ICML 2023 Papers — Page 3
International Conference on Machine Learning · 1828 papers
Block Subsampled Randomized Hadamard Transform for Nyström Approximation on Distributed Architectures
Oleg Balabanov (Sorbonne Université), Victor Lederer
Gaussian SplattingTabular
🎯 What it does: A block sampling random Hadamard transform (Block SRHT) is designed and applied to Nystrom low-rank approximation in a distributed environment;
Blockwise Stochastic Variance-Reduced Methods with Parallel Speedup for Multi-Block Bilevel Optimization
Quanqi Hu (Texas A&M University), Tianbao Yang (Texas A&M University)
Recommendation SystemOptimizationHyperparameter SearchTabular
🎯 What it does: Two block-level stochastic variance reduction algorithms (BSVRB v1 and v2) for the multi-block Bayesian optimization (MBBO) problem are proposed, achieving the same optimal iteration complexity as single-block standard Bayesian optimization, and allowing for parallel sampling of multiple blocks and samples at each step, thus avoiding the computation of the inverse Hessian matrix in high dimensions.
Blossom: an Anytime Algorithm for Computing Optimal Decision Trees
Emir Demirović (Delft University of Technology), Louis Jean (Jolibrain)
OptimizationTabular
🎯 What it does: This paper proposes a dynamic programming algorithm called Blossom for solving optimal decision trees under a given maximum depth constraint, continuously returning better available trees during the search process.
BNN-DP: Robustness Certification of Bayesian Neural Networks via Dynamic Programming
Steven Adams (Delft University of Technology), Luca Laurenti (Delft University of Technology)
ClassificationOptimizationReinforcement LearningImageTabular
🎯 What it does: This paper proposes BNN-DP, an efficient algorithmic framework that models the adversarial robustness problem of Bayesian Neural Networks (BNN) as Dynamic Programming (DP); it provides lower and upper bounds on predictions over the input set T by hierarchical recursion of expected values at each layer.
Boosting Graph Contrastive Learning via Graph Contrastive Saliency
Chunyu Wei (Tsinghua University), LU FANG
Representation LearningGraph Neural NetworkContrastive LearningGraph
🎯 What it does: This paper proposes an unsupervised graph contrastive learning framework called GCS, which can automatically identify semantic substructures in graphs through a gradient-based approach and generate semantically consistent positive samples and hard negative samples for contrastive learning.
Boosting Offline Reinforcement Learning with Action Preference Query
Qisen Yang (Tsinghua University), Gao Huang (Tsinghua University)
Reinforcement LearningTabular
🎯 What it does: This paper proposes an offline reinforcement learning training framework called OAP based on action preference queries, which dynamically adjusts policy constraints on offline data without the need for online interaction, thereby improving policy performance.
Bootstrap in High Dimension with Low Computation
Henry Lam (Columbia University), Zhenyuan Liu (Columbia University)
Computational EfficiencyText
🎯 What it does: This study investigates the 'cheap bootstrap' method that uses very few (even just one) resampling in high-dimensional environments, providing the finite sample coverage error theory for this method and validating its effectiveness through numerical experiments.
Bootstrapped Representations in Reinforcement Learning
Charline Le Lan (University of Oxford), Will Dabney (Google DeepMind)
Representation LearningReinforcement Learning
🎯 What it does: This study investigates the state representation learned through bootstrapped auxiliary tasks in reinforcement learning, providing theoretical analysis and empirical validation.
BPipe: Memory-Balanced Pipeline Parallelism for Training Large Language Models
Taebum Kim (FriendliAI Inc), Byung-Gon Chun (FriendliAI Inc)
TransformerLarge Language ModelText
🎯 What it does: This paper proposes BPIPE, a pipeline parallel training method that achieves GPU memory balancing through activation transfer, enhancing the training efficiency of large language models.
Brainformers: Trading Simplicity for Efficiency
Yanqi Zhou (Google Deepmind), Jeff Dean (Google Deepmind)
Computational EfficiencyTransformerMixture of ExpertsText
🎯 What it does: A non-uniform, sparsifiable Transformer architecture called Brainformer is proposed, which derives complex layer sequences through evolutionary search.
Brauer's Group Equivariant Neural Networks
Edward Pearce-Crump (Imperial College London)
🎯 What it does: This paper provides a complete characterization of all learnable linear group equivariant layer functions (i.e., G-equivariant mappings from tensor spaces of arbitrary rank to tensor spaces of another rank) through the Schur-Weyl duality of Brauer algebras and the groups SO, O, and Sp, and presents explicit matrix bases (or generating sets) for them under standard/symplectic bases.
Building Neural Networks on Matrix Manifolds: A Gyrovector Space Approach
Xuan Son Nguyen (CY Cergy Paris Université), Shuo Yang (CY Cergy Paris Université)
ClassificationRecognitionVideoGraph
🎯 What it does: This paper studies how to construct neural networks on Symmetric Positive Definite (SPD) and Grassmann manifolds, proposing to extend gyrovector space theory to these two types of matrix manifolds, and based on this, defines basic operations and layers such as inner product, gyro angle, gyro isometries, and polynomial logistic regression (MLR) on SPD.
Buying Information for Stochastic Optimization
Mingchen Ma (University of Wisconsin Madison), Christos Tzamos (University of Wisconsin Madison)
Optimization
🎯 What it does: This paper studies how to purchase information for stochastic optimization and formalizes this problem as an online learning problem. Assuming the learner has an oracle for the original optimization problem, a deterministic algorithm with a competitive ratio of 2 and a randomized algorithm with a competitive ratio of e/e(ε-1) are designed for information purchasing.
Byzantine-Robust Learning on Heterogeneous Data via Gradient Splitting
Yuchen Liu (Zhejiang University), Gang Chen (Zhejiang University)
OptimizationFederated LearningImage
🎯 What it does: This paper proposes a gradient-splitting based Byzantine-Robust Federated Learning method called GAS, which addresses the high-dimensional disaster and gradient heterogeneity issues in non-IID environments.
CAB: Comprehensive Attention Benchmarking on Long Sequence Modeling
Jun Zhang (Shanghai AI Laboratory), Lingpeng Kong (The University of Hong Kong)
TransformerTextPoint CloudTime SeriesBenchmark
🎯 What it does: This paper proposes a comprehensive attention benchmark (CAB) for long sequence modeling and evaluates nine mainstream efficient attention architectures under four fine-grained attention modes (non-causal self-attention, causal self-attention, non-causal cross-attention, causal cross-attention).
Calibrating Multimodal Learning
Huan Ma (Tianjin University), Qinghua Hu
ClassificationRecognitionMultimodality
🎯 What it does: This study experimentally found that existing multimodal learning methods are unreliable in confidence estimation, often resulting in increased confidence when certain modalities are removed;
Can Forward Gradient Match Backpropagation?
Louis Fournier (Sorbonne Universite), Edouard Oyallon (Sorbonne Universite)
OptimizationConvolutional Neural NetworkReinforcement LearningImage
🎯 What it does: This paper studies a method for training deep networks using Forward Gradient without the use of backpropagation, and systematically evaluates the impact of different gradient targets (Global, Local, Intermediate) and gradient guesses (Random, NTK, Local) on training effectiveness.
Can Large Language Models Reason about Program Invariants?
Kexin Pei (Columbia University), Pengcheng Yin (Google Research)
AI Code AssistantTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
🎯 What it does: This paper achieves static invariant inference by fine-tuning a pre-trained large language model to directly predict the loop/function entry/exit point invariants at specified program points based on source code.
Can Neural Network Memorization Be Localized?
Pratyush Maini (Carnegie Mellon University), Chiyuan Zhang (Google Research)
ClassificationOptimizationConvolutional Neural NetworkTransformerSupervised Fine-TuningImage
🎯 What it does: This paper studies the spatial localization of the 'memorization' phenomenon in neural networks, proving that memorization is not concentrated in the last few layers of the model, but rather dispersed across a few neurons in different layers. It proposes an 'Example-Tied Dropout' method that directs memorization towards a pre-specified set of neurons.
Can We Scale Transformers to Predict Parameters of Diverse ImageNet Models?
Boris Knyazev (Samsung), Simon Lacoste-Julien (Mila)
OptimizationRepresentation LearningTransformerImage
🎯 What it does: Designed and trained a large-scale Graph Hyper-Network (GHN-3) that can quickly predict ImageNet pre-trained parameters and use these parameters to initialize models of various architectures, thereby reducing pre-training costs.
CataBEEM: Integrating Latent Interaction Categories in Node-wise Community Detection Models for Network Data
Yuhua Zhang (University of Michigan), Walter H. Dempsey (University of Michigan)
Graph Neural NetworkGraph
🎯 What it does: A CataBEEM model under an edge-exchangeable framework is proposed to simultaneously learn latent categories at the interaction level and community structures at the node level, achieving scalable variational EM inference on large-scale interaction networks.
Causal Bounds in Quasi-Markovian Graphs
Madhumitha Shridharan (Columbia University), Garud Iyengar (Columbia University)
OptimizationGraph Neural NetworkGraph
🎯 What it does: The study presents a multi-linear programming framework that can obtain upper and lower bounds for the case of unidentifiable causal queries in quasi-Markovian causal graphs. Through structural transformations, the polynomial objective degree is reduced from the total number of C-components to only the number of intervened C-components. Additionally, a Frank-Wolfe based convergent heuristic is proposed, which can quickly obtain bounds containing the true value on large-scale graphs.
Causal Discovery with Latent Confounders Based on Higher-Order Cumulants
Ruichu Cai (Guangdong University of Technology), Kun Zhang (Mohamed bin Zayed University of Artificial Intelligence)
Computational EfficiencyTabularTime SeriesFinance Related
🎯 What it does: This study investigates the closed-form OICA estimation using higher-order cumulants in linear non-Gaussian causal models with potential confounding variables, and proposes a recursive algorithm based on the One-Latent-Component condition to identify confounders and determine causal direction.
Causal Isotonic Calibration for Heterogeneous Treatment Effects
Lars van der Laan (University of Washington), Alex Luedtke (University of Washington)
Tabular
🎯 What it does: A non-parametric double-robust causal isotonic calibration method is proposed, along with the implementation of cross-calibration.
Causal Modeling of Policy Interventions From Treatment–Outcome Sequences
Çağlar Hızlı (Aalto University), Pekka Marttinen (Aalto University)
TabularTime SeriesBiomedical Data
🎯 What it does: Construct a joint treatment-outcome model that combines marked point processes and conditional Gaussian processes to achieve causal inference of treatment policies in continuous time.
Causal Proxy Models for Concept-based Model Explanations
Zhengxuan Wu (Stanford University), Christopher Potts (Stanford University)
Explainability and InterpretabilityKnowledge DistillationText
🎯 What it does: This paper proposes a method to explain the causal concepts of NLP black box models by training a proxy model.
Causal Strategic Classification: A Tale of Two Shifts
Guy Horowitz (Technion Israel Institute of Technology), Nir Rosenfeld (Technion Israel Institute of Technology)
ClassificationOptimizationTabularFinance Related
🎯 What it does: This paper studies the strategic classification problem where users can change features and these changes affect the true labels, and proposes a learning framework and algorithm that incorporates causal relationships.
Causal Structure Learning for Latent Intervened Non-stationary Data
Chenxi Liu (Zhejiang University), Kun Kuang (Zhejiang University)
Time Series
🎯 What it does: In non-stationary time series, causal structure learning is conducted under potential interventions without domain labels, proposing the LIN method to simultaneously recover domain indices and causal graphs.
Cell-Free Latent Go-Explore
Quentin Gallouédec (Universite Lyon), Emmanuel Dellandrea (Universite Lyon)
Robotic IntelligenceReinforcement LearningSequential
🎯 What it does: Proposes the Latent Go-Explore (LGE) method, which utilizes learned latent representations to achieve exploration strategies without cell-based or domain knowledge, constructing target sampling and subgoal paths.
Certified Robust Neural Networks: Generalization and Corruption Resistance
Mohammed Amine Bennouna (Massachusetts Institute of Technology), Bart Van Parys (Massachusetts Institute of Technology)
OptimizationAdversarial AttackConvolutional Neural NetworkSupervised Fine-TuningImage
🎯 What it does: A holistic robust (HR) training framework is proposed, which can simultaneously resist evasion attacks, poisoning attacks, and statistical errors, and theoretically provides evidence and protection methods against robust overfitting.
Certifying Ensembles: A General Certification Theory with S-Lipschitzness
Aleksandar Petrov (University of Oxford), Adel Bibi (University of Oxford)
ClassificationOptimizationAdversarial AttackImage
🎯 What it does: This paper introduces the S-Lipschitz property, based on which a more precise robustness certification theory is provided, and it is used to analyze the robustness of ensemble classifiers.
Chameleon: Adapting to Peer Images for Planting Durable Backdoors in Federated Learning
Yanbo Dai (Hong Kong University of Science and Technology), Songze Li (Hong Kong University of Science and Technology)
Federated LearningAdversarial AttackConvolutional Neural NetworkContrastive LearningImage
🎯 What it does: An attack method named Chameleon is proposed, which implants a more persistent visual backdoor in federated learning.
Change is Hard: A Closer Look at Subpopulation Shift
Yuzhe Yang (Massachusetts Institute of Technology), Marzyeh Ghassemi (Massachusetts Institute of Technology)
Domain AdaptationData-Centric LearningImageTextBiomedical DataBenchmark
🎯 What it does: This paper systematically evaluates the impact of subpopulation shift on various real datasets and constructs a comprehensive benchmark that includes 20 advanced algorithms and 12 cross-domain datasets.
Chemically Transferable Generative Backmapping of Coarse-Grained Proteins
Soojung Yang (Massachusetts Institute of Technology), Rafael Gomez-Bombarelli
GenerationProtein Structure PredictionAuto EncoderBiomedical Data
🎯 What it does: A generative post-mapping method has been developed that can quickly recover all atomic-level details from coarse-grained (only containing α carbon) structures;
CHiLS: Zero-Shot Image Classification with Hierarchical Label Sets
Zachary Novack (University of California), Saurabh Garg (Carnegie Mellon University)
ClassificationTransformerContrastive LearningImage
🎯 What it does: The CHiLS method is proposed, which enhances the zero-shot image classification performance of CLIP by utilizing hierarchical label sets.
ChiPFormer: Transferable Chip Placement via Offline Decision Transformer
Yao Lai (University of Hong Kong), Ping Luo (University of Hong Kong)
OptimizationTransformerReinforcement LearningGraph
🎯 What it does: This paper proposes ChiPFormer, a chip layout method based on offline decision transformers;
CircuitNet: A Generic Neural Network to Realize Universal Circuit Motif Modeling
Yansen Wang (Microsoft Research Asia), Dongsheng Li (Microsoft Research Asia)
ClassificationOptimizationConvolutional Neural NetworkRecurrent Neural NetworkTransformerReinforcement LearningImageTime Series
🎯 What it does: A general neural network architecture named CircuitNet is proposed, which achieves unified modeling of four basic circuit patterns (feedforward, interconnection, feedback, lateral) in biological neural networks by introducing densely connected 'Circuit Primitive Units' (CMU) into the neural network.
ClimaX: A foundation model for weather and climate
Tung Nguyen (University of California Los Angeles), Aditya Grover (Microsoft)
TransformerTime Series
🎯 What it does: A Transformer-based model named ClimaX was trained, undergoing unsupervised pre-training on multivariate, multi-spatial-temporal scale climate and weather data, and fine-tuned for various downstream tasks including global forecasting, regional forecasting, sub-seasonal to seasonal prediction, climate projection, and downscaling.
CLIPood: Generalizing CLIP to Out-of-Distributions
Yang Shu (Tsinghua University), Mingsheng Long (Tsinghua University)
Domain AdaptationTransformerSupervised Fine-TuningPrompt EngineeringContrastive LearningImageTextMultimodality
🎯 What it does: Fine-tune the CLIP model with a focus on achieving better generalization capabilities under the OOD setting where domain shift and new classes (open class) coexist in downstream tasks.
Cluster Explanation via Polyhedral Descriptions
Connor Lawless (Cornell University), Oktay Gunluk (Cornell University)
OptimizationExplainability and InterpretabilityTabular
🎯 What it does: This paper proposes a polytope-based clustering explanation framework that describes clustering results by constructing polytopes around each cluster.
ClusterFuG: Clustering Fully connected Graphs by Multicut
Ahmed Abbas (Max Planck Institute for Informatics), Paul Swoboda (University of Mannheim)
SegmentationRetrievalGraph Neural NetworkContrastive LearningImageGraph
🎯 What it does: This paper proposes a dense clustering model based on multi-cut (weighted correlation clustering) on complete graphs, utilizing the inner product of node features to construct edge costs, and presents several efficient solving methods (Dense-GAEC, incremental nearest neighbor updates, lazy edge contraction, etc.), enabling clustering of large-scale graphs without the need to pre-specify the graph topology.
CLUSTSEG: Clustering for Universal Segmentation
James Chenhao Liang, Wenguan Wang (Zhejiang University)
SegmentationTransformerImage
🎯 What it does: This paper proposes CLUSTSEG, a unified framework based on transformers that can simultaneously perform superpixel, semantic, instance, and panoptic segmentation tasks.
CLUTR: Curriculum Learning via Unsupervised Task Representation Learning
Abdus Salam Azad (University of California), Ion Stoica (University of California)
Representation LearningRecurrent Neural NetworkReinforcement LearningAuto EncoderSequential
🎯 What it does: Developed a curriculum learning algorithm CLUTR based on unsupervised task representation learning for unsupervised environment design in reinforcement learning.
CO-BED: Information-Theoretic Contextual Optimization via Bayesian Experimental Design
Desi R. Ivanova (University of Oxford), Adam Foster (Microsoft Research)
OptimizationReinforcement LearningTabular
🎯 What it does: A general context optimization framework CO-BED is proposed, which utilizes the expected information gain from Bayesian experimental design (CMV-EIG) to guide experimental design, thereby achieving better contextual decisions during the deployment phase.
Coarse-to-Fine: a Hierarchical Diffusion Model for Molecule Generation in 3D
Bo Qiang (Peking University), Yanyan Lan (Tsinghua University)
GenerationDrug DiscoveryGraph Neural NetworkDiffusion modelGraph
🎯 What it does: A hierarchical diffusion model (HierDiff) is proposed, which first generates 3D molecular geometric structures at the fragment level and then decodes them into atomic-level molecular constructs.
Cocktail Party Attack: Breaking Aggregation-Based Privacy in Federated Learning Using Independent Component Analysis
Sanjay Kariyappa (Georgia Institute of Technology), Hsien-Hsin S. Lee (Intel)
Federated LearningSafty and PrivacyAdversarial AttackConvolutional Neural NetworkImage
🎯 What it does: This paper proposes the Cocktail Party Attack (CPA), which transforms the problem of recovering aggregated gradients into a blind source separation (BSS) problem and uses independent component analysis (ICA) to recover private inputs from aggregated gradients/weight updates, achieving a breakthrough in the privacy of federated learning aggregation.
CocktailSGD: Fine-tuning Foundation Models over 500Mbps Networks
Jue WANG, Ce Zhang (ETH Zurich)
CompressionOptimizationTransformerLarge Language ModelSupervised Fine-TuningTextChain-of-Thought
🎯 What it does: The COCKTAILSGD framework is proposed, integrating three compression techniques: random sparsity, Top-K sparsity, and quantization, achieving fine-tuning of a 20B parameter LLM over a 500 Mbps network, with a communication volume reduction of 117× and a training speed that is only 1.2× slower.
CoCo: A Coupled Contrastive Framework for Unsupervised Domain Adaptive Graph Classification
Nan Yin (Mohamed bin Zayed University of Artificial Intelligence), Xiao Luo (University of California)
ClassificationDomain AdaptationGraph Neural NetworkContrastive LearningGraph
🎯 What it does: The CoCo framework is proposed, which utilizes Graph Convolutional Networks (GCN) and Hierarchical Graph Kernel Networks (HGKN) to jointly learn graph structures through two branches, combined with cross-branch and cross-domain contrastive learning to achieve unsupervised domain adaptive graph classification.
CodeIPPrompt: Intellectual Property Infringement Assessment of Code Language Models
Zhiyuan Yu (Washington University in St. Louis), Chaowei Xiao (Arizona State University)
GenerationAI Code AssistantTransformerSupervised Fine-TuningPrompt EngineeringText
🎯 What it does: Developed the CODEIPPROMPT platform to automate the assessment of the infringement risk of licensed code by code generation language models; the platform constructs prompts by extracting function signatures and comments from GitHub licensed code repositories and uses similarity detection tools to evaluate the similarity between generated code and original licensed code.
Coder Reviewer Reranking for Code Generation
Tianyi Zhang (Stanford University), Sida Wang
GenerationAI Code AssistantTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: This paper studies and implements a reordering method based on a pre-trained code generation model (Coder) and a reverse evaluation model (Reviewer) called Coder-Reviewer reordering. It estimates the likelihood of instruction-to-program mappings through prompts without additional training.
CoDi: Co-evolving Contrastive Diffusion Models for Mixed-type Tabular Synthesis
Chaejeong Lee (Yonsei University), Noseong Park (Yonsei University)
GenerationData SynthesisDiffusion modelContrastive LearningTabular
🎯 What it does: This paper proposes a dual-modal diffusion model that targets both continuous and discrete variables, capable of synthesizing high-quality, rich, and fast-generating mixed-type tabular data.
Coin Sampling: Gradient-Based Bayesian Inference without Learning Rates
Louis Sharrock (Lancaster University), Christopher Nemeth (Lancaster University)
Tabular
🎯 What it does: A class of learning rate-free particle sampling method called Coin Sampling is proposed, which implements sampling from the Bayesian posterior based on coin betting techniques and Wasserstein gradient flow.
COLA: Orchestrating Error Coding and Learning for Robust Neural Network Inference Against Hardware Defects
Anlan Yu (Lehigh University), Wujie Wen (Lehigh University)
ClassificationOptimizationConvolutional Neural NetworkImage
🎯 What it does: A framework named COLA is proposed, aimed at improving the clean accuracy and robust accuracy of deep neural networks using error-correcting output codes (ECOC) in hardware defect environments by reducing error correlation.
Cold Analysis of Rao-Blackwellized Straight-Through Gumbel-Softmax Gradient Estimator
Alexander Shekhovtsov (Czech Technical University in Prague)
Image
🎯 What it does: This study proposes a closed-form solution for the Gumbel-Rao estimator in the zero-temperature limit, referred to as the ZGR estimator.
Collaborative Causal Inference with Fair Incentives
Rui Qiao (National University of Singapore), Bryan Kian Hsiang Low (National University of Singapore)
Biomedical Data
🎯 What it does: This paper proposes a collaborative causal inference framework that incentivizes self-interested parties to share data and obtain more accurate causal effect estimates through data value assessment and a fair reward mechanism.
Collaborative Multi-Agent Heterogeneous Multi-Armed Bandits
Ronshee Chawla (University of Texas), R. Srikant (University of Illinois)
OptimizationReinforcement Learning from Human FeedbackTabular
🎯 What it does: A collaborative multi-agent heterogeneous multi-armed bandit model is proposed, supporting N agents to learn M K-armed bandits respectively, and two decentralized algorithms are designed for two communication scenarios (no context and partially known context).
Combinatorial Neural Bandits
Taehyun Hwang (Seoul National University), Min-hwan Oh (Seoul National University)
Recommendation SystemOptimizationReinforcement LearningTabular
🎯 What it does: This paper studies a contextual combinatorial bandit problem and proposes the Combinatorial Neural UCB (CN-UCB) and Combinatorial Neural Thompson Sampling (CN-TS) algorithms, utilizing deep neural networks to approximate the unknown scoring function.
COMCAT: Towards Efficient Compression and Customization of Attention-Based Vision Models
Jinqi Xiao (Rutgers University), Bo Yuan (Snap Inc.)
GenerationCompressionTransformerDiffusion modelImageText
🎯 What it does: This paper proposes a method for compressing visual Transformers and personalizing text-to-image diffusion models based on head-level low-rank approximation of attention layers.
Communication-Constrained Bandits under Additive Gaussian Noise
Prathamesh Mayekar (National University of Singapore), Vincent Tan
Reinforcement Learning
🎯 What it does: This paper studies the distributed stochastic multi-armed bandit (SMAB) problem under an AWGN channel, proposing a coding scheme and a multi-stage upper bound algorithm, and providing lower and upper bounds parameterized by the signal-to-noise ratio (SNR).
Communication-Efficient Federated Hypergradient Computation via Aggregated Iterative Differentiation
Peiyao Xiao (University at Buffalo), Kaiyi Ji (University at Buffalo)
OptimizationFederated LearningRepresentation LearningAuto EncoderImage
🎯 What it does: Proposed the Aggregated Iterative Differentiation (AggITD) method for estimating supergradients in federated bilevel optimization, and based on this, designed the FBO-AggITD algorithm.
Comparison of meta-learners for estimating multi-valued treatment heterogeneous effects
Naoufal Acharki (Ecole polytechnique), Josselin Garnier (Ecole polytechnique)
Meta LearningTabular
🎯 What it does: This paper studies the estimation of heterogeneous treatment effects (CATE) under multi-valued treatment (multiple treatment levels), explores the performance of various meta-learners (T, S, M, DR, X, R) in this scenario, and provides theoretical error bounds along with empirical validation.
Competing for Shareable Arms in Multi-Player Multi-Armed Bandits
Renzhe Xu (Tsinghua University), Peng Cui (Tsinghua University)
OptimizationReinforcement Learning
🎯 What it does: A shared reward model for multi-player multi-armed bandits is designed, and the SMAA algorithm is proposed to implement the learning and competition of self-interested players.
Competitive Gradient Optimization
Abhijeet Vyas (Purdue University), Kamyar Azizzadenesheli (Nvidia)
Optimization
🎯 What it does: Proposes a Competitive Gradient Optimization (CGO) method to solve zero-sum two-player minimax problems; by incorporating cross second-order terms between players in the gradient update, the algorithm can converge in non-convex, non-concave, and non-weak-MVI scenarios.
Complementary Attention for Multi-Agent Reinforcement Learning
Jianzhun Shao (Tsinghua University), Xiangyang Ji (Tsinghua University)
Recurrent Neural NetworkReinforcement LearningSequential
🎯 What it does: In multi-agent reinforcement learning, this paper proposes a new framework called CAMA to simultaneously address the issues of attention dispersion and information insufficiency caused by dynamic team sizes and partial observability. The framework divides environmental entities into high-attention and low-attention parts through an entity partitioning module, enhancing attention on high-attention entities using inverse models, and using a global coach to generate compressed information for low-attention entities to achieve attention compensation, thereby enabling more robust and efficient collaborative decision-making.
Complexity of Block Coordinate Descent with Proximal Regularization and Applications to Wasserstein CP-dictionary Learning
Dohyun Kwon (University of Seoul), Hanbaek Lyu (University of Wisconsin Madison)
OptimizationTabular
🎯 What it does: This paper studies the Block Coordinate Descent method with proximal regularization (BCD-PR) and establishes theoretical worst-case complexity bounds for it. An effective Wasserstein CP dictionary learning algorithm is proposed, aimed at finding a set of base probability distributions that can well approximate a given d-dimensional joint probability distribution.
Composer: Creative and Controllable Image Synthesis with Composable Conditions
Lianghua Huang (Alibaba Group), Jingren Zhou (Alibaba Group)
GenerationData SynthesisDiffusion modelImage
🎯 What it does: Composer is constructed, a multi-conditional diffusion model that enables controllable and creative image synthesis by combining the decomposed representations of images.
Compositional Exemplars for In-context Learning
Jiacheng Ye (University of Hong Kong), Lingpeng Kong (University of Hong Kong)
GenerationRetrievalTransformerLarge Language ModelContrastive LearningText
🎯 What it does: This paper proposes a global example subset selection method called CEIL based on Conditional DPP, which utilizes the output probabilities of large language models for contrastive learning to automatically select the most helpful demonstration example set in In-Context Learning.
Compositional Score Modeling for Simulation-Based Inference
Tomas Geffner (University of Massachusetts), Andriy Mnih (DeepMind)
Score-based ModelBenchmark
🎯 What it does: A conditional score-based SBI method (F-NPSE, PF-NPSE) has been developed, which efficiently approximates the posterior for any number of observations by learning from single-observation posterior scores and aggregating scores during inference, all while using only single-observation training samples.
Compressed Decentralized Proximal Stochastic Gradient Method for Nonconvex Composite Problems with Heterogeneous Data
Yonggui Yan (Rensselaer Polytechnic Institute), Yangyang Xu (Rensselaer Polytechnic Institute)
OptimizationConvolutional Neural NetworkImage
🎯 What it does: Two decentralized atomic gradient tracking methods, DProxSGT and its compressed version CDProxSGT, are proposed to address the optimization problem of non-convex sparse/non-smooth objective functions under heterogeneous data distributions.
Compressing Tabular Data via Latent Variable Estimation
Andrea Montanari (Project N), Eric Weiner (Project N)
CompressionTabular
🎯 What it does: A lossless compression method for tabular data based on latent variable estimation is proposed, which achieves higher compression rates by first clustering the latent variables of rows and columns, then using LZ/ZSTD or ANS encoding within the same latent block, and finally encoding the latent variables themselves.
Computational Asymmetries in Robust Classification
Samuele Marro (University of Bologna), Michele Lombardi (University of Bologna)
ClassificationAdversarial AttackConvolutional Neural NetworkImageBenchmark
🎯 What it does: The paper explores the computational asymmetry of robust classification from both theoretical and experimental perspectives, proving that attacking ReLU classifiers is NP-hard, while finding robust parameters during the training phase is ΣP₂-hard; it proposes a defense method based on attacks called Counter-Attack and provides its complexity; further experiments demonstrate that heuristic attacks can reliably approach the decision boundary distance, and the UG100 dataset is released.
Computational Doob h-transforms for Online Filtering of Discretely Observed Diffusions
Nicolas Chopin (Institut Polytechnique de Paris), Alexandre H. Thiery (National University of Singapore)
OptimizationTime SeriesStochastic Differential Equation
🎯 What it does: This paper proposes a method based on the computational Doob h-transformation (CDT) to approximate the global adaptive particle filter for discrete observation diffusion processes.
Computationally Efficient PAC RL in POMDPs with Latent Determinism and Conditional Embeddings
Masatoshi Uehara (Cornell University), Wen Sun (Cornell University)
Computational EfficiencyReinforcement LearningTabular
🎯 What it does: This paper proposes a PAC learning algorithm for POMDPs with deterministic hidden state transitions and observations embedded in a conditional Hilbert space, capable of obtaining optimal policies with polynomial sample and computational complexity in large-scale or continuous state/observation spaces.
Concept-based Explanations for Out-of-Distribution Detectors
Jihye Choi (University of Wisconsin), Atul Prakash (University of Michigan)
Anomaly DetectionExplainability and InterpretabilityConvolutional Neural NetworkAuto EncoderImage
🎯 What it does: This paper proposes an unsupervised explanation framework based on high-level concepts to explain the decisions of black-box OOD detectors and provides corresponding concept importance scores.
ConCerNet: A Contrastive Learning Based Framework for Automated Conservation Law Discovery and Trustworthy Dynamical System Prediction
Wang Zhang (Massachusetts Institute of Technology), Lam M. Nguyen (IBM Research)
OptimizationRepresentation LearningAuto EncoderContrastive LearningTime SeriesPhysics Related
🎯 What it does: By comparing system trajectories through contrastive learning, conserved quantities are automatically discovered, and a dynamical network with conservation properties is constructed.
Concurrent Shuffle Differential Privacy Under Continual Observation
Jay Tenenbaum (Google Research), Uri Stemmer (Blavatnik School of Computer Science, Tel Aviv University)
Safty and Privacy
🎯 What it does: This paper proposes a new Concurrent Shuffle Model and presents an algorithm for the Private Sum with Continuous Observations (PSCO) problem under this model, achieving an error rate that can reach polylog. It also proves that this error rate matches both the lower and upper bounds for any number of concurrent shuffles k.
Conditional Graph Information Bottleneck for Molecular Relational Learning
Namkyeong Lee (KAIST), Chanyoung Park (KAIST)
Drug DiscoveryGraph Neural NetworkContrastive LearningGraph
🎯 What it does: This paper proposes a molecular relationship learning framework based on Conditional Graph Information Bottleneck (CGIB), which can automatically detect and utilize the minimal core subgraph required for the interaction between two molecules when given a pair of molecules, thereby predicting the interaction behavior of the molecular pair.
Conditional Tree Matching for Inference-Time Adaptation of Tree Prediction Models
Harshit Varma (Indian Institute of Technology Bombay), Sunita Sarawagi (Indian Institute of Technology Bombay)
TransformerText
🎯 What it does: CTREEOT is proposed, a differentiable and convergent conditional tree matching algorithm that adapts tree prediction models using a small number of cases during inference.
Conditionally Strongly Log-Concave Generative Models
Florentin Guth (Ecole Normale Superieure), Stéphane Mallat (Collège de France)
GenerationScore-based ModelImageMultimodalityPhysics Related
🎯 What it does: This paper proposes a Conditional Strong Log-Convex (CSLC) generative model that decomposes high-dimensional data into several conditional log-convex sub-distributions through adaptive orthogonal projection, achieving theoretically guaranteed learning and sampling.
Cones: Concept Neurons in Diffusion Models for Customized Generation
Zhiheng Liu (University of Science and Technology of China), Yang Cao (University of Science and Technology of China)
GenerationData SynthesisTransformerDiffusion modelImage
🎯 What it does: Locate and utilize a small cluster of 'concept neurons' in pre-trained text-to-image diffusion models to achieve parameter-free generation of multiple subjects.
Confidence and Dispersity Speak: Characterizing Prediction Matrix for Unsupervised Accuracy Estimation
Weijian Deng (Australian National University), Liang Zheng (Australian National University)
Domain AdaptationContrastive LearningImage
🎯 What it does: Evaluate the model's accuracy under distribution shift without labels, studying the confidence and dispersion of the prediction matrix and unifying their quantification.
Conformal Inference is (almost) Free for Neural Networks Trained with Early Stopping
Ziyi Liang (University of Southern California), Matteo Sesia (University of Southern California)
ClassificationAnomaly DetectionConvolutional Neural NetworkImageTabular
🎯 What it does: This paper proposes a method that combines early stopping with conformal inference, called Conformalized Early Stopping (CES). It achieves reliable predictions with no additional data partitioning by completing model selection and calibration on the same hold-out sample, making it nearly 'free'.
Conformal Prediction for Federated Uncertainty Quantification Under Label Shift
Vincent Plassier (Ecole Polytechnique), Maxim Panov (Technology Innovation Institute)
Federated LearningSafty and PrivacyImage
🎯 What it does: The DP-FedCP method is proposed in the context of federated learning, utilizing distributed non-iid label distributions to construct a credible ensemble prediction set.
Conformal Prediction Sets for Graph Neural Networks
Soroush H. Zargarbashi (CISPA Helmholtz Center for Information Security), Aleksandar Bojchevski (University of Cologne)
ClassificationComputational EfficiencyGraph Neural NetworkDiffusion modelGraph
🎯 What it does: A distribution-independent synthetic prediction set (DAPS) method is proposed to provide reliable confidence sets for Graph Neural Networks (GNNs).
Conformal Prediction with Missing Values
Margaux Zaffran (INRIA), Yaniv Romano (Technion Israel Institute of Technology)
TabularBiomedical Data
🎯 What it does: This paper proposes a method to construct prediction intervals using missing value imputation + CP (split CP) in the context of missing covariates, and explores its performance in terms of marginal coverage and mask condition coverage (MCV).
Conformalization of Sparse Generalized Linear Models
Etash Kumar Guha (Georgia Institute of Technology), Xiaoming Huo (Georgia Institute of Technology)
OptimizationTabular
🎯 What it does: A method is proposed to efficiently construct conformal prediction sets for sparse generalized linear models using numerical continuity (homotopy).
Consistency Models
Yang Song (OpenAI), Ilya Sutskever (OpenAI)
GenerationKnowledge DistillationDiffusion modelScore-based ModelImageOrdinary Differential Equation
🎯 What it does: A new class of generative models called consistency models is proposed, which can directly map noise to data in one or few steps and supports zero-shot editing.
Consistency of Multiple Kernel Clustering
Weixuan Liang (National University of Defense Technology), En Zhu (National University of Defense Technology)
OptimizationTabular
🎯 What it does: This paper conducts a theoretical study on the consistency of kernel weights in Multiple Kernel Clustering (MKC) and proposes a scalable Large-Scale SimpleMKKM algorithm, providing non-asymptotic consistency upper bounds, difference upper bounds, and generalization risk upper bounds.
Constant Matters: Fine-grained Error Bound on Differentially Private Continual Observation
Hendrik Fichtenberger (Google), Jalaj Upadhyay (Rutgers University)
Safty and PrivacyGaussian SplattingTime Series
🎯 What it does: This paper studies the differential privacy counting problem under continuous observation models, providing an explicit constant upper bound on the error and offering a corresponding lower triangular decomposition.
Constrained Causal Bayesian Optimization
Virginia Aglietti (DeepMind), Silvia Chiappa (DeepMind)
OptimizationGraphTabularElectronic Health Records
🎯 What it does: This paper proposes a constrained causal Bayesian optimization (cCBO) method under a known causal graph, aimed at finding the optimal intervention set that satisfies multiple constraints to optimize the expected value of the target variable.
Constrained Decision Transformer for Offline Safe Reinforcement Learning
Zuxin Liu (Carnegie Mellon University), Ding Zhao (Carnegie Mellon University)
Robotic IntelligenceTransformerReinforcement LearningSequential
🎯 What it does: This paper studies the problem of offline safe reinforcement learning, proposing the concept of ε-approximation to evaluate task difficulty, and designing a method based on the Transformer Decision Transformer (CDT) that can learn adjustable safety constraint policies from offline data.
Constrained Efficient Global Optimization of Expensive Black-box Functions
Wenjie Xu (Automatic Control Laboratory EPFL), Colin Jones
OptimizationTabular
🎯 What it does: The CONFIG algorithm is proposed to solve the efficient global optimization problem of constrained expensive black-box functions.
Constrained Monotonic Neural Networks
Davor Runje (Airt Research), Sharath M Shankaranarayana (Algebra University)
ClassificationOptimizationTabular
🎯 What it does: A Monotonic Dense Unit is proposed, which can approximate any continuous monotonic function without increasing additional parameters or training steps by using three derived activation functions (original, concave, saturated) in the hidden layers and applying sign constraints on the weights.
Constrained Optimization via Exact Augmented Lagrangian and Randomized Iterative Sketching
Ilgee Hong (University of Chicago), mladen kolar
Optimization
🎯 What it does: An adaptive randomized quasi-Newton (AdaSketch-Newton) algorithm is proposed for solving non-convex equality constrained optimization problems. The algorithm solves the Lagrangian Newton equations at each step using randomized iterative projection and selects the step size through an exact augmented Lagrangian penalty function.
Constrained Phi-Equilibria
Martino Bernasconi (Politecnico di Milano), Nicola Gatti (Politecnico di Milano)
OptimizationComputational EfficiencyTabular
🎯 What it does: This paper introduces and computationally describes the constrained Phi equilibrium, which is a more general concept than the constrained correlated equilibrium, applicable to normal form games.
Context Consistency Regularization for Label Sparsity in Time Series
Yooju Shin (KAIST), Byung Suk Lee (University of Vermont)
ClassificationTime Series
🎯 What it does: A cross-window consistency regularization framework called CrossMatch is proposed, based on context-augmented enhancement, for time series classification with sparse labels.
Context-Aware Bayesian Network Actor-Critic Methods for Cooperative Multi-Agent Reinforcement Learning
Dingyang Chen (University of South Carolina), Qi Zhang (University of South Carolina)
Reinforcement Learning
🎯 What it does: This paper proposes the use of Bayesian Networks (BN) in cooperative multi-agent reinforcement learning to introduce correlations between agent actions. Under this framework, it derives the gradient formula for BN combination strategies and proves that gradient ascent under tabular softmax parameterization can globally converge to Nash and optimal fully correlated strategies. It then presents a practical algorithm for differentiable DAG learning, which can transform any multi-agent actor-critic method (such as MAPPO) into a BN combination strategy and implement it in partially observable scenarios.
Contextual Combinatorial Bandits with Probabilistically Triggered Arms
Xutong Liu (Chinese University of Hong Kong), Wei Chen (Microsoft Research)
Recommendation SystemOptimizationTabular
🎯 What it does: This study investigates context-combinatorial multi-armed bandits with probabilistic triggering (C2 MAB-T), proposing two new algorithms and providing improved theoretical bounds and experimental validation.
Contextual Conservative Interleaving Bandits
Kei Takemura (NEC Corporation)
OptimizationReinforcement LearningTabular
🎯 What it does: A meta-algorithm GCW is proposed to address the context-constrained interleaved multi-armed bandit (CCIB) problem, which combines time-varying features, a selectable action set, baseline actions, and stage constraints.
Contextual Reliability: When Different Features Matter in Different Contexts
Gaurav Rohit Ghosal (University of California), Aditi Raghunathan (Carnegie Mellon University)
Autonomous DrivingConvolutional Neural NetworkReinforcement LearningImageTabular
🎯 What it does: This paper proposes a new framework called 'Contextual Reliability' to address the issue of using different features for models in different environments/contexts, implemented through a two-stage Explicit Non-spurious feature Prediction (ENP) method;