International Conference on Machine Learning Β· 421 papers
Disentangled Multi-Fidelity Deep Bayesian Active Learning
Dongxia Wu (University of California San Diego), Rose Yu (University of California San Diego)
CodeTabularPhysics Related
π― What it does: A multi-fidelity deep Bayesian active learning framework D-MFDAL is proposed to learn high-fidelity simulation models from multi-fidelity data.
π― What it does: Proposes an unsupervised multi-graph representation learning framework that explicitly separates public information from private information and enhances the complementarity of private information through contrastive learning.
Divide and Conquer Dynamic Programming: An Almost Linear Time Change Point Detection Methodology in High Dimensions
Wanshan Li (Carnegie Mellon University), Alessandro Rinaldo (Carnegie Mellon University)
CodeAnomaly DetectionOptimizationTime SeriesBiomedical DataFinance Related
π― What it does: An algorithm named Divide and Conquer Dynamic Programming (DCDP) is proposed for efficiently detecting and locating multiple change points in high-dimensional time series, covering various scenarios such as mean, linear regression, and Gaussian graphical models.
Do Machine Learning Models Learn Statistical Rules Inferred from Data?
Aaditya Naik (University of Pennsylvania), Eric Wong (University of Pennsylvania)
CodeDomain AdaptationAnomaly DetectionAutonomous DrivingOptimizationBiomedical DataFinance Related
π― What it does: This paper proposes a method for automatically generating statistical quantile rules and achieving unsupervised model adaptation through violation penalties during testing.
π― What it does: A dynamic SGD step size formula without learning rate parameters (Distance over Gradients, DOG) is proposed, achieving parameter-independent adaptive updates;
Domain Adaptation for Time Series Under Feature and Label Shifts
Huan He (Harvard University), Marinka Zitnik (Harvard University)
CodeDomain AdaptationConvolutional Neural NetworkMultimodalityTime Series
π― What it does: The RAINCOAT method is proposed for unsupervised domain adaptation of time series, addressing the challenges of feature shift and label shift.
π― What it does: A Doubly Adversarial Federated Bandits framework is proposed, and the FEDEXP3 algorithm is designed to achieve sublinear regret convergence in a distributed manner without exchanging raw information.
DP-Fast MH: Private, Fast, and Accurate Metropolis-Hastings for Large-Scale Bayesian Inference
Wanrong Zhang (Harvard University), Ruqi Zhang
CodeSafty and PrivacyComputational EfficiencyTabular
π― What it does: A Metropolis-Hastings algorithm called DP-Fast MH is proposed, which can achieve differential privacy in large-scale Bayesian inference while using only small batches of data in most iterations.
DRew: Dynamically Rewired Message Passing with Delay
Benjamin Gutteridge (University of Oxford), Francesco Di Giovanni (University of Cambridge)
CodeGraph Neural NetworkGraph
π― What it does: This paper proposes a dynamic reconnection message passing framework DRew (and its delayed version Ξ½DRew), which gradually expands the connectivity of the graph at each layer, allowing nodes to interact layer by layer according to distance, thereby alleviating the over-compression problem of traditional MPNNs.
DSGD-CECA: Decentralized SGD with Communication-Optimal Exact Consensus Algorithm
Lisang Ding (University of California), Wotao Yin (Alibaba US)
CodeOptimizationFederated LearningTabular
π― What it does: A new decentralized stochastic gradient descent algorithm, DSGD-CECA, is proposed, which allows an arbitrary number of agents to collaborate in training, overcoming the limitations of previous algorithms regarding the number of agents.
π― What it does: This paper proposes Dual Focal Loss, which simultaneously considers the logit of the true label and the second largest logit during the training of classification models to balance overconfidence and underconfidence, thereby improving the model's probability calibration performance.
π― What it does: A Dual Propagation algorithm for Dyadic Neurons is proposed, utilizing single-stage closed-form layerwise inference to achieve contrastive Hebbian learning.
π― What it does: This paper proposes DualHSIC, a general loss that can be seamlessly integrated into existing replay-based continual learning methods to alleviate task interference and enhance task-invariant knowledge sharing.
π― What it does: Proposes the DUET method, which utilizes transformation-variant and content-variant information in 2D structured representation learning MSSL.
π― What it does: This paper proposes a dynamics-based neural network architecture (DyN) that transforms traditional weight networks into a form represented solely by the dynamic states of neurons and path integrals, and learns through entropy reduction.
π― What it does: This paper proposes the Embedding Clustering Regularization Topic Model (ECRTM), which addresses the topic collapse issue by incorporating embedding clustering regularization into neural topic models.
Efficient and Equivariant Graph Networks for Predicting Quantum Hamiltonian
Haiyang Yu (Texas A&M University), Shuiwang Ji (Texas A&M University)
CodeComputational EfficiencyDrug DiscoveryGraph Neural NetworkGraphPhysics Related
π― What it does: Proposed and implemented the SE(3)-equivariant graph network QHNet for efficient and accurate prediction of molecular Hamiltonian matrices.
Efficient Approximations of Complete Interatomic Potentials for Crystal Property Prediction
Yuchao Lin (Texas A&M University), Shuiwang Ji (Texas A&M University)
CodeGraph Neural NetworkTabularPhysics Related
π― What it does: A crystal material property prediction method based on the complete summation of interatomic potential energies is proposedβPotNet.
Efficient displacement convex optimization with particle gradient descent
Hadi Daneshmand (Massachusetts Institute of Technology), Chi Jin (Princeton University)
CodeOptimization
π― What it does: This paper studies the convergence properties and complexity of Particle Gradient Descent (PGD) with a finite number of particles on displacement convex functions, providing theoretical upper bounds for achievable approximation errors and optimization errors. It also validates experimental results on instances such as energy distance and uniformly distributed neural networks on a two-dimensional circle.
Efficient Learning of Mesh-Based Physical Simulation with Bi-Stride Multi-Scale Graph Neural Network
Yadi Cao (University of California Los Angeles), Chenfanfu Jiang (University of California Los Angeles)
CodeGraph Neural NetworkMeshGraphPhysics Related
π― What it does: For large-scale grid physical simulations, a trainable and efficient multi-scale graph neural network framework has been constructed, allowing the model to maintain low latency during inference and enabling rapid updates at different time steps.
Efficient Personalized Federated Learning via Sparse Model-Adaptation
Daoyuan Chen (Alibaba Group), Yaliang Li (Alibaba Group)
CodeFederated LearningExplainability and InterpretabilityComputational EfficiencyImage
π― What it does: This paper proposes a framework for adaptively generating sparse personalized models in federated learning, called pFedGate, which utilizes a lightweight trainable gating layer to achieve model sparsification at the client level;
π― What it does: This paper proposes a defense model based on information discard and robust representation recovery, which destroys the spatial structure of noise by randomly generating complementary masks on adversarial samples, and utilizes adversarial games to guide the removal of non-robust features and recover robust representations.
π― What it does: In a feeding task with periodic lighting, the study investigates how deep reinforcement learning agents spontaneously generate and internalize circadian rhythms, thereby achieving prediction and adaptation to environmental cycles.
End-to-end Differentiable Clustering with Associative Memories
Bishwajit Saha (Rensselaer Polytechnic Institute), Parikshit Ram (IBM Research)
CodeTabular
π― What it does: A differentiable clustering algorithm ClAM is implemented using Associative Memory (AM) dynamics and energy functions, retaining hard cluster assignments.
π― What it does: This paper proposes an end-to-end framework for learning decision mappings directly from data and provides its Bayesian interpretation. Based on this interpretation, new training algorithms for empirical risk minimization (ERM) and distributionally robust optimization (DRO) are designed.
End-to-End Multi-Object Detection with a Regularized Mixture Model
Jaeyoung Yoo (NAVER WEBTOON AI), Nojun Kwak (Seoul National University)
CodeObject DetectionMixture of ExpertsImage
π― What it does: An end-to-end multi-object detection framework D-RMM based on a mixed model is proposed, which directly learns the target distribution through negative log-likelihood.
Enhancing Activity Prediction Models in Drug Discovery with the Ability to Understand Human Language
Philipp Seidl (Johannes Kepler University), GΓΌnter Klambauer (Johannes Kepler University)
CodeDrug DiscoveryTransformerContrastive LearningTextBiomedical Data
π― What it does: A model called CLAMP is proposed, which can achieve zero/few-shot activity prediction by understanding experimental text during inference.
π― What it does: This paper proposes a multi-agent reinforcement learning framework named EnDi, which utilizes natural language manuals and target instructions to represent environmental entities, allowing each agent to independently define sub-goals (self and others), thereby achieving entity-level sub-goal allocation and collaborative decision-making.
Equivariant Architectures for Learning in Deep Weight Spaces
Aviv Navon (Bar Ilan University), Haggai Maron (Nvidia)
CodeDomain AdaptationContrastive LearningImage
π― What it does: This paper studies and proposes a deep network weight space equivariant network architecture called DWSNet, which can directly handle sequences of network weights and is suitable for tasks such as weight editing and domain transfer.
π― What it does: A multi-task deep ensemble framework named CMDE is proposed to learn shared and group-specific information from control and treatment groups, thereby estimating the Conditional Average Treatment Effect (CATE).
Estimating Heterogeneous Treatment Effects: Mutual Information Bounds and Learning Algorithms
Xingzhuo Guo (Tsinghua University), Mingsheng Long (Tsinghua University)
CodeOptimizationRepresentation LearningAdversarial AttackGenerative Adversarial NetworkContrastive LearningTabularBiomedical Data
π― What it does: This paper proposes the use of mutual information to select bias and derives a new error bound, subsequently designing the MitNet algorithm to estimate heterogeneous treatment effects.
Estimating the Contamination Factor's Distribution in Unsupervised Anomaly Detection
Lorenzo Perini (KU Leuven), Arto Klami (University of Helsinki)
CodeAnomaly DetectionTabular
π― What it does: A method for estimating the posterior distribution of the contamination factor (anomaly ratio) in unsupervised anomaly detection is proposed.
Estimation Beyond Data Reweighting: Kernel Method of Moments
Heiner Kremer (Max Planck Institute for Intelligent Systems), Jia-Jie Zhu (Weierstrass Institute for Applied Analysis and Stochastics)
CodeOptimization
π― What it does: This paper proposes the Kernel Method of Moments (KMM), a method based on Maximum Mean Discrepancy (MMD), which breaks through the limitations of traditional methods that only reweight empirical distributions by approximating continuous distributions to satisfy (conditional) moment constraints.
π― What it does: Two unsupervised denoising evaluation metrics (uMSE, uPSNR) that can be computed based solely on noisy images are proposed and validated, with experiments conducted on various real and synthetic datasets.
π― What it does: A value function based on 'final discount' and an LTL counterfactual experience replay method are designed to learn policies that satisfy LTL specifications without the need for sparse rewards.
Exploring the Benefits of Training Expert Language Models over Instruction Tuning
Joel Jang (Korea Advanced Institute of Science and Technology), Minjoon Seo (Korea Advanced Institute of Science and Technology)
CodeTransformerLarge Language ModelSupervised Fine-TuningMixture of ExpertsTextRetrieval-Augmented Generation
π― What it does: This study investigates the generalization ability of Expert Language Models (Expert LM) on unseen tasks after single-task training, comparing it with Multi-Task Instruction Fine-Tuning Models (MT LM).
π― What it does: The research model targets the accessibility of indiscriminate data poisoning attacks and proposes a computable threshold Ο, demonstrating its threshold effect across different models and datasets.
π― What it does: This paper proposes a fast discretization inference framework called FaDIn for estimating multivariate Hawkes processes with finite support general parametric kernels.
FAENet: Frame Averaging Equivariant GNN for Materials Modeling
Alexandre AGM Duval, David Rolnick (McGill University)
CodeGraph Neural NetworkGraph
π― What it does: In the material modeling task, a general framework is proposed to achieve E(3) equivariance through Stochastic Frame Averaging (SFA), and within this framework, a lightweight, fast, and expressive graph neural network FAENet is designed; this network does not require structural constraints on symmetry and directly utilizes atomic relative position information; SFA enhances the model's symmetry at the data level.
Jaakko Peltonen (Tampere University), Jyrki Nummenmaa (Tampere University)
CodeRetrievalOptimizationExplainability and InterpretabilityTabular
π― What it does: Two information retrieval-based neighbor embedding methods, Fair-NeRV and Fair-t-NeRV, are proposed for achieving unbiased nonlinear dimensionality reduction and visualization.
FARE: Provably Fair Representation Learning with Practical Certificates
Nikola JovanoviΔ (ETH Zurich), Martin Vechev (ETH Zurich)
CodeRepresentation LearningAuto EncoderTabular
π― What it does: The FARE method is proposed, which uses a constrained encoder to generate fair representations and provides a confidence upper bound for any subsequent classifier (fairness proof).
π― What it does: Two sampling algorithms are proposed in distributed k-clustering (including outliers), significantly reducing local running time and eliminating dependence on instance aspect ratio.
Fast Combinatorial Algorithms for Min Max Correlation Clustering
Sami Davies (Northwestern University), Heather Newman (Carnegie Mellon University)
CodeOptimizationGraph
π― What it does: A fast combinatorial algorithm for Min Max correlation clustering is proposed, which can achieve a constant factor approximation on complete graphs and run in linear or near-linear time.
Fast Online Value-Maximizing Prediction Sets with Conformal Cost Control
Zhen Lin (University of Illinois), Jimeng Sun (University of Illinois)
CodeRecommendation SystemOptimizationTabularBiomedical DataElectronic Health Records
π― What it does: This paper proposes an online multi-label prediction framework called FavMac, which maximizes business value while satisfying user-defined cost constraints.
Fast Private Kernel Density Estimation via Locality Sensitive Quantization
Tal Wagner (Amazon), Nina Mishra (Amazon)
CodeSafty and PrivacyComputational EfficiencyTabular
π― What it does: A framework based on Local Sensitivity Quantization (LSQ) is proposed for efficiently achieving differential privacy kernel density estimation (DP-KDE) for high-dimensional data.
π― What it does: A DSNO network is constructed using neural operators and temporal convolution blocks, allowing for single-step prediction of the probability flow trajectory of the diffusion model, enabling sampling to be completed with a single forward pass of the model.
π― What it does: The FedCR framework is proposed in multi-client personalized federated learning, utilizing cross-client common representation learning and CMI regularization to enhance the generalization performance of local models.
FedDisco: Federated Learning with Discrepancy-Aware Collaboration
Rui Ye (Shanghai Jiao Tong University), Yanfeng Wang (Shanghai AI Laboratory)
CodeFederated LearningImageText
π― What it does: In response to the heterogeneity of class distribution in federated learning, a deviation-based aggregation weight method called FedDisco is proposed.
π― What it does: During the training process of neural networks, the author proposes using low-bit quantization of gradients (i.e., approximating the derivative of the activation function as a piecewise constant) to replace the storage of the complete input tensor, thereby significantly reducing memory usage.
π― What it does: This paper proposes Graph Complementary Learning (GOAL), which completes missing intra-class/inter-class edges and uses new graph convolution for node classification after completion.
FlexRound: Learnable Rounding based on Element-wise Division for Post-Training Quantization
Jung Hyun Lee (NAVER Cloud), Dongsoo Lee (NAVER Cloud)
CodeCompressionOptimizationTransformerLarge Language ModelImageText
π― What it does: In this study, the authors propose a learnable quantization rounding method called FlexRound based on element-wise division, aimed at post-training quantization (PTQ) to enhance the performance of quantized models.
Flipping Coins to Estimate Pseudocounts for Exploration in Reinforcement Learning
Sam Lobel (Brown University), George Konidaris (Brown University)
CodeReinforcement LearningImageSequential
π― What it does: By training a neural network to predict the square of the sample mean of the Rademacher distribution (coin toss), we estimate the pseudocount of states, thereby providing a count-based exploration reward for reinforcement learning agents.
π― What it does: The (Stochastic) Forward-Backward Gaussian Variational Inference (FB-GVI) algorithm is proposed to solve Gaussian variational inference problems in the Bures-Wasserstein space.
Free-Form Variational Inference for Gaussian Process State-Space Models
Xuhui Fan (University of Newcastle), Scott A Sisson
CodeTime SeriesBenchmark
π― What it does: A Gaussian Process State Space Model (GPSSM) inference method based on free-form variational inference (FFVD) is proposed, which can simultaneously capture the posterior correlation between state trajectories and inducing variables without making parametric assumptions.
From Hypergraph Energy Functions to Hypergraph Neural Networks
Yuxin Wang (Fudan University), David Wipf (Amazon)
CodeClassificationOptimizationExplainability and InterpretabilityGraph Neural NetworkGraph
π― What it does: A class of hypergraph learning framework based on energy functions is designed, and trainable node embeddings are obtained through a two-layer optimization, ultimately constructing PhenomNN and its simplified version PhenomNNsimple.
CodeRobotic IntelligenceTransformerLarge Language ModelReinforcement LearningTabular
π― What it does: A Pretrained Decision Transformer (PDT) model is constructed, utilizing future trajectory information for Transformer pre-training on reward-free offline data, followed by fine-tuning on reward-based tasks through a return prediction network, achieving efficient unsupervised pre-training and rapid adaptation to downstream tasks.
π― What it does: A Guided Adversarial Training (GAT) method is proposed to enhance the model's adversarial robustness by incorporating self-supervised or domain knowledge auxiliary tasks on a small amount of training data, and introducing gradient curvature regularization and multi-objective weight scheduling in adversarial training.
π― What it does: This paper proposes a Gaussian process prior (EPGP) constructed using the Ehrenpreis-Palamodov fundamental principle, which can generate samples that satisfy PDE constraints for any system of linear constant coefficient partial differential equations, and presents a sparse version (S-EPGP);
Arjun Karuvally (University of Massachusetts Amherst), Hava T Siegelmann
CodeSequential
π― What it does: Proposes the General Sequential Episodic Memory Model (GSEMM), which introduces delayed coupling and multiple time scales based on the Hopfield network to achieve a dynamic energy landscape that can store and retrieve memory sequences.
π― What it does: This paper introduces the concept of 'Generalization to Unseen Domains' (GOTU) and studies the reasoning ability of neural networks on unseen samples when part of the data distribution is completely ignored. Theoretical and experimental analyses are conducted on architectures such as random feature models, diagonal linear networks, and Transformers, revealing that they tend to learn minimum degree interpolators (MD interpolators) in unseen domains. A degree-based curriculum learning algorithm, Degree-Curriculum, is proposed to accelerate learning and improve length generalization.
Generalized Disparate Impact for Configurable Fairness Solutions in ML
Luca Giuliani (University of Bologna), Michele Lombardi (University of Bologna)
CodeOptimizationExplainability and InterpretabilityTabular
π― What it does: A configurable family of GeDI metrics is proposed to measure the functional dependency between continuously protected attributes and model outputs, and fairness constraints are implemented through an optimization framework.
π― What it does: A generative graph detection framework is proposed, defining four training/testing scenarios, and conducting binary classification experiments on generated graphs and real graphs using three models (end-to-end GCN, contrastive learning, metric learning).
Jianke Yang (University of California San Diego), Rose Yu (University of California San Diego)
CodeGenerationData SynthesisGraph Neural NetworkGenerative Adversarial NetworkTabularTime SeriesPhysics Related
π― What it does: This paper presents LieGAN, a framework based on Generative Adversarial Networks that can automatically discover the symmetries of continuous Lie groups and their discrete subgroups from data and generate corresponding Lie algebra bases.
Generative Causal Representation Learning for Out-of-Distribution Motion Forecasting
Shayan Shirahmad Gale Bagi (University of Waterloo), Mark Crowley (University of Waterloo)
CodeGenerationDomain AdaptationRepresentation LearningAuto EncoderGenerative Adversarial NetworkMultimodalityTime Series
π― What it does: A Generative Causal Representation Learning (GCRL) framework is proposed for human trajectory prediction in cross-domain and noisy environments.
π― What it does: A framework for offline black-box optimization called BONET is proposed, which uses a self-supervised generative model to learn the dynamics of the optimizer.
π― What it does: Using a pre-trained diffusion model as a data prior, GibbsDDRM is proposed to solve the blind linear inverse problem through a partially collapsed Gibbs sampler, which simultaneously estimates unknown data and unknown linear operator parameters.
Ali Hatamizadeh (NVIDIA), Pavlo Molchanov (NVIDIA)
CodeObject DetectionSegmentationTransformerImage
π― What it does: Proposes the GC ViT architecture, which integrates local window self-attention with global query tokens, combined with a convolutional token generator and an improved downsampling module, achieving efficient context modeling and parameter utilization.
π― What it does: Analyzed Elman-type recurrent neural networks (RNNs) and their training under mean-field conditions, demonstrating that the gradient descent training dynamics of RNNs converge to the corresponding mean-field equations in the width limit, and proving that under certain weight initialization assumptions, the fixed point of the infinite width limit dynamics is globally optimal.
π― What it does: Proposes Global Contrastive Batch Sampling (GCBS), which optimizes sample arrangement to approximate global contrastive loss, avoiding the high cost of hard negative mining.
GNN&GBDT-Guided Fast Optimizing Framework for Large-scale Integer Programming
Huigen Ye (Tsinghua University), Yu Jiang (Meituan)
CodeOptimizationGraph Neural NetworkGraph
π― What it does: A fast optimization framework based on multi-task GNN and GBDT is proposed to solve large-scale integer programming problems using small-scale optimizers.
GNOT: A General Neural Operator Transformer for Operator Learning
Zhongkai Hao (Tsinghua University), Jun Zhu (Tsinghua University)
CodeTransformerMixture of Experts
π― What it does: A general neural operator framework GNOT based on Transformer is proposed for learning the solution operator of PDEs, capable of handling irregular grids, multiple input functions, and multi-scale problems.
π― What it does: A scalable global Transformer (GOAT) is proposed for large-scale graph node classification tasks, capable of handling both homogeneous and heterogeneous graphs simultaneously.
π― What it does: The researchers proposed a gradient-driven Wang-Landau sampler to sample the output distribution of neural networks across the entire input space, thereby obtaining a complete output histogram and corresponding representative input samples.
GRAFENNE: Learning on Graphs with Heterogeneous and Dynamic Feature Sets
Shubham Gupta (Indian Institute of Technology Delhi), Srikanta J. Bedathur
CodeClassificationGraph Neural NetworkGraph
π― What it does: Proposes the GRAFENNE framework to address the learning and continual learning issues of graph neural networks under heterogeneous node features and dynamic changes.
Graph Neural Networks with Learnable and Optimal Polynomial Bases
Yuhe Guo (Renmin University of China), Zhewei Wei (Renmin University of China)
CodeGraph Neural NetworkGraph
π― 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.
π― 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.
π― 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.
CodeGenerationRetrievalTransformerLarge Language ModelVision Language ModelImageTextMultimodalityRetrieval-Augmented Generation
π― 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)
CodeRobotic IntelligenceTransformerLarge Language ModelReinforcement LearningAgentic AIText
π― 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)
CodeTime SeriesPhysics Related
π― 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.
π― 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.
π― 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.
CodeExplainability and InterpretabilityComputational EfficiencyConvolutional Neural NetworkRecurrent Neural NetworkImageTabular
π― 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)
CodeOptimizationFederated LearningSafty and PrivacyComputational EfficiencyImage
π― 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.
π― 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.
π― 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-dimensional Location Estimation via Norm Concentration for Subgamma Vectors
Shivam Gupta (University of Texas at Austin), Eric Price (University of Texas at Austin)
Code
π― 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;
Sean R. Sinclair (Cornell University), Adith Swaminathan (Microsoft Research)
CodeOptimizationReinforcement LearningTabularTime Series
π― What it does: This paper studies the Exogenous Markov Decision Process (Exo-MDP) and proposes a framework for accelerating learning using retrospective informationβHindsight Learning (HL).
π― 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.
π― 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).
Identifiability and Generalizability in Constrained Inverse Reinforcement Learning
Andreas Schlaginhaufen (Ecole Polytechnique Federale de Lausanne), Maryam Kamgarpour (Ecole Polytechnique Federale de Lausanne)
CodeOptimizationReinforcement Learning
π― 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.
ILLUME: Rationalizing Vision-Language Models through Human Interactions
Manuel Brack (German Center for Artificial Intelligence), Kristian Kersting (German Center for Artificial Intelligence)
CodeExplainability and InterpretabilityTransformerLarge Language ModelPrompt EngineeringVision Language ModelImageTextMultimodality
π― 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.
π― 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 Expert Predictions with Conformal Prediction
Eleni Straitouri (Max Planck Institute for Software Systems), Manuel Gomez Rodriguez (Max Planck Institute for Software Systems)
CodeClassificationRecommendation SystemImage
π― 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.