ICLR 2024 Papers — Page 10
International Conference on Learning Representations · 2260 papers
Horizon-free Reinforcement Learning in Adversarial Linear Mixture MDPs
Kaixuan Ji (University of California), Quanquan Gu (University of California)
Reinforcement Learning
🎯 What it does: The HF-O-PS2 algorithm is proposed for achieving horizon-free reinforcement learning in linear mixture MDPs when facing adversarial rewards.
How connectivity structure shapes rich and lazy learning in neural circuits
Yuhan Helena Liu (University of Washington), Guillaume Lajoie (Mila - Quebec AI Institute)
Recurrent Neural NetworkSequential
🎯 What it does: This study investigates the impact of effective rank of initial weights on the learning patterns of neural networks. Through theoretical derivation and experiments on RNN cognitive tasks, it is found that low-rank initialization often leads to richer learning.
How do Language Models Bind Entities in Context?
Jiahai Feng (University of California Berkeley), Jacob Steinhardt (University of California Berkeley)
TransformerLarge Language ModelText
🎯 What it does: The study investigates how language models bind entities to attributes in context and identifies a general internal mechanism called binding ID.
How Do Transformers Learn In-Context Beyond Simple Functions? A Case Study on Learning with Representations
Tianyu Guo (University of California Berkeley), Yu Bai (Salesforce AI Research)
Representation LearningTransformerTabular
🎯 What it does: This paper explores the capabilities of Transformers in context learning tasks that involve representation functions, demonstrating that they can achieve representations at lower levels and linear in-context learning (ICL) at higher levels.
How Does Unlabeled Data Provably Help Out-of-Distribution Detection?
Xuefeng Du (University of Wisconsin-Madison), Yixuan Li (University of Wisconsin-Madison)
Anomaly DetectionImage
🎯 What it does: Proposes the SAL framework, which first filters candidate anomaly samples from unlabeled field data, and then trains a binary classification OOD detector using these samples along with labeled ID samples.
How I Warped Your Noise: a Temporally-Correlated Noise Prior for Diffusion Models
Pascal Chang (ETH Zurich), Vinicius C. Azevedo (ETH Zurich)
RestorationGenerationData SynthesisDiffusion modelOptical FlowImageVideo
🎯 What it does: A time-dependent noise prior (∫-noise) is proposed, which treats discrete noise as the integral of an infinitely resolved white noise field, and based on this, derives a noise transmission equation to achieve time deformation that preserves noise distribution in diffusion models.
How Many Pretraining Tasks Are Needed for In-Context Learning of Linear Regression?
Jingfeng Wu (University of California Berkeley), Peter Bartlett
TransformerTabular
🎯 What it does: This study investigates the contextual learning ability of a single-layer linear attention model under Gaussian priors in linear regression tasks with limited task pre-training.
How Over-Parameterization Slows Down Gradient Descent in Matrix Sensing: The Curses of Symmetry and Initialization
Nuoya Xiong (Tsinghua University), Simon Shaolei Du
Optimization
🎯 What it does: The study investigates the impact of parameterization on the convergence speed of gradient descent in matrix sensing, providing upper and lower bounds for symmetric and asymmetric parameterization, and proposing a single-step adjustment that makes the convergence rate independent of the initialization scale.
How Realistic Is Your Synthetic Data? Constraining Deep Generative Models for Tabular Data
Mihaela C Stoian, Eleonora Giunchiglia (Vienna University of Technology)
GenerationData SynthesisGenerative Adversarial NetworkTabular
🎯 What it does: This paper proposes a method to automatically convert linear inequality constraints into a differentiable constraint layer and embed it into various deep generative models (GAN, CTGAN, TableGAN, TVAE, GOGGLE) to generate synthetic tabular data that satisfies the constraints.
How to Capture Higher-order Correlations? Generalizing Matrix Softmax Attention to Kronecker Computation
Josh Alman (Columbia University), Zhao Song (Adobe Research)
Computational EfficiencyTransformer
🎯 What it does: This study proposes an extension of the attention mechanism aimed at capturing the correlations between triplets, addressing the issue that traditional transformer attention mechanisms cannot detect triplet connections.
How to Catch an AI Liar: Lie Detection in Black-Box LLMs by Asking Unrelated Questions
Lorenzo Pacchiardi (University of Oxford), Jan M. Brauner (University of Oxford)
ClassificationTransformerLarge Language ModelText
🎯 What it does: A lie detection method is proposed that only uses black-box access, based on asking the model a set of yes/no questions unrelated to lies, and then using logistic regression to discriminate the response patterns.
How to Fine-Tune Vision Models with SGD
Ananya Kumar (Stanford University), Suriya Gunasekar (Microsoft)
OptimizationComputational EfficiencyConvolutional Neural NetworkTransformerSupervised Fine-TuningImage
🎯 What it does: This paper studies the performance differences between SGD and AdamW in fine-tuning visual models and proposes a simple technique of freezing the embedding layer, which significantly enhances the robustness and efficiency of SGD.
How Well Do Supervised 3D Models Transfer to Medical Imaging Tasks?
Wenxuan Li (Johns Hopkins University), Zongwei Zhou (Johns Hopkins University)
ClassificationSegmentationConvolutional Neural NetworkTransformerSupervised Fine-TuningImageBiomedical DataComputed Tomography
🎯 What it does: We constructed and released the largest 3D CT dataset, AbdomenAtlas 1.1 (9,262 CT scans, 25 anatomical structures, and 7 tumor pseudo-labels), and trained and publicly released a series of supervised pre-trained models called SuPreM using this dataset. We then evaluated their transfer learning performance on various 3D medical image segmentation and classification tasks.
Human Feedback is not Gold Standard
Tom Hosking (University of Edinburgh), Max Bartolo (Cohere)
Reinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: By conducting human evaluations on the text generated by large language models (LLMs), this study explores whether a single preference score can cover all important types of errors and analyzes potential biases and confounding factors in the annotations.
Human Motion Diffusion as a Generative Prior
Yoni Shafir, Amit Haim Bermano
GenerationData SynthesisTransformerDiffusion modelVideoSequential
🎯 What it does: This paper proposes three combination methods based on pre-trained motion diffusion models: sequential combination (DoubleTake) for generating actions of arbitrary length; parallel combination (ComMDM) for achieving two-person interactive actions; and model combination (DiffusionBlending) for multi-dimensional fine-grained control.
Hybrid Directional Graph Neural Network for Molecules
Junyi An (Nanjing University), Furao Shen (Nanjing University)
Drug DiscoveryGraph Neural NetworkGraph
🎯 What it does: This paper proposes the Hybrid Directional Graph Neural Network (HDGNN), which enhances molecular property prediction performance by integrating strict equivariant operations with learnable modules to balance equivariance and expressive power.
Hybrid Distillation: Connecting Masked Autoencoders with Contrastive Learners
Bowen Shi (Shanghai Jiao Tong University), Qi Tian (Huawei Inc.)
Object DetectionSegmentationKnowledge DistillationTransformerAuto EncoderContrastive LearningImage
🎯 What it does: Proposes the Hybrid Distill framework, which utilizes both CL/supervised teacher and MIM teacher to guide the student model's learning;
Hybrid Internal Model: Learning Agile Legged Locomotion with Simulated Robot Response
Junfeng Long (Shanghai AI Laboratory), Jiangmiao Pang (Shanghai AI Laboratory)
Robotic IntelligenceReinforcement LearningContrastive Learning
🎯 What it does: This paper proposes the Hybrid Internal Model (HIM), which utilizes the robot's own responses to estimate the dynamics of the external environment, thereby achieving global walking control for legged robots without relying on external sensors.
Hybrid LLM: Cost-Efficient and Quality-Aware Query Routing
Dujian Ding (University of British Columbia), Ahmed Hassan Awadallah
TransformerLarge Language ModelMixture of ExpertsText
🎯 What it does: A quality-aware hybrid large language model inference framework is proposed, which allocates 'easy' queries to small models and 'difficult' queries to large models through a router, in order to reduce costs while maintaining high quality;
Hybrid Sharing for Multi-Label Image Classification
Zihao Yin (Nanjing University), Junfeng Zhang (Nanjing University)
ClassificationRecognitionTransformerMixture of ExpertsImage
🎯 What it does: This paper proposes a multi-label image classification framework called Hybrid Sharing Query (HSQ), which combines task-specific experts with shared experts using a Mixture of Experts (MoE) structure and performs adaptive fusion through a gating network to reduce negative transfer caused by label heterogeneity, thereby improving the recognition performance both overall and for each label.
HypeBoy: Generative Self-Supervised Representation Learning on Hypergraphs
Sunwoo Kim (Korea Advanced Institute of Science and Technology), Kijung Shin (Korea Advanced Institute of Science and Technology)
Representation LearningGraph Neural NetworkGraph
🎯 What it does: A generative self-supervised learning task based on hyperedge filling is designed, and the corresponding pre-training framework HYPEBOY is implemented to enhance the performance of hypergraph neural networks in node classification and hyperedge prediction tasks.
Hyper Evidential Deep Learning to Quantify Composite Classification Uncertainty
Changbin Li (University of Texas at Dallas), Feng Chen (Virginia Tech)
ClassificationConvolutional Neural NetworkImage
🎯 What it does: A deep network framework capable of handling composite category labels present in the training set is proposed—Hyper Evidential Neural Network (HENN). This framework quantifies the confidence and uncertainty (especially ambiguity) of classification results by learning to predict hyper-opinions, and it can provide predictions for single or composite categories.
HyperAttention: Long-context Attention in Near-Linear Time
Insu Han (Yale University), Amir Zandieh (Independent Researcher)
OptimizationComputational EfficiencyTransformerLarge Language ModelText
🎯 What it does: Designed HyperAttention, a near-linear time attention approximation algorithm that can accelerate the self-attention computation of Transformers under large contexts.
Hypergraph Dynamic System
Jielong Yan (Tsinghua University), Yue Gao (Tsinghua University)
ClassificationGraph Neural NetworkGraphOrdinary Differential Equation
🎯 What it does: A control-diffusion hypergraph dynamic system (HDS ODE) is proposed, achieving controllable and stable diffusion of vertex and hyperedge representations in the hypergraph through a continuous ODE form, and implemented as a trainable multi-layer network for vertex classification.
HyperHuman: Hyper-Realistic Human Generation with Latent Structural Diffusion
Xian Liu (Chinese University of Hong Kong), Sergey Tulyakov (Snap Inc)
GenerationData SynthesisPose EstimationDepth EstimationDiffusion modelImage
🎯 What it does: Proposes the HyperHuman framework, which utilizes a joint Latent Structural Diffusion Model and Structure-Guided Refiner to achieve controllable generation from text + skeleton to high-resolution portraits;
HYPO: Hyperspherical Out-Of-Distribution Generalization
Haoyue Bai (University of Wisconsin Madison), Yixuan Li (University of Wisconsin Madison)
Domain AdaptationRepresentation LearningImage
🎯 What it does: Proposes the HYPO framework, which learns domain-invariant representations on the unit hypersphere to reduce intra-class domain variation and maximize inter-class separation.
Hypothesis Search: Inductive Reasoning with Language Models
Ruocheng Wang (Stanford University), Noah Goodman
AI Code AssistantTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: A reasoning pipeline based on large language models is proposed: first, the model generates natural language hypotheses, then these hypotheses are converted into executable Python code to verify whether the code meets the given training examples, and finally, the code is used for reasoning on new inputs.
I-PHYRE: Interactive Physical Reasoning
Shiqian Li (Peking University), Yixin Zhu (Peking University)
Robotic IntelligenceReinforcement LearningBenchmarkPhysics Related
🎯 What it does: An interactive physical reasoning benchmark I-PHYRE is proposed and implemented to evaluate agents' physical reasoning capabilities under multi-step and immediate interventions.
IceFormer: Accelerated Inference with Long-Sequence Transformers on CPUs
Yuzhen Mao (Simon Fraser University), Ke Li (Simon Fraser University)
Computational EfficiencyTransformerSequentialBenchmark
🎯 What it does: A method called IceFormer is proposed to accelerate the inference of long sequence Transformer models on CPUs, addressing the computational complexity issues of existing Transformer models when handling long sequences.
IDEAL: Influence-Driven Selective Annotations Empower In-Context Learners in Large Language Models
Shaokun Zhang (Pennsylvania State University), Tongliang Liu (University of Sydney)
ClassificationOptimizationTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: This paper proposes an influence-driven selective labeling method called IDEAL, which is used to select a small subset of examples for manual labeling from a large pool of unlabeled data, in order to reduce labeling costs in in-context learning (ICL) and improve model performance.
Idempotence and Perceptual Image Compression
Tongda Xu (Tsinghua University), Ya-Qin Zhang (Tsinghua University)
CompressionDiffusion modelGenerative Adversarial NetworkImage
🎯 What it does: A new perceptual image compression method is proposed, utilizing a pre-trained unconditional generative model for inversion while keeping the bitstream unchanged, and incorporating idempotency constraints;
Idempotent Generative Network
Assaf Shocher (Google Research), Alexei A Efros
GenerationData SynthesisAuto EncoderGenerative Adversarial NetworkImage
🎯 What it does: This paper proposes an Idempotent Generative Network (IGN) that trains the network to project noise or any input to the target distribution (such as natural images) in a single forward pass, supporting multiple iterations for refinement. It can also directly map data from different distributions, such as distorted images, grayscale, or sketches, onto the learned image manifold.
Identifiable Latent Polynomial Causal Models through the Lens of Change
Yuhang Liu (Australian Institute for Machine Learning, University of Adelaide), Javen Qinfeng Shi (Australian Institute for Machine Learning, University of Adelaide)
ImageBiomedical DataMagnetic Resonance Imaging
🎯 What it does: This paper extends the scope of potential causal models, exploring nonlinear causal relationships and proposing a new empirical estimation method to learn consistent potential causal representations.
Identifying Policy Gradient Subspaces
Jan Schneider (Max Planck Institute for Intelligent Systems), Dieter Büchler (Max Planck Institute for Intelligent Systems)
OptimizationReinforcement LearningSequential
🎯 What it does: The study investigates whether the gradient in policy gradient-based deep reinforcement learning (PPO, SAC) focuses on a low-dimensional subspace in a few high-curvature directions and examines the stability of this subspace during the training process.
Identifying Representations for Intervention Extrapolation
Sorawit Saengkyongam (ETH Zurich), Jonas Peters (ETH Zurich)
Representation LearningAuto EncoderTabular
🎯 What it does: This paper proposes the Rep4Ex framework, which achieves predictions of unobserved intervention extrapolation through identifiable latent representations.
Identifying the Risks of LM Agents with an LM-Emulated Sandbox
Yangjun Ruan (University of Toronto), Tatsunori Hashimoto (Stanford University)
OptimizationSafty and PrivacyTransformerLarge Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: The ToolEmu framework is proposed, utilizing large language models (such as GPT-4) to simulate tool execution and sandboxes, thereby enabling large-scale and scalable assessment of safety risks for language model agents. An automated safety evaluator and a helpful evaluator were also constructed, and based on this, a safety benchmark was created that includes 36 toolkits and 144 test cases.
iGraphMix: Input Graph Mixup Method for Node Classification
Jongwon Jeong (KRAFTON), Kim Jin Seon
ClassificationData SynthesisGraph Neural NetworkGraph
🎯 What it does: This paper proposes an input Mixup method called iGraphMix, specifically designed for node classification tasks in graph neural networks, to synthesize virtual nodes and their adjacency relationships during training, thereby achieving data augmentation.
Illusory Attacks: Information-theoretic detectability matters in adversarial attacks
Tim Franzmeyer (University of Oxford), Christian Schroeder de Witt (University of Oxford)
Adversarial AttackReinforcement LearningSequential
🎯 What it does: A new observation space adversarial attack called illusionary attack is proposed, utilizing information-theoretic detectability constraints to make the attack both effective and difficult to be automatically detected or recognized by humans.
Image Background Serves as Good Proxy for Out-of-distribution Data
Sen Pei (ByteDance Inc)
ClassificationAnomaly DetectionConvolutional Neural NetworkContrastive LearningImage
🎯 What it does: A self-supervised sampling-based OOD detection framework SSOD is proposed, which can generate natural OOD supervision signals using the background information of ID images without collecting explicit OOD samples, and achieve end-to-end training.
Image Clustering Conditioned on Text Criteria
Sehyun Kwon (Seoul National University), Kangwook Lee (University of Wisconsin-Madison)
ClassificationRecognitionTransformerLarge Language ModelPrompt EngineeringVision Language ModelImageText
🎯 What it does: A text-conditioned image clustering method called IC TC is proposed, allowing users to directly control clustering results through natural language descriptions.
Image Clustering via the Principle of Rate Reduction in the Age of Pretrained Models
Tianzhe Chu (University of California), Yi Ma (Hong Kong University)
Representation LearningTransformerContrastive LearningImage
🎯 What it does: This paper proposes an image clustering pipeline CPP that utilizes features from the large-scale pre-trained model CLIP, combined with the principle of Maximum Coding Rate Reduction (MCR).
Image Inpainting via Iteratively Decoupled Probabilistic Modeling
Wenbo Li (Huawei Noah's Ark Lab), Zhe Lin (Huawei Noah's Ark Lab)
RestorationGenerationConvolutional Neural NetworkGenerative Adversarial NetworkImage
🎯 What it does: The Pixel Spread Model (PSM) is proposed, which completes large gap image inpainting through iterative propagation of reliable pixels.
Image Inpainting via Tractable Steering of Diffusion Models
Anji Liu (University of California), Guy Van den Broeck (University of California)
RestorationGenerationDiffusion modelGenerative Adversarial NetworkImage
🎯 What it does: This paper proposes the use of Probabilistic Circuits to guide the sampling process of diffusion models, achieving high-quality image inpainting.
Image Translation as Diffusion Visual Programmers
Cheng Han (Rochester Institute of Technology), Dongfang Liu (University of California, Los Angeles)
Image TranslationTransformerLarge Language ModelDiffusion modelImageText
🎯 What it does: This paper proposes the Diffusion Visual Programmer (DVP), which combines a conditionally flexible diffusion model with GPT visual programming to achieve controllable and interpretable image translation and local editing.
Image2Sentence based Asymmetrical Zero-shot Composed Image Retrieval
Yongchao Du (University of Science and Technology of China), Houqiang Li (University of Science and Technology of China)
RetrievalKnowledge DistillationTransformerVision Language ModelContrastive LearningImageTextMultimodality
🎯 What it does: Proposes an Asymmetric Zero-Shot Synthetic Image Retrieval (ISA) framework based on Image2Sentence;
ImageNet-OOD: Deciphering Modern Out-of-Distribution Detection Algorithms
William Yang (Princeton University), Olga Russakovsky (Princeton University)
Anomaly DetectionConvolutional Neural NetworkImageBenchmark
🎯 What it does: A dataset for ImageNet-OOD has been constructed, and various OOD detection methods have been evaluated.
ImagenHub: Standardizing the evaluation of conditional image generation models
Max Ku (University of Waterloo), Wenhu Chen (University of Waterloo)
GenerationData SynthesisImageBenchmark
🎯 What it does: This paper presents ImagenHub, a library for standardized evaluation of conditional image generation models, aimed at addressing inconsistencies in existing model evaluations.
Imitation Learning from Observation with Automatic Discount Scheduling
Yuyang Liu (Tsinghua University), Yang Gao (Tsinghua University)
Robotic IntelligenceReinforcement Learning
🎯 What it does: A new imitation learning framework is proposed, aimed at addressing the problem of learning from observation (ILfO), particularly for tasks with progress-dependent characteristics.
Impact of Computation in Integral Reinforcement Learning for Continuous-Time Control
Wenhan Cao (Tsinghua University), Wei Pan (University of Manchester)
OptimizationReinforcement LearningTime Series
🎯 What it does: This study investigates the impact of using different numerical integration methods (trapezoidal rule and Bayesian quadrature) on control performance during the policy evaluation phase of Integral Reinforcement Learning (IntRL) under continuous-time control, and provides convergence rate analysis and simulation validation.
Implicit bias of SGD in $L_2$-regularized linear DNNs: One-way jumps from high to low rank
Zihan Wang (New York University), Arthur Jacot (New York University)
OptimizationTabular
🎯 What it does: The study investigates the training of deep linear networks on matrix completion tasks under L2 regularization, where the implicit bias of SGD causes parameters to transition unidirectionally from high-rank critical points to low-rank critical points.
Implicit Gaussian process representation of vector fields over arbitrary latent manifolds
Robert Peach, Adam Gosztolai (EPFL)
ClassificationRestorationSuper ResolutionMeshBiomedical DataAlzheimer's Disease
🎯 What it does: This paper studies a new Gaussian Process (GP) model—RVGP—for learning vector fields on unknown Riemannian manifolds and performing super-resolution and missing region completion.
Implicit Maximum a Posteriori Filtering via Adaptive Optimization
Gianluca Bencomo, Thomas L. Griffiths (Princeton University)
OptimizationConvolutional Neural NetworkImageTime Series
🎯 What it does: This paper proposes Implicit Maximum A Posteriori (IMAP) filtering, treating the Bayesian filtering problem as an optimization of a time-varying objective function. It utilizes adaptive optimizers (such as Adam, RMSProp, etc.) to perform K steps of gradient updates at each time step to obtain state estimates.
Implicit Neural Representation Inference for Low-Dimensional Bayesian Deep Learning
Panagiotis Dimitrakopoulos (University of Ioannina), Christophoros Nikou (University of Ioannina)
Tabular
🎯 What it does: A low-dimensional Bayesian deep learning framework is proposed: the network weights are decomposed into a deterministic part w and a product of a random factor ξ encoded by implicit neural representations (INR); Bayesian inference is performed on the INR parameters (Laplace, SWAG, or RealNVP), and MAP inference is conducted on w, thereby achieving probabilistic inference and uncertainty estimation for the entire network.
Implicit Neural Representations and the Algebra of Complex Wavelets
T Mitchell Roddenberry, Richard Baraniuk
RestorationImageTime Series
🎯 What it does: This paper studies the use of complex wavelet activation functions in implicit neural representations (INR) and proposes a network structure based on segmentation (scaling + wavelet), combined with waveform modal maxima (WMM) initialization, to enhance signal approximation and image denoising effects.
Implicit regularization of deep residual networks towards neural ODEs
Pierre Marion (Sorbonne Université), Gérard Biau (Sorbonne Université)
ImageOrdinary Differential Equation
🎯 What it does: This study investigates the training dynamics of deep residual networks, proving that with appropriate scaling and weight initialization, as the network depth approaches infinity, the gradient flow training process remains a discretization of neural ODEs, achieving global convergence under wide networks and the Polyak-Łojasiewicz condition.
ImplicitSLIM and How it Improves Embedding-based Collaborative Filtering
Ilya Shenbin (Steklov Mathematical Institute of Russian Academy of Sciences), Sergey Nikolenko (Steklov Mathematical Institute of Russian Academy of Sciences)
Recommendation SystemTabular
🎯 What it does: This paper proposes ImplicitSLIM, an unsupervised learning method that extracts embeddings from SLIM-like models for collaborative filtering.
Improved Active Learning via Dependent Leverage Score Sampling
Atsushi Shimizu (New York University), Jonathan Weare (New York University)
TabularPhysics Related
🎯 What it does: This paper proposes an active learning method that combines leverage scores and Pivotal sampling, targeting linear regression problems in low-dimensional space structures (especially polynomial approximation), which can effectively select samples without observing all labels.
Improved algorithm and bounds for successive projection
Jiashun Jin (Carnegie Mellon University), Jingming Wang (Harvard University)
OptimizationSupervised Fine-TuningPoint Cloud
🎯 What it does: An improved vertex finding algorithm, pp-SPA, is proposed to address the performance degradation of traditional SPA under strong noise or outliers.
Improved Analysis of Sparse Linear Regression in Local Differential Privacy Model
Liyang Zhu (King Abdullah University of Science and Technology), Di Wang (King Abdullah University of Science and Technology)
OptimizationSafty and Privacy
🎯 What it does: This paper studies the estimation error of sparse linear regression under the local differential privacy model, providing lower and upper bounds in the k-sparse case, and proposes the first efficient non-interactive algorithm.
Improved Efficiency Based on Learned Saccade and Continuous Scene Reconstruction From Foveated Visual Sampling
Jiayang Liu (Syracuse University), Qinru Qiu (Syracuse University)
ClassificationRecognitionComputational EfficiencyConvolutional Neural NetworkRecurrent Neural NetworkReinforcement LearningImage
🎯 What it does: This study proposes an active scene reconstruction architecture based on the human eye's retinal perspective and saccadic mechanism, utilizing multiple viewpoint sampling (high-resolution central vision + sparse peripheral sampling) and reinforcement learning-controlled saccadic actions to reconstruct the original image while maintaining classification accuracy, and reducing pixel usage by approximately 90%.
Improved Probabilistic Image-Text Representations
Sanghyuk Chun (NAVER AI Lab)
RetrievalContrastive LearningImageText
🎯 What it does: An improved probabilistic image-text matching model PCME++ is proposed, which captures the many-to-many relationships and uncertainties brought by label sparsity through probabilistic embedding.
Improved Regret Bounds for Non-Convex Online-Within-Online Meta Learning
Jiechao Guan (Hong Kong University of Science and Technology), Hui Xiong (Hong Kong University of Science and Technology)
OptimizationMeta Learning
🎯 What it does: This study investigates the improved regret bounds for initialization and step size in non-convex online meta-learning (OWO), providing stronger theoretical guarantees.
Improved sampling via learned diffusions
Lorenz Richter (Zuse Institute Berlin), Julius Berner (California Institute of Technology)
Diffusion modelStochastic Differential Equation
🎯 What it does: A unified perspective on path space is proposed, utilizing time-reversed controlled diffusion processes to achieve sampling from prior distributions to target distributions, and extending it to arbitrary path space divergences; within this framework, a new log-variance loss is introduced to address issues such as mode collapse and high variance associated with traditional KL loss.
Improved statistical and computational complexity of the mean-field Langevin dynamics under structured data
Atsushi Nitanda (Agency for Science, Technology and Research), Denny Wu (New York University)
OptimizationComputational EfficiencyTabularStochastic Differential Equation
🎯 What it does: This study investigates the learning complexity of training a two-layer neural network to learn k-sparse parity functions under structured (anisotropic) data using Mean Field Langevin Dynamics (MFLD), and provides theoretical guarantees for convergence and generalization.
Improved Techniques for Training Consistency Models
Yang Song (OpenAI), Prafulla Dhariwal (OpenAI)
GenerationData SynthesisDiffusion modelImage
🎯 What it does: An improved consistency model training (iCT) method is proposed, enabling the model to learn high-quality image samples directly from raw data without using auxiliary networks such as knowledge distillation or LPIPS.
Improving Convergence and Generalization Using Parameter Symmetries
Bo Zhao (University of California San Diego), Rose Yu (University of California San Diego)
OptimizationMeta LearningImage
🎯 What it does: This paper proposes a method of symmetric transformation in parameter space (Teleportation), which accelerates convergence and improves generalization by jumping to points with larger gradients on the level surfaces during the gradient descent process.
Improving Domain Generalization with Domain Relations
Huaxiu Yao (Stanford University), Chelsea Finn (Stanford University)
Domain AdaptationDrug DiscoverySupervised Fine-TuningImageTabular
🎯 What it does: This paper proposes a method to improve domain transfer generalization using domain relationships, called D G. It first learns a set of domain-specific models during the training phase, and then infers by weighting these models based on the similarity between domains during the testing phase.
Improving equilibrium propagation without weight symmetry through Jacobian homeostasis
Axel Laborieux (Friedrich Miescher Institute for Biomedical Research), Friedemann Zenke (University of Basel)
ClassificationOptimizationConvolutional Neural NetworkImageOrdinary Differential Equation
🎯 What it does: A general framework for using holographic equipotential propagation (hEP) under the condition of no weight symmetry is proposed, and the gradient estimation bias is reduced through Jacobian homeostasis.
Improving Generalization of Alignment with Human Preferences through Group Invariant Learning
Rui Zheng (Fudan University), Xuanjing Huang (Fudan University)
OptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: A framework is proposed that automatically partitions data through group-invariant learning and dynamically adjusts KL penalties during RLHF training to enhance the generalization and training stability of language models in different domains.
Improving Intrinsic Exploration by Creating Stationary Objectives
Roger Creus Castanyer (Mila Quebec AI Institute), Glen Berseth (Mila Quebec AI Institute)
Reinforcement LearningImage
🎯 What it does: This paper proposes the SOFE (Stationary Objectives for Exploration) framework, which transforms the originally non-stationary exploration rewards into stationary rewards by incorporating sufficient statistics of exploration rewards (such as counts, pseudo-counts, or generative model parameters) into the state, making POMDPs observable MDPs, thereby enabling deep RL to learn exploration strategies more stably and efficiently.
Improving LoRA in Privacy-preserving Federated Learning
Youbang Sun (Northeastern University), Bolin Ding (Alibaba Group)
Federated LearningSafty and PrivacyComputational EfficiencyTransformerLarge Language ModelImageText
🎯 What it does: This study investigates the performance of LoRA in privacy-preserving federated learning and proposes the FFA-LoRA method.
Improving Non-Transferable Representation Learning by Harnessing Content and Style
Ziming Hong (University of Sydney), Tongliang Liu (University of Sydney)
Domain AdaptationKnowledge DistillationRepresentation LearningAuto EncoderImage
🎯 What it does: This paper proposes an H-NTL method that separates content (C) from style (S) through a causal model and trains the model using dual-path knowledge distillation, achieving high performance in the source domain while significantly decreasing performance in the target domain, thus realizing non-transferable representation learning.
Improving Offline RL by Blending Heuristics
Sinong Geng (Princeton University), Ching-An Cheng (Microsoft Research)
Reinforcement LearningTabular
🎯 What it does: Proposes HUBL, which partially replaces Bootstrapping by mixing heuristic values from Monte Carlo returns into the Bellman operator of offline RL, rewriting rewards and discounts.
Improving protein optimization with smoothed fitness landscapes
Andrew Kirjner (Massachusetts Institute of Technology), Ila R Fiete
OptimizationDrug DiscoveryConvolutional Neural NetworkBiomedical DataBenchmark
🎯 What it does: A protein fitness landscape smoothing method based on graph signal processing is proposed, and the GGS algorithm is constructed in conjunction with Gibbs sampling for protein optimization under limited noise data.
Improving the Convergence of Dynamic NeRFs via Optimal Transport
Sameera Ramasinghe (Amazon), Anton van den Hengel (Amazon)
GenerationOptimizationNeural Radiance FieldOptical FlowImage
🎯 What it does: This paper proposes a lightweight regularization based on optimal transport (sliced Wasserstein) to improve the convergence and rendering quality of dynamic NeRF, achieving this without the need for additional networks or preprocessing, making it easy to integrate into any existing architecture.
IMPUS: Image Morphing with Perceptually-Uniform Sampling Using Diffusion Models
Zhaoyuan Yang (General Electric Research), Richard Hartley (Australian National University)
GenerationData SynthesisDiffusion modelImageOrdinary Differential Equation
🎯 What it does: This paper proposes an image deformation method called IMPUS based on diffusion models, which achieves smooth, direct, and realistic transitions between two images through linear interpolation in the optimized CLIP text embedding space and latent space, combined with probabilistic flow ODE.
In defense of parameter sharing for model-compression
Aditya Desai (Rice University), Anshumali Shrivastava (Rice University)
CompressionConvolutional Neural NetworkImage
🎯 What it does: Evaluate the memory-accuracy trade-off of parameter sharing and pruning in model compression, and propose an improved version of stochastic parameter sharing (STABLE-RPS) to address the stability and Pareto continuity issues of the existing ROAST method.
In-context Autoencoder for Context Compression in a Large Language Model
Tao Ge (Microsoft Corporation), Furu Wei (Microsoft Corporation)
GenerationCompressionTransformerLarge Language ModelPrompt EngineeringAuto EncoderText
🎯 What it does: Designed and implemented the In-Context Autoencoder (ICAE), which uses a lightweight LoRA encoder to compress long contexts into a small number of memory slots, allowing the original LLM to be used as a decoder to complete generation tasks directly with these slots.
In-context Exploration-Exploitation for Reinforcement Learning
Zhenwen Dai (Spotify Research), Sina Ghiassian (Spotify Research)
TransformerReinforcement LearningSequential
🎯 What it does: An In-Context Exploration-Exploitation (ICEE) algorithm is proposed for online policy learning during inference.
In-Context Learning Dynamics with Random Binary Sequences
Eric J Bigelow, Tomer Ullman
TransformerLarge Language ModelText
🎯 What it does: By using random binary sequences as context in large language models, this study investigates the dynamics of context learning and explores the model's learning behavior regarding subjective randomness and formal languages.
In-Context Learning Learns Label Relationships but Is Not Conventional Learning
Jannik Kossen (University of Oxford), Tom Rainforth (University of Oxford)
ClassificationTransformerLarge Language ModelTextFinance Related
🎯 What it does: Systematically evaluate the dependence of in-context learning (ICL) on label information in large-scale language models, exploring how models utilize example labels to infer input-label relationships;
In-Context Learning through the Bayesian Prism
Madhur Panwar (Microsoft Research), Navin Goyal (Microsoft Research)
TransformerLarge Language Model
🎯 What it does: This paper experimentally explores the Bayesian perspective on in-context learning (ICL) in large-scale language models, extending to a hierarchical ICL (HMICL) setting with multi-task mixing. It verifies whether Transformers can approximate Bayesian predictors across various families of linear and nonlinear functions (such as linear regression, sparse regression, neural networks, decision trees, Fourier series, quadratic monomials, GMMs, etc.) and studies their generalization and simplification bias on unseen tasks.
In-Context Pretraining: Language Modeling Beyond Document Boundaries
Weijia Shi (Meta AI), Mike Lewis (Meta AI)
GenerationRetrievalTransformerLarge Language ModelTextRetrieval-Augmented Generation
🎯 What it does: A pre-training strategy called IN-CONTEXT PRETRAINING is proposed, allowing the language model to concatenate semantically related documents in order during pre-training, enabling cross-document reasoning and referencing during generation.
Incentive-Aware Federated Learning with Training-Time Model Rewards
Zhaoxuan Wu (National University of Singapore), Bryan Kian Hsiang Low (National University of Singapore)
Federated LearningImageText
🎯 What it does: An incentive mechanism IAFL based on model rewards during training is proposed, along with its implementation and theoretical analysis;
Incentivized Truthful Communication for Federated Bandits
Zhepei Wei (University of Virginia), Hongning Wang (University of Virginia)
OptimizationFederated Learning
🎯 What it does: This paper proposes an incentive-compatible federated bandit communication protocol named TRUTH-FEDBAN, which allows distributed clients to achieve truthful cost reporting through incentive payments while participating in federated learning.
Increasing Model Capacity for Free: A Simple Strategy for Parameter Efficient Fine-tuning
Haobo SONG, Tao Lin (Westlake University)
CompressionOptimizationTransformerSupervised Fine-TuningText
🎯 What it does: This paper proposes a plugin framework named CAPABOOST, which implements parallel multi-branching of weights by using random binary masks for sparse masking of shared weights in the PEFT module, thereby enhancing the effective rank and capacity of the model without increasing parameters or FLOP.
Incremental Randomized Smoothing Certification
Shubham Ugare (University of Illinois Urbana-Champaign), Sasa Misailovic (University of Illinois Urbana-Champaign)
ClassificationOptimizationComputational EfficiencyConvolutional Neural NetworkImage
🎯 What it does: The Incremental Randomized Smoothing (IRS) method is proposed, which can quickly re-verify the robustness of an already certified DNN after approximate modifications such as quantization or pruning.
Independent-Set Design of Experiments for Estimating Treatment and Spillover Effects under Network Interference
Chencheng Cai (Washington State University), Edoardo Airoldi
OptimizationGraph Neural NetworkGraph
🎯 What it does: This paper proposes an experimental design based on network independent sets to estimate treatment effects and spillover effects in network experiments with interference.
Inducing High Energy-Latency of Large Vision-Language Models with Verbose Images
Kuofeng Gao (Tsinghua University), Wei Liu (Tencent Data Platform)
GenerationComputational EfficiencyAdversarial AttackTransformerVision Language ModelImageText
🎯 What it does: A method for energy consumption and latency attacks on large-scale visual-language models (VLM) has been designed and implemented—verbose images, which induce the model to generate extremely long text sequences, significantly increasing energy consumption and latency during inference.
Influencer Backdoor Attack on Semantic Segmentation
Haoheng Lan (Dartmouth College), Hengshuang Zhao (The University of Hong Kong)
SegmentationAdversarial AttackConvolutional Neural NetworkTransformerImage
🎯 What it does: This study investigates backdoor attacks on semantic segmentation models and proposes the Influencer Backdoor Attack (IBA), which induces the model to misclassify target categories by injecting small triggers on non-target pixels.
InfoBatch: Lossless Training Speed Up by Unbiased Dynamic Data Pruning
Ziheng Qin (National University of Singapore), Yang You (National University of Singapore)
OptimizationComputational EfficiencyConvolutional Neural NetworkTransformerSupervised Fine-TuningImage
🎯 What it does: The InfoBatch framework is proposed, achieving lossless training acceleration through dynamic unbiased sample pruning and weight recalibration.
InfoCon: Concept Discovery with Generative and Discriminative Informativeness
Ruizhe Liu (Peking University), Yanchao Yang (University of Hong Kong)
Robotic IntelligenceTransformerReinforcement LearningSequentialBenchmark
🎯 What it does: This paper proposes InfoCon, a self-supervised method that automatically discovers manipulation concepts from unlabeled demonstration trajectories through generative and discriminative information metrics, segments the trajectories into subsequences, generates key states, and subsequently uses these concepts to guide policy training.
Information Bottleneck Analysis of Deep Neural Networks via Lossy Compression
Ivan Butakov (Skolkovo Institute of Science and Technology), Kirill Andreev (Skolkovo Institute of Science and Technology)
CompressionRepresentation LearningConvolutional Neural NetworkAuto EncoderImage
🎯 What it does: A framework for mutual information estimation based on random networks and compression techniques is proposed, and this framework is applied to perform information bottleneck analysis on the MNIST convolutional network.
Information Retention via Learning Supplemental Features
Zhipeng Xie (Fudan University), Yahe Li (Fudan University)
ClassificationDomain AdaptationConvolutional Neural NetworkSupervised Fine-TuningImageTextTabular
🎯 What it does: Proposes the principle of information retention and designs a three-stage supervised learning framework InfoR‑LSF to simultaneously learn main features and supplementary features, aiming to retain as much relevant information as possible.
Inherently Interpretable Time Series Classification via Multiple Instance Learning
Joseph Early (University of Southampton), Niall Twomey (Amazon Prime Video)
ClassificationExplainability and InterpretabilityConvolutional Neural NetworkTime Series
🎯 What it does: A framework called MILLET is proposed, which integrates Multiple Instance Learning (MIL) into Time Series Classification (TSC), allowing the model itself to provide interpretable point predictions, thereby enhancing interpretability.
Initializing Models with Larger Ones
Zhiqiu Xu (University of Pennsylvania), Zhuang Liu (Meta AI Research)
ClassificationComputational EfficiencyKnowledge DistillationImage
🎯 What it does: A method is proposed to initialize a small model by selecting a subset of weights from a large model.
Inner Classifier-Free Guidance and Its Taylor Expansion for Diffusion Models
Shikun Sun (Tsinghua University), Qi Tian (Huawei Inc.)
GenerationDiffusion modelImageStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: Proposes the Inner Classifier Free Guidance (ICFG) and its second-order Taylor expansion to improve the guidance strategy of diffusion models;
Input-gradient space particle inference for neural network ensembles
Trung Trinh (Aalto University), Samuel Kaski (Aalto University)
ClassificationOptimizationImage
🎯 What it does: A molecular particle variational inference method called FoRDE is proposed, which enhances the functional diversity of neural network ensembles based on the input gradient space.
Ins-DetCLIP: Aligning Detection Model to Follow Human-Language Instruction
Renjie Pi (Hong Kong University of Science and Technology), Hang Xu (Hong Kong University of Science and Technology)
Object DetectionTransformerLarge Language ModelVision Language ModelContrastive LearningImageTextMultimodalityBenchmark
🎯 What it does: This paper proposes the Instruction-Oriented Object Detection (IOD) task and constructs the IOD-Bench dataset, followed by the design of the Ins-DetCLIP model to achieve object detection based on natural language instructions.
InsertNeRF: Instilling Generalizability into NeRF with HyperNet Modules
Yanqi Bao (Nanjing University), Yang Gao (Nanjing University)
GenerationData SynthesisNeural Radiance FieldImage
🎯 What it does: By inserting a pluggable HyperNet module into the NeRF framework, dynamic generation of scene-specific network weights is achieved using reference scene features, thereby enabling the generalization of NeRF.