NeurIPS 2023 Papers — Page 30
Conference on Neural Information Processing Systems · 3218 papers
This Looks Like Those: Illuminating Prototypical Concepts Using Multiple Visualizations
Chiyu Ma (Dartmouth), Cynthia Rudin (Duke)
ClassificationExplainability and InterpretabilityConvolutional Neural NetworkImage
🎯 What it does: This paper proposes the ProtoConcepts method, which transforms the single image prototype in traditional prototype networks into a multi-image visual concept, using spherical prototypes to capture similar features of multiple training samples, thereby achieving a more intuitive 'this class versus those' explanation.
Thought Cloning: Learning to Think while Acting by Imitating Human Thinking
Shengran Hu (University of British Columbia), Jeff Clune (University of British Columbia)
Robotic IntelligenceRecurrent Neural NetworkTransformerReinforcement LearningAgentic AITextSequential
🎯 What it does: A framework for imitation learning called Thought Cloning is proposed, allowing agents to mimic human thought processes while executing actions.
Three Iterations of (d − 1)-WL Test Distinguish Non Isometric Clouds of d-dimensional Points
Valentino delle Rose, Pablo Barcelo
Representation LearningGraph Neural NetworkPoint CloudPhysics Related
🎯 What it does: The paper provides a theoretical analysis of the geometric Weisfeiler–Leman (WL) test for identifying the isomorphism of point clouds in d-dimensional Euclidean space, proving that (d‑1)-WL can completely distinguish any point cloud after 3 iterations, while d-WL can achieve this in just one iteration.
Three Towers: Flexible Contrastive Learning with Pretrained Image Models
Jannik Kossen (University of Oxford), Effrosyni Kokiopoulou
ClassificationRetrievalTransformerContrastive LearningImageText
🎯 What it does: Proposes the Three Towers method, which integrates pre-trained image models and self-supervised learning in contrastive learning.
Three-Way Trade-Off in Multi-Objective Learning: Optimization, Generalization and Conflict-Avoidance
Lisha Chen (Rensselaer Polytechnic Institute), Tianyi Chen (Rensselaer Polytechnic Institute)
OptimizationTabular
🎯 What it does: This study investigates the three-way trade-off in multi-objective learning (optimization, generalization, conflict avoidance), proposes the MoDo dual-sampling algorithm, and provides theoretical analysis and experimental validation.
Thrust: Adaptively Propels Large Language Models with External Knowledge
Xinran Zhao (Carnegie Mellon University), Jianshu Chen (Tencent)
ClassificationRetrievalTransformerLarge Language ModelTextRetrieval-Augmented Generation
🎯 What it does: This paper studies how to adaptively retrieve external knowledge in large-scale pre-trained language models (PTLMs), only performing retrieval when the internal knowledge of the model is insufficient, and proposes an instance-level metric called Thrust to evaluate the adequacy of knowledge for a specific instance.
TIES-Merging: Resolving Interference When Merging Models
Prateek Yadav (University of North Carolina at Chapel Hill), Mohit Bansal (University of North Carolina at Chapel Hill)
TransformerSupervised Fine-TuningText
🎯 What it does: A new model merging method called TIES-Merging is proposed, which can merge multiple task-specific fine-tuned models into a multi-task model without additional training, significantly reducing performance loss caused by parameter interference.
Tight Bounds for Volumetric Spanners and Applications
Aditya Bhaskara (University of Utah), Ali Vakilian (Toyota Technological Institute at Chicago)
Optimization
🎯 What it does: This paper proposes and proves that a volumetric spanner and well-conditioned spanning subsets can be constructed using a local search algorithm with a single swap. It provides near-optimal size upper bounds for ℓ₂ and ℓ_p (p∈[1,∞)) and utilizes this result to construct a coreset of size O(d/ε) for the Minimum Volume Enclosing Ellipsoid (MVEE) problem.
Tight Risk Bounds for Gradient Descent on Separable Data
Matan Schliserman (Tel Aviv University), Tomer Koren (Google Research)
Optimization
🎯 What it does: This study investigates the generalization risk of unregularized gradient descent (GD) in separable linear classification problems, providing tight upper and lower bounds for any smooth convex loss function, and clarifying the relationship between risk and factors such as the number of gradient steps, sample size, data spacing, and the tail decay rate of the loss.
Time Series as Images: Vision Transformer for Irregularly Sampled Time Series
Zekun Li (University of California, Santa Barbara), Xifeng Yan (University of California, Santa Barbara)
ClassificationTransformerTime SeriesBiomedical Data
🎯 What it does: Convert irregularly sampled multivariate time series into line graph images and classify them using a pre-trained Vision Transformer.
Time Series Kernels based on Nonlinear Vector AutoRegressive Delay Embeddings
Giovanni De Felice (University of Liverpool), Vladimir Gusev
ClassificationComputational EfficiencyTime Series
🎯 What it does: This paper proposes a time series kernel based on Nonlinear Vector Autoregression (NVAR) delay embedding, utilizing random subsample embedding and ridge regression reading to extract system dynamics features and construct similarity.
Time-Independent Information-Theoretic Generalization Bounds for SGLD
Futoshi Futami (Osaka University / RIKEN AIP), Masahiro Fujisawa (RIKEN AIP)
OptimizationStochastic Differential Equation
🎯 What it does: This paper studies the generalization error of noise-free Stochastic Gradient Langevin Dynamics (SGLD), providing a new information-theoretic generalization upper bound that is time-independent and decreases with sample size. It also presents for the first time the sub-exponential property and corresponding upper bound when evaluating generalization using the same training loss, further deriving an improved excess risk bound.
Time-Reversed Dissipation Induces Duality Between Minimizing Gradient Norm and Function Value
Jaeyeon Kim (Seoul National University), Ernest K. Ryu (Seoul National University)
Optimization
🎯 What it does: The H-duality theory is proposed, proving a one-to-one correspondence between optimization methods that minimize function values and those that minimize gradient norms. This duality is used to derive new gradient norm degradation algorithms (such as Super FISTA-G) and their continuous-time versions.
Time-uniform confidence bands for the CDF under nonstationarity
Paul Mineiro (Microsoft Research), Steven R Howard
Time Series
🎯 What it does: This paper proposes a time-uniform and value-uniform confidence band for the cumulative distribution function (CDF) of non-stationary, data-dependent univariate sequences, and extends this method to importance-weighted counterfactual distribution estimation.
Timewarp: Transferable Acceleration of Molecular Dynamics by Learning Time-Coarsened Dynamics
Leon Klein (Freie Universität Berlin), Ryota Tomioka (Microsoft Research AI4Science)
Computational EfficiencyDrug DiscoveryTransformerFlow-based ModelTime SeriesSequential
🎯 What it does: A transferable accelerated molecular dynamics (MD) sampling method named Timewarp is proposed, which generates large-scale structural changes by learning time-coarse-grained dynamics, and then uses Metropolis-Hastings correction within a Markov Chain Monte Carlo (MCMC) framework to sample the thermal equilibrium distribution in an unbiased manner.
TMT-VIS: Taxonomy-aware Multi-dataset Joint Training for Video Instance Segmentation
rongkun Zheng, Hengshuang Zhao (University of Hong Kong)
Object DetectionSegmentationTransformerVideo
🎯 What it does: This paper proposes a model named TMT-VIS, which improves the video instance segmentation task of joint training on multiple datasets by utilizing taxonomy (category hierarchy) information.
To Repeat or Not To Repeat: Insights from Scaling LLM under Token-Crisis
Fuzhao Xue (National University of Singapore), Yang You (National University of Singapore)
TransformerLarge Language ModelMixture of ExpertsText
🎯 What it does: This paper conducts a systematic empirical study on the impact of reusing pre-trained data on LLM training under the token crisis, exploring the causes of overfitting, key factors, regularization methods, and MoE tuning strategies.
To Stay or Not to Stay in the Pre-train Basin: Insights on Ensembling in Transfer Learning
Ildus Sadrtdinov (HSE University), Ekaterina Lobacheva (Independent researcher)
ClassificationDomain AdaptationConvolutional Neural NetworkTransformerSupervised Fine-TuningContrastive LearningImage
🎯 What it does: This paper studies the performance improvement of a model ensemble using only a single pre-trained checkpoint in the context of transfer learning, and proposes a parallel version of Snapshot Ensembling (StarSSE) to better explore model diversity within the pre-trained basin.
TOA: Task-oriented Active VQA
Xiaoying Xing (Northwestern University), Ying Wu (Northwestern University)
TransformerLarge Language ModelPrompt EngineeringVision Language ModelImageText
🎯 What it does: A task-oriented active visual question answering (TOA) framework is proposed, which utilizes large language models to actively generate hypotheses and obtain evidence verification through visual functions, ultimately answering questions that require external knowledge.
Token-Scaled Logit Distillation for Ternary Weight Generative Language Models
Minsoo Kim (Hanyang University), Jungwook Choi (Hanyang University)
GenerationKnowledge DistillationTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: For generative language models (GLMs), a method for quantization-aware training (QAT) under low precision (especially ternary quantization) is proposed, named Token-Scaled Logit Distillation (TSLD).
Toolformer: Language Models Can Teach Themselves to Use Tools
Timo Schick (Meta), Thomas Scialom
TransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
🎯 What it does: Using large-scale language models to learn to call tools (such as search, computation, translation, etc.) in a self-supervised manner, improving performance on zero-shot tasks.
ToolkenGPT: Augmenting Frozen Language Models with Massive Tools via Tool Embeddings
Shibo Hao (University of California San Diego), Zhiting Hu (University of California San Diego)
GenerationRobotic IntelligenceTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
🎯 What it does: This paper proposes ToolkenGPT, which allows a frozen LLM to call tools during text generation in the same way as generating ordinary words by learning a token (toolken) embedding for each tool.
Tools for Verifying Neural Models' Training Data
Dami Choi (University of Toronto), David Duvenaud (University of Toronto)
TransformerLarge Language ModelText
🎯 What it does: A Proof-of-Training-Data protocol is proposed to verify the source of training data, utilizing model checkpoints, memoized detection, and preset random seeds to assist third parties in validating the dataset used for training.
Top-Ambiguity Samples Matter: Understanding Why Deep Ensemble Works in Selective Classification
Qiang Ding (Chinese Academy of Sciences), Ping Luo (Chinese Academy of Sciences)
ClassificationConvolutional Neural NetworkTransformerImageText
🎯 What it does: This paper demonstrates through theoretical analysis and experimental validation that in selective classification tasks, Deep Ensemble can achieve lower selective risk than a single model within the coverage range, revealing that the performance improvement mainly comes from the ensemble effect of high-ambiguity samples.
Topological Obstructions and How to Avoid Them
Babak Esmaeili (Eindhoven University of Technology), Jan-Willem van de Meent (University of Amsterdam)
OptimizationRepresentation LearningFlow-based ModelAuto EncoderImageMultimodality
🎯 What it does: This study investigates the topological optimization obstacles encountered when learning geometric structure latent space encoders and proposes a variational flow with multimodal regularization (GF-VAE) to overcome these obstacles.
Topological Parallax: A Geometric Specification for Deep Perception Models
Abraham David Smith, Paul Bendich (Geometric Data Analytics)
Depth EstimationPoint Cloud
🎯 What it does: A method called 'topological parallax' based on topological data analysis is proposed to estimate the geometric structure of deep learning models and compare it with training data.
Topological RANSAC for instance verification and retrieval without fine-tuning
Guoyuan An (Electronics and Telecommunications Research Institute), Sung-eui Yoon
RetrievalContrastive LearningImage
🎯 What it does: A topology-based RANSAC framework is proposed for instance verification and retrieval, completely independent of task-specific fine-tuning datasets.
Topology-Aware Uncertainty for Image Segmentation
Saumya Gupta (Stony Brook University), Chao Chen (Stony Brook University)
SegmentationGraph Neural NetworkImage
🎯 What it does: In response to the weak segmentation of curved structures, this paper proposes a topology-based structural-level uncertainty quantification method;
TopoSRL: Topology preserving self-supervised Simplicial Representation Learning
Hiren Madhu (Indian Institute of Science), Sundeep Prabhakar Chepuri (Indian Institute of Science)
Representation LearningGraph Neural NetworkSpiking Neural NetworkContrastive LearningGraph
🎯 What it does: A self-supervised learning framework named TopoSRL has been developed, specifically designed for simplicial complexes, capable of capturing higher-order interactions and preserving topological information in the absence of labeled data.
TopP&R: Robust Support Estimation Approach for Evaluating Fidelity and Diversity in Generative Models
Pum Jun Kim (Ulsan National Institute of Science and Technology), Jaejun Yoo (Ulsan National Institute of Science and Technology)
GenerationData SynthesisGenerative Adversarial NetworkImage
🎯 What it does: A Topological Precision and Recall (TopP&R) evaluation metric is proposed to robustly estimate the support set of generative models, thereby providing a more reliable assessment of the authenticity and diversity of samples.
Toward Better PAC-Bayes Bounds for Uniformly Stable Algorithms
Sijia Zhou (University of Birmingham), Ata Kaban (University of Birmingham)
Optimization
🎯 What it does: This paper proposes a more precise PAC-Bayes generalization error upper bound for randomized learning algorithms (such as SGD and RCD), significantly improving the convergence rate of previous conclusions (approximately increased to √n times).
Toward Re-Identifying Any Animal
Bingliang Jiao (Northwestern Polytechnical University), Yanning Zhang (Northwestern Polytechnical University)
RecognitionRetrievalTransformerPrompt EngineeringVision Language ModelImageMultimodality
🎯 What it does: The task of 'Re-identify Any Animal in the Wild (ReID-AW)' is proposed, and a general ReID model capable of handling any wild animal category is constructed.
Toward Understanding Generative Data Augmentation
Chenyu Zheng (Renmin University of China), Chongxuan Li (Renmin University of China)
GenerationData SynthesisConvolutional Neural NetworkDiffusion modelGenerative Adversarial NetworkImage
🎯 What it does: A theoretical framework for Generative Data Augmentation (GDA) is proposed, deriving the generalization error upper bound under non-i.i.d. conditions using algorithm stability methods, and applying it to specific analyses of Gaussian Mixture Models and Generative Adversarial Networks (GANs), along with experimental validation.
Towards a fuller understanding of neurons with Clustered Compositional Explanations
Biagio La Rosa (Sapienza University of Rome), Roberto Capobianco (Sony AI)
Explainability and InterpretabilityComputational EfficiencyConvolutional Neural NetworkImage
🎯 What it does: The Clustered Compositional Explanations (CCE) algorithm is proposed, which generates logical formula explanations for neurons by clustering unit activations and using an improved heuristic search (MMESH) on each cluster, providing explanations over a broader range of activations. Based on this, a systematic analysis of the phenomena of 'non-specificity', 'gradual specificity', and 'ambiguity' of neurons under different activation intervals is conducted.
Towards A Richer 2D Understanding of Hands at Scale
Tianyi Cheng (University of Michigan), David Fouhey (University of Michigan)
Object DetectionSegmentationConvolutional Neural NetworkSupervised Fine-TuningImage
🎯 What it does: The Hands23 dataset and its model are proposed, achieving unified prediction of hand, object, second object detection, segmentation, contact state, grasp type, and tool usage in 2D images.
Towards a Unified Analysis of Kernel-based Methods Under Covariate Shift
Xingdong Feng (Shanghai University of Finance and Economics), Jingnan Zhang (University of Science and Technology of China)
ClassificationDomain AdaptationTabular
🎯 What it does: Under covariate shift, a unified theoretical analysis of kernel methods in RKHS space is conducted, providing convergence rates for different families of loss functions.
Towards a Unified Framework of Contrastive Learning for Disentangled Representations
Stefan Matthes (fortiss GmbH), Hao Shen (fortiss GmbH)
Anomaly DetectionAutonomous DrivingRepresentation LearningContrastive LearningImage
🎯 What it does: A unified contrastive learning framework is proposed for learning separable representations.
Towards Accelerated Model Training via Bayesian Data Selection
Zhijie Deng (Shanghai Jiao Tong University), Jun Zhu (Tsinghua University)
Computational EfficiencyData-Centric LearningTransformerContrastive LearningImage
🎯 What it does: A training acceleration method based on Bayesian data selection is proposed, utilizing zero-shot predictors to replace additional validation sets, filtering training data to eliminate noise and redundant samples, thereby improving model training efficiency.
Towards Anytime Classification in Early-Exit Architectures by Enforcing Conditional Monotonicity
Metod Jazbec (University of Amsterdam), Eric Nalisnick (University of Amsterdam)
ClassificationMixture of ExpertsImage
🎯 What it does: A post-hoc Product-of-Experts (PoE) transformation is proposed, which allows early exit neural networks to exhibit conditional monotonicity in computation time (i.e., the quality of predictions does not decrease at each step).
Towards Automated Circuit Discovery for Mechanistic Interpretability
Arthur Conmy (Independent), Adrià Garriga-Alonso (FAR AI)
Explainability and InterpretabilityTransformerGraph
🎯 What it does: This paper proposes an automated circuit discovery method called ACDC, which helps researchers identify subgraphs (circuits) in Transformer models that implement specific behaviors.
Towards Better Dynamic Graph Learning: New Architecture and Unified Library
Le Yu (Beihang University), Weifeng Lv (Beihang University)
Representation LearningGraph Neural NetworkTransformerGraphTime Series
🎯 What it does: A dynamic graph learning architecture based on Transformer, DyGFormer, and a unified duration graph learning library, DyGLib, are proposed to achieve reproducible, efficient, and scalable dynamic graph representation learning.
Towards Characterizing the First-order Query Complexity of Learning (Approximate) Nash Equilibria in Zero-sum Matrix Games
Hedi Hadiji (Universite Paris-Saclay), Wouter M Koolen
Optimization
🎯 What it does: This paper studies how to learn (approximately) Nash equilibria in finite action zero-sum matrix games under a first-order query model, and provides new lower and upper bound results.
Towards Combinatorial Generalization for Catalysts: A Kohn-Sham Charge-Density Approach
Phil Pope, David Jacobs (University of Maryland)
Graph Neural NetworkGraphPhysics Related
🎯 What it does: Train and evaluate a pointwise prediction model based on Kohn–Sham charge density to accelerate the self-consistent field (SCF) iterations of DFT, verifying its ability to achieve combinatorial generalization on mixed elements;
Towards Consistent Video Editing with Text-to-Image Diffusion Models
Zicheng Zhang (University of Chinese Academy of Sciences), Luoqi Liu (Meitu Inc.)
GenerationData SynthesisDiffusion modelVideoText
🎯 What it does: Proposes the EI 2 model, utilizing a text-image diffusion model for single video editing and enhancing temporal and semantic consistency.
Towards Data-Agnostic Pruning At Initialization: What Makes a Good Sparse Mask?
Hoang Pham (FPT Software AI Center), Long Tran-Thanh (University of Warwick)
OptimizationImage
🎯 What it does: The study investigates methods for network pruning during the initialization phase and proposes an evaluation framework based on subnetwork topology.
Towards Data-Algorithm Dependent Generalization: a Case Study on Overparameterized Linear Regression
Jing Xu (Tsinghua University), Andrew C Yao
OptimizationTabular
🎯 What it does: This paper proposes and studies the concept of data-algorithm compatibility, and uses gradient descent in over-parameterized linear regression to prove its sufficient conditions, demonstrating the positive impact of early stopping on generalization.
Towards Distribution-Agnostic Generalized Category Discovery
Jianhong Bai (Zhejiang University), Haoji Hu (Angelalign Technology Inc.)
ClassificationTransformerContrastive LearningImage
🎯 What it does: In the context of long-tail imbalance and open-world scenarios, a distribution-independent universal category discovery task that simultaneously handles known and unknown categories is proposed, along with the BaCon framework based on adversarial co-learning.
Towards Efficient and Accurate Winograd Convolution via Full Quantization
Chen Tianqi, Jian Cheng (Institute of Automation, Chinese Academy of Sciences)
OptimizationComputational EfficiencyConvolutional Neural NetworkSupervised Fine-TuningImage
🎯 What it does: This paper proposes a post-training quantization-based unified optimization method for Winograd convolution called PAW, along with a fully quantized factorized scale quantization (FSQ) to achieve efficient and accuracy-robust Winograd convolution.
Towards Efficient Image Compression Without Autoregressive Models
Muhammad Salman Ali (Kyung Hee University), Hui Yong Kim (Kyung Hee University)
CompressionConvolutional Neural NetworkTransformerImageVideo
🎯 What it does: By incorporating correlation loss during training, the latent variable space becomes more decorrelated, thereby improving the rate-distortion performance of hyperprior-based image compression without increasing model structure or inference time.
Towards Efficient Pre-Trained Language Model via Feature Correlation Distillation
Kun Huang (Ant Group), Meng Wang (Ant Group)
Computational EfficiencyKnowledge DistillationTransformerLarge Language ModelText
🎯 What it does: By constructing token-level and sample-level correlations on the output features of the Transformer block, Feature Correlation Distillation (FCD) is proposed to compress large pre-trained language models.
Towards Evaluating Transfer-based Attacks Systematically, Practically, and Fairly
Qizhang Li (Harbin Institute of Technology), Hao Chen (University of California Davis)
ClassificationAdversarial AttackConvolutional Neural NetworkTransformerImageBenchmark
🎯 What it does: A transfer attack benchmark named TA-Bench is proposed and implemented, covering over 30 existing transfer attack methods and conducting unified evaluations on 25 mainstream victim models.
Towards Foundation Models for Scientific Machine Learning: Characterizing Scaling and Transfer Behavior
Shashank Subramanian (Lawrence Berkeley National Lab), Amir Gholami (Institute for Computational Science and Engineering)
Domain AdaptationOptimizationComputational EfficiencyRepresentation LearningData-Centric LearningTransformerSupervised Fine-TuningTabularTime SeriesPhysics RelatedOrdinary Differential Equation
🎯 What it does: This paper constructs a large-scale and diverse pre-training dataset on various two-dimensional partial differential equations (Poisson equation, transport diffusion equation, and Helmholtz equation) and uses the Fourier Neural Operator (FNO) for pre-training to study the transfer learning behavior under different model scales, downstream data scales, variations in physical parameters, and multi-operator pre-training.
Towards Free Data Selection with General-Purpose Models
Yichen Xie (University of California Berkeley), Wei Zhan (University of California Berkeley)
Object DetectionSegmentationData-Centric LearningTransformerContrastive LearningImage
🎯 What it does: A data selection method called FreeSel is designed, which does not require task-specific models, is unsupervised, and extracts semantic patterns through a pre-trained visual Transformer for distance sampling, achieving efficient data labeling across different visual tasks.
Towards Generic Semi-Supervised Framework for Volumetric Medical Image Segmentation
Haonan Wang (Hong Kong University of Science and Technology), Xiaomeng Li (Hong Kong University of Science and Technology)
SegmentationDomain AdaptationKnowledge DistillationDiffusion modelImageBiomedical DataMagnetic Resonance ImagingComputed Tomography
🎯 What it does: A general semi-supervised framework (Aggregating & Decoupling, A&D) is proposed for volumetric medical image segmentation, capable of handling four scenarios: SSL, class-imbalanced SSL, UDA, and SemiDG.
Towards Higher Ranks via Adversarial Weight Pruning
Yuchuan Tian (Peking University), Yunhe Wang (Huawei)
Object DetectionOptimizationConvolutional Neural NetworkTransformerGenerative Adversarial NetworkImage
🎯 What it does: The paper proposes a rank-based adversarial weight pruning method (RPG) that enhances model performance by maintaining a high rank of the sparse weight matrix under high sparsity rates.
Towards Hybrid-grained Feature Interaction Selection for Deep Sparse Network
Fuyuan Lyu (McGill University), Xue Liu (McGill University)
Recommendation SystemOptimizationTabular
🎯 What it does: A mixed granularity (field-level and value-level) feature interaction selection method called OptFeature is proposed and implemented in a Deep Sparse Network (DSN).
Towards In-context Scene Understanding
Ivana Balazevic, Olivier J Henaff
SegmentationDepth EstimationRetrievalContrastive LearningImage
🎯 What it does: The paper studies in-context learning in visual tasks and proposes achieving dense scene understanding through nearest neighbor retrieval, eliminating the need for task-specific decoders or fine-tuning.
Towards Label Position Bias in Graph Neural Networks
Haoyu Han (Michigan State University), Jiliang Tang (Michigan State University)
Graph Neural NetworkGraph
🎯 What it does: This paper reveals and mitigates the Label Position Bias in graph neural networks, proposing a metric based on the Label Proximity Score (LPS) and enhancing model performance and fairness through Label Position Unbiased Structure Learning (LPSL).
Towards Label-free Scene Understanding by Vision Foundation Models
Runnan Chen (University of Hong Kong), Wenping Wang (Texas A&M University)
Object DetectionSegmentationContrastive LearningImagePoint Cloud
🎯 What it does: This paper explores how to utilize the visual foundation models CLIP and SAM to achieve label-free 2D/3D scene understanding, and proposes a Cross-Modal Noise Supervision (CNS) framework.
Towards Last-layer Retraining for Group Robustness with Fewer Annotations
Tyler LaBonte (Georgia Institute of Technology), Abhishek Kumar (Google DeepMind)
Supervised Fine-TuningTextBenchmark
🎯 What it does: The study investigates the last layer retraining method in the absence of group labels and proposes the SELF technique with no group labels and only a small number of category labels.
Towards Optimal Caching and Model Selection for Large Model Inference
Banghua Zhu (University of California Berkeley), Jiantao Jiao (University of California Berkeley)
OptimizationLarge Language ModelText
🎯 What it does: A joint optimization framework that combines caching and model multiplexing is proposed to reduce the inference cost and latency of large models.
Towards Optimal Effective Resistance Estimation
Rajat Vadiraj Dwaraknath (Stanford University), Aaron Sidford (Stanford University)
OptimizationGraph Neural Network
🎯 What it does: A new efficient resistance estimation and compression algorithm is proposed, providing upper and lower bounds for effective resistance concerning all edges or all vertex pairs in the graph.
Towards Personalized Federated Learning via Heterogeneous Model Reassembly
Jiaqi Wang (Pennsylvania State University), Fenglong Ma (Pennsylvania State University)
Federated LearningKnowledge DistillationImage
🎯 What it does: A personalized federated learning framework called pFedHR is proposed, which addresses the collaboration and personalization issues arising from the heterogeneity of model structures across different clients.
Towards Revealing the Mystery behind Chain of Thought: A Theoretical Perspective
Guhao Feng (Peking University), Liwei Wang (Peking University)
TransformerLarge Language ModelTextChain-of-Thought
🎯 What it does: The paper explores the theoretical mechanisms and performance of Chain of Thought (CoT) in large language models, analyzing its impact on arithmetic, equations, and dynamic programming problems.
Towards Robust and Expressive Whole-body Human Pose and Shape Estimation
Hui En Pang (Nanyang Technological University), Ziwei Liu (Nanyang Technological University)
Pose EstimationConvolutional Neural NetworkContrastive LearningImage
🎯 What it does: The RoboSMPLX framework is proposed to achieve full-body pose and shape estimation from monocular images, enhancing robustness and expressiveness.
Towards robust and generalizable representations of extracellular data using contrastive learning
Ankit Vishnubhotla (Columbia University), Cole Lincoln Hurwitz
ClassificationRepresentation LearningTransformerContrastive LearningBiomedical Data
🎯 What it does: The CEED framework is proposed, which generates robust features through contrastive learning of extracellular waveforms for spike sorting and cell type classification.
Towards Self-Interpretable Graph-Level Anomaly Detection
Yixin Liu (Monash University), Shirui Pan (Griffith University)
Anomaly DetectionExplainability and InterpretabilityGraph Neural NetworkContrastive LearningGraph
🎯 What it does: This paper studies the interpretable layer anomaly detection model SIGNET, which can provide anomaly scores and subgraph explanations under unsupervised conditions.
Towards Semi-Structured Automatic ICD Coding via Tree-based Contrastive Learning
Chang Lu (Stevens Institute of Technology), Yue Ning (Stevens Institute of Technology)
ClassificationContrastive LearningTextBiomedical DataElectronic Health Records
🎯 What it does: This paper proposes an automatic ICD code assignment method based on semi-structured clinical notes. It first automatically extracts and segments the titles and paragraphs of the notes, and then performs contrastive learning and masked training at the paragraph level to reduce note variability and enhance multi-label ICD coding performance.
Towards Stable Backdoor Purification through Feature Shift Tuning
Rui Min (Hong Kong University of Science and Technology), Minhao Cheng (Hong Kong University of Science and Technology)
ClassificationAdversarial AttackConvolutional Neural NetworkTransformerSupervised Fine-TuningImage
🎯 What it does: A backdoor removal method based on Feature Shift Tuning (FST) is proposed, specifically targeting backdoor attacks on deep learning models in low contamination scenarios.
Towards Symmetry-Aware Generation of Periodic Materials
Youzhi Luo (Texas A&M University), Shuiwang Ji (Texas A&M University)
GenerationOptimizationGraph Neural NetworkDiffusion modelScore-based ModelAuto EncoderGraphPhysics Related
🎯 What it does: A framework named SyMat for generating symmetric perception periodic materials is proposed, capable of generating chemical compositions, lattice parameters, and atomic coordinates that satisfy physical symmetry constraints.
Towards Test-Time Refusals via Concept Negation
Peiran Dong (Hong Kong Polytechnic University), Ziming Liu (Hong Kong Polytechnic University)
GenerationData SynthesisDiffusion modelImage
🎯 What it does: A framework for concept negation during inference, called PROTORE, is proposed, which can suppress user-specified negative concepts when generating images.
Towards the Difficulty for a Deep Neural Network to Learn Concepts of Different Complexities
Dongrui Liu (Shanghai Jiao Tong University), Quanshi Zhang (Shanghai Jiao Tong University)
Convolutional Neural NetworkRecurrent Neural NetworkImageTabular
🎯 What it does: This paper proves through theoretical derivation and experimental validation that deep neural networks are more capable of learning low-order interaction concepts (i.e., simple concepts) and explains that the complexity of concepts has an exponential relationship with learning difficulty.
Towards Unbounded Machine Unlearning
Meghdad Kurmanji (University of Warwick), Eleni Triantafillou (Google DeepMind)
OptimizationSafty and PrivacyConvolutional Neural NetworkContrastive LearningImage
🎯 What it does: This study investigates the 'forgetting' function of deep learning models and proposes a general forgetting method called SCRUB to address three practical needs: debiasing, eliminating label confusion, and user privacy.
Towards Understanding the Dynamics of Gaussian-Stein Variational Gradient Descent
Tianle Liu (Harvard University), Natesh S. Pillai (University of California)
OptimizationTabularStochastic Differential Equation
🎯 What it does: This paper conducts a theoretical analysis of the dynamics of Stein Variational Gradient Descent (SVGD) with a bilinear kernel on Gaussian targets, and presents the corresponding Gaussian-SVGD algorithm and its convergence properties.
Tracking Most Significant Shifts in Nonparametric Contextual Bandits
Joe Suk (Columbia University), Samory Kpotufe (Columbia University)
OptimizationReinforcement Learning
🎯 What it does: This paper studies the non-stationary case of contextual multi-armed bandits (Lipschitz contextual bandits) under a reward function that satisfies 1-smoothness, where the reward distribution varies over time and context space. The authors first provide an optimal dynamic loss lower bound characterized by the global change count L and total variation V, proving that existing methods cannot achieve this lower bound. They then introduce the concept of 'experienced significant shifts,' a more granular measure of non-stationarity that only accounts for significant changes in the best arm observed in the context. Based on this measure, they design an adaptive algorithm CMETA that can achieve a dynamic loss upper bound matching the lower bound without prior knowledge of L, V, or the timing of shifts.
Tracr: Compiled Transformers as a Laboratory for Interpretability
David Lindner (Google DeepMind), Vladimir Mikulik (Google DeepMind)
CompressionExplainability and InterpretabilityTransformerTabular
🎯 What it does: Tracr is proposed and implemented, which is a tool that compiles human-readable RASP programs into Transformer weights, capable of generating Transformer models with known structures for interpretability experiments and evaluations.
Trade-off Between Efficiency and Consistency for Removal-based Explanations
Yifan Zhang (Tsinghua University), Yang Yuan (Tsinghua University)
Explainability and InterpretabilityComputational EfficiencyImageText
🎯 What it does: This paper studies explanation methods based on feature removal and proposes the Incompatibility Triplet Theorem, indicating that interpretability, efficiency, and consistency cannot be satisfied simultaneously; under this framework, a lower error interpreter is designed.
Trading-off price for data quality to achieve fair online allocation
Mathieu Molina (Inria), Vianney Perchet (Inria)
Recommendation SystemOptimizationReinforcement Learning
🎯 What it does: A joint algorithm for online allocation problems is proposed, achieving long-term fairness and revenue maximization by purchasing data sources of different qualities without observing the protected attributes.
Train 'n Trade: Foundations of Parameter Markets
Tzu-Heng Huang (University of Wisconsin-Madison), Frederic Sala (University of Wisconsin-Madison)
OptimizationFederated LearningConvolutional Neural NetworkImage
🎯 What it does: This paper proposes a feasible parameter market framework that allows organizations to trade and exchange sets of neural network parameters without sharing the complete model, thereby reducing the cost of large-scale model training and accelerating convergence.
Train Faster, Perform Better: Modular Adaptive Training in Over-Parameterized Models
Yubin Shi (Fudan University), Li Shang (Fudan University)
TransformerImageText
🎯 What it does: This paper studies the modular-level learning dynamics of parameterized models, proposing a modular neural tangent kernel (mNTK) and designing a modular adaptive training (MAT) strategy based on its principal eigenvalue.
Train Hard, Fight Easy: Robust Meta Reinforcement Learning
Ido Greenberg (Technion), Eli Meirom (Nvidia Research)
Meta LearningReinforcement LearningSequential
🎯 What it does: This paper proposes a robust optimization framework for Meta-RL that utilizes CVaR (Conditional Value at Risk) to measure and enhance the minimum returns across different tasks, thereby achieving risk-sensitive Meta-RL.
Train Once and Explain Everywhere: Pre-training Interpretable Graph Neural Networks
Jun Yin (Central South University), Senzhang Wang (Central South University)
Explainability and InterpretabilityGraph Neural NetworkGraph
🎯 What it does: This paper proposes a pre-trained interpretable graph neural network π-GNN, which is first pre-trained on a synthetic graph dataset with ground truth explanations and then fine-tuned on various real graph datasets and tasks, achieving interpretability and predictive performance across datasets and tasks.
Train Once, Get a Family: State-Adaptive Balances for Offline-to-Online Reinforcement Learning
Shenzhi Wang (Tsinghua University), Gao Huang (Tsinghua University)
Reinforcement LearningTabularBenchmark
🎯 What it does: This paper proposes the FamO2O framework, which addresses the distribution shift problem in offline to online reinforcement learning by training various conservative/aggressive strategies and adaptively selecting the appropriate strategy based on the state.
Training biologically plausible recurrent neural networks on cognitive tasks with long-term dependencies
Wayne WM Soo, Xiao-Jing Wang (New York University)
Recurrent Neural NetworkTime SeriesSequential
🎯 What it does: This paper proposes and evaluates several methods for training biologically interpretable leak-type RNNs to learn long-term dependency tasks, including Coarse Discretization (CD), Skip Connection with Time (SCTT), and Aligned Skip Connection (DASC).
Training Chain-of-Thought via Latent-Variable Inference
Du Phan (Google), Rif A. Saurous (Google)
OptimizationComputational EfficiencyTransformerLarge Language ModelPrompt EngineeringTextChain-of-Thought
🎯 What it does: This paper proposes a training framework called TRICE based on MCMC-EM, which automatically generates effective reasoning chains by maximizing the marginal log-likelihood of answers in Chain of Thought (CoT), reducing reliance on manually annotated reasoning chains.
Training Energy-Based Normalizing Flow with Score-Matching Objectives
Chen-Hao Chao (National Tsing Hua University), Chun-Yi Lee (National Tsing Hua University)
GenerationData SynthesisOptimizationComputational EfficiencyScore-based ModelFlow-based ModelImage
🎯 What it does: This paper proposes Energy-Based Regularization Flow (EBFlow), which treats flow models as energy models and uses score matching objectives for training, thereby avoiding the computation of the Jacobian determinant of linear layers while maintaining expressive power.
Training Fully Connected Neural Networks is $\exists\mathbb{R}$-Complete
Daniel Bertschinger (ETH Zurich), Simon Weber (ETH Zurich)
OptimizationTabular
🎯 What it does: This paper proves that training a fully connected two-layer ReLU neural network (with both input and output being 2-dimensional) to minimize the error for a given sample is an ∃R-complete decision problem.
Training Neural Networks is NP-Hard in Fixed Dimension
Vincent Froese (Technische Universität Berlin), Christoph Hertrich (London School of Economics and Political Science)
Optimization
🎯 What it does: This paper studies the parameterized complexity of the training problem for two-layer neural networks (ReLU and linear threshold activation) under fixed input dimensions and the number of hidden neurons, providing proofs of NP-hardness and W[1]-hardness; it also presents a fixed-parameter tractable algorithm for the case of convex functions.
Training neural operators to preserve invariant measures of chaotic attractors
Ruoxi Jiang (University of Chicago), Rebecca Willett (University of Chicago)
Contrastive LearningTime SeriesPhysics Related
🎯 What it does: This paper proposes training neural operators through optimal transport or contrastive learning to maintain the invariant measure of chaotic attractors across multiple environments, thereby improving long-term statistical properties.
Training on Foveated Images Improves Robustness to Adversarial Attacks
Muhammad A Shah, Bhiksha Raj (Carnegie Mellon University)
OptimizationAdversarial AttackConvolutional Neural NetworkImage
🎯 What it does: A method for simulating low-resolution image preprocessing of the human retinal periphery, called R-Blur, is proposed, and a DNN is trained based on this to enhance robustness against adversarial attacks and common noise.
Training Private Models That Know What They Don’t Know
Stephan Rabanser (University of Toronto), Nicolas Papernot (Google DeepMind)
ClassificationSafty and PrivacyImage
🎯 What it does: This paper studies how to perform Selective Classification (SC) under the constraints of Differential Privacy (DP) and evaluates the performance of different SC methods under various privacy budgets through experiments.
Training shallow ReLU networks on noisy data using hinge loss: when do we overfit and is it benign?
Erin George (University of California Los Angeles), Deanna Needell (University of California Los Angeles)
Tabular
🎯 What it does: This paper studies the overfitting behavior when training a two-layer ReLU network using gradient descent and hinge loss on linearly separable data with label noise.
Training Transformers with 4-bit Integers
Haocheng Xi (Tsinghua University), Jun Zhu (Tsinghua University)
TransformerImageText
🎯 What it does: A full low-precision training method using INT4 integer matrix multiplication in Transformer training has been developed.
Training Transitive and Commutative Multimodal Transformers with LoReTTa
Manuel Tran (Roche Diagnostics GmbH), Eldad Klaiman (Roche Diagnostics GmbH)
ClassificationTransformerReinforcement LearningMultimodalityBiomedical Data
🎯 What it does: A self-supervised multimodal pre-training framework called LoReTTa is proposed to handle data with missing modality combinations.
Training Your Image Restoration Network Better with Random Weight Network as Optimization Function
Man Zhou (Nanyang Technological University), Chongyi Li (Nankai University)
RestorationSuper ResolutionConvolutional Neural NetworkImage
🎯 What it does: Proposes using random weight networks as a constraint function to improve the training method of image restoration networks, compatible with existing models and without increasing additional computational costs.
Training-free Diffusion Model Adaptation for Variable-Sized Text-to-Image Synthesis
Zhiyu Jin (Fudan University), Xiangyang Xue (Fudan University)
GenerationData SynthesisTransformerDiffusion modelImageText
🎯 What it does: Proposes a training-independent scaling factor to improve text-to-image synthesis of diffusion models at different resolutions.
Trajectory Alignment: Understanding the Edge of Stability Phenomenon via Bifurcation Theory
Minhak Song (Korea Advanced Institute of Science and Technology), Chulhee Yun (Korea Advanced Institute of Science and Technology)
OptimizationTabular
🎯 What it does: This paper studies the trajectory behavior of gradient descent when hyperparameter values are close to the Edge of Stability (EoS) by introducing a normalized reparameterization method, and finds that trajectories of GD with different initializations align to the same bifurcation diagram under second-order approximation.
Trans-Dimensional Generative Modeling via Jump Diffusion Models
Andrew Campbell (University of Oxford), Arnaud Doucet (University of Oxford)
GenerationData SynthesisRobotic IntelligenceTransformerDiffusion modelScore-based ModelVideoSequential
🎯 What it does: A jump diffusion model is proposed, which can simultaneously model the dimensions (quantity) and states (values) of data during the generation process, thereby directly generating data with variable dimensions.
Transfer learning for atomistic simulations using GNNs and kernel mean embeddings
John Isak Texas Falk, Massimiliano Pontil (Istituto Italiano di Tecnologia)
Graph Neural NetworkGraphPhysics Related
🎯 What it does: A transfer learning method based on kernel ridge regression (MEKRR) is proposed to model the potential energy surface of atomic systems by combining features from pre-trained graph neural networks (SCN, SchNet) with kernel mean embedding.
Transfer Learning with Affine Model Transformation
Shunya Minami (Institute of Statistical Mathematics), Ryo Yoshida (Institute of Statistical Mathematics)
Domain AdaptationOptimizationTabular
🎯 What it does: This paper proposes a supervised transfer learning framework based on affine model transformation, which can simultaneously capture cross-domain commonalities and specific factors through affine mapping between source domain features and target domain outputs.