ICLR 2024 Papers — Page 7
International Conference on Learning Representations · 2260 papers
Efficient Backpropagation with Variance Controlled Adaptive Sampling
Ziteng Wang (Tsinghua University), Jun Zhu (Tsinghua University)
OptimizationComputational EfficiencyTransformerSupervised Fine-TuningImageText
🎯 What it does: An adaptive sampling algorithm based on variational control (VCAS) is designed, which generates unbiased approximate gradients during backpropagation using fine-grained data and token-level sampling, significantly reducing training costs.
Efficient Continual Finite-Sum Minimization
Ioannis Mavrothalassitis (École Polytechnique Fédérale de Lausanne), Volkan Cevher (École Polytechnique Fédérale de Lausanne)
OptimizationTabularStochastic Differential Equation
🎯 What it does: A new continuous finite sum minimization method called CSVRG is proposed, which can achieve optimization with approximately O(n/ε^{1/3}) gradient queries while ensuring that each prefix sum meets ε accuracy.
Efficient ConvBN Blocks for Transfer Learning and Beyond
Kaichao You (Tsinghua University), Mingsheng Long (Tsinghua University)
ClassificationObject DetectionComputational EfficiencyConvolutional Neural NetworkImage
🎯 What it does: This paper proposes the Tune mode, which dynamically combines convolution and BatchNorm, maintaining the training stability of the Eval mode while approximating the computational efficiency of the Deploy mode, specifically optimizing the ConvBN block for transfer learning scenarios.
Efficient Dynamics Modeling in Interactive Environments with Koopman Theory
Arnab Kumar Mondal (Mila), Siamak Ravanbakhsh (Mila)
OptimizationComputational EfficiencyRecurrent Neural NetworkTransformerReinforcement LearningAuto EncoderTime SeriesSequential
🎯 What it does: A linear latent dynamic model based on Koopman theory is proposed for efficient long-range prediction and modeling of interactive environment dynamics, and it is integrated into model predictive control and model-free reinforcement learning.
Efficient Episodic Memory Utilization of Cooperative Multi-Agent Reinforcement Learning
Hyungho Na (Korea Advanced Institute of Science and Technology), Il-chul Moon
Reinforcement LearningSequential
🎯 What it does: This paper proposes an efficient episodic memory utilization framework EMU for collaborative multi-agent reinforcement learning, aimed at accelerating learning and avoiding local optima.
Efficient Heterogeneous Meta-Learning via Channel Shuffling Modulation
Minh Hoang (Carnegie Mellon University), Carl Kingsford (Carnegie Mellon University)
Meta LearningConvolutional Neural NetworkImageMultimodality
🎯 What it does: This paper proposes a heterogeneous meta-learning method under multi-modal task distributions, utilizing differentiable channel reordering routing to achieve task-specific network configurations.
Efficient Integrators for Diffusion Generative Models
Kushagra Pandey (University of California), Stephan Mandt (University of California)
GenerationComputational EfficiencyDiffusion modelImageStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: Efficient deterministic and stochastic sampling methods have been designed for diffusion generative models (especially the enhanced PSLD model), significantly reducing the number of NFE (network function evaluations) and improving sampling speed.
Efficient Inverse Multiagent Learning
Denizalp Goktas (Brown University), Sumitra Ganesh (JP Morgan Chase)
OptimizationReinforcement LearningTime SeriesFinance Related
🎯 What it does: This paper proposes a framework for formulating the inverse game (inverse multi-agent learning) problem as a zero-sum min-max optimization problem, and provides methods for solving inverse Nash equilibria and first-order, second-order simulacrum learning in polynomial time under scenarios such as convex-concave games and finite/continuous Markov games.
Efficient local linearity regularization to overcome catastrophic overfitting
Elias Abad Rocamora (École Polytechnique Fédérale de Lausanne), Volkan Cevher (École Polytechnique Fédérale de Lausanne)
OptimizationComputational EfficiencyAdversarial AttackConvolutional Neural NetworkImage
🎯 What it does: This paper proposes an efficient local linear regularization method called ELLE, combined with adaptive regularization ELLE-A, to address the problem of catastrophic overfitting in single-step adversarial training.
Efficient Modulation for Vision Networks
Xu Ma (Northeastern University), Lu Yuan (Microsoft)
ClassificationObject DetectionSegmentationConvolutional Neural NetworkImage
🎯 What it does: An efficient modulation module named EfficientMod is proposed, and a new convolutional network architecture is built based on this module.
Efficient Multi-agent Reinforcement Learning by Planning
Qihan Liu (Tsinghua University), Chongjie Zhang (Washington University in St. Louis)
Reinforcement LearningBenchmark
🎯 What it does: This paper proposes MAZero, a model-based multi-agent reinforcement learning algorithm that combines the ideas of MuZero and achieves significant sample efficiency improvements on the SMAC benchmark.
Efficient Planning with Latent Diffusion
Wenhao Li (Tongji University)
Robotic IntelligenceReinforcement LearningDiffusion modelAuto EncoderSequential
🎯 What it does: This paper studies an offline reinforcement learning framework based on a continuous latent diffusion model—LatentDiffuser, which achieves learning and planning in the latent action space.
Efficient Score Matching with Deep Equilibrium Layers
Yuhao Huang (University of Utah), Bao Wang (Lehigh University)
GenerationComputational EfficiencyScore-based ModelImageTabular
🎯 What it does: This paper proposes embedding Deep Equilibrium Networks (DEQ) into score matching models to address the memory and computational bottlenecks of traditional score matching in high-dimensional data and deep networks.
Efficient Sharpness-Aware Minimization for Molecular Graph Transformer Models
Yili Wang (Jilin University), Xin Wang (Jilin University)
OptimizationDrug DiscoveryGraph Neural NetworkTransformerGraphTabular
🎯 What it does: An efficient sharpness-aware minimization algorithm named GraphSAM is proposed to improve the generalization performance of unpretrained molecular graph Transformer models.
Efficient Streaming Language Models with Attention Sinks
Guangxuan Xiao (Massachusetts Institute of Technology), Mike Lewis (Meta AI)
TransformerLarge Language ModelText
🎯 What it does: The StreamingLLM framework is proposed, utilizing attention sink and sliding window KV cache to enable pre-trained LLMs to efficiently infer on infinitely long texts.
Efficient Subgraph GNNs by Learning Effective Selection Policies
Beatrice Bevilacqua (Purdue University), Haggai Maron (Technion and NVIDIA Research)
Computational EfficiencyRepresentation LearningGraph Neural NetworkGraph
🎯 What it does: This paper studies a framework called Policy-Learn, which is used to learn subgraph selection strategies to reduce the computational cost of subgraph GNNs while maintaining expressive power.
Efficient Video Diffusion Models via Content-Frame Motion-Latent Decomposition
Sihyun Yu (KAIST), Anima Anandkumar (California Institute of Technology)
GenerationData SynthesisComputational EfficiencyTransformerDiffusion modelAuto EncoderVideo
🎯 What it does: This paper proposes an efficient video diffusion model CMD that encodes videos into content frames and low-dimensional motion latent variables, achieving video generation through the refinement of a pre-trained image diffusion model and a lightweight motion diffusion model.
Efficient-3Dim: Learning a Generalizable Single-image Novel-view Synthesizer in One Day
Yifan Jiang (Apple), Liangliang Cao (Apple)
GenerationData SynthesisComputational EfficiencyTransformerDiffusion modelImage
🎯 What it does: This paper proposes a new single-image novel view synthesis framework named Efficient-3DiM, which can complete training within a day, significantly improving efficiency.
EfficientDM: Efficient Quantization-Aware Fine-Tuning of Low-Bit Diffusion Models
Yefei He (Zhejiang University), Bohan Zhuang (Monash University)
GenerationData SynthesisOptimizationComputational EfficiencyKnowledge DistillationDiffusion modelImage
🎯 What it does: A data-independent, parameter-efficient low-bit diffusion model fine-tuning framework called EfficientDM is proposed, utilizing Quantization-Aware Low-Rank Adapters (QALoRA) and noise distillation to achieve model quantization.
Efficiently Computing Similarities to Private Datasets
Arturs Backurs (Microsoft Research), Jakub Tarnawski (Microsoft Research)
Safty and PrivacyComputational EfficiencyImage
🎯 What it does: A general differential privacy data structure is proposed for efficiently computing the sum of any similarity function (kernel function or distance function) with private datasets.
Elastic Feature Consolidation For Cold Start Exemplar-Free Incremental Learning
Simone Magistri (University of Florence), Andrew D. Bagdanov (University of Florence)
ClassificationKnowledge DistillationRepresentation LearningConvolutional Neural NetworkImage
🎯 What it does: This paper proposes the Elastic Feature Consolidation (EFC) method, which utilizes the Empirical Feature Matrix (EFM) to impose elastic constraints on feature drift, and combines it with the asymmetric prototype replay loss (PR-ACE) to enhance the model's plasticity and stability in sample-free incremental learning.
Elucidating the design space of classifier-guided diffusion generation
Jiajun Ma (Hong Kong University of Science and Technology), Jiacheng Sun (Huawei Noah's Ark Lab)
GenerationData SynthesisDiffusion modelImage
🎯 What it does: In conditional generation with diffusion models, this study explores how to utilize existing pre-trained classifiers for training-independent guidance to enhance sample quality and controllability.
Elucidating the Exposure Bias in Diffusion Models
Mang Ning (Utrecht University), Itir Onal Ertugrul (Utrecht University)
GenerationData SynthesisDiffusion modelImageOrdinary Differential Equation
🎯 What it does: Research and address the exposure bias issue in the sampling phase of diffusion models, proposing a no-training, plug-in Epsilon Scaling method to improve sampling trajectories.
Embarrassingly Simple Dataset Distillation
Yunzhen Feng (New York University), Julia Kempe (New York University)
Data SynthesisOptimizationKnowledge DistillationRecurrent Neural NetworkReinforcement LearningImage
🎯 What it does: Through a dual-layer optimization framework, large datasets are compressed to generate a small number of synthetic samples that achieve comparable test performance to the original dataset.
Embodied Active Defense: Leveraging Recurrent Feedback to Counter Adversarial Patches
Lingxuan Wu (Tsinghua University), Jun Zhu (Tsinghua University)
Object DetectionAdversarial AttackRobotic IntelligenceRecurrent Neural NetworkTransformerReinforcement LearningAgentic AIImage
🎯 What it does: This paper proposes an Embodied Active Defense (EAD) framework based on active perception, which actively collects multi-view information through a repetitive perception and action decision loop to eliminate the interference of adversarial patches in 3D physical environments.
EmerDiff: Emerging Pixel-level Semantic Knowledge in Diffusion Models
Koichi Namekata (University of Toronto), Seung Wook Kim (University of Toronto)
SegmentationGenerationDiffusion modelImage
🎯 What it does: Utilizing a pre-trained diffusion model (Stable Diffusion) to directly generate high-resolution fine-grained semantic segmentation maps without any additional training;
Emergent Communication with Conversational Repair
Mitja Nikolaus (Centre de recherche en cognitive)
Recurrent Neural NetworkReinforcement LearningImage
🎯 What it does: In the Lewis signaling game, noise and a bidirectional feedback channel are introduced, training RNN senders and receivers so that the receiver can send back clarification or confirmation information in the presence of noise, thereby studying the impact of conversational repair on language emergence.
Emergent mechanisms for long timescales depend on training curriculum and affect performance in memory tasks
Sina Khajehabdollahi (University of Tbingen), Anna Levina (Max Planck Institute for Biological Cybernetics)
Recurrent Neural NetworkSequential
🎯 What it does: This paper studies the learning mechanisms and performance of different curricula (single-head vs. multi-head) in N-parity and N-delay matching tasks by training an RNN with a learnable single-neuron time constant τ and recurrent weights.
EmerNeRF: Emergent Spatial-Temporal Scene Decomposition via Self-Supervision
Jiawei Yang (University of Southern California), Yue Wang (University of Southern California)
Autonomous DrivingTransformerNeural Radiance FieldVideoPoint CloudBenchmark
🎯 What it does: We propose EmerNeRF, a self-supervised 4D neural field framework for simultaneously reconstructing static and dynamic geometry, appearance, and motion in autonomous driving scenarios, enhancing semantic perception.
EMO: EARTH MOVER DISTANCE OPTIMIZATION FOR AUTO-REGRESSIVE LANGUAGE MODELING
Siyu Ren (Shanghai Jiao Tong University), Kenny Q. Zhu (University of Texas at Arlington)
GenerationOptimizationTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: An automatic regression language model optimization method based on the upper bound of Earth Mover's Distance (EMD) called EMO is proposed, achieving distribution calibration of the pre-trained model during the lightweight fine-tuning phase.
Empirical Analysis of Model Selection for Heterogeneous Causal Effect Estimation
Divyat Mahajan (Mila), Vasilis Syrgkanis (Stanford University)
OptimizationMeta LearningTabular
🎯 What it does: This paper focuses on the selection of Conditional Average Treatment Effect (CATE) models, systematically evaluating 34 proxy indicators, and proposes a new strategy based on AutoML hyperparameter tuning, two-level selection, and causal integration.
Empirical Likelihood for Fair Classification
Pangpang Liu (Purdue University), Yichuan Zhao (Georgia State University)
ClassificationSupervised Fine-TuningTabularFinance Related
🎯 What it does: This paper proposes a method to construct confidence regions for the covariance of sensitive attributes and decision boundaries using empirical likelihood, thereby considering estimation uncertainty in fair classification and incorporating it as a fairness constraint.
Emu: Generative Pretraining in Multimodality
Quan Sun (Beijing Academy of Artificial Intelligence), Xinlong Wang (Beijing Academy of Artificial Intelligence)
GenerationTransformerLarge Language ModelVision Language ModelDiffusion modelImageVideoTextMultimodality
🎯 What it does: A multimodal generative pre-training model called Emu is proposed, which can simultaneously handle images, text, and videos, achieving multi-tasking such as image generation, image description, and question answering.
Enabling Efficient Equivariant Operations in the Fourier Basis via Gaunt Tensor Products
Shengjie Luo (Peking University), Aditi S. Krishnapriyan (University of California)
Computational EfficiencyTabularPhysics Related
🎯 What it does: This paper proposes the Gaunt Tensor Product, which utilizes Gaunt coefficients to transform E(3) equivariant tensor products into spherical function products, significantly reducing computational complexity through two-dimensional Fourier transforms and FFT acceleration.
Enabling Lanuguage Models to Implicitly Learn Self-Improvement
Ziqi Wang (University of Illinois Urbana-Champaign), Heng Ji (Google)
Reinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningText
🎯 What it does: A framework (PIT) is proposed that allows large language models to improve their output quality through implicit learning without the need for manually defined evaluation criteria, and achieves self-improvement by comparing with original responses.
Encoding Unitig-level Assembly Graphs with Heterophilous Constraints for Metagenomic Contigs Binning
Hansheng Xue (Australian National University), Vaibhav Rajan (National University of Singapore)
Graph Neural NetworkAuto EncoderGraph
🎯 What it does: A metagenomic contig binning tool named UNITIGBIN has been developed, which utilizes the heterogeneity constraint of single-copy marker genes to perform graph neural network learning and binning directly on the unitig-level assembly graph.
End-to-End (Instance)-Image Goal Navigation through Correspondence as an Emergent Phenomenon
Guillaume Bono (Naver Labs Europe), Christian Wolf (Naver Labs Europe)
Robotic IntelligenceTransformerReinforcement LearningImage
🎯 What it does: For the visual target navigation tasks of ImageNav and Instance-ImageNav, a dual-encoder (dual-view ViT) structure is proposed, and a two-step pre-training (cross-view completion CroCo and relative pose + visibility estimation RPEV) is conducted to provide priors for the perception module, followed by training an integrated end-to-end navigation policy in RL.
Energy-based Automated Model Evaluation
Ru Peng (Zhejiang University), Junbo Zhao (Zhejiang University)
ClassificationDomain AdaptationImageText
🎯 What it does: This paper researches and implements an automatic model evaluation (AutoEval) method that does not require labels or training, using the Meta-Distribution Energy (MDE) metric constructed from energy models to predict the accuracy of models on unlabeled OOD datasets.
Energy-Based Concept Bottleneck Models: Unifying Prediction, Concept Intervention, and Probabilistic Interpretations
Xinyue Xu (Hong Kong University of Science and Technology), Xiaomeng Li (Hong Kong University of Science and Technology)
ClassificationExplainability and InterpretabilityConvolutional Neural NetworkImage
🎯 What it does: Proposes the Energy-based Concept Bottleneck Model (ECBM), unifying concept prediction, concept intervention, and conditional explanation;
Energy-conserving equivariant GNN for elasticity of lattice architected metamaterials
Ivan Grega (University of Cambridge), Vikram Deshpande
Graph Neural NetworkMixture of ExpertsGraphPhysics Related
🎯 What it does: A SE(3) equivariant GNN model has been developed to predict the fourth-order elastic tensor of lattice materials while maintaining energy conservation, and a large-scale structure-property dataset is provided.
Energy-guided Entropic Neural Optimal Transport
Petr Mokrov (Skolkovo Institute of Science and Technology), Evgeny Burnaev (Artificial Intelligence Research Institute)
Image TranslationOptimizationGenerative Adversarial NetworkImage
🎯 What it does: Proposes an entropy-regularized optimal transport (EOT) solving method based on an energy model, which directly generates conditional transport plans using a latent network trained with an energy function;
Enhanced Face Recognition using Intra-class Incoherence Constraint
Yuanqing Huang (Ant Group), Lei Wang (Ant Group)
RecognitionKnowledge DistillationConvolutional Neural NetworkContrastive LearningImage
🎯 What it does: By performing orthogonal decomposition and recombination of facial feature space, innovative sub-features are introduced to enhance facial recognition accuracy.
Enhancing Contrastive Learning for Ordinal Regression via Ordinal Content Preserved Data Augmentation
Jiyang Zheng (University of Sydney), Tongliang Liu (University of Sydney)
ClassificationGenerationData SynthesisRepresentation LearningGenerative Adversarial NetworkContrastive LearningImage
🎯 What it does: A generative enhancement method based on the principle of minimal change is proposed, replacing traditional strong augmentation, preserving the key ordinal content in ordinal regression tasks, and enhancing model performance through contrastive learning.
Enhancing Group Fairness in Online Settings Using Oblique Decision Forests
Somnath Basu Roy Chowdhury (University of North Carolina at Chapel Hill), Snigdha Chaturvedi (Google Research)
ClassificationOptimizationComputational EfficiencyImageTextTabular
🎯 What it does: Proposes the Aranyani framework, which achieves group fair decision-making through oblique decision forests in an online learning environment.
Enhancing High-Resolution 3D Generation through Pixel-wise Gradient Clipping
Zijie Pan (Fudan University), Li Zhang (Fudan University)
GenerationData SynthesisDiffusion modelScore-based ModelImage
🎯 What it does: This paper proposes the Pixel-wise Gradient Clipping (PGC) technique, which enhances the texture quality of high-resolution 3D generative models by clipping gradients at the pixel level.
Enhancing Human Experience in Human-Agent Collaboration: A Human-Centered Modeling Approach Based on Positive Human Gain
Yiming Gao (Tencent AI Lab), Wei Liu (Tencent AI Lab)
Recurrent Neural NetworkReinforcement LearningSequential
🎯 What it does: A reinforcement learning method based on positive human gains (RLHG) is proposed, allowing cooperative agents to actively enhance the gaming experience of human players.
Enhancing Human-AI Collaboration Through Logic-Guided Reasoning
Chengzhi Cao (University of Science and Technology of China), Shuang Li (Chinese University of Hong Kong)
Robotic IntelligenceReinforcement Learning from Human FeedbackReinforcement LearningVideo
🎯 What it does: A framework based on logical rules and Theory of Mind (ToM) is proposed to enhance human-machine collaboration and cognitive reasoning.
Enhancing Instance-Level Image Classification with Set-Level Labels
Renyu Zhang (University of Chicago), Robert L. Grossman (University of Chicago)
ClassificationContrastive LearningImageBiomedical Data
🎯 What it does: A general framework (FACILE) is proposed to enhance instance-level image classification performance using set-level coarse labels.
Enhancing Neural Subset Selection: Integrating Background Information into Set Representations
Binghui Xie (Chinese University of Hong Kong), James Cheng (Chinese University of Hong Kong)
Recommendation SystemAnomaly DetectionDrug DiscoveryTransformerImageTabular
🎯 What it does: The INSET model is proposed for the neural subset selection task, which enhances the accuracy of subset selection by incorporating superset information into the subset representation to learn invariant sufficient statistics of the conditional distribution P(Y|S,V).
Enhancing Neural Training via a Correlated Dynamics Model
Jonathan Brokman (Technion Israel Institute of Technology), Guy Gilboa (Technion Israel Institute of Technology)
ClassificationSegmentationGenerationFederated LearningComputational EfficiencyConvolutional Neural NetworkTransformerGenerative Adversarial NetworkImage
🎯 What it does: A low-dimensional training dynamic modeling method based on weight trajectory correlation, called Correlation Mode Decomposition (CMD), is proposed, along with online and embedded variants, enabling real-time capture of network training dynamics without significant computational or storage overhead.
Enhancing One-Shot Federated Learning Through Data and Ensemble Co-Boosting
Rong Dai (University of Science and Technology of China), Bo Han (Hong Kong Baptist University)
Federated LearningKnowledge DistillationGenerative Adversarial NetworkImage
🎯 What it does: A consistency boosting method called Co‑Boosting is proposed, which enhances the performance of the server model in one-shot federated learning by generating high-quality hard samples and dynamically reweighting client models to form a stronger teacher ensemble.
Enhancing Small Medical Learners with Privacy-preserving Contextual Prompting
Xinlu Zhang (University of California), Linda Ruth Petzold (University of California)
OptimizationSafty and PrivacyTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextBiomedical Data
🎯 What it does: An end-to-end process is proposed to utilize large language models to generate medical context in privacy-constrained scenarios to enhance the performance of small medical language models.
Enhancing Tail Performance in Extreme Classifiers by Label Variance Reduction
Anirudh Buvanesh (Microsoft), Manik Varma (Microsoft)
ClassificationKnowledge DistillationText
🎯 What it does: This paper proposes the LEVER framework, which reduces label variance in extreme classifiers through knowledge distillation, thereby improving the prediction performance of tail labels.
Enhancing Transfer Learning with Flexible Nonparametric Posterior Sampling
Hyungi Lee (KAIST), Juho Lee (KAIST)
Domain AdaptationOptimizationSupervised Fine-TuningImageText
🎯 What it does: A non-parametric learning-based posterior sampling method NPTL is proposed to enhance the performance of Bayesian model averaging in transfer learning.
Enhancing Transferable Adversarial Attacks on Vision Transformers through Gradient Normalization Scaling and High-Frequency Adaptation
Zhiyu Zhu (University of Sydney), Huaming Chen (University of Sydney)
Adversarial AttackTransformerImage
🎯 What it does: For transferable adversarial attacks on Vision Transformers, two modules are proposed: Gradient Normalization Scaling (GNS) and High-Frequency Adaptation (HFA), which utilize fine-grained gradient normalization/scaling and frequency domain high-frequency masking to enhance the transferability of adversarial samples.
Ensemble Distillation for Unsupervised Constituency Parsing
Behzad Shayegh (University of Alberta), Lili Mou (University of Alberta)
Domain AdaptationComputational EfficiencyKnowledge DistillationRecurrent Neural NetworkSupervised Fine-TuningText
🎯 What it does: This study investigates ensemble and distillation methods for unsupervised constituent syntax parsing, generating an 'average' tree from the results of various existing unsupervised parsers, and then using this tree as a pseudo-label to train a student model to improve inference efficiency.
Entity-Centric Reinforcement Learning for Object Manipulation from Pixels
Dan Haramati (Technion Israel Institute of Technology), Aviv Tamar (Technion Israel Institute of Technology)
Robotic IntelligenceTransformerReinforcement LearningImage
🎯 What it does: This paper proposes an entity-centered reinforcement learning framework for multi-object visual reinforcement learning, capable of learning object manipulation under target conditions from pixel images.
Entropy Coding of Unordered Data Structures
Julius Kunze (University College London), James Townsend (University of Amsterdam)
CompressionGraph
🎯 What it does: This paper studies a general method for compressing unordered data structures—shuffle coding—used for optimal lossless compression of unordered objects such as graphs, sets, and multisets.
Entropy is not Enough for Test-Time Adaptation: From the Perspective of Disentangled Factors
Jonghyun Lee (Seoul National University), Sungroh Yoon (Seoul National University)
ClassificationDomain AdaptationImage
🎯 What it does: In response to the phenomenon where low-entropy samples during Test-Time Adaptation (TTA) may lead to errors, this paper proposes a new adaptive method called DeYO and introduces Pseudo Label Probability Difference (PLPD) as a confidence metric to filter harmful samples.
Entropy-MCMC: Sampling from Flat Basins with Ease
Bolian Li (Purdue University), Ruqi Zhang (Purdue University)
ClassificationOptimizationConvolutional Neural NetworkImageStochastic Differential Equation
🎯 What it does: A new MCMC method (Entropy-MCMC, EMCMC) is proposed, which introduces auxiliary guiding variables to bias sampling towards the flat valleys in the posterior distribution of deep networks, resulting in samples with better generalization ability.
Epitopological learning and Cannistraci-Hebb network shape intelligence brain-inspired theory for ultra-sparse advantage in deep learning
Yingtao Zhang (Tsinghua University), Carlo Vittorio Cannistraci (Tsinghua University)
ClassificationOptimizationConvolutional Neural NetworkImage
🎯 What it does: A brain-inspired network shape intelligence framework called Epitopological Sparse Meta-deep Learning (ESML) is proposed, which is extremely sparse (only 1% connectivity) and includes a four-step training method called Cannistraci-Hebb Training (CHT), achieving training with only 1% connectivity while surpassing the performance of fully connected networks.
EQA-MX: Embodied Question Answering using Multimodal Expression
Md Mofijul Islam (University of Virginia), Tariq Iqbal (McGill University)
Robotic IntelligenceTransformerVision Language ModelMultimodality
🎯 What it does: A large-scale multimodal embedded question-answering dataset EQA-MX has been constructed, and the VQ-Fusion model has been proposed to fuse multi-perspective visual and textual information, thereby completing eight novel embedded question-answering tasks.
EquiformerV2: Improved Equivariant Transformer for Scaling to Higher-Degree Representations
Yi-Lun Liao (Massachusetts Institute of Technology), Tess Smidt (Massachusetts Institute of Technology)
TransformerSupervised Fine-TuningPoint Cloud
🎯 What it does: Proposes EquiformerV2, an improved equivariant Transformer for predicting energy and forces in 3D atomic systems;
Equivariant Matrix Function Neural Networks
Ilyes Batatia (University of Cambridge), Felix Andreas Faber
Computational EfficiencyDrug DiscoveryGraph Neural NetworkGraph
🎯 What it does: A novel matrix function neural network (MFN) is proposed for efficiently modeling non-local many-body interactions in graph structures.
Equivariant Scalar Fields for Molecular Docking with Fast Fourier Transforms
Bowen Jing (Massachusetts Institute of Technology), Bonnie Berger (Massachusetts Institute of Technology)
OptimizationDrug DiscoveryGraph Neural NetworkGraph
🎯 What it does: A molecular docking scoring function based on equivariant scalar fields is proposed, and FFT is used to quickly optimize ligand poses.
Error Feedback Reloaded: From Quadratic to Arithmetic Mean of Smoothness Constants
Peter Richtárik (AI Initiative KAUST), Konstantin Pavlovich Burlachenko (AI Initiative KAUST)
OptimizationFederated LearningTabular
🎯 What it does: This paper theoretically improves the commonly used error feedback mechanism EF21 in distributed training, replacing the square root arithmetic mean (L_QM) with the square arithmetic mean (L_AM) of the smoothness parameter, thereby enhancing the convergence speed in data heterogeneous scenarios.
Error Norm Truncation: Robust Training in the Presence of Data Noise for Text Generation Models
Tianjian Li (Johns Hopkins University), Kenton Murray (Johns Hopkins University)
GenerationTransformerSupervised Fine-TuningText
🎯 What it does: This paper studies how to improve the robustness of text generation models to noisy data during training and proposes the Error Norm Truncation (ENT) method, which truncates noisy words based on the L2 error between the model's predicted distribution and a one-hot vector during the training process.
Escape Sky-high Cost: Early-stopping Self-Consistency for Multi-step Reasoning
Yiwei Li (Beijing Institute of Technology), Kan Li (Beijing Institute of Technology)
GenerationOptimizationComputational EfficiencyTransformerPrompt EngineeringTextChain-of-Thought
🎯 What it does: This paper proposes an Early Stopping Consistency (ESC) sampling strategy that significantly reduces the number of generated samples while retaining the performance of Self-Consistency (SC).
Estimating Conditional Mutual Information for Dynamic Feature Selection
Soham Gadgil (University of Washington), Su-In Lee (University of Washington)
TransformerReinforcement LearningImageTabularBiomedical DataAlzheimer's Disease
🎯 What it does: The DIME method is proposed, which estimates conditional mutual information without a generative model by using loss to improve regression targets through discriminative training of predictors and value networks, achieving dynamic feature selection.
Estimating Shape Distances on Neural Representations with Limited Samples
Dean A Pospisil, Alex H Williams
Biomedical Data
🎯 What it does: This study investigates the statistical properties of shape distance estimation represented by neural networks under high-dimensional and sample-limited conditions, and proposes a moment estimator that adjusts the bias-variance trade-off.
Eureka: Human-Level Reward Design via Coding Large Language Models
Yecheng Jason Ma (University of Pennsylvania), Anima Anandkumar (NVIDIA)
Robotic IntelligenceReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: EUREKA is proposed, a general reward design algorithm based on large language models (LLM) and evolutionary search, capable of automatically generating executable reward code without the need for task-specific prompts. It continuously improves through multiple rounds of evolution and reward reflection, ultimately achieving human expert-level reward performance and even making breakthrough progress in complex grasping and rotation tasks.
Evaluating Language Model Agency Through Negotiations
Tim Ruben Davidson, Robert West (École Polytechnique Fédérale de Lausanne)
TransformerLarge Language ModelAgentic AIText
🎯 What it does: This paper proposes and implements a dynamic evaluation framework based on structured negotiation games to simultaneously measure the agent capabilities, dialogue coherence, and alignment level of language models (LMs).
Evaluating Large Language Models at Evaluating Instruction Following
Zhiyuan Zeng (Tsinghua University), Danqi Chen (Princeton University)
TransformerLarge Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: A meta-evaluation benchmark specifically designed to assess the performance of large language model (LLM) evaluators in instruction following, called LLMBAR, has been created, and various LLM evaluators and prompting strategies have been systematically evaluated on this benchmark.
Evaluating Representation Learning on the Protein Structure Universe
Arian Rokkum Jamasb, Tom Leon Blundell
Representation LearningProtein Structure PredictionGraph Neural NetworkGraphBiomedical DataBenchmark
🎯 What it does: A ProteinWorkshop benchmark was constructed, unifying the implementation of large-scale pre-training corpora (AlphaFoldDB), various characterization schemes, geometric GNN architectures, and numerous downstream tasks (structure, function, and node-level predictions), and systematically evaluating the models.
Evaluating the Zero-shot Robustness of Instruction-tuned Language Models
Jiuding Sun, Byron C Wallace
TransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
🎯 What it does: This study investigates the robustness of instruction-tuned large language models in zero-shot scenarios regarding instruction rephrasing. A total of 319 manually crafted new instructions were collected to evaluate the model's sensitivity to unseen instructions, and a method was proposed to enhance robustness through soft prefixes and KL alignment loss.
EventRPG: Event Data Augmentation with Relevance Propagation Guidance
Mingyuan Sun (Hong Kong University of Science and Technology), Renjing Xu (Hong Kong University of Science and Technology)
ClassificationRecognitionSpiking Neural NetworkTime Series
🎯 What it does: This paper proposes a method called EventRPG, which utilizes correlation propagation in SNN to generate CAM and saliency maps, guiding the Drop and Mix augmentation of event data.
Evoke: Evoking Critical Thinking Abilities in LLMs via Reviewer-Author Prompt Editing
Xinyu Hu (Microsoft), Denis X Charles
Recommendation SystemTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: Proposes the Evoke framework, which utilizes a feedback loop of author and reviewer instances from the same large language model to iteratively edit and evaluate prompts, and implements automatic prompt engineering with a hard sample selector;
EX-Graph: A Pioneering Dataset Bridging Ethereum and X
Qian Wang (National University of Singapore), Bingsheng He (National University of Singapore)
Graph Neural NetworkGraph
🎯 What it does: The EX-Graph dataset was constructed, integrating the Ethereum transaction graph with the X (formerly Twitter) follower network, and verified X accounts were linked to Ethereum addresses through OpenSea; statistical analysis was performed on this heterogeneous graph, and experiments were conducted on on-chain tasks (link prediction, shuffle transaction detection) and cross-chain matching prediction tasks.
ExeDec: Execution Decomposition for Compositional Generalization in Neural Program Synthesis
Kensen Shi (Google DeepMind), Charles Sutton (Google DeepMind)
OptimizationAI Code AssistantTransformerText
🎯 What it does: A program synthesis method based on executing sub-goals, ExeDec, is proposed, which utilizes program execution states to decompose tasks step by step, enhancing the combinatorial generalization ability of program synthesis.
Expected flow networks in stochastic environments and two-player zero-sum games
Marco Jiralerspong (Mila Quebec AI Institute), Nikolay Malkin (Mila Quebec AI Institute)
OptimizationReinforcement LearningFlow-based ModelSequential
🎯 What it does: This paper proposes Expectation Flow Networks (EFlowNets) and Adversarial Flow Networks (AFlowNets), extending Generative Flow Networks (GFlowNets) to stochastic environments and two-player zero-sum games.
Experimental Design for Multi-Channel Imaging via Task-Driven Feature Selection
Stefano B. Blumberg (University College London), Daniel C. Alexander (University College London)
OptimizationComputational EfficiencyImageBiomedical DataMagnetic Resonance ImagingDiffusion Tensor Imaging
🎯 What it does: A task-driven multi-channel imaging experiment design method called TADRED is proposed, which can automatically select the smallest set of image channels while preserving user-specified task performance.
Explaining Kernel Clustering via Decision Trees
Maximilian Fleissner (Technical University of Munich), Debarghya Ghoshdastidar (Technical University of Munich)
OptimizationExplainability and InterpretabilityTabular
🎯 What it does: This paper proposes Kernel IMM and its greedy extensions Kernel ExKMC and Kernel Expand, using axis-aligned decision trees to approximate the clustering results of kernel k-means, achieving interpretable kernel clustering.
Explaining Time Series via Contrastive and Locally Sparse Perturbations
Zichuan Liu (Nanjing University), Qingsong Wen (Alibaba Group)
Explainability and InterpretabilityContrastive LearningTime SeriesBiomedical DataElectronic Health Records
🎯 What it does: This paper proposes ContraLSP, a time series explanation method based on contrastive learning and local sparse gates, which can generate adversarial perturbations while maintaining distribution consistency and learning binarized, smooth masks.
Exploiting Causal Graph Priors with Posterior Sampling for Reinforcement Learning
Mirco Mutti (Technion), Giorgia Ramponi (ETH Zurich)
Reinforcement Learning
🎯 What it does: An algorithm C-PSRL is proposed for reinforcement learning with partial causal graph priors, which can simultaneously learn causal structures and transition models.
Exploring Diffusion Time-steps for Unsupervised Representation Learning
Zhongqi Yue (Nanyang Technological University), Hanwang Zhang (Nanyang Technological University)
GenerationRepresentation LearningDiffusion modelAuto EncoderContrastive LearningImage
🎯 What it does: Using the time steps of the diffusion model as a prior, the encoder is trained to let different sub-vectors correspond to the attribute loss of different time steps, thereby learning unsupervised decomposable features;
Exploring Effective Stimulus Encoding via Vision System Modeling for Visual Prostheses
Chuanqing Wang (Westlake University), Mohamad Sawan (Westlake University)
OptimizationComputational EfficiencyRecurrent Neural NetworkSpiking Neural NetworkImage
🎯 What it does: Designed and implemented an end-to-end visual prosthesis stimulation optimization framework called StimuSEE, which uses V1 neuron spike information as supervision to generate stimulation patterns similar to normal vision.
Exploring Target Representations for Masked Autoencoders
xingbin liu, Rongrong Ji (Xiamen University)
ClassificationObject DetectionSegmentationKnowledge DistillationRepresentation LearningTransformerAuto EncoderImage
🎯 What it does: Through multi-stage masked knowledge distillation (dBOT), excellent visual representations can be obtained using a randomly initialized teacher model without the need for carefully designed target representations during pre-training.
Exploring the cloud of feature interaction scores in a Rashomon set
Sichao Li (Australian National University), Amanda S Barnard
TransformerImageTabular
🎯 What it does: This paper proposes Feature Interaction Score (FIS) and studies its distribution cloud (FISC) within the Rashomon set, while also providing a greedy search algorithm and Halo/Swarm visualization methods.
Exploring the Common Appearance-Boundary Adaptation for Nighttime Optical Flow
Hanyu Zhou (Huazhong University of Science and Technology), Luxin Yan (Huazhong University of Science and Technology)
Domain AdaptationContrastive LearningOptical FlowImageVideo
🎯 What it does: To address the challenges of optical flow estimation caused by weakened illumination and increased noise at night, the authors propose a 'Common Appearance-Boundary Adaptation Framework' constructed from daytime images and event cameras. This framework transfers global appearance motion knowledge and local boundary motion knowledge in two common latent spaces of reflectance and spatiotemporal gradients to enhance the accuracy of optical flow at night.
Exploring the Promise and Limits of Real-Time Recurrent Learning
Kazuki Irie (Center for Brain Science Harvard University), Jürgen Schmidhuber (Swiss AI Lab IDSIA USI & SUPSI)
Recurrent Neural NetworkReinforcement LearningSequential
🎯 What it does: This paper studies a real-time recurrent learning (RTRL) method that can be used in real environments, proposing an element-wise recurrent eLSTM implementation and comparing it with truncated BPTT in multi-task reinforcement learning.
Exploring Weight Balancing on Long-Tailed Recognition Problem
Naoya Hasegawa (University of Tokyo), Issei Sato (University of Tokyo)
ClassificationRecognitionConvolutional Neural NetworkImage
🎯 What it does: Analyze and simplify the weight balancing method in long-tail recognition, proposing that performance can be improved with only single-stage training.
Exposing Text-Image Inconsistency Using Diffusion Models
Mingzhen Huang (University at Buffalo), Siwei Lyu (University at Buffalo)
SegmentationAnomaly DetectionExplainability and InterpretabilityDiffusion modelImageTextMultimodality
🎯 What it does: A text-image inconsistency localization method based on diffusion models, D-TIIL, is proposed, which can locate inconsistent words and pixel regions between images and text, and provide consistency scores.
Expressive Losses for Verified Robustness via Convex Combinations
Alessandro De Palma (Inria), Alessio Lomuscio (Imperial College London)
OptimizationAdversarial AttackConvolutional Neural NetworkImage
🎯 What it does: A tunable parameter 'expressive loss' framework is proposed, implemented with three convex combinations (CC-IBP, MTL-IBP, Exp-IBP), achieving a better balance between accuracy and robustness in validating robust training.
Expressivity of ReLU-Networks under Convex Relaxations
Maximilian Baader (ETH Zurich), Martin Vechev (ETH Zurich)
🎯 What it does: This paper conducts a theoretical analysis of commonly used convex relaxation methods (IBP, DEEPPOLY, Δ, Multi‑Neuron) in the context of ReLU networks, proving that more precise relaxations on univariate functions can encode a larger class of continuous piecewise linear (CPWL) functions, and revealing the fundamental limitations of single-neuron relaxations on multivariate functions.
Extending Power of Nature from Binary to Real-Valued Graph Learning in Real World
Chunshu Wu (University of Rochester), Tong Geng (Pacific Northwestern National Laboratory)
Graph Neural NetworkSupervised Fine-TuningGraphTime Series
🎯 What it does: A natural force graph learning framework NP-GL is proposed, achieving high accuracy and extremely fast inference on four types of spatiotemporal tasks.
f-FERM: A Scalable Framework for Robust Fair Empirical Risk Minimization
Sina Baharlouei (University of Southern California), Meisam Razaviyayn (University of Southern California)
OptimizationTabular
🎯 What it does: A differentiable fair empirical risk minimization framework based on f-divergence, called f-FERM, is proposed, achieving a convergent unbiased gradient estimate under stochastic mini-batch training, followed by constructing a distributionally robust version using L1/∞ uncertainty sets.
Facing the Elephant in the Room: Visual Prompt Tuning or Full finetuning?
Cheng Han (Rochester Institute of Technology), Dongfang Liu (Rochester Institute of Technology)
ClassificationTransformerPrompt EngineeringImage
🎯 What it does: Conducted systematic experiments on 19 visual tasks to compare the performance of Visual Prompt Tuning (VPT) and Full Fine-Tuning (FT), and explored the reasons and conditions for the advantages and disadvantages of both methods.
Fair and Efficient Contribution Valuation for Vertical Federated Learning
Zhenan Fan (University of British Columbia), Yong Zhang (Huawei Technologies Canada)
Federated LearningTabular
🎯 What it does: A fair and effective contribution assessment method for vertical federated learning, VerFedSV, is proposed, and feasible computation is achieved in both synchronous and asynchronous scenarios.
Fair Classifiers that Abstain without Harm
Tongxin Yin (University of Michigan), Yang Liu (ByteDance Research)
ClassificationOptimizationTabular
🎯 What it does: A post-processing rejectable classification framework (FAN) is proposed, which determines whether to reject predictions through integer programming and trains the model to achieve automatic rejection based on this.
FairerCLIP: Debiasing CLIP's Zero-Shot Predictions using Functions in RKHSs
Sepehr Dehdashtian (Michigan State University), Vishnu Boddeti
ClassificationOptimizationComputational EfficiencyTransformerVision Language ModelContrastive LearningImage
🎯 What it does: Proposes the FairerCLIP method, which end-to-end debiases CLIP zero-shot predictions through RKHS, balancing implicit and exogenous associations, and supports both labeled and unlabeled learning.