ICML 2024 Papers — Page 25
International Conference on Machine Learning · 2610 papers
Toward Adaptive Reasoning in Large Language Models with Thought Rollback
Sijia Chen (University of Toronto), Baochun Li (University of Toronto)
TransformerLarge Language ModelPrompt EngineeringTextChain-of-Thought
🎯 What it does: The Thought Rollback (TR) framework is proposed, allowing LLMs to adaptively backtrack and correct their reasoning processes, significantly improving the quality of multi-step reasoning.
Toward Availability Attacks in 3D Point Clouds
Yifan Zhu (Academy of Mathematics and Systems Science, Chinese Academy of Sciences), Xiao-Shan Gao (Academy of Mathematics and Systems Science, Chinese Academy of Sciences)
OptimizationAdversarial AttackPoint Cloud
🎯 What it does: A poisoning method based on Feature Collision Error Minimization (FC-EM) is proposed for the availability attack on 3D point clouds.
Towards a Better Theoretical Understanding of Independent Subnetwork Training
Egor Shulgin (King Abdullah University of Science and Technology), Peter Richtárik (King Abdullah University of Science and Technology)
OptimizationFederated LearningComputational EfficiencyConvolutional Neural NetworkImage
🎯 What it does: In the context of distributed and federated learning, the theoretical convergence properties of the Independent Subnetwork Training (IST) method are proposed and analyzed in depth, particularly for quadratic objective functions and considering model heterogeneity.
Towards a Self-contained Data-driven Global Weather Forecasting Framework
Yi Xiao (Tsinghua University), Wanli Ouyang (Shanghai Artificial Intelligence Laboratory)
OptimizationAuto EncoderTime Series
🎯 What it does: An AI-embedded 4D variational data assimilation algorithm (AI-embedded 4DVar) is proposed and implemented, which is coupled with the global AI weather forecasting model FengWu to construct a self-consistent full-scale weather forecasting framework.
Towards an Understanding of Stepwise Inference in Transformers: A Synthetic Graph Navigation Model
Mikail Khona (Massachusetts Institute of Technology), Hidenori Tanaka (Harvard University)
TransformerGraph
🎯 What it does: This study investigates the stepwise inference mechanism of Transformers in synthetic graph navigation tasks.
Towards AutoAI: Optimizing a Machine Learning System with Black-box and Differentiable Components
Zhiliang Chen (National University of Singapore), Bryan Kian Hsiang Low (National University of Singapore)
OptimizationTabularSequential
🎯 What it does: This paper proposes the AutoAI framework and designs the A-BAD-BO algorithm, which combines Bayesian optimization with local loss information from differentiable components to jointly optimize complex machine learning systems containing black-box modules.
Towards Causal Foundation Model: on Duality between Optimal Balancing and Attention
Jiaqi Zhang (Massachusetts Institute of Technology), Chao Ma (Microsoft Research)
TransformerTabular
🎯 What it does: The Causal Inference with Attention (CInA) method is proposed, which utilizes self-attention and self-supervised learning to learn covariate balancing weights on multiple unlabeled datasets, thereby achieving zero-shot causal effect inference on new data.
Towards Certified Unlearning for Deep Neural Networks
Binchi Zhang (University of Virginia), Jundong Li (University of Virginia)
OptimizationSafty and PrivacyComputational EfficiencyImage
🎯 What it does: This paper proposes a complete framework for provably 'forgetting' data in deep neural networks. It first uses a modified Newton update to estimate the model after removing samples, then adds noise to meet the differential privacy error bound, ensuring that the distribution difference with the retrained model is controllable.
Towards Compositionality in Concept Learning
Adam Stein (University of Pennsylvania), Eric Wong (University of Pennsylvania)
OptimizationRepresentation LearningMultimodality
🎯 What it does: Automatically extract composable concept representations in an unsupervised environment and achieve concept composability through a new algorithm CCE (Compositional Concept Extraction).
Towards efficient deep spiking neural networks construction with spiking activity based pruning
Yaxin Li (Dalian University of Technology), Gang Pan (Zhejiang University)
Spiking Neural NetworkImage
🎯 What it does: A structured pruning framework based on synaptic activity (SCA) is proposed, which dynamically reduces and regenerates convolutional kernels during the training process, automatically learning lightweight SNN structures.
Towards Efficient Exact Optimization of Language Model Alignment
Haozhe Ji (Tsinghua University), Minlie Huang (Tsinghua University)
OptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: This paper proposes a precise optimization algorithm named EXO, which treats the alignment problem of language models as a probability matching issue. It directly optimizes the RLHF objective by minimizing the reverse KL divergence, avoiding the use of traditional RL during the alignment process.
Towards Efficient Spiking Transformer: a Token Sparsification Framework for Training and Inference Acceleration
Zhengyang Zhuge (Chinese Academy of Sciences), Jian Cheng (Chinese Academy of Sciences)
Computational EfficiencySpiking Neural NetworkTransformerImage
🎯 What it does: This paper proposes a Token Sparsification framework named STATA for efficient training and inference of Spiking Transformers.
Towards Efficient Training and Evaluation of Robust Models against $l_0$ Bounded Adversarial Perturbations
Xuyang Zhong (City University of Hong Kong), Chen Liu (City University of Hong Kong)
OptimizationAdversarial AttackImage
🎯 What it does: This paper proposes a white-box attack method based on sparse-PGD, which can efficiently generate l0-constrained sparse adversarial perturbations and combines with black-box attacks to form sAA for comprehensive evaluation; at the same time, this attack is used for adversarial training, resulting in a model with the strongest robustness against sparse attacks.
Towards General Algorithm Discovery for Combinatorial Optimization: Learning Symbolic Branching Policy from Bipartite Graph
Yufei Kuang (University of Science and Technology of China), Feng Wu (University of Science and Technology of China)
OptimizationTransformerReinforcement LearningGraph
🎯 What it does: This paper proposes a graph-based symbolic discovery framework GS4CO, which is used to directly learn interpretable branching variable selection strategies from the bipartite representation of mixed-integer linear programming.
Towards General Neural Surrogate Solvers with Specialized Neural Accelerators
Chenkai Mao (Stanford), Jonathan Fan
Neural Radiance FieldPhysics Related
🎯 What it does: A method called SNAP-DDM is proposed, which combines neural operators with domain decomposition to solve PDEs on subdomains using specialized neural operators, achieving efficient simulation for arbitrary domain sizes and boundary conditions.
Towards Generalization beyond Pointwise Learning: A Unified Information-theoretic Perspective
Yuxin Dong (Xi'an Jiaotong University), Chen Li (Xi'an Jiaotong University)
ClassificationRepresentation LearningConvolutional Neural NetworkTransformerContrastive LearningImage
🎯 What it does: An information-theoretic generalization error upper bound for non-pointwise learning (such as contrastive, triplet, quadruplet, and other multi-sample losses) is proposed, unifying the coverage from pointwise to higher-order learning paradigms.
Towards Global Optimality for Practical Average Reward Reinforcement Learning without Mixing Time Oracles
Bhrij Patel (University of Maryland), Amrit Bedi
OptimizationReinforcement LearningTabular
🎯 What it does: A policy gradient method is proposed for the average reward Markov decision process that does not require prior information on mixing time, and it is proven to converge to the global optimal solution.
Towards Interpretable Deep Local Learning with Successive Gradient Reconciliation
Yibo Yang (King Abdullah University of Science and Technology), Bernard Ghanem (King Abdullah University of Science and Technology)
OptimizationExplainability and InterpretabilityComputational EfficiencyConvolutional Neural NetworkTransformerImage
🎯 What it does: A new local learning strategy called Successive Gradient Reconciliation (SGR) is proposed, which aligns the gradients of neighboring layers at each layer to better align local error updates with global objectives, thereby improving the convergence and effectiveness of non-greedy hierarchical learning.
Towards Modular LLMs by Building and Reusing a Library of LoRAs
Oleksiy Ostapenko (Microsoft Research), Alessandro Sordoni (Microsoft Research)
TransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This paper constructs and reuses a LoRA-based adapter library to achieve modularity in large language models (LLMs), enabling zero-shot inference or few-shot supervised fine-tuning on unseen tasks.
Towards Neural Architecture Search through Hierarchical Generative Modeling
Lichuan Xiang (University of Warwick), Hongkai Wen (University of Warwick)
Neural Architecture SearchGraph Neural NetworkTransformerFlow-based ModelGenerative Adversarial NetworkImage
🎯 What it does: A hierarchical generative model method is proposed to efficiently explore a vast search space for NAS tasks using interpretable microstructure clustering and a macrostructure generator.
Towards Optimal Adversarial Robust Q-learning with Bellman Infinity-error
Haoran Li (University of Chinese Academy of Sciences), Shichen Liao (University of Chinese Academy of Sciences)
OptimizationReinforcement LearningVideo
🎯 What it does: This study investigates the existence of optimal robust policies in the State Adversarial Markov Decision Process (SA-MDP) and proves that the Bellman optimal policy is indeed the optimal robust policy. It then proposes the Consistent Adversarial Robust DQN (CAR-DQN) algorithm, which utilizes the Bellman ∞-error.
Towards Realistic Model Selection for Semi-supervised Learning
Muyang Li (University of Sydney), Tongliang Liu (University of Sydney)
ClassificationOptimizationImage
🎯 What it does: This paper proposes a model selection method for semi-supervised learning called SLAM, which can evaluate the generalization performance of models using training data without the need for a validation set.
Towards Resource-friendly, Extensible and Stable Incomplete Multi-view Clustering
Shengju Yu (National University of Defense Technology), Xinwang Liu (National University of Defense Technology)
OptimizationComputational EfficiencyTabular
🎯 What it does: A non-complete multi-view clustering algorithm named ToRES is proposed, which directly learns discrete clustering labels using prototype-sample similarity, avoiding the high cost and hyperparameter dependence of self-expressive similarity matrices.
Towards Robust Model-Based Reinforcement Learning Against Adversarial Corruption
Chenlu Ye (Hong Kong University of Science and Technology), Tong Zhang (University of Illinois Urbana-Champaign)
Adversarial AttackReinforcement Learning
🎯 What it does: In model-based reinforcement learning, a new robust algorithm has been designed to address state transition probabilities controlled by adversaries;
Towards Scalable and Versatile Weight Space Learning
Konstantin Schürholt, Damian Borth
GenerationRepresentation LearningConvolutional Neural NetworkTransformerAuto EncoderContrastive LearningImage
🎯 What it does: The SANE (Sequential Autoencoder for Neural Embeddings) model is proposed, which learns a low-dimensional representation of the weight space that can be cross-task and scalable to large models, and supports predicting model quality and generating new weights.
Towards the Theory of Unsupervised Federated Learning: Non-asymptotic Analysis of Federated EM Algorithms
Ye Tian (Columbia University), Yang Feng (New York University)
Federated LearningTabular
🎯 What it does: A federated gradient EM (FedGrEM) algorithm for mixed models is proposed for unsupervised federated learning, capable of handling task heterogeneity and a small number of contaminated tasks.
Towards Theoretical Understanding of Learning Large-scale Dependent Data via Random Features
Chao Wang (Shanghai University of Finance and Economics), Caixing Wang (Shanghai University of Finance and Economics)
Time SeriesSequential
🎯 What it does: This paper studies the theoretical properties of kernel ridge regression with random features (KRR-RF) under large dependent data (τ-mixing processes), proving that optimal learning rates can be achieved under exponential decay conditions, while only suboptimal results can be obtained under polynomial decay.
Towards Theoretical Understandings of Self-Consuming Generative Models
Shi Fu (University of Science and Technology of China), Dacheng Tao (Nanyang Technological University)
GenerationData SynthesisDiffusion model
🎯 What it does: This paper constructs a theoretical framework for self-consumption generative model training and provides an upper bound on the total variation distance between the generated distribution of subsequent models and the original true distribution under various mixed data sampling strategies.
Towards Understanding Inductive Bias in Transformers: A View From Infinity
Itay Lavie (Hebrew University of Jerusalem), Zohar Ringel (Hebrew University of Jerusalem)
TransformerText
🎯 What it does: This paper studies the inductive bias of Transformers under the infinite-width Gaussian process limit, utilizing representation theory of symmetric groups to analyze their expressible functions and learnability.
Towards Understanding the Word Sensitivity of Attention Layers: A Study via Random Features
Simone Bombari (Institute of Science and Technology), Marco Mondelli
TransformerLarge Language ModelText
🎯 What it does: This paper uses random features as a model to theoretically analyze and quantify the sensitivity of the attention layer to word changes.
Towards Unified Multi-granularity Text Detection with Interactive Attention
Xingyu Wan (Baidu), Jingdong Wang (Baidu)
Object DetectionSegmentationRepresentation LearningTransformerPrompt EngineeringImage
🎯 What it does: A unified multi-granularity text detection framework DAT is proposed, capable of simultaneously performing scene text detection, document layout analysis, and page detection.
Trainable Transformer in Transformer
Abhishek Panigrahi (Princeton University), Sanjeev Arora (Princeton University)
TransformerSupervised Fine-TuningText
🎯 What it does: TINT is proposed, a construction that can simulate and fine-tune the internal large Transformer model in a single forward pass.
Trained Random Forests Completely Reveal your Dataset
Julien Ferry (Polytechnique Montreal), Thibaut Vidal (Polytechnique Montreal)
OptimizationAdversarial AttackData-Centric LearningTabular
🎯 What it does: This paper proposes a maximum likelihood dataset reconstruction attack based on constraint programming, which can reconstruct the training set using only the structure and node count information of the random forest.
Training Greedy Policy for Proposal Batch Selection in Expensive Multi-Objective Combinatorial Optimization
Deokjae Lee (Seoul National University), Kyunghyun Cho (New York University)
OptimizationReinforcement LearningSequential
🎯 What it does: In the expensive multi-objective combinatorial optimization (MOCO) problem, a greedy strategy was trained to directly perform subset selection on a batch of candidate sets within the combinatorial space, thereby optimizing the batch acquisition function.
Training Large Language Models for Reasoning through Reverse Curriculum Reinforcement Learning
Zhiheng Xi (Fudan University), Xuanjing Huang (Fudan University)
Large Language ModelReinforcement LearningText
🎯 What it does: This paper proposes a training method R³ based on inverse curriculum reinforcement learning, which achieves stepwise supervision in reasoning tasks using only the final outcome for supervision.
Training-Free Long-Context Scaling of Large Language Models
Chenxin An (University of Hong Kong), Lingpeng Kong
TransformerLarge Language ModelText
🎯 What it does: This paper proposes a Dual Chunk Attention (DCA) mechanism that does not require further training, achieving efficient inference of LLaMA2 70B with context lengths exceeding 100k.
Transferable Facial Privacy Protection against Blind Face Restoration via Domain-Consistent Adversarial Obfuscation
Kui Zhang (University of Science and Technology of China), Nenghai Yu (University of Science and Technology of China)
RestorationSafty and PrivacyAdversarial AttackGenerative Adversarial NetworkImage
🎯 What it does: A transferable facial privacy protection method based on domain-consistent adversarial perturbations (DomCo) is proposed, which can maintain the anonymity of facial images under blind face restoration model attacks.
Transferring Knowledge From Large Foundation Models to Small Downstream Models
Shikai Qiu (New York University), Andrew Gordon Wilson (New York University)
Computational EfficiencyKnowledge DistillationRepresentation LearningContrastive LearningImageTextMultimodality
🎯 What it does: This paper proposes Adaptive Feature Transfer (AFT), which transfers task-related knowledge from large foundational models to smaller, lower-cost models by adaptively regularizing pre-trained features in a small downstream model.
Transformers are SSMs: Generalized Models and Efficient Algorithms Through Structured State Space Duality
Tri Dao (Princeton University), Albert Gu (Carnegie Mellon University)
TransformerLarge Language ModelText
🎯 What it does: A structured state space duality (SSD) framework is proposed, linking structured state space models (SSM) with attention mechanisms through semi-separable matrices, and based on this, a more efficient SSD algorithm and Mamba-2 architecture are designed.
Transformers Get Stable: An End-to-End Signal Propagation Theory for Language Models
Akhil Kedia (Samsung Research), Haejun Lee (Samsung Research)
TransformerLarge Language ModelImageTextAudio
🎯 What it does: This paper proposes a unified signal propagation theory and provides closed-form expressions for the first and second moments during the forward and backward propagation of each layer in the Transformer, aimed at explaining and alleviating issues such as gradient vanishing/explosion, rank collapse, and instability of attention scores in deep Transformers.
Transformers Implement Functional Gradient Descent to Learn Non-Linear Functions In Context
Xiang Cheng (Massachusetts Institute of Technology), Suvrit Sra (Technical University of Munich)
OptimizationTransformerTabular
🎯 What it does: This study investigates how the Transformer achieves gradient descent in function space under nonlinear activation and uses this mechanism for contextual learning of nonlinear functions.
Transformers Learn Nonlinear Features In Context: Nonconvex Mean-field Dynamics on the Attention Landscape
Juno Kim (University of Tokyo), Taiji Suzuki (University of Tokyo)
OptimizationTransformerTabular
🎯 What it does: This study investigates the mean field optimization of the MLP + linear attention structure in Transformers, proving that it can learn nonlinear features in in-context feature learning tasks and converge to a global optimum.
Transformers Provably Learn Sparse Token Selection While Fully-Connected Nets Cannot
Zixuan Wang (Princeton University), Jason D. Lee (Columbia University)
Transformer
🎯 What it does: This paper studies the learning capability of transformers in the sparse token selection task (STS q) and demonstrates that transformers outperform fully connected networks (FCN) in this task.
Transformers, parallel computation, and logarithmic depth
Clayton Sanford (Columbia University), Matus Telgarsky (New York University)
TransformerSequential
🎯 What it does: Proved the bidirectional simulation relationship between Transformer and Massively Parallel Computation (MPC) models, and utilized this relationship to demonstrate computational tasks that can be solved at logarithmic depth (such as graph connectivity and k-hop induction heads), thereby revealing that parallel computation is the core advantage of Transformers;
Transforming and Combining Rewards for Aligning Large Language Models
Zihao Wang (University of Chicago), Victor Veitch (Google DeepMind)
OptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: Proposed a log-sigmoid centering transformation (LSC-transform) for the reward model to improve the alignment effect of RLHF.
Transitional Uncertainty with Layered Intermediate Predictions
Ryan Benkert (Georgia Institute of Technology), Ghassan AlRegib (Georgia Institute of Technology)
ClassificationAnomaly DetectionConvolutional Neural NetworkImageBiomedical DataComputed Tomography
🎯 What it does: This paper proposes a new method for single-channel uncertainty estimation called TULIP. This method inserts internal classifiers (Shallow-Deep networks) into the intermediate layers of the network, utilizing the features of these intermediate outputs to complete classification and uncertainty estimation in a single forward inference.
Translating Subgraphs to Nodes Makes Simple GNNs Strong and Efficient for Subgraph Representation Learning
Dongkwan Kim (KAIST), Alice Oh (KAIST)
Computational EfficiencyRepresentation LearningGraph Neural NetworkGraphBenchmark
🎯 What it does: Map subgraphs to nodes, construct a coarse-grained graph, and transform subgraph-level tasks into node-level tasks;
Translation Equivariant Transformer Neural Processes
Matthew Ashman (University of Cambridge), Richard E. Turner (Microsoft Research AI for Science)
TransformerImageTime Series
🎯 What it does: Proposed the Translation Equivariant Transformer Neural Process (TE-TNP, TE-PT-TNP), achieving unbiased predictions for spatial and temporal translation-invariant data in a multi-task setting.
Transolver: A Fast Transformer Solver for PDEs on General Geometries
Haixu Wu (Tsinghua University), Mingsheng Long (Tsinghua University)
TransformerMeshBenchmarkPhysics Related
🎯 What it does: This paper presents Transolver, a Transformer structure utilizing Physics-Attention, capable of quickly solving partial differential equations on complex geometric meshes.
Transport of Algebraic Structure to Latent Embeddings
Samuel Pfrommer (University of California), Somayeh Sojoudi (University of California)
Data SynthesisRepresentation LearningFlow-based ModelAuto EncoderPoint Cloud
🎯 What it does: Design and validate an operational method that maintains the source algebraic structure in potential embedding spaces, with experiments specifically targeting set algebra.
TravelPlanner: A Benchmark for Real-World Planning with Language Agents
Jian Xie (Fudan University), Yu Su (Ohio State University)
TransformerLarge Language ModelAgentic AITextBenchmarkChain-of-Thought
🎯 What it does: Proposes the TravelPlanner benchmark to evaluate the capabilities of language model-driven agents in real-world multi-constraint travel planning tasks.
Triadic-OCD: Asynchronous Online Change Detection with Provable Robustness, Optimality, and Convergence
Yancheng Huang (Tongji University), Leian Chen (Columbia University)
Anomaly DetectionOptimizationTabularTime Series
🎯 What it does: An asynchronous distributed online change detection framework, Triadic-OCD, is proposed, which can timely identify change points in data streams under uncertain system parameters.
Triple Changes Estimator for Targeted Policies
Sina Akbari (École Polytechnique Fédérale de Lausanne), Negar Kiyavash (École Polytechnique Fédérale de Lausanne)
TabularElectronic Health Records
🎯 What it does: This paper proposes and derives a Triple Changes estimator for identifying potential outcome distributions and their average treatment effects in policy evaluation targeted at subpopulations.
Triplet Interaction Improves Graph Transformers: Accurate Molecular Graph Learning with Triplet Graph Transformers
Md Shamim Hussain (Rensselaer Polytechnic Institute), Dharmashankar Subramanian (IBM)
Drug DiscoveryGraph Neural NetworkTransformerGraph
🎯 What it does: A Triplet Graph Transformer (TGT) is designed, which directly implements third-order information flow between triplet nodes through a triplet interaction mechanism, and constructs a three-stage training process for distance predictor and attribute predictor, supporting random inference and uncertainty estimation.
Tripod: Three Complementary Inductive Biases for Disentangled Representation Learning
Kyle Hsu (Stanford University), Jiajun Wu (Stanford University)
Representation LearningAuto EncoderImage
🎯 What it does: The Tripod model is proposed, which integrates three complementary inductive biases to achieve unsupervised decomposable representation learning.
TroVE: Inducing Verifiable and Efficient Toolboxes for Solving Programmatic Tasks
Zhiruo Wang, Daniel Fried (Carnegie Mellon University)
OptimizationAI Code AssistantTransformerLarge Language ModelPrompt EngineeringTabular
🎯 What it does: Dynamically generate, maintain, and prune a reusable high-level function toolbox through LLM under unsupervised and untrained conditions, in order to produce more concise and accurate programmatic answers.
Truly No-Regret Learning in Constrained MDPs
Adrian Müller (École Polytechnique Fédérale de Lausanne), Niao He (ETH Zurich)
OptimizationReinforcement Learning
🎯 What it does: In the unknown finite-horizon constrained Markov decision process, a regularized frontier-successor algorithm is designed and analyzed, achieving true sublinear convergence without strong regret.
Trust Regions for Explanations via Black-Box Probabilistic Certification
Amit Dhurandhar (IBM Research), Karthikeyan Natesan Ramamurthy (IBM Research)
OptimizationExplainability and InterpretabilityImageTabular
🎯 What it does: This paper proposes the problem of verifying the credibility of explanations within a black-box model through limited queries, and presents algorithms for three incremental query strategies.
Trust the Model Where It Trusts Itself - Model-Based Actor-Critic with Uncertainty-Aware Rollout Adaption
Bernd Frauenknecht (RWTH Aachen University), Sebastian Trimpe (RWTH Aachen University)
Robotic IntelligenceReinforcement LearningSequential
🎯 What it does: A Dyna-style model-based reinforcement learning algorithm called MACURA is proposed, which is based on model uncertainty adaptive episode length.
Trustless Audits without Revealing Data or Models
Suppakit Waiwitlikhit (Stanford University), Daniel Kang (University of Illinois Urbana-Champaign)
Recommendation SystemConvolutional Neural NetworkImage
🎯 What it does: The ZKAUDIT protocol is proposed, which enables trustworthy auditing of deep learning models and training processes without disclosing model weights or training data.
Trustworthy Actionable Perturbations
Jesse Friedbaum (University of Arizona), Ravi Tandon (University of Arizona)
OptimizationExplainability and InterpretabilityAdversarial AttackContrastive LearningTabularFinance Related
🎯 What it does: A new framework for trustworthy actionable perturbations (TAP) is proposed, which provides individual action recommendations with minimal cost and maximum benefit while ensuring changes in true class probabilities without being misclassified as adversarial examples.
Trustworthy Alignment of Retrieval-Augmented Large Language Models via Reinforcement Learning
Zongmeng Zhang (University of Science and Technology of China), Houqiang Li
RetrievalTransformerLarge Language ModelReinforcement LearningTextRetrieval-Augmented Generation
🎯 What it does: In retrieval-augmented large language models, the model's reliance on external retrieval evidence is enhanced to a 'trustworthy' state through reinforcement learning, meaning that responses are based solely on retrieved context rather than internal parameter knowledge.
TSLANet: Rethinking Transformers for Time Series Representation Learning
Emadeldeen Eldele (Agency for Science Technology and Research), Xiaoli Li (Agency for Science Technology and Research)
ClassificationAnomaly DetectionRepresentation LearningConvolutional Neural NetworkTransformerTime SeriesSequentialBiomedical Data
🎯 What it does: This paper proposes TSLANet, a lightweight convolutional network capable of simultaneously performing time series classification, prediction, and anomaly detection.
Tuning-free Estimation and Inference of Cumulative Distribution Function under Local Differential Privacy
Yi Liu (University of Alberta), Linglong Kong (University of Alberta)
Safty and Privacy
🎯 What it does: A parameter-free algorithm for estimating cumulative distribution functions under local differential privacy is proposed, which transforms LDP data into a current state problem using random response and employs constrained monotonic maximum likelihood regression to obtain estimators.
Tuning-Free Stochastic Optimization
Ahmed Khaled (Princeton University), Chi Jin (Princeton University)
Optimization
🎯 What it does: This paper studies the feasibility and limits of implementing tuning-free algorithms in stochastic optimization, proposes a definition of 'tuning-free', and provides upper bounds, lower bounds, and conditions under different function classes (smooth convex, Lipschitz convex, smooth non-convex) and domains (bounded, unbounded).
Turnstile $\ell_p$ leverage score sampling with applications
Alexander Munteanu (TU Dortmund University), Simon Omlor (TU Dortmund University)
OptimizationTabularBenchmark
🎯 What it does: An algorithm is proposed for sampling matrix rows according to the ℓp norm under the turnstile data stream model, providing approximate row vectors and sampling probabilities, and further generalizing it to ℓp leverage sampling, constructing an ε-coreset for various regression tasks.
TVE: Learning Meta-attribution for Transferable Vision Explainer
Guanchu Wang (Rice University), Xia Hu (Rice University)
Explainability and InterpretabilityMeta LearningTransformerImage
🎯 What it does: A transferable visual explainer (TVE) is proposed, which generates explanations across tasks and models through pre-training to learn meta-attribution.
Two Fists, One Heart: Multi-Objective Optimization Based Strategy Fusion for Long-tailed Learning
Zhe Zhao (University of Science and Technology of China), Yang Wang (University of Science and Technology of China)
OptimizationImage
🎯 What it does: This study investigates the trade-off between head and tail class performance in long-tail learning and proposes a multi-objective optimization-based strategy fusion framework (MOOSF) to enhance multi-class balance performance.
Two Heads are Actually Better than One: Towards Better Adversarial Robustness via Transduction and Rejection
Nils Palumbo (University of Wisconsin Madison), Somesh Jha (University of Wisconsin Madison)
Domain AdaptationAdversarial AttackConvolutional Neural NetworkGenerative Adversarial NetworkImage
🎯 What it does: A robust defense method TLDR that combines transduction and rejection is proposed;
Two Heads Are Better Than One: Boosting Graph Sparse Training via Semantic and Topological Awareness
Guibin Zhang (Tongji University), Tianlong Chen (Massachusetts Institute of Technology)
Graph Neural NetworkGraph
🎯 What it does: Proposes the Graph Sparse Training (GST) framework, which achieves efficient graph sparsification by dynamically sparsifying the graph structure during training, combining semantic and topological information.
Two Stones Hit One Bird: Bilevel Positional Encoding for Better Length Extrapolation
Zhenyu He (Peking University), Di He (Peking University)
TransformerText
🎯 What it does: This paper proposes Bilevel Positional Encoding (BiPE), a dual-layer position encoding that combines intra-segment encoding and inter-segment encoding to enhance the length extrapolation capability of Transformers on long sequences.
Two Tales of Single-Phase Contrastive Hebbian Learning
Rasmus Høier (Chalmers University of Technology), Christopher Zach (Chalmers University of Technology)
ClassificationOptimizationAdversarial AttackContrastive LearningImage
🎯 What it does: Proposes and improves the single-phase comparative Hebbian learning algorithm (Dual Propagation), providing its theoretical foundation and introducing a new variant DP⊤ that is robust to asymmetric shocks.
Two-sided Competing Matching Recommendation Markets With Quota and Complementary Preferences Constraints
Yuantong Li (University of California Los Angeles), Xiaowu Dai (University of California Los Angeles)
Recommendation SystemReinforcement Learning
🎯 What it does: This paper proposes a multi-agent multi-type matching algorithm MMTS based on Thompson Sampling, aimed at solving the bilateral matching recommendation problem with quota and complementary preference constraints.
Two-Stage Shadow Inclusion Estimation: An IV Approach for Causal Inference under Latent Confounding and Collider Bias
Baohong Li (Zhejiang University), Kun Kuang (Zhejiang University)
Reinforcement Learning
🎯 What it does: A two-stage shadow variable embedding (2SSI) method is proposed to simultaneously address potential confounding and collision bias in observational data; the first stage processes IV regression to obtain residuals and learn decomposed representations, while the second stage uses the residuals and representations as shadow variables for outcome regression.
Two-timescale Derivative Free Optimization for Performative Prediction with Markovian Data
Haitong LIU, Hoi To Wai
OptimizationTime SeriesSequential
🎯 What it does: The research addresses non-convex optimization problems under decision-dependent data distribution (performative prediction) and proposes a zero-gradient (DFO) method to find approximate stationary solutions.
UGrid: An Efficient-And-Rigorous Neural Multigrid Solver for Linear PDEs
Xi Han (Stony Brook University), Hong Qin (Stony Brook University)
Convolutional Neural NetworkMeshPhysics Related
🎯 What it does: A neural multigrid solver UGrid based on U-Net and multigrid fusion is proposed and implemented, providing rigorous convergence and correctness proofs for linear PDEs.
ULAREF: A Unified Label Refinement Framework for Learning with Inaccurate Supervision
Congyu Qiao (Southeast University), Xin Geng (Southeast University)
ClassificationSegmentationContrastive LearningImage
🎯 What it does: A unified label refinement framework ULAREF is proposed, which predicts reliability through global detection and performs local enhancement on unreliable samples, achieving a unified treatment of noisy labels and partial labels.
ULTRAFEEDBACK: Boosting Language Models with Scaled AI Feedback
Ganqu Cui (Tsinghua University), Maosong Sun (Tsinghua University)
Reinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningText
🎯 What it does: A large-scale AI feedback dataset called ULTRAFEEDBACK is proposed, and this dataset is used to align open-source large language models.
Unbiased Multi-Label Learning from Crowdsourced Annotations
Mingxuan Xia (Zhejiang University), Haobo Wang (Zhejiang University)
ClassificationAuto EncoderImage
🎯 What it does: Train a multi-label classifier using crowdsourced annotation data in the absence of real labels.
Uncertainty Estimation by Density Aware Evidential Deep Learning
Taeseong Yoon (KAIST), Heeyoung Kim (KAIST)
ClassificationAnomaly DetectionConvolutional Neural NetworkImage
🎯 What it does: The Density Aware Evidential Deep Learning (DAEDL) method is proposed, which improves the performance of traditional Evidential Deep Learning (EDL) in terms of OOD detection and classification accuracy.
Uncertainty for Active Learning on Graphs
Dominik Fuchsgruber (Technical University of Munich), Stephan Günnemann (Technical University of Munich)
ClassificationOptimizationGraph Neural NetworkGraph
🎯 What it does: This study evaluates the effectiveness of uncertainty sampling in active learning for graph node classification and proposes a true uncertainty measure based on generative models along with feasible approximation methods.
Uncertainty-Aware Reward-Free Exploration with General Function Approximation
Junkai Zhang (University of California), Quanquan Gu (University of California)
Reinforcement Learning
🎯 What it does: This paper proposes a reward-free exploration algorithm based on general function approximation, GFA-RFE, which utilizes uncertainty-aware intrinsic rewards and weighted regression to achieve efficient exploration.
Understanding Adam Optimizer via Online Learning of Updates: Adam is FTRL in Disguise
Kwangjun Ahn (Massachusetts Institute of Technology), Yan Dai (Tsinghua University)
OptimizationTabular
🎯 What it does: The paper maps the Adam optimizer to the discounted FTRL framework of online learning, explaining the roles of its momentum and exponential weighted average.
Understanding and Diagnosing Deep Reinforcement Learning
Ezgi Korkmaz (University College London)
Adversarial AttackReinforcement LearningVideo
🎯 What it does: A diagnostic method based on non-Lipschitz directions (RA-NLD) is proposed, which systematically reveals the distribution of unstable directions and non-robust features in space-time by performing principal component analysis on the gradient information of deep reinforcement learning policies.
Understanding Diffusion Models by Feynman's Path Integral
Yuji Hirono (Kyoto University), Kenji Fukushima (University of Tokyo)
GenerationData SynthesisDiffusion modelScore-based ModelTabularPhysics RelatedStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: Reformulate score-based diffusion models within the Feynman path integral framework, derive the reverse SDE and loss function, and introduce the interpolation parameter h, corresponding it to the Planck constant. Utilize WKB expansion for quantitative analysis of the negative log-likelihood of noisy sampling.
Understanding Finetuning for Factual Knowledge Extraction
Gaurav Rohit Ghosal (Carnegie Mellon University), Aditi Raghunathan (Carnegie Mellon University)
TransformerSupervised Fine-TuningText
🎯 What it does: This study investigates the impact of fine-tuning known and unknown facts in pre-trained knowledge on the factuality of downstream question answering.
Understanding Forgetting in Continual Learning with Linear Regression
Meng Ding (State University of New York at Buffalo), Jinhui Xu (King Abdullah University of Science and Technology)
🎯 What it does: Through theoretical analysis and experimental research, the upper and lower bounds of forgetting in multi-step SGD under linear regression in continual learning are provided, and the effects of task order, step size, data scale, and dimensionality on forgetting are verified.
Understanding Heterophily for Graph Neural Networks
Junfu Wang (Beihang University), Yunhong Wang (Beihang University)
ClassificationGraph Neural NetworkGraph
🎯 What it does: This paper models the structure of heterogeneous graphs by constructing Heterophilous Stochastic Block Models (HSBM), systematically analyzes the performance of Graph Convolutional Networks (GCN) in multi-class node classification tasks, and provides theoretical insights into the effects of heterogeneity, neighborhood inconsistency, and multi-layer convolution on separability.
Understanding Inter-Concept Relationships in Concept-Based Models
Naveen Janaki Raman (Carnegie Mellon University), Mateja Jamnik (University of Cambridge)
ClassificationExplainability and InterpretabilityImage
🎯 What it does: This paper studies whether Concept-Based Models can capture the relationships between concepts and proposes a new algorithm to enhance the accuracy of concept interventions by utilizing these relationships.
Understanding MLP-Mixer as a wide and sparse MLP
Tomohiro Hayase (Metaverse Lab, Cluster Inc.), Ryo Karakida (Artificial Intelligence Research Center, AIST)
Image
🎯 What it does: By reinterpreting the mixing layer of MLP-Mixer as an extremely wide and sparse MLP through vectorization and the Kronecker product, the implicit L1 regularization characteristics are revealed and linked to the Monarch matrix. Experimental validation shows its similarity to sparse weight MLP in the width-sparsity relationship, and an extensible random permutation mixer (RP-Mixer) and PK family are proposed to achieve larger width architectures.
Understanding Reasoning Ability of Language Models From the Perspective of Reasoning Paths Aggregation
Xinyi Wang (University of California), William Yang Wang (University of California)
TransformerLarge Language ModelSupervised Fine-TuningTextGraphChain-of-Thought
🎯 What it does: This study proposes that language models can achieve reasoning capabilities by aggregating indirect reasoning paths, viewing pre-trained corpora as random walk paths on a knowledge/inference graph, and conducts experimental validation on two types of tasks: knowledge graph and chain reasoning.
Understanding Retrieval-Augmented Task Adaptation for Vision-Language Models
Yifei Ming (University of Wisconsin Madison), Yixuan Li (University of Wisconsin Madison)
ClassificationRetrievalDomain AdaptationTransformerVision Language ModelImageMultimodalityRetrieval-Augmented Generation
🎯 What it does: The study investigates the impact of retrieval-enhanced task adaptation methods on visual-language models in low-data scenarios, systematically evaluating the effects of retrieval methods and logical integration on performance.
Understanding Server-Assisted Federated Learning in the Presence of Incomplete Client Participation
Haibo Yang (Rochester Institute of Technology), Jia Liu (Ohio State University)
Federated LearningImage
🎯 What it does: Analyzed the traditional federated learning's non-PAC learnability under incomplete client participation conditions, proposed the Server-Assisted Federated Learning (SA-FL) framework, proved its ability to recover PAC learnability, and designed a new SAFARI algorithm, providing convergence rate proofs in non-convex and strongly convex scenarios.
Understanding Stochastic Natural Gradient Variational Inference
Kaiwen Wu (University of Pennsylvania), Jacob R. Gardner (University of Pennsylvania)
OptimizationTabular
🎯 What it does: This paper studies the non-asymptotic convergence rate of stochastic natural gradient variational inference (NGVI), proving that it can achieve O(1/T) under conjugate likelihood and exploring the non-conjugate case.
Understanding the Effects of Iterative Prompting on Truthfulness
Satyapriya Krishna (Harvard University), Himabindu Lakkaraju (Harvard University)
TransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: This study systematically evaluates the impact of iterative prompting on the truthfulness of large language models (LLMs) and designs two improved prompts (Improved Prompt-1 and Improved Prompt-2) to address the apology and error phenomena caused by the traditional 'Are you sure?' prompt, further enhancing the model's accuracy and calibration performance on TruthfulQA.
Understanding the Impact of Introducing Constraints at Inference Time on Generalization Error
Masaaki Nishino (NTT Corporation), Norihito Yasuda (NTT Corporation)
ClassificationOptimization
🎯 What it does: Analyzes the impact of adding constraints during inference on the relative generalization error of multi-class models, providing theoretical conditions under which the error is maintained.
Understanding the Learning Dynamics of Alignment with Human Feedback
Shawn Im (University of Wisconsin Madison), Yixuan Li (University of Wisconsin Madison)
OptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: The theoretical analysis of the learning dynamics of Direct Preference Optimization (DPO) on large language models is conducted, with experimental validation on models such as Llama-2 and Mistral.
Understanding the Training Speedup from Sampling with Approximate Losses
Rudrajit Das (University of Texas at Austin), sujay sanghavi
OptimizationComputational EfficiencyConvolutional Neural NetworkTransformerSupervised Fine-TuningImageText
🎯 What it does: The paper explores the use of approximate loss for sample selection to accelerate training through theoretical analysis and experiments, and proposes a lightweight sample screening method SIFT based on early exit.
Understanding Unimodal Bias in Multimodal Deep Linear Networks
Yedi Zhang (University College London), Andrew M Saxe
Multimodality
🎯 What it does: This study investigates the unimodal bias in multimodal deep linear networks, deriving the relationship between the duration of the unimodal phase and the depth of the fusion layer, data statistics, and initialization through the analysis of gradient descent dynamics.
UniAudio: Towards Universal Audio Generation with Large Language Models
Dongchao Yang (Chinese University of Hong Kong), Helen M. Meng
GenerationData SynthesisTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringMultimodalityAudio
🎯 What it does: A general audio generation model called UniAudio based on large language models has been constructed, capable of uniformly handling 11 audio generation tasks including TTS, music, acoustic editing, and voice conversion, and supports lightweight fine-tuning for new tasks.