ICML 2023 Papers — Page 9
International Conference on Machine Learning · 1828 papers
In Search for a Generalizable Method for Source Free Domain Adaptation
Malik Boudiaf (ETS Montreal), Eleni Triantafillou (Google Research)
Domain AdaptationKnowledge DistillationAudio
🎯 What it does: The study investigates the performance of source-free domain adaptation (SFDA) in bioacoustic tasks and proposes a new method.
In Search of Insights, Not Magic Bullets: Towards Demystification of the Model Selection Dilemma in Heterogeneous Treatment Effect Estimation
Alicia Curth (University of Cambridge), Mihaela van der Schaar (University of Cambridge)
Tabular
🎯 What it does: This paper conducts a systematic experimental study on the model selection problem of heterogeneous treatment effect estimation (CATE) in causal inference, comparing the advantages and disadvantages of three types of evaluation criteria: factual prediction, plug-in surrogate, and pseudo outcome surrogate.
IncDSI: Incrementally Updatable Document Retrieval
Varsha Kishore (Cornell University), Kilian Q Weinberger
RetrievalOptimizationTransformerLarge Language ModelText
🎯 What it does: A retrieval method called IncDSI is proposed, which allows for the real-time addition of new documents after model training.
Incentivizing Exploration with Linear Contexts and Combinatorial Actions
Mark Sellke (Amazon Core AI)
OptimizationReinforcement Learning
🎯 What it does: This paper studies how to achieve Bayesian Incentive Compatibility (BIC) exploration through Thompson sampling in linear contextual and combinatorial bandit problems.
Individually Fair Learning with One-Sided Feedback
Yahav Bechavod (Hebrew University and University of Pennsylvania), Aaron Roth (University of Pennsylvania)
OptimizationTabularFinance Related
🎯 What it does: This paper studies an online learning problem where the learner can only observe the true labels of positive prediction instances. The learner's goal is to maximize accuracy while achieving individual fairness. A novel auditing scheme is proposed that leverages feedback from dynamically selected multiple auditors to explore the trade-off between accuracy and fairness.
Inferring Relational Potentials in Interacting Systems
Armand Comas, Octavia Camps (Northeastern University)
Anomaly DetectionOptimizationGraph Neural NetworkReinforcement LearningGraphTime SeriesSequentialPhysics Related
🎯 What it does: An unsupervised method based on energy functions, NIIP, is proposed to infer and characterize relational potential energy in interactive systems, thereby enabling trajectory prediction and manipulation.
Infinite Action Contextual Bandits with Reusable Data Exhaust
Mark Rucker (University of Virginia), Paul Mineiro (Microsoft Research)
OptimizationReinforcement Learning from Human FeedbackTabular
🎯 What it does: The CappedIGW algorithm is proposed, achieving context robustness under smooth costs in the context of infinite actions, and can generate reusable data vapor.
Inflow, Outflow, and Reciprocity in Machine Learning
Mukund Sundararajan (Google), Walid Krichene (Google Research)
Recommendation SystemAnomaly DetectionTabularBiomedical Data
🎯 What it does: This paper proposes a framework to measure the individual data contribution (outflow) and benefit (inflow) in machine learning systems and assess their balance (reciprocity), providing theoretical proof and experimental validation.
InfoDiffusion: Representation Learning Using Information Maximizing Diffusion Models
Yingheng Wang (Cornell University), Volodymyr Kuleshov (Cornell University)
GenerationRepresentation LearningDiffusion modelAuto EncoderImage
🎯 What it does: This paper proposes InfoDiffusion, a framework for auxiliary variable representation learning based on diffusion models, which can simultaneously generate high-quality samples and learn a semantically interpretable low-dimensional latent space.
InfoOT: Information Maximizing Optimal Transport
Ching-Yao Chuang (Massachusetts Institute of Technology), David Alvarez-Melis (Microsoft Research)
Domain AdaptationOptimizationImagePoint Cloud
🎯 What it does: An optimal transport framework called InfoOT is proposed, which utilizes mutual information from kernel density estimation to introduce global structure;
Information-Theoretic State Space Model for Multi-View Reinforcement Learning
HyeongJoo Hwang (KAIST), Kee-Eung Kim (KAIST)
Robotic IntelligenceReinforcement LearningAuto EncoderImage
🎯 What it does: This paper proposes Fuse2Control (F2C), an information-theoretic multi-view reinforcement learning framework that uses Total Correlation and Conditional Variational Information Bottleneck (CVIB) to learn low-dimensional hidden states under missing views, and achieves linearly scalable observation fusion through inverse variance weighting.
Infusing Lattice Symmetry Priors in Attention Mechanisms for Sample-Efficient Abstract Geometric Reasoning
Mattia Atzeni (École Polytechnique Fédérale de Lausanne), Andreas Loukas (Prescient Design)
Convolutional Neural NetworkTransformerMixture of Experts
🎯 What it does: Proposes LATFORMER, which achieves differentiable learning of geometric transformations by incorporating a soft mask based on grid symmetry into the attention mechanism;
InGram: Inductive Knowledge Graph Embedding via Relation Graphs
Jaejun Lee (KAIST), Joyce Jiyoung Whang (KAIST)
Graph Neural NetworkGraph
🎯 What it does: A knowledge graph embedding method named INGRAM is proposed, which can simultaneously generate embedding vectors for new relations and new entities during inference, achieving complete inductive reasoning.
Input Perturbation Reduces Exposure Bias in Diffusion Models
Mang Ning (Utrecht University), Rita Cucchiara (University of Modena and Reggio Emilia)
GenerationData SynthesisComputational EfficiencyDiffusion modelImage
🎯 What it does: This work simulates the prediction error during inference by perturbing the input x_t during the training phase of DDPM, thereby reducing exposure bias and improving image generation quality as well as training/inference speed.
Input uncertainty propagation through trained neural networks
Paul Monchot (Ecole Polytechnique), Nicolas Fischer (National Laboratory of Metrology and Testing)
Anomaly DetectionComputational EfficiencyConvolutional Neural NetworkImageBiomedical Data
🎯 What it does: This paper proposes a Gaussian Mixture Model (GMM) propagation method based on Wasserstein distance (WGMprop) for propagating input uncertainty in a trained neural network and estimating the output probability density function.
Instant Soup: Cheap Pruning Ensembles in A Single Pass Can Draw Lottery Tickets from Large Models
AJAY KUMAR JAISWAL, Zhangyang Wang (University of Texas at Austin)
OptimizationComputational EfficiencyTransformerSupervised Fine-TuningImageText
🎯 What it does: This paper proposes two methods: Instant Soup Pruning (ISP) and Instant Model Soup (IMS). By generating multiple weak masks at once and performing denoising averaging, it quickly extracts sparse subnetworks from large-scale pre-trained models (CLIP ViT-B32 and BERT BASE) and validates their effectiveness on various visual and language tasks.
Instrumental Variable Estimation of Average Partial Causal Effects
Yuta Kawakami (Yokohama National University), Jin Tian (Iowa State University)
TabularFinance Related
🎯 What it does: This paper proposes two estimation methods for the average partial causal effect (APCE) of continuous treatment variables under the instrumental variable (IV) setting (non-parametric Picard iteration and parametric basis function regression), and proves their good properties and statistical/computational characteristics.
Integrating Prior Knowledge in Contrastive Learning with Kernel
Benoit Dufumier (NeuroSpin CEA Université Paris-Saclay), Pietro Gori (LTCI Télécom Paris)
ClassificationRepresentation LearningGenerative Adversarial NetworkContrastive LearningImageBiomedical Data
🎯 What it does: A decoupled uniformity loss is proposed, utilizing kernel methods to integrate prior knowledge (generative models or weak attributes) to enhance the quality of contrastive learning representations and eliminate the negative-positive coupling problem.
Interactive Object Placement with Reinforcement Learning
Shengping Zhang (Harbin Institute of Technology), Rongrong Ji (Xiamen University)
GenerationData SynthesisRecurrent Neural NetworkTransformerReinforcement LearningImage
🎯 What it does: An interactive object placement method called IOPRE is proposed, which utilizes reinforcement learning to achieve interactive and customizable image synthesis by gradually adjusting the position and size of foreground objects.
Internally Rewarded Reinforcement Learning
Mengdi Li (University of Hamburg), Stefan Wermter (University of Hamburg)
Reinforcement LearningImage
🎯 What it does: This paper studies the problem of Internal Reward Reinforcement Learning (IRRL) and proposes a clipped linear reward function to reduce reward noise and stabilize the training process.
Internet Explorer: Targeted Representation Learning on the Open Web
Alexander Cong Li, Deepak Pathak (Carnegie Mellon University)
RetrievalRepresentation LearningLarge Language ModelContrastive LearningImage
🎯 What it does: A framework for online self-supervised learning named Internet Explorer is proposed, which actively retrieves internet images using search engines and gradually trains models to improve the representation quality for specific small-scale target datasets.
Interpolation for Robust Learning: Data Augmentation on Wasserstein Geodesics
Jiacheng Zhu (Carnegie Mellon University), Ding Zhao (University of Illinois)
OptimizationAdversarial AttackConvolutional Neural NetworkAuto EncoderGenerative Adversarial NetworkImage
🎯 What it does: Proposes a worst-case Wasserstein barycenter based on Wasserstein geodesics for data augmentation and regularization, thereby enhancing the model's adversarial robustness.
Interpretable Neural-Symbolic Concept Reasoning
Pietro Barbiero (University of Cambridge), Giuseppe Marra (KU Leuven)
Explainability and InterpretabilityImageGraph
🎯 What it does: Proposes Deep Concept Reasoner (DCR), an interpretable neural-symbolic model based on concept embeddings, which achieves interpretability in task prediction by learning fuzzy logic rules and executing them on concept truth values;
Interval Bound Interpolation for Few-shot Learning with Few Tasks
Shounak Datta (Indian Statistical Institute), Swagatam Das (Indian Statistical Institute)
ClassificationMeta LearningImage
🎯 What it does: In few-shot learning with a small number of tasks, Interval Boundary Propagation (IBP) is used to maintain task neighborhoods, and artificial tasks are generated by interpolating between the tasks and their interval boundaries to enhance the model's generalization performance.
Interventional Causal Representation Learning
Kartik Ahuja (Meta AI), Yoshua Bengio (Mila-Quebec AI Institute)
Representation LearningAuto EncoderImage
🎯 What it does: The study utilizes intervention data to achieve causal representation learning, proposing the identification of latent factors through geometric features supported by latent variables (such as support independence).
Intrinsic Sliced Wasserstein Distances for Comparing Collections of Probability Distributions on Manifolds and Graphs
Raif M. Rustamov (Amazon), Subhabrata Majumdar (AI Risk and Vulnerability Alliance)
Graph
🎯 What it does: This paper proposes an intrinsic sliced Wasserstein distance (ISW₂) for comparing sets of probability distributions on manifolds and graphs, and based on this, designs an interpretable hypothesis testing method;
Invariance in Policy Optimisation and Partial Identifiability in Reward Learning
Joar Max Viktor Skalse (Oxford University), Adam Gleave (FAR AI, Inc.)
OptimizationReinforcement Learning from Human Feedback
🎯 What it does: This paper formalizes and provides a theorem-based description of the recognizability and uncertainty of various common data sources in reward learning (such as expert demonstrations, trajectory comparisons, etc.) and downstream tasks (such as policy optimization). It constructs a lattice structure based on 'invariance' to compare the relative information quantity and uncertainty tolerance between different data sources and applications.
Invariant Slot Attention: Object Discovery with Slot-Centric Reference Frames
Ondrej Biza (Northeastern University), Thomas Kipf (Google Research)
Object DetectionSegmentationAutonomous DrivingConvolutional Neural NetworkImageVideo
🎯 What it does: This paper proposes a Slot Attention model called Invariant Slot Attention (ISA) that achieves spatial symmetry (translation, scaling, rotation) using the reference frame of each object, thereby enabling object representations to be invariant to geometric transformations in unsupervised scene decomposition.
Inverse Reinforcement Learning without Reinforcement Learning
Gokul Swamy (Carnegie Mellon University), Steven Wu
Reinforcement LearningGenerative Adversarial NetworkSequential
🎯 What it does: Two new algorithms (MMDP and NRMM) are proposed to accelerate inverse reinforcement learning using expert state distribution, along with theoretical proofs and experimental validation.
Investigating the Role of Model-Based Learning in Exploration and Transfer
Jacob C Walker, Jessica B Hamrick
Robotic IntelligenceMeta LearningReinforcement LearningWorld ModelTabular
🎯 What it does: The study explores the role of model-based learning in exploration and transfer through unsupervised pre-training in a reward-free environment.
IRNeXt: Rethinking Convolutional Network Design for Image Restoration
Yuning Cui (Sun Yat-sen University), Alois Knoll (Technical University of Munich)
RestorationConvolutional Neural NetworkImage
🎯 What it does: This paper proposes a convolution-based image inpainting network called IRNeXt.
Is Consensus Acceleration Possible in Decentralized Optimization over Slowly Time-Varying Networks?
Dmitry Metelev (Moscow Institute of Physics and Technology), Alexander Gasnikov (Skolkovo Institute of Science and Technology)
OptimizationGraph
🎯 What it does: This paper studies the communication complexity of decentralized optimization on slowly varying networks, deriving lower bounds under different network change rates (polynomial, logarithmic, constant), and proposes an accelerated consensus algorithm in time-varying graphs with a common connected skeleton.
Is Learning Summary Statistics Necessary for Likelihood-free Inference?
Yanzhi Chen (Cambridge University), Adrian Weller (Cambridge University)
TabularTime SeriesSequentialPhysics Related
🎯 What it does: A new method for learning slice-based sufficient statistics (SSS) is studied, and a posterior inference algorithm SNL+SSS is proposed based on it.
Is Overfitting Necessary for Implicit Video Representation?
Hee Min Choi (Samsung Advanced Institute of Technology), Dokwan Oh (Samsung Advanced Institute of Technology)
RestorationCompressionVideo
🎯 What it does: A method for implicit neural representation of videos without full training based on the strong lottery hypothesis is proposed, utilizing randomly initialized networks to learn multi-layer sparse subnetworks and learnable scaling factors for video encoding.
Iterative Approximate Cross-Validation
Yuetian Luo (University of Chicago), Rina Barber
OptimizationTabular
🎯 What it does: An Iterative Approximate Cross-Validation (IACV) method is proposed, which can efficiently approximate the leave-one-out cross-validation results when iterative solvers such as gradient descent, stochastic gradient descent, and proximal gradient descent have not converged.
JAWS-X: Addressing Efficiency Bottlenecks of Conformal Prediction Under Standard and Feedback Covariate Shift
Drew Prinster (Johns Hopkins University), Anqi Liu (Johns Hopkins University)
Computational EfficiencyDrug DiscoveryTabularBiomedical Data
🎯 What it does: For the scenarios of standard and feedback covariate shift (SCS and FCS), JAW-FCS and its two relaxed computational versions (JAWK-LOO and WCV+) are proposed to achieve distribution-independent finite sample confidence interval estimation.
Jump-Start Reinforcement Learning
Ikechukwu Uchendu (Google), Karol Hausman (Stanford University)
Reinforcement LearningTabularBenchmark
🎯 What it does: This paper proposes Jump-Start Reinforcement Learning (JSRL), which accelerates value-based reinforcement learning by first using an existing guiding policy for roll-in and then allowing the learning policy to continue exploring.
K-SHAP: Policy Clustering Algorithm for Anonymous Multi-Agent State-Action Pairs
Andrea Coletta (J.P. Morgan), Tucker Balch (J.P. Morgan)
OptimizationExplainability and InterpretabilityReinforcement LearningAgentic AITabularTime SeriesFinance Related
🎯 What it does: An algorithm named K-SHAP is proposed to identify different strategies through clustering in the state-action pairs of anonymous multi-agent systems.
KDEformer: Accelerating Transformers via Kernel Density Estimation
Amir Zandieh (Max Planck Institute for Informatics), Amin Karbasi (Yale University)
GenerationComputational EfficiencyTransformerImageText
🎯 What it does: This paper proposes an algorithm called KDEformer that utilizes Kernel Density Estimation (KDE) to accelerate Transformer attention, reducing the attention computation from quadratic time/space to nearly linear/sub-quadratic.
Kernel Logistic Regression Approximation of an Understandable ReLU Neural Network
Marie Guyomard (University Cote d'Azur), Lionel Fillatre (University Cote d'Azur)
OptimizationExplainability and InterpretabilityTabular
🎯 What it does: An interpretable ReLU neural network called SATURNN is proposed, modeled as a sum of additive univariate spline functions. It is proven that as the network width approaches infinity, it can be linearized and is equivalent to kernel logistic regression, ensuring unique convergence.
Kernel QuantTree
Diego Stucchi (Politecnico di Milano), Giacomo Boracchi (Politecnico di Milano)
Anomaly DetectionTabular
🎯 What it does: This paper studies a non-parametric multivariate data batch change detection algorithm called Kernel QuantTree (KQT).
Kernel Sufficient Dimension Reduction and Variable Selection for Compositional Data via Amalgamation
Junyoung Park (Korea Advanced Institute of Science and Technology), Cheolwoo Park (Korea Advanced Institute of Science and Technology)
TabularBiomedical Data
🎯 What it does: A variable selection method based on amalgamation and sufficient dimension reduction (SDR) is proposed to handle compositional data with a large number of zeros and high dimensionality.
Label differential privacy and private training data release
Robert Istvan Busa-Fekete (Google Research), Sergei Vassilvitskii (Google Research)
Safty and PrivacyTabular
🎯 What it does: This study investigates differential privacy mechanisms for sharing training data in machine learning environments, aiming to protect the privacy of user labels while learning accurate predictive models.
Label Distributionally Robust Losses for Multi-class Classification: Consistency, Robustness and Adaptivity
Dixian Zhu (University of Iowa), Tianbao Yang (Texas A&M University)
ClassificationOptimizationContrastive LearningImage
🎯 What it does: Proposes and analyzes the Label Distribution Robust (LDR) loss function, proving its Allk consistency and robustness for multi-class classification, and designs an adaptive temperature ALDR-KL loss;
Langevin Thompson Sampling with Logarithmic Communication: Bandits and Reinforcement Learning
Amin Karbasi (Yale University), Siddharth Mitra (Yale University)
OptimizationReinforcement LearningTabularStochastic Differential Equation
🎯 What it does: This paper proposes the use of Langevin Monte Carlo (LMC) methods to approximate the posterior in a batched environment, enabling the application of Thompson Sampling (TS) in multi-armed bandits (MAB) and infinite reinforcement learning (RL) without the need for conjugate distribution assumptions.
Language Instructed Reinforcement Learning for Human-AI Coordination
Hengyuan Hu (Stanford University), Dorsa Sadigh (Stanford University)
TransformerLarge Language ModelReinforcement LearningText
🎯 What it does: This paper proposes an 'instructRL' framework that utilizes large language models to generate prior strategies based on natural language instructions, aimed at guiding multi-agent reinforcement learning to achieve human-desired cooperative equilibria.
Large Language Models Can Be Easily Distracted by Irrelevant Context
Freda Shi (Google DeepMind), Denny Zhou (Google DeepMind)
TransformerLarge Language ModelPrompt EngineeringTextChain-of-Thought
🎯 What it does: This study investigates the interference of large language models when irrelevant context is present and constructs a new arithmetic reasoning dataset, GSM-IC, based on GSM8K to evaluate various prompting techniques (such as COT, LTM, PROGRAM, Self-Consistency, instruction-based prompting, etc.).
Large Language Models Struggle to Learn Long-Tail Knowledge
Nikhil Kandpal (University of North Carolina), Colin Raffel
TransformerLarge Language ModelTextRetrieval-Augmented Generation
🎯 What it does: This study investigates the relationship between the memory of knowledge in large language models and the frequency of related documents in their pre-training data in closed-book question answering tasks.
Last Switch Dependent Bandits with Monotone Payoff Functions
Ayoub Foussoul (Columbia University), assaf zeevi
OptimizationReinforcement Learning from Human FeedbackReinforcement Learning
🎯 What it does: This paper studies the Last Switch Dependency (LSD) gambling machine model and proposes an efficient constant approximation algorithm to solve the NP-hard problem of computing the optimal arm-pulling strategy under complete knowledge of the model.
Latent Traversals in Generative Models as Potential Flows
Yue Song (University of Trento), Max Welling (University of Amsterdam)
GenerationData SynthesisGenerative Adversarial NetworkImage
🎯 What it does: This paper proposes modeling the semantic transfer of latent space as the gradient flow of a learned dynamic potential field, achieving nonlinear and interpretable latent traversal.
Layered State Discovery for Incremental Autonomous Exploration
Liyu Chen (University of Southern California), Matteo Pirotta (Meta)
Robotic IntelligenceReinforcement LearningTabular
🎯 What it does: Under the framework of self-exploration, a hierarchical state discovery method and the Layered Autonomous Exploration (LAE) algorithm are proposed to achieve efficient learning of an incrementally controllable set of states, significantly reducing sample complexity and applicable to countably infinite state spaces.
Lazy Agents: A New Perspective on Solving Sparse Reward Problem in Multi-agent Reinforcement Learning
Boyin Liu (Chinese Academy of Sciences), Du Zhang (Macau University of Science and Technology)
Reinforcement Learning
🎯 What it does: A framework called LAIES is proposed from the perspective of 'lazy agents', utilizing causal reasoning to define lazy agents and teams, and enhancing MARL learning under sparse rewards through intrinsic rewards for individual and collaborative diligence (IDI, CDI).
LazyGNN: Large-Scale Graph Neural Networks via Lazy Propagation
Rui Xue (North Carolina State University), Xiaorui Liu (North Carolina State University)
ClassificationComputational EfficiencyGraph Neural NetworkGraph
🎯 What it does: This paper proposes LazyGNN, a shallow graph neural network that accumulates information through lazy forward and backward propagation during multi-round training, significantly reducing the computational and storage overhead caused by neighborhood explosion while maintaining long-distance dependencies.
LeadFL: Client Self-Defense against Model Poisoning in Federated Learning
Chaoyi Zhu (Delft University of Technology), Lydia Y. Chen (Delft University of Technology)
OptimizationFederated LearningSafty and PrivacyImage
🎯 What it does: A client-side self-defense method called LeadFL is proposed, which utilizes the regularization of the local gradient Hessian matrix to suppress the persistent effects of model poisoning attacks and can work in conjunction with existing server-side defenses.
Learn to Accumulate Evidence from All Training Samples: Theory and Practice
Deep Shankar Pandey (Rochester Institute of Technology), Qi Yu (Rochester Institute of Technology)
ClassificationConvolutional Neural NetworkSupervised Fine-TuningImage
🎯 What it does: This study investigates the flaw of evidence deep learning models in failing to learn from zero evidence samples during training and proposes a vacuum-based positive evidence regularization method (RED) to address this issue.
Learnability and Algorithm for Continual Learning
Gyuhak Kim (University of Illinois at Chicago), Bing Liu (University of Illinois at Chicago)
ClassificationRecognitionTransformerSupervised Fine-TuningImage
🎯 What it does: This paper studies and proves the learnability of Class Incremental Learning (CIL), and based on this, proposes a new CIL algorithm called ROW, which is based on replay, OOD detection, and WP heads.
Learning Affinity with Hyperbolic Representation for Spatial Propagation
Jin-Hwi Park (Gwangju Institute of Science and Technology), Hae-Gon Jeon (Gwangju Institute of Science and Technology)
RestorationSegmentationImage
🎯 What it does: In spatial propagation tasks, this paper introduces hyperbolic space to learn pixel-level affinity graphs, thereby achieving more accurate pixel relationship transmission.
Learning Antidote Data to Individual Unfairness
Peizhao Li (Brandeis University), Hongfu Liu (Brandeis University)
Data SynthesisGenerative Adversarial NetworkTabular
🎯 What it does: This paper studies how to enhance individual fairness and maintain predictive utility by generating comparable samples, referred to as 'antidotes', that resemble the original data distribution on tabular data.
Learning Belief Representations for Partially Observable Deep RL
Andrew Wang (University of Toronto), Sheila A. McIlraith (University of Toronto)
Representation LearningReinforcement LearningAuto EncoderImageSequential
🎯 What it does: This paper proposes a deep partially observable RL algorithm named Believer, which learns a Bayesian representation using the complete state information observable during training and relies only on observations during deployment;
Learning Compiler Pass Orders using Coreset and Normalized Value Prediction
Youwei Liang (University of California), Yuandong Tian (Meta AI)
CompressionOptimizationGraph Neural NetworkReinforcement LearningGraph
🎯 What it does: This paper first constructs a core set containing 50 efficient compiler pass sequences, and then utilizes Normalized Value Prediction to train a Graph Edge Attention Network (GEAN) to encode programs, achieving better code size compression within 45 compiler passes.
Learning Control by Iterative Inversion
Gal Leibovich (Intel Labs), Aviv Tamar (Technion)
Robotic IntelligenceTransformerSupervised Fine-TuningReinforcement LearningVideo
🎯 What it does: An Iterative Inversion algorithm is proposed, which can learn the inverse function through the evaluation of the forward function and supervised learning, with only the desired output distribution available and no input-output pairs, and apply it to high-dimensional control tasks.
Learning Control-Oriented Dynamical Structure from Data
Spencer M. Richards (Stanford University), Marco Pavone (Stanford University)
Reinforcement LearningTime Series
🎯 What it does: This paper proposes a state-dependent coefficient decomposition for learning control vector fields from limited input-output data, and based on this, implements nonlinear tracking control.
Learning Controllable Degradation for Real-World Super-Resolution via Constrained Flows
Seobin Park (University of Texas at Austin), Tae Hyun Kim (Hanyang University)
RestorationSuper ResolutionFlow-based ModelImage
🎯 What it does: The study investigates how to generate realistic low-resolution images with continuous degradation levels through latent space interpolation, thereby expanding the super-resolution dataset of real scenes.
Learning Deductive Reasoning from Synthetic Corpus based on Formal Logic
Terufumi Morishita (Hitachi), Yasuhiro Sogawa (Hitachi)
TransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This study proposes a synthetic reasoning corpus framework FLD based on formal logic axioms, using this corpus to train language models to enhance their deductive reasoning capabilities.
Learning Deep Time-index Models for Time Series Forecasting
Gerald Woo (Salesforce Research), Steven Hoi (Salesforce Research)
OptimizationMeta LearningTime Series
🎯 What it does: A deep time indexing model called DeepTime based on meta-optimization is proposed to learn generalizable function forms in long-term time series forecasting.
Learning Dense Correspondences between Photos and Sketches
Xuanchen Lu (University of California), Judith E Fan
RetrievalConvolutional Neural NetworkContrastive LearningImageBenchmark
🎯 What it does: A self-supervised method is proposed to learn the dense correspondence between photos and sketches, and a new annotated benchmark dataset PSC6k is constructed.
Learning Distributions over Quantum Measurement Outcomes
Weiyuan Gong (Tsinghua University), Scott Aaronson (University of Texas at Austin)
Physics Related
🎯 What it does: This paper proposes an online shadow imaging program for learning probability distributions from quantum measurements with K outcomes, addressing the problem of estimating the probability distribution of measurement results under a quantum state ρ and M unknown quantum measurements.
Learning Dynamic Query Combinations for Transformer-based Object Detection and Segmentation
Yiming Cui (University of Florida), Haichao Yu (ByteDance Inc.)
Object DetectionSegmentationTransformerReinforcement LearningImageVideo
🎯 What it does: For DETR and its variants, a method is proposed to dynamically generate query combinations based on high-order semantic information from images (modulated queries) to replace fixed queries, enhancing detection and segmentation performance.
Learning Expressive Priors for Generalization and Uncertainty Estimation in Neural Networks
Dominik Schnaus (Technical University Munich), Rudolph Triebel (German Aerospace Center)
ClassificationImage
🎯 What it does: This paper proposes a learnable prior method based on Bayesian neural networks, using the posterior of trained tasks as the prior for new tasks to enhance generalization and uncertainty estimation.
Learning for Edge-Weighted Online Bipartite Matching with Robustness Guarantees
Pengfei Li (University of California), Shaolei Ren (University of California)
Recommendation SystemReinforcement LearningTabular
🎯 What it does: A novel online switching framework combining reinforcement learning and expert algorithms, called LOMAR, is proposed for the online bipartite matching problem with edge weights.
Learning Functional Distributions with Private Labels
Changlong Wu (Purdue University), Wojciech Szpankowski (Purdue University)
🎯 What it does: The study learns the functional distribution from features to label distribution under the condition of label noise disturbance, and provides the optimal lower and upper bounds of risk for online learning.
Learning GFlowNets From Partial Episodes For Improved Convergence And Stability
Kanika Madan (Mila Quebec AI Institute), Nikolay Malkin (Mila Quebec AI Institute)
GenerationReinforcement LearningSequential
🎯 What it does: A new GFlowNet training objective SubTB(λ) is proposed, which improves the convergence and stability of the sampler by learning partial trajectories.
Learning Globally Smooth Functions on Manifolds
Juan Cervino (University of Pennsylvania), Alejandro Ribeiro (University of Pennsylvania)
OptimizationRobotic IntelligencePoint Cloud
🎯 What it does: This paper proposes a method for learning globally smooth functions on data manifolds by constraining the manifold Lipschitz constant.
Learning Hidden Markov Models When the Locations of Missing Observations are Unknown
Binyamin Perets (Technion Israel Institute of Technology), Shie Mannor (Technion Israel Institute of Technology)
SequentialBiomedical Data
🎯 What it does: This paper proposes a Gibbs sampling-based HMM learning framework that can reconstruct hidden Markov models in the presence of unknown missing observation locations, supporting both ignorable and non-ignorable missing probabilities.
Learning in POMDPs is Sample-Efficient with Hindsight Observability
Jonathan Lee, Tong Zhang (HKUST)
Reinforcement Learning
🎯 What it does: The Hindsight Observable Markov Decision Process (HOMDP) framework is proposed, which studies the POMDP problem where the latent states are observable during the training phase but only partially observable during the testing phase.
Learning Instance-Specific Augmentations by Capturing Local Invariances
Ning Miao (University of Oxford), Hyunjik Kim (DeepMind)
ClassificationRepresentation LearningData-Centric LearningConvolutional Neural NetworkContrastive LearningImage
🎯 What it does: InstaAug is proposed, a method for automatically generating data augmentation by learning the transformation distribution specific to the input; it can be end-to-end trained together with downstream models during training or learned separately on pre-trained models.
Learning Intuitive Policies Using Action Features
Mingwei Ma (Ubiquant Investment), Jakob Nicolaus Foerster
TransformerReinforcement LearningTabular
🎯 What it does: This paper studies the use of the semantic relationship between action features and observation features in multi-agent coordination tasks to learn intuitive and interpretable strategies through self-attention networks.
Learning Lightweight Object Detectors via Multi-Teacher Progressive Distillation
Shengcao Cao (University of Illinois Urbana-Champaign), Liangyan Gui
Object DetectionKnowledge DistillationConvolutional Neural NetworkTransformerImage
🎯 What it does: Learn a lightweight object detector through Multi-Teacher Progressive Distillation (MTPD), gradually transferring knowledge from a series of teacher models to the student model.
Learning Mixtures of Gaussians with Censored Data
Wai Ming Tai (University of Chicago), Bryon Aragam (University of Chicago)
Mixture of Experts
🎯 What it does: This paper studies the parameter estimation problem of one-dimensional Gaussian mixture models under censored data and provides finite sample convergence guarantees.
Learning Mixtures of Markov Chains and MDPs
Chinmaya Kausik (University of Michigan), Ambuj Tewari (University of Michigan)
Reinforcement LearningTabularTime Series
🎯 What it does: An offline algorithm is proposed to learn mixed Markov chains/MDPs with only short trajectories, and to perform trajectory clustering and model estimation.
Learning Neural Constitutive Laws from Motion Observations for Generalizable PDE Dynamics
Pingchuan Ma (Massachusetts Institute of Technology), Wojciech Matusik (Massachusetts Institute of Technology)
VideoPhysics Related
🎯 What it does: A hybrid framework NCLaw combining neural networks and partial differential equations (PDE) is proposed to learn generalizable constitutive laws from motion observations, thereby inferring dynamic behaviors under multiphysics and large deformations.
Learning Neural PDE Solvers with Parameter-Guided Channel Attention
Makoto Takamoto (NEC Laboratories Europe), Mathias Niepert (University of Stuttgart)
Convolutional Neural NetworkPhysics RelatedOrdinary Differential Equation
🎯 What it does: A parameter-guided channel attention module (CAPE) and a curriculum learning strategy are proposed to help the neural network PDE solver adapt to unseen PDE parameters.
Learning Noisy OR Bayesian Networks with Max-Product Belief Propagation
Antoine Dedieu (DeepMind), Miguel Lazaro-Gredilla
OptimizationReinforcement LearningText
🎯 What it does: This work proposes a noise-OR Bayesian network (BN) learning framework based on Parallel Max-Product inference, providing a linear time update implementation, and further utilizes GPU acceleration to achieve stochastic gradient training on large-scale dense data.
Learning Perturbations to Explain Time Series Predictions
Joseph Enguehard (Babylon Health)
Explainability and InterpretabilityRecurrent Neural NetworkTime SeriesBiomedical DataElectronic Health Records
🎯 What it does: This paper proposes a time series prediction explanation method based on learnable perturbations, using masks and GRU networks to jointly learn perturbations to more accurately capture the importance of temporal features.
Learning Physical Models that Can Respect Conservation Laws
Derek Hansen (University of Michigan), Michael W. Mahoney (Amazon Supply Chain Optimization Technologies)
Physics Related
🎯 What it does: A two-step framework called PROBCONSERV is proposed, which first uses a probabilistic model (such as Attentive Neural Process) to predict the mean and variance of the PDE solution, and then applies a Bayesian update using the integral form of the conservation constraint to obtain predictions that satisfy the conservation laws.
Learning Preconditioners for Conjugate Gradient PDE Solvers
Yichen Li (Massachusetts Institute of Technology), Wojciech Matusik (Massachusetts Institute of Technology)
Graph Neural NetworkGraphPhysics Related
🎯 What it does: This paper proposes a learning-based preconditioner based on graph neural networks to accelerate the conjugate gradient (PCG) solution of sparse positive definite systems arising from linear partial differential equations (PDEs).
Learning Prescriptive ReLU Networks
Wei Sun (IBM Research), Asterios Tsiourvas (Massachusetts Institute of Technology)
OptimizationExplainability and InterpretabilityDrug DiscoveryTabularBiomedical Data
🎯 What it does: This paper proposes an interpretable ReLU neural network called P-ReLU, which is used to learn the optimal policy for multiple treatment options from observational data and can directly embed constraints.
Learning Rate Schedules in the Presence of Distribution Shift
Matthew Fahrbach (Google Research), Pratik Worah (Google Research)
OptimizationBiomedical DataStochastic Differential Equation
🎯 What it does: The study focuses on online SGD learning with data distribution changing over time, designing adaptive learning rate schedules to minimize dynamic loss, and proposing optimal learning rate schemes for linear, convex, and non-convex losses.
Learning Regions of Interest for Bayesian Optimization with Adaptive Level-Set Estimation
Fengxue Zhang (University of Chicago), Yuxin Chen (University of Chicago)
OptimizationAuto EncoderTabular
🎯 What it does: This paper proposes a Bayesian optimization framework called BALLET based on adaptive hierarchical estimation, which first identifies high-confidence regions of interest (ROI) using a global GP, and then performs fine optimization within that region using a local GP.
Learning Representations without Compositional Assumptions
Tennison Liu (University of Cambridge), Mihaela van der Schaar (Alan Turing Institute)
Representation LearningGraph Neural NetworkAuto EncoderTabularBiomedical Data
🎯 What it does: Unsupervised representation learning for multi-view tabular data without relying on predefined combination assumptions.
Learning Signed Distance Functions from Noisy 3D Point Clouds via Noise to Noise Mapping
Baorui Ma (Tsinghua University), Zhizhong Han (Wayne State University)
GenerationData SynthesisPoint Cloud
🎯 What it does: Learning SDF from noisy point clouds using noise-to-noise mapping, without the need for clean point clouds or labels.
Learning Subpocket Prototypes for Generalizable Structure-based Drug Design
ZAIXI ZHANG, Qi Liu (University of Science and Technology of China)
Drug DiscoveryGraph Neural NetworkTransformerGraphBiomedical Data
🎯 What it does: The DrugGPS method is proposed, which utilizes sub-pocket prototypes and a global sub-pocket-molecule fragment interaction graph to generate high-affinity molecules at the molecular fragment level in structure-based drug design.
Learning Temporally AbstractWorld Models without Online Experimentation
Benjamin Freed (Carnegie Mellon University), Howie Choset (Carnegie Mellon University)
Robotic IntelligenceReinforcement LearningWorld ModelTabular
🎯 What it does: Proposes a method that learns a set of skills and a spatiotemporal abstract world model using only offline data, achieving zero-shot online planning;
Learning the Dynamics of Sparsely Observed Interacting Systems
Linus Bleistein (Inria Paris), Agathe Guilloux (Centre de Recherche des Cordeliers)
Time SeriesOrdinary Differential Equation
🎯 What it does: This study investigates how to learn the dynamics of a target time series from sparse observations of feature time series, and achieves predictions through controlled differential equations (CDE) and path signature theory.
Learning the Right Layers a Data-Driven Layer-Aggregation Strategy for Semi-Supervised Learning on Multilayer Graphs
Sara Venturini (University of Padova), Francesco Tudisco (Gran Sasso Science Institute)
ClassificationOptimizationGraph Neural NetworkGraph
🎯 What it does: For semi-supervised learning on multilayer graphs, a parameter-free, data-driven hierarchical aggregation strategy is designed to enhance node classification performance by learning nonlinear weighted combinations across different layers.
Learning to acquire novel cognitive tasks with evolution, plasticity and meta-meta-learning
Thomas Miconi
Meta LearningRecurrent Neural NetworkReinforcement LearningSequential
🎯 What it does: Through evolutionary, plasticity, and reward-modulated recurrent neural networks, the ability to autonomously acquire new cognitive tasks in a single task exposure has been achieved.
Learning to Bid in Repeated First-Price Auctions with Budgets
Qian Wang (Peking University), Yuqing Kong (Peking University)
OptimizationReinforcement Learning from Human FeedbackReinforcement LearningTabular
🎯 What it does: This study investigates how to learn the optimal bidding strategy under budget constraints in repeated first-price auctions.
Learning to Boost Training by Periodic Nowcasting Near Future Weights
Jinhyeok Jang (Electronics and Telecommunications Research Institute), ByungOk Han (Electronics and Telecommunications Research Institute)
OptimizationComputational EfficiencyConvolutional Neural NetworkImage
🎯 What it does: A module named Weight Nowcaster Network (WNN) is proposed, which periodically predicts the future weights of the network, allowing for skipping several training epochs and significantly reducing training time.
Learning to Decouple Complex Systems
Zihan Zhou (Chinese University of Hong Kong), Tianshu Yu (Chinese University of Hong Kong)
Anomaly DetectionOptimizationRecurrent Neural NetworkVideoTime SeriesSequentialStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: This paper proposes a model that decomposes complex systems into several relatively independent latent subsystems in sparse and chaotic sequential data, and introduces a meta-system for their interactions.
Learning to Design Analog Circuits to Meet Threshold Specifications
Dmitrii Krylov (University of California), Roy Fox (University of California)
Supervised Fine-TuningTabular
🎯 What it does: Proposes an automated analog circuit design method based on supervised learning, utilizing simulation data to generate a training set and predict circuit parameters that meet threshold specifications.