ICML 2023 Papers — Page 17
International Conference on Machine Learning · 1828 papers
Surrogate Model Extension (SME): A Fast and Accurate Weight Update Attack on Federated Learning
Junyi Zhu (KU Leuven), Matthew B. Blaschko (KU Leuven)
Federated LearningComputational EfficiencyAdversarial AttackImage
🎯 What it does: A new gradient inversion attack is proposed - Surrogate Model Extension (SME), which can quickly and accurately recover local training data from multi-step weight updates transmitted in federated learning.
Surrogate Module Learning: Reduce the Gradient Error Accumulation in Training Spiking Neural Networks
Shikuang Deng (University of Electronic Science and Technology of China), Shi Gu (University of Electronic Science and Technology of China)
Knowledge DistillationSpiking Neural NetworkImage
🎯 What it does: A training framework called Surrogate Module Learning (SML) is proposed, which constructs a more accurate gradient propagation path by inserting auxiliary modules (Surrogate Module) into SNNs, thereby reducing the accumulation of surrogate gradient errors.
SWARM Parallelism: Training Large Models Can Be Surprisingly Communication-Efficient
Max Ryabinin (HSE University), Alexander Borzunov (HSE University)
TransformerLarge Language ModelText
🎯 What it does: The SWARM parallel algorithm is proposed, which can efficiently train trillion-parameter Transformer models on low-speed, heterogeneous, and failure-prone devices.
Symmetry-Aware Robot Design with Structured Subgroups
Heng Dong (Tsinghua University), Chongjie Zhang (Tsinghua University)
Robotic IntelligenceGraph Neural NetworkReinforcement LearningSequential
🎯 What it does: This paper proposes a Symmetry-Aware Robot Design (SARD) framework that generates controllable and efficient robots by searching for and utilizing symmetry during the robot design process.
Synergies between Disentanglement and Sparsity: Generalization and Identifiability in Multi-Task Learning
Sebastien Lachapelle, Quentin Bertrand (Mila)
OptimizationRepresentation LearningMeta LearningSupervised Fine-TuningImage
🎯 What it does: This study investigates the synergy between decoupled representations and sparse task-specific predictors in multi-task learning, demonstrating that sparse regularization can enhance the generalization performance of decoupled representations, and provides an interpretable identification theory along with a feasible bi-level optimization method.
Synthetic data for model selection
Alon Shoshan (Amazon), Gerard Medioni
Data SynthesisHyperparameter SearchGenerative Adversarial NetworkImage
🎯 What it does: Using synthetic images instead of traditional validation sets to achieve model selection (including early stopping, random seed selection, and hyperparameter search)
Synthetic Data, Real Errors: How (Not) to Publish and Use Synthetic Data
Boris van Breugel (University of Cambridge), Mihaela van der Schaar (Alan Turing Institute)
GenerationData SynthesisGenerative Adversarial NetworkTabularBiomedical Data
🎯 What it does: This paper addresses the errors present in synthetic data and proposes a Deep Generative Ensemble (DGE) framework;
Synthetic Prompting: Generating Chain-of-Thought Demonstrations for Large Language Models
Zhihong Shao (Tsinghua University), Weizhu Chen (Microsoft Azure AI)
TransformerLarge Language ModelPrompt EngineeringTextChain-of-Thought
🎯 What it does: Using a limited number of manual examples and leveraging the generative capabilities of large language models, we automatically synthesize more chain reasoning examples through forward-backward iteration, selecting the most complex and diverse examples during reasoning to enhance performance.
System Identification of Neural Systems: If We Got It Right, Would We Know?
Yena Han (Massachusetts Institute of Technology), Brian Cheung (Massachusetts Institute of Technology)
RecognitionConvolutional Neural NetworkRecurrent Neural NetworkTransformerImage
🎯 What it does: This paper evaluates the reliability of system identification methods such as linear regression and Centered Kernel Alignment (CKA) in recognizing different neural network architectures by simulating the brain on known artificial networks.
TabDDPM: Modelling Tabular Data with Diffusion Models
Akim Kotelnikov (Higher School of Economics), Artem Babenko (Yandex)
GenerationData SynthesisDiffusion modelTabular
🎯 What it does: A diffusion model named TabDDPM is proposed, specifically designed for generating tabular data with mixed numerical and categorical features.
TabLeak: Tabular Data Leakage in Federated Learning
Mark Vero (ETH Zurich), Martin Vechev (ETH Zurich)
Federated LearningSafty and PrivacyTabular
🎯 What it does: A reconstruction attack method for tabular data in federated learning, called TabLeak, is proposed, which can recover a large amount of private data from gradient updates under different batch sizes and training protocols.
Taming graph kernels with random features
Krzysztof Marcin Choromanski (Google DeepMind), Krzysztof Marcin Choromanski (Columbia University)
OptimizationComputational EfficiencyGraph Neural NetworkGraph
🎯 What it does: This paper proposes a Graph Random Features (GRFs) framework, which constructs unbiased graph kernel approximations using random walks, significantly reducing the traditional graph kernel's O(N³) computational complexity.
TAN Without a Burn: Scaling Laws of DP-SGD
Tom Sander (Meta AI), Alexandre Sablayrolles (Meta AI)
OptimizationSafty and PrivacyComputational EfficiencyConvolutional Neural NetworkSupervised Fine-TuningImage
🎯 What it does: This study explores how to optimize differential privacy (DP) methods during the training of deep neural networks by decoupling privacy analysis and experimental behavior, particularly using Total Amount of Noise (TAN) to reduce computational resource requirements.
Target-Aware Generative Augmentations for Single-Shot Adaptation
Kowshik Thopalli (Lawrence Livermore National Laboratory), Jayaraman J. Thiagarajan (Arizona State University)
GenerationData SynthesisDomain AdaptationGenerative Adversarial NetworkImage
🎯 What it does: This paper proposes a technique for adaptive adjustment of the source model based on a single target sample, utilizing target domain information to generate synthetic data tailored for the target domain, thereby enhancing the model's generalization ability to the target domain.
Target-based Surrogates for Stochastic Optimization
Jonathan Wilder Lavington (University of British Columbia), Nicolas Le Roux (Microsoft Research)
OptimizationReinforcement LearningTabularSequential
🎯 What it does: This paper proposes to construct a global upper bound surrogate function (target-smoothness surrogate) in scenarios where gradients are expensive, and to effectively approximate the original objective function through multi-step parameter updates, presenting the corresponding SSO (Stochastic Surrogate Optimization) algorithm.
Task-specific experimental design for treatment effect estimation
Bethany Connolly (Faculty), Christopher Frye (Faculty)
TabularBiomedical Data
🎯 What it does: This paper proposes an experimental design method for downstream tasks, achieving improved sample efficiency in causal effect estimation through an adaptive sampling strategy.
Task-Specific Skill Localization in Fine-tuned Language Models
Abhishek Panigrahi (Princeton University), Sanjeev Arora (Princeton University)
TransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
🎯 What it does: Research and prove that there exists a very small subset of parameters (about 0.01%) in fine-tuned language models that can retain over 95% of task performance, and propose a method for locating these 'skill' areas through model grafting.
Taxonomy-Structured Domain Adaptation
Tianyi Liu (Rutgers University), Hao Wang (Rutgers University)
Domain AdaptationGenerative Adversarial NetworkImage
🎯 What it does: A taxonomic-structured domain adaptation method (TSDA) is proposed, which balances domain invariance and domain similarity in adversarial learning by utilizing the hierarchical classification structure of domains.
Team Belief DAG: Generalizing the Sequence Form to Team Games for Fast Computation of Correlated Team Max-Min Equilibria via Regret Minimization
Brian Hu Zhang (Carnegie Mellon University), Tuomas Sandholm (Carnegie Mellon University)
OptimizationComputational EfficiencyReinforcement LearningSequential
🎯 What it does: This paper proposes the Team Belief DAG (TB-DAG) representation, which maps the team strategy space to a log-convex polytope, and uses a CFR-based online annealing algorithm to solve the cooperative max-min equilibrium on it.
Temporal Label Smoothing for Early Event Prediction
Hugo Yèche (ETH Zürich), Rita Kuznetsova (ETH Zürich)
ClassificationAnomaly DetectionTime SeriesBiomedical DataElectronic Health Records
🎯 What it does: This paper proposes Temporal Label Smoothing (TLS), a label smoothing method that utilizes temporal structure and maintains the monotonicity of predictions in early event forecasting.
Temporally Consistent Transformers for Video Generation
Wilson Yan (University of California Berkeley), Pieter Abbeel (University of California Berkeley)
GenerationData SynthesisTransformerGenerative Adversarial NetworkVideo
🎯 What it does: A Transformer video generation model capable of capturing long-term consistency is proposed, and three large long-term datasets are constructed for evaluation.
Tensor Decompositions Meet Control Theory: Learning General Mixtures of Linear Dynamical Systems
Ainesh Bakshi (Massachusetts Institute of Technology), morris yau
Time SeriesSequential
🎯 What it does: An algorithm based on tensor decomposition is proposed to learn the parameters and mixing weights of Mixture of Linear Dynamical Systems (M-LDS) and achieve parameter identification in partially observable situations.
Tensor Gaussian Process with Contraction for Multi-Channel Imaging Analysis
Hu Sun (University of Michigan), Yang Chen (University of Michigan)
Image TranslationRestorationOptimizationExplainability and InterpretabilityGaussian SplattingImage
🎯 What it does: This paper proposes a framework that combines tensor contraction with Tensor Gaussian Process (Tensor‑GPST) for scalar regression prediction of multi-channel images, achieving interpretable low-dimensional tensor representations through anisotropic total variation regularization.
Test-time Adaptation with Slot-Centric Models
Mihir Prabhudesai (Carnegie Mellon University), Katerina Fragkiadaki (Carnegie Mellon University)
SegmentationDomain AdaptationTransformerAuto EncoderImageMultimodalityPoint Cloud
🎯 What it does: A semi-supervised slot center model, Slot-TTA, is proposed, which utilizes reconstruction (or cross-view synthesis) to perform one-time gradient descent on each test sample, achieving test-time adaptation for scene deconstruction.
Test-Time Style Shifting: Handling Arbitrary Styles in Domain Generalization
Jungwuk Park (Korea Advanced Institute of Science and Technology), Jaekyun Moon (Korea Advanced Institute of Science and Technology)
ClassificationRetrievalDomain AdaptationConvolutional Neural NetworkContrastive LearningImage
🎯 What it does: Two methods are proposed: Test-Time Style Shifting and Style Balancing, to achieve style alignment for unknown target domains in domain generalization without model updates, significantly improving prediction performance across any style domain.
Text Generation with Diffusion Language Models: A Pre-training Approach with Continuous Paragraph Denoise
Zhenghao Lin (Xiamen University), Weizhu Chen (Microsoft)
GenerationTransformerLarge Language ModelDiffusion modelText
🎯 What it does: The language model GENIE, based on a diffusion model, uses a Continuous Paragraph Denoising (CPD) task during the pre-training phase, aiming to achieve text generation.
Text-To-4D Dynamic Scene Generation
Uriel Singer (Meta AI), Yaniv Taigman (Meta AI)
GenerationData SynthesisDiffusion modelNeural Radiance FieldVideoText
🎯 What it does: Generate 4D dynamic scenes (3D + time) that can be rendered from any perspective based on text descriptions, and can be embedded into 3D environments or games.
Text-To-Concept (and Back) via Cross-Model Alignment
Mazda Moayeri (University of Maryland), Soheil Feizi (Meta AI)
ClassificationRetrievalTransformerLarge Language ModelContrastive LearningImageText
🎯 What it does: This paper proposes a method that requires training only a single linear layer to align the feature space of any offline visual model to the CLIP space, thereby achieving a mapping from text to concept vectors. Based on this, it enables zero-shot classification, concept bottleneck models, dataset distribution diagnosis, concept logic retrieval, and reverse decoding from concepts to text.
TGRL: An Algorithm for Teacher Guided Reinforcement Learning
Idan Shenfeld (Massachusetts Institute of Technology), Pulkit Agrawal (Massachusetts Institute of Technology)
Reinforcement LearningMultimodality
🎯 What it does: This paper proposes a Teacher-Guided Reinforcement Learning (TGRL) algorithm that dynamically balances teacher supervision and reward signals, automatically adjusting their weights to achieve better performance across various tasks.
The Acquisition of Physical Knowledge in Generative Neural Networks
Luca M. Schulze Buschoff (Max Planck Institute for Biological Cybernetics), Marcel Binz (Max Planck Institute for Biological Cybernetics)
GenerationData SynthesisOptimizationVideoPhysics Related
🎯 What it does: Using the β-VAE (RSSM) framework, we implement two human development hypotheses, 'random optimization' and 'complexity increment', to train an unsupervised video generation model. We evaluate its understanding of physical rules using the violation of expectation (VOE) method and compare its learning trajectory with the developmental stages of children in support, occlusion, and collision events.
The Benefits of Mixup for Feature Learning
Difan Zou (University of Hong Kong), Quanquan Gu (University of California)
Representation LearningConvolutional Neural NetworkImage
🎯 What it does: This paper theoretically analyzes the advantages of Mixup training from the perspective of feature learning, proving that it can effectively learn rare features and improve generalization.
The Benefits of Model-Based Generalization in Reinforcement Learning
Kenny John Young, Jürgen Schmidhuber (International Institute of Applied Systems Analysis)
Reinforcement LearningTabular
🎯 What it does: Through theoretical proof (Theorem 1.1) and large-scale experiments, this paper evaluates and demonstrates the significant improvement in sampling efficiency of reinforcement learning when using learned models to generate imagined experiences in environments with factorized structures.
The Blessing of Heterogeneity in Federated Q-Learning: Linear Speedup and Beyond
Jiin Woo (Carnegie Mellon University), Yuejie Chi (Carnegie Mellon University)
Federated LearningReinforcement Learning
🎯 What it does: This paper studies federated Q-learning, aiming to learn the optimal Q function by periodically aggregating local Q estimates trained on local data. It focuses on infinite-horizon tabular Markov decision processes and provides sample complexity guarantees for both synchronous and asynchronous federated Q-learning.
The case for 4-bit precision: k-bit Inference Scaling Laws
Tim Dettmers (University of Washington), Luke Zettlemoyer (University of Washington)
TransformerLarge Language ModelText
🎯 What it does: This study investigates the scaling laws of zero-shot inference in large language models under different bit precisions, comparing the impact of 8, 6, 4, and 3-bit quantization on zero-shot accuracy and model bit count.
The Catalog Problem: Clustering and Ordering Variable-Sized Sets
Mateusz Maria Jurewicz (Tjek A/S), Leon Derczynski (University of Washington)
TransformerSupervised Fine-TuningTabular
🎯 What it does: The Catalog Problem is proposed, which involves predicting a variable number of ordered clusters in a set of arbitrary size, addressing the joint task of set clustering and sorting.
The Computational Complexity of Concise Hypersphere Classification
Eduard Eiben (Royal Holloway University of London), Stefan Szeider (Algorithms and Complexity Group TU Wien)
ClassificationOptimization
🎯 What it does: This paper conducts a systematic complexity study of the binary data hypersphere classification (BHC) problem from the perspective of parameterized complexity, proving its NP-hardness, FPT, and XP results under various parameters, and providing corresponding algorithms and lower bounds.
The Dormant Neuron Phenomenon in Deep Reinforcement Learning
Ghada Sokar (Eindhoven University of Technology), Utku Evci (Google DeepMind)
Reinforcement LearningVideo
🎯 What it does: Identifying and mitigating the 'dormant neuron' phenomenon in deep reinforcement learning
The Edge of Orthogonality: A Simple View of What Makes BYOL Tick
Pierre Harvey Richemond, Felix Hill (Google DeepMind)
OptimizationRepresentation LearningConvolutional Neural NetworkContrastive LearningImage
🎯 What it does: This paper clarifies that the linear predictor in BYOL approximates an orthogonal projection through linear algebra analysis, thereby explaining its non-collapse mechanism and proposing various fixed-form predictors based on closed-form matrix solutions.
The Fast Johnson-Lindenstrauss Transform Is Even Faster
Ora Nova Fandina (Aarhus University), Kasper Green Larsen (Aarhus University)
🎯 What it does: This paper provides a detailed analysis of the classic Fast Johnson-Lindenstrauss transform, offering a more compact sparsity parameter and demonstrating a significant reduction in its embedding time.
The Flan Collection: Designing Data and Methods for Effective Instruction Tuning
Shayne Longpre (Massachusetts Institute of Technology), Adam Roberts (Google)
TransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextChain-of-Thought
🎯 What it does: This study investigates the Flan 2022 instruction fine-tuning collection, systematically breaking down its methods, public data, and templates, and demonstrating its significant improvement on the T5 model.
The Hessian perspective into the Nature of Convolutional Neural Networks
Sidak Pal Singh (ETH Zurich), Bernhard Schölkopf (MPI for Intelligent Systems)
Convolutional Neural NetworkImage
🎯 What it does: This paper theoretically analyzes the structure and rank of the Hessian of the loss function of convolutional neural networks (CNNs), revealing the impact of CNN architecture on the Hessian and providing an upper bound.
The Ideal Continual Learner: An Agent That Never Forgets
Liangzu Peng (Johns Hopkins University), Rene Vidal (NORCE Norwegian Research Centre)
Tabular
🎯 What it does: Proposes the Ideal Continual Learner (ICL) framework, theoretically proving that under the assumption of shared multi-task models, catastrophic forgetting can be completely avoided, and unifying it with existing methods such as memory, regularization, and expansion;
The Impact of Exploration on Convergence and Performance of Multi-Agent Q-Learning Dynamics
Aamal Hussain (Imperial College London), Dario Paccagnan (Imperial College London)
OptimizationReinforcement LearningGraph
🎯 What it does: This study investigates how the exploration rate affects the convergence of learning and system payoffs in the dynamic of multi-agent Q-learning (SQL) in online games, proving that even if it does not converge, the dynamics remain within the QRE neighborhood, and a higher exploration rate reduces payoffs.
The Implicit Regularization of Dynamical Stability in Stochastic Gradient Descent
Lei Wu (Peking University), Weijie J Su
Optimization
🎯 What it does: This paper studies the implicit regularization of Stochastic Gradient Descent (SGD), particularly analyzing the characteristics of SGD through dynamic stability.
The Monge Gap: A Regularizer to Learn All Transport Maps
Théo Uscidda (ENSAE), marco cuturi
OptimizationDrug DiscoveryBiomedical Data
🎯 What it does: A new regularization method called Monge gap is proposed for learning optimal transport mappings under arbitrary costs without imposing strict constraints on the network structure.
The Numerical Stability of Hyperbolic Representation Learning
Gal Mishne (University of California San Diego), Sheng Yang (Harvard University)
OptimizationRepresentation LearningTabularBiomedical Data
🎯 What it does: Analyzed the numerical representation capacity and gradient vanishing problem of the Poincaré ball and Lorentz model under 64-bit floating point, and proposed Euclidean parameterization to overcome these limitations.
The Optimal Approximation Factors in Misspecified Off-Policy Value Function Estimation
Philip Amortila (University of Illinois), Csaba Szepesvari
Reinforcement Learning
🎯 What it does: This study investigates the error amplification factor of linear value function estimation in offline reinforcement learning and provides the optimal approximation ratio for state coverage and feature differentiation.
The Persistent Laplacian for Data Science: Evaluating Higher-Order Persistent Spectral Representations of Data
Thomas Davies (University of Southampton), Ruben Sanchez-Garcia (Halıcıoğlu Data Science Institute University of California San Diego)
ClassificationRepresentation LearningImage
🎯 What it does: This paper proposes and experimentally evaluates the Persistent Laplacian as a data embedding feature vector, exploring its effectiveness in downstream machine learning tasks.
The Power of Learned Locally Linear Models for Nonlinear Policy Optimization
Daniel Pfrommer (Massachusetts Institute of Technology), Stephen Tu (Google Brain)
OptimizationRobotic IntelligenceReinforcement Learning from Human FeedbackReinforcement LearningSequential
🎯 What it does: This paper proposes a learning-based trajectory optimization algorithm that utilizes local linear models to iteratively estimate dynamics and perform iLQR-style updates in nonlinear control.
The Power of Preconditioning in Overparameterized Low-Rank Matrix Sensing
Xingyu Xu (Carnegie Mellon University), Cong Ma (University of Chicago)
OptimizationTabular
🎯 What it does: A preconditioned gradient descent algorithm ScaledGD(λ) is proposed for the low-rank matrix sensing problem, where the rank is unknown and the matrix may have a large condition number, to quickly converge from a small random initialization to the true low-rank matrix.
The Power of Uniform Sampling for k-Median
Lingxiao Huang (Nanjing University), Jianing Lou (Peking University)
OptimizationTabular
🎯 What it does: This paper studies the theoretical and experimental performance of using uniform sampling for dimensionality reduction in the k-Median clustering problem, providing lower and approximate upper bounds on the sample size, and proving that uniform sampling can form an ε-weak core set.
The Price of Differential Privacy under Continual Observation
Palak Jain (Boston University), Adam Smith (Boston University)
Safty and Privacy
🎯 What it does: This paper studies the error limits of differential privacy mechanisms under the continual release model and provides strong lower bounds for two fundamental problems (MaxSum and SumSelect); it also proposes and analyzes a data stream model that allows for adaptive input selection.
The Regret of Exploration and the Control of Bad Episodes in Reinforcement Learning
Victor Boone (University Grenoble Alpes), Bruno Gaujal (University Grenoble Alpes)
Reinforcement Learning
🎯 What it does: This paper proposes a new performance metric for reinforcement learning (RL) algorithms, called the regret of exploration, to measure the asymptotic cost of exploration. Additionally, a new performance test (PT) is introduced to terminate episodes in RL optimistic algorithms, which is based on the performance of the current policy compared to the best policy in the current confidence set.
The Role of Entropy and Reconstruction in Multi-View Self-Supervised Learning
Borja Rodríguez Gálvez (KTH Royal Institute of Technology), Luca Zappella (Apple)
OptimizationRepresentation LearningConvolutional Neural NetworkContrastive LearningImage
🎯 What it does: This paper studies the impact of entropy and reconstruction on mutual information in Multi-View Self-Supervised Learning (MVSSL), proposing and validating an optimization framework based on the ER (Entropy + Reconstruction) lower bound.
The Saddle-Point Method in Differential Privacy
Wael Alghamdi (Harvard University), Lalitha Sankar (Arizona State University)
OptimizationSafty and Privacy
🎯 What it does: This paper proposes a Saddle-Point Accountant (SPA) based on the saddle point method, which can accurately estimate differential privacy parameters in constant time under large compositions (many iterations);
The SSL Interplay: Augmentations, Inductive Bias, and Generalization
Vivien Cabannes (Meta AI), Alberto Bietti (Meta AI)
Representation LearningData-Centric LearningContrastive LearningImage
🎯 What it does: A theoretical framework is proposed to systematically analyze the interactions between data augmentation, network structure (induced bias), and training algorithms in self-supervised learning (SSL), along with an analysis of generalization performance for pre-training and downstream tasks.
The Statistical Benefits of Quantile Temporal-Difference Learning for Value Estimation
Mark Rowland (DeepMind), Will Dabney (Google Research)
Reinforcement LearningTabular
🎯 What it does: This study investigates the statistical advantages of Quantile Temporal Difference Learning (QTD) in estimating expected returns in tabular policy evaluation and systematically compares it with traditional TD.
The Statistical Scope of Multicalibration
Georgy Noarov (University of Pennsylvania), Aaron Roth (University of Pennsylvania)
OptimizationFinance Related
🎯 What it does: This paper proposes a connection between multicalibration and property elicitation, providing a complete theory on the feasibility of multicalibration for continuous attributes. It also presents general batch and online algorithms, extending to conditionally extractable two-dimensional attributes, and conducts application and negation analysis on financial risk measures such as CVaR and deformation risk measures.
The Test of Tests: A Framework for Differentially Private Hypothesis Testing
Zeki Kazan (Duke University), Andrew Bray (University of California Berkeley)
Safty and PrivacyTabularBiomedical Data
🎯 What it does: This paper proposes a general 'Test of Tests' framework that can black-box convert any traditional hypothesis test into a differentially private version, providing complete implementation and theoretical analysis.
The Unintended Consequences of Discount Regularization: Improving Regularization in Certainty Equivalence Reinforcement Learning
Sarah Rathnam (Harvard University), Finale Doshi-Velez (Harvard University)
Reinforcement LearningBiomedical Data
🎯 What it does: This study investigates the unintended consequences of discount regularization in confidence-equivalent reinforcement learning and proposes a state-action specific regularization method.
The Unreasonable Effectiveness of Few-shot Learning for Machine Translation
Xavier Garcia (Google DeepMind), Orhan Firat (Google DeepMind)
TransformerText
🎯 What it does: Achieved unsupervised machine translation during the inference phase using only 5 high-quality translation examples, and compared it with traditional supervised models.
The Value of Out-of-Distribution Data
Ashwin De Silva (Johns Hopkins University), Joshua T Vogelstein
Domain AdaptationConvolutional Neural NetworkImage
🎯 What it does: The study investigates the impact of incorporating out-of-domain (OOD) samples into the training data on the generalization error of the target task, finding that the error exhibits a non-monotonic relationship with the number of OOD samples.
The Virtues of Laziness in Model-based RL: A Unified Objective and Algorithms
Anirudh Vemula (Aurora Innovation), Sanjiban Choudhury (Cornell University)
Reinforcement Learning
🎯 What it does: A new unified objective based on model-based reinforcement learning is proposed (Performance Difference via Advantage in Model, PDAM), and two 'lazy' algorithms, LAMPS and LAMPS-MM, are designed to utilize this objective, significantly reducing the computational overhead of finding the optimal policy in learning models, while eliminating the target mismatch problem between model fitting and policy computation through value instantaneous matching.
The Wisdom of Hindsight Makes Language Models Better Instruction Followers
Tianjun Zhang (University of California), Joseph E. Gonzalez (University of California)
Reinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningContrastive LearningText
🎯 What it does: A two-stage reward-free learning algorithm based on Hindsight Instruction Relabeling (HIR) is proposed, which achieves alignment by having the language model relabel instructions on its own generated outputs.
Theoretical Behavior of XAI Methods in the Presence of Suppressor Variables
Rick Wilming (Technische Universität Berlin), Stefan Haufe (Charité Universitätsmedizin Berlin)
Explainability and InterpretabilityTabular
🎯 What it does: This paper studies the theoretical behavior of explainable artificial intelligence (XAI) methods in the presence of suppressor variables, particularly how to analyze the performance of these methods in simple binary classification problems.
Theoretical Bounds on the Network Community Profile from Low-rank Semi-definite Programming
Yufan Huang (Purdue University), David F. Gleich
OptimizationGraph
🎯 What it does: Research and provide computable low-order lower bounds, derive the µ-directed network community profile (NCP);
Theoretical Guarantees of Learning Ensembling Strategies with Applications to Time Series Forecasting
Hilaf Hasson (Amazon), Youngsuk Park (Amazon)
TabularTime Series
🎯 What it does: Proved the theoretical guarantees of stacking generalization and proposed a learnable weighted ensemble method for probabilistic time series forecasting, capable of selecting the optimal stacking model from a finite-dimensional family through cross-validation.
Theory on Forgetting and Generalization of Continual Learning
Sen Lin (Ohio State University), Ness Shroff (Ohio State University)
Convolutional Neural NetworkImage
🎯 What it does: Theoretical derivation of catastrophic forgetting and generalization error in continual learning under over-parameterized linear models is presented, providing a closed-form expectation formula. Further analysis is conducted on the impact of hyper-parameterization, task similarity, and task order on forgetting and generalization error, with theoretical insights transferred to practical deep network experiments.
Thompson Sampling for High-Dimensional Sparse Linear Contextual Bandits
Sunrit Chakraborty (University of Michigan), Ambuj Tewari (University of Michigan)
OptimizationReinforcement Learning from Human FeedbackBiomedical Data
🎯 What it does: A Thompson sampling algorithm for high-dimensional sparse linear contextual bandits has been developed, and an approximately optimal upper bound on cumulative regret has been provided.
Thompson Sampling with Diffusion Generative Prior
Yu-Guan Hsieh (Université Grenoble Alpes), Patrick Blöbaum (Amazon)
Recommendation SystemOptimizationMeta LearningReinforcement LearningDiffusion modelTabular
🎯 What it does: This study investigates the use of denoising diffusion models for the prior in meta-learning multi-armed bandits and designs a Thompson Sampling algorithm based on this prior.
Thompson Sampling with Less Exploration is Fast and Optimal
Tianyuan Jin (National University of Singapore), Pan Xu (Duke University)
OptimizationReinforcement Learning from Human FeedbackTabular
🎯 What it does: The ϵ-Exploring Thompson Sampling (ϵ-TS) algorithm is proposed, improving traditional Thompson Sampling by primarily using empirical mean greedy selection of arms most of the time, and only performing posterior sampling with probability ϵ, thereby reducing the number of explorations.
TIDE: Time Derivative Diffusion for Deep Learning on Graphs
Maysam Behmanesh (Ecole Polytechnique), Maks Ovsjanikov
Graph Neural NetworkGraph
🎯 What it does: A graph neural network TIDE based on learnable temporal derivative diffusion is proposed, improving the traditional message passing framework to achieve global information propagation and avoid over-smoothing.
Tied-Augment: Controlling Representation Similarity Improves Data Augmentation
Emirhan Kurtuluş, Ekin Dogus Cubuk
Representation LearningData-Centric LearningConvolutional Neural NetworkContrastive LearningImage
🎯 What it does: Proposes the Tied-Augment framework, which uses a dual-branch network with shared weights during training to apply two types (or the same type) of random augmentations to the same image. In addition to the conventional cross-entropy loss, it incorporates a feature similarity loss to make representations under different augmented views more similar or controllable, thereby enhancing the effectiveness of data augmentation.
Tight and fast generalization error bound of graph embedding in metric space
Atsushi Suzuki (University of Tokyo), Kenji Yamanishi (University of Tokyo)
Graph Neural NetworkGraph
🎯 What it does: This paper proposes a new upper bound on the generalization error of graph embeddings, achieved by evaluating the local Rademacher complexity of the model, indicating that the performance of graph embeddings in non-Euclidean metric spaces surpasses that suggested by existing upper bounds.
Tight Certification of Adversarially Trained Neural Networks via Nonconvex Low-Rank Semidefinite Relaxations
Hong-Ming Chiu (University of Illinois), Richard Y. Zhang (University of Illinois)
OptimizationAdversarial AttackConvolutional Neural NetworkImage
🎯 What it does: A non-convex neural network robustness provable method based on low-rank SDP constraints is proposed, overcoming the traditional LP approximation barrier.
Tight Data Access Bounds for Private Top-$k$ Selection
Hao WU, Anthony Wirth (University of Melbourne)
Safty and PrivacyComputational Efficiency
🎯 What it does: A Top-k selection algorithm under the differential privacy model is proposed, aiming to minimize the number of data accesses;
Tight Regret Bounds for Single-pass Streaming Multi-armed Bandits
Chen Wang (Rutgers University)
Reinforcement LearningTabular
🎯 What it does: In the single-channel streaming multi-armed bandit, the paper presents optimal upper and lower bounds and provides an expected zero-loss algorithm with a memory requirement of log*(K);
Tighter Analysis for ProxSkip
Zhengmian Hu (University of Maryland), Heng Huang (University of Pittsburgh)
OptimizationOrdinary Differential Equation
🎯 What it does: Conduct a more rigorous convergence analysis of ProxSkip and its variant algorithms, propose new Lyapunov quantities, and achieve better convergence rates;
Tighter Bounds on the Expressivity of Transformer Encoders
David Chiang (University of Notre Dame), Anand Pillay (University of Notre Dame)
Transformer
🎯 What it does: A new first-order logic FOC_{≤,MOD} is proposed, and it is proven that the languages recognizable by fixed-precision Transformer encoders are exactly equivalent to the languages definable by this logic, thereby tightening the theoretical upper and lower bounds of the expressive power of Transformers.
Tighter Information-Theoretic Generalization Bounds from Supersamples
Ziqiao Wang (University of Ottawa), Yongyi Mao (University of Ottawa)
Contrastive LearningTabular
🎯 What it does: Under the Supersample setting, various tighter upper bounds on generalization error are proposed using information-theoretic methods such as loss difference, single loss, and Rademacher sequences.
Tighter Lower Bounds for Shuffling SGD: Random Permutations and Beyond
Jaeyoung Cha (KAIST), Chulhee Yun (KAIST)
Optimization
🎯 What it does: This paper provides a tighter lower bound analysis for random reshuffling in non-replacement stochastic gradient descent (SGD) and permutation-based SGD, proving that the weight-averaged iteration can match the known upper bound, thereby demonstrating that the GraB algorithm is optimal with respect to all factors (especially condition number κ, sample size n, and number of iterations K).
Tilted Sparse Additive Models
Yingjie Wang (China University of Petroleum), Dacheng Tao (University of Sydney)
ClassificationOptimizationTabular
🎯 What it does: Proposes the Tilted Sparse Additive Model (T-SpAM) and conducts experimental validation on tasks such as regression, classification, imbalanced classification, and multi-objective learning.
TIPS: Topologically Important Path Sampling for Anytime Neural Networks
Guihong Li (University of Texas at Austin), Radu Marculescu (University of Texas at Austin)
OptimizationConvolutional Neural NetworkImage
🎯 What it does: By modeling the training process of the Anytime Neural Network (AnytimeNN) as a Discrete Time Markov Chain (DTMC), two topological metrics, Topological Accumulated Score (TAS) and Topological Path Score (TPS), were designed. Subsequently, a Topologically Important Path Sampling (TIPS) training strategy was proposed, which automatically searches for Pareto-optimal sub-networks under different hardware budgets.
Topological Point Cloud Clustering
Vincent Peter Grande, Michael T Schaub
Point Cloud
🎯 What it does: A point cloud topological clustering method TPCC based on Hodge-Laplace zero features is proposed to identify clusters corresponding to different topological features in point clouds.
Topological Singularity Detection at Multiple Scales
Julius Von Rohrscheidt, Bastian Rieck (Helmholtz Munich)
Anomaly DetectionImagePoint CloudBiomedical Data
🎯 What it does: An unsupervised framework TARDIS is proposed, which uses multi-scale persistent local homology (PLH) to simultaneously estimate the local intrinsic dimension (PID) and Euclidity of points, thereby detecting singular points in point clouds and assessing the 'manifoldness' of the data;
Topologically Faithful Image Segmentation via Induced Matching of Persistence Barcodes
Nico Daniel Stucki, Ulrich Bauer (Technical University of Munich)
SegmentationImage
🎯 What it does: A topology-preserving image segmentation method based on induced matching is proposed, which can accurately match the persistent homology barcodes of predictions and ground truth spatially, thereby defining a differentiable Betti matching error as the segmentation loss;
Total Variation Graph Neural Networks
Jonas Berg Hansen (UiT Arctic University of Norway), Filippo Maria Bianchi (UiT Arctic University of Norway)
ClassificationGraph Neural NetworkGraph
🎯 What it does: This paper proposes a graph neural network (TVGNN) that utilizes graph total variation (GTV) as an unsupervised loss for vertex clustering and graph classification.
Toward Efficient Gradient-Based Value Estimation
Arsalan Sharifnassab (University of Alberta), Richard S. Sutton
OptimizationReinforcement LearningTabular
🎯 What it does: A new gradient-based value estimation algorithm RANS is proposed, combining the Gauss-Newton approach and 'outlier-splitting' technology, aimed at accelerating and stabilizing gradient methods based on MSBE.
Toward Large Kernel Models
Amirhesam Abedsoltan (University of California San Diego), Parthe Pandit (University of California San Diego)
ImageAudio
🎯 What it does: This paper proposes a new EigenPro 3.0 algorithm for training a general kernel model that can independently scale model size and training sample size.
Towards a better understanding of representation dynamics under TD-learning
Yunhao Tang (Google DeepMind), Remi Munos (Google DeepMind)
Representation LearningReinforcement LearningTabularSequentialOrdinary Differential Equation
🎯 What it does: This paper provides a theoretical analysis of the dynamics of end-to-end updates in TD learning and proposes that using random reward prediction as an auxiliary task can effectively learn useful representations.
Towards a Persistence Diagram that is Robust to Noise and Varied Densities
Hang Zhang (Nanjing University), Ye Zhu (Deakin University)
ClassificationRecognitionAnomaly DetectionBiomedical Data
🎯 What it does: A new data-dependent kernel (Lambda-kernel) and corresponding filtering function (Lambda-filter) are proposed to construct persistent diagrams (PD) that are robust to both noise and uneven density.
Towards Better Graph Representation Learning with Parameterized Decomposition & Filtering
Mingqi Yang (Dalian University of Technology), Bryan Hooi (National University of Singapore)
Representation LearningDrug DiscoveryGraph Neural NetworkGraph
🎯 What it does: A Parameterized-(DF) framework is proposed, unifying and extending existing GNN models, achieving more flexible graph representations through learning the decomposition (D) and filtering (F) of graph matrices.
Towards Bridging the Gaps between the Right to Explanation and the Right to be Forgotten
Satyapriya Krishna (Harvard University), Himabindu Lakkaraju (Harvard University)
OptimizationSafty and PrivacyExplainability and InterpretabilityConvolutional Neural NetworkTabular
🎯 What it does: An algorithmic framework named ROCERF is proposed for generating still valid interpretability (i.e., adversarial) counterfactual explanations when model updates occur due to the execution of the 'right to be forgotten'.
Towards Coherent Image Inpainting Using Denoising Diffusion Implicit Models
Guanhua Zhang (University of California Santa Barbara), Shiyu Chang (University of California Santa Barbara)
RestorationDiffusion modelImage
🎯 What it does: A diffusion model-based image inpainting method called COPAINT is proposed, achieving globally consistent inpainting while maintaining constraints from reference images.
Towards Constituting Mathematical Structures for Learning to Optimize
Jialin Liu (Alibaba Group), HanQin Cai (University of Central Florida)
OptimizationRecurrent Neural NetworkImageTabular
🎯 What it does: A structured learning optimizer based on mathematical conditions is proposed, with the constraint update rule in the form of a preconditioner + bias.
Towards Controlled Data Augmentations for Active Learning
Jianan Yang (Zhejiang University), Junbo Zhao
ClassificationData-Centric LearningImage
🎯 What it does: A proactive learning framework named CAMPAL is proposed, which can enhance the sample selection effectiveness of active learning through a controllable data augmentation strategy.
Towards credible visual model interpretation with path attribution
NAVEED AKHTAR, Mohammad A. A. K. Jalwana (University of Western Australia)
Explainability and InterpretabilityConvolutional Neural NetworkTransformerImage
🎯 What it does: An improved path attribution method is proposed to address issues such as counterintuitive attribution, baseline uncertainty, and path feature ambiguity that arise when traditional path attribution is used in deep visual models, making the explanation results more credible and interpretable.
Towards Deep Attention in Graph Neural Networks: Problems and Remedies
Soo Yong Lee (Korea Advanced Institute of Science and Technology), Kijung Shin (Korea Advanced Institute of Science and Technology)
Explainability and InterpretabilityComputational EfficiencyGraph Neural NetworkGraph
🎯 What it does: This paper addresses the interpretability and performance bottlenecks of deep graph attention networks by proposing a new AERO-GNN architecture.
Towards Explaining Distribution Shifts
Sean Kulinski (Purdue University), David I. Inouye (Purdue University)
Domain AdaptationExplainability and InterpretabilityGenerative Adversarial NetworkImageTextTabular
🎯 What it does: A distribution shift explanation framework based on interpretable transport maps is proposed, providing k-sparse, k-clustering, and interpretable mappings for image adversarial generation.
Towards Learning Geometric Eigen-Lengths Crucial for Fitting Tasks
Yijia Weng (Stanford University), Leonidas Guibas (Stanford University)
Robotic IntelligencePoint Cloud
🎯 What it does: This paper proposes a method to determine whether an object can complete different assembly/placement tasks by learning low-dimensional geometric Eigen-lengths, without the need for explicit geometric modeling.