ICML 2024 Papers — Page 12
International Conference on Machine Learning · 2610 papers
Implicit Representations via Operator Learning
Sourav Pal (University of Wisconsin Madison), Vikas Singh (University of Wisconsin Madison)
RestorationRepresentation LearningImageVideoBiomedical DataMagnetic Resonance Imaging
🎯 What it does: This paper proposes an implicit neural representation method based on operator learning (O-INR), treating the coordinate-value mapping as an operator between function spaces, rather than traditional MLP predictions.
Improved Bounds for Pure Private Agnostic Learning: Item-Level and User-Level Privacy
Bo Li (Hong Kong University of Science and Technology), Peng Ye (Hong Kong University of Science and Technology)
Safty and Privacy
🎯 What it does: This study investigates pure private learning under a non-biased learning model, with a particular focus on the sample complexity of project-level and user-level privacy, and proposes improved upper bounds.
Improved Communication-Privacy Trade-offs in $L_2$ Mean Estimation under Streaming Differential Privacy
Wei-Ning Chen (Stanford University), Zheng Xu (Google)
Federated LearningSafty and PrivacyImageText
🎯 What it does: A method for L2 mean estimation under the constraints of communication compression and central differential privacy is proposed, and this method is extended to the streaming differential privacy (DP-FTRL) environment.
Improved Differentially Private and Lazy Online Convex Optimization: Lower Regret without Smoothness Requirements
Naman Agarwal (Google DeepMind), Abhradeep Guha Thakurta
OptimizationSafty and Privacy
🎯 What it does: A differential privacy online convex optimization algorithm is proposed without the need for smoothness assumptions, utilizing Gibbs sampling with Euclidean regularization to achieve low risk.
Improved Dimensionality Dependence for Zeroth-Order Optimisation over Cross-Polytopes
Weijia Shao (Federal Institute for Occupational Safety and Health)
Optimization
🎯 What it does: This paper proposes a new algorithm for zero-order (gradient-free) optimization on cross-polytopes, significantly reducing dimensional dependence;
Improved Generalization of Weight Space Networks via Augmentations
Aviv Shamsian (Bar Ilan University), Haggai Maron (Technion)
ClassificationRepresentation LearningContrastive LearningImagePoint Cloud
🎯 What it does: This paper studies the serious overfitting problem that arises in Deep Weight Space (DWS) learning and enhances the model's generalization ability through data augmentation techniques.
Improved Modelling of Federated Datasets using Mixtures-of-Dirichlet-Multinomials
Jonathan Scott (Institute of Science and Technology Austria), Áine Cahill (Apple)
Federated LearningSafty and PrivacyImageTabular
🎯 What it does: A completely federated, privacy-friendly algorithm is proposed, which learns the true client distribution using the Mixture-of-Dirichlet-Multinomials (MDM) model, and utilizes this distribution to partition server proxy data into realistic simulated clients, thereby enhancing the authenticity of server-side training simulations.
Improved Operator Learning by Orthogonal Attention
Zipeng Xiao (Shanghai Jiao Tong University), Hang Su (Tsinghua University)
TransformerTabularBenchmark
🎯 What it does: A neural operator based on an orthogonal attention mechanism (ONO) is proposed to approximate the solution family of PDEs.
Improved Stability and Generalization Guarantees of the Decentralized SGD Algorithm
Batiste Le bars, Giovanni Neglia (Inria Universite Cote d'Azur)
OptimizationGraph
🎯 What it does: This paper conducts a systematic stability analysis of the generalization error of the decentralized stochastic gradient descent (D-SGD) algorithm and proposes a new optimal upper bound for the generalization error.
Improving Accuracy-robustness Trade-off via Pixel Reweighted Adversarial Training
Jiacheng Zhang (University of Melbourne), Tongliang Liu (University of Sydney)
Adversarial AttackConvolutional Neural NetworkGenerative Adversarial NetworkImage
🎯 What it does: Proposes Pixel Weighted Adversarial Training (PART), which enhances the accuracy-robustness trade-off by reducing the perturbation upper limit for low-impact pixels.
Improving Adversarial Energy-Based Model via Diffusion Process
Cong Geng (Vivo Mobile Communication Company Limited), Bo Li (Vivo Mobile Communication Company Limited)
GenerationData SynthesisDiffusion modelGenerative Adversarial NetworkImage
🎯 What it does: A training framework that embeds the Energy-Based Model (EBM) into the denoising diffusion process (DDAEBM) is proposed, achieving MCMC-free training and efficient sampling.
Improving Antibody Humanness Prediction using Patent Data
Talip Ucar (AstraZeneca), Pietro Sormanni (University of Cambridge)
Drug DiscoveryTransformerContrastive LearningBiomedical Data
🎯 What it does: Using antibody sequences and functional labels from patent databases, a multi-stage contrastive learning and fine-tuning method was employed to construct an antibody 'humanness' prediction model (SelfPAD).
Improving Computational Complexity in Statistical Models with Local Curvature Information
Pedram Akbarian (University of Texas at Austin), Nhat Ho (University of Texas at Austin)
OptimizationComputational EfficiencyTabular
🎯 What it does: This paper proposes a normalized gradient descent algorithm (NormGD) that uses the maximum eigenvalue of the Hessian of the sample loss function as a step size scaling factor, and proves that it can achieve the optimal statistical radius within logarithmic iterations under uniform loss functions, greatly enhancing computational efficiency in singular statistical models.
Improving Context Understanding in Multimodal Large Language Models via Multimodal Composition Learning
Wei Li (Zhejiang University), Mohan Kankanhalli (National University of Singapore)
RetrievalTransformerLarge Language ModelContrastive LearningImageTextMultimodality
🎯 What it does: This paper proposes a Multimodal Combination Learning (MCL) method that utilizes a frozen LLM and a CLIP visual encoder. By training on generated Multimodal Combination data (MMC), it aims to improve the understanding and generation of complex multimodal queries through visual-language mapping and context retrieval tasks.
Improving Diffusion Models for Inverse Problems Using Optimal Posterior Covariance
Xinyu Peng (Shanghai Jiao Tong University), Hongkai Xiong (Shanghai Jiao Tong University)
RestorationSuper ResolutionDiffusion modelImageOrdinary Differential Equation
🎯 What it does: This paper presents a unified perspective that equates the implementation of all zero-copy diffusion models in inverse problems to the use of an isotropic Gaussian approximation for the intractable denoising posterior, and based on this, optimizes the posterior covariance through maximum likelihood estimation to improve existing methods.
Improving Equivariant Graph Neural Networks on Large Geometric Graphs via Virtual Nodes Learning
Yuelin Zhang (Renmin University of China), Wenbing Huang (Renmin University of China)
Graph Neural NetworkGraph
🎯 What it does: We propose FastEGNN, which learns a set of ordered virtual nodes to enable E(3)-equivariant GNNs to maintain global message passing on large-scale geometric graphs while balancing efficiency and accuracy.
Improving Factuality and Reasoning in Language Models through Multiagent Debate
Yilun Du (Massachusetts Institute of Technology), Igor Mordatch (Google DeepMind)
TransformerLarge Language ModelPrompt EngineeringTextChain-of-Thought
🎯 What it does: A multi-agent debate framework is proposed, allowing multiple instances of language models to collaboratively generate and refine answers in multi-turn discussions, enhancing reasoning and factual accuracy.
Improving fine-grained understanding in image-text pre-training
Ioana Bica (Google DeepMind), Jovana Mitrovic
Object DetectionSegmentationRetrievalTransformerVision Language ModelContrastive LearningImageTextMultimodality
🎯 What it does: This paper proposes a novel image-text pre-training method called SPARC, which utilizes sparse alignment to generate language-aggregated visual embeddings and performs fine-grained contrastive learning on a single image-text pair, supplemented by global contrastive loss to learn multi-scale representations.
Improving Generalization in Offline Reinforcement Learning via Adversarial Data Splitting
Da Wang (Shanxi University), Jiye Liang (Shanxi University)
Meta LearningReinforcement LearningTabular
🎯 What it does: A framework for Adversarial Data Splitting (ADS) is proposed, which utilizes the distribution shift of training/validation subsets to enhance the generalization performance of offline reinforcement learning models.
Improving Gradient-Guided Nested Sampling for Posterior Inference
Pablo Lemos (Dreamfold), Laurence Perreault-Levasseur (Perimeter Institute)
OptimizationComputational EfficiencyGenerative Adversarial NetworkTabularStochastic Differential Equation
🎯 What it does: A gradient-guided nested sampling algorithm GGNS is designed, combining techniques such as differentiable programming, Hamiltonian slice sampling, clustering decomposition, and dynamic parallelization to achieve efficient sampling and model comparison in high-dimensional multimodal Bayesian inference.
Improving Group Robustness on Spurious Correlation Requires Preciser Group Inference
Yujin Han (University of Hong Kong), Difan Zou (University of Hong Kong)
ClassificationDomain AdaptationContrastive LearningImageText
🎯 What it does: The GIC (Group Inference via data Comparison) method is proposed, which accurately infers group labels by utilizing the distribution differences between the training set and the comparison set, thereby improving the worst group accuracy in tasks with spurious correlations.
Improving Instruction Following in Language Models through Proxy-Based Uncertainty Estimation
JoonHo Lee (Samsung SDS), Seungjai Min (Samsung SDS)
OptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningText
🎯 What it does: A framework for uncertainty assessment based on agents is proposed, constructing an Uncertainty-Aware Reward Model (URM) and applying it to data selection and training objectives (UDPO, UCPO), thereby enhancing the instruction-following capability of language models.
Improving Interpretation Faithfulness for Vision Transformers
Lijie Hu (King Abdullah University of Science and Technology), Di Wang (King Abdullah University of Science and Technology)
Explainability and InterpretabilityTransformerDiffusion modelImage
🎯 What it does: A new method called FViTs (Faithful Vision Transformers) is proposed, aimed at improving the interpretability of Vision Transformers (ViTs) and addressing their vulnerability issues under input perturbations.
Improving Neural Additive Models with Bayesian Principles
Kouroche Bouchiat (ETH Zurich), Vincent Fortuin (Munich Center for Machine Learning)
ClassificationAnomaly DetectionAuto EncoderTabularBiomedical DataElectronic Health Records
🎯 What it does: An improved Neural Additive Model (NAM) is introduced, incorporating a Bayesian framework that provides credible intervals, automatic feature selection, and interaction selection.
Improving Neural Logic Machines via Failure Reflection
Zhiming Li (Nanyang Technological University), Yang Liu (Nanyang Technological University)
Reinforcement LearningGraph
🎯 What it does: This paper proposes a Failure Reflection Guided Regularizer (FRGR) based on error root cause mining and penalties, and applies it to the training of neural logic machines to enhance the model's reasoning performance.
Improving Open-Ended Text Generation via Adaptive Decoding
Wenhong Zhu (Shanghai Jiao Tong University), Rui Wang (Shanghai Jiao Tong University)
GenerationTransformerLarge Language ModelText
🎯 What it does: An adaptive decoding algorithm is proposed, which dynamically determines the size of the candidate set for the next token based on entropy and confidence increments to enhance the diversity and coherence of open-ended text generation.
Improving Prototypical Visual Explanations with Reward Reweighing, Reselection, and Retraining
Aaron Jiaxun Li, Bin Yu (University of California, Berkeley)
Explainability and InterpretabilityConvolutional Neural NetworkReinforcement LearningImage
🎯 What it does: The R3 framework is proposed to enhance the interpretability and predictive accuracy of ProtoPNet through three steps: reward weighting, prototype re-selection, and re-training.
Improving Robustness to Multiple Spurious Correlations by Multi-Objective Optimization
Nayeong Kim (POSTECH), Suha Kwak (POSTECH)
ClassificationOptimizationConvolutional Neural NetworkImageBenchmark
🎯 What it does: A multi-objective optimization-based debiasing training framework is proposed, which groups training data and dynamically adjusts group weights with the goal of minimizing the loss across all groups simultaneously, thereby achieving an unbiased and accurate classifier in the presence of multiple spurious correlations in the data.
Improving SAM Requires Rethinking its Optimization Formulation
Wanyun Xie (Ecole Polytechnique Federale de Lausanne), Volkan Cevher (Ecole Polytechnique Federale de Lausanne)
ClassificationOptimizationTransformerSupervised Fine-TuningImageText
🎯 What it does: Transform the original Sharpness-Aware Minimization (SAM) from a zero-sum game into a bi-level optimization framework called BiSAM, which directly constructs perturbation strategies based on 0-1 loss.
Improving Sample Efficiency of Model-Free Algorithms for Zero-Sum Markov Games
Songtao Feng (University of Florida), Yingbin Liang (Ohio State University)
Reinforcement Learning
🎯 What it does: A new model-free algorithm based on min-gap reference-advantage decomposition Q-learning is proposed for learning approximate Nash equilibria in two-player zero-sum Markov games.
Improving Sharpness-Aware Minimization by Lookahead
Runsheng Yu (Hong Kong University of Science and Technology), James Kwok
OptimizationTransformerSupervised Fine-TuningImageText
🎯 What it does: This paper proposes the introduction of the Lookahead mechanism into Sharpness-Aware Minimization (SAM) to achieve flatter and more robustly converging optimal solutions.
Improving Token-Based World Models with Parallel Observation Prediction
Lior Cohen (Technion - Israel Institute of Technology), Shie Mannor (Technion - Israel Institute of Technology)
Recurrent Neural NetworkReinforcement LearningWorld ModelVideoBenchmark
🎯 What it does: This paper proposes an improved token-based world model, designs and implements a Parallel Observation Prediction (POP) mechanism, and constructs a new REM agent.
Improving Transformers with Dynamically Composable Multi-Head Attention
Da Xiao (Beijing University of Posts and Telecommunications), xingyuan yuan
TransformerLarge Language ModelImageText
🎯 What it does: A dynamic composable multi-head attention (DCMHA) mechanism is proposed to replace the traditional MHA, enhancing the expressive power of the Transformer and reducing head redundancy.
In value-based deep reinforcement learning, a pruned network is a good network
Johan Samir Obando Ceron, Pablo Samuel Castro (Google DeepMind)
Reinforcement LearningVideo
🎯 What it does: This paper studies the use of gradual magnitude pruning in value-based deep reinforcement learning to improve network parameter utilization and agent performance, exploring its effects across various agents, training modes (online, offline, low data, Actor-Critic), and different network widths.
In-context Convergence of Transformers
Yu Huang (University of Pennsylvania), Yingbin Liang (Ohio State University)
Transformer
🎯 What it does: Analyzed the context learning process and convergence of single-layer Transformers using the Softmax attention mechanism and trained through gradient descent for linear function classes;
In-Context Decision Transformer: Reinforcement Learning via Hierarchical Chain-of-Thought
Sili Huang (Jilin University), Bo Yang (Jilin University)
TransformerReinforcement LearningTabularSequentialChain-of-Thought
🎯 What it does: This paper proposes an IDT model for high-level decision generation in offline reinforcement learning using autoregressive Transformers, which can autonomously improve through trial and error during testing.
In-Context Freeze-Thaw Bayesian Optimization for Hyperparameter Optimization
Herilalaina Rakotoarison (University of Freiburg), Frank Hutter (University of Freiburg)
OptimizationHyperparameter SearchTransformerTabular
🎯 What it does: A Transformer-based Freeze-Thaw Bayesian Optimization (ifBO) framework is proposed for hyperparameter optimization in deep learning without online training.
In-Context Language Learning: Architectures and Algorithms
Ekin Akyürek, Jacob Andreas (Massachusetts Institute of Technology)
Recurrent Neural NetworkTransformerTextBenchmark
🎯 What it does: This paper systematically evaluates the performance of different neural sequence models on the In-context Learning (ICL) task for regular languages by constructing a new benchmark called REGBENCH. It finds that the Transformer significantly outperforms models such as RNN and CNN in such tasks. Further analysis of the attention mechanism in the Transformer reveals that it achieves high-order n-gram statistics through an 'n-gram head'. Subsequently, the paper proposes adding fixed n-gram attention heads to other models to enhance their ICL capabilities, and validates the positive impact of this improvement on language modeling using real corpus SlimPajama.
In-Context Learning Agents Are Asymmetric Belief Updaters
Johannes A. Schubert (Max Planck Institute for Biological Cybernetics), Eric Schulz (Max Planck Institute for Biological Cybernetics)
Meta LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningAgentic AIText
🎯 What it does: This study investigates the contextual learning dynamics of large language models (LLMs) in the 2AFC task, finding that their belief updates exhibit asymmetry: they place more emphasis on positive prediction errors when there is agency and provide feedback in the reverse direction when complete feedback is given; when there is no agency, no bias occurs; an idealized model trained with Meta-RL validates that these behaviors are justifiable.
In-context Learning on Function Classes Unveiled for Transformers
Zhijie Wang (Shanghai Jiao Tong University), Shuai Li (Shanghai Jiao Tong University)
Transformer
🎯 What it does: This paper studies the performance of Transformer-based neural sequence models in context learning, exploring how Transformers learn different types of function classes in context. Through construction, it is proven that a class of Transformers with different activation functions can achieve approximate gradient descent on neural network parameters, and provides upper bounds on the number of heads, hidden dimensions, and layers required for Transformers.
In-Context Principle Learning from Mistakes
Tianjun Zhang (University of California Berkeley), Uri Alon (Google DeepMind)
TransformerLarge Language ModelPrompt EngineeringTextChain-of-Thought
🎯 What it does: This paper proposes the LEAP method, which first allows large language models to actively generate errors based on a small number of given examples, and then generates low-level and high-level task-specific principles through self-reflection, incorporating these principles into prompts to enhance the model's reasoning performance on unseen samples.
In-Context Reinforcement Learning for Variable Action Spaces
Viacheslav Sinii (Tinkoff), Sergey Kolesnikov
TransformerReinforcement LearningContrastive LearningSequential
🎯 What it does: A Transformer model named Headless-AD is proposed, which removes the output linear head, uses random orthogonal action embeddings, and directly predicts embeddings, enabling in-context reinforcement learning in any discrete action space (size, semantics, order) after a single training session.
In-Context Sharpness as Alerts: An Inner Representation Perspective for Hallucination Mitigation
Shiqi Chen (City University of Hong Kong), Junxian He (Hong Kong University of Science and Technology)
TransformerLarge Language ModelText
🎯 What it does: This paper proposes 'in-context sharpness' as an unsupervised metric for identifying and correcting hallucinations by analyzing the activation patterns in the hidden states of large language models, and designs an Activation Decoding method based on this metric.
In-Context Unlearning: Language Models as Few-Shot Unlearners
Martin Pawelczyk (Harvard University), Himabindu Lakkaraju (Harvard University)
OptimizationData-Centric LearningTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: This paper proposes an 'In-Context Unlearning' method that does not require access to model parameters. It achieves forgetting specific training points by adding target training samples in the context during inference, flipping their labels, and supplementing with several correctly labeled samples.
In-context Vectors: Making In Context Learning More Effective and Controllable Through Latent Space Steering
Sheng Liu (Stanford University), James Y. Zou
TransformerPrompt EngineeringContrastive LearningText
🎯 What it does: A method called 'In-context Vector (ICV)' is proposed, which extracts task information from examples and directly adds this vector to the latent representations of all layers during inference, replacing the traditional example preset approach to achieve zero-training context learning.
Incentivized Learning in Principal-Agent Bandit Games
Antoine Scheid (Ecole Polytechnique), Alain Oliviero Durmus
OptimizationReinforcement LearningTabular
🎯 What it does: Designed and implemented two algorithms, IPA and Contextual IPA, to learn incentive strategies in a principal-agent game environment to maximize the principal's revenue.
Incorporating Information into Shapley Values: Reweighting via a Maximum Entropy Approach
Darya Biparva (University of Minnesota), Donatello Materassi (University of Minnesota)
OptimizationExplainability and InterpretabilityTabular
🎯 What it does: This paper studies how to incorporate prior causal information into Shapley values, proposing a weighted Shapley method based on the maximum entropy principle, and introducing two new interpretative mechanisms: Markov Blanket Shapley and Ancestor Shapley, to achieve observational and interventional feature attribution, respectively.
Incorporating probabilistic domain knowledge into deep multiple instance learning
Ghadi S. Al Hajj (University of Oslo), Geir Kjetil Sandve (University of Oslo)
ClassificationObject DetectionConvolutional Neural NetworkSupervised Fine-TuningTabular
🎯 What it does: Design DeeMILIP under the multi-instance learning framework, utilizing available instance-level probabilistic prior knowledge to enhance the performance of deep models.
Incremental Topological Ordering and Cycle Detection with Predictions
Samuel McCauley (Williams), Shikha Singh (Williams)
Machine LearningGraphTime Series
🎯 What it does: A data structure for incremental topological sorting and cycle detection based on machine learning prediction has been designed, and two implementations, ideal learning sorting and learning DFS, have been proposed.
Indirectly Parameterized Concrete Autoencoders
Alfred Nilsson (KTH Royal Institute of Technology), Hossein Azizpour (KTH Royal Institute of Technology)
ClassificationRepresentation LearningAuto EncoderImageAudio
🎯 What it does: Proposes Indirectly Parameterized Concrete Autoencoders (IP-CAE), which improve traditional Concrete Autoencoders (CAE) for differentiable feature selection.
Individual Contributions as Intrinsic Exploration Scaffolds for Multi-agent Reinforcement Learning
Xinran Li (Hong Kong University of Science and Technology), Jun Zhang (Hong Kong University of Science and Technology)
Reinforcement LearningSequential
🎯 What it does: This paper proposes an Individual Contribution-based Exploration Scaffold (ICES) method, which evaluates the impact of each agent's actions on global potential state transitions to guide their exploration behavior in multi-agent reinforcement learning.
Individual Fairness in Graph Decomposition
Kamesh Munagala (Duke University), Govind S. Sankar (Duke University)
Graph Neural NetworkGraph
🎯 What it does: A low diameter decomposition method is proposed in planar graphs to achieve individual fairness, and the shortcomings of the classical KPR algorithm are analyzed.
Individualized Privacy Accounting via Subsampling with Applications in Combinatorial Optimization
Badih Ghazi (Google Research), Adam Sealfon (Google Research)
OptimizationSafty and Privacy
🎯 What it does: A technique is proposed that processes through a single-sided differential privacy mechanism with subsampling, making it satisfy two-sided differential privacy, and thus constructs a pure differential privacy combinatorial optimization algorithm.
Inexact Newton-type Methods for Optimisation with Nonnegativity Constraints
Oscar Smee (University of Queensland), Fred Roosta (University of Queensland)
OptimizationImageText
🎯 What it does: A Newton-MR algorithm based on a two-metric projection framework is designed and analyzed to solve large-scale non-convex optimization problems with non-negative constraints.
InferCept: Efficient Intercept Support for Augmented Large Language Model Inference
Reyna Abhyankar (University of California), Yiying Zhang (University of California)
TransformerLarge Language ModelText
🎯 What it does: In response to interruptions from external tools or human-computer interactions during augmented large language model (LLM) inference, the INFERCEPT framework is proposed, which can efficiently handle interception events while maintaining context without loss.
Inferring Change Points in High-Dimensional Linear Regression via Approximate Message Passing
Gabriel Arpino (University of Cambridge), Ramji Venkataramanan (University of Cambridge)
Anomaly DetectionOptimizationTime Series
🎯 What it does: A high-dimensional linear regression change point detection method based on Approximate Message Passing (AMP) is proposed, which can simultaneously estimate signal parameters and change point locations, and provide uncertainty quantification.
Inferring Dynamic Networks from Marginals with Iterative Proportional Fitting
Serina Chang (Stanford University), Johan Ugander (Stanford University)
Graph Neural NetworkGraphTabular
🎯 What it does: A dynamic network inference framework based on Iterative Proportional Fitting (IPF) is proposed, combining IPF with statistical models to provide theoretical explanations and error analysis;
Inferring the Long-Term Causal Effects of Long-Term Treatments from Short-Term Experiments
Allen Tran (Netflix), Nathan Kallus (Cornell University)
Reinforcement LearningTabularElectronic Health Records
🎯 What it does: The study investigates how to infer the long-term causal effects of sustained interventions based solely on short-term experimental data;
InfiAgent-DABench: Evaluating Agents on Data Analysis Tasks
Xueyu Hu (Zhejiang University), Fei Wu
Data-Centric LearningTransformerLarge Language ModelAgentic AIPrompt EngineeringTabularBenchmark
🎯 What it does: Designed and released InfiAgent-DABench to evaluate the end-to-end performance of LLM agents in data analysis tasks, and built a dedicated agent DAAgent based on this benchmark.
Infinite-Horizon Distributionally Robust Regret-Optimal Control
Taylan Kargin (California Institute of Technology), Babak Hassibi (California Institute of Technology)
Optimization
🎯 What it does: This paper studies the Wasserstein-2 distributionally robust regret optimal control of infinite time-domain linear quadratic systems, solving the steady-state worst-case expected regret minimization problem.
InfoNet: Neural Estimation of Mutual Information without Test-Time Optimization
Zhengyang Hu (University of Hong Kong), Yanchao Yang (University of Hong Kong)
OptimizationComputational EfficiencyTransformerVideo
🎯 What it does: A neural network called InfoNet is proposed, which directly outputs the mutual information estimate of any data stream, avoiding optimization during testing;
Information Complexity of Stochastic Convex Optimization: Applications to Generalization, Memorization, and Tracing
Idan Attias (Ben Gurion University), Daniel M. Roy (University of Toronto)
Optimization
🎯 What it does: This paper studies the relationship between the accuracy of learning algorithms and their 'memory' capability of training samples in Stochastic Convex Optimization (SCO), and measures the information revealed by learning algorithms about the training data using Conditional Mutual Information (CMI).
Information Flow in Self-Supervised Learning
Zhiquan Tan (Tsinghua University), Yifan Zhang (Tsinghua University)
Representation LearningTransformerContrastive LearningImage
🎯 What it does: This study investigates the implicit maximization of the losses from contrastive learning and feature decorrelation methods in the framework of matrix information theory, introducing matrix entropy regularization into MAE and proposing MMAE;
Information-Directed Pessimism for Offline Reinforcement Learning
Alec Koppel (J.P. Morgan), Sumitra Ganesh
Reinforcement LearningTabular
🎯 What it does: This paper proposes an offline reinforcement learning lazy policy optimization framework IDP based on discrete Stein mismatch degree, and presents two algorithms: IDP-VI and IDP-Q.
Inherent Trade-Offs between Diversity and Stability in Multi-Task Benchmarks
Guanhua Zhang (Max Planck Institute for Intelligent Systems), Moritz Hardt (Max Planck Institute for Intelligent Systems)
TextBenchmark
🎯 What it does: Study the diversity and stability of multi-task benchmarks, revealing the inherent trade-off between the two;
Initial Guessing Bias: How Untrained Networks Favor Some Classes
Emanuele Francazi (École Polytechnique Fédérale de Lausanne), Marco Baity-Jesi (Eawag)
ClassificationObject DetectionConvolutional Neural NetworkTransformerImage
🎯 What it does: This paper proposes and theoretically analyzes the 'Initial Guess Bias (IGB)' present in untrained deep neural networks, which refers to the network giving unbalanced predictions for different categories before seeing any data.
Instruction Tuning for Secure Code Generation
Jingxuan He (ETH Zurich), Martin Vechev (ETH Zurich)
GenerationSafty and PrivacyAI Code AssistantTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This paper proposes SafeCoder, which combines security refinement and instruction tuning to enhance the security of code generated by LLMs during the instruction tuning phase while maintaining functional correctness.
InstructRetro: Instruction Tuning post Retrieval-Augmented Pretraining
Boxin Wang (NVIDIA), Bryan Catanzaro (NVIDIA)
GenerationRetrievalTransformerLarge Language ModelSupervised Fine-TuningTextRetrieval-Augmented Generation
🎯 What it does: This paper presents Retro 48B—the largest retrieval-augmented pre-trained language model—and conducts instruction fine-tuning based on it to obtain InstructRetro.
InstructSpeech: Following Speech Editing Instructions via Large Language Models
Rongjie Huang (Zhejiang University), Zhou Zhao (Zhejiang University)
GenerationTransformerLarge Language ModelAudio
🎯 What it does: Construct a triplet dataset of instruction-speech-editing to train a large language model, InstructSpeech, enabling the editing of speech's semantic and acoustic properties according to natural language instructions.
InstructZero: Efficient Instruction Optimization for Black-Box Large Language Models
Lichang Chen (University of Maryland), Tianyi Zhou (University of Maryland)
OptimizationTransformerLarge Language ModelReinforcement LearningPrompt EngineeringText
🎯 What it does: The INSTRUCTZERO method is proposed, which utilizes low-dimensional soft prompts to drive open-source LLMs to generate interpretable instructions. These instructions are then delivered to a black-box LLM for zero-shot evaluation, and the soft prompts are continuously improved through a Bayesian optimization loop to find the optimal instructions.
Integrated Hardware Architecture and Device Placement Search
Irene Wang (Georgia Institute of Technology), Divya Mahajan (Georgia Institute of Technology)
OptimizationTransformerLarge Language ModelText
🎯 What it does: This study proposes the PHAZE framework, which achieves joint optimization of hardware accelerator architecture and distributed training model allocation strategies, automatically generating optimal accelerator configurations and device placement schemes.
Integrating Global Context Contrast and Local Sensitivity for Blind Image Quality Assessment
Xudong Li (Xiamen University), Rongrong Ji (Xiamen University)
Knowledge DistillationTransformerContrastive LearningImage
🎯 What it does: This paper proposes CSIQA, a no-reference image quality assessment framework that integrates global context contrast and local sensitivity.
Integrating Multimodal Data for Joint Generative Modeling of Complex Dynamics
Manuel Brenner (Central Institute of Mental Health Medical Faculty Mannheim Heidelberg University), Daniel Durstewitz (Central Institute of Mental Health Medical Faculty Mannheim Heidelberg University)
GenerationData SynthesisOptimizationRecurrent Neural NetworkAuto EncoderMultimodalityTime SeriesMagnetic Resonance Imaging
🎯 What it does: A multi-modal variational autoencoder-based framework for generating sparse teacher forcing signals (MTF) is proposed, capable of reconstructing dynamical systems from discrete or multi-modal time series.
Interacting Diffusion Processes for Event Sequence Forecasting
Mai Zeng (McGill University), Mark Coates (McGill University)
TransformerDiffusion modelTime SeriesSequential
🎯 What it does: A generative model utilizing the Cross-Diffusion process is proposed, which directly predicts future multi-step event sequences (time intervals and event types), avoiding the error accumulation of traditional autoregressive methods.
Interaction-based Retrieval-augmented Diffusion Models for Protein-specific 3D Molecule Generation
Zhilin Huang (Shenzhen International Graduate School Tsinghua University), Wenming Yang (Shenzhen International Graduate School Tsinghua University)
GenerationData SynthesisDrug DiscoveryGraph Neural NetworkDiffusion modelBiomedical DataRetrieval-Augmented Generation
🎯 What it does: A 3D molecular diffusion model based on interactive retrieval enhancement, IRDIFF, is proposed for drug molecule generation targeting protein sites.
InterLUDE: Interactions between Labeled and Unlabeled Data to Enhance Semi-Supervised Learning
Zhe Huang (Tufts University), Michael C Hughes
ClassificationRepresentation LearningTransformerContrastive LearningImageBiomedical DataBenchmark
🎯 What it does: This paper proposes InterLUDE, a semi-supervised learning framework that interacts with labeled and unlabeled samples in the embedding space and loss function.
Interplay of ROC and Precision-Recall AUCs: Theoretical Limits and Practical Implications in Binary Classification
Martin Mihelich (Glanceable), Charles Dognin (Glanceable)
ClassificationAnomaly Detection
🎯 What it does: This paper theoretically derives the minimum and maximum values of the area under the precision-recall curve (AUC PR) given the area under the ROC curve (AUC ROC); conversely, it provides precise upper and lower bounds between the two metrics.
Interpretability Illusions in the Generalization of Simplified Models
Dan Friedman (Princeton University), Asma Ghandeharioun (Google Research)
Explainability and InterpretabilityTransformerLarge Language ModelText
🎯 What it does: By simplifying the Transformer language model using SVD, k-means clustering, and one-hot attention, the study evaluates the matching degree of the simplified proxy and the original model in out-of-distribution generalization on Dyck language and code completion tasks.
Interpretable Deep Clustering for Tabular Data
Jonathan Svirsky (Bar Ilan University), Ofir Lindenbaum (Bar Ilan University)
Explainability and InterpretabilityAuto EncoderImageTabular
🎯 What it does: An interpretable deep clustering framework, IDC, is proposed, which can perform clustering on tabular data and provide sample-level and cluster-level feature explanations.
InterpreTabNet: Distilling Predictive Signals from Tabular Data by Salient Feature Interpretation
Jacob Yoke Hong Si (University of Toronto), Rahul Krishnan
Explainability and InterpretabilityKnowledge DistillationTransformerLarge Language ModelPrompt EngineeringAuto EncoderTabular
🎯 What it does: This paper proposes InterpreTabNet, an improved TabNet model that utilizes Gumbel-Softmax sampling and KL divergence regularization to generate sparse, non-overlapping attention masks, thereby enhancing the interpretability of tabular data; it also combines GPT-4 to provide natural language explanations for the masks.
Interpreting and Improving Diffusion Models from an Optimization Perspective
Frank Permenter (Toyota Research Institute), Chenyang Yuan (Toyota Research Institute)
GenerationOptimizationDiffusion modelImageText
🎯 What it does: This paper views the denoising process of diffusion models as an approximate projection, equating diffusion sampling to gradient descent on the Euclidean distance function, thereby providing a convergence analysis of DDIM sampling and proposing a second-order sampler based on gradient estimation, which can generate high-quality samples with only 5-10 function evaluations.
Interpreting and Improving Large Language Models in Arithmetic Calculation
Wei Zhang (University of Science and Technology of China), Jieping Ye (Alibaba Cloud)
TransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: In large language models, a small number of attention heads focusing on arithmetic operations and subsequent MLP layers were identified through causal intervention and ablation experiments. Precise fine-tuning based on these key components was conducted to enhance arithmetic reasoning abilities, with minimal impact on other tasks.
Interpreting Equivariant Representations
Andreas Abildtrup Hansen (Technical University of Denmark), Aasa Feragen
Explainability and InterpretabilityRepresentation LearningGraph Neural NetworkGraph
🎯 What it does: This paper studies the ambiguity of the equivariant latent space implicitly obtained in equivariant networks and proposes obtaining unambiguous latent representations by projecting onto invariant spaces, thereby improving the interpretability and performance of downstream tasks.
Intersecting-Boundary-Sensitive Fingerprinting for Tampering Detection of DNN Models
Bai Xiaofan, Hai Jin (Huazhong University of Science and Technology)
Anomaly DetectionOptimizationConvolutional Neural NetworkTransformerImage
🎯 What it does: This paper proposes the Intersecting-Boundary-Sensitive Fingerprinting (IBSF) method for detecting whether a deployed DNN model has been tampered with in a black-box environment where only the model's top-1 prediction label is available.
Intersectional Unfairness Discovery
Gezheng Xu (University of Western Ontario), Changjian Shui (Vector Institute)
GenerationData SynthesisReinforcement LearningAuto EncoderImageText
🎯 What it does: This paper proposes viewing the discovery of cross-group unfairness as a generative task and designs a Bias-Guided Generative Network (BGGN) that efficiently generates high-bias subgroups of multi-attribute cross groups using bias values as rewards.
Invariant Risk Minimization Is A Total Variation Model
Zhao-Rong Lai (Jinan University), Weiwen Wang (Jinan University)
Domain AdaptationContrastive LearningTabularTime Series
🎯 What it does: The invariant risk minimization (IRM) is theorized as a total variation (TV) model, and IRM-TVℓ1 and Minimax-TVℓ1 based on TVℓ1 are proposed for robust OOD generalization with or without environment labels.
Inverse-Variance Weighting for Estimation of Heterogeneous Treatment Effects
Aaron Fisher (Genentech)
🎯 What it does: This paper proposes and systematically evaluates the impact of inverse variance weighting (IVW) on the stability and effectiveness of the pseudo-outcome regression (POR) method when estimating conditional average treatment effects (CATE);
Investigating Pre-Training Objectives for Generalization in Vision-Based Reinforcement Learning
Donghu Kim (KAIST), Jaegul Choo (KAIST)
Convolutional Neural NetworkReinforcement LearningImageVideoBenchmark
🎯 What it does: This study investigates the impact of different pre-training objectives on the generalization performance of models in various distribution environments (ID, Near-OOD, Far-OOD) in visual reinforcement learning, and conducts a systematic evaluation through a unified experimental framework.
INViT: A Generalizable Routing Problem Solver with Invariant Nested View Transformer
Han Fang (Shanghai Jiao Tong University), Yutong Ban (Shanghai Jiao Tong University)
OptimizationTransformerReinforcement LearningGraph
🎯 What it does: A generalizable routing problem solver based on Transformer, INViT, is proposed, which can achieve near-optimal solutions on larger-scale and differently distributed instances while being trained only on small-scale, uniformly distributed instances.
IOI: Invisible One-Iteration Adversarial Attack on No-Reference Image- and Video-Quality Metrics
Ekaterina Shumitskaya (Lomonosov Moscow State University), Dmitriy S. Vatolin (Lomonosov Moscow State University)
Adversarial AttackImageVideo
🎯 What it does: An imperceptible one-iteration adversarial attack (IOI) is proposed for no-reference image and video quality assessment metrics.
Irregular Multivariate Time Series Forecasting: A Transformable Patching Graph Neural Networks Approach
Weijia Zhang (Hong Kong University of Science and Technology), Hui Xiong (Hong Kong University of Science and Technology)
Graph Neural NetworkTransformerTime SeriesBiomedical DataElectronic Health Records
🎯 What it does: A Transformable Patch Graph Neural Network (T-PATCHGNN) is proposed for the prediction of irregular multivariate time series.
Is DPO Superior to PPO for LLM Alignment? A Comprehensive Study
Shusheng Xu (Tsinghua University), Yi Wu
OptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: A comparative analysis of the performance of DPO and PPO in aligning large language models is conducted, systematically exploring the limitations of DPO and proposing key techniques to enhance the effectiveness of PPO.
Is Epistemic Uncertainty Faithfully Represented by Evidential Deep Learning Methods?
Mira Juergens, Willem Waegeman (Ghent University)
🎯 What it does: This paper systematically evaluates whether Evidential Deep Learning (EDL) methods can truly quantify epistemic uncertainty from both theoretical and experimental perspectives.
Is In-Context Learning in Large Language Models Bayesian? A Martingale Perspective
Fabian Falck (University of Oxford), Christopher C. Holmes
Large Language ModelText
🎯 What it does: This study investigates whether large language models (LLMs) approximate Bayesian inference in in-context learning (ICL) and conducts a theoretical examination using Markov properties.
Is Inverse Reinforcement Learning Harder than Standard Reinforcement Learning? A Theoretical Perspective
Lei Zhao (University of Science and Technology of China), Yu Bai (Salesforce AI Research)
Reinforcement Learning
🎯 What it does: Under both offline and online learning frameworks, two efficient algorithms for solving the Inverse Reinforcement Learning (IRL) problem (RLP and RLE) are designed and proven, along with their sample complexity upper and lower bounds. Stronger evaluation metrics are proposed, and theoretical guarantees for reward transfer are achieved.
Is Kernel Prediction More Powerful than Gating in Convolutional Neural Networks?
Lorenz K Muller
RestorationConvolutional Neural NetworkImage
🎯 What it does: This paper explores the equivalence and practicality of two methods for dynamic weight prediction in convolutional neural networks: HyperNetwork and gating mechanisms, and applies them to the task of image denoising.
Is Temperature Sample Efficient for Softmax Gaussian Mixture of Experts?
Huy Nguyen (University of Texas at Austin), Nhat Ho (University of Texas at Austin)
OptimizationMixture of ExpertsTabular
🎯 What it does: This paper studies the convergence properties of parameter estimation when using a temperature parameter in the dense-to-sparse softmax gating of the Gaussian Mixture of Experts model, and proposes an activation dense-to-sparse gating by adding an activation function before the softmax to improve sample efficiency.
Isometric Representation Learning for Disentangled Latent Space of Diffusion Models
Jaehoon Hahm (Seoul National University), Joonseok Lee (Google Research)
GenerationRepresentation LearningDiffusion modelImage
🎯 What it does: A new diffusion model called Isometric Diffusion is proposed, which guides the model to learn a geometrically reasonable latent space through geometric regularization, achieving better decoupling of the latent space.
Iterated Denoising Energy Matching for Sampling from Boltzmann Densities
Tara Akhound-Sadegh (Mila Quebec AI Institute), Alexander Tong (Mila Quebec AI Institute)
Diffusion modelGraphPhysics RelatedStochastic Differential Equation
🎯 What it does: An iterative denoising energy matching (iDEM) algorithm is proposed for efficient sampling from an unknown normalized Boltzmann distribution.
Iterative Data Smoothing: Mitigating Reward Overfitting and Overoptimization in RLHF
Banghua Zhu (University of California), Jiantao Jiao (University of California)
OptimizationReinforcement Learning from Human FeedbackReinforcement LearningText
🎯 What it does: An Iterative Data Smoothing (IDS) algorithm is proposed to address the issues of reward overfitting and over-optimization in RLHF.