ICML 2023 Papers — Page 8
International Conference on Machine Learning · 1828 papers
Graph Neural Networks can Recover the Hidden Features Solely from the Graph Structure
Ryoma Sato (Kyoto University)
Graph Neural NetworkGraph
🎯 What it does: The study proves that graph neural networks can recover hidden node features from the graph structure in the absence of any explicit node features.
Graph Neural Networks with Learnable and Optimal Polynomial Bases
Yuhe Guo (Renmin University of China), Zhewei Wei (Renmin University of China)
Graph Neural NetworkGraph
🎯 What it does: Two spectral graph neural network models, FavardGNN and OptBasisGNN, are proposed, which can learn and utilize optimal polynomial bases during training to achieve graph convolution filtering.
Graph Neural Tangent Kernel: Convergence on Large Graphs
Sanjukta Krishnagopal (University of California, Los Angeles), Luana Ruiz (Massachusetts Institute of Technology)
Recommendation SystemGraph Neural NetworkGraph
🎯 What it does: This study investigates the learning dynamics of graph neural networks on large-scale graphs, proposing and proving that the Graph Neural Tangent Kernel (GNTK) converges to the graphon limit kernel (WNTK) as the graph size increases, and demonstrates its spectral convergence.
Graph Positional Encoding via Random Feature Propagation
Moshe Eliasof (Ben-Gurion University of the Negev), Haggai Maron (NVIDIA Research)
ClassificationRepresentation LearningGraph Neural NetworkGraph
🎯 What it does: A novel graph positional encoding method RFP is proposed, which generates feature trajectories through iterative propagation and normalization using random features;
Graph Reinforcement Learning for Network Control via Bi-Level Optimization
Daniele Gammelli (Stanford University), Francisco C. Pereira (Technical University of Denmark)
OptimizationGraph Neural NetworkReinforcement LearningGraph
🎯 What it does: By modeling the network control problem as reinforcement learning and introducing a bi-level optimization framework, an automated algorithm learning for large-scale dynamic network control has been achieved.
Graph Switching Dynamical Systems
Yongtuo Liu (University of Amsterdam), Efstratios Gavves (University of Amsterdam)
Recurrent Neural NetworkGraph Neural NetworkAuto EncoderTime SeriesSequentialOrdinary Differential Equation
🎯 What it does: A graphical switching dynamical system framework named GRASS is proposed to learn the pattern switching behavior of multi-object interactions using dynamic graphs.
GraphCleaner: Detecting Mislabelled Samples in Popular Graph Learning Benchmarks
Yuwen Li (National University of Singapore), Bryan Hooi (National University of Singapore)
ClassificationAnomaly DetectionGraph Neural NetworkGraphBenchmark
🎯 What it does: Detecting and correcting mislabels in graph learning benchmarks, a post-processing framework GraphCleaner is proposed to identify and correct node label errors.
Graphically Structured Diffusion Models
Christian Dietrich Weilbach (University of British Columbia), Frank Wood (University of British Columbia)
GenerationData SynthesisGraph Neural NetworkTransformerDiffusion modelGraph
🎯 What it does: A Graphically Structured Diffusion Model (GSDM) is proposed, which automatically learns a generative model that satisfies specific constraints by mapping the problem's graph structure into the attention mechanism of the diffusion model.
GREAD: Graph Neural Reaction-Diffusion Networks
Jeongwhan Choi (Yonsei University), Sung-Bae Cho (Yonsei University)
Graph Neural NetworkGraphOrdinary Differential Equation
🎯 What it does: A graph neural network framework GREAD based on reaction-diffusion equations is proposed to handle the propagation and interaction of node features on graphs.
Great Models Think Alike: Improving Model Reliability via Inter-Model Latent Agreement
Ailin Deng (National University of Singapore), Bryan Hooi (National University of Singapore)
ClassificationAnomaly DetectionTransformerContrastive LearningImage
🎯 What it does: This study proposes an interactive model latent space consistency framework that utilizes the neighborhood similarity between the base model and the newly trained model to assess the reliability of the latter, and enhances error detection performance through post-processing temperature scaling.
Grounding Language Models to Images for Multimodal Inputs and Outputs
Jing Yu Koh (Carnegie Mellon University), Daniel Fried (Carnegie Mellon University)
GenerationRetrievalTransformerLarge Language ModelVision Language ModelImageTextMultimodalityRetrieval-Augmented Generation
🎯 What it does: After freezing the pre-trained text language model, linear mapping and retrieval tokens are used to enable it to handle any interleaved image-text input, and to insert retrieved images during generation, thus achieving multimodal input and output.
Grounding Large Language Models in Interactive Environments with Online Reinforcement Learning
Thomas Carta (Inria), Pierre-Yves Oudeyer (Inria)
Robotic IntelligenceTransformerLarge Language ModelReinforcement LearningAgentic AIText
🎯 What it does: This paper proposes a method called GLAM that uses large language models (LLMs) as policies and achieves functional alignment through online reinforcement learning, training LLMs to complete navigation and manipulation tasks in a text-based interactive environment.
Group Equivariant Fourier Neural Operators for Partial Differential Equations
Jacob Helwig (Texas A&M University), Shuiwang Ji (Texas A&M University)
Time SeriesPhysics Related
🎯 What it does: This paper proposes a Fourier Neural Operator with group equivariance (G-FNO) implemented in the frequency domain, which can simultaneously maintain equivariance to rotation, translation, and reflection, thereby better solving partial differential equations.
GuardHFL: Privacy Guardian for Heterogeneous Federated Learning
Hanxiao Chen (University of Electronic Science and Technology of China), Xilin Zhang (University of Electronic Science and Technology of China)
Federated LearningSafty and PrivacyComputational EfficiencyImage
🎯 What it does: Proposes GuardHFL, a privacy-preserving heterogeneous federated learning framework that addresses the issues of query sample, model parameter, and response prediction leakage.
Guiding Pretraining in Reinforcement Learning with Large Language Models
Yuqing Du (University of California), Jacob Andreas (Massachusetts Institute of Technology)
TransformerLarge Language ModelReinforcement LearningText
🎯 What it does: This paper proposes a reward-free pre-training method for reinforcement learning (ELLM) that utilizes pre-trained large language models (LLMs) to generate executable, contextually relevant, and commonsense goals.
H-Likelihood Approach to Deep Neural Networks with Temporal-Spatial Random Effects for High-Cardinality Categorical Features
Hangbin Lee (Seoul National University), Youngjo Lee (Seoul National University)
Large Language ModelTabular
🎯 What it does: The paper proposes a new method based on h-likelihood for simultaneously estimating mean parameters, variance components, and random effects in deep neural networks, addressing the issue of obtaining accurate maximum likelihood estimates that traditional integration or variational methods struggle with.
Half-Hop: A graph upsampling approach for slowing down message passing
Mehdi Azabou (Georgia Tech), Eva L Dyer
ClassificationRepresentation LearningGraph Neural NetworkContrastive LearningGraph
🎯 What it does: A half-hop upsampling technique is proposed that inserts slow nodes on graph edges to slow down message propagation, alleviate oversmoothing, and enhance the performance of graph neural networks.
Hardness of Independent Learning and Sparse Equilibrium Computation in Markov Games
Dylan J Foster, Sham M. Kakade (Harvard University)
Reinforcement Learning
🎯 What it does: The paper proves that independent learning cannot achieve no-regret strategies in general Markov games, which is infeasible both computationally and statistically.
Hardware-Aware Compression with Random Operation Access Specific Tile (ROAST) Hashing
Aditya Desai (Rice University), Anshumali Shrivastava (ThirdAI Corp)
ClassificationCompressionImageText
🎯 What it does: A hardware-oriented parameter sharing compression method called ROAST is proposed, which achieves model compression through block-level hashing while keeping memory usage during training controllable.
Harmonic Neural Networks
Atiyo Ghosh (PASQAL SAS), Vincent Emanuel Elfving
Neural Architecture SearchPhysics Related
🎯 What it does: This paper introduces prior biases for harmonic functions in classical and quantum neural networks, and presents models that can be implemented in two-dimensional simply connected, two-dimensional multiply connected, and arbitrary dimensions (Holomorphic, Multiholomorphic, CurlNet, etc.).
HarsanyiNet: Computing Accurate Shapley Values in a Single Forward Propagation
Lu Chen (Shanghai Jiao Tong University), Quanshi Zhang (Shanghai Jiao Tong University)
Explainability and InterpretabilityComputational EfficiencyConvolutional Neural NetworkRecurrent Neural NetworkImageTabular
🎯 What it does: A network architecture named HarsanyiNet is proposed, which can simultaneously perform model inference and accurately compute the Shapley values of input variables in a single forward pass.
HETAL: Efficient Privacy-preserving Transfer Learning with Homomorphic Encryption
Seewoo Lee (University of California), Mun-Kyu Lee (Inha University)
OptimizationFederated LearningSafty and PrivacyComputational EfficiencyImage
🎯 What it does: HETAL is proposed, which fully implements the encrypted training process of transfer learning and completes model fine-tuning and early stopping on the server side.
Hidden Symmetries of ReLU Networks
Elisenda Grigsby, David Rolnick (Mila Quebec AI Institute)
Convolutional Neural Network
🎯 What it does: This study investigates the redundancy relationship between the parameter space and function space of fully connected ReLU networks, proving the existence of parameter settings without hidden symmetry in architectures where the width of the hidden layer is not less than the input dimension. It systematically enumerates several geometric mechanisms that lead to hidden symmetry; subsequently, it calculates the functional dimension during the initialization phase to statistically assess the occurrence probability of hidden symmetry under different network depths and widths.
Hiding Data Helps: On the Benefits of Masking for Sparse Coding
Muthu Chidambaram (Duke University), Rong Ge (Duke University)
🎯 What it does: The study addresses the issue of overcomplete dictionary learning in the presence of noise, where the standard objective is prone to overfitting, and proposes a masking objective to improve dictionary recovery.
Hiera: A Hierarchical Vision Transformer without the Bells-and-Whistles
Chaitanya Ryali (Meta AI), Christoph Feichtenhofer (Meta AI)
ClassificationRecognitionObject DetectionSegmentationTransformerAuto EncoderImageVideo
🎯 What it does: A minimal hierarchical visual Transformer (Hiera) has been constructed by removing 'bells-and-whistles' such as convolutions, cross-window/shifted window, and relative position embeddings, utilizing only MAE self-supervised pre-training to learn spatial biases, ultimately achieving faster and more accurate visual models.
Hierarchical Diffusion for Offline Decision Making
Wenhao Li (Chinese University of Hong Kong), Hongyuan Zha (Chinese University of Hong Kong)
OptimizationRobotic IntelligenceTransformerReinforcement LearningDiffusion modelTabularSequentialFinance Related
🎯 What it does: A hierarchical diffusion framework HDMI is proposed and implemented for offline decision-making, which can first discover sub-goals and then generate complete action sequences, thereby addressing issues such as long delays, sparse rewards, and sub-goal contamination.
Hierarchical Grammar-Induced Geometry for Data-Efficient Molecular Property Prediction
Minghao Guo (Massachusetts Institute of Technology), Wojciech Matusik (Massachusetts Institute of Technology)
Drug DiscoveryGraph Neural NetworkGraph
🎯 What it does: This paper proposes a geometric model of molecular structures based on a learnable hierarchical molecular grammar, achieving data-efficient molecular property prediction through graph neural diffusion on this geometry.
Hierarchical Imitation Learning with Vector Quantized Models
Kalle Kujanpää (Aalto University), Alexander Ilin (Aalto University)
Robotic IntelligenceReinforcement LearningSequential
🎯 What it does: By unsupervised segmentation of expert trajectories and learning sub-goals, using VQ-VAE to generate sub-goals, and combining search algorithms for hierarchical planning, imitation learning was achieved without upper-level supervision.
Hierarchical Neural Coding for Controllable CAD Model Generation
Xiang Xu (Simon Fraser University), Yasutaka Furukawa
GenerationData SynthesisTransformerAuto EncoderPoint Cloud
🎯 What it does: A three-layer hierarchical neural encoding is proposed, utilizing VQ-VAE to learn global, local, and curve geometric codes of CAD models, and controlling CAD generation and auto-completion through a two-stage autoregressive Transformer.
Hierarchical Programmatic Reinforcement Learning via Learning to Compose Programs
Guan-Ting Liu (National Taiwan University), Shao-Hua Sun (National Taiwan University)
Recurrent Neural NetworkReinforcement LearningSequential
🎯 What it does: This paper proposes a Hierarchical Programmatic Reinforcement Learning framework (HPRL), which learns a program embedding space and generates task-solving programs by sequentially combining programs through a meta-policy.
Hierarchies of Reward Machines
Daniel Furelos-Blanco (Imperial College London), Alessandra Russo (Imperial College London)
Reinforcement Learning
🎯 What it does: In reinforcement learning tasks, the authors propose the Hierarchical Reward Machine (HRM), a finite state machine that can form a hierarchy by invoking other reward machines to represent the reward function of the task and achieve subtask decomposition.
High Fidelity Image Counterfactuals with Probabilistic Causal Models
Fabio De Sousa Ribeiro (Imperial College London), Ben Glocker (Imperial College London)
Image TranslationGenerationData SynthesisAuto EncoderImageBiomedical DataMagnetic Resonance Imaging
🎯 What it does: A deep generative framework based on probabilistic Markov structure causal models (SCM) is proposed for generating high-fidelity counterfactuals and estimating direct, indirect, and total causal effects.
High Probability Convergence of Stochastic Gradient Methods
Zijian Liu (New York University), Huy Nguyen
Optimization
🎯 What it does: A high-probability convergence analysis is provided for stochastic gradient methods (stochastic mirror descent, accelerated stochastic mirror descent, SGD, AdaGrad-Norm) with sub-Gaussian noise, where the convergence upper bound depends only on the distance from the starting point to the optimal solution, rather than the domain diameter.
High-dimensional Clustering onto Hamiltonian Cycle
Tianyi Huang (Westlake University), Zhengjun Zhang (University of Chinese Academy of Sciences)
Anomaly DetectionOptimizationAuto EncoderImageTime Series
🎯 What it does: A method called HCHC is proposed to map high-dimensional clustering results onto Hamiltonian cycles, and a deep clustering model GLDC is used to generate relative probability distributions for sample visualization, while revealing clusters, cluster similarities, and outliers.
High-dimensional Location Estimation via Norm Concentration for Subgamma Vectors
Shivam Gupta (University of Texas at Austin), Eric Price (University of Texas at Austin)
🎯 What it does: A high-dimensional location estimation method is proposed that smooths samples by adding Gaussian noise and uses maximum likelihood estimation (MLE), providing a non-asymptotic error upper bound;
High-Probability Bounds for Stochastic Optimization and Variational Inequalities: the Case of Unbounded Variance
Abdurakhmon Sadiev (King Abdullah University of Science and Technology), Peter Richtárik (King Abdullah University of Science and Technology)
Optimization
🎯 What it does: This paper proposes and analyzes gradient clipping and acceleration methods in the context of high-probability convergence analysis with unbounded variance.
Hindsight Learning for MDPs with Exogenous Inputs
Sean R. Sinclair (Cornell University), Adith Swaminathan (Microsoft Research)
OptimizationReinforcement LearningTabularTime Series
🎯 What it does: This paper studies the Exogenous Markov Decision Process (Exo-MDP) and proposes a framework for accelerating learning using retrospective information—Hindsight Learning (HL).
Homomorphism AutoEncoder --- Learning Group Structured Representations from Observed Transitions
Hamza Keurti (Max Planck Institute for Intelligent Systems), Bernhard Schölkopf (Max Planck Institute for Intelligent Systems)
Representation LearningReinforcement Learning from Human FeedbackAuto EncoderImage
🎯 What it does: Designed and trained an autoencoder named Homomorphism Autoencoder (HAE), which learns group representations in the latent space, allowing the agent to learn group-structured and decomposable representations from observed transitions without relying on prior group structures or specific action sets. It is also proven that using two-step latent prediction and reconstruction loss ensures that the learned representations satisfy homogeneity and decoupling properties.
HOPE: High-order Graph ODE For Modeling Interacting Dynamics
Xiao Luo (University of California), Yizhou Sun (Peking University)
Graph Neural NetworkGraphTime SeriesOrdinary Differential Equation
🎯 What it does: A high-order graph neural ODE model HOPE is proposed for continuous-time modeling of dynamic interaction systems.
Horizon-Free and Variance-Dependent Reinforcement Learning for Latent Markov Decision Processes
Runlong Zhou (University of Washington), Simon Shaolei Du
Reinforcement Learning
🎯 What it does: Two reinforcement learning algorithms based on models and values are proposed to solve the problem of achieving approximately windowless, variance-related unbounded scheduling in LMDP (with hidden context) environments.
Horizon-free Learning for Markov Decision Processes and Games: Stochastically Bounded Rewards and Improved Bounds
Shengshi Li (Peking University), Lin Yang
Reinforcement LearningSequential
🎯 What it does: This paper proposes a new relaxation of the expected boundedness assumption to study the issue of ignoring time dependence in reinforcement learning and presents a general algorithmic framework applicable to Markov decision processes (MDPs) and games.
How Bad is Top-$K$ Recommendation under Competing Content Creators?
Fan Yao (University of Virginia), Haifeng Xu (University of Chicago)
Recommendation SystemReinforcement LearningTabular
🎯 What it does: Analyzes the impact of competition among content creators on user welfare in recommendation systems, providing upper and lower bounds on the mismatch of social welfare under optimal and self-interested behavior (PoA < 2).
How Do Transformers Learn Topic Structure: Towards a Mechanistic Understanding
Yuchen Li (Carnegie Mellon University), Andrej Risteski (Carnegie Mellon University)
TransformerText
🎯 What it does: This study investigates how a single-layer Transformer learns the thematic structure of words through the embedding layer and self-attention layer under the distribution of topic models (LDA), providing a theoretical explanation for the training dynamics.
How Does Information Bottleneck Help Deep Learning?
Kenji Kawaguchi (National University of Singapore), Jiaoyang Huang (University of Pennsylvania)
OptimizationRepresentation LearningConvolutional Neural NetworkImage
🎯 What it does: The paper provides a rigorous theoretical proof of the information bottleneck principle in deep learning based on statistical learning theory, deriving a sample complexity upper bound that includes mutual information terms, and experimentally validating its explanatory power for generalization error.
How Jellyfish Characterise Alternating Group Equivariant Neural Networks
Edward Pearce-Crump (Imperial College London)
🎯 What it does: A complete characterization of all possible forms of equivariant linear layers of the alternating group Aₙ on the tensor power space in Rⁿ is provided, along with the corresponding matrix basis.
How Many Perturbations Break This Model? Evaluating Robustness Beyond Adversarial Accuracy
Raphael Olivier (Carnegie Mellon University), Bhiksha Raj (Carnegie Mellon University)
Adversarial AttackConvolutional Neural NetworkImage
🎯 What it does: This paper proposes and evaluates a new robustness metric—adversarial sparsity—to measure the number of perturbations that can successfully attack a model under a given perturbation budget.
How much does Initialization Affect Generalization?
Sameera Ramasinghe (University of Adelaide), Simon Lucey (University of Adelaide)
Convolutional Neural NetworkImage
🎯 What it does: The study investigates the impact of network initialization on generalization performance, proposing a weak spectral bias theorem that reveals high frequencies can lead to overfitting.
How Powerful are Shallow Neural Networks with Bandlimited Random Weights?
Ming Li (Zhejiang Normal University), Jiye Liang (Shanxi University)
TabularTime Series
🎯 What it does: This paper studies the expressive power of shallow neural networks with finite bandwidth random weights and proposes and proves a lower bound on the approximation error of the network when the hidden parameters are constrained.
How to address monotonicity for model risk management?
Dangxing Chen (Duke Kunshan University), Weicheng Ye (Duke Kunshan University)
ClassificationExplainability and InterpretabilityTabularFinance Related
🎯 What it does: This study proposes interpretable Monotonic Groves of Neural Additive Models (MGNAMs) in the context of model risk management, achieving a unified constraint and verification of three types of monotonicity (individual feature monotonicity, weak symmetric monotonicity, and strong symmetric monotonicity).
How to Trust Your Diffusion Model: A Convex Optimization Approach to Conformal Risk Control
Jacopo Teneggi (Johns Hopkins University), Jeremias Sulam (Johns Hopkins University)
OptimizationDiffusion modelImageBiomedical DataComputed Tomography
🎯 What it does: The paper proposes a credibility assessment method for diffusion models, providing entry-level confidence intervals for future samples, and minimizes interval length while maintaining error control through a new multidimensional risk control technique (K-RCPS).
Human-Timescale Adaptation in an Open-Ended Task Space
Jakob Bauer (DeepMind), Lei M Zhang (DeepMind)
Knowledge DistillationMeta LearningTransformerReinforcement LearningSequential
🎯 What it does: This paper proposes a Transformer-based Meta-RL algorithm called Adaptive Agent (AdA), which can quickly adapt to unseen 3D tasks in a human-like manner within minutes and achieve high scores.
Hybrid Energy Based Model in the Feature Space for Out-of-Distribution Detection
Marc Lafon (Cedric Laboratory Cnam), Nicolas THOME
ClassificationAnomaly DetectionImageStochastic Differential Equation
🎯 What it does: Proposes the HEAT model, which uses a Hybrid Energy Model (EBM) to estimate the density of ID samples in the feature space of a pre-trained network, thereby achieving efficient OOD detection.
Hyena Hierarchy: Towards Larger Convolutional Language Models
Michael Poli (Stanford University), Christopher Re
TransformerLarge Language ModelText
🎯 What it does: Proposes the Hyena operator as a sub-quadratic alternative to attention, using implicit long convolution and gating to achieve autoregressive language modeling.
Hyperbolic Diffusion Embedding and Distance for Hierarchical Representation Learning
Ya-Wei Eileen Lin (Technion), Ronen Talmon (Technion)
Representation LearningDiffusion modelGraphTabularBiomedical Data
🎯 What it does: A multi-scale embedding and distance method based on diffusion geometry and hyperbolic geometry (HDE/HDD) is proposed, which can recover hierarchical relationships from observational data without prior hierarchical structure.
Hyperbolic Image-text Representations
Karan Desai (University of Michigan), Shanmukha Ramakrishna Vedantam
RetrievalRepresentation LearningTransformerVision Language ModelContrastive LearningImageText
🎯 What it does: Designed and trained a hyperbolic space-based image-text contrastive learning model MERU to capture the visual-language hierarchy.
Hyperbolic Representation Learning: Revisiting and Advancing
Menglin Yang (Chinese University of Hong Kong), Irwin King (Chinese University of Hong Kong)
Representation LearningGraph Neural NetworkGraph
🎯 What it does: The HIE method is proposed, which enhances the quality of the hierarchical structure of embeddings in hyper-surface representation learning through root alignment and hierarchical stretching.
Hyperparameters in Reinforcement Learning and How To Tune Them
Theresa Eimer (Leibniz University Hannover), Roberta Raileanu (Meta AI)
Hyperparameter SearchReinforcement Learning
🎯 What it does: Systematically study the impact of hyperparameters in reinforcement learning (RL) on performance, and propose a process and best practices for automatically tuning RL using existing AutoML hyperparameter optimization (HPO) tools (such as DEHB, PB2, BGT, etc.).
HyperTuning: Toward Adapting Large Language Models without Back-propagation
Jason Phang (New York University), Weizhu Chen (Microsoft Azure AI)
OptimizationData-Centric LearningTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: By introducing hypermodels for gradient-free model adaptation, the HyperTuning framework is proposed, which implements two T5-based hypermodels: HyperT5-Prefix and HyperT5-LoRA.
Hypervolume Knowledge Gradient: A Lookahead Approach for Multi-Objective Bayesian Optimization with Partial Information
Sam Daulton, Eytan Bakshy (Meta)
Recommendation SystemOptimizationTabular
🎯 What it does: A first-order forward-looking multi-objective Bayesian optimization method named Hypervolume Knowledge Gradient (HV-KG) is proposed, which can make decisions under partial information (such as decoupled evaluation and multi-fidelity evaluation).
Hypothesis Transfer Learning with Surrogate Classification Losses: Generalization Bounds through Algorithmic Stability
Anass Aghbalou (Telecom Paris), Guillaume Staerman (Universite Paris-Saclay)
Domain AdaptationOptimizationTabular
🎯 What it does: This paper studies the theoretical properties of Hypothesis Transfer Learning (HTL) in binary classification tasks, providing generalization error and excess risk bounds based on algorithm stability without complexity constraints, and deriving specific constants for common loss functions such as exponential, logarithmic, MSE, squared hinge, and softplus.
I$^2$SB: Image-to-Image Schrödinger Bridge
Guan-Horng Liu (Georgia Institute of Technology), Anima Anandkumar (California Institute of Technology)
Image TranslationRestorationSuper ResolutionDiffusion modelScore-based ModelImage
🎯 What it does: A new conditional diffusion model I2SB is proposed, which directly learns the nonlinear diffusion bridge between clear images and degraded images.
Identifiability and Generalizability in Constrained Inverse Reinforcement Learning
Andreas Schlaginhaufen (Ecole Polytechnique Federale de Lausanne), Maryam Kamgarpour (Ecole Polytechnique Federale de Lausanne)
OptimizationReinforcement Learning
🎯 What it does: A theoretical framework for constrained inverse reinforcement learning is proposed, analyzing the identifiability and generalizability of the reward function, and providing finite sample error guarantees, followed by validation of the theory in a grid world experiment.
Identifiability of Label Noise Transition Matrix
Yang Liu (University of California, Santa Cruz), Kun Zhang (Carnegie Mellon University)
Data-Centric LearningContrastive LearningImage
🎯 What it does: Investigate the identifiability of instance-level label noise transfer matrices and provide necessary and sufficient conditions.
Identification of the Adversary from a Single Adversarial Example
Minhao Cheng (Hong Kong University of Science and Technology), Pin-Yu Chen (IBM Research)
Adversarial AttackConvolutional Neural NetworkImage
🎯 What it does: A framework based on random mask watermarking is proposed, utilizing a single adversarial sample to trace the contaminated model replica used for its generation.
Identifying Interpretable Subspaces in Image Representations
Neha Kalibhat (University of Maryland), Soheil Feizi (University of Maryland)
Explainability and InterpretabilityRepresentation LearningVision Language ModelContrastive LearningImageText
🎯 What it does: This paper proposes the FALCON framework for automatically interpreting the concepts encoded by features in image model representations, eliminating the need for manual annotation and explicit alignment.
Identifying Useful Learnwares for Heterogeneous Label Spaces
Lan-Zhe Guo (Nanjing University), Zhi-Hua Zhou (Nanjing University)
ClassificationRecognitionKnowledge DistillationImage
🎯 What it does: A method for learnware specification and recognition under heterogeneous label spaces is proposed, combining class-level linear proxies with Reduced Kernel Mean Embedding (RKME), and employing the Hungarian algorithm for precise model reuse.
ILLUME: Rationalizing Vision-Language Models through Human Interactions
Manuel Brack (German Center for Artificial Intelligence), Kristian Kersting (German Center for Artificial Intelligence)
Explainability and InterpretabilityTransformerLarge Language ModelPrompt EngineeringVision Language ModelImageTextMultimodality
🎯 What it does: ILLUME is proposed, which transfers the reasoning and interpretability capabilities of language models to the framework of visual-language models through interactive sampling, human preference filtering, and Adapter fine-tuning.
Image generation with shortest path diffusion
Ayan Das (MediaTek Research), Alberto Bernacchia (MediaTek Research)
GenerationDiffusion modelImage
🎯 What it does: A shortest path diffusion model (SPD) based on entropy information geometry is proposed, which determines the image erosion process by calculating the shortest path in the probability distribution space, thereby improving the generation quality of the diffusion model.
Image Restoration with Mean-Reverting Stochastic Differential Equations
Ziwei Luo (Uppsala University), Thomas B. Schön (Uppsala University)
RestorationImageStochastic Differential Equation
🎯 What it does: A general image restoration method without prior knowledge is achieved through stepwise reverse simulation using mean-reverting stochastic differential equations (mean-reverting SDE) for tasks such as image denoising, deraining, and deblurring.
Image Shortcut Squeezing: Countering Perturbative Availability Poisons with Compression
Zhuoran Liu (Radboud University), Martha Larson (Radboud University)
CompressionAdversarial AttackImage
🎯 What it does: Proposes Image Shortcut Squeezing (ISS), a simple adversarial perturbation availability poisoning (PAP) defense method based on image compression;
Implicit Graph Neural Networks: A Monotone Operator Viewpoint
Justin Baker (University of Utah), Bao Wang (University of Utah)
ClassificationOptimizationGraph Neural NetworkGraph
🎯 What it does: This paper proposes MIGNN (Monotone Operator Implicit Graph Neural Network), which constructs a learnable implicit graph neural network capable of long-range dependencies through new well-posedness analysis, monotonic and orthogonal parameterization, and a splitting algorithm based on Anderson acceleration.
Implicit Jacobian regularization weighted with impurity of probability output
Sungyoon Lee (Hanyang University), Jaewook Lee (Seoul National University)
ClassificationOptimizationAdversarial AttackConvolutional Neural NetworkImage
🎯 What it does: This study investigates the implicit regularization effect of SGD on the logit-weight Jacobian norm in neural networks and proposes an explicit Jacobian regularization method based on this theory.
Implicit Neural Spatial Representations for Time-dependent PDEs
Honglin Chen (Columbia University), Peter Yichen Chen (Massachusetts Institute of Technology)
Time SeriesPhysics RelatedStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: Using implicit neural spatial representation (INSR) as a spatial discretization method for time-dependent partial differential equations (PDEs), and directly solving the PDEs in conjunction with classical time integrators.
Implicit Regularization Leads to Benign Overfitting for Sparse Linear Regression
Mo Zhou (Duke University), Rong Ge (Duke University)
Tabular
🎯 What it does: This paper studies how implicit regularization in sparse linear regression leads to benign overfitting and proposes a new model parameterization method that can train an interpolator with approximately optimal test loss through gradient descent.
Importance Weighted Expectation-Maximization for Protein Sequence Design
Zhenqiao Song (University of California), Lei Li (University of California)
GenerationOptimizationDrug DiscoveryTransformerBiomedical Data
🎯 What it does: A protein sequence design framework based on importance-weighted expectation maximization, IsEM-Pro, is proposed.
Improved Active Multi-Task Representation Learning via Lasso
Yiping Wang (Zhejiang University), Simon Shaolei Du
Representation LearningImage
🎯 What it does: This paper proposes an active multi-task representation learning strategy based on L1 regularization (L1-A-MTRL), which achieves more efficient source task sampling and representation learning through sparse task selection.
Improved Algorithms for Multi-period Multi-class Packing Problems with Bandit Feedback
Wonyoung Kim (Columbia University), assaf zeevi
OptimizationReinforcement Learning
🎯 What it does: This paper studies the Linear Context Multi-class Multi-period Packing Problem (LMMP), with the goal of packing items such that the total consumption vector is below a given budget vector while maximizing total value. The problem considers a decision-making process with bandit feedback.
Improved Algorithms for White-Box Adversarial Streams
Ying Feng (Carnegie Mellon University), David Woodruff
OptimizationAdversarial Attack
🎯 What it does: A series of streaming algorithms under the white-box adversarial flow model are proposed, capable of recovering structures such as vector sparsity, matrix low-rank, tensor low-rank, and sparse + low-rank (robust PCA) within limited space, and based on this, the space complexity of ℓ0 estimation, rank determination, and maximum matching problems is improved.
Improved Analysis of Score-based Generative Modeling: User-Friendly Bounds under Minimal Smoothness Assumptions
Hongrui Chen (Peking University), Jianfeng Lu (Duke University)
GenerationData SynthesisOptimizationScore-based ModelStochastic Differential Equation
🎯 What it does: This paper provides a theoretical analysis of score-based generative models (SGM) under the assumption of minimal smoothness, and concludes that SGM can converge to the data distribution with polynomial complexity under the premise of having only L2-accurate score estimators and finite second moments.
Improved Learning-Augmented Algorithms for the Multi-Option Ski Rental Problem via Best-Possible Competitive Analysis
Yongho Shin (Yonsei University), Hyung-Chan An (Yonsei University)
OptimizationTabular
🎯 What it does: A learning-enhanced algorithm for the multi-option ski rental problem is proposed, providing a randomized scheme for the first time and achieving an optimal competitive ratio.
Improved Online Conformal Prediction via Strongly Adaptive Online Learning
Aadyot Bhatnagar (Salesforce), Yu Bai (Salesforce)
ClassificationOptimizationMixture of ExpertsTime Series
🎯 What it does: This paper studies the problem of uncertainty quantification in online environments and proposes a new online conformal prediction method aimed at handling situations where data distribution changes over time.
Improved Online Learning Algorithms for CTR Prediction in Ad Auctions
Zhe Feng (Google), Zixin Zhou (Stanford University)
Recommendation SystemOptimization
🎯 What it does: This paper addresses the problem of real-time advertising bidding with unknown click-through rates by proposing two types of online learning mechanisms: a UCB-based prediction method for short-sighted advertisers (who only care about immediate returns) and a global IC mechanism that combines UCB with explore-then-commit for non-short-sighted advertisers (who focus on long-term returns). Both mechanisms are implemented within the framework of the Vickrey-Clarke-Groves second-price auction to achieve click-through rate prediction and revenue maximization.
Improved Policy Evaluation for Randomized Trials of Algorithmic Resource Allocation
Aditya Mate (Harvard University), Milind Tambe (Harvard University)
TabularBiomedical Data
🎯 What it does: An improved method for evaluating randomized experiments is proposed, which reduces estimation variance by reallocating observable counterfactuals to participants after the experiment ends.
Improved Regret for Efficient Online Reinforcement Learning with Linear Function Approximation
Uri Sherman (Tel Aviv University), Yishay Mansour (Tel Aviv University)
OptimizationComputational EfficiencyReinforcement Learning
🎯 What it does: This study investigates reinforcement learning with linear function approximation and adversarial change cost functions, proposing a computationally efficient policy optimization algorithm suitable for complex settings with unknown dynamics and bandit feedback.
Improved Techniques for Maximum Likelihood Estimation for Diffusion ODEs
Kaiwen Zheng (Tsinghua University), Jun Zhu (Tsinghua University)
GenerationData SynthesisDiffusion modelImageStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: This paper proposes improved techniques to enhance the maximum likelihood estimation performance of diffusion ODE models, further addressing numerical and matching issues during the training and evaluation processes.
Improving Adversarial Robustness by Putting More Regularizations on Less Robust Samples
Dongyoon Yang (Seoul National University), Yongdai Kim (Seoul National University)
Adversarial AttackConvolutional Neural NetworkGenerative Adversarial NetworkImage
🎯 What it does: A robust risk upper bound-based adversarial training algorithm ARoW is proposed, which enhances model robustness by applying stronger regularization to easily attackable samples.
Improving Adversarial Robustness of Deep Equilibrium Models with Explicit Regulations Along the Neural Dynamics
Zonghan Yang (Tsinghua University), Yang Liu (Tsinghua University)
OptimizationAdversarial AttackConvolutional Neural NetworkImageOrdinary Differential Equation
🎯 What it does: In the Deep Equilibrium (DEQ) model, two new adversarial defense strategies are introduced: ① reducing prediction entropy by gradually updating the input at each step of the neural dynamics; ② explicitly constraining the internal dynamics of DEQ by randomly selecting intermediate states to compute the loss during adversarial training.
Improving Adversarial Robustness Through the Contrastive-Guided Diffusion Process
Yidong Ouyang (Chinese University of Hong Kong), Guang Cheng (University of California)
Data SynthesisAdversarial AttackDiffusion modelContrastive LearningImage
🎯 What it does: This paper proposes a method for generating adversarially robust data using contrastive learning to guide diffusion models, called Contrastive-Guided Diffusion Process (Contrastive-DP), and theoretically proves that more distinguishable synthetic distributions under Gaussian settings can enhance adversarial robustness;
Improving Bi-level Optimization Based Methods with Inspiration from Humans' Classroom Study Techniques
Pengtao Xie (University of California San Diego)
OptimizationNeural Architecture SearchImageTextBiomedical Data
🎯 What it does: A general framework named Skillearn is proposed, which transforms human classroom learning techniques (such as the Feynman technique and peer questioning) into executable multi-layer optimization training strategies, applied to neural architecture search and data reweighting.
Improving Expert Predictions with Conformal Prediction
Eleni Straitouri (Max Planck Institute for Software Systems), Manuel Gomez Rodriguez (Max Planck Institute for Software Systems)
ClassificationRecommendation SystemImage
🎯 What it does: An automatic decision support system based on isometric prediction has been constructed, which provides experts with a label prediction set and forces them to select answers from the set to improve their accuracy in multi-class classification tasks.
Improving Fair Training under Correlation Shifts
Yuji Roh (KAIST), Changho Suh (KAIST)
ClassificationOptimizationTabular
🎯 What it does: This paper proposes a framework to enhance model fairness when the correlation between labels and sensitive groups shifts. It first adjusts the (y,z) correlation of the training data through preprocessing, and then employs existing in-processing fairness learning algorithms.
Improving Graph Generation by Restricting Graph Bandwidth
Nathaniel Lee Diamant (Genentech), Gabriele Scalia (Genentech)
GenerationData SynthesisComputational EfficiencyGraph Neural NetworkGraph
🎯 What it does: This paper proposes reducing the output space and computational complexity of graph generation models by limiting the bandwidth of the graph (BwR), utilizing the Cuthill-McKee permutation to obtain a low-bandwidth adjacency matrix, and applying it to existing graph generation methods.
Improving Graph Neural Networks with Learnable Propagation Operators
Moshe Eliasof (Ben-Gurion University), Eran Treister (Ben-Gurion University)
Graph Neural NetworkGraph
🎯 What it does: A learnable propagation weight ω graph neural network ω GNN is proposed, capable of mixing smoothing and sharpening propagation operators.
Improving Hyperparameter Learning under Approximate Inference in Gaussian Process Models
Rui Li (Aalto University), Arno Solin (Aalto University)
Hyperparameter SearchTabular
🎯 What it does: A mixed training method that combines variational inference and expectation propagation is proposed, using EP-style marginal likelihood to learn hyperparameters, improving hyperparameter learning in non-conjugate Gaussian process models.
Improving l1-Certified Robustness via Randomized Smoothing by Leveraging Box Constraints
Vaclav Voracek (University of Tuebingen), Matthias Hein (University of Tuebingen)
OptimizationImage
🎯 What it does: By utilizing box constraints in the image domain, the provable robustness against ℓ1 error is improved under the random smoothing framework.
Improving Medical Predictions by Irregular Multimodal Electronic Health Records Modeling
Xinlu Zhang (University of California), Linda Ruth Petzold (University of California)
ClassificationAnomaly DetectionTransformerMultimodalityTime SeriesBiomedical DataElectronic Health Records
🎯 What it does: A unified modeling framework for multimodal irregular electronic health records is proposed, modeling irregularities in time series, clinical notes, and multimodal fusion separately, and applied to early prediction tasks in the ICU.
Improving Statistical Fidelity for Neural Image Compression with Implicit Local Likelihood Models
Matthew J. Muckley, Jakob Verbeek (Inria)
CompressionAuto EncoderGenerative Adversarial NetworkImage
🎯 What it does: An improved statistical fidelity neural image compression model MS-ILLM is proposed, which combines a non-binary discriminator that generates local labels using VQ-VAE.
Improving the Model Consistency of Decentralized Federated Learning
Yifan Shi (Tsinghua University), Dacheng Tao (Sydney University)
Federated LearningSupervised Fine-TuningImage
🎯 What it does: This paper proposes two decentralized federated learning (DFL) algorithms—DFedSAM and DFedSAM-MGS, aimed at reducing the consistency differences between client models and improving generalization performance;
Improving Visual Prompt Tuning for Self-supervised Vision Transformers
Seungryong Yoo (Seoul National University), Sungroh Yoon (Seoul National University)
ClassificationSegmentationTransformerPrompt EngineeringContrastive LearningImage
🎯 What it does: A gated visual prompt tuning method is proposed to enable self-supervised ViT to adapt more effectively to downstream tasks.
In or Out? Fixing ImageNet Out-of-Distribution Detection Evaluation
Julian Bitterwolf (University Tuebingen), Matthias Hein (University Tuebingen)
Anomaly DetectionTransformerImage
🎯 What it does: This paper first identifies and quantifies the ID object contamination issue present in the existing ImageNet-1K OOD test set, and then constructs a completely ID object-free NINCO OOD dataset (5,879 images, 64 classes) and designs various OOD unit tests. Based on this, a systematic evaluation is conducted on multiple models (ViT, ConvNext, EfficientNet, CLIP, etc.) and various OOD detection methods (MSP, MaxLogit, Energy, Mahalanobis, ViM, RMaha, Cosine, etc.).