arXivSub Start free trial

NeurIPS 2023 Papers — Page 31

Conference on Neural Information Processing Systems · 3218 papers

Transferable Adversarial Robustness for Categorical Data via Universal Robust Embeddings

Klim Kireev (École Polytechnique Fédérale de Lausanne), Nicolas Flammarion (École Polytechnique Fédérale de Lausanne)

ClassificationAnomaly DetectionOptimizationAdversarial AttackTabularFinance Related

🎯 What it does: A trainable robust neural network and a general robust embedding transferable to tree models are proposed for tabular data with categorical features, supporting a realistic threat model based on financial costs.

Transformed Low-Rank Parameterization Can Help Robust Generalization for Tensor Neural Networks

Andong Wang (RIKEN Artificial Intelligence Project), Qibin Zhao (RIKEN Artificial Intelligence Project)

ClassificationAdversarial AttackImage

🎯 What it does: This paper proposes and theoretically analyzes the standard and adversarial generalization performance of neural networks with t-product layers (t-NN), and enhances robustness by introducing low-rank parameterization.

Transformer as a hippocampal memory consolidation model based on NMDAR-inspired nonlinearity

Dong-Kyum Kim (Institute for Basic Science), C. Justin Lee (Institute for Basic Science)

TransformerSequential

🎯 What it does: A novel activation function based on the nonlinear dynamics of NMDA receptors is proposed, which is embedded in the feedforward network of the Transformer to simulate the hippocampal memory integration process and enhance long-term reference memory.

Transformer-based Planning for Symbolic Regression

Parshin Shojaee (Virginia Tech), Chandan K. Reddy (Virginia Tech)

TransformerReinforcement LearningTabularBenchmark

🎯 What it does: This paper proposes TPSR, which combines Monte Carlo Tree Search (MCTS) planning with pre-trained Transformers to introduce non-differentiable feedback on fitting accuracy and complexity in the generated program sequences during symbolic regression.

Transformers are uninterpretable with myopic methods: a case study with bounded Dyck grammars

Kaiyue Wen (Tsinghua University), Andrej Risteski (Carnegie Mellon University)

Explainability and InterpretabilityTransformerSequential

🎯 What it does: This paper conducts theoretical and experimental research on the Transformer in the synthetic grammar task of Dyck language, demonstrating that merely examining attention patterns or single weights cannot explain the model's functionality.

Transformers as Statisticians: Provable In-Context Learning with In-Context Algorithm Selection

Yu Bai (Salesforce Research), Song Mei (University of California Berkeley)

TransformerTabular

🎯 What it does: This paper proves and implements that transformers can approximately execute various classic machine learning algorithms (ridge regression, Lasso, generalized linear models, two-layer neural network gradient descent) and an adaptive algorithm selection process in in-context learning (ICL), ultimately achieving near-Bayes optimal predictions.

Transformers learn through gradual rank increase

Enric Boix-Adserà (Massachusetts Institute of Technology), Joshua M. Susskind (Apple)

TransformerImageText

🎯 What it does: During the training process of the Transformer, an incremental learning dynamic was discovered and proven, indicating that the difference between weights and initial values gradually increases with training, and the rank progressively improves.

Transformers learn to implement preconditioned gradient descent for in-context learning

Kwangjun Ahn (Massachusetts Institute of Technology), Suvrit Sra (Technical University of Munich)

OptimizationTransformerTabular

🎯 What it does: The study investigates the training of Transformer on random linear regression instances, proving that a single-layer Transformer corresponds to a global optimum with one-step preconditioned gradient descent, while deeper layers correspond to multi-step preconditioned gradient methods.

Transformers over Directed Acyclic Graphs

Yuankai Luo (Beihang University), Lei Shi (Beihang University)

Graph Neural NetworkTransformerGraph

🎯 What it does: An improved Transformer architecture for directed acyclic graphs (DAGs) is proposed, enabling it to capture partial sequential relationships of DAGs.

TransHP: Image Classification with Hierarchical Prompting

Wenhao Wang (University of Technology Sydney), Yi Yang (Zhejiang University)

ClassificationTransformerPrompt EngineeringImage

🎯 What it does: This paper proposes a Transformer with Hierarchical Prompting (TransHP), which introduces a hierarchical prompting mechanism in visual Transformers. It first predicts coarse categories through intermediate layers and then injects the corresponding prompt tokens into subsequent layers to enhance fine category classification performance.

Transient Neural Radiance Fields for Lidar View Synthesis and 3D Reconstruction

Anagh Malik (University of Toronto), David B. Lindell (University of Toronto)

Data SynthesisDepth EstimationNeural Radiance FieldPoint Cloud

🎯 What it does: By constructing a transient neural radiance field model, we perform view synthesis and 3D reconstruction of the time-domain photon counting histogram from single-photon lidar.

Transition-constant Normalization for Image Enhancement

Jie Huang (University of Science and Technology of China), Feng Zhao (University of Science and Technology of China)

RestorationImage

🎯 What it does: A new normalization operation called Transition-constant Normalization (TCN) is proposed for image enhancement tasks, balancing illumination consistency modeling and the reversibility of information transfer.

Transitivity Recovering Decompositions: Interpretable and Robust Fine-Grained Relationships

Abhra Chaudhuri (University of Exeter), Anjan Dutta (The Alan Turing Institute)

ClassificationExplainability and InterpretabilityGraph Neural NetworkTransformerImageAgriculture Related

🎯 What it does: An interpretable fine-grained visual classification method called TRD has been developed, which deconstructs abstract relational representations into graph structures and learns directly in the graph space, achieving complete interpretability at both the instance and category levels.

Transportability for Bandits with Data from Different Environments

Alexis Bellot (Google DeepMind), Silvia Chiappa (Google DeepMind)

Reinforcement LearningTabular

🎯 What it does: A method is proposed to improve multi-armed bandit (MAB) learning by utilizing batch data from different environments and causal structure assumptions;

Tree of Thoughts: Deliberate Problem Solving with Large Language Models

Shunyu Yao (Princeton University), Karthik R Narasimhan

TransformerLarge Language ModelPrompt EngineeringTextChain-of-Thought

🎯 What it does: Proposes the Tree of Thoughts (ToT) framework, which extends the reasoning of large language models from the word level to multi-path exploration and self-evaluation at the 'thought' level through tree search, in order to enhance the problem-solving ability for complex issues.

Tree Variational Autoencoders

Laura Manduchi (ETH Zurich), Julia E Vogt

GenerationData SynthesisAuto EncoderContrastive LearningImageText

🎯 What it does: This paper proposes TreeVAE, a generative hierarchical clustering model that achieves hierarchical partitioning of samples and generates corresponding cluster samples by learning tree-structured posterior distributions in an unsupervised manner.

Tree-Based Diffusion Schrödinger Bridge with Applications to Wasserstein Barycenters

Maxence Noble (École polytechnique), Alain Durmus

Diffusion modelImageTabularStochastic Differential Equation

🎯 What it does: This paper proposes an extended Diffusion Schrödinger Bridge based on a tree structure (TreeDSB) for solving the optimal transport problem with multi-marginal entropy regularization in continuous state spaces.

Tree-Rings Watermarks: Invisible Fingerprints for Diffusion Images

Yuxin Wen (University of Maryland), Tom Goldstein (University of Maryland)

GenerationData SynthesisDiffusion modelImage

🎯 What it does: During the sampling process of the diffusion model, an 'tree ring' pattern is embedded in the Fourier domain of the initial noise vector, achieving an invisible watermark on the generated content without altering the generated images;

TRIAGE: Characterizing and auditing training data for improved regression

Nabeel Seedat (University of Cambridge), Mihaela van der Schaar (University of Cambridge)

Data-Centric LearningTabularBiomedical Data

🎯 What it does: The TRIAGE framework is proposed for training data representation and auditing in regression tasks.

Trial matching: capturing variability with data-constrained spiking neural networks

Christos Sourmpis (École Polytechnique Fédérale de Lausanne), Guillaume Bellec (École Polytechnique Fédérale de Lausanne)

Spiking Neural NetworkTime Series

🎯 What it does: By constructing a large-scale data-constrained recursive synaptic neural network, we fit multi-session electrophysiological recordings to capture the trial-to-trial variability of neural activity and behavior.

Triangulation Residual Loss for Data-efficient 3D Pose Estimation

Jiachen Zhao (Tsinghua University), Qionghai Dai (Tsinghua University)

Pose EstimationImage

🎯 What it does: A Triangulation Residual Loss (TR loss) is proposed for multi-view 3D pose estimation, achieving unsupervised global geometric consistency training by minimizing the distance between 3D estimated points and rays from all views.

Triple Eagle: Simple, Fast and Practical Budget-Feasible Mechanisms

Kai Han (Soochow University), Shuang Cui (University of Science and Technology of China)

Recommendation SystemOptimizationGraph

🎯 What it does: A budget-feasible mechanism framework named TripleEagle is proposed, which designs random, deterministic, and non-monotonic versions of the mechanism for selection problems under budget constraints for monotonic and non-monotonic submodular functions, capable of completion within one price query and O(n) value queries.

TriRE: A Multi-Mechanism Learning Paradigm for Continual Knowledge Retention and Promotion

Preetha Vijayan (NavInfo Europe), Elahe Arani (Eindhoven University of Technology)

Image

🎯 What it does: A three-stage continual learning paradigm called TriRE is proposed, utilizing three mechanisms: retain, revise, and rewind to achieve the preservation, revision, and reshaping of task knowledge.

TrojLLM: A Black-box Trojan Prompt Attack on Large Language Models

Jiaqi Xue (University of Central Florida), Qian Lou (University of Central Florida)

ClassificationAdversarial AttackTransformerLarge Language ModelReinforcement LearningPrompt EngineeringText

🎯 What it does: This paper proposes a black-box backdoor trigger attack framework for large language model (LLM) APIs called TrojLLM, which can automatically generate general and covert trigger words and embed them into discrete prompts, leading to malicious manipulation of LLM outputs.

Truly Scale-Equivariant Deep Nets with Fourier Layers

Md Ashiqur Rahman (Purdue University), Raymond A. Yeh (Purdue University)

ClassificationConvolutional Neural NetworkImage

🎯 What it does: This paper proposes a truly scale-equivariant deep network implemented using Fourier layers, which can maintain zero error under scale variations;

Truncated Affinity Maximization: One-class Homophily Modeling for Graph Anomaly Detection

Hezhe Qiao (Singapore Management University), Guansong Pang (Singapore Management University)

Anomaly DetectionGraph Neural NetworkGraph

🎯 What it does: In unsupervised graph anomaly detection, a local node affinity anomaly scoring based on one-class homophily is proposed, and specialized node representations are learned through the Truncated Affinity Maximization (TAM) method, significantly improving anomaly detection performance.

Truncating Trajectories in Monte Carlo Policy Evaluation: an Adaptive Approach

Riccardo Poiani (Politecnico di Milano), Marcello Restelli (Politecnico di Milano)

OptimizationReinforcement Learning

🎯 What it does: This paper studies how to minimize estimation variance by truncating trajectories and dynamically adjusting sampling strategies when evaluating policy performance under a limited interaction budget using Monte Carlo methods, and proposes an adaptive algorithm called RIDO.

Trust Region-Based Safe Distributional Reinforcement Learning for Multiple Constraints

Dohyeong Kim (Seoul National University), Songhwai Oh (Seoul National University)

OptimizationSafty and PrivacyRobotic IntelligenceReinforcement Learning

🎯 What it does: A trust-region based safe distributed actor-critic algorithm (SDAC) is proposed, capable of training robots in multi-constraint and risk-averse reinforcement learning environments.

Trust Your $\nabla$: Gradient-based Intervention Targeting for Causal Discovery

Mateusz Olko (Warsaw University), Piotr Miłoś (Institute of Mathematics, Polish Academy of Sciences)

OptimizationGraph Neural NetworkGraph

🎯 What it does: A gradient-based experimental design method GIT is proposed to select the most informative single-node intervention targets in causal discovery, enabling rapid recovery of causal structures under low sample sizes.

Tuning Multi-mode Token-level Prompt Alignment across Modalities

Dongsheng Wang (Xidian University), Hanwang Zhang (Nanyang Technological University)

ClassificationDomain AdaptationTransformerPrompt EngineeringVision Language ModelImageMultimodality

🎯 What it does: This paper proposes a method for simultaneously learning multi-modal visual and textual prompts in a multi-modal vision-language model, achieving token-level alignment through hierarchical optimal transport (OT) to enhance the performance of few-shot image classification.

Two Heads are Better Than One: A Simple Exploration Framework for Efficient Multi-Agent Reinforcement Learning

Jiahui Li (Zhejiang University), Long Chen (HKUST)

Reinforcement Learning

🎯 What it does: A COIN (Curiosity + Influence) exploration framework is proposed, which combines mutual information-driven influence rewards with prediction error-driven curiosity rewards in multi-agent reinforcement learning to enhance exploration efficiency.

Two Sides of One Coin: the Limits of Untuned SGD and the Power of Adaptive Methods

Junchi YANG, Niao He (ETH Zurich)

OptimizationConvolutional Neural NetworkImage

🎯 What it does: This study investigates the convergence properties of unadjusted SGD and various adaptive methods in the presence of unknown smooth parameters, revealing the exponential risk of gradient explosion in unadjusted SGD and proving that adaptive methods can eliminate this exponential factor.

Two Sides of The Same Coin: Bridging Deep Equilibrium Models and Neural ODEs via Homotopy Continuation

Shutong Ding (ShanghaiTech University), Ye Shi (ShanghaiTech University)

ClassificationExplainability and InterpretabilityComputational EfficiencyImageOrdinary Differential Equation

🎯 What it does: The HomoODE model is proposed, unifying Deep Equilibrium Models (DEQs) and Neural ODEs through the same theory (homotopy continuation method), and based on this, new implicit networks are constructed; at the same time, a strategy of sharing learnable initial points is proposed to accelerate the solving process, which explains the effects of Augmented Neural ODE.

Two-Stage Learning to Defer with Multiple Experts

Anqi Mao (Courant Institute), Yutao Zhong (Courant Institute)

ClassificationOptimizationConvolutional Neural NetworkMixture of ExpertsImage

🎯 What it does: This paper studies a two-stage learning to defer framework with multiple experts: first, a predictor is trained using standard multi-class loss, and then a new surrogate loss is used to learn the deferral function, which decides how to allocate inputs to different experts.

Two-Stage Predict+Optimize for MILPs with Unknown Parameters in Constraints

Xinyi HU, Jimmy H.M. Lee (University of Wisconsin-Madison)

OptimizationTabular

🎯 What it does: A Two-Stage Predict+Optimize framework is proposed to address mixed-integer linear programming (MILP) problems with unknown parameters in constraints, along with a general end-to-end gradient training algorithm.

Type-to-Track: Retrieve Any Object via Prompt-based Tracking

Pha Nguyen (University of Arkansas), Khoa Luu (University of Arkansas)

Object DetectionObject TrackingTransformerVideoText

🎯 What it does: This paper proposes a multi-object tracking paradigm based on natural language descriptions (Type-to-Track) and constructs a large-scale Grounded MOT dataset called GroOT, while designing an efficient single-stage Transformer tracking model named MENDER.

UE4-NeRF:Neural Radiance Field for Real-Time Rendering of Large-Scale Scene

Jiaming Gu (Xidian University), Mohammed Bennamoun (Western Australia)

GenerationOptimizationComputational EfficiencyNeural Radiance FieldPoint CloudMeshAgriculture Related

🎯 What it does: A real-time interactive large-scale scene rendering system based on UE4, called UE4-NeRF, has been implemented.

UltraRE: Enhancing RecEraser for Recommendation Unlearning via Error Decomposition

Yuyuan Li (Zhejiang University), Jun Wang (OPPO Research Institute)

Recommendation SystemTabular

🎯 What it does: This paper proposes a new recommendation system machine learning-free framework called UltraRE, aimed at improving model utility and learning-free efficiency while ensuring complete learning-free integrity.

Unbalanced Low-rank Optimal Transport Solvers

Meyer Scetbon (Microsoft Research), marco cuturi

OptimizationBiomedical Data

🎯 What it does: A low-rank unbalanced optimal transport (Unbalanced Low-Rank OT) solver is proposed and implemented, balancing low-rank acceleration with unbalanced regularization.

Unbiased Compression Saves Communication in Distributed Optimization: When and How Much?

Yutong He (Peking University), Kun Yuan (Peking University)

OptimizationTabular

🎯 What it does: The study establishes the lower and upper bounds of total communication costs when using unbiased and mutually independent compressors in distributed optimization, proving that independent unbiased compression can reduce the total communication cost to Θ(√min{n,κ}) times, and presents the approximately optimal ADIANA algorithm.

Unbiased constrained sampling with Self-Concordant Barrier Hamiltonian Monte Carlo

Maxence Noble (Institut Polytechnique de Paris), Alain Durmus

OptimizationTabularStochastic Differential EquationOrdinary Differential Equation

🎯 What it does: This paper proposes the Barrier Hamiltonian Monte Carlo (BHMC) algorithm, specifically designed for sampling in Riemannian subspaces constructed by self-diagonal barrier functions; at the numerical implementation level, n-BHMC is introduced, incorporating an 'involution checking step' to eliminate the bias issues caused by implicit integration in traditional RMHMC.

Unbiased learning of deep generative models with structured discrete representations

Harry Bendekgey (University of California), Erik B. Sudderth (University of California)

GenerationOptimizationAuto EncoderMultimodalityTime SeriesAudio

🎯 What it does: A structured variational autoencoder (SVAE) capable of unbiased learning is proposed, achieving a fusion of deep generative models and graphical models, and addressing optimization challenges during training.

Unbounded Differentially Private Quantile and Maximum Estimation

David Durfee (Anonym Inc)

Safty and PrivacyTabular

🎯 What it does: A differential privacy quantile and maximum value estimation algorithm (UQE) for unbounded data has been developed, which implements guess-and-check for quantiles through multiple calls to AboveThreshold and is used for private summation and mean calculation.

Uncertainty Estimation for Safety-critical Scene Segmentation via Fine-grained Reward Maximization

Hongzheng Yang (Chinese University of Hong Kong), Qi Dou (Chinese University of Hong Kong)

SegmentationReinforcement LearningBiomedical DataMagnetic Resonance Imaging

🎯 What it does: A framework for fine-grained reward maximization (FGRM) is proposed to fine-tune a pre-trained evidence learning segmentation model using reinforcement learning, aiming for reliable uncertainty estimation in safety-critical scenarios.

Uncertainty Quantification over Graph with Conformalized Graph Neural Networks

Kexin Huang (Stanford University), Jure Leskovec (Stanford University)

ClassificationOptimizationGraph Neural NetworkSupervised Fine-TuningGraph

🎯 What it does: A framework based on synthetic prediction, CF-GNN, is proposed in graph neural networks for uncertainty quantification in node classification and regression tasks, providing theoretical guarantees for coverage.

Uncertainty Quantification via Neural Posterior Principal Components

Elias Nehme (Technion - Israel Institute of Technology), Tomer Michaeli (Technion - Israel Institute of Technology)

Image TranslationRestorationSuper ResolutionImageBiomedical Data

🎯 What it does: A neural network method for predicting the posterior distribution of principal components in image restoration problems (NPPC) is proposed, which provides instance-adaptive uncertainty directions for image inversion tasks (denoising, interpolation, super-resolution, coloring, cell image translation);

Uncertainty-Aware Alignment Network for Cross-Domain Video-Text Retrieval

Xiaoshuai Hao (Samsung Research China), Wanqian Zhang (Institute of Information Engineering)

RetrievalDomain AdaptationVideoText

🎯 What it does: In the unsupervised cross-domain video-text retrieval task, an Uncertainty-aware Alignment Network (UAN) is proposed, which significantly improves retrieval performance through self-supervised matching and distribution alignment to achieve multi-to-one relationship learning in the target domain.

Uncertainty-Aware Instance Reweighting for Off-Policy Learning

Xiaoying Zhang (ByteDance Research), Hang Li (ByteDance Research)

Recommendation SystemReinforcement LearningTabular

🎯 What it does: This paper proposes an Uncertainty-Aware Inverse Propensity Score (UIPS) estimator to improve policy evaluation and optimization in offline policy learning with unknown logging policies.

Unconstrained Dynamic Regret via Sparse Coding

Zhiyu Zhang (Harvard University), Ioannis Paschalidis

OptimizationTime Series

🎯 What it does: This paper proposes a general sparse coding framework to obtain an adaptive dynamic regret upper bound in unconstrained online convex optimization (OCO); it also provides a specific algorithm and a better dynamic regret lower bound based on the Haar wavelet dictionary.

Uncoupled and Convergent Learning in Two-Player Zero-Sum Markov Games with Bandit Feedback

Yang Cai (Yale University), Weiqiang Zheng (Yale University)

Reinforcement Learning

🎯 What it does: A fully decoupled, convergent, and rational non-asymptotic learning algorithm is proposed for two-dimensional zero-sum Markov games, providing optimal convergence rates for matrix games, irreducible Markov games, and general Markov games.

Uncovering and Quantifying Social Biases in Code Generation

Yan Liu (Microsoft Research), Tsung-Yi Ho (Peking University)

GenerationAI Code AssistantTransformerLarge Language ModelPrompt EngineeringText

🎯 What it does: This paper studies the issue of social bias in pre-trained code generation models, proposing a new code prompt template to activate model biases and quantifying the generated code for evaluation.

Uncovering Meanings of Embeddings via Partial Orthogonality

Yibo Jiang (University of Chicago), Victor Veitch (University of Chicago)

Representation LearningText

🎯 What it does: By introducing the theory of partial orthogonality, this paper explores the semantic independence in text embedding vectors and defines the generalized Markov boundary. Subsequently, methods such as random subspace projection are used to validate its semantic implications on CLIP embeddings.

Uncovering motifs of concurrent signaling across multiple neuronal populations

Evren Gokcen (Carnegie Mellon University), Byron M. Yu (Carnegie Mellon University)

Time Series

🎯 What it does: A multi-population delayed latent variable model (mDLAG) is proposed to simultaneously characterize the common and local dynamics of multiple neural populations and their delayed signal flow.

Uncovering Prototypical Knowledge for Weakly Open-Vocabulary Semantic Segmentation

Fei Zhang (Shanghai Jiao Tong University), Yanfeng Wang (Shanghai Jiao Tong University)

SegmentationTransformerContrastive LearningImageTextMultimodality

🎯 What it does: This paper proposes a prototype knowledge-based supervised method that improves the grouped token mechanism in ViT to achieve weakly supervised vocabulary semantic segmentation.

Uncovering the Hidden Dynamics of Video Self-supervised Learning under Distribution Shifts

Pritam Sarkar (Queen's University), Ali Etemad (Google Research)

Domain AdaptationRepresentation LearningTransformerContrastive LearningVideoBenchmark

🎯 What it does: Conduct systematic experiments on six mainstream video self-supervised learning methods (Contrastive, Non-Contrastive, Generative) under various real distribution shifts (background, viewpoint, actor, data source, zero-shot, open set), and construct 17 InD–OoD dataset pairs to comprehensively evaluate their robustness and dynamic behavior.

Understanding and Addressing the Pitfalls of Bisimulation-based Representations in Offline Reinforcement Learning

Hongyu Zang (Beijing Institute of Technology), Romain Laroche (Wayve)

Reinforcement LearningTabular

🎯 What it does: Research and improve the bisimulation-based state representation method in offline reinforcement learning to address the issues of estimation failure and feature collapse caused by missing offline data.

Understanding and Improving Ensemble Adversarial Defense

Yian Deng (University of Manchester), Tingting Mu (University of Manchester)

Adversarial AttackGenerative Adversarial NetworkImage

🎯 What it does: Theoretical proof is presented that ensemble adversarial defense outperforms single models, and the iGAT method is proposed to enhance ensemble robustness through global adversarial sample allocation and regularization.

Understanding and Improving Feature Learning for Out-of-Distribution Generalization

Yongqiang Chen (Chinese University of Hong Kong), James Cheng (Chinese University of Hong Kong)

Domain AdaptationRepresentation LearningConvolutional Neural NetworkImage

🎯 What it does: This paper explores the behavior of ERM and OOD objectives in feature learning through theoretical analysis and experiments, and proposes the Feature Augmented Training (FeAT) iterative method to obtain richer features and enhance OOD generalization performance.

Understanding and Mitigating Copying in Diffusion Models

Gowthami Somepalli (University of Maryland), Tom Goldstein (University of Maryland)

GenerationDiffusion modelImageText

🎯 What it does: Analyzes the replication behavior of diffusion models under text conditions and proposes various de-duplication strategies during training and inference.

Understanding Contrastive Learning via Distributionally Robust Optimization

Junkang Wu (University of Science and Technology of China), Xiangnan He (University of Science and Technology of China)

OptimizationRepresentation LearningContrastive LearningImageText

🎯 What it does: By viewing contrastive learning as distributionally robust optimization, this paper provides a theoretical analysis that explains the tolerance of contrastive learning to negative sample sampling bias and proposes a new weighted InfoNCE loss—ADNCE—to alleviate issues of excessive conservativeness and sensitivity to outliers.

Understanding Deep Gradient Leakage via Inversion Influence Functions

Haobo Zhang (Michigan State University), Jiayu Zhou (Michigan State University)

Safty and PrivacyAdversarial AttackConvolutional Neural NetworkTransformerImageText

🎯 What it does: The problem of Deep Gradient Leakage (DGL) is analyzed by proposing and validating the Inverse Influence Function (I2F), providing both empirical and theoretical results.

Understanding Diffusion Objectives as the ELBO with Simple Data Augmentation

Diederik P Kingma, Ruiqi Gao (Google DeepMind)

GenerationData SynthesisDiffusion modelImage

🎯 What it does: Clarified that the training objective of the diffusion model is essentially a weighted integral of the ELBO at different noise levels, and proved that when the weight function is monotonic, it can be equivalent to the ELBO of data augmentation with additive Gaussian noise.

Understanding Few-Shot Learning: Measuring Task Relatedness and Adaptation Difficulty via Attributes

Minyang Hu (Institute of Computing Technology, Chinese Academy of Sciences), Xilin CHEN

Meta LearningImage

🎯 What it does: Proposed and validated the Task Attribute Distance (TAD) metric to measure the correlation between training tasks and new tasks, as well as the adaptation difficulty of new tasks;

Understanding How Consistency Works in Federated Learning via Stage-wise Relaxed Initialization

Yan Sun (University of Sydney), Dacheng Tao (University of Sydney)

Federated LearningImage

🎯 What it does: A new federated learning method called FedInit is proposed, which employs a phased personalized relaxation initialization to enhance local consistency.

Understanding Multi-phase Optimization Dynamics and Rich Nonlinear Behaviors of ReLU Networks

Mingze Wang (Peking University), Chao Ma (Stanford University)

Optimization

🎯 What it does: This paper provides a complete theoretical analysis of the optimization dynamics of a two-layer ReLU network under gradient flow (GF) training when facing linearly separable data with small angles.

Understanding Neural Network Binarization with Forward and Backward Proximal Quantizers

Yiwei Lu (University of Waterloo), Vahid Partovi Nia (Huawei)

OptimizationTransformerImage

🎯 What it does: ProxConnect++ (PC++) framework is proposed, unifying and extending the forward-backward quantizers of binary networks;

Understanding the detrimental class-level effects of data augmentation

Polina Kirichenko (New York University), Andrew Gordon Wilson (Meta AI)

ClassificationData-Centric LearningConvolutional Neural NetworkTransformerImage

🎯 What it does: This study investigates and quantifies the negative impact of data augmentation on the accuracy of each category in ImageNet, corrects label noise through multi-label annotations, summarizes the types of categories that lead to confusion, and proposes a simple data augmentation strategy based on class-level confusion information, significantly improving the accuracy of affected categories while maintaining overall accuracy relatively unchanged.

Understanding the Latent Space of Diffusion Models through the Lens of Riemannian Geometry

Yong-Hyun Park (Seoul National University), Youngjung Uh (Seoul National University)

GenerationDiffusion modelImageText

🎯 What it does: This paper analyzes the latent space of diffusion models through pullback metrics, extracts local latent bases, and utilizes them to achieve image editing in the latent space at a single moment. It further studies the evolution of the latent structure with diffusion steps and the impact of text prompts.

Understanding the Limitations of Deep Models for Molecular property prediction: Insights and Solutions

Jun Xia (Westlake University), Stan Z. Li (Westlake University)

Drug DiscoveryConvolutional Neural NetworkRecurrent Neural NetworkGraph Neural NetworkTransformerGraphTabularBiomedical DataBenchmark

🎯 What it does: A systematic benchmark evaluation of 12 representative models on 15 datasets for molecular property prediction tasks is conducted, and an Independent Feature Mapping (IFM) method is proposed to enhance the performance of deep models.

Understanding, Predicting and Better Resolving Q-Value Divergence in Offline-RL

Yang Yue (Tsinghua University), Gao Huang (Tsinghua University)

Reinforcement LearningTabular

🎯 What it does: This paper studies the divergence problem of Q-value estimation in offline reinforcement learning, proposes a Self-Excitation mechanism to explain the root cause of divergence, and introduces the SEEM (Self-Excitation Eigenvalue) metric to predict divergence.

Undirected Probabilistic Model for Tensor Decomposition

Zerui Tao (Tokyo University of Agriculture and Technology), Qibin Zhao (RIKEN AIP)

Contrastive LearningMultimodalityTime Series

🎯 What it does: A framework for undirected tensor decomposition is constructed through deep energy-based models (EBM) to jointly learn tensor observations and latent factors, thereby achieving probabilistic modeling of non-Gaussian, multimodal data.

Unexpected Improvements to Expected Improvement for Bayesian Optimization

Sebastian Ament (Meta), Eytan Bakshy (Meta)

OptimizationTabular

🎯 What it does: A LogEI family is proposed in Bayesian optimization, which numerically reconstructs the traditional Expected Improvement and its extensions (CEI, qEI, qEHVI) to avoid the gradient vanishing problem in numerical calculations, making it easier for gradient optimization.

Uni-ControlNet: All-in-One Control to Text-to-Image Diffusion Models

Shihao Zhao (University of Hong Kong), Kwan-Yee K. Wong (University of Hong Kong)

GenerationData SynthesisDiffusion modelImageText

🎯 What it does: This paper proposes Uni-ControlNet, a unified framework that can simultaneously utilize various local (such as edges, depth, segmentation, etc.) and global (such as CLIP image embeddings) control signals within a single model, enabling composable text-to-image diffusion model control.

Uni3DETR: Unified 3D Detection Transformer

Zhenyu Wang (Tsinghua University), Shengjin Wang (Tsinghua University)

Object DetectionAutonomous DrivingTransformerPoint Cloud

🎯 What it does: This paper proposes Uni3DETR, a unified 3D object detection framework capable of handling both indoor and outdoor point clouds.

UniControl: A Unified Diffusion Model for Controllable Visual Generation In the Wild

Can Qin (Northeastern University), Ran Xu (Salesforce AI Research)

SegmentationGenerationData SynthesisMixture of ExpertsDiffusion modelImageText

🎯 What it does: We propose and train UniControl, a unified diffusion model capable of handling multiple visual conditions (edges, segmentation, depth, skeletons, etc.) and text prompts to achieve controllable image generation.

Unified 3D Segmenter As Prototypical Classifiers

Zheyun Qin (Shandong University), Xiankai Lu (Shandong University)

SegmentationTransformerPoint Cloud

🎯 What it does: A prototype-based unified framework called PROTOSEG is proposed, which unifies semantic, instance, and panoptic segmentation tasks into classification problems of different granularities, using Transformers to extract point embeddings and achieve classification through dynamic prototype association and updating.

Unified Embedding: Battle-Tested Feature Representations for Web-Scale ML Systems

Benjamin Coleman (Google DeepMind), Derek Zhiyuan Cheng (Google DeepMind)

Recommendation SystemOptimizationTabularBenchmark

🎯 What it does: Proposes a Feature Multiplexing framework that allows multiple classification features to share the same embedding space, and based on this, designs a Unified Embedding that significantly reduces model parameters and latency.

Unified Enhancement of Privacy Bounds for Mixture Mechanisms via $f$-Differential Privacy

Chendi Wang (University of Pennsylvania), Weijie J Su

Safty and Privacy

🎯 What it does: A unified f-DP analysis framework is proposed to accurately evaluate the privacy boundaries of mixed mechanisms (such as shuffling models and single-step DP-GD with random initialization) and provides a closed-form trade-off function.

Unified Lower Bounds for Interactive High-dimensional Estimation under Information Constraints

Jayadev Acharya (Cornell University), Himanshu Tyagi (Indian Institute of Science)

🎯 What it does: This paper considers distributed parameter estimation under information constraints and proposes a unified framework that can derive tight minimax lower bounds for different parameter distribution families, applicable to both continuous and discrete distributions, and effective under any glyph[lscript] p loss.

Unified Off-Policy Learning to Rank: a Reinforcement Learning Perspective

Zeyu Zhang (University of Science and Technology of China), Mengdi Wang (Princeton University)

Recommendation SystemOptimizationTransformerReinforcement LearningTabular

🎯 What it does: Unified the problem of Off-Policy Learning to Rank as a Markov Decision Process (MDP) and directly learned the optimal ranking policy through offline reinforcement learning (RL).

Unified Segment-to-Segment Framework for Simultaneous Sequence Generation

Shaolei Zhang (Chinese Academy of Sciences), Yang Feng (Chinese Academy of Sciences)

RecognitionGenerationTransformerTextAudio

🎯 What it does: A unified paragraph-to-paragraph framework (Seg2Seg) is proposed, which introduces latent paragraphs as a bridge to learn source-target adaptive mapping in real-time sequence generation (such as streaming ASR, synchronous MT, and synchronous ST) and achieves multi-task learning.

Uniform Convergence with Square-Root Lipschitz Loss

Lijia Zhou (University of Chicago), Nathan Srebro (Toyota Technological Institute at Chicago)

Tabular

🎯 What it does: This paper presents a uniform convergence guarantee suitable for 'square root-Lipschitz' loss functions, extending previous theories based on smoothness (Lipschitz of derivatives);

Uniform-in-Time Wasserstein Stability Bounds for (Noisy) Stochastic Gradient Descent

Lingjiong Zhu (Florida State University), Umut Simsekli (Inria Paris)

OptimizationStochastic Differential Equation

🎯 What it does: This paper studies the Wasserstein stability of stochastic gradient descent (SGD) and its noisy versions under different convexity assumptions, providing a time-uniform stability upper bound.

Unifying GANs and Score-Based Diffusion as Generative Particle Models

Jean-Yves Franceschi (Criteo AI Lab), Alain Rakotomamonjy (Criteo AI Lab)

GenerationData SynthesisDiffusion modelScore-based ModelFlow-based ModelGenerative Adversarial NetworkImageStochastic Differential EquationOrdinary Differential Equation

🎯 What it does: A unified particle model framework that integrates GANs and score-based diffusion models is proposed, under which two novel generative models—Score GAN and Discriminator Flow—are designed.

Unifying Predictions of Deterministic and Stochastic Physics in Mesh-reduced Space with Sequential Flow Generative Model

Luning Sun (Lawrence Livermore National Laboratory), Liping Liu

Graph Neural NetworkTransformerFlow-based ModelAuto EncoderMeshTime SeriesPhysics Related

🎯 What it does: A unified deep learning framework is proposed, capable of simultaneously predicting the time evolution of deterministic and stochastic physical systems on unstructured grids. This framework compresses spatial states into low-dimensional latent variables using a graph autoencoder, then generates conditional vectors with a transformer, and subsequently propagates to generate subsequent states in the latent space through Conditional RealNVP, ultimately decoding back to the physical grid.

UniPC: A Unified Predictor-Corrector Framework for Fast Sampling of Diffusion Models

Wenliang Zhao (Tsinghua University), Jiwen Lu (Tsinghua University)

GenerationComputational EfficiencyDiffusion modelImageOrdinary Differential Equation

🎯 What it does: A unified prediction-correction framework called UniPC is proposed to accelerate the sampling of diffusion probabilistic models.

UniT: A Unified Look at Certified Robust Training against Text Adversarial Perturbation

Muchao Ye (Pennsylvania State University), Fenglong Ma (Pennsylvania State University)

ClassificationAdversarial AttackTransformerLarge Language ModelText

🎯 What it does: A unified text attack robust training framework called UniT is proposed, and a Decomposed Regularization (DR) loss is designed to enhance the robustness of the base model.

UniTSFace: Unified Threshold Integrated Sample-to-Sample Loss for Face Recognition

Qiufu Li (Shenzhen University), Jinming Duan (University of Birmingham)

RecognitionConvolutional Neural NetworkContrastive LearningImage

🎯 What it does: A unified threshold integrated sample-to-sample loss (USS loss) is proposed, which learns a unified threshold to distinguish between positive and negative face pairs;

Universal Gradient Descent Ascent Method for Nonconvex-Nonconcave Minimax Optimization

Taoli Zheng (Chinese University of Hong Kong), Jiajin Li (Stanford University)

Optimization

🎯 What it does: This paper proposes a Dual Smoothing Gradient Descent Ascent (DS-GDA) algorithm to solve minimax optimization problems that are non-convex-concave, non-convex-convex, and convex-non-concave, and implements a single-loop structure.

Universal Online Learning with Gradient Variations: A Multi-layer Online Ensemble Approach

Yu-Hu Yan (Nanjing University), Zhi-Hua Zhou (Nanjing University)

OptimizationReinforcement Learning

🎯 What it does: A multi-layer online aggregation framework is proposed, achieving a general convergence guarantee for gradient variations in online convex optimization with unknown curvature.

Universal Prompt Tuning for Graph Neural Networks

Taoran Fang (Zhejiang University), Lei CHEN

Graph Neural NetworkPrompt EngineeringGraphBiomedical Data

🎯 What it does: A general prompt tuning method for pre-trained graph neural networks (Graph Prompt Feature, GPF and its variant GPF-plus) is proposed, which adapts to downstream tasks by adding learnable prompt vectors in the graph node feature space without modifying the model itself.

Universality and Limitations of Prompt Tuning

Yihan Wang (University of California Los Angeles), Cho-Jui Hsieh (Google)

TransformerPrompt EngineeringTextSequential

🎯 What it does: This paper studies the theoretical properties of prompt tuning in Transformers, proving that it is a universal approximator under specifically constructed Transformers, and reveals its expressive limitations in both single-layer and multi-layer networks.

Universality laws for Gaussian mixtures in generalized linear models

Yatin Dandi (EPFL), Lenka Zdeborova (EPFL)

OptimizationGenerative Adversarial NetworkImage

🎯 What it does: In the high-dimensional limit, it is proven that mixed Gaussian distribution data and equivalent Gaussian mixture models share the same statistical properties under generalized linear models, ERM, Gibbs sampling, and ensemble tasks, including training error, generalization error, and the geometric characteristics of the estimator.

Unleash the Potential of Image Branch for Cross-modal 3D Object Detection

Yifan Zhang (City University of Hong Kong), Guoliang Xing (Chinese University of Hong Kong)

Object DetectionAutonomous DrivingConvolutional Neural NetworkImagePoint Cloud

🎯 What it does: A cross-modal 3D object detection framework UPIDet is proposed, which utilizes image information to enhance point cloud detection performance.

Unleashing the Full Potential of Product Quantization for Large-Scale Image Retrieval

Yu Liang (Hunan University), Xiaoyu Wang (Hong Kong University of Science and Technology)

RetrievalConvolutional Neural NetworkImage

🎯 What it does: A product quantization framework FPPQ based on deep learning is proposed, which utilizes a softmax differentiable PQ branch to learn category-level PQ codes, aiming to enhance the performance of large-scale image retrieval.

Unleashing the Power of Graph Data Augmentation on Covariate Distribution Shift

Yongduo Sui (University of Science and Technology of China), Xiangnan He (Ant Group)

ClassificationDomain AdaptationGraph Neural NetworkGenerative Adversarial NetworkGraph

🎯 What it does: A framework named Adversarial Invariant Augmentation (AIA) is proposed for graph data augmentation to address the covariate shift problem in graph classification tasks; it enhances environmental features differentially while keeping stable features unchanged, improving the model's generalization performance in unseen environments.

Unleashing the Power of Randomization in Auditing Differentially Private ML

Krishna Pillutla (Google Research), Sewoong Oh (University of Washington)

Safty and PrivacyGaussian SplattingTabular

🎯 What it does: Audit differential privacy machine learning algorithms and propose adding multiple random 'canaries' to the dataset for multiple statistical tests.

Unlimiformer: Long-Range Transformers with Unlimited Length Input

Amanda Bertsch (Carnegie Mellon University), Matthew R. Gormley (Carnegie Mellon University)

TransformerTextRetrieval-Augmented Generation

🎯 What it does: The existing pre-trained encoder-decoder Transformer is modified to use k-NN retrieval to focus only on the top k most relevant keys in the cross-attention, enabling the processing of inputs of infinite length.

Unlocking Deterministic Robustness Certification on ImageNet

Kai Hu (Carnegie Mellon University), Matt Fredrikson (Carnegie Mellon University)

ClassificationOptimizationAdversarial AttackConvolutional Neural NetworkDiffusion modelImage

🎯 What it does: Design and train the LiResNet architecture and EMMA loss to achieve provably robust deep networks.

Unlocking Feature Visualization for Deep Network with MAgnitude Constrained Optimization

Thomas FEL, Thomas Serre (Brown University)

GenerationOptimizationExplainability and InterpretabilityConvolutional Neural NetworkTransformerImage

🎯 What it does: A new feature visualization method called MACO is proposed, which utilizes amplitude constraints of the Fourier spectrum and phase optimization to generate natural image explanations.