Conference on Neural Information Processing Systems · 1376 papers
DesCo: Learning Object Recognition with Rich Language Descriptions
Liunian Harold Li (University of California Los Angeles), Kai-Wei Chang (University of California Los Angeles)
CodeRecognitionObject DetectionTransformerLarge Language ModelPrompt EngineeringVision Language ModelImageTextMultimodality
🎯 What it does: This paper proposes a language-supervised object detection method that utilizes rich language descriptions to train a vision-language model for better object recognition and localization.
Describe, Explain, Plan and Select: Interactive Planning with LLMs Enables Open-World Multi-Task Agents
Zihao Wang (Peking University), Yitao Liang (Peking University)
CodeTransformerLarge Language ModelAgentic AIPrompt EngineeringTextBenchmark
🎯 What it does: An interactive planning framework named DEPS is proposed to improve the execution of open-world tasks through description, explanation, planning, and selection.
🎯 What it does: The DESSERT algorithm is proposed and implemented for approximate search in set-to-set queries on vector collections, significantly improving retrieval speed.
Detecting Any Human-Object Interaction Relationship: Universal HOI Detector with Spatial Prompt Learning on Foundation Models
Yichao Cao (Southeast University), Chang Xu (University of Sydney)
CodeRecognitionObject DetectionTransformerLarge Language ModelPrompt EngineeringVision Language ModelImageText
🎯 What it does: Developed UniHOI, a general human-object interaction (HOI) detection framework that utilizes visual-language foundation models and large language models for spatial prompt learning, enabling the recognition of any interaction relationship.
🎯 What it does: Proposes the Diff-Instruct framework, which utilizes the knowledge of pre-trained diffusion models to guide the training of any implicit generative model in a data-free manner.
DIFFER:Decomposing Individual Reward for Fair Experience Replay in Multi-Agent Reinforcement Learning
Xunhan Hu (University of Science and Technology of China), Houqiang Li (University of Science and Technology of China)
CodeReinforcement LearningTabularBenchmark
🎯 What it does: A framework called DIFFER is constructed to decompose team rewards into individual rewards, utilizing individual TD-error to achieve fair experience replay, thereby enhancing the learning efficiency and fairness of multi-agent reinforcement learning.
🎯 What it does: A differentiable random partition model (DRPM) is proposed, which can end-to-end partition set elements into an unknown number of subsets and is applied to variational clustering, weakly supervised generative factor separation, and multi-task learning.
Differentiable Sampling of Categorical Distributions Using the CatLog-Derivative Trick
Lennert De Smet (KU Leuven), Pedro Zuidberg Dos Martires (Örebro University)
CodeOptimizationReinforcement LearningTabular
🎯 What it does: The CatLog-Derivative trick is proposed, and based on it, an unbiased gradient estimator for independent multivariate categorical distributions, called IndeCateR, is developed.
Differentiable sorting for censored time-to-event data.
Andre Vauvelle (University College London), Spiros Denaxas (University College London)
CodeTabularBiomedical DataElectronic Health Records
🎯 What it does: This paper proposes Diffsurv, a differentiable ranking method capable of handling censored time-to-event data for risk prediction in survival analysis.
Differentially Private Decoupled Graph Convolutions for Multigranular Topology Protection
Eli Chien (University of Illinois at Urbana-Champaign), Olgica Milenkovic (University of Illinois at Urbana-Champaign)
CodeSafty and PrivacyGraph Neural NetworkGraph
🎯 What it does: This paper proposes a Graph Differential Privacy (GDP) framework for graph learning and designs a Decoupled Graph Convolution (DPDGC) based on it to achieve multi-granularity graph topology protection.
CodeClassificationSafty and PrivacyRepresentation LearningGenerative Adversarial NetworkContrastive LearningImageBiomedical Data
🎯 What it does: Using synthetic images generated by stochastic processes to learn visual priors, combined with the three-stage differential privacy training framework DP-RandP, enhances the performance of private image classification.
CodeProtein Structure PredictionGraph Neural NetworkDiffusion modelBiomedical Data
🎯 What it does: We propose DiffPack, a method based on an autoregressive torsional diffusion model to predict protein side chain conformations, modeling distributions directly in the torsional angle space.
DiffTraj: Generating GPS Trajectory with Diffusion Probabilistic Model
Yuanshao Zhu (Southern University of Science and Technology), James Yu
CodeGenerationSafty and PrivacyConvolutional Neural NetworkDiffusion modelTime Series
🎯 What it does: This paper proposes a GPS trajectory generation method called DiffTraj based on a diffusion probability model, aiming to achieve privacy protection while maintaining the practicality of the trajectories.
🎯 What it does: This paper proposes and verifies the existence of a diffusion redundancy phenomenon in the hidden layers of pre-trained models, meaning that randomly selecting a sufficient number of neurons can approximate the complete representation of the layer and downstream performance.
Diffusion Model for Graph Inverse Problems: Towards Effective Source Localization on Complex Networks
Xin Yan (Northeastern University), Qiang He (Northeastern University)
CodeGraph Neural NetworkDiffusion modelGraph
🎯 What it does: A discrete diffusion denoising model DDMSL is proposed to simultaneously achieve source node localization and information propagation path recovery in graph inverse problems.
🎯 What it does: A three-stage Dual Pseudo Training (DPT) method is proposed, which utilizes a semi-supervised classifier to generate pseudo-labels, trains a conditional diffusion model to generate pseudo-images, and then retrains the classifier using these pseudo-images, thereby achieving few-label conditional generation and semi-supervised classification.
🎯 What it does: Proposes the MagDM framework, which uses magnetic transformation to map asymmetric kernels to Hermitian kernels, and defines corresponding integral operators, diffusion distances, and mappings to achieve dimensionality reduction for asymmetric data.
🎯 What it does: This paper proposes a new Schrödinger Bridge (SB) solving method—Iterative Markovian Fitting (IMF), and based on this method, introduces the Diffusion Schrödinger Bridge Matching (DSBM) algorithm, which can efficiently learn SB and generate samples;
🎯 What it does: A method for generating adversarial examples based on diffusion models, called Diff-PGD, is proposed, which maintains a high attack success rate while making adversarial examples more aligned with the distribution of natural images.
Diffusion-SS3D: Diffusion Model for Semi-supervised 3D Object Detection
Cheng-Ju Ho (National Yang Ming Chiao Tung University), Yi-Hsuan Tsai (Google)
CodeObject DetectionDiffusion modelPoint Cloud
🎯 What it does: This paper proposes Diffusion-SS3D, a method that utilizes diffusion models to generate high-quality pseudo-labels in semi-supervised 3D object detection. It is integrated into a teacher-student framework, using noise perturbation of object size and category distribution for denoising to obtain more reliable 3D bounding boxes.
Haoxing Chen (Ant Group), Weiqiang Wang (Ant Group)
CodeRecognitionImage TranslationGenerationLarge Language ModelDiffusion modelImageText
🎯 What it does: This paper proposes DiffUTE, a universal text editing method based on diffusion models, which can replace or modify text in any image while maintaining a realistic background.
🎯 What it does: A combinatorial optimization solver DIFUSCO based on graph diffusion models has been designed and implemented, capable of solving NP-complete problems such as TSP and MIS without relying on manual heuristics.
DIN-SQL: Decomposed In-Context Learning of Text-to-SQL with Self-Correction
Mohammadreza Pourreza (University of Alberta), Davood Rafiei (University of Alberta)
CodeTransformerLarge Language ModelPrompt EngineeringTextBenchmarkChain-of-Thought
🎯 What it does: This paper proposes a few-shot prompting method that decomposes the natural language to SQL task into four modules (schema linking, query classification and decomposition, SQL generation, and automatic error correction), significantly improving the execution accuracy of LLMs.
🎯 What it does: Developed DinoSR, which combines masked language modeling, self-supervised distillation, and online clustering to learn speech representations and generate discrete phoneme units.
🎯 What it does: This paper proposes an inference framework that directly uses the Direct Diffusion Bridge (DDB) in inverse problems, and incorporates data consistency correction to form the CDDB method.
Direct Preference Optimization: Your Language Model is Secretly a Reward Model
Rafael Rafailov (Stanford University), Chelsea Finn (Stanford University)
CodeOptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningText
🎯 What it does: A direct preference optimization (DPO) method is proposed that does not require reinforcement learning, using binary cross-entropy to directly train language models to meet human preferences.
Direct Preference-based Policy Optimization without Reward Modeling
Gaon An (Seoul National University), Hyun Oh Song (Seoul National University)
CodeOptimizationReinforcement Learning from Human FeedbackReinforcement LearningContrastive LearningSequential
🎯 What it does: An algorithm for learning strategies directly from human preference labels (DPPO) is proposed, eliminating the need to construct a reward model and optimizing the strategy directly through contrastive learning.
Direct Training of SNN using Local Zeroth Order Method
Bhaskar Mukhoty (Mohamed bin Zayed University of Artificial Intelligence), Bin Gu (Jilin University)
CodeOptimizationSpiking Neural NetworkImage
🎯 What it does: A direct training algorithm for SNN based on the local zero-order method (LOCALZO) is proposed, utilizing random sampling to approximate the gradient of the spike function, achieving gradient estimation and sparsification in backpropagation;
🎯 What it does: A new direction-oriented multi-objective optimization (direction-oriented MOO) target is proposed, and based on this target, two stochastic gradient algorithms, SDMGrad and SDMGrad-OS, are designed to solve Pareto stable points in non-convex scenarios.
🎯 What it does: A Directional Diffusion Model (DDM) is proposed for unsupervised graph representation learning, and its effectiveness is validated on various graph classification and node classification tasks.
🎯 What it does: This paper proposes an open-source semi-supervised learning framework named TIDA, which aims to learn multi-level classification contexts (from subcategories to target categories to supercategories) through self-supervised methods in scenarios where there are insufficient known category samples and unknown categories, thereby improving the quality of feature representation and pseudo-labels.
DISCOVER: Making Vision Networks Interpretable via Competition and Dissection
Konstantinos P. Panousis (Cyprus University of Technology), Sotirios Chatzis (Cyprus University of Technology)
CodeExplainability and InterpretabilityConvolutional Neural NetworkTransformerVision Language ModelImageMultimodality
🎯 What it does: By introducing a random local winner-takes-all (LWTA) layer in the visual network, activation sparsity is achieved, and then using CLIP-Dissect to provide textual descriptions for each neuron, a post-hoc explanation framework (DISCOVER) is constructed, enhancing the interpretability of the network;
Discovering General Reinforcement Learning Algorithms with Adversarial Environment Design
Matthew Thomas Jackson, Jakob Nicolaus Foerster
CodeMeta LearningReinforcement LearningTabular
🎯 What it does: The GROOVE method is proposed, which transfers Unsupervised Environment Design (UED) to Policy Meta Optimization (PMO), utilizing adversarial environment generation and adaptive curating to enhance the generalization ability of RL algorithms in unknown environments.
Discovering Hierarchical Achievements in Reinforcement Learning via Contrastive Learning
Seungyong Moon (Seoul National University), Hyun Oh Song (Seoul National University)
CodeReinforcement LearningContrastive Learning
🎯 What it does: This paper proposes a method called 'Achievement Distillation' that jointly uses contrastive learning during the reinforcement learning process, helping the model to learn and predict hierarchical achievements without an explicit planning module, thereby enabling efficient long-range exploration and achievement unlocking in procedurally generated environments.
Discriminative Calibration: Check Bayesian Computation from Simulations and Flexible Classifier
Yuling Yao (Flatiron Institute), Justin Domke (University of Massachusetts)
CodeClassificationOptimization
🎯 What it does: This paper proposes a classifier-based Bayesian inference diagnostic method—Discriminative Calibration, which replaces traditional ranking statistics with learnable features, directly constructs label mappings from simulated data, and trains classifiers to estimate the error and divergence of the posterior distribution.
DisDiff: Unsupervised Disentanglement of Diffusion Probabilistic Models
Tao Yang (Xi'an Jiaotong University), Nanning Zheng (Xi'an Jiaotong University)
CodeGenerationData SynthesisExplainability and InterpretabilityDiffusion modelImage
🎯 What it does: This paper proposes an unsupervised method called DisDiff, which performs factor disentanglement on pre-trained diffusion probabilistic models, automatically discovering and representing the intrinsic interpretable factors of images, and decomposing the diffusion gradient field into sub-gradient fields corresponding to each factor.
🎯 What it does: A Disentangled Counterfactual Learning (DCL) framework is proposed, which decomposes videos into static and dynamic factors and constructs contrastive learning and causal relationships to achieve physical audiovisual common sense reasoning.
Disentangling Cognitive Diagnosis with Limited Exercise Labels
Xiangzhi Chen (Hefei University of Technology), Meng Wang (Hefei University of Technology)
CodeExplainability and InterpretabilityRepresentation LearningAuto EncoderTabular
🎯 What it does: A model DCD for interpretable cognitive diagnosis is proposed, which can diagnose students' mastery of knowledge concepts even with a small amount of labeled practice problems and in the absence of a complete Q matrix.
Dissecting Chain-of-Thought: Compositionality through In-Context Filtering and Learning
Yingcong Li (University of California), Samet Oymak (University of Michigan)
CodeTransformerChain-of-Thought
🎯 What it does: This paper studies the mechanism of Chain of Thought (CoT) in Transformers, focusing on its impact on self-supervised learning multi-layer perceptrons (MLPs). It demonstrates through theory and experiments that CoT can be decomposed into two stages: filtering and context learning, significantly reducing sample complexity.
🎯 What it does: A class of distance-restricted FWL(2) graph neural networks (d-DRFWL(2) GNN) is proposed, and it is proven to maintain efficiency while having the ability to count cycles of length 6 and above.
Distilling Out-of-Distribution Robustness from Vision-Language Foundation Models
Andy Zhou (University of Illinois), Haohan Wang (University of Illinois)
CodeKnowledge DistillationAdversarial AttackVision Language ModelContrastive LearningImageMultimodality
🎯 What it does: The study uses large visual language foundation models (such as CLIP) as teachers to enhance the out-of-distribution robustness of small visual models through knowledge distillation and discrete adversarial distillation (DAD).
Distribution-Free Model-Agnostic Regression Calibration via Nonparametric Methods
Shang Liu (Imperial College London), Xiaocheng Li (Imperial College London)
CodeTabularTime Series
🎯 What it does: A two-step nonparametric regression calibration method is proposed, which estimates quantiles of the residuals from a pre-trained regression model to achieve individual calibration.
Distributional Learning of Variational AutoEncoder: Application to Synthetic Data Generation
Seunghwan An (University of Seoul), Jong-June Jeon (University of Seoul)
CodeGenerationData SynthesisSafty and PrivacyAuto EncoderTabular
🎯 What it does: A distribution learning-based variational autoencoder (DistVAE) is proposed, which directly estimates the conditional cumulative distribution function (CDF) of continuous variables through an infinite mixture of asymmetric Laplace distributions, thus no longer being limited by traditional Gaussian decoders.
Xin-Qiang Cai (University of Tokyo), Ashley Juan Llorens
CodeOptimizationReinforcement LearningBenchmark
🎯 What it does: This paper proposes Distributed Pareto Optimal Multi-Objective Reinforcement Learning (DPMORL), which addresses distributional preferences in multi-objective decision-making by learning Distributed Pareto Optimal (DPO) policies.
🎯 What it does: A distributionally robust optimization (DRO) framework is proposed to learn the undirected skeleton of discrete Bayesian networks in the presence of noise or contaminated data.
🎯 What it does: Proposed Diversified Outlier Exposure (DivOE), which enhances OOD detection by extrapolating information from auxiliary outlier samples during the fine-tuning process to expand the OOB distribution.
Diversify Your Vision Datasets with Automatic Diffusion-based Augmentation
Lisa Dunlap (University of California Berkeley), Trevor Darrell (University of California Berkeley)
CodeData SynthesisDomain AdaptationVision Language ModelDiffusion modelImage
🎯 What it does: An automated language-guided image enhancement method called ALIA is proposed, which first generates task-independent domain descriptions using visual and language models, then edits the original image through a diffusion model, and maintains data quality through semantic and confidence filtering.
🎯 What it does: This paper proposes a Spatial-Temporal Diversification Network (STDN) that learns diversified video features through spatial grouping and spatial-temporal relationship modeling to enhance video domain generalization performance.
Does a sparse ReLU network training problem always admit an optimum ?
TUNG QUOC LE, Elisa Riccietti (University of Lyon)
CodeOptimizationTabular
🎯 What it does: This paper studies the existence of optimal solutions for training sparse ReLU neural networks, proving that under certain sparse structures, optimal solutions may not exist, and providing necessary and sufficient conditions; it also proves that single hidden layer sparse ReLU networks can always achieve optimal solutions on finite or continuous domains; and provides a closure determination algorithm based on real algebraic geometry.
🎯 What it does: A structural broadcast graph dataset distillation framework SGDD is proposed, which can generate extremely small-scale synthetic graphs while preserving the original graph structure information.
🎯 What it does: This paper studies the feasibility of achieving Invariant graph representation learning in graph data lacking environmental labels and proposes the GALA framework based on an environmental assistant model.
Does Localization Inform Editing? Surprising Differences in Causality-Based Localization vs. Knowledge Editing in Language Models
Peter Hase (Google Research), Asma Ghandeharioun (Google Research)
CodeTransformerLarge Language ModelText
🎯 What it does: The relationship between causal tracing localization and model editing success rate was studied, finding no significant correlation between the two.
Ning Liu (Global Engineering and Materials), Yue Yu (Lehigh University)
CodeTabularTime SeriesPhysics Related
🎯 What it does: A domain-agnostic Fourier neural operator (DAFNO) is proposed for learning PDE solvers for arbitrary geometries and topological changes.
🎯 What it does: This paper studies a generative model for transferring from the source domain to the target domain under very few samples, namely few-shot generative domain adaptation (GDA).
🎯 What it does: This paper proposes a new method for dataset ownership verification, utilizing domain watermarking to ensure that models trained on protected datasets perform correctly on specific 'hard' samples, rather than using traditional backdoor attack methods.
Don’t just prune by magnitude! Your mask topology is a secret weapon
Duc N.M Hoang (University of Texas at Austin), Zhangyang Wang (University of Texas at Austin)
CodeImage
🎯 What it does: This study investigates the spectral properties of sparse neural networks, proposes a correlation between weighted spectral distance and performance, constructs a full spectral coordinate, and based on this, designs a sparsification method PAGS/PEGS under initialization or lightweight pre-training.
Don’t Stop Pretraining? Make Prompt-based Fine-tuning Powerful Learner
Zhengxiang Shi (University College London), Aldo Lipani (University College London)
CodeTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: Proposes Prompt-based Continued Pre-training (PCP), which introduces task-related text and prompt templates during the continued pre-training phase, improving traditional Task Adaptive Pre-training (TAPT).
DoReMi: Optimizing Data Mixtures Speeds Up Language Model Pretraining
Sang Michael Xie (Google DeepMind), Adams Wei Yu (Google DeepMind)
CodeOptimizationComputational EfficiencyLarge Language ModelText
🎯 What it does: This paper proposes the DoReMi algorithm, which automatically finds the optimal domain weights through distributed robust optimization (DRO) using a small model, thereby improving the pre-training efficiency of large language models.
DOSE: Diffusion Dropout with Adaptive Prior for Speech Enhancement
Wenxin Tai (University of Electronic Science and Technology of China), Ting Zhong (University of Electronic Science and Technology of China)
CodeRestorationGenerationDiffusion modelAudio
🎯 What it does: A speech enhancement method based on diffusion models, DOSE, is proposed, which significantly improves the utilization of conditional information and addresses the issue of conditional collapse by randomly dropping intermediate samples during training and using adaptive priors.
🎯 What it does: A Doubly Robust Self-Training method is proposed, which improves the self-training loss function so that the model can revert to training only on labeled data when the teacher model is inaccurate, regardless of the quality of the pseudo-labels, and fully utilize pseudo-label data when the teacher model is accurate.
DoWG Unleashed: An Efficient Universal Parameter-Free Gradient Descent Method
Ahmed Khaled (Princeton University), Chi Jin (Princeton University)
CodeOptimizationImage
🎯 What it does: A new non-parametric gradient descent optimizer called DoWG (Distance-based Weighted Gradient) is proposed, which matches the convergence rate of optimally tuned gradient descent in convex optimization without the need to adjust any parameters.
DP-Mix: Mixup-based Data Augmentation for Differentially Private Learning
Wenxuan Bao (University of Florida), Vincent Bindschaedler (University of Florida)
CodeSafty and PrivacyDiffusion modelImage
🎯 What it does: The study explores the feasibility of using the data augmentation technique mixup with multi-sample data in differential privacy learning and proposes two mixup data augmentation methods based on self-augmentation and diffusion models, namely DP-MIXSELF and DP-MIXDIFF.
🎯 What it does: This paper presents a novel fast diffusion ODE solver DPM-Solver-v3, which significantly reduces first-order discretization errors and further improves the quality of few-step sampling by reparameterizing the ODE through the introduction of Empirical Model Statistics (EMS).
DRAUC: An Instance-wise Distributionally Robust AUC Optimization Framework
Siran Dai (Institute of Information Engineering, Chinese Academy of Sciences), Qingming Huang (University of Chinese Academy of Sciences)
CodeOptimizationImage
🎯 What it does: An instance-level distributionally robust AUC (DRAUC) optimization framework is proposed to address the computational intractability issues arising from the combination of AUC and distributionally robust optimization.
🎯 What it does: This paper proposes the DREAM-OOD framework, which utilizes diffusion models to sample low-likelihood embeddings in a text-conditioned latent space and decode them into high-resolution, realistic out-of-distribution (OOD) images, thereby providing automated OOD data for training; the method can also be extended to generate in-distribution (ID) images.
Drift doesn't Matter: Dynamic Decomposition with Diffusion Reconstruction for Unstable Multivariate Time Series Anomaly Detection
Chengsen Wang (Beijing University of Posts and Telecommunications), Jianxin Liao (Beijing University of Posts and Telecommunications)
CodeAnomaly DetectionTransformerDiffusion modelTime Series
🎯 What it does: A heterogeneous multivariate time series anomaly detection framework named D3R is proposed, which first achieves dynamic decomposition of long-period unstable sequences through data-time mixed attention and offset subtraction, and then constructs an information bottleneck using noise diffusion for direct reconstruction, obtaining anomaly scores from the reconstruction error.
🎯 What it does: Proposes the DropCompute method, which reduces computational variance and enhances robustness through threshold clipping in distributed synchronous training.
🎯 What it does: This paper proposes DropPos, a self-supervised pre-training task based on position recovery, allowing the Vision Transformer to predict the positions of visible image patches solely based on visual information.
DrugCLIP: Contrastive Protein-Molecule Representation Learning for Virtual Screening
Bowen Gao (Tsinghua University), Yanyan Lan (Tsinghua University)
CodeRetrievalRepresentation LearningDrug DiscoveryTransformerContrastive LearningBiomedical Data
🎯 What it does: We propose DrugCLIP, which rephrases virtual screening as a dense retrieval task, using contrastive learning to learn a shared representation of protein pockets and small molecules, and enabling rapid retrieval on pre-encoded molecular embeddings.
DSR: Dynamical Surface Representation as Implicit Neural Networks for Protein
Daiwen Sun (Renmin University of China), Qiwei Ye (Beijing Academy of Artificial Intelligence)
CodeProtein Structure PredictionPoint Cloud
🎯 What it does: In protein dynamics, an implicit neural network is used to learn a continuous representation of protein surfaces in three-dimensional space and time, directly predicting the signed distance function (SDF) from point clouds, enabling the reconstruction and prediction of large-scale protein trajectories.
🎯 What it does: The Dual Mean-Teacher (DMT) framework is proposed, utilizing a dual teacher-student structure for semi-supervised learning in audio-video source localization tasks, enhancing localization accuracy through noise filtering and pseudo-label cross-validation.
DYffusion: A Dynamics-informed Diffusion Model for Spatiotemporal Forecasting
Salva Rühling Cachay (University of California San Diego), Rose Yu (University of California San Diego)
CodeDiffusion modelTime SeriesPhysics Related
🎯 What it does: A dynamic information diffusion model named DYffusion is proposed for probabilistic prediction of high-dimensional spatiotemporal sequences.
Dynamic Personalized Federated Learning with Adaptive Differential Privacy
Xiyuan Yang (Wuhan University), Mang Ye (Wuhan University)
CodeFederated LearningSafty and PrivacyConvolutional Neural NetworkImage
🎯 What it does: A method called FedDPA is proposed, which achieves dynamic Fisher information personalization and adaptive constraints in a personalized federated learning framework with user-level differential privacy, thereby improving model performance and convergence speed.
🎯 What it does: The Chase method is proposed, seamlessly transforming the advantages of unstructured dynamic sparse training (DST) into GPU-friendly channel-level sparse networks while maintaining or even improving model accuracy.
🎯 What it does: A dynamic tensor decomposition model based on neural diffusion-reaction processes (DEMOTE) is proposed to learn embeddings of entities that change over time and predict entry values of sparse multidimensional time series data.
🎯 What it does: This paper proposes a Dynamic Masking Discriminator (DMD) that adaptively adjusts the discriminator during GAN training through an online continuous learning mechanism, enhancing the learning effectiveness of the generator.
Echoes Beyond Points: Unleashing the Power of Raw Radar Data in Multi-modality Fusion
Yang Liu (Chinese Academy of Sciences Institute of Automation), Zhaoxiang Zhang (Chinese Academy of Sciences Institute of Automation and University of Chinese Academy of Sciences)
🎯 What it does: The EchoFusion method is proposed, which directly fuses raw radar time-domain/frequency-domain data with images to achieve efficient cross-modal fusion in the BEV space, thereby enhancing object detection performance.
🎯 What it does: This paper proposes and applies an ecosystem-level analysis method to assess the comprehensive impact of multi-model ensembles deployed in different scenarios on individuals, revealing systemic failures and homogenized outcomes.
Effective Bayesian Heteroscedastic Regression with Deep Neural Networks
Alexander Immer (ETH Zurich), Julia E Vogt
CodeAuto EncoderGaussian SplattingTabular
🎯 What it does: A deep heteroscedastic regression model based on natural parameterization is proposed, and a scalable Laplace approximation is introduced to achieve automatic Bayesian regularization and model uncertainty estimation.
Effective Human-AI Teams via Learned Natural Language Rules and Onboarding
Hussein Mozannar (Massachusetts Institute of Technology), David Sontag (Massachusetts Institute of Technology)
CodeObject DetectionRecommendation SystemAutonomous DrivingTransformerLarge Language ModelContrastive LearningImageText
🎯 What it does: A framework called IntegrAI is proposed for automatically learning and mastering natural language rules, aimed at guiding humans in making three types of decisions—ignore, rely, or collaborate—when working with AI.
🎯 What it does: A targeted adversarial attack for self-supervised learning is proposed, which selects the most confusing samples by combining entropy and similarity, allowing the model to achieve better robustness under unlabeled conditions.
🎯 What it does: By constructing three major benchmark datasets for activation functions and utilizing the Fisher information matrix features along with activation function outputs to create a low-dimensional proxy space, efficient optimization of activation functions was achieved through simple regression search, discovering various activation functions that surpass ReLU (including sigmoidal forms).
🎯 What it does: The Adapter Re-Composing (ARC) method is proposed, which achieves efficient fine-tuning of Vision Transformers by sharing low-rank adapter parameters.
🎯 What it does: A robust sensitivity-aware core subset selection (RCS) framework is proposed, which utilizes an unlabeled representative difference (RD) metric to select a subset that enables adversarial contrastive learning (ACL) to produce robust representations, thereby significantly accelerating ACL training.
Efficient Bayesian Learning Curve Extrapolation using Prior-Data Fitted Networks
Steven Adriaensen (University of Freiburg), Frank Hutter (University of Freiburg)
CodeTransformerTime SeriesBenchmark
🎯 What it does: This study investigates the use of Prior-Data Fitting Networks (PFN) for Bayesian learning curve extrapolation, proposing and implementing the LC-PFN model, which is further applied to online early stopping decisions.
Jishnu Ray Chowdhury (University of Illinois Chicago), Cornelia Caragea (University of Illinois Chicago)
CodeRecurrent Neural NetworkText
🎯 What it does: An efficient Beam Tree Recursive Neural Network (EBT-RvNN) is proposed and extended into a structured model (EBT-GAU) that can contextualize tokens.
Efficient Equivariant Transfer Learning from Pretrained Models
Sourya Basu (University of Illinois at Urbana-Champaign), Lav R. Varshney (University of Illinois at Urbana-Champaign)
CodeClassificationOptimizationTransformerLarge Language ModelReinforcement LearningImageText
🎯 What it does: Proposes λ-equitune and its special case equizero, utilizing pre-trained models to achieve weighted group equivariant transfer learning across various tasks.
🎯 What it does: A model-based reinforcement learning algorithm for continuous time systems, OCORL, is proposed, which utilizes ODE to describe dynamics, Gaussian Process (GP) to estimate the model, and employs an optimistic principle for policy optimization. Various measurement selection strategies (MSS) are designed and proven to achieve asymptotic no-regret under the GP model.
Efficient Learning of Linear Graph Neural Networks via Node Subsampling
Seiyun Shin (University of Illinois), Han Zhao (University of Illinois)
CodeGraph Neural NetworkGraph
🎯 What it does: A training algorithm for linear graph convolutional networks based on node subsampling and leverage score sampling is designed, requiring only O(nd log n) entries of A for training, avoiding the O(nd²) computational cost.
🎯 What it does: A low-rank backpropagation method called LBP-WHT is proposed, which utilizes the Walsh-Hadamard transform to perform gradient matrix multiplication in a low-rank space, significantly reducing the computational load of Vision Transformers during the fine-tuning process.
🎯 What it does: An efficient Evolutionary Meta-Neuro Heuristic (EMNH) framework is proposed for Multi-Objective Combinatorial Optimization Problems (MOCOP), which generates Pareto approximate solutions by first training a meta-model and then quickly fine-tuning it for different weight vectors.
Efficient Robust Bayesian Optimization for Arbitrary Uncertain inputs
Lin Yang (Huawei Noah's Ark Lab), Zhitang Chen (Huawei Noah's Ark Lab)
CodeOptimizationRobotic IntelligenceTabular
🎯 What it does: This paper proposes a robust Bayesian optimization algorithm called AIRBO, which addresses the uncertainty of arbitrary distribution inputs and aims to find optimal solutions that perform stably under uncertain inputs.
Efficient Test-Time Adaptation for Super-Resolution with Second-Order Degradation and Reconstruction
Zeshuai Deng (South China University of Technology), Mingkui Tan (South China University of Technology)
CodeRestorationSuper ResolutionImage
🎯 What it does: A super-resolution test-time adaptation framework (SRTTA) is proposed, which can quickly adjust a pre-trained model to generate high-quality HR images for unknown degradation of test images without training data.