNNs on larger graphs.
Enhancing Storage and Computational Efficiency in Federated Multimodal Learning for Large-Scale Models
Zixin Zhang (Tianjin University of Technology), Changsheng Xu (Chinese Academy of Sciences)
Federated LearningComputational EfficiencyKnowledge DistillationTransformerSupervised Fine-TuningMultimodality
🎯 What it does: The M2 FEDSA framework is proposed for the efficient deployment of large multimodal models in federated learning, addressing client storage and computation limitations.
Enhancing Sufficient Dimension Reduction via Hellinger Correlation
SeungBeom Hong, Jun Song (Korea University)
Tabular
🎯 What it does: A sufficient dimension reduction (SDR) method based on Hellinger correlation is proposed and applied to single-index models; this method can serve as an enhancement to existing SDR algorithms (SIR, SAVE, DR, MAVE, etc.).
Enhancing Trajectory Prediction through Self-Supervised Waypoint Distortion Prediction
Pranav singh chib, Pravendra Singh (Indian Institute of Technology)
Autonomous DrivingRepresentation LearningContrastive LearningTime Series
🎯 What it does: A trajectory prediction framework SSWDP based on self-supervised waypoint distortion prediction is proposed, which enhances the model's representation learning using clean and distorted views.
Enhancing Value Function Estimation through First-Order State-Action Dynamics in Offline Reinforcement Learning
Yun-Hsuan Lien (National Yang Ming Chiao Tung University), Yu-Shuen Wang (National Yang Ming Chiao Tung University)
Reinforcement LearningTabularBenchmark
🎯 What it does: Improved value function estimation in offline reinforcement learning, combining the discrete Bellman equation with the gradient consistency constraint of the continuous HJB equation.
Enhancing Vision Transformer: Amplifying Non-Linearity in Feedforward Network Module
Yixing Xu (Advanced Micro Devices), Emad Barsoum (Advanced Micro Devices)
ClassificationObject DetectionSegmentationTransformerSupervised Fine-TuningImage
🎯 What it does: An improved FFN module in Vision Transformer is proposed, utilizing arbitrary GeLU and spatial depth convolution to enhance non-linearity, thereby reducing hidden dimensions, FLOPs, and parameters while maintaining or even improving classification accuracy.
Ensemble Pruning for Out-of-distribution Generalization
Fengchun Qiao (University of Delaware), Xi Peng (University of Delaware)
Domain AdaptationOptimizationImage
🎯 What it does: A model predictive relationship topology-based integrated pruning framework is proposed, which enhances out-of-distribution generalization performance by selecting diverse subsets in the absence of target labels.
Entropy-Reinforced Planning with Large Language Models for Drug Discovery
Xuefeng Liu (University of Chicago), Rick L. Stevens (Argonne National Laboratory)
Drug DiscoveryTransformerLarge Language ModelReinforcement LearningBiomedical Data
🎯 What it does: A Transformer decoding algorithm based on entropy reinforcement planning (ERP) is proposed for drug molecule generation.
Environment Design for Inverse Reinforcement Learning
Thomas Kleine Buening (Alan Turing Institute), Christos Dimitrakakis (University of Neuchâtel)
Robotic IntelligenceReinforcement Learning from Human FeedbackReinforcement LearningAgentic AISequential
🎯 What it does: A proactive environment design framework is proposed, which quickly learns unknown reward functions by collecting expert trajectories in a series of different demonstration environments.
Envisioning Outlier Exposure by Large Language Models for Out-of-Distribution Detection
Chentao Cao (Hong Kong Baptist University), Bo Han (Hong Kong Baptist University)
Anomaly DetectionTransformerLarge Language ModelPrompt EngineeringVision Language ModelImage
🎯 What it does: Using large language models to generate potential anomaly categories and combining them with CLIP to achieve zero-shot OOD detection.
EquiAV: Leveraging Equivariance for Audio-Visual Contrastive Learning
Jongsuk Kim (Korea Advanced Institute of Science and Technology), Joon Son Chung (Korea Advanced Institute of Science and Technology)
RetrievalRepresentation LearningTransformerContrastive LearningVideoMultimodalityAudio
🎯 What it does: Proposes the EquiAV framework, which combines equivariance and self-supervised contrastive learning for joint representation learning of audio and visual modalities.
Equilibrium of Data Markets with Externality
Safwan Hossain (Harvard University), Yiling Chen (Harvard University)
Tabular
🎯 What it does: This study investigates the impact of buyer externalities on game equilibrium and social welfare in fixed-price data markets, and proposes a transaction cost intervention scheme based on predicted externalities.
EquiPocket: an E(3)-Equivariant Geometric Graph Neural Network for Ligand Binding Site Prediction
yang zhang, Wenbing Huang (Renmin University of China)
Drug DiscoveryProtein Structure PredictionGraph Neural NetworkGraphBiomedical Data
🎯 What it does: Designed and implemented a geometry-based E(3)-equivariant geometric neural network called EquiPocket to predict ligand binding sites of proteins.
Equivariance via Minimal Frame Averaging for More Symmetries and Efficiency
Yuchao Lin (Texas A&M University), Shuiwang Ji (Texas A&M University)
Computational EfficiencyGraph Neural NetworkGraph
🎯 What it does: A Minimal Frame Averaging (MFA) framework is proposed, which achieves precise equivariance through the construction of a minimal frame with a single forward inference.
Equivariant Deep Weight Space Alignment
Aviv Navon (Bar Ilan University), Haggai Maron (Technion)
OptimizationFederated LearningConvolutional Neural NetworkImage
🎯 What it does: DEEP-ALIGN is proposed, a model that utilizes equivariant networks to learn weight alignment, providing high-quality alignment after a single forward pass.
Equivariant Diffusion for Crystal Structure Prediction
Peijia Lin (Sun Yat-sen University), Yutong Lu (Sun Yat-sen University)
GenerationData SynthesisOptimizationDrug DiscoveryGraph Neural NetworkDiffusion modelScore-based ModelGraph
🎯 What it does: Developed an equivariant diffusion model that maintains invariance under transformations such as lattice permutation and periodic translation in crystal structure prediction.
Equivariant Frames and the Impossibility of Continuous Canonicalization
Nadav Dym (Technion), Jonathan W. Siegel
ClassificationPoint Cloud
🎯 What it does: This study investigates the relationship between frame averaging and equivariant networks, proving that commonly used unweighted frames often lead to continuity destruction. It proposes weighted and weakly equivariant robust frames, constructing robust frames that can maintain continuity under three common group actions (S_n, SO_d, O_d).
Equivariant Graph Neural Operator for Modeling 3D Dynamics
Minkai Xu (Stanford University), Anima Anandkumar (California Institute of Technology)
Graph Neural NetworkGraphTime SeriesPhysics Related
🎯 What it does: A new Equivariant Graph Neural Operator (EGNO) is proposed, which simulates three-dimensional dynamics by learning SE(3) equivariant time trajectory mappings, replacing traditional GNN methods that only predict the next state.
Erasing the Bias: Fine-Tuning Foundation Models for Semi-Supervised Learning
Kai Gan (Southeast University), Tong Wei (Southeast University)
ClassificationRecognitionTransformerSupervised Fine-TuningImage
🎯 What it does: A new semi-supervised learning method called FINESSL is proposed, which improves performance in low-label scenarios by fine-tuning a pre-trained base model using PEFT.
ERQ: Error Reduction for Post-Training Quantization of Vision Transformers
Yunshan Zhong (Xiamen University), Rongrong Ji (Xiamen University)
ClassificationObject DetectionSegmentationTransformerImage
🎯 What it does: A two-step post-training quantization (PTQ) framework called ERQ is proposed, which significantly reduces the quantization error of ViT and improves model accuracy by first using Aqer (Ridge regression optimized activation error) and then Wqer (layer-wise quantization + Rounding Refinement + Ridge regression optimized weight error).
Error Feedback Can Accurately Compress Preconditioners
Ionut-Vlad Modoranu (Institute of Science and Technology Austria), Dan Alistarh (Institute of Science and Technology Austria)
CompressionOptimizationConvolutional Neural NetworkTransformerSupervised Fine-TuningImageText
🎯 What it does: This paper proposes a method that utilizes an Error Feedback mechanism to compress the gradient history window of full matrix preprocessors (such as M-FAC and GGT), significantly reducing memory consumption without affecting convergence.
ESM All-Atom: Multi-Scale Protein Language Model for Unified Molecular Modeling
Kangjie Zheng (Peking University), Hao Zhou (Tsinghua University)
Drug DiscoveryProtein Structure PredictionTransformerLarge Language ModelSupervised Fine-TuningBiomedical Data
🎯 What it does: A multi-scale protein language model ESM-AA is proposed, capable of unified modeling at both residue and atomic scales, suitable for tasks involving protein-small molecule interactions.
ESNet: Evolution and Succession Network for High-Resolution Salient Object Detection
Hongyu Liu (Beijing Jiaotong University), Wei Zhang (Shandong University)
Object DetectionConvolutional Neural NetworkImage
🎯 What it does: A two-stage high-resolution salient object detection network ESNet is proposed, which first achieves detail-preserving saliency localization on low-resolution images, and then performs lightweight detail refinement on high-resolution images.
Estimating Barycenters of Distributions with Neural Optimal Transport
Alexander Kolesov (Skolkovo Institute of Science and Technology), Alexander Korotin (Artificial Intelligence Research Institute)
GenerationOptimizationGenerative Adversarial NetworkImage
🎯 What it does: A dual-layer adversarial optimization framework based on neural networks is proposed for computing the Wasserstein barycenter with general weak cost functions, and an error upper bound is provided.
Estimating Canopy Height at Scale
Jan Pauls (University of Münster), Fabian Gieseke (University of Copenhagen)
SegmentationGenerationOptimizationConvolutional Neural NetworkImageAgriculture Related
🎯 What it does: Based on Sentinel-1/2 satellite imagery and GEDI LiDAR measurements, a fully convolutional network was trained using U-Net+ResNet50 to generate global canopy height maps with a resolution of 10 meters.
Estimating Distributional Treatment Effects in Randomized Experiments: Machine Learning for Variance Reduction
Undral Byambadalai (CyberAgent), Shota Yasui (CyberAgent)
Supervised Fine-TuningTabular
🎯 What it does: A regression adjustment-based estimation method is proposed for estimating distributional treatment effects (DTE, PTE, QTE) in randomized experiments and achieving variance reduction.
Estimating the Permanent by Nesting Importance Sampling
Juha Harviainen (University of Helsinki), Mikko Koivisto (University of Helsinki)
OptimizationGraph
🎯 What it does: The NIS method utilizing nested upper bounds for importance sampling is proposed for high-dimensional integration and counting problems.
Estimating Unknown Population Sizes Using the Hypergeometric Distribution
Liam Hodgson (McGill University), Danilo Bzdok (Mila Quebec Artificial Intelligence Institute)
TextTabular
🎯 What it does: Using the likelihood function of the hypergeometric distribution combined with the variational autoencoder framework to estimate the distribution of unknown population and category sizes under low sampling conditions.
Et Tu Certifications: Robustness Certificates Yield Better Adversarial Examples
Andrew Craig Cullen, Benjamin I. P. Rubinstein
Adversarial AttackImage
🎯 What it does: A framework for generating smaller adversarial examples using model robustness certificates (Certification Aware Attack) is proposed, and it is proven to significantly reduce the perturbation magnitude across various datasets.
ETHER: Efficient Finetuning of Large-Scale Models with Hyperplane Reflections
Massimo Bini (University of Tuebingen), Anna Khoreva (Bosch Center for Artificial Intelligence)
GenerationData SynthesisOptimizationTransformerSupervised Fine-TuningDiffusion modelContrastive LearningImageText
🎯 What it does: A parameter-efficient fine-tuning method based on hyperplane reflection (Householder transformation) called ETHER and its relaxed version ETHER+ is proposed, which can efficiently adapt large pre-trained models without altering the model's pre-trained knowledge.
Eureka-Moments in Transformers: Multi-Step Tasks Reveal Softmax Induced Optimization Problems
David T Hoffmann, Thomas Brox (University of Freiburg)
OptimizationTransformerImageText
🎯 What it does: This study investigates the 'Eureka moments' that occur in multi-step tasks with Transformers, revealing the gradient imbalance problem caused by Softmax, and proposes improvements such as temperature scheduling and NormSoftmax.
Evaluating and Analyzing Relationship Hallucinations in Large Vision-Language Models
Mingrui Wu (Xiamen University), Rongrong Ji (Xiamen University)
RecognitionObject DetectionSegmentationTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelImageTextMultimodalityBenchmark
🎯 What it does: A benchmark for relationship hallucination in large visual language models (LVLM) called R-Bench is proposed and constructed, and various LVLMs are evaluated and analyzed on this benchmark.
Evaluating Model Bias Requires Characterizing its Mistakes
Isabela Albuquerque (Google DeepMind), Olivia Wiles (Google DeepMind)
ClassificationData SynthesisVision Language ModelImageMultimodality
🎯 What it does: A new bias measurement method called SKEWSIZE is proposed and validated to measure the differences in error distribution among different subgroups in classification models.
Evaluating Quantized Large Language Models
Shiyao Li (Tsinghua University), Yu Wang (Tsinghua University)
OptimizationTransformerLarge Language ModelTextBenchmark
🎯 What it does: A systematic evaluation of post-training quantization (PTQ) was conducted on large language models ranging from 125M to 180B across 11 major categories, examining the quantization effects of different tensor types (weights, activations, KV cache) on five task types (basic NLP, emergent abilities, reliability, dialogue, long text).
Evaluation of LLMs on Syntax-Aware Code Fill-in-the-Middle Tasks
Linyuan Gong (University of California), Alvin Cheung (University of California)
AI Code AssistantTransformerLarge Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: A new multilingual, executable grammar-aware Fill-in-the-Middle (FIM) benchmark SAFIM is proposed to evaluate the capabilities of large language models in code completion.
Evaluation of Test-Time Adaptation Under Computational Time Constraints
Motasem Alfarra (King Abdullah University of Science and Technology), Bernard Ghanem (King Abdullah University of Science and Technology)
Domain AdaptationComputational EfficiencyConvolutional Neural NetworkTransformerSupervised Fine-TuningImage
🎯 What it does: This paper proposes an online evaluation protocol that considers inference speed, penalizing the number of sample adaptations for test-time adaptation (TTA) methods in real-time streams.
Evaluation of Trajectory Distribution Predictions with Energy Score
Novin Shahroudi (University of Tartu), Meelis Kull (University of Tartu)
Time SeriesSequential
🎯 What it does: This paper conducts a theoretical analysis and experimental validation of evaluation metrics for trajectory distribution prediction, proposing a rigorous and effective evaluation method based on the Energy Score.
EVEREST: Efficient Masked Video Autoencoder by Removing Redundant Spatiotemporal Tokens
Sunil Hwang (Korea Military Academy), Sung Ju Hwang (KAIST)
Computational EfficiencyRepresentation LearningTransformerAuto EncoderVideo
🎯 What it does: This paper proposes an efficient Masked Video Autoencoder—EVEREST, which significantly reduces the computational and memory consumption of video self-supervised pre-training by utilizing non-parametric redundancy robust token selection and adaptive frame sampling.
EvGGS: A Collaborative Learning Framework for Event-based Generalizable Gaussian Splatting
Jiaxu Wang (Hong Kong University of Science and Technology), Renjing Xu (Hong Kong University of Science and Technology)
RestorationData SynthesisDepth EstimationConvolutional Neural NetworkGaussian SplattingPoint Cloud
🎯 What it does: Proposes the EvGGS framework, achieving 3D Gaussian reconstruction that can be generalized to unknown scenarios from raw event streams.
EvIL: Evolution Strategies for Generalisable Imitation Learning
Silvia Sapora (University of Oxford), Jakob Nicolaus Foerster (University of Oxford)
OptimizationRobotic IntelligenceReinforcement LearningSequential
🎯 What it does: This paper proposes an improved method that combines inverse reinforcement learning and evolutionary strategies, called EvIL, to enhance the effectiveness and efficiency of retraining policies from reward functions in new environments.
EvoluNet: Advancing Dynamic Non-IID Transfer Learning on Graphs
Haohui Wang (Virginia Tech), Dawei Zhou (Virginia Tech)
Domain AdaptationRepresentation LearningGraph Neural NetworkTransformerGraphTime Series
🎯 What it does: This paper studies the problem of transfer learning on dynamic non-IID graphs and proposes the EVOLUNET framework to achieve knowledge transfer from source graphs to target graphs.
Evolution of Heuristics: Towards Efficient Automatic Algorithm Design Using Large Language Model
Fei Liu (City University of Hong Kong), Qingfu Zhang (City University of Hong Kong)
OptimizationTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: A framework called Evolution of Heuristics (EoH) is proposed, which utilizes large language models (LLM) and evolutionary computation (EC) to co-evolve the natural language thoughts and code implementations of heuristic algorithms for automated heuristic design.
Evolution-Inspired Loss Functions for Protein Representation Learning
Chengyue Gong (University of Texas at Austin), Daniel Jesus Diaz
Representation LearningProtein Structure PredictionGraph Neural NetworkTransformerSupervised Fine-TuningBiomedical Data
🎯 What it does: In protein representation learning, the authors propose a new self-supervised training objective—EvoRank, which utilizes the amino acid distribution in multiple sequence alignments (MSA) for ranking learning, aiming to enhance the zero-shot prediction capability of protein variant effects.
Evolving Subnetwork Training for Large Language Models
Hanqi Li (Shanghai Jiao Tong University), Kai Yu (Shanghai Jiao Tong University)
TransformerLarge Language ModelText
🎯 What it does: A new training paradigm called Evolving Subnetwork Training (EST) is proposed, which reduces training costs by gradually expanding the subnetwork extracted from large language models during the training process.
EvoRainbow: Combining Improvements in Evolutionary Reinforcement Learning for Policy Search
Pengyi Li (Tianjin University), Jianye HAO
Robotic IntelligenceReinforcement LearningTabularBenchmark
🎯 What it does: This paper studies how to combine evolutionary algorithms with reinforcement learning for efficient policy search;
EvTexture: Event-driven Texture Enhancement for Video Super-Resolution
Dachun Kai (University of Science and Technology of China), Xiaoyan Sun (University of Science and Technology of China)
RestorationSuper ResolutionConvolutional Neural NetworkRecurrent Neural NetworkOptical FlowVideo
🎯 What it does: Designed and implemented a multi-branch video super-resolution network EvTexture based on event cameras, specifically utilizing high-frequency event information to restore video texture details.
Exact Conversion of In-Context Learning to Model Weights in Linearized-Attention Transformers
Brian K Chen (National University of Singapore), Kenji Kawaguchi (National University of Singapore)
TransformerText
🎯 What it does: This paper proposes a method to permanently encode in-context learning (ICL) information into model weights by adding bias terms to linearized attention Transformers.
Exact Soft Analytical Side-Channel Attacks using Tractable Circuits
Thomas Wedenig (Graz University of Technology), Robert Peharz (Graz University of Technology)
OptimizationSafty and PrivacyTime Series
🎯 What it does: This paper proposes an exact side-channel attack algorithm based on compileable circuits (SDD/PSDD) called ExSASCA, which replaces traditional loop belief propagation to achieve precise inference of the first round of AES-128 encryption.
ExCP: Extreme LLM Checkpoint Compression via Weight-Momentum Joint Shrinking
Wenshuo Li (Huawei Noah's Ark Lab), Yunhe Wang (Huawei Noah's Ark Lab)
CompressionTransformerLarge Language ModelImageText
🎯 What it does: Proposes the ExCP framework, which performs extreme compression of checkpoints during the training process of LLMs, achieving nearly lossless compression using residuals, weight-momentum joint pruning, and non-uniform quantization.
Executable Code Actions Elicit Better LLM Agents
Xingyao Wang (University of Illinois Urbana-Champaign), Heng Ji (University of Illinois Urbana-Champaign)
AI Code AssistantTransformerLarge Language ModelSupervised Fine-TuningAgentic AIText
🎯 What it does: This paper proposes the use of executable Python code as an action interface for LLM agents (CodeAct), enabling multi-turn interaction and self-debugging.
Expand-and-Cluster: Parameter Recovery of Neural Networks
Flavio Martinelli (École Polytechnique Fédérale de Lausanne), Johanni Brea (École Polytechnique Fédérale de Lausanne)
Image
🎯 What it does: By probing the input-output mapping of the target network, a recognizable recovery of neural network parameters and structure was achieved using the expand-and-cluster method.
Expert Proximity as Surrogate Rewards for Single Demonstration Imitation Learning
Chia-Cheng Chiang (National Tsing Hua University), Chun-Yi Lee (National Tsing Hua University)
Robotic IntelligenceReinforcement LearningSequential
🎯 What it does: In single demonstration imitation learning, the authors propose the TDIL method, which guides the agent to converge faster by constructing a dense reward function based on a transition discriminator.
Experts Don't Cheat: Learning What You Don't Know By Predicting Pairs
Daniel D. Johnson (Google DeepMind), Chris J. Maddison (University of Toronto)
ClassificationReinforcement LearningImageSequential
🎯 What it does: This paper proposes a second-order calibration model by allowing the model to predict two independent answers (pairs) for the same input during training and permitting the model to 'cheat' by observing one of the answers, thereby estimating the model's own missing information about the true distribution.
Explain Temporal Black-Box Models via Functional Decomposition
Linxiao Yang (Alibaba Group), Liang Sun (Alibaba Group)
Anomaly DetectionExplainability and InterpretabilityAuto EncoderTime SeriesElectronic Health Records
🎯 What it does: This paper proposes FDTempExplainer, a model-agnostic time series explanation method based on functional decomposition, which can simultaneously quantify the main effects and pure interaction effects at each moment, and fairly allocate the interaction effects to the participating moments.
Explaining Graph Neural Networks via Structure-aware Interaction Index
Ngoc Bui (Yale University), Rex Ying (Yale University)
Explainability and InterpretabilityGraph Neural NetworkGraph
🎯 What it does: The Myerson-Taylor interaction index is proposed to provide importance scores for higher-order interactions in graph neural networks while maintaining graph structure constraints, and based on this index, the MAGE (Myerson-Taylor Structure-Aware Graph Explainer) interpreter is designed.
Explaining Probabilistic Models with Distributional Values
Luca Franceschi (Amazon Web Services), Matthias Seeger (Helsing)
Explainability and InterpretabilityConvolutional Neural NetworkTransformerImageTextTabular
🎯 What it does: A distributed values method is proposed, extending traditional game theory XAI to explain the distributed outputs of probabilistic models rather than just scalar results.
Exploiting Code Symmetries for Learning Program Semantics
Kexin Pei (University of Chicago), Suman Jana (Columbia University)
OptimizationAI Code AssistantGraph Neural NetworkTransformerGraph
🎯 What it does: This paper proposes a framework based on code symmetry and implements the SYMC model, utilizing equivariant layers and invariant predictions with self-attention to ensure consistent predictions under semantically invariant code transformations (such as reordering, optimization, and obfuscation).