π― What it does: A framework for constructing confidence intervals and eliminating uncertainty for over-parameterized neural networks is proposed, which is centered on training the network only twice (the base network and the artificial label network). The Procedural-Noise-Correcting (PNC) predictor is used to achieve procedural noise removal; combined with lightweight resampling (batching and cheap bootstrap), it results in confidence intervals with controllable statistical coverage.
Efficiently incorporating quintuple interactions into geometric deep learning force fields
Zun Wang (Microsoft Research AI4Science), Bin Shao (Microsoft Research AI4Science)
CodeGraph Neural NetworkGraphPhysics Related
π― What it does: This paper studies a graph neural network named QuinNet, which can efficiently and explicitly incorporate five-body interaction force field models.
π― What it does: By introducing the Anomaly Adversarial Example Regularization (AAER) method, catastrophic overfitting (CO) in single-step adversarial training is eliminated, and model robustness is enhanced.
π― What it does: This paper proposes the Domain Bias Eliminator (DBE) framework, which eliminates representation bias and degradation issues caused by uneven data domains in federated learning by introducing two modules: Personalized Representation Bias Memory (PRBM) and Mean Regularization (MR) into the local model.
Emergent and Predictable Memorization in Large Language Models
Stella Biderman (Booz Allen Hamilton), Edward Raff (Stability AI)
CodeTransformerLarge Language ModelText
π― What it does: This paper studies how to predict whether a model will remember specific training data by observing the memory performance of smaller models or intermediate checkpoints before training large language models.
π― What it does: EMMA-X is proposed, a cross-language pre-training method based on the EM framework, which learns universal sentence representations by utilizing a large amount of non-parallel multilingual data through bidirectional supervision of a GMM classifier and a cross-language encoder.
π― What it does: This paper proposes and implements an adversarial InfoNCE loss (AdvInfoNCE) for collaborative filtering (CF), enhancing the robustness and generalization ability of recommendation models through fine-grained hardness learning of difficult negative samples.
π― What it does: An Energy Discrepancy (ED) loss is proposed for training energy models without the need for MCMC or gradient information, allowing training in both Euclidean and discrete spaces.
π― What it does: This paper proposes a readout layer based on a visual attention mechanism, combined with Energy-Guided Diffusion (EGG) technology, to generate more natural and cross-model generalizable Most Exciting Images (MEI) and image reconstructions.
π― What it does: The Energy Transformer (ET) architecture is proposed, which combines attention mechanisms, energy models, and associative memory. By designing an energy function, the Transformer block is transformed into a recursive energy descending process for tasks such as image completion, graph anomaly detection, and graph classification.
π― What it does: A systematic comparison of seven energy-based learning algorithms (CL, P-EP, N-EP, C-EP, P-CpL, N-CpL, C-CpL) on deep convolutional Hopfield networks is conducted, proposing asynchronous energy minimization and low-precision solving, achieving faster and more accurate results than existing methods.
π― What it does: An anomaly detection method based on an energy model, MPDR, is proposed. By projecting diffusion perturbations on a low-dimensional manifold learned by an autoencoder and training the EBM using recovery likelihood maximization, the anomaly detection performance for various data types is significantly improved.
Khai Nguyen (University of Texas at Austin), Nhat Ho (University of Texas at Austin)
CodeData SynthesisOptimizationImagePoint Cloud
π― What it does: An energy-based slice Wasserstein distance (EBSW) is proposed, which adaptively selects the projection direction without optimization by setting the slice distribution to be a non-parametric distribution proportional to the energy of the one-dimensional Wasserstein distance.
π― What it does: Proposed and implemented Adversarial Invariant Regularization based on Causal Inference (AIR), which forces the model to be insensitive to style factors in adversarial contrastive learning, thereby obtaining more robust and transferable representations.
Enhancing Knowledge Transfer for Task Incremental Learning with Data-free Subnetwork
Qiang Gao (Southwestern University of Finance and Economics), Fan Zhou (University of Electronic Science and Technology of China)
CodeClassificationKnowledge DistillationImage
π― What it does: This paper proposes a data-independent subnetwork (DSN) method for task incremental learning, achieving forward and backward knowledge transfer through neuron-level masking and data-independent replay, thereby avoiding catastrophic forgetting.
Enhancing Minority Classes by Mixing: An Adaptative Optimal Transport Approach for Long-tailed Classification
Jintong Gao (Jilin University), Dan dan Guo
CodeClassificationContrastive LearningImage
π― What it does: An adaptive image mixing method based on optimal transport, OTmix, is proposed to enhance the performance of minority classes in long-tail classification.
π― What it does: Proposed the VaSSO method to stabilize the adversarial perturbations of the Sharpness-Aware Optimizer (SAM) by suppressing variance;
π― What it does: The FAPAT framework is designed to enhance anonymous session sequence encoding using frequent attribute graph patterns, thereby improving user intent capture and next item prediction in session recommendations.
Ensemble-based Deep Reinforcement Learning for Vehicle Routing Problems under Distribution Shift
Yuan Jiang (Nanyang Technological University), Jie Zhang (Nanyang Technological University)
CodeOptimizationReinforcement LearningTabular
π― What it does: A set-based deep reinforcement learning framework (EL-DRL) is proposed to solve vehicle routing problems (TSP, CVRP) under distribution drift.
π― What it does: An end-to-end neural network algorithm based on the SchrΓΆdinger bridge is proposed to solve the entropy-regularized optimal transport (EOT) plan between continuous probability distributions, supporting small entropy coefficients and capable of being trained in one go.
Entropy-based Training Methods for Scalable Neural Implicit Samplers
Weijian Luo (Peking University), Zhihua Zhang (Peking University)
CodeScore-based ModelImageTabular
π― What it does: A trainable neural implicit sampler is proposed, capable of sampling directly from an unnormalized target distribution through a single forward pass.
Environment-Aware Dynamic Graph Learning for Out-of-Distribution Generalization
Haonan Yuan (Beihang University), Jianxin Li (Beihang University)
CodeDomain AdaptationRecommendation SystemGraph Neural NetworkAuto EncoderGraphTime Series
π― What it does: Proposes the EAGLE framework to achieve adaptive generalization to distribution drift on dynamic graphs, targeting node-level future link prediction tasks.
π― What it does: A new decentralized learning algorithm called Epidemic Learning (EL) is proposed, which achieves dynamic topology through random sampling and communication with several nodes in each round;
Episodic Multi-Task Learning with Heterogeneous Neural Processes
Jiayi Shen (University of Amsterdam), Marcel Worring (University of Amsterdam)
CodeMeta LearningTransformerTabular
π― What it does: This paper proposes a heterogeneous neural process (HNPs) aimed at combining meta-learning and multi-task learning, which simultaneously handles multiple related and heterogeneous tasks in each meta-training/testing round to address the issue of insufficient data.
Fangxin Wang (University of Illinois Chicago), Philip S. Yu (University of Illinois Chicago)
CodeTabular
π― What it does: This paper proposes a new fairness metricβEqual Opportunity Coverage (EOC) in the context of uncertain regression, and presents a post-processing method BFQR, which aims to achieve consistent coverage within bins while maintaining overall coverage and minimizing the width of prediction intervals.
π― What it does: By adding a learnable normalization network in front of a pre-trained large model, it achieves equivariance to specific transformations (such as rotation) while maintaining original performance;
π― What it does: A geometric generative model based on flow matching, EquiFM, is proposed to simultaneously generate molecular atom types and three-dimensional coordinates, addressing the issues of probabilistic dynamics instability and slow sampling speed in traditional diffusion models.
Equivariant Neural Operator Learning with Graphon Convolution
Chaoran Cheng (University of Illinois Urbana-Champaign), Jian Peng (University of Illinois Urbana-Champaign)
CodeGraph Neural NetworkGraphPhysics Related
π― What it does: A covariant neural operator that combines coefficient learning and coordinate residual layers is proposed, capable of learning mappings between continuous functions in 3D Euclidean space while ensuring SE(3) covariance.
π― What it does: This paper proposes Equivariant Probabilistic Neural Simulation (EPNS), a method for stochastic spatiotemporal dynamic autoregressive probabilistic simulation that maintains symmetry constraints.
Error Discovery By Clustering Influence Embeddings
Fulton Wang (Meta), Narine Kokhlikyan (Meta)
CodeAnomaly DetectionExplainability and InterpretabilityImageTextTime SeriesBiomedical Data
π― What it does: This paper proposes a slice discovery method based on influence embedding, called InfEmbed, to identify subsets of samples in the test set that perform poorly and share the same error causes.
ESSEN: Improving Evolution State Estimation for Temporal Networks using Von Neumann Entropy
Qiyao Huang (Xiamen University), Edwin Hancock (University of York)
CodeGraph Neural NetworkMixture of ExpertsContrastive LearningGraphTime Series
π― What it does: The ESSEN framework is proposed, utilizing approximate von Neumann entropy and thermodynamic temperature to sense the evolving state of time-varying networks, and achieving node representation and link prediction through entropy-aware attention, virtual evolution node learning, and a Mixture of Thermodynamic Experts decoder.
π― What it does: Three NystrΓΆm random projection-based Koopman operator kernel methods (KRR, PCR, RRR) are proposed, achieving scalability for large-scale dynamical system learning.
Estimating Noise Correlations Across Continuous Conditions With Wishart Processes
Amin Nejatbakhsh (New York University), Alex H Williams (Flatiron Institute)
CodeTabularSequential
π― What it does: A probabilistic model based on the Wishart process is proposed and implemented to estimate the noise covariance of neurons under continuously parameterized experimental conditions, and it is capable of interpolation predictions under unseen conditions.
Evaluating Cognitive Maps and Planning in Large Language Models with CogEval
Ida Momennejad (Microsoft Research), Jonathan Larson (Microsoft Research)
CodeTransformerLarge Language ModelPrompt EngineeringTextGraph
π― What it does: The CogEval protocol is proposed for systematically evaluating the cognitive mapping and planning abilities of large language models.
Evaluating Neuron Interpretation Methods of NLP Models
Yimin Fan (Chinese University of Hong Kong), Hassan Sajjad (Dalhousie University)
CodeExplainability and InterpretabilityTransformerText
π― What it does: A systematic comparison of various neuron interpretation methods is conducted, and a voting compatibility-based evaluation framework is proposed.
Evaluating Post-hoc Explanations for Graph Neural Networks via Robustness Analysis
Junfeng Fang (University of Science and Technology of China), Xiangnan He (University of Science and Technology of China)
CodeAnomaly DetectionExplainability and InterpretabilityAdversarial AttackGraph Neural NetworkGraph
π― What it does: An evaluation metric OAR based on adversarial robustness and OOD reweighting is proposed to measure the reliability of subgraph explanations in graph neural networks.
Nino Scherrer (FAR AI), David Blei (Columbia University)
CodeTransformerLarge Language ModelPrompt EngineeringText
π― What it does: This paper studies the internal moral beliefs of large language models (LLMs) by designing and executing a large-scale moral scenario questionnaire, and quantifying the choices and uncertainties of LLMs using statistical methods.
CodeExplainability and InterpretabilityGraph Neural NetworkMultimodalityTime SeriesElectrocardiogram
π― What it does: A framework for assessing the robustness of interpretability methods based on model symmetry is proposed, defining invariance and equivariance metrics for explanations and providing improvement strategies.
π― What it does: A framework for evolutionary connectivity that uses only inference is proposed to train recurrent spiking neural networks (RSNN) with 1-bit sparse connections.
π― What it does: This paper proposes a continuous domain generalization method called EvoS, which addresses domain distribution that gradually drifts over time and achieves generalization to future domains;
Exact Optimality of Communication-Privacy-Utility Tradeoffs in Distributed Mean Estimation
Berivan Isik (Stanford University), Albert No (Hongik University)
CodeOptimizationFederated LearningSafty and PrivacyTabular
π― What it does: In scenarios such as federated learning, this study investigates distributed mean estimation under limited communication and local differential privacy (Ξ΅-LDP) constraints, and proposes an algorithm that achieves perfectly accurate optimal error.
Expanding Small-Scale Datasets with Guided Imagination
Yifan Zhang (National University of Singapore), Jiashi Feng (ByteDance)
CodeGenerationData SynthesisDiffusion modelImageBiomedical Data
π― What it does: This paper proposes a guided imagination framework (GIF) based on a prior generative model, aimed at automatically generating labeled new samples to expand datasets on small-scale datasets.
Expert load matters: operating networks at high accuracy and low manual effort
Sara Sangalli (ETH ZΓΌrich), Ender Konukoglu (ETH ZΓΌrich)
CodeClassificationOptimizationImageBiomedical Data
π― What it does: A new multi-class loss function called AUCOCLoss is proposed, aimed at simultaneously improving model accuracy and reducing the number of samples requiring human expert intervention.
Explain Any Concept: Segment Anything Meets Concept-Based Explanation
Ao Sun (Hong Kong University of Science and Technology), Shuai Wang (Hong Kong University of Science and Technology)
CodeSegmentationExplainability and InterpretabilityConvolutional Neural NetworkImage
π― What it does: The EAC (Explain Any Concept) method is proposed, which utilizes SAM to automatically extract instance concepts from images and provides efficient and interpretable concept-level explanations for the target DNN through Shapley values and a lightweight Per-Input Equivalence (PIE) surrogate.
Explainable Brain Age Prediction using coVariance Neural Networks
Saurabh Sihag (University of Pennsylvania), Alejandro Ribeiro (University of Pennsylvania)
CodeAnomaly DetectionExplainability and InterpretabilityGraph Neural NetworkBiomedical DataMagnetic Resonance ImagingAlzheimer's Disease
π― What it does: Using the Variational Neural Network (VNN) to predict age based on cortical thickness features, and by analyzing the distribution of model residuals across different brain regions, we obtain interpretable brain age differences (ββAge) and their association with clinical scores of Alzheimer's disease (AD).
Explaining Predictive Uncertainty with Information Theoretic Shapley Values
David Watson, Ido Guy (Meta)
CodeExplainability and InterpretabilityTextTabular
π― What it does: A framework based on information theory is proposed, utilizing an improved Shapley value to explain the uncertainty of model outputs, along with corresponding inference and implementation methods.
π― What it does: This paper constructs a reparameterization method that preserves Lipschitz constants, mapping various common first-order Lipschitz network structures (such as MonDEQ, SLL, CPL, AOL, etc.) to Lipschitz-bounded DEQ (LBEN), achieving an improvement in the provable robustness of DEQ models.
π― What it does: A framework named Point-In-Context (PIC) is proposed, which implements context learning for 3D point clouds and constructs a benchmark dataset that includes four tasks: reconstruction, denoising, registration, and part segmentation.
π― What it does: The ExpGen algorithm is proposed, achieving better generalization in zero-shot reinforcement learning through the integration of maximum entropy exploration and reward aggregation.
CodeOptimizationExplainability and InterpretabilityTabularBiomedical DataElectronic Health RecordsFinance Related
π― What it does: This paper studies the Rashomon set of sparse additive models (GAM), proposes an efficient approximation method based on the maximum volume ellipsoid, and utilizes this approximation to address interactive tasks such as variable importance, monotonicity constraints, and model editing.
Exploring Diverse In-Context Configurations for Image Captioning
Xu Yang (Southeast University), Xin Geng (Chinese University of Hong Kong)
CodeGenerationRetrievalTransformerVision Language ModelImageText
π― What it does: In the image description task, a systematic exploration of the impact of multimodal context configurations is conducted, proposing four image selection and four caption allocation strategies to construct few-shot prompt sequences.
π― What it does: This paper studies the relationship of loss functions under time-based training strategies, proving that rate-coded loss can be mapped to an equivalent time-coded form, and based on this, proposes a more suitable enhanced counting loss; at the same time, it transfers the scale parameter in weight normalization to threshold learning to further stabilize training.
π― What it does: Analyzed the differences in sampling quality between ODE (probability flow) and SDE (stochastic diffusion) in diffusion models, providing a theoretical derivation of error propagation under different noise intensities.
π― What it does: This paper evaluates 41 models based on four types of datasets (CIFAR10, ImageNet1k, FFHQ, LSUN-Bedroom) including diffusion, GAN, VAE, flow, Transformer, and consistency models through large-scale human experiments and multidimensional metrics. It systematically verifies the low correlation between traditional metrics (such as FID) and human real perception, and proposes that replacing Inception-V3 with a self-supervised feature extractor (especially DINOv2-ViT-L/14) significantly enhances the correlation between metrics and human evaluations.
Fabian Jogl (Vienna University of Technology), Thomas GΓ€rtner
CodeGraph Neural NetworkGraph
π― What it does: This paper studies how to simulate various non-standard message passing graph neural networks (GNNs) through graph transformations in standard message passing (MPNN) without losing expressiveness, proposing the concepts of strong/weak simulation and providing feasible automated transformation methods.
π― What it does: A foundational model for few-shot experimental design, ExPT, is proposed, which utilizes unlabeled data for synthetic pre-training and quickly adapts to the target function through contextual learning.
Extending the Design Space of Graph Neural Networks by Rethinking Folklore Weisfeiler-Lehman
Jiarui Feng (Washington University in St. Louis), Yixin Chen (Washington University in St. Louis)
CodeGraph Neural NetworkGraph
π― What it does: A general (k, t)-FWL+ framework is proposed, which extends the tuple aggregation method and neighborhood definition of k-FWL, and implements a practical N2-GNN under this framework, achieving an expressive power close to 3-WL with only O(nΒ²) space.
Extraction and Recovery of Spatio-Temporal Structure in Latent Dynamics Alignment with Diffusion Models
Yule Wang (Georgia Institute of Technology), Anqi Wu (Georgia Institute of Technology)
CodeDomain AdaptationTransformerDiffusion modelAuto EncoderTime Series
π― What it does: Proposes the ERDiff method, which extracts the complete spatiotemporal structure of latent dynamics using a diffusion model in the source domain, and then recovers this structure through maximum likelihood alignment in the target domain, achieving unsupervised neural distribution alignment.
π― What it does: Proposes the Extremal Transport (ET) theory for unpaired image domain translation, introducing an algorithm that can be approximated by Incomplete Transport (IT);
Qizhi Pei (Renmin University of China), Rui Yan (Renmin University of China)
CodeDrug DiscoveryProtein Structure PredictionGraph Neural NetworkBiomedical Data
π― What it does: An end-to-end protein-ligand binding prediction framework called FABind is proposed, which integrates pocket prediction and docking, capable of directly providing ligand binding sites and conformations.
FACE: Evaluating Natural Language Generation with Fourier Analysis of Cross-Entropy
Zuhao Yang (Nanyang Technological University), Kefan Chen (Nanyang Technological University)
CodeGenerationTransformerLarge Language ModelText
π― What it does: This paper proposes the FACE (Fourier Analysis of Cross-Entropy) metric, which quantifies the similarity in the frequency domain between the cross-entropy sequences of model-generated text and human text through a fast Fourier transform, thereby assessing the quality of natural language generation.
π― What it does: Reconstruct high-quality Face NeRF from a single face image using the frozen 3D GAN generator EG3D, and achieve semantically driven 3D editing and lighting reconstruction through text prompts.
Facilitating Graph Neural Networks with Random Walk on Simplicial Complexes
Cai Zhou (Tsinghua University), Muhan Zhang (Peking University)
CodeGraph Neural NetworkGraph
π― What it does: This paper proposes EdgeRWSE, a structure encoding based on edge-level random walks, and the first edge-level positional encoding Hodge1Lap based on the Hodge 1-Laplacian spectrum by extending random walks to k-dimensional simplices (including edges and higher-order simplices). It further introduces a cross-random walk framework that unifies various simplex-based GNN methods.
π― What it does: This paper studies the shortcomings of multimodal contrastive learning in scenarios with uneven distribution of task-related information (low shared, high unique) and proposes the FACTORCL method, which can separate and learn shared and unique task-related representations.
π― What it does: This paper studies the fairness issue in Canonical Correlation Analysis (CCA) and proposes a Fair CCA (F-CCA) framework that eliminates bias by minimizing the correlation difference error between different protected attribute groups, providing both multi-objective and single-objective optimization algorithms.
π― What it does: The FairLISA framework is proposed to achieve fair user modeling when only a subset of users possess sensitive attribute information.
Fairly Recommending with Social Attributes: A Flexible and Controllable Optimization Approach
Jinqiu Jin (University of Science and Technology of China), Xiangnan He (University of Science and Technology of China)
CodeRecommendation SystemOptimizationVideoTabular
π― What it does: Proposed a Neighborhood SP/NEO indicator for item-side group fairness (IGF) based on social attributes, and designed a multi-objective optimization algorithm SoFA that balances direct exposure and social benefits while maintaining recommendation accuracy.
Loukas Kavouras (Athena Research Center), Ioannis Emiris (National and Kapodistrian University of Athens)
CodeOptimizationExplainability and InterpretabilityComputational EfficiencyTabular
π― What it does: The FACTS framework is proposed, which uses counterfactual explanations to assess subgroup fairness and measure the difficulty of obtaining remedies.
Faith and Fate: Limits of Transformers on Compositionality
Nouha Dziri (Allen Institute for Artificial Intelligence), Yejin Choi (University of Washington)
CodeTransformerLarge Language ModelText
π― What it does: This paper systematically evaluates the performance of large Transformers on three typical combinatorial reasoning tasks (multi-digit multiplication, logic grid puzzles, dynamic programming subset sum) by constructing a computational graph of the algorithms, revealing that they often complete tasks through subgraph matching rather than true systematic reasoning during multi-step reasoning.
π― What it does: An efficient multi-task learning optimizer FAMO is proposed, capable of achieving balanced loss reduction for tasks in O(1) space and time.
π― What it does: Conduct large-scale experiments and comparisons on pruning criteria in dynamic sparse training (DST) to explore their impact on model performance, update frequency, and network structure.
Fast and Simple Spectral Clustering in Theory and Practice
Peter Macgregor (University of Edinburgh)
CodeOptimizationComputational EfficiencyGraph
π― What it does: A fast spectral clustering algorithm is proposed that uses power iteration to generate O(log k) dimensional random projections, avoiding the expensive computation of k eigenvectors in traditional algorithms.
π― What it does: A randomized algorithm utilizing kernel density estimation (KDE) as a black box has been designed and implemented, capable of quickly constructing a sparse similarity graph from an input point set while compressing the number of edges to nearly linear, all while preserving the clustering structure.
Fast Optimal Locally Private Mean Estimation via Random Projections
Hilal Asi (Apple Inc), Kunal Talwar (Apple Inc)
CodeSafty and PrivacyComputational EfficiencyTabular
π― What it does: This paper studies the problem of local private mean estimation for high-dimensional vectors and proposes a new algorithmic framework called ProjUnit, which can achieve efficient mean estimation while preserving privacy.
π― What it does: A new Wasserstein distance proxy called min-SWGG is proposed, which constructs the distance and corresponding transport plan by projecting the two distributions onto one dimension and matching the order after projection; a closed-form Wasserstein computation formula is also provided when the supporting distribution is on a line.
π― What it does: This paper proposes two improved versions of the Partition Learning Bloom Filter (PLBF), namely fast PLBF and fast PLBF++, which reduce the number of dynamic programming (DP) table constructions and utilize matrix monotonicity, lowering the construction time of the original PLBF from O(NkΒ³) to O(NkΒ²) and O(Nk log N + NkΒ²), while maintaining or closely approaching the original memory efficiency and false positive rate.
Fast Rank-1 Lattice Targeted Sampling for Black-box Optimization
Yueming Lyu (Agency for Science Technology and Research)
CodeOptimizationTabular
π― What it does: A fast target sampling method using random Rank-1 grids (RLTS) is proposed, which is embedded into the sampling phase of black-box optimization, achieving efficient GP training and inference with O(n log n) complexity, significantly improving query efficiency.
Fast Scalable and Accurate Discovery of DAGs Using the Best Order Score Search and Grow Shrink Trees
Bryan Andrews (University of Minnesota), Erich Kummerfeld (University of Minnesota)
CodeBiomedical DataMagnetic Resonance Imaging
π― What it does: A DAG learning algorithm BOSS based on permutation search is proposed, and a Grow-Shrink Tree (GST) cache structure is designed for efficient construction and scoring of DAGs, verifying its scalability and accuracy on large-scale highly connected variables (such as fMRI).
π― What it does: Proposes Fast Trainable Projection (FTP), an algorithm that achieves robust fine-tuning through learning layer-wise projection constraints.
Faster Discrete Convex Function Minimization with Predictions: The M-Convex Case
Taihei Oki (University of Tokyo), Shinsaku Sakaue (University of Tokyo)
CodeOptimization
π― What it does: A framework for predicting the acceleration of M-convex function minimization using machine learning is proposed and implemented, with improved time complexity specifically for Laminar, Nested, and Box problems.
π― What it does: A Feature Likelihood Divergence (FLD) evaluation metric is proposed to simultaneously measure the novelty, fidelity, and diversity of samples generated by models;
Feature Selection in the Contrastive Analysis Setting
Ethan Weinberger (University of Washington), Su-In Lee (University of Washington)
CodeAuto EncoderContrastive LearningBiomedical Data
π― What it does: This study investigates the feature selection problem in the context of contrastive analysis and proposes a two-stage contrastive autoencoder combined with a stochastic gate for differentiable feature selection (CFS).
π― What it does: In sample-free category incremental learning, the FeCAM method is proposed, which significantly improves performance by freezing the feature extractor, constructing class prototypes and covariance matrices, and using Mahalanobis distance for classification.
Fed-CO$_{2}$: Cooperation of Online and Offline Models for Severe Data Heterogeneity in Federated Learning
Zhongyi Cai (ShanghaiTech University), Jingya Wang (ShanghaiTech University)
CodeFederated LearningKnowledge DistillationImage
π― What it does: A unified federated learning framework, Fed-CO2, is proposed to simultaneously address two severe data heterogeneity issues: label distribution skew and feature skew. It achieves the integration of global knowledge and local exclusive knowledge through the collaboration of online and offline models.
Fed-GraB: Federated Long-tailed Learning with Self-Adjusting Gradient Balancer
Zikai Xiao (Zhejiang University), Zuozhu Liu (Zhejiang University)
CodeClassificationFederated LearningImage
π― What it does: Proposes the Fed-GraB framework, which combines a Global Long-Tail Prior Analyzer (DPA) and an Adaptive Gradient Balancer (SGB) to address the problem of Federated Long-Tail Learning (Fed-LT).
Xidong Wu (University of Pittsburgh), Heng Huang (University of Maryland)
CodeOptimizationFederated LearningTabular
π― What it does: This paper proposes a distributed algorithm for Federated Conditional Stochastic Optimization (FCSO), including FCSG, FCSG-M, and Acc-FCSG-M, and provides their theoretical convergence, sample complexity, and communication complexity.
π― What it does: This paper proposes a Bayesian personalized federated learning framework called Meta-Variational Dropout (MetaVD), which uses a hypernetwork to predict the dropout rate for each client, achieving model personalization and compression in non-IID and data-scarce scenarios.
Federated Learning with Bilateral Curation for Partially Class-Disjoint Data
Ziqing Fan (Shanghai Jiao Tong University), Yanfeng Wang (Shanghai Jiao Tong University)
CodeFederated LearningImage
π― What it does: This paper proposes a federated learning framework called FedGELA for the scenario of partially complete category data (PCDD), aiming to simultaneously improve the performance of global tasks and local personalization tasks.
π― What it does: EPISODE++ is proposed, a federated learning algorithm that achieves linear acceleration and lower communication costs under client sampling, data heterogeneity, and unbounded smoothness conditions;
π― What it does: The FedFed framework is proposed to alleviate data heterogeneity in Federated Learning through performance-sensitive feature sharing with feature separation and differential privacy protection.
FedGCN: Convergence-Communication Tradeoffs in Federated Training of Graph Convolutional Networks
Yuhang Yao (Carnegie Mellon University), Carlee Joe-Wong (Carnegie Mellon University)
CodeFederated LearningSafty and PrivacyComputational EfficiencyGraph Neural NetworkGraph
π― What it does: This paper proposes FedGCN, which uses a federated learning framework for semi-supervised node classification on a single large graph, significantly reducing communication overhead during the training process.
Royson Lee (University of Cambridge), Nicholas Donald Lane
CodeFederated LearningMeta LearningImageAudio
π― What it does: In the framework of federated learning, FedL2P is proposed to automatically generate batch normalization statistics weights and hierarchical learning rates for each client by learning two types of meta-networks (BNNet and LRNet), achieving personalized fine-tuning strategies.
FedNAR: Federated Optimization with Normalized Annealing Regularization
Junbo Li (Mohamed bin Zayed University of Artificial Intelligence), Hongyi Wang (Carnegie Mellon University)
CodeOptimizationFederated LearningImageText
π― What it does: The study investigates the impact of weight decay on convergence and generalization in federated learning, and proposes the FedNAR algorithm to adaptively control weight decay.
Few-Shot Class-Incremental Learning via Training-Free Prototype Calibration
Qi-Wei Wang (Nanjing University), Han-Jia Ye (Nanjing University)
CodeClassificationRecognitionImage
π― What it does: This paper proposes a prototype calibration method called TEEN, which does not require additional training, to address the issue of new classes being misclassified into old classes in few-shot class incremental learning.
Finding Counterfactually Optimal Action Sequences in Continuous State Spaces
Stratis Tsirtsis (Max Planck Institute for Software Systems), Manuel Gomez Rodriguez
CodeOptimizationReinforcement Learning from Human FeedbackReinforcement LearningTime SeriesBiomedical DataElectronic Health Records
π― What it does: This study investigates the problem of finding counterfactual optimal action sequences in continuous state spaces, constructing a framework based on continuous MDP and reversible structural causal models (SCM).
Finding Local Minima Efficiently in Decentralized Optimization
Wenhan Xian (University of Maryland), Heng Huang (University of Maryland)
CodeRecommendation SystemOptimizationTabular
π― What it does: A decentralized stochastic gradient algorithm named PEDESTAL is proposed, which can efficiently escape saddle points and find local optimal solutions for non-convex optimization problems.
Finding Order in Chaos: A Novel Data Augmentation Method for Time Series in Contrastive Learning
Berken Utku Demirel (ETH Zurich), Christian Holz (ETH Zurich)
CodeClassificationRecognitionData-Centric LearningAuto EncoderContrastive LearningTime Series
π― What it does: A custom mixup data augmentation method specifically designed for quasi-periodic time series data in self-supervised contrastive learning is proposed, avoiding the destruction of signal information caused by traditional mixup.
π― What it does: This paper proposes HC-Net, which uses a differentiable spherical transformation to project ground panoramic images into bird's-eye views, and directly aligns satellite images through a single iterative depth homomorphic estimator with relevant perception, achieving accurate GPS positioning and orientation.