NeurIPS 2024 Papers — Page 26
Conference on Neural Information Processing Systems · 4035 papers
Nuclear Norm Regularization for Deep Learning
Christopher Scarvelis (Massachusetts Institute of Technology), Justin Solomon (Massachusetts Institute of Technology)
RestorationConvolutional Neural NetworkImage
🎯 What it does: This study investigates the use of nuclear norm regularization (Jacobian nuclear norm) to constrain deep learning models and proposes a scalable method for Jacobian-free estimation.
NVRC: Neural Video Representation Compression
Ho Man Kwan (University of Bristol), David Bull (University of Bristol)
CompressionVideo
🎯 What it does: An end-to-end optimized INR video compression framework NVRC is proposed, which achieves significant improvements in compression efficiency through hierarchical compression of neural representation, quantization parameters, and entropy model parameters.
OASIS: Conditional Distribution Shaping for Offline Safe Reinforcement Learning
Yihang Yao (Carnegie Mellon University), Ding Zhao (Google DeepMind)
Safty and PrivacyRobotic IntelligenceReinforcement LearningDiffusion modelTabular
🎯 What it does: The OASIS method is proposed, utilizing conditional diffusion models to shape the distribution of offline safe reinforcement learning datasets to address the issue of safe data mismatch.
Object segmentation from common fate: Motion energy processing enables human-like zero-shot generalization to random dot stimuli
Matthias Tangemann (University of Tübingen), Matthias Bethge (University of Tübingen)
Object DetectionSegmentationConvolutional Neural NetworkOptical FlowVideo
🎯 What it does: This study investigates the zero-shot motion segmentation task, comparing the performance of traditional optical flow models with neuroscience-inspired motion energy models on random point stimuli.
Observational Scaling Laws and the Predictability of Langauge Model Performance
Yangjun Ruan (University of Toronto), Tatsunori Hashimoto (Stanford University)
Computational EfficiencyLarge Language ModelTextBenchmarkChain-of-Thought
🎯 What it does: This paper proposes an Observational Scaling Law based on publicly available model standard benchmark scores, allowing for scale predictions across model families without the need to retrain the models.
OccamLLM: Fast and Exact Language Model Arithmetic in a Single Step
Owen M Dugan, Marin Soljacic (Massachusetts Institute of Technology)
TransformerLarge Language ModelTextBenchmark
🎯 What it does: This paper presents OccamLLM, a framework for achieving precise arithmetic in single-step autoregression.
OccFusion: Rendering Occluded Humans with Generative Diffusion Priors
Adam Sun (Stanford University), Ehsan Adeli (Stanford University)
GenerationOptimizationDiffusion modelScore-based ModelGaussian SplattingVideo
🎯 What it does: To address the issue of rendering occluded human bodies in monocular videos, a three-stage pipeline called OccFusion is proposed: first, a complete human silhouette is generated using a diffusion model and pose priors; second, optimization is performed on a 3D Gaussian scattering model, utilizing Score-Distillation Sampling to regularize human geometry in both pose space and canonical space; finally, in-context inpainting is employed to refine the appearance, ultimately achieving fast, high-quality complete human rendering.
Occupancy-based Policy Gradient: Estimation, Convergence, and Optimality
Audrey Huang (University of Illinois), Nan Jiang (University of Illinois)
OptimizationReinforcement Learning
🎯 What it does: A policy gradient method that relies solely on the occupancy function is proposed, capable of directly optimizing policies in both online and offline reinforcement learning.
Octopus: A Multi-modal LLM with Parallel Recognition and Sequential Understanding
Chuyang Zhao (Baidu), Yifan Sun (Baidu)
RecognitionObject DetectionComputational EfficiencyTransformerLarge Language ModelVision Language ModelImageTextMultimodality
🎯 What it does: This paper proposes a multimodal large language model called Octopus, which divides visual recognition and understanding into two stages: parallel recognition and sequential understanding, enhancing reasoning efficiency and accuracy.
OctreeOcc: Efficient and Multi-Granularity Occupancy Prediction Using Octree Queries
Yuhang Lu (ShanghaiTech University), Yuexin Ma (ShanghaiTech University)
SegmentationAutonomous DrivingComputational EfficiencyPoint Cloud
🎯 What it does: This paper proposes OctreeOcc, a 3D occupancy prediction framework that utilizes an octree structure, capable of adaptively modeling objects of varying sizes and complexities at multiple granularities.
ODGEN: Domain-specific Object Detection Data Generation with Diffusion Models
JingYuan Zhu, Huimin Ma
Object DetectionGenerationData SynthesisDiffusion modelImageMagnetic Resonance ImagingComputed Tomography
🎯 What it does: This paper proposes ODGEN, which combines bounding boxes and text prompts on a pre-trained diffusion model for high-quality, controllable image synthesis, aimed at generating domain-specific object detection training data.
ODGS: 3D Scene Reconstruction from Omnidirectional Images with 3D Gaussian Splattings
Suyoung Lee (Seoul National University), Kyoung Mu Lee (Seoul National University)
RestorationGenerationComputational EfficiencyGaussian SplattingSimultaneous Localization and MappingImage
🎯 What it does: A 3D high-quality scene reconstruction method for panoramic images, ODGS, has been developed, utilizing 3D Gaussian splatting for fast rendering.
Off-Dynamics Reinforcement Learning via Domain Adaptation and Reward Augmented Imitation
Yihong Guo (Johns Hopkins University), Anqi Liu (Johns Hopkins University)
Domain AdaptationReinforcement LearningGenerative Adversarial NetworkSequential
🎯 What it does: The DARC policy trained using modified rewards in the source domain is transferred to the target domain through observational imitation learning (GAIfO), forming an offline dynamic transfer reinforcement learning framework (DARAIL).
Off-policy estimation with adaptively collected data: the power of online learning
Jeonghwan Lee (University of Chicago), Cong Ma (University of Chicago)
OptimizationReinforcement Learning
🎯 What it does: This paper proposes an AIPW (doubly robust) estimation framework for adaptive sampling data and provides a non-asymptotic error upper bound, addressing the off-policy value estimation problem in offline reinforcement learning and causal inference under discrete and continuous action spaces.
Off-Policy Selection for Initiating Human-Centric Experimental Design
Ge Gao (Stanford University), Min Chi (North Carolina State University)
OptimizationReinforcement LearningAuto EncoderTabularSequentialBiomedical Data
🎯 What it does: A framework for offline strategy selection (FPS) based on subgroup partitioning and VAE data augmentation is designed for new participants in human-centered systems, allowing the deployment of a specific RL strategy using only the initial state.
Offline Behavior Distillation
Shiye Lei (University of Sydney), Dacheng Tao (Nanyang Technological University)
OptimizationKnowledge DistillationReinforcement LearningTabularBenchmark
🎯 What it does: This study investigates how to compress massive optimal offline RL data into a small amount of expert behavior data, thereby quickly obtaining high-performance policies through behavior cloning.
Offline Multitask Representation Learning for Reinforcement Learning
Haque Ishfaq (Mila McGill University), Doina Precup (Mila McGill University)
Representation LearningReinforcement Learning
🎯 What it does: This paper proposes an offline multi-task representation learning algorithm called MORL, which learns shared representations from offline data of multiple tasks in low-rank MDP environments and significantly improves learning efficiency in downstream tasks (reward-free, offline, online).
Offline Oracle-Efficient Learning for Contextual MDPs via Layerwise Exploration-Exploitation Tradeoff
Jian Qian (Massachusetts Institute of Technology), David Simchi-Levi (Massachusetts Institute of Technology)
OptimizationReinforcement LearningTabular
🎯 What it does: The LOLIPOP algorithm is proposed, which achieves optimal learning of hierarchical context MDP through offline density estimation oracle.
Offline Reinforcement Learning with OOD State Correction and OOD Action Suppression
Yixiu Mao (Tsinghua University), Xiangyang Ji (Tsinghua University)
Reinforcement LearningTabular
🎯 What it does: This paper proposes an offline reinforcement learning method named SCAS, which unifies the solutions to the problems of out-of-distribution (OOD) states and out-of-distribution (OOD) actions that arise in offline datasets.
Oja's Algorithm for Streaming Sparse PCA
Syamantak Kumar (University of Texas at Austin), Purnamrita Sarkar (University of Texas at Austin)
OptimizationComputational Efficiency
🎯 What it does: A sparse PCA algorithm is proposed that performs a single pass with O(nd) time complexity and O(d) space complexity, obtaining sparse principal components by thresholding the output of the Oja algorithm.
OMG-LLaVA: Bridging Image-level, Object-level, Pixel-level Reasoning and Understanding
Tao Zhang (Wuhan University), Shuicheng YAN
Object DetectionSegmentationConvolutional Neural NetworkTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelImageTextMultimodality
🎯 What it does: Integrated the general segmentation model OMG-Seg with a large language model (LLM) to construct a single architecture that achieves multimodal understanding and reasoning at the image level, object level, and pixel level;
Omnigrasp: Grasping Diverse Objects with Simulated Humanoids
Zhengyi Luo (Carnegie Mellon University), Weipeng Xu (ETH Zurich)
Robotic IntelligenceReinforcement LearningMultimodality
🎯 What it does: Train a humanoid robot capable of controlling full-body movements with both hands in a simulated environment, to grasp various objects and move along a given trajectory.
OmniJARVIS: Unified Vision-Language-Action Tokenization Enables Open-World Instruction Following Agents
Zihao Wang (Peking University), Yitao Liang (Peking University)
Robotic IntelligenceReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningVision-Language-Action ModelTextMultimodality
🎯 What it does: This paper presents OmniJARVIS, a unified Vision-Language-Action (VLA) model that can achieve instruction following, planning, and execution in the Minecraft open-world environment through chain reasoning and action tagging.
OmniTokenizer: A Joint Image-Video Tokenizer for Visual Generation
Junke Wang (Fudan University), Yu-Gang Jiang (Fudan University)
GenerationTransformerDiffusion modelImageVideo
🎯 What it does: This paper proposes OmniTokenizer, a transformer-based tokenizer that can uniformly handle images and videos, and achieves visual generation based on this.
On $f$-Divergence Principled Domain Adaptation: An Improved Framework
Ziqiao Wang (Tongji University), Yongyi Mao (University of Ottawa)
Domain AdaptationImage
🎯 What it does: This paper proposes an improved domain adaptation theory and algorithm framework based on f-divergence.
On Affine Homotopy between Language Encoders
Robin Chan, Ryan Cotterell (ETH Zürich)
ClassificationRepresentation LearningTransformerLarge Language ModelText
🎯 What it does: This paper proposes and theorizes a framework for affine homotopy comparison between language encoders, defining a semi-metric space and a pre-ordering relation to measure the intrinsic similarity of encoders, and proves that this intrinsic similarity can upper bound the extrinsic similarity of downstream tasks.
On Causal Discovery in the Presence of Deterministic Relations
Loka Li (Mohamed bin Zayed University of Artificial Intelligence), Kun Zhang (Carnegie Mellon University)
Drug DiscoveryScore-based ModelTabular
🎯 What it does: This paper proposes a causal discovery framework called DGES that can address errors caused by violations of faithfulness in traditional methods by detecting deterministic clusters, using an improved Greedy Equivalent Search, and performing local exact searches.
On conditional diffusion models for PDE simulations
Aliaksandra Shysheya (University of Cambridge), Emile Mathieu (University of Cambridge)
Diffusion modelTime SeriesPhysics Related
🎯 What it does: Utilize conditional diffusion models to achieve predictions and data assimilation of partial differential equations (PDEs);
On Convergence of Adam for Stochastic Optimization under Relaxed Assumptions
Yusu Hong (Zhejiang University), Junhong Lin (Zhejiang University)
Optimization
🎯 What it does: This paper provides a rigorous proof of the convergence of the Adam algorithm under generalized noise models (including almost deterministic affine variance noise, bounded noise, and sub-Gaussian noise) for non-convex smooth optimization problems. It presents a high-probability upper bound on the gradient norm and maintains the same convergence rate under the general (L, L0q)-smooth conditions.
On Differentially Private Subspace Estimation in a Distribution-Free Setting
Eliad Tsfadia (Georgetown University)
Safty and PrivacyTabular
🎯 What it does: This paper studies how to estimate the top-k dimensional subspace of a dataset under no distribution assumptions using differential privacy, and proposes two smooth measures to assess the 'ease' of the dataset through multiplicative singular-value gaps, providing upper and lower bounds, along with practical algorithms.
On Differentially Private U Statistics
Kamalika Chaudhuri (University of California San Diego), Purnamrita Sarkar (University of Texas at Austin)
Safty and PrivacyGraph
🎯 What it does: A new differential privacy U-statistic estimation method is proposed, which can achieve nearly optimal private error under both non-degenerate sub-Gaussian kernels and degenerate bounded kernels, and corresponding lower bounds are provided.
On Divergence Measures for Training GFlowNets
Tiago Silva, Diego Mesquita (Getulio Vargas Foundation)
GenerationData SynthesisReinforcement LearningSequential
🎯 What it does: This paper explores methods to change the training objective of GFlowNet to traditional f-divergence (KL, Renyi, Tsallis, etc.), and based on this, designs a control variate technique to reduce gradient estimation variance, validating the effectiveness of this paradigm in various synthetic tasks.
On Feature Learning in Structured State Space Models
Leena Chennuru Vankadara (Amazon), Volkan Cevher (Amazon)
Representation LearningText
🎯 What it does: Analyzed the signal propagation of structured state space models (such as Mamba) in the infinite width limit, derived the maximum update parameterization (µP-SSM) that can achieve feature learning, and experimentally validated its effectiveness.
On Giant's Shoulders: Effortless Weak to Strong by Dynamic Logits Fusion
Chenghao Fan (Huazhong University of Science and Technology), Yu Cheng (Chinese University of Hong Kong)
OptimizationKnowledge DistillationTransformerLarge Language ModelMixture of ExpertsTextBenchmark
🎯 What it does: This paper proposes a dynamic numerical fusion method that integrates small models specialized for multiple tasks into a large model with weighted fusion during inference, achieving strong model knowledge transfer without gradient updates.
On improved Conditioning Mechanisms and Pre-training Strategies for Diffusion Models
Tariq Berrada (Meta), Michal Drozdzal (Meta)
GenerationData SynthesisTransformerDiffusion modelImageMultimodality
🎯 What it does: Research and implementation of a more efficient diffusion model training scheme, proposing a mechanism to decouple semantics and control conditions, and optimizing the pre-training transfer strategy;
On Learning Multi-Modal Forgery Representation for Diffusion Generated Video Detection
Xiufeng Song (Shanghai Jiao Tong University), Xiaohong Liu (Shanghai Jiao Tong University)
RecognitionGenerationAnomaly DetectionTransformerLarge Language ModelDiffusion modelVideoMultimodality
🎯 What it does: A detection framework for video generation based on diffusion models, called MM-Det, is proposed and implemented, which combines multi-modal forgery representation and spatiotemporal attention mechanisms to achieve video-level forgery detection.
On Mesa-Optimization in Autoregressively Trained Transformers: Emergence and Capability
Chenyu Zheng (Renmin University of China), Chongxuan Li (Tsinghua University)
OptimizationTransformerTime SeriesSequentialOrdinary Differential Equation
🎯 What it does: This study investigates whether a 'Mesa-optimizer' appears in autoregressive pre-trained transformers and conducts a rigorous analysis of its training dynamics through gradient flow on a single-layer linear self-attention model; it proves that under specific data distribution conditions, the model converges to a structure that implements a single gradient descent and explores its capacity limits.
On Neural Networks as Infinite Tree-Structured Probabilistic Graphical Models
Boyao Li (Duke University), David Page (Duke University)
Tabular
🎯 What it does: This paper proposes mapping arbitrary structured deep neural networks (DNNs) to infinite-width tree-structured probabilistic graphical models (PGMs), proving that the forward propagation of DNNs is equivalent to the exact inference of the PGM, and based on this, designs a new Hamiltonian Monte Carlo (HMC) fine-tuning algorithm;
On provable privacy vulnerabilities of graph representations
Ruofan Wu (Ant Group), Weiqiang Wang (Ant Group)
Safty and PrivacyRepresentation LearningAdversarial AttackGraph Neural NetworkContrastive LearningGraph
🎯 What it does: This paper studies privacy vulnerabilities in graph representation learning, conducting theoretical analysis and experimental validation on similarity-based edge reconstruction attacks (SERA), and exploring the feasibility of using noise aggregation (NAG) to counter this attack.
On Sampling Strategies for Spectral Model Sharding
Denis Korzhenkov (Qualcomm AI Research), Christos Louizos (Qualcomm AI Research)
OptimizationFederated LearningTransformerImageText
🎯 What it does: Two SVD-based spectral model partition sampling strategies (Unbiased and Collective) are proposed, combined with auxiliary weights and learning rate clipping to achieve more efficient federated learning on heterogeneous devices.
On scalable oversight with weak LLMs judging strong LLMs
Zachary Kenton (Google DeepMind), Rohin Shah (Google DeepMind)
TransformerLarge Language ModelReinforcement LearningPrompt EngineeringTextMultimodality
🎯 What it does: This study investigates the reasoning performance of scalable supervision protocols such as Debate and Consultancy when a weak large language model (LLM) acts as a judge supervising a strong LLM agent, covering extractive, closed, and multimodal multitasking.
On Socially Fair Low-Rank Approximation and Column Subset Selection
Zhao Song (Simons Institute for the Theory of Computing UC Berkeley), Samson Zhou (Texas A&M University)
OptimizationTabularFinance Related
🎯 What it does: This paper studies the low-rank approximation (fair LRA) and column subset selection (fair CSS) problems under the constraint of 'social fairness'. It proves that constant factor approximation is not achievable in polynomial time and provides a (1+ε) approximation algorithm with a time complexity of 2 poly(k) for constant numbers, as well as a bicriteria algorithm that is acceptable in polynomial time. The algorithm's advantages in fairness and efficiency are validated on real data.
On Softmax Direct Preference Optimization for Recommendation
Yuxin Chen (National University of Singapore), Tat-Seng Chua (National University of Singapore)
Recommendation SystemOptimizationRecurrent Neural NetworkSupervised Fine-TuningSequential
🎯 What it does: This paper proposes the Softmax-DPO (S-DPO) loss, which aligns preferences for language models using multiple negative samples to enhance the personalized ranking performance of recommendation systems.
On Sparse Canonical Correlation Analysis
Yongchun Li (University of Tennessee), Weijun Xie (Georgia Tech)
OptimizationTabularBiomedical Data
🎯 What it does: This paper studies the theory and algorithms of Sparse Canonical Correlation Analysis (SCCA), proposing a combinatorial and MISDP formulation and providing both approximate and exact solution methods.
On Statistical Rates and Provably Efficient Criteria of Latent Diffusion Transformers (DiTs)
Jerry Yao-Chieh Hu (Northwestern University), Han Liu (Northwestern University)
TransformerDiffusion model
🎯 What it does: This paper studies the statistical and computational limits of the latent diffusion Transformer (DiT) under low-dimensional linear latent spaces.
On the Ability of Developers' Training Data Preservation of Learnware
Hao-Yi Lei (Nanjing University), Zhi-Hua Zhou (Nanjing University)
RetrievalOptimizationSafty and Privacy
🎯 What it does: This paper theoretically analyzes the data retention capability of the RKME specification used in the Learnware paradigm, proving that under reasonable sizes, RKME neither leaks the original data nor loses sufficient information for model retrieval.
On the Adversarial Robustness of Benjamini Hochberg
Louis Chen, Matan Seri (Naval Postgraduate School)
Adversarial AttackTabularFinance Related
🎯 What it does: This paper studies the robustness of the Benjamini–Hochberg (BH) multiple testing procedure in the face of adversarial perturbations and proposes a simple perturbation algorithm INCREASE‑c, which aims to maximize the false discovery rate (FDR) of BH by changing at most c p-values during the testing phase.
On the Benefits of Public Representations for Private Transfer Learning under Distribution Shift
Pratiksha Thaker (Carnegie Mellon University), Virginia Smith (Carnegie Mellon University)
Domain AdaptationFederated LearningSafty and PrivacyTransformerContrastive LearningImageBiomedical Data
🎯 What it does: The study investigates how public pre-training can enhance the effectiveness of differential privacy (DP) transfer learning in the presence of significant distribution shifts between public and private tasks, and demonstrates its validity through experiments and theoretical proofs.
On the cohesion and separability of average-link for hierarchical agglomerative clustering
Eduardo Sany Laber, Miguel A. Batista
🎯 What it does: This paper conducts a systematic theoretical and experimental study on the clustering performance of the average-link method in metric spaces, proposing and analyzing various clustering evaluation metrics that reflect both cohesion and separability.
On the Comparison between Multi-modal and Single-modal Contrastive Learning
Wei Huang (RIKEN AIP), Taiji Suzuki (University of Tokyo)
OptimizationRepresentation LearningContrastive LearningImageTextMultimodality
🎯 What it does: This paper proposes a unified theoretical framework for feature learning, comparing the optimization processes and downstream task generalization performance of single-modal and multi-modal contrastive learning under gradient descent training, and experimentally validates the theoretical conclusions.
On the Complexity of Identification in Linear Structural Causal Models
Julian Dörfler (Saarland University), Maciej Liskiewicz (University of Lübeck)
Computational EfficiencyTabular
🎯 What it does: This paper proposes a new algorithm for parameter identification in linear structural causal models, aiming to estimate model parameters from observed data and the graphical structure assumptions of the model.
On the Complexity of Learning Sparse Functions with Statistical and Gradient Queries
Nirmit Joshi (Toyota Technological Institute at Chicago), Nathan Srebro (Toyota Technological Institute at Chicago)
Tabular
🎯 What it does: This paper studies the complexity of learning sparse functions (juntas) using gradient queries under a general product distribution and proposes a Differentiable Learning Query (DLQ) model.
On the Complexity of Teaching a Family of Linear Behavior Cloning Learners
Shubham Kumar Bharti, Jerry Zhu
Optimization
🎯 What it does: This paper studies how to achieve optimal teaching for a family of linear behavior cloning (BC) learners using the smallest scale of demonstration data, thereby enabling all consistent learners to learn the target policy from a teaching dataset.
On the Computational Complexity of Private High-dimensional Model Selection
Saptarshi Roy (University of Michigan), Ambuj Tewari (University of Michigan)
OptimizationSafty and PrivacyTabular
🎯 What it does: Proposed different private optimal subset selection algorithms under high-dimensional sparse linear regression models, implementing model selection by combining the exponential mechanism and Metropolis–Hastings sampling.
On the Computational Landscape of Replicable Learning
Alkis Kalavasis (Yale University), Felix Zhou (Yale University)
Supervised Fine-TuningSequential
🎯 What it does: This paper studies the computational aspects of algorithmic replicability, exploring the computational connections between replicability and learning paradigms such as online learning, private learning, and SQ learning.
On the Convergence of Loss and Uncertainty-based Active Learning Algorithms
Daniel Haimovich (Meta), Milan Vojnovic (London School of Economics)
OptimizationTabular
🎯 What it does: The study investigates the convergence and sample efficiency of SGD algorithms based on loss or uncertainty sampling, and proposes the Adaptive-Weight Sampling (AWS) algorithm.
On the Curses of Future and History in Future-dependent Value Functions for Off-policy Evaluation
Yuheng Zhang (University of Illinois Urbana-Champaign), Nan Jiang (University of Illinois Urbana-Champaign)
Reinforcement Learning
🎯 What it does: This paper studies the theory and algorithms for offline evaluation (OPE) in partially observable environments (POMDPs), proposing a new coverage assumption to avoid the exponential 'time step' dependence of traditional methods.
On the Efficiency of ERM in Feature Learning
Ayoub El Hanchi (University of Toronto), Murat A Erdogdu
OptimizationRepresentation LearningSupervised Fine-TuningTabular
🎯 What it does: This study investigates the performance of empirical risk minimization (ERM) in regression problems, particularly in joint learning induced by a set of feature mappings.
On the Expressive Power of Tree-Structured Probabilistic Circuits
Lang Yin (University of Illinois), Han Zhao (University of Illinois)
🎯 What it does: This paper studies the expressive power of tree-structured probabilistic circuits (PCs) and proves that for n variables, there exists a quasi-polynomial upper bound of n^O(log n) for tree-structured PCs that can compute the same probability distribution. Additionally, under depth constraints, there is a super-polynomial separation between tree-structured PCs and DAG-structured PCs.
On the Expressivity and Sample Complexity of Node-Individualized Graph Neural Networks
Paolo Pellizzoni (Max Planck Institute of Biochemistry), Karsten Borgwardt (Max Planck Institute of Biochemistry)
Graph Neural NetworkGraphBiomedical Data
🎯 What it does: This study analyzes the expressiveness and sample complexity of Graph Neural Networks (GNN) using node personalization techniques, proposing upper bounds on VC dimension and covering number, and designing a low VC dimension EGONN architecture.
On the Identifiability of Hybrid Deep Generative Models: Meta-Learning as a Solution
Yubo Ye (Zhejiang University), Linwei Wang (Rochester Institute of Technology)
GenerationMeta LearningAuto EncoderVideoPhysics Related
🎯 What it does: Investigate the identifiability of Hybrid Deep Generative Models (Hybrid-DGMs) and propose a meta-learning based identifiable Hybrid-DGM framework;
On the Identifiability of Poisson Branching Structural Causal Model Using Probability Generating Function
Yu Xiang (Guangdong University of Technology), Zhifeng Hao (Shantou University)
Tabular
🎯 What it does: This paper studies the identifiability of the Poisson Branching Structure Causal Model (PB-SCM), proposing the use of Probability Generating Functions (PGF) to solve its closed-form solution, and designing a rank test algorithm based on the local features of that solution to learn the causal structure.
On the Impact of Feature Heterophily on Link Prediction with Graph Neural Networks
Jiong Zhu (University of Michigan), Danai Koutra (University of Michigan)
Graph Neural NetworkGraph
🎯 What it does: This study investigates the impact of feature heterogeneity on graph neural networks in link prediction and systematically analyzes the design of encoders and decoders at different levels of feature homogeneity.
On the Impacts of the Random Initialization in the Neural Tangent Kernel Theory
Guhan Chen (Tsinghua University), Qian Lin (Tsinghua University)
Image
🎯 What it does: This paper explores the impact of random initialization of neural networks on the Neural Tangent Kernel (NTK) theory, pointing out that this impact has been overlooked in existing research.
On the Inductive Bias of Stacking Towards Improving Reasoning
Nikunj Saunshi (Google Research), Sanjiv Kumar (Google Research)
OptimizationComputational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This study investigates the impact of hierarchical stacking training on the training efficiency and inference performance of large language models, and proposes the MIDAS method, which replicates intermediate layers.
On the Limitations of Fractal Dimension as a Measure of Generalization
Charlie Tan, Anthea Monod (Imperial College London)
Convolutional Neural NetworkImageTabular
🎯 What it does: Conduct large-scale experimental evaluations of the generalization metric based on persistent homology (PH) dimensions, examining its correlation with generalization error and uncovering potential failure modes.
On the Minimax Regret for Contextual Linear Bandits and Multi-Armed Bandits with Expert Advice
Shinji Ito (University of Tokyo)
Reinforcement Learning
🎯 What it does: This paper conducts a theoretical analysis of two types of extended multi-armed bandit problems: one is the bandit with expert advice (BwE), and the other is the contextual linear bandit (CLB).
On the Necessity of Collaboration for Online Model Selection with Decentralized Data
Junfan Li (Harbin Institute of Technology), Irwin King (Chinese University of Hong Kong)
OptimizationFederated LearningTabular
🎯 What it does: The study investigates whether online model selection (OMS) requires collaboration in a distributed data environment and provides a computational constraint threshold for the necessity of collaboration.
On the Noise Robustness of In-Context Learning for Text Generation
Hongfu Gao (Southern University of Science and Technology), Hongxin Wei (Southern University of Science and Technology)
GenerationText
🎯 What it does: Analyzes and addresses the decline in contextual learning performance caused by noise in demonstration data for text generation tasks, proposing the Local Perplexity Ranking (LPR) method to enhance noise robustness.
On the Optimal Time Complexities in Decentralized Stochastic Asynchronous Optimization
Alexander Tyurin (King Abdullah University of Science and Technology), Peter Richtárik (King Abdullah University of Science and Technology)
OptimizationTabularStochastic Differential Equation
🎯 What it does: A new lower bound on time complexity is presented for distributed asynchronous stochastic optimization, addressing the limitations of computation and communication speed, along with the introduction of the Fragile SGD and Amelie SGD algorithms that can converge in any heterogeneous computing/communication environment.
On the Optimality of Dilated Entropy and Lower Bounds for Online Learning in Extensive-Form Games
Zhiyuan Fan (Massachusetts Institute of Technology), Gabriele Farina (Massachusetts Institute of Technology)
OptimizationReinforcement Learning from Human Feedback
🎯 What it does: The paper studies the use of Dilated Entropy (DilEnt) as a Distance Generating Function (DGF) in Extensive Form Games (EFG) and proves that it can achieve optimal convergence rates in first-order methods (OMD, Clairvoyant OMD, etc.). By introducing a tree-structured dual norm, the paper provides upper bounds on the strong convexity of DilEnt and the optimal ratio of the decision space diameter, along with corresponding lower bounds, indicating that this ratio is nearly optimal.
On the Parameter Identifiability of Partially Observed Linear Causal Models
Xinshuai Dong (Carnegie Mellon University), Kun Zhang (Carnegie Mellon University)
Optimization
🎯 What it does: This paper studies the parameter identifiability of partially observable linear causal models and proposes a new parameter estimation method based on this.
On the Power of Decision Trees in Auto-Regressive Language Modeling
Yulu Gan (Massachusetts Institute of Technology), eran malach
GenerationExplainability and InterpretabilityTransformerTextChain-of-Thought
🎯 What it does: Proposes and implements Auto-Regressive Decision Trees (ARDTs) as an interpretable language model framework, completing both theoretical analysis and empirical experiments.
On the Power of Small-size Graph Neural Networks for Linear Programming
Qian Li (Shenzhen International Center For Industrial And Applied Mathematics), Ruoyu Sun (Shenzhen International Center For Industrial And Applied Mathematics)
OptimizationGraph Neural NetworkGraph
🎯 What it does: A graph neural network architecture GD-Net is proposed for Packing and Covering linear programming, and it is proven that a polylogarithmic depth and constant width GNN can achieve arbitrarily precise approximate solutions.
On the Robustness of Spectral Algorithms for Semirandom Stochastic Block Models
Aditya Bhaskara (University of Utah), Weronika Wrzos-Kaminska (École Polytechnique Fédérale de Lausanne)
Graph
🎯 What it does: The robustness of spectral algorithms in the semi-random Stochastic Block Model is studied, proving that under certain models, the unnormalized Laplacian can achieve exact recovery, while the normalized Laplacian will incur a constant error.
On the Role of Attention Masks and LayerNorm in Transformers
Xinyi Wu (Massachusetts Institute of Technology), Ali Jadbabaie (Massachusetts Institute of Technology)
TransformerText
🎯 What it does: This paper studies the self-attention mechanism in Transformers, exploring the impact of attention masks and LayerNorm on the rank collapse of token representations, and provides new insights through theoretical analysis and constructed counterexamples.
On the Role of Information Structure in Reinforcement Learning for Partially-Observable Sequential Teams and Games
Awni Altabaa (Yale University), Zhuoran Yang (Yale University)
Reinforcement LearningSequential
🎯 What it does: A new framework for explicitly modeling information structures in reinforcement learning (POST/POSG) is proposed, along with a statistical complexity upper bound based on this structure.
On the Saturation Effects of Spectral Algorithms in Large Dimensions
Weihao Lu (Tsinghua University), Qian Lin (Tsinghua University)
🎯 What it does: The study investigates the saturation effect of spectral algorithms (including Kernel Ridge Regression (KRR) and gradient flow) under over-smooth regression functions in high-dimensional conditions (n ≍ d^γ), providing the asymptotic optimal convergence rate of gradient flow and improving the information lower bound.
On the Scalability of Certified Adversarial Robustness with Generated Data
Thomas Altstidl (Friedrich Alexander University Erlangen Nuremberg), Leo Schwinn (Technical University of Munich)
ClassificationData SynthesisAdversarial AttackDiffusion modelImage
🎯 What it does: By incorporating data generated by diffusion models into the training process, the certified robustness based on deterministic Lipschitz-bound has been improved, and a systematic evaluation of the impact of factors such as data volume, model size, and training cycles on robustness has been conducted.
On the Scalability of GNNs for Molecular Graphs
Maciej Sypetkowski (Valence Labs), Dominique Beaini (Valence Labs)
Drug DiscoveryGraph Neural NetworkTransformerSupervised Fine-TuningGraph
🎯 What it does: This paper systematically studies the scalability of molecular graph neural networks under large-scale supervised pre-training, comparing three architectures: MPNN++, Transformer, and GPS++. It also proposes the MolGPS foundational model for multi-fingerprint detection.
On the Sparsity of the Strong Lottery Ticket Hypothesis
Emanuele Natale (Universite Cote dAzur), Frederik Mallmann-Trenn (King's College London)
🎯 What it does: The paper theoretically proves the quantitative relationship between the sparsity of subnetworks and the over-parameterization of the original network in the Strong Lottery Ticket Hypothesis, and for the first time provides strict upper and lower bounds for the Random Fixed-Size Subset Sum (RFSS); subsequently, it uses this result to present a strong lottery ticket hypothesis theorem with controllable sparsity for dense networks and equivariant networks.
On the Stability and Generalization of Meta-Learning
Yunjuan Wang (Johns Hopkins University), Raman Arora (Johns Hopkins University)
Federated LearningMeta Learning
🎯 What it does: A new unified definition of Meta-stability is proposed, analyzing the generalization performance of Meta-learning algorithms and providing a tighter upper bound on generalization error.
On the Surprising Effectiveness of Attention Transfer for Vision Transformers
Alexander Cong Li, Xinlei Chen (Facebook AI Research)
ClassificationObject DetectionDomain AdaptationKnowledge DistillationTransformerSupervised Fine-TuningImage
🎯 What it does: A method is proposed and validated that replaces traditional Fine-Tuning by only transferring attention patterns (Attention Transfer), where the student model uses the attention information from the pre-trained teacher while learning its own features, significantly improving the performance of ViT on downstream tasks.
On the Target-kernel Alignment: a Unified Analysis with Kernel Complexity
Chao Wang (Shanghai University of Finance and Economics), Junhui Wang (Chinese University of Hong Kong)
OptimizationTabular
🎯 What it does: This paper studies the impact of the objective function and kernel matrix on the performance of various kernel methods (general loss function family) and presents a theoretical analysis of the Truncated Kernel Method (TKM) at different alignment levels.
On the Use of Anchoring for Training Vision Models
Vivek Narayanaswamy (Lawrence Livermore National Laboratory), Jayaraman J. Thiagarajan (Lawrence Livermore National Laboratory)
ClassificationAnomaly DetectionConvolutional Neural NetworkTransformerSupervised Fine-TuningImage
🎯 What it does: Proposed and evaluated the Anchoring training protocol, combined with a new reference masking regularization to enhance the generalization and safety of visual models.
On the Worst Prompt Performance of Large Language Models
Bowen Cao (Chinese University of Hong Kong), Wai Lam (Chinese University of Hong Kong)
TransformerLarge Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: This paper proposes a new benchmark, ROBUSTALPACAEVAL, to evaluate the worst performance of LLMs under prompts with varying semantic equivalence and fluency, and conducts large-scale experiments on ChatGPT and six open-source LLMs.
On Tractable $\Phi$-Equilibria in Non-Concave Games
Yang Cai (Yale University), Weiqiang Zheng (Yale University)
Optimization
🎯 What it does: This paper studies computable Φ-equilibria in non-convex games and proposes efficient online learning algorithms to approximate different types of Φ-equilibria, including finite global changes, projections, convex combinations, and interpolated local changes.
On Weak Regret Analysis for Dueling Bandits
El Mehdi Saad (King Abdullah University of Science and Technology), Nicolas Verzelen (Institut National de la Recherche Agronomique)
Recommendation SystemOptimization
🎯 What it does: This paper studies how to minimize weak regret in dueling bandits in the presence of a Condorcet winner, proposing two new algorithms WR-TINF and WR-EXP3-IX, providing corresponding upper bounds and proving lower bounds that demonstrate their optimality in certain instances.
On-Road Object Importance Estimation: A New Dataset and A Model with Multi-Fold Top-Down Guidance
Zhixiong Nan (Chongqing University), Tao Xiang (Chongqing University)
Object DetectionAutonomous DrivingConvolutional Neural NetworkRecurrent Neural NetworkSupervised Fine-TuningVideoMultimodality
🎯 What it does: This paper proposes a method for estimating object importance in road scenes based on driver perspective videos and releases a larger Traffic Object Importance (TOI) dataset.
Once Read is Enough: Domain-specific Pretraining-free Language Models with Cluster-guided Sparse Experts for Long-tail Domain Knowledge
Fang Dong (Fudan University), Li Shang (Fudan University)
TransformerLarge Language ModelMixture of ExpertsTextBiomedical DataFinance Related
🎯 What it does: Proposes the Cluster-guided Sparse Expert (CSE) layer, which allocates rare domain knowledge to specialized expert networks through clustering during pre-training, achieving a language model without domain-specific pre-training;
One for All: Multi-Domain Joint Training for Point Cloud Based 3D Object Detection
Zhenyu Wang, Shengjin Wang
Object DetectionDomain AdaptationAutonomous DrivingPoint Cloud
🎯 What it does: This paper presents OneDet3D, a unified 3D object detection model capable of simultaneously processing indoor and outdoor point clouds after a single multi-domain joint training, using a single set of parameters for multi-domain detection.
One Sample Fits All: Approximating All Probabilistic Values Simultaneously and Efficiently
Weida Li (National University of Singapore), Yaoliang Yu (University of Waterloo)
OptimizationExplainability and InterpretabilityComputational EfficiencyTabular
🎯 What it does: A framework (OFA) is proposed that can approximate all probability values simultaneously through a single sampling, providing an optimal sampling vector that ensures no amplification factor and fully utilizes the samples;
One Token to Seg Them All: Language Instructed Reasoning Segmentation in Videos
Zechen Bai (National University of Singapore), Mike Zheng Shou (National University of Singapore)
SegmentationTransformerLarge Language ModelVision Language ModelVideoMultimodality
🎯 What it does: This paper presents VideoLISA, a video instruction reasoning segmentation model based on a multimodal large language model, capable of generating cross-frame semantic segmentation masks based on language descriptions.
One-Layer Transformer Provably Learns One-Nearest Neighbor In Context
Zihao Li (Princeton University), Mengdi Wang (Princeton University)
ClassificationOptimizationTransformerTabular
🎯 What it does: This study investigates a single-layer Transformer (containing only the Softmax Attention layer) learning and approximating a nearest neighbor (1-NN) predictor in the In-Context Learning scenario, providing a theoretical proof of gradient descent converging to zero loss, and demonstrating that it maintains 1-NN behavior under distribution shift.
One-shot Federated Learning via Synthetic Distiller-Distillate Communication
Junyuan Zhang (National University of Singapore), Xinchao Wang (National University of Singapore)
Federated LearningSafty and PrivacyKnowledge DistillationConvolutional Neural NetworkRecurrent Neural NetworkAuto EncoderImageBiomedical DataMagnetic Resonance Imaging
🎯 What it does: Design and implement a one-round federated learning framework FedSD2C, which uses V-information core set extraction, local Fourier amplitude perturbation, and a pre-trained autoencoder to synthesize privacy-preserving distilled samples, and sends these samples directly to the server for model training.
One-Shot Safety Alignment for Large Language Models via Optimal Dualization
Xinmeng Huang (University of Pennsylvania), Dongsheng Ding (University of Pennsylvania)
OptimizationSafty and PrivacyReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: A one-shot safe alignment method is proposed, transforming the constrained language model alignment problem into an unconstrained alignment problem. The optimal Lagrange multipliers are solved through a closed-form dual function, achieving compliance with safety constraints with only one training session.
One-Step Diffusion Distillation through Score Implicit Matching
Weijian Luo (Peking University), Guo-Jun Qi
GenerationKnowledge DistillationDiffusion modelScore-based ModelImageText
🎯 What it does: This paper proposes a data-free distillation method called Score Implicit Matching (SIM), which compresses multi-step diffusion models into a single-step generator.
One-Step Effective Diffusion Network for Real-World Image Super-Resolution
Rongyuan Wu (Hong Kong Polytechnic University), Lei Zhang (Hong Kong Polytechnic University)
RestorationSuper ResolutionKnowledge DistillationDiffusion modelImage
🎯 What it does: OSEDiff is proposed, a one-shot diffusion network that uses low-quality images directly as the diffusion starting point to recover high-quality images, eliminating the uncertainty of random noise and significantly reducing computational costs.
One-to-Multiple: A Progressive Style Transfer Unsupervised Domain-Adaptive Framework for Kidney Tumor Segmentation
Kai Hu (Xiangtan University), Xieping Gao (Hunan Normal University)
SegmentationDomain AdaptationGenerative Adversarial NetworkImageBiomedical DataMagnetic Resonance ImagingComputed Tomography
🎯 What it does: A one-to-many unsupervised domain adaptation framework called PSTUDA is proposed for multi-sequence MRI kidney tumor segmentation.
One-to-Normal: Anomaly Personalization for Few-shot Anomaly Detection
Yiyue Li (West China Biomedical Big Data Center), Qicheng Lao (Beijing University of Posts and Telecommunications)
Anomaly DetectionPrompt EngineeringDiffusion modelContrastive LearningImageBiomedical DataMagnetic Resonance ImagingComputed Tomography
🎯 What it does: A one-to-one 'one-to-normal' personalized method is proposed for few-shot anomaly detection, utilizing a customized diffusion model to transform query images into defect-free versions, and enhancing accuracy through triplet contrastive reasoning.