NeurIPS 2024 Papers — Page 18
Conference on Neural Information Processing Systems · 4035 papers
Identifying Functionally Important Features with End-to-End Sparse Dictionary Learning
Dan Braun (Apollo Research), Lee Sharkey (ML Alignment and Theory Scholars University of Queensland)
Explainability and InterpretabilityComputational EfficiencyPrompt EngineeringAuto EncoderText
🎯 What it does: This paper proposes and implements an end-to-end sparse dictionary learning (e2e SAE) method to train sparse autoencoders, enabling the learned features to have functional importance in network performance.
Identifying General Mechanism Shifts in Linear Causal Representations
Tianyu Chen (University of Texas at Austin), Pradeep Kumar Ravikumar
Representation LearningTabular
🎯 What it does: This study considers the setting of linear causal representation learning, aiming to identify changes in potential causal mechanisms across multiple environments, particularly in the context of imperfect interventions to identify changes between latent factors.
Identifying Latent State-Transition Processes for Individualized Reinforcement Learning
Yuewen Sun (Mohamed bin Zayed University of Artificial Intelligence), Kun Zhang (Mohamed bin Zayed University of Artificial Intelligence)
Reinforcement LearningAuto EncoderSequential
🎯 What it does: This paper proposes a recognition framework for the latent state transition process in personalized reinforcement learning, defines the iMDP model, and provides identifiability theory.
Identifying Selections for Unsupervised Subtask Discovery
Yiwen Qiu (Carnegie Mellon University), Kun Zhang (Carnegie Mellon University)
Autonomous DrivingReinforcement LearningSequential
🎯 What it does: This paper proposes viewing sub-tasks from a causal perspective of selecting variables and designs a sequence non-negative matrix factorization (seq-NMF) method based on this perspective to automatically discover sub-goals and sub-tasks from expert demonstrations, and subsequently uses these sub-goals to construct augmented strategies for cross-task transfer.
Identifying Spatio-Temporal Drivers of Extreme Events
Mohamad Hakam Shams Eddin (Institute of Computer Science, University of Bonn), Juergen Gall (Institute of Computer Science, University of Bonn)
Anomaly DetectionConvolutional Neural NetworkTransformerTime SeriesAgriculture Related
🎯 What it does: An end-to-end network is proposed that can identify spatiotemporal driving factors related to extreme events (such as agricultural drought) from multivariate climate data and use these driving factors to predict future extreme events.
Identity Decoupling for Multi-Subject Personalization of Text-to-Image Models
Sangwon Jang (KAIST), Sung Ju Hwang (DeepAuto.ai)
SegmentationGenerationSupervised Fine-TuningDiffusion modelImageText
🎯 What it does: Proposes the MuDI framework to achieve personalized text generation for multiple agents, eliminating identity mixing.
IDGen: Item Discrimination Induced Prompt Generation for LLM Evaluation
Fan Lin (Tencent), Yu Zhang (SouthEast University)
GenerationData SynthesisTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextChain-of-Thought
🎯 What it does: A framework for prompt generation based on Item Discrimination (ID) theory has been designed and implemented to dynamically and sustainably produce high-discriminative, appropriately difficult LLM evaluation data.
Idiographic Personality Gaussian Process for Psychological Assessment
Yehu Chen (Washington University in St Louis), Roman Garnett (Washington University in St Louis)
Time SeriesSequential
🎯 What it does: This paper proposes a psychological assessment framework that combines nomothetic and idiographic perspectives—Idiographic Personality Gaussian Process (IPGP)—which can learn the underlying psychological structure of individuals from longitudinally serialized ordinal survey data and predict actual responses.
If You Want to Be Robust, Be Wary of Initialization
Sofiane ENNADIR, El houcine Bergou
Adversarial AttackConvolutional Neural NetworkGraph Neural NetworkImageGraph
🎯 What it does: This paper studies the impact of weight initialization and training cycles on the adversarial robustness of Graph Neural Networks (GNNs) and Deep Neural Networks (DNNs).
IF-Font: Ideographic Description Sequence-Following Font Generation
Xinping Chen (Fuzhou University), Wenzhong Guo (Fuzhou University)
GenerationTransformerGenerative Adversarial NetworkContrastive LearningImage
🎯 What it does: The IF-Font framework is proposed, transforming font generation into an autoregressive token prediction task based on Ideographic Description Sequences (IDS), achieving high-quality reproduction of font styles.
IllumiNeRF: 3D Relighting Without Inverse Rendering
Xiaoming Zhao (Google Research), Philipp Henzler (Google Research)
RestorationGenerationData SynthesisDiffusion modelNeural Radiance FieldImage
🎯 What it does: A non-inverse rendering 3D reconstruction method based on diffusion models, IllumiNeRF, is proposed, which can generate new views that can be rendered under any target lighting from a set of observed images in unknown lighting conditions.
Image Copy Detection for Diffusion Models
Wenhao Wang (University of Technology Sydney), Yi Yang (Zhejiang University)
GenerationTransformerDiffusion modelImage
🎯 What it does: This paper proposes the image copy detection task ICDiff for diffusion models and constructs the D-Rep dataset, which contains 40,000 pairs of real and generated images. It then introduces the PDF-Embedding method, which converts the copy level into a probability density function.
Image Reconstruction Via Autoencoding Sequential Deep Image Prior
Ismail Alkhouri, Saiprasad Ravishankar (Michigan State University)
RestorationConvolutional Neural NetworkAuto EncoderImageMagnetic Resonance ImagingComputed Tomography
🎯 What it does: An unsupervised image restoration method called aSeqDIP is proposed, which does not require a pre-trained model and utilizes the network structure itself. This method achieves gradual denoising and reconstruction by updating network weights in stages and adaptively updating the input.
Image Understanding Makes for A Good Tokenizer for Image Generation
Luting Wang (ByteDance), Bingyi Kang (ByteDance)
GenerationKnowledge DistillationTransformerGenerative Adversarial NetworkContrastive LearningImage
🎯 What it does: This study investigates how to transfer knowledge from image understanding (IU) models to image generation (IG) by training an image tokenizer through feature reconstruction, thereby improving generation quality.
Image-aware Evaluation of Generated Medical Reports
Gefen Dawidowicz (Technion Israel Institute of Technology), Ayellet Tal (Technion Israel Institute of Technology)
GenerationData SynthesisContrastive LearningImageTextMultimodalityElectronic Health Records
🎯 What it does: A new medical report generation evaluation metric, V LScore, is proposed, which measures the consistency between the generated report, the true report, and the image by utilizing the area of the triangle formed in the shared embedding space.
Images that Sound: Composing Images and Sounds on a Single Canvas
Ziyang Chen, Andrew Owens
GenerationData SynthesisDiffusion modelImageMultimodalityAudio
🎯 What it does: Utilizing pre-trained image diffusion models and audio diffusion models for parallel denoising in a shared latent space, generating visual spectrograms that can be viewed as images and played as audio (i.e., 'audible images');
IMAGPose: A Unified Conditional Framework for Pose-Guided Person Generation
Fei Shen (Nanjing University of Science and Technology), Jinhui Tang (Nanjing University of Science and Technology)
GenerationData SynthesisPose EstimationRetrievalDiffusion modelAuto EncoderImage
🎯 What it does: A unified conditional framework IMAGPose is proposed for pose-guided human image generation, capable of generating multiple target images with different poses at once, and can also generate target images using multi-view source images.
Imitating Language via Scalable Inverse Reinforcement Learning
Markus Wulfmeier (Google DeepMind), Martin Riedmiller (Google DeepMind)
Reinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: This paper explores the application of Inverse Reinforcement Learning (IRL) to the fine-tuning of large language models (LLMs), proposing a method to leverage the advantages of IRL in an offline manner without the need for online sampling by rewriting IQLearn in the form of Maximum Likelihood Estimation (MLE) with Temporal Difference (TD) regularization.
Immiscible Diffusion: Accelerating Diffusion Training with Noise Assignment
Yiheng Li, Chenfeng Xu (University of California Berkeley)
GenerationData SynthesisComputational EfficiencyDiffusion modelImage
🎯 What it does: A one-line code Immiscible Diffusion method is proposed, which accelerates the training of diffusion models through batch image-noise allocation.
ImOV3D: Learning Open Vocabulary Point Clouds 3D Object Detection from Only 2D Images
Timing Yang (Shanghai Qi Zhi Institute), Li Yi (Shanghai AI Lab)
Object DetectionDepth EstimationLarge Language ModelSupervised Fine-TuningContrastive LearningImagePoint Cloud
🎯 What it does: This study proposes the ImOV3D framework, which generates pseudo 3D point clouds and pseudo 3D annotations from 2D images through depth estimation and rendering, and trains an open vocabulary 3D object detection model using pseudo-multimodal representation (image-point cloud), significantly reducing the modality gap between training and testing.
Implicit Bias of Mirror Flow on Separable Data
Scott Pesme (École Polytechnique Fédérale de Lausanne), Nicolas Flammarion (École Polytechnique Fédérale de Lausanne)
ClassificationOptimizationTabularStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: A continuous time analysis of mirror flow for linearly separable classification problems is conducted, proving that its iterations converge directionally to the ϕ∞-maximum margin solution related to the shape of the potential function.
Implicit Curriculum in Procgen Made Explicit
Zhenxiong Tan (National University of Singapore), Xinchao Wang (National University of Singapore)
Reinforcement Learning
🎯 What it does: A controllable context version of Procgen, C-Procgen, was constructed, and a fine-grained analysis of the learning process under multi-level training revealed the existence of implicit curriculum learning.
Implicit Multimodal Alignment: On the Generalization of Frozen LLMs to Multimodal Inputs
Mustafa Shukor (Sorbonne University), Matthieu Cord (Valeo)
CompressionComputational EfficiencyRepresentation LearningTransformerLarge Language ModelVision Language ModelImageVideoMultimodalityAudio
🎯 What it does: This study investigates the internal representations of frozen large language models when receiving multimodal inputs such as images, videos, and audio, revealing different spatial distributions of perception and text tokens within the model, a high overlap in weight activations, and an implicit multimodal alignment phenomenon (IMA) that naturally emerges during training and inference.
Implicit Optimization Bias of Next-token Prediction in Linear Models
Christos Thrampoulidis (University of British Columbia)
OptimizationText
🎯 What it does: This paper studies the implicit optimization bias in the next word prediction (NTP) task trained under linear models using gradient descent or regularization paths. It proposes necessary conditions for achieving the entropy lower bound (NTP H-compatibility and NTP-separability) and proves that the gradient descent iteration direction ultimately converges to the maximum margin solution of NTP-SVM, while in the data subspace it converges to a finite solution that satisfies a system of linear equations.
Implicit Regularization of Decentralized Gradient Descent for Sparse Regression
Tongle Wu (Pennsylvania State University), Ying Sun (Pennsylvania State University)
OptimizationGraph
🎯 What it does: This study investigates the implicit regularization effect of decentralized gradient descent in over-parameterized sparse regression without regularization, and proves that it can achieve the same statistical optimality as centralized methods under the conditions of RIP and network connectivity.
Implicit Regularization of Sharpness-Aware Minimization for Scale-Invariant Problems
Bingcong Li (ETH Zurich), Niao He (ETH Zurich)
OptimizationComputational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This paper studies the implicit regularization of Sharpness-Aware Minimization (SAM) in scale-invariant problems, proposes a new balancedness metric, and based on this, designs a more computationally efficient BAR (Balancedness-Aware Regularization) algorithm. Experiments on various language models (RoBERTa, GPT-2, OPT-1.3B) for LoRA fine-tuning and few-shot learning validate the advantages of BAR in terms of performance and computational cost.
Implicit Regularization Paths of Weighted Neural Representations
Jin-Hong Du (Carnegie Mellon University), Pratik Patil (University of California Berkeley)
ClassificationOptimizationHyperparameter SearchConvolutional Neural NetworkImage
🎯 What it does: This paper studies the implicit regularization path of ridge regression after weighting (e.g., subsampling, resampling, etc.) on pre-trained features, and proves that along appropriate paths, the weighted ridge estimator has first and second-order equivalence in terms of degrees of freedom and prediction error compared to the full sample ridge estimator.
Implicitly Guided Design with PropEn: Match your Data to Follow the Gradient
Natasa Tagasovska, Andreas Loukas (New York University)
GenerationData SynthesisOptimizationAuto EncoderTabularSequential
🎯 What it does: A matching-based implicit guidance method called PropEn is proposed for attribute enhancement of designs in low-sample environments.
Imprecise Label Learning: A Unified Framework for Learning with Various Imprecise Label Configurations
Hao Chen (Carnegie Mellon University), Bhiksha Raj (Carnegie Mellon University)
ClassificationImageText
🎯 What it does: This paper proposes a unified 'Imprecise Label Learning' (ILL) framework that uses the EM algorithm to handle various incomplete or erroneous label scenarios (partial labels, semi-supervised, noisy labels, and their mixtures) within the framework of maximum likelihood estimation, without the need to design separate methods for each type of label uncertainty.
Improved Algorithms for Contextual Dynamic Pricing
Matilde Tullii (FairPlay Team CREST ENSAE), Vianney Perchet (FairPlay Team CREST ENSAE)
Optimization
🎯 What it does: This paper proposes a new algorithmic framework called VAPE for achieving adaptive valuation and price elimination in dynamic pricing problems with contextual information.
Improved Analysis for Bandit Learning in Matching Markets
Fang Kong (Southern University of Science and Technology), Shuai Li (Shanghai Jiao Tong University)
Recommendation SystemOptimizationReinforcement LearningTabular
🎯 What it does: An adaptive exploration-matching algorithm (AETGS-E) is proposed to learn player preferences in a two-sided matching market and achieve player-optimal stable matching.
Improved Bayes Regret Bounds for Multi-Task Hierarchical Bayesian Bandit Algorithms
Jiechao Guan (Hong Kong University of Science and Technology), Hui Xiong (Hong Kong University of Science and Technology)
Reinforcement Learning
🎯 What it does: This paper proposes an improved Bayesian regret bound for multi-task hierarchical Bayesian gambling algorithms, particularly in multi-task Gaussian linear gambling and semi-gambling settings.
Improved Distribution Matching Distillation for Fast Image Synthesis
Tianwei Yin (Massachusetts Institute of Technology), William T. Freeman (Massachusetts Institute of Technology)
GenerationData SynthesisKnowledge DistillationDiffusion modelGenerative Adversarial NetworkImage
🎯 What it does: The distribution matching distillation method DMD is improved to have a non-regressive loss, incorporate GAN objectives, support multi-step sampling, and eliminate training-inference mismatches, resulting in higher quality one-step or four-step image generators.
Improved Few-Shot Jailbreaking Can Circumvent Aligned Language Models and Their Defenses
Xiaosen Zheng (Sea AI Lab), Min Lin (Sea AI Lab)
Adversarial AttackTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: This paper proposes an improved few-shot jailbreak attack method—I-FSJ, which can achieve a high success rate of jailbreak with a very low number of queries (1-8 times) on open-source aligned LLMs with limited context windows (≤8192).
Improved Generation of Adversarial Examples Against Safety-aligned LLMs
Qizhang Li (Harbin Institute of Technology), Hao Chen (UC Davis)
Adversarial AttackLarge Language ModelPrompt EngineeringText
🎯 What it does: Conduct white-box attacks on large language models for safety alignment, proposing an improved gradient-based adversarial example generation method.
Improved Guarantees for Fully Dynamic $k$-Center Clustering with Outliers in General Metric Spaces
Leyla Biabani (Eindhoven University of Technology), Melanie Schmidt (Heinrich Heine University Dusseldorf)
Optimization
🎯 What it does: This paper proposes a fully dynamic k-center clustering algorithm (including z outliers) that can maintain a 4+ε approximate solution in any metric space.
Improved learning rates in multi-unit uniform price auctions
Marius Potfer (ENSAE), Cheng Wan (EDF R&D)
OptimizationReinforcement Learning
🎯 What it does: A new bidding space representation is proposed to address the online learning problem in multi-unit uniform price auctions, along with a corresponding online learning algorithm that can achieve optimal or near-optimal regret rates under different feedback models (bandit, full-information, all-winner);
Improved off-policy training of diffusion samplers
Marcin Sendera (Mila, Université de Montréal), Nikolay Malkin (Mila, Université de Montréal)
GenerationData SynthesisOptimizationDiffusion modelImageTabularStochastic Differential Equation
🎯 What it does: This paper studies how to train diffusion models (neural SDE) for sampling given only an unnormalized energy function, and unifies it with methods such as continuous generative flow networks (GFlowNet).
Improved Particle Approximation Error for Mean Field Neural Networks
Atsushi Nitanda (Agency for Science Technology and Research Singapore)
Optimization
🎯 What it does: An improved result on particle approximation error is presented in Mean Field Lagrangian Dynamics (MFLD), proving that the error is O(1/N), eliminating the dependence on the logarithmic Sobolev inequality constant;
Improved Regret for Bandit Convex Optimization with Delayed Feedback
Yuanyu Wan (Zhejiang University), Lijun Zhang (Nanjing University)
OptimizationTabular
🎯 What it does: This paper studies bandit convex optimization (BCO) with delayed feedback and proposes a new algorithm D-FTBL, proving that its regret bound in general cases is O(√nT^(3/4) + √dT).
Improved Regret of Linear Ensemble Sampling
Harin Lee (Seoul National University), Min-hwan Oh (Seoul National University)
OptimizationReinforcement Learning from Human Feedback
🎯 What it does: This paper proposes an improved Linear Ensemble Sampling algorithm and proves that with a logarithmic scale model set (m=Ω(K log T)), it can achieve a frequentist regret upper bound of O~(d^{3/2}√T) over T time steps, which is comparable to state-of-the-art stochastic linear bandit algorithms (such as Thompson Sampling and LinPHE).
Improved Sample Complexity Bounds for Diffusion Model Training
Shivam Gupta (University of Texas at Austin), Zhiyang Xun (University of Texas at Austin)
GenerationData SynthesisDiffusion modelScore-based Model
🎯 What it does: This paper studies the sample complexity of training diffusion models and proposes that by using the standard Score-matching objective to train fully connected ReLU networks, it is possible to learn sufficiently accurate scores with a polynomial sample size, thereby achieving DDPM sampling.
Improved Sample Complexity for Multiclass PAC Learning
Steve Hanneke (Purdue University), Qian Zhang (Purdue University)
ClassificationOptimization
🎯 What it does: This paper studies the optimal sample complexity of multi-class PAC learning and proposes two approaches, list learning and pivot shifting, to narrow the known upper and lower bounds.
Improving Adaptivity via Over-Parameterization in Sequence Models
Yicheng Li (Tsinghua University), Qian Lin (Beijing Academy of Artificial Intelligence)
OptimizationRecurrent Neural NetworkSequential
🎯 What it does: This paper studies the use of over-parameterized gradient descent in sequence models to dynamically learn the eigenvalues of kernel functions, thereby enhancing the generalization performance of non-parametric regression.
Improving Adversarial Robust Fairness via Anti-Bias Soft Label Distillation
Shiji Zhao (Institute of Artificial Intelligence Beihang University), Xingxing Wei (Alibaba Group)
ClassificationKnowledge DistillationAdversarial AttackConvolutional Neural NetworkImage
🎯 What it does: A soft label distillation method based on category temperature adjustment is proposed to improve the robustness and fairness of deep networks under adversarial attacks.
Improving Alignment and Robustness with Circuit Breakers
Andy Zou (Gray Swan AI), Dan Hendrycks
GenerationRepresentation LearningAdversarial AttackMultimodality
🎯 What it does: By inserting a 'circuit breaker' into the internal representation of the model, the internal representation is rerouted and interrupted in advance when generating harmful content, preventing the model from producing harmful outputs.
Improving Context-Aware Preference Modeling for Language Models
Silviu Pitis (University of Toronto), Alessandro Sordoni (Microsoft Research)
Recommendation SystemTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: A two-step context-aware preference modeling framework is proposed and validated, utilizing explicit context to eliminate ambiguous preferences, and a dataset called RPR (Reasonable Preference Reversal) is constructed to evaluate the model's sensitivity to context.
Improving Decision Sparsity
Yiyang Sun (Duke University), Cynthia Rudin (Duke University)
OptimizationExplainability and InterpretabilityTabularBiomedical Data
🎯 What it does: This paper proposes and evaluates various improved Sparse Explanation Value (SEV) methods to provide closer, more trustworthy, and sparser decision explanations in machine learning models.
Improving Deep Learning Optimization through Constrained Parameter Regularization
Jörg K.H. Franke (University of Freiburg), Frank Hutter (University of Freiburg)
ClassificationSegmentationOptimizationImageTextBiomedical Data
🎯 What it does: Proposes Constrained Parameter Regularization (CPR), transforming regularization into a constrained optimization problem for each parameter matrix, and achieving adaptive regularization strength through the augmented Lagrangian method.
Improving Deep Reinforcement Learning by Reducing the Chain Effect of Value and Policy Churn
Hongyao Tang (Mila - Quebec Artificial Intelligence Institute University of Montreal), Glen Berseth (Mila - Quebec Artificial Intelligence Institute University of Montreal)
Reinforcement Learning
🎯 What it does: This study investigates the churn (output drift) of value networks and policy networks in deep reinforcement learning and its chain amplification effect, proposing a pluggable regularization method called CHAIN to mitigate churn and enhance learning performance.
Improving Environment Novelty Quantification for Effective Unsupervised Environment Design
Jayden Teoh (Singapore Management University), Pradeep Varakantham (Singapore Management University)
OptimizationRobotic IntelligenceReinforcement LearningSequential
🎯 What it does: A new environmental novelty assessment framework based on student state-action coverage, CENIE, is proposed and embedded into existing UED algorithms (PLR-CENIE, ACCEL-CENIE) to enhance the effectiveness of unsupervised environment design.
Improving Equivariant Model Training via Constraint Relaxation
Stefanos Pertigkiozoglou (University of Pennsylvania), Kostas Daniilidis (University of Pennsylvania)
OptimizationGraph Neural NetworkPoint Cloud
🎯 What it does: A training framework is proposed that relaxes the equivariance constraints during the training phase and restores equivariance during the inference phase to enhance the optimization performance of equivariant neural networks.
Improving Generalization and Convergence by Enhancing Implicit Regularization
Mingze Wang (Peking University), Lei Wu (Peking University)
OptimizationComputational EfficiencyConvolutional Neural NetworkTransformerLarge Language ModelImageText
🎯 What it does: An Implicit Regularization Enhancement (IRE) framework is proposed, which accelerates gradient updates in flat directions during training, leading to faster convergence to flat minima, thereby enhancing the model's generalization ability and training speed.
Improving Generalization in Federated Learning with Model-Data Mutual Information Regularization: A Posterior Inference Approach
Hao Zhang (Shanghai Jiao Tong University), Hongkai Xiong (Shanghai Jiao Tong University)
Federated LearningSafty and PrivacyImageStochastic Differential Equation
🎯 What it does: A posterior inference framework FedMDMI is proposed in federated learning through model-data mutual information regularization to enhance the model's generalization ability and uncertainty estimation.
Improving Generalization of Dynamic Graph Learning via Environment Prompt
Kuo Yang (University of Science and Technology of China), Yang Wang (University of Science and Technology of China)
Graph Neural NetworkPrompt EngineeringGraphTime Series
🎯 What it does: A dynamic learning framework EpoD based on prompt learning and structural causal models is proposed to address the dual challenges of environmental inference and utilization.
Improving Gloss-free Sign Language Translation by Reducing Representation Density
Jinhui Ye (Hong Kong University of Science and Technology), Hui Xiong (Hong Kong University of Science and Technology)
RecognitionRepresentation LearningContrastive LearningVideo
🎯 What it does: This paper identifies the representation density problem in non-Gloss sign language translation and proposes a lightweight contrastive learning strategy, SignCL, to reduce feature density and improve translation performance.
Improving Linear System Solvers for Hyperparameter Optimisation in Iterative Gaussian Processes
Jihao Andreas Lin (University of Cambridge), José Miguel Hernández-Lobato (University of Cambridge)
OptimizationComputational EfficiencyHyperparameter SearchTabular
🎯 What it does: This paper addresses the hyperparameter optimization of Gaussian processes (GP) for large datasets and proposes three techniques to improve the linear system solver, significantly enhancing computational efficiency.
Improving Neural Network Surface Processing with Principal Curvatures
Josquin Harrison (Inria Sophia Antipolis), Maxime Sermesant (Inria Sophia Antipolis)
ClassificationSegmentationPoint Cloud
🎯 What it does: Using principal curvature as an input feature in neural network surface processing tasks to enhance the model's perception of surface geometry.
Improving Neural ODE Training with Temporal Adaptive Batch Normalization
Su Zheng (Chinese University of Hong Kong), Martin D. Wong (Massachusetts Institute of Technology)
ClassificationOptimizationImageTime SeriesPhysics RelatedOrdinary Differential Equation
🎯 What it does: Proposed and implemented a time-adaptive batch normalization (TA-BN) for Neural ODEs, addressing the mismatch issue of traditional BN in continuous time models.
Improving Robustness of 3D Point Cloud Recognition from a Fourier Perspective
Yibo Miao (Chinese Academy of Sciences), Xiao-Shan Gao (Tsinghua University)
RecognitionAdversarial AttackGraph Neural NetworkPoint Cloud
🎯 What it does: This study investigates the robustness of 3D point cloud recognition models under distortions such as noise and proposes adversarial training in the frequency domain to enhance model robustness.
Improving robustness to corruptions with multiplicative weight perturbations
Trung Trinh (Aalto University), Samuel Kaski (Aalto University)
OptimizationAdversarial AttackTransformerGaussian SplattingImage
🎯 What it does: A method is proposed to enhance the model's robustness to various image distortions (noise, lighting, compression, etc.) by multiplying the network weights by random multiplicative noise (DAMP) during training.
Improving self-training under distribution shifts via anchored confidence with theoretical guarantees
Taejong Joo (Northwestern University), Diego Klabjan (Northwestern University)
Domain AdaptationKnowledge DistillationImage
🎯 What it does: Proposes an Anchored Confidence (AnCon) method that utilizes the temporal consistency of reliable predictions for self-training to improve performance under distribution shifts.
Improving Sparse Decomposition of Language Model Activations with Gated Sparse Autoencoders
Senthooran Rajamanoharan (Google DeepMind), Neel Nanda (Google DeepMind)
Explainability and InterpretabilityComputational EfficiencyRepresentation LearningTransformerAuto EncoderText
🎯 What it does: A Gated Sparse Autoencoder (Gated SAE) is proposed for efficiently and sparsely learning interpretable features in language model activations.
Improving Subgroup Robustness via Data Selection
Saachi Jain (Massachusetts Institute of Technology), Aleksander Madry (Massachusetts Institute of Technology)
ClassificationData-Centric LearningImage
🎯 What it does: A data model-based training sample selection method called D3M is proposed, which enhances subgroup robustness by identifying and removing training samples that negatively impact the performance of minority groups.
Improving Temporal Link Prediction via Temporal Walk Matrix Projection
Xiaodong Lu (Beihang University), Weifeng Lv (Beihang University)
Computational EfficiencyRepresentation LearningGraph Neural NetworkGraphBenchmark
🎯 What it does: This paper proposes a new temporal walk matrix projection model TPNet, which combines a time-decayed temporal walk matrix with random feature projection to implicitly maintain node representations for efficient relative encoding.
Improving the Learning Capability of Small-size Image Restoration Network by Deep Fourier Shifting
Man Zhou (Aerospace Information Research Institute, Chinese Academy of Sciences University of Science and Technology of China)
RestorationSuper ResolutionConvolutional Neural NetworkImage
🎯 What it does: This paper proposes a Deep Fourier Shifting operator based on Fourier cycles to replace traditional convolution units, aiming to enhance the performance of low-level image restoration tasks.
Improving the Training of Rectified Flows
Sangyun Lee (Carnegie Mellon University), Giulia Fanti (Carnegie Mellon University)
GenerationKnowledge DistillationDiffusion modelRectified FlowImage
🎯 What it does: Improved the training process of Rectified Flow, enabling it to compete with knowledge distillation methods under low NFE conditions.
Improving the Worst-Case Bidirectional Communication Complexity for Nonconvex Distributed Optimization under Function Similarity
Kaja Gruntkowska (King Abdullah University of Science and Technology), Peter Richtárik (King Abdullah University of Science and Technology)
OptimizationTabular
🎯 What it does: A new distributed optimization communication compression method is proposed, focusing on improving the downlink communication complexity from the server to the worker nodes, and based on this, a bidirectional compression algorithm M3 is constructed;
Improving Viewpoint-Independent Object-Centric Representations through Active Viewpoint Selection
Yinxuan Huang (Fudan University), Xiangyang Xue (Fudan University)
SegmentationGenerationRepresentation LearningDiffusion modelImage
🎯 What it does: This paper proposes the AVS multi-view object-centered learning model and implements an active view selection strategy based on information gain to improve object representation and image generation.
Improving Visual Prompt Tuning by Gaussian Neighborhood Minimization for Long-Tailed Visual Recognition
Mengke Li (Guangming Laboratory), Hui Huang (Shenzhen University)
ClassificationRecognitionOptimizationTransformerPrompt EngineeringImage
🎯 What it does: This paper proposes Gaussian Neighborhood Minimization Prompt Tuning (GNM-PT), which uses Gaussian neighborhood mean loss in Visual Prompt Tuning (VPT) to smooth the loss landscape and enhance the generalization performance of long-tail visual recognition.
In Pursuit of Causal Label Correlations for Multi-label Image Recognition
Zhao-Min Chen (Wenzhou University), Sixian Chan (Zhejiang University of Technology)
RecognitionTransformerImage
🎯 What it does: This paper proposes a multi-label image recognition method based on causal intervention, utilizing Transformer to decouple label features and modeling causal label associations through cross-attention, thereby suppressing misleading co-occurrence relationships.
In-and-Out: Algorithmic Diffusion for Sampling Convex Bodies
Yunbum Kook (Georgia Institute of Technology), Matthew Shunshi Zhang
Stochastic Differential Equation
🎯 What it does: A new random walk algorithm called In-and-Out is proposed for uniform sampling in high-dimensional convex bodies, with convergence guarantees under Rényi divergence.
In-Context Learning of a Linear Transformer Block: Benefits of the MLP Component and One-Step GD Initialization
Ruiqi Zhang (University of California Berkeley), Peter Bartlett
TransformerLarge Language ModelTabular
🎯 What it does: This paper studies the performance of Linear Transformer Blocks (LTB) — composed of linear attention and linear MLP — on linear regression (Gaussian prior with non-zero mean) problems in the context of in-context learning (ICL) tasks, and demonstrates that it can achieve ICL risk close to Bayesian optimality.
In-Context Learning State Vector with Inner and Momentum Optimization
Dongfang Li (Harbin Institute of Technology), Min Zhang (Harbin Institute of Technology)
OptimizationTransformerLarge Language ModelText
🎯 What it does: This paper introduces the concept of 'state vector' by treating the compressed vector of In-Context Learning (ICL) as parameters for gradient descent training, and optimizes and aggregates it to enhance the performance of LLMs in zero-shot and few-shot tasks.
In-Context Learning with Representations: Contextual Generalization of Trained Transformers
Tong Yang (Carnegie Mellon University), Yuejie Chi (Carnegie Mellon University)
TransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: This paper studies the ability of pre-trained large language models in in-context learning (ICL), particularly how transformers generalize to unseen examples in prompts. By analyzing the training dynamics of a single-layer multi-head transformer, it explores how to perform non-linear regression tasks in context learning under partially labeled prompts.
In-Context Learning with Transformers: Softmax Attention Adapts to Function Lipschitzness
Liam Collins (University of Texas at Austin), Sanjay Shakkottai (University of Texas at Austin)
Transformer
🎯 What it does: This study investigates how pre-trained softmax attention adapts the Lipschitz constant of functions and noise in context learning to achieve nearest neighbor regression.
In-Context Symmetries: Self-Supervised Learning through Contextual World Models
Sharut Gupta (Massachusetts Institute of Technology), Stefanie Jegelka (Massachusetts Institute of Technology)
ClassificationRetrievalRepresentation LearningTransformerContrastive LearningImageTabularBiomedical Data
🎯 What it does: A self-supervised learning framework called CONTEXTSSL is proposed, which can adaptively achieve equivariant or invariant representations for different transformations based on task requirements through a small amount of contextual memory.
In-N-Out: Lifting 2D Diffusion Prior for 3D Object Removal via Tuning-Free Latents Alignment
Dongting Hu (University of Melbourne), Mingming Gong (Mohamed bin Zayed University of Artificial Intelligence)
RestorationGenerationDiffusion modelNeural Radiance FieldImage
🎯 What it does: Based on the 2D diffusion model, this paper proposes generating multi-view consistent completed images through Explicit Latent Variable Alignment (ELA) and Implicit Latent Variable Alignment (ILA), and optimizes NeRF using Patch-Hybrid loss to achieve high-quality 3D object removal.
In-Trajectory Inverse Reinforcement Learning: Learn Incrementally Before an Ongoing Trajectory Terminates
Shicheng Liu (Pennsylvania State University), Minghui Zhu (Pennsylvania State University)
Robotic IntelligenceReinforcement LearningTime SeriesSequentialFinance Related
🎯 What it does: A framework for 'intra-trajectory inverse reinforcement learning' (MERIT-IRL) is proposed, which can update the reward function and corresponding policy in real-time while the expert trajectory is still ongoing.
Incentivizing Quality Text Generation via Statistical Contracts
Eden Saig (Technion Israel Institute of Technology), Inbal Talgam-Cohen (Tel Aviv University)
GenerationOptimizationTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: A 'pay-for-performance' scheme based on contract design is proposed to incentivize large language models (LLMs) to produce high-quality text and avoid moral hazards.
Incorporating Surrogate Gradient Norm to Improve Offline Optimization Techniques
Manh Cuong Dao (Hanoi University of Science and Technology), Trong Nghia Hoang (Washington State University)
Optimization
🎯 What it does: This paper proposes a model-agnostic regularization method based on model sharpness (IGNITE) to improve the training of surrogate models for offline optimizers, thereby enhancing the performance of offline optimization tasks.
Incorporating Test-Time Optimization into Training with Dual Networks for Human Mesh Recovery
Yongwei Nie (South China University of Technology), Xuemiao Xu (South China University of Technology)
Pose EstimationOptimizationMeta LearningImageMesh
🎯 What it does: A meta-learning framework that integrates test-time optimization into the training process is proposed, utilizing a dual network structure to unify training and testing objectives, thereby improving the accuracy of human mesh reconstruction from a single image.
Incremental Learning of Retrievable Skills For Efficient Continual Task Adaptation
Daehee Lee (Sungkyunkwan University), Honguk Woo (Sungkyunkwan University)
Robotic IntelligenceMeta LearningTransformerReinforcement LearningMultimodality
🎯 What it does: A framework for incremental learning with retrievable skills based on adapters, called IsCiL, has been developed to achieve efficient task adaptation in continuous imitation learning.
INDICT: Code Generation with Internal Dialogues of Critiques for Both Security and Helpfulness
Hung Le (Salesforce Research), Silvio Savarese (Salesforce Research)
Safty and PrivacyAI Code AssistantTransformerLarge Language ModelAgentic AIText
🎯 What it does: A framework named INDICT is proposed, which utilizes two internal critics (security critic and practicality critic) for collaborative review and improvement of code generated by large language models;
Induced Model Matching: Restricted Models Help Train Full-Featured Models
Usama Muneeb (University of Illinois Chicago), Mesrob I Ohannessian
Knowledge DistillationRecurrent Neural NetworkTransformerLarge Language ModelReinforcement LearningTextSequential
🎯 What it does: This paper proposes a method called Induced Model Matching (IMM), which enhances the performance of a full-feature model by aligning its induced version with an existing constrained feature model during the training process.
Inductive biases of multi-task learning and finetuning: multiple regimes of feature reuse
Samuel Lippl (Columbia University), Jack Lindsey (Anthropic)
Convolutional Neural NetworkTransformerImage
🎯 What it does: The theoretical derivation of the implicit inductive bias of multi-task learning (MTL) and pre-training + fine-tuning (PT + FT) in diagonal linear networks and single hidden layer ReLU networks, validated through teacher-student experiments and deep CNN/ViT experiments.
Inevitable Trade-off between Watermark Strength and Speculative Sampling Efficiency for Language Models
Zhengmian Hu (University of Maryland), Heng Huang (University of Maryland)
GenerationOptimizationTransformerLarge Language ModelText
🎯 What it does: This paper studies the combination of unbiased watermarking and speculative sampling in large language models, proposing a dual re-weighting framework and proving that both cannot simultaneously maintain optimal watermark strength and sampling efficiency.
Inexact Augmented Lagrangian Methods for Conic Optimization: Quadratic Growth and Linear Convergence
Feng-Yi Liao (University of California San Diego), Yang Zheng (University of California San Diego)
OptimizationGraph
🎯 What it does: This paper proves that under strict complementarity conditions, both the primal and dual problems of semidefinite programming satisfy quadratic growth and error bounds, thereby demonstrating that the inexact augmented Lagrangian method (ALM) achieves linear convergence for both primal and dual iterations.
Inference of Neural Dynamics Using Switching Recurrent Neural Networks
Yongxu Zhang (Yale University), Shreya Saxena (Yale University)
Recurrent Neural NetworkTime SeriesSequential
🎯 What it does: A model named Switching Recurrent Neural Network (SRNN) is proposed and implemented to automatically identify discrete states from neural time series data and reconstruct nonlinear neural dynamics.
Inference via Interpolation: Contrastive Representations Provably Enable Planning and Inference
Benjamin Eysenbach (Princeton University), Sergey Levine (University of California Berkeley)
OptimizationRepresentation LearningRobotic IntelligenceContrastive LearningTime SeriesSequential
🎯 What it does: This paper utilizes representations obtained from regularized temporal contrastive learning and proves that these representations form a Gauss-Markov chain under certain assumptions, allowing for predictions of future states, inferences of intermediate states, and reasoning tasks such as path planning through simple methods like linear interpolation or low-dimensional matrix inversion.
Inferring Neural Signed Distance Functions by Overfitting on Single Noisy Point Clouds through Finetuning Data-Driven based Priors
Chao Chen (Tsinghua University), Zhizhong Han (Wayne State University)
GenerationAutonomous DrivingOptimizationAuto EncoderPoint Cloud
🎯 What it does: A neural SDF learning method that combines data-driven priors and overfitting strategies is proposed, using local noise-noise statistical inference to fine-tune the prior on a single noisy point cloud, thereby recovering high-quality implicit surfaces without the need for signed distance supervision, clean point clouds, or normal information.
Inferring stochastic low-rank recurrent neural networks from neural data
Matthijs Pals (University of Tübingen), Jakob H. Macke (Max Planck Institute for Intelligent Systems)
GenerationExplainability and InterpretabilityRecurrent Neural NetworkTime SeriesSequentialBiomedical DataElectrocardiogramStochastic Differential Equation
🎯 What it does: Using the variational sequential Monte Carlo (SMC) method to fit recurrent neural networks (RNNs) with low-rank structure and stochastic transitions, generating interpretable generative models that can match the variability of neural data.
Infinite Limits of Multi-head Transformer Dynamics
Blake Bordelon (Harvard University), Cengiz Pehlevan (Harvard University)
OptimizationRepresentation LearningTransformerImageText
🎯 What it does: This paper analyzes the training dynamics of Transformers in feature learning mode using Dynamical Mean Field Theory (DMFT) under three infinite limits (key/query dimension N→∞, number of heads H→∞, number of layers L→∞), and provides corresponding parameterization (such as μP scaling) and learning rate trade-offs. Subsequently, numerical experiments were conducted on visual and language models to validate theoretical predictions and explore the impact of different scalings on performance.
Infinite-Dimensional Feature Interaction
Chenhui Xu (George Mason University), Xiang Chen (Peking University)
ClassificationObject DetectionSegmentationConvolutional Neural NetworkImage
🎯 What it does: Designed the InfiNet network, achieving infinite-dimensional interactions in the feature interaction space by using RBF kernels instead of element-wise multiplication.
Inflationary Flows: Calibrated Bayesian Inference with Diffusion-Based Models
Daniela F De Albuquerque, John Pearson (Duke University)
GenerationData SynthesisCompressionDiffusion modelImageStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: This paper studies a generative model for low-dimensional Bayesian inference that is calibratable and identifiable, utilizing a diffusion model based on probabilistic flow ODEs (inflationary flows).
InfLLM: Training-Free Long-Context Extrapolation for LLMs with an Efficient Context Memory
Chaojun Xiao (Tsinghua University), Maosong Sun (Tsinghua University)
TransformerLarge Language ModelTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Proposed a training-free InfLLM method that handles ultra-long sequences through sliding window attention combined with external context memory.
InfoRM: Mitigating Reward Hacking in RLHF via Information-Theoretic Reward Modeling
Yuchun Miao (Wuhan University), Dacheng Tao (Nanyang Technological University)
OptimizationReinforcement Learning from Human FeedbackReinforcement LearningText
🎯 What it does: Proposes the InfoRM information bottleneck reward model to address the issue of reward hacking in RLHF, and designs the Cluster Separation Index (CSI) metric for real-time detection and prevention of over-optimization.
Information Re-Organization Improves Reasoning in Large Language Models
Xiaoxia Cheng (Zhejiang University), Weiming Lu (Zhejiang University)
TransformerLarge Language ModelReinforcement LearningTextChain-of-Thought
🎯 What it does: Reorganize information from the context, first extract logical relationships, then prune noise, and use the reorganized information for reasoning.
Information-theoretic Generalization Analysis for Expected Calibration Error
Futoshi Futami (Osaka University), Masahiro Fujisawa (RIKEN AIP)
ClassificationConvolutional Neural NetworkRecurrent Neural NetworkImage
🎯 What it does: This paper analyzes the total bias of expected calibration error (ECE) in binary classification models using uniform width (UWB) and uniform mass (UMB) binning, proposes a tighter bias upper bound, and derives the optimal number of bins; it also provides a generalization error upper bound for ECE and true calibration error (TCE) using information theory methods, applying it to the recalibration of reused training data.
Information-theoretic Limits of Online Classification with Noisy Labels
Changlong Wu (Purdue University), Wojciech Szpankowski (Purdue University)
ClassificationOptimization
🎯 What it does: This paper studies the scenario in online classification tasks where true labels are contaminated by noise while features are adversarially generated, providing the optimal minimization risk under the noise mechanism and its lower bound.