arXivSub Start free trial

NeurIPS 2024 Papers with Code β€” Page 17

Conference on Neural Information Processing Systems Β· 1874 papers

Structured Multi-Track Accompaniment Arrangement via Style Prior Modelling

Jingwei Zhao (National University of Singapore), Ye Wang (National University of Singapore)

CodeGenerationData SynthesisTransformerAuto EncoderAudio

🎯 What it does: A two-stage music accompaniment arrangement system is proposed, which first converts the main melody and chords into piano accompaniment, and then uses a style prior model to convert the piano accompaniment into a multi-track full orchestra arrangement.

SubgDiff: A Subgraph Diffusion Model to Improve Molecular Representation Learning

Jiying Zhang (International Digital Economy Academy), Yu Li (International Digital Economy Academy)

CodeRepresentation LearningDrug DiscoveryGraph Neural NetworkDiffusion modelGraphBiomedical Data

🎯 What it does: This paper proposes SubgDiff, a molecular representation learning method that incorporates subgraph information into diffusion models.

Subject-driven Text-to-Image Generation via Preference-based Reinforcement Learning

Yanting Miao (University of Waterloo), Yeqing Li (Google)

CodeGenerationData SynthesisOptimizationReinforcement Learning from Human FeedbackReinforcement LearningDiffusion modelImageTextMultimodality

🎯 What it does: Proposes the λ-Harmonic reward function and Reward Preference Optimization (RPO) scheme for lightweight fine-tuning of pre-trained diffusion models with a small number of reference images to achieve theme-driven text-to-image generation.

Subwords as Skills: Tokenization for Sparse-Reward Reinforcement Learning

David Yunis (Toyota Technological Institute at Chicago), Matthew Walter (Toyota Technological Institute at Chicago)

CodeReinforcement Learning

🎯 What it does: Quickly extract limited interpretable skills from demonstration actions using BPE subword tokenization, and use these skills as a discrete action space for sparse reward reinforcement learning.

Superposed Decoding: Multiple Generations from a Single Autoregressive Inference Pass

Ethan Shen (University of Washington), Aditya Kusupati (University of Washington)

CodeGenerationComputational EfficiencyTransformerLarge Language ModelText

🎯 What it does: Proposes the Superposed Decoding algorithm, which utilizes token embedding superposition and n-gram interpolation to generate k drafts in a single autoregressive inference, significantly reducing computational costs.

Supervised Kernel Thinning

Albert Gong (Cornell University), Raaz Dwivedi (Cornell University)

CodeCompressionComputational EfficiencyTabular

🎯 What it does: A supervised compression method based on Kernel Thinning is proposed to accelerate the training and inference of Nadaraya-Watson and Kernel Ridge Regression.

SuperVLAD: Compact and Robust Image Descriptors for Visual Place Recognition

Feng Lu (Tsinghua University), Chun Yuan (Tsinghua University)

CodeRecognitionRetrievalDomain AdaptationTransformerImage

🎯 What it does: A novel global descriptor called SuperVLAD is proposed, which achieves more compact and robust visual location recognition features by removing cluster centers and using very few clusters. Additionally, a 1-Cluster VLAD is designed to obtain extremely low-dimensional descriptors.

Supra-Laplacian Encoding for Transformer on Dynamic Graphs

Yannis Karmim (Conservatoire National des Arts et MΓ©tiers), Nicolas THOME

CodeGraph Neural NetworkTransformerGraphTime Series

🎯 What it does: Developed the SLATE model, which combines supra-Laplacian encoding with a fully connected Transformer to address dynamic link prediction.

SVFT: Parameter-Efficient Fine-Tuning with Singular Vectors

Vijay Lingam (University of Texas at Austin), sujay sanghavi

CodeTransformerSupervised Fine-TuningImageTextBenchmark

🎯 What it does: For parameter-efficient fine-tuning (PEFT) of large-scale pre-trained models, a method called SVFT is proposed, which utilizes a sparse weighted combination of the singular vectors of the weight matrix to update the weights.

Swift Sampler: Efficient Learning of Sampler by 10 Parameters

Jiawei Yao (University of Washington), Canran Xiao (Central South University)

CodeOptimizationHyperparameter SearchReinforcement LearningImage

🎯 What it does: This paper proposes an algorithm for an automated search data sampler (sampler) - Swift Sampler (SS), which accelerates model convergence and improves final performance by assigning sampling probabilities to each sample during training.

SwitchHead: Accelerating Transformers with Mixture-of-Experts Attention

RΓ³bert CsordΓ‘s (Stanford University), JΓΌrgen Schmidhuber (AI Initiative KAUST)

CodeOptimizationComputational EfficiencyTransformerMixture of ExpertsText

🎯 What it does: This paper proposes SwitchHead, a Transformer architecture that applies Mixture-of-Experts to the attention layer, significantly reducing computational and memory requirements while maintaining language modeling performance comparable to dense models.

SymILO: A Symmetry-Aware Learning Framework for Integer Linear Optimization

Qian Chen (Chinese University of Hong Kong), Tsung-Hui Chang (Chinese University of Hong Kong)

CodeOptimizationGraph Neural NetworkGraph

🎯 What it does: The SymILO framework is proposed, which improves optimal solution prediction by introducing the symmetry of ILP in supervised learning.

Symmetry-Informed Governing Equation Discovery

Jianke Yang (University of California San Diego), Rose Yu (University of California San Diego)

CodeOptimizationTime SeriesPhysics RelatedOrdinary Differential Equation

🎯 What it does: This paper proposes embedding time-invariant symmetry (Lie point symmetry) into the symbolic equation discovery process. It first derives the invariance constraints of ODE flow mappings, and then explicitly solves linear symmetry constraints or incorporates symmetry regularization in methods such as sparse regression (SINDy) and genetic programming (GP), further enhancing the success rate of discovery and model simplicity under noisy data.

SyncVIS: Synchronized Video Instance Segmentation

rongkun Zheng, Hengshuang Zhao (University of Hong Kong)

CodeObject DetectionSegmentationTransformerVideo

🎯 What it does: A synchronous video instance segmentation framework SyncVIS is proposed, which enhances the modeling capability for complex videos by synchronously processing video-level and frame-level queries.

Synthetic Programming Elicitation for Text-to-Code in Very Low-Resource Programming and Formal Languages

Federico Mora (University of California Berkeley), Sanjit A. Seshia (University of California Berkeley)

CodeGenerationAI Code AssistantLarge Language ModelPrompt EngineeringText

🎯 What it does: A text-to-code generation method called SPEAC is proposed for very low-resource programming languages, utilizing intermediate languages and automatic repair to achieve syntactic correctness.

TabEBM: A Tabular Data Augmentation Method with Distinct Class-Specific Energy-Based Models

Andrei Margeloiu (University of Cambridge), Mateja Jamnik (University of Cambridge)

CodeClassificationData SynthesisGenerative Adversarial NetworkTabularStochastic Differential Equation

🎯 What it does: This paper proposes TabEBM, an energy model-based table data augmentation method that generates synthetic samples by constructing independent EBMs for each category to enhance the performance of downstream classifiers.

TabPedia: Towards Comprehensive Visual Table Understanding with Concept Synergy

Weichao Zhao (University of Science and Technology of China), Can Huang (ByteDance)

CodeRecognitionObject DetectionTransformerLarge Language ModelVision Language ModelTabular

🎯 What it does: TabPedia is proposed, a unified visual table understanding large visual language model that can simultaneously perform table detection, structure recognition, location querying, and question-answering tasks.

Tackling Uncertain Correspondences for Multi-Modal Entity Alignment

Liyi Chen (University of Science and Technology of China), Hui Xiong (Hong Kong University of Science and Technology)

CodeTransformerLarge Language ModelAuto EncoderContrastive LearningMultimodality

🎯 What it does: A new model TMEA is proposed for entity alignment in multimodal knowledge graphs, focusing on addressing the uncertainty of correspondence both between different modalities and within the same modality.

TAIA: Large Language Models are Out-of-Distribution Data Learners

Shuyang Jiang (Fudan University), Yu Wang (Shanghai Jiao Tong University)

CodeTransformerLarge Language ModelSupervised Fine-TuningText

🎯 What it does: The TAIA method is proposed, which trains all parameters during LLM fine-tuning but retains only the updated attention parameters during inference, reducing reliance on high-quality domain data and enhancing robustness to OOD data.

Talking Heads: Understanding Inter-Layer Communication in Transformer Language Models

Jack Merullo (Brown University), Ellie Pavlick (Brown University)

CodeRetrievalExplainability and InterpretabilityTransformerLarge Language ModelPrompt EngineeringText

🎯 What it does: By performing singular value decomposition (SVD) on the weights of GPT-2 and Pythia, low-rank communication channels are identified, and interventions are conducted to verify their causal impact on model behavior, further explaining prompt sensitivity and improving performance on list recall tasks.

TALoS: Enhancing Semantic Scene Completion via Test-time Adaptation on the Line of Sight

Hyun-Kurl Jang (Korea Advanced Institute of Science and Technology), Kuk-Jin Yoon (Korea Advanced Institute of Science and Technology)

CodeSegmentationDomain AdaptationAutonomous DrivingContrastive LearningPoint Cloud

🎯 What it does: This paper proposes TALoS, which achieves test-time adaptation for SSC by utilizing line-of-sight information from multi-temporal LiDAR observations.

Taming Diffusion Prior for Image Super-Resolution with Domain Shift SDEs

Qinpeng Cui (Advanced Micro Devices), Emad Barsoum (Advanced Micro Devices)

CodeRestorationSuper ResolutionDomain AdaptationDiffusion modelImageStochastic Differential Equation

🎯 What it does: This paper proposes DoSSR, a domain transfer super-resolution framework based on a pre-trained diffusion model, which directly initiates the diffusion process from low-resolution images and achieves a smooth transition from low resolution to high resolution through a domain shift equation.

Taming Generative Diffusion Prior for Universal Blind Image Restoration

Siwei Tu (Fudan University), Ben Fei (Chinese University of Hong Kong)

CodeRestorationDiffusion modelImage

🎯 What it does: This paper proposes a general framework for blind image restoration using diffusion models under unknown degradation conditions, called BIR-D.

Taming the Long Tail in Human Mobility Prediction

Xiaohang Xu (University of Tokyo), Kaoru Sezaki (University of Tokyo)

CodeRecommendation SystemSafty and PrivacyGraph Neural NetworkTransformerGraph

🎯 What it does: The LoTNext framework is proposed to address the long-tail distribution problem in predicting the next location (POI) for humans, improving the prediction accuracy for low-visit-frequency places.

Tangent Space Causal Inference: Leveraging Vector Fields for Causal Discovery in Dynamical Systems

Kurt Butler (Stony Brook University), Petar Djuric

CodeTime SeriesOrdinary Differential Equation

🎯 What it does: A new causal inference method is proposed - Tangent Space Causal Inference (TSCI), which detects causal relationships by comparing the tangent vector fields of dynamic system embedding spaces.

Target-Guided Adversarial Point Cloud Transformer Towards Recognition Against Real-world Corruptions

Jie Wang (Beijing Institute of Technology), Jianan Li (Chinese University of Hong Kong)

CodeRecognitionAdversarial AttackTransformerPrompt EngineeringPoint Cloud

🎯 What it does: This paper proposes an adversarial Transformer model for 3D point cloud recognition, named APCT, aimed at enhancing robustness against real-world noise and corruptions.

Targeted Sequential Indirect Experiment Design

Elisabeth Ailer (Technical University of Munich), Niki Kilbertus (Technical University of Munich)

CodeOptimizationTabular

🎯 What it does: This paper proposes an adaptive strategy for designing indirect experiments to optimally obtain information about the target queries regarding the true mechanisms, particularly in the presence of confounding factors and multidimensional nonlinearity.

TARP-VP: Towards Evaluation of Transferred Adversarial Robustness and Privacy on Label Mapping Visual Prompting Models

Zhen Chen (University of Liverpool), Wenjie Ruan (University of Science and Technology of China)

CodeClassificationSafty and PrivacyAdversarial AttackConvolutional Neural NetworkTransformerPrompt EngineeringImage

🎯 What it does: This study investigates the security of the Label Mapping Visual Prompt (LM-VP) model, systematically assessing its robustness against transfer adversarial attacks and the privacy leakage risk from Membership Inference Attacks (MIA).

TARSS-Net: Temporal-Aware Radar Semantic Segmentation Network

Youcheng Zhang (Intelligent Science and Technology Academy of CASIC), Zhe Ma (Shenzhen International Graduate School Tsinghua University)

CodeSegmentationConvolutional Neural NetworkTransformerImage

🎯 What it does: A new framework for radar semantic segmentation, TARSS-Net, is proposed, which utilizes the TRAM module for target historical association learning and aggregation of temporal information to enhance the performance of radar multi-view semantic segmentation.

Task-Agnostic Machine-Learning-Assisted Inference

Jiacheng Miao (University of Wisconsin Madison), Qiongshi Lu (University of Wisconsin Madison)

CodeTabularBiomedical Data

🎯 What it does: A task-agnostic machine learning-assisted inference framework, PSPS, is proposed, allowing any existing statistical inference method based on labeled data to combine unlabeled data and pre-trained ML models to achieve effective and more accurate inference results.

Task-oriented Time Series Imputation Evaluation via Generalized Representers

Zhixian Wang (University of Hong Kong), Yi Wang (Alibaba Group)

CodeAnomaly DetectionOptimizationTime Series

🎯 What it does: This paper addresses the issue of imputing missing values in time series by proposing a task-oriented (using prediction tasks as an example) evaluation framework. This framework can estimate the impact of each imputed value on downstream tasks without retraining the prediction model, and it combines the advantages of different imputation methods based on this impact to achieve better imputation results.

Task-recency bias strikes back: Adapting covariances in Exemplar-Free Class Incremental Learning

Grzegorz RypeΕ›Δ‡ (Warsaw University of Technology), BartΕ‚omiej Twardowski (Autonomous University of Barcelona)

CodeClassificationKnowledge DistillationGaussian SplattingImage

🎯 What it does: The AdaGauss method is proposed for category incremental learning without sample storage, dynamically adapting the Gaussian distribution for each task and alleviating dimensional collapse and task proximity bias through regularization.

Temporal Sentence Grounding with Relevance Feedback in Videos

Jianfeng Dong (Zhejiang Gongshang University), Meng Wang (Hefei University of Technology)

CodeRetrievalRecurrent Neural NetworkVideoText

🎯 What it does: The Temporal Sentence Grounding with Relevance Feedback (TSG-RF) task is proposed, which can determine whether there are segments in a video related to a query sentence, and if so, provide precise time boundaries; otherwise, it gives feedback of 'no relevant segments'.

Temporal-Difference Learning Using Distributed Error Signals

Jonas Guan (University of Toronto), William A Cunningham

CodeReinforcement LearningSequential

🎯 What it does: A new deep Q-learning algorithm called ARTIFICIAL DOPAMINE (AD) is designed, which learns using only distributed hierarchical TD errors without the need for backpropagation.

Test-Time Adaptation Induces Stronger Accuracy and Agreement-on-the-Line

Eungyeup Kim (Carnegie Mellon University), J Zico Kolter

CodeDomain AdaptationImage

🎯 What it does: This study investigates the impact of Test-Time Adaptation (TTA) on model accuracy and Agreement-on-the-Line (ACL/AGL) under different distribution shifts.

Test-Time Dynamic Image Fusion

Bing Cao (Tianjin University), Qinghua Hu (Tianjin University)

CodeRestorationConvolutional Neural NetworkImageMultimodalityBiomedical DataMagnetic Resonance Imaging

🎯 What it does: This paper proposes a Test-Time Dynamic Image Fusion (TTD) method, which significantly improves fusion quality by dynamically fusing multi-source images based on pixel-level relative dominance (RD) weights derived from reconstruction loss.

Testing Semantic Importance via Betting

Jacopo Teneggi (Johns Hopkins University), Jeremias Sulam (Johns Hopkins University)

CodeExplainability and InterpretabilityVision Language ModelImage

🎯 What it does: A semantic concept importance testing framework for black-box models (C-SKIT and X-SKIT) was constructed using the theory of conditional independence and sequential betting methods, achieving global and local importance ranking along with control of false discovery rates.

Text-DiFuse: An Interactive Multi-Modal Image Fusion Framework based on Text-modulated Diffusion Model

Hao Zhang (Wuhan University), Jiayi Ma

CodeImage TranslationRestorationDiffusion modelImageMultimodality

🎯 What it does: This paper proposes an interactive multimodal image fusion framework called Text-DiFuse based on a text-controlled diffusion model, which can perform denoising, color cast removal, and multimodal information fusion in the presence of composite distortions (such as noise, color cast, underexposure, etc.), while also supporting the emphasis of targets of interest through natural language instructions.

Text-Guided Attention is All You Need for Zero-Shot Robustness in Vision-Language Models

Lu Yu (Tianjin University of Technology), Changsheng Xu (Institute of Automation, University of Chinese Academy of Sciences)

CodeExplainability and InterpretabilityAdversarial AttackTransformerVision Language ModelImageMultimodality

🎯 What it does: A new framework TGA-ZSR is proposed, which enhances the zero-shot robustness of visual-language models through a text-guided attention mechanism while maintaining performance on clean samples.

Text-Infused Attention and Foreground-Aware Modeling for Zero-Shot Temporal Action Detection

Yearang Lee (Korea University), Seong-Whan Lee (Korea University)

CodeRecognitionObject DetectionTransformerVision Language ModelVideoTextMultimodality

🎯 What it does: A zero-shot temporal action detection method that integrates text and video information is proposed to address the common sub-action bias problem.

Text2CAD: Generating Sequential CAD Designs from Beginner-to-Expert Level Text Prompts

Mohammad Sadil Khan (DFKI), Muhammad Zeshan Afzal (DFKI)

CodeGenerationTransformerVision Language ModelTextSequential

🎯 What it does: Proposes the Text2CAD framework, which automatically generates parametric CAD models using text prompts, supporting different levels of description from beginners to experts;

Text2NKG: Fine-Grained N-ary Relation Extraction for N-ary relational Knowledge Graph Construction

Haoran Luo (Beijing University of Posts and Telecommunications), Anh Tuan Luu (Nanyang Technological University)

CodeClassificationComputational EfficiencyKnowledge DistillationTransformerSupervised Fine-TuningText

🎯 What it does: Proposes the Text2NKG framework for fine-grained n-ary relation extraction from natural language text, and uses the extraction results to construct four common types of NKGs (hyper-relational, event-based, role-based, hypergraph-based).

TextCtrl: Diffusion-based Scene Text Editing with Prior Guidance Control

Weichao Zeng (Institute of Information Engineering, Chinese Academy of Sciences), Yu Zhou (Nankai University)

CodeRecognitionGenerationData SynthesisDiffusion modelImage

🎯 What it does: This paper proposes a scene text editing method called TextCtrl based on diffusion models, achieving high-fidelity editing through prior style decoupling and character structure representation.

Textual Training for the Hassle-Free Removal of Unwanted Visual Data: Case Studies on OOD and Hateful Image Detection

Saehyung Lee (Seoul National University), Sungroh Yoon (Seoul National University)

CodeClassificationAnomaly DetectionTransformerPrompt EngineeringVision Language ModelImageText

🎯 What it does: A framework named HFTT (Hassle‑Free Textual Training) is proposed, which trains a classifier to detect undesirable visual content (such as OOD samples and hate images) using only text data and a pre-trained vision-language model.

TFG: Unified Training-Free Guidance for Diffusion Models

Haotian Ye (Stanford University), Stefano Ermon (Peking University)

CodeGenerationData SynthesisOptimizationHyperparameter SearchDiffusion modelImageBenchmarkAudio

🎯 What it does: This paper proposes a unified training-free guidance framework (TFG) for achieving conditional generation on diffusion models.

The ALCHEmist: Automated Labeling 500x CHEaper than LLM Data Annotators

Tzu-Heng Huang (University of Wisconsin Madison), Frederic Sala (University of Wisconsin Madison)

CodeData-Centric LearningTransformerLarge Language ModelImageTextMultimodality

🎯 What it does: The Alchemist system is proposed, which utilizes large language models to generate executable annotation programs, replacing the implementation of data annotation through individual API calls.

The Benefits of Balance: From Information Projections to Variance Reduction

Lang Liu (University of Washington), Zaid Harchaoui (University of Washington)

CodeClassificationRepresentation LearningTransformerContrastive LearningImageTextMultimodality

🎯 What it does: This paper explains and utilizes multi-source data balancing in self-supervised learning through balancing multimodal data (e.g., Sinkhorn iteration), demonstrating that this process can significantly reduce the variance of target estimation and providing a non-asymptotic error upper bound.

The Best of Both Worlds: On the Dilemma of Out-of-distribution Detection

Qingyang Zhang (Tianjin University), Changqing Zhang (Tianjin University)

CodeClassificationAnomaly DetectionImage

🎯 What it does: A new Decoupled Uncertainty Learning (DUL) framework is proposed to address the contradiction in existing OOD detection methods, which improve detection performance while causing a decline in OOD generalization performance.

The Dormant Neuron Phenomenon in Multi-Agent Reinforcement Learning Value Factorization

Haoyuan Qin (Xiamen University), Siqi Shen (Xiamen University)

CodeReinforcement Learning

🎯 What it does: This paper studies the phenomenon of 'dormant neurons' in value decomposition methods within multi-agent reinforcement learning and proposes a new parameter perturbation method called ReBorn.

The Empirical Impact of Neural Parameter Symmetries, or Lack Thereof

Derek Lim (Massachusetts Institute of Technology), Stefanie Jegelka (Technical University of Munich)

CodeOptimizationMeta LearningGraph Neural NetworkImageGraph

🎯 What it does: This study proposes two structures to eliminate the symmetry in the parameter space of deep networks and experimentally evaluates their impact on phenomena such as linear mode connectivity, Bayesian inference, meta-networks, and monotonic linear interpolation.

The Evolution of Statistical Induction Heads: In-Context Learning Markov Chains

Ezra Edelman (University of Pennsylvania), Surbhi Goel (University of Pennsylvania)

CodeTransformerSequential

🎯 What it does: This study investigates the learning dynamics of the Transformer in in-context learning tasks based on random Markov chains, revealing the process by which the model gradually achieves statistical inference of pairs through a 'statistical induction head'.

The Expressive Capacity of State Space Models: A Formal Language Perspective

Yash Sarrof (Saarland University), Michael Hahn (Saarland University)

CodeRecurrent Neural NetworkTransformerSequential

🎯 What it does: The theoretical analysis of the expressive power of linear state space models (SSM) is conducted, comparing it with Transformers and RNNs, and studying its representability in regular languages, counting languages, and bounded hierarchical structures.

The Feature Speed Formula: a flexible approach to scale hyper-parameters of deep neural networks

LΓ©naΓ―c Chizat (Γ‰cole Polytechnique FΓ©dΓ©rale de Lausanne), Praneeth Netrapalli (Google DeepMind)

CodeOptimizationHyperparameter Search

🎯 What it does: A feature speed formula is proposed, and based on this, the feature learning rate of deep networks is analyzed, leading to the design of an automated hyperparameter scaling method.

The GAN is dead; long live the GAN! A Modern GAN Baseline

Nick Huang, James Tompkin (Brown University)

CodeGenerationData SynthesisConvolutional Neural NetworkGenerative Adversarial NetworkImage

🎯 What it does: This paper studies a more stable GAN training method and builds a simplified baseline R3GAN that does not require traditional techniques.

The Group Robustness is in the Details: Revisiting Finetuning under Spurious Correlations

Tyler LaBonte (Georgia Institute of Technology), Abhishek Kumar (Georgia Institute of Technology)

CodeClassificationConvolutional Neural NetworkTransformerSupervised Fine-TuningImageTextBenchmark

🎯 What it does: Conducted detailed experiments on the fine-tuning process of deep models in the presence of spurious correlations, studying the effects of class balance, model size, and spectral imbalance on the worst-group accuracy, and proposed a mixed balance method.

The High Line: Exact Risk and Learning Rate Curves of Stochastic Adaptive Learning Rate Algorithms

Elizabeth Collins-Woodfin (McGill University), Courtney Paquette (Google DeepMind)

CodeOptimizationTabularStochastic Differential EquationOrdinary Differential Equation

🎯 What it does: This paper develops a theoretical framework for high-dimensional optimization problems, providing precise analytical expressions for the risk and learning rate curves using a single pass of Stochastic Gradient Descent (SGD) with adaptive learning rates, and conducts an in-depth analysis of idealized Exact Line Search and AdaGrad-Norm algorithms on least squares problems.

The Impact of Initialization on LoRA Finetuning Dynamics

Soufiane Hayou (University of California Berkeley), Bin Yu (University of California Berkeley)

CodeTransformerSupervised Fine-TuningText

🎯 What it does: Study the impact of two random initialization schemes in LoRA fine-tuning on training dynamics and performance.

The Implicit Bias of Gradient Descent toward Collaboration between Layers: A Dynamic Analysis of Multilayer Perceptions

Zheng Wang (University of Exeter), Wenjie Ruan (University of Science and Technology of China)

CodeOptimizationAdversarial AttackConvolutional Neural NetworkImage

🎯 What it does: This study investigates the implicit bias of gradient descent on inter-layer cooperation regarding adversarial robustness, proposing a correlation measure to assess inter-layer cooperation.

The Importance of Being Scalable: Improving the Speed and Accuracy of Neural Network Interatomic Potentials Across Chemical Domains

Eric Qu (University of California Berkeley), Aditi S. Krishnapriyan (University of California Berkeley)

CodeGraph Neural NetworkTransformerTabular

🎯 What it does: This study investigates the scalability of neural network interatomic potentials (NNIP) and proposes an efficient scalable attention interatomic potential (EScAIP) model, which is evaluated on multiple chemical datasets.

The Ladder in Chaos: Improving Policy Learning by Harnessing the Parameter Evolving Path in A Low-dimensional Space

Hongyao Tang (Tianjin University), Jianye HAO

CodeReinforcement LearningTime Series

🎯 What it does: This study investigates the learning paths of the policy networks of typical deep reinforcement learning agents, discovering that they evolve in low-dimensional subspaces. The Policy Path Trimming and Boosting (PPTB) method is proposed to trim minor directions and reinforce the main direction, significantly improving the learning performance of TD3, RAD, and DoubleDQN.

The Mamba in the Llama: Distilling and Accelerating Hybrid Models

Junxiong Wang (Cornell University), Tri Dao (Princeton University)

CodeOptimizationComputational EfficiencyKnowledge DistillationRecurrent Neural NetworkTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningText

🎯 What it does: Distill large Transformer models (such as Llama-3 8B Instruct) into a linear RNN (Mamba) mixture model through weight mapping and a stepwise mixing structure, and based on this, propose hardware-aware multi-step speculative decoding to accelerate inference.

The Map Equation Goes Neural: Mapping Network Flows with Graph Neural Networks

Christopher BlΓΆcker (University of Zurich), Ingo Scholtes (Julius-Maximilians-UniversitΓ€t WΓΌrzburg)

CodeOptimizationGraph Neural NetworkGraph

🎯 What it does: Rewrite the objective function of the Map Equation in information theory into a differentiable tensor form, making it a loss function that can be directly used by Graph Neural Networks (GNNs), thus achieving end-to-end unsupervised community detection.

The Road Less Scheduled

Aaron Defazio (Meta), Ashok Cutkosky (Boston University)

CodeOptimizationConvolutional Neural NetworkTransformerSupervised Fine-TuningImageTextMagnetic Resonance ImagingBenchmark

🎯 What it does: An optimization method that does not require a preset learning rate schedule is proposedβ€”Schedule-Free learning;

The Surprising Effectiveness of SP Voting with Partial Preferences

Hadi Hosseini (Penn State University), Amrit Puhan (Penn State University)

CodeTabular

🎯 What it does: A scalable voting algorithm based on the idea of 'surprising popularity' is proposed, which recovers the true ranking using local preferences instead of complete rankings;

The Unmet Promise of Synthetic Training Images: Using Retrieved Real Images Performs Better

Scott Geng (University of Washington), Ranjay Krishna (Allen Institute for AI)

CodeClassificationRetrievalTransformerSupervised Fine-TuningDiffusion modelImageRetrieval-Augmented Generation

🎯 What it does: This paper studies the comparison of the effects of fine-tuning models using synthetic images generated by generative models versus real images retrieved directly from upstream datasets in visual classification tasks, demonstrating that retrieval methods often enhance model performance more effectively.

Theoretical Foundations of Deep Selective State-Space Models

Nicola Muca Cirone (Imperial College London), Terry Lyons (University of Oxford)

CodeRecurrent Neural NetworkTransformerSequentialBenchmarkStochastic Differential EquationOrdinary Differential Equation

🎯 What it does: This paper studies the Linear Control Differential Equation (CDE) model and provides a theoretical foundation for its expressiveness in sequential tasks, explaining the advantages of modern selectable state space models (SSMs) such as Mamba.

Thinking Forward: Memory-Efficient Federated Finetuning of Language Models

Kunjal Panchal (University of Massachusetts), Hui Guan (University of Massachusetts)

CodeFederated LearningComputational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningText

🎯 What it does: This paper proposes a method for low-memory fine-tuning of LLMs in a federated learning environment by splitting trainable weights across clients and using forward-mode automatic differentiation.

This Too Shall Pass: Removing Stale Observations in Dynamic Bayesian Optimization

Anthony Bardou (Ecole polytechnique fΓ©dΓ©rale de Lausanne), Giovanni Ranieri (Ecole polytechnique fΓ©dΓ©rale de Lausanne)

CodeOptimizationTabularTime SeriesBenchmark

🎯 What it does: A dynamic Bayesian optimization (DBO) algorithm W-DBO is proposed, which can online identify and remove observation points that are irrelevant to future predictions.

Thompson Sampling For Combinatorial Bandits: Polynomial Regret and Mismatched Sampling Paradox

Raymond Zhang (Universite Paris Saclay), Richard Combes (Universite Paris Saclay)

CodeOptimizationReinforcement Learning from Human FeedbackReinforcement LearningTabular

🎯 What it does: This paper proposes an improved Thompson Sampling (BG-CTS) algorithm for the linear combination of semi-bandit problems and provides its polynomial upper bound non-asymptotic convergence rate.

Time-FFM: Towards LM-Empowered Federated Foundation Model for Time Series Forecasting

Qingxiang Liu (Hong Kong University of Science and Technology), Yuxuan Liang (Hong Kong University of Science and Technology)

CodeFederated LearningSafty and PrivacyTransformerLarge Language ModelPrompt EngineeringTime Series

🎯 What it does: A federated foundation model TIME-FFM based on a pre-trained language model has been constructed for cross-domain time series prediction, addressing data privacy and multi-domain heterogeneity issues.

Time-Varying LoRA: Towards Effective Cross-Domain Fine-Tuning of Diffusion Models

Zhan Zhuang (Southern University of Science and Technology), Yu Zhang (Southern University of Science and Technology)

CodeGenerationDomain AdaptationDiffusion modelContrastive LearningImage

🎯 What it does: A time-varying low-rank adapter called Terra is proposed for cross-domain generation and fine-tuning of diffusion models, capable of achieving continuous domain flow generation and interpolation on a low-rank module.

TimeXer: Empowering Transformers for Time Series Forecasting with Exogenous Variables

Yuxuan Wang (Tsinghua University), Mingsheng Long (Tsinghua University)

CodeTransformerTime SeriesFinance Related

🎯 What it does: The TimeXer model is proposed, utilizing the self-attention and cross-attention of the Transformer to encode endogenous time series in patches and exogenous sequences in variates, and connects the two through globally learnable tokens, achieving precise integration of exogenous variables in time series forecasting.

TinyLUT: Tiny Look-Up Table for Efficient Image Restoration at the Edge

Huanan LI, Zhangming Zhu (Xidian University)

CodeRestorationSuper ResolutionConvolutional Neural NetworkImage

🎯 What it does: Proposes the TinyLUT framework, which utilizes a separation mapping strategy and a dynamic discretization mechanism to map convolutional networks into LUTs with minimal storage, achieving efficient image restoration.

TinyTTA: Efficient Test-time Adaptation via Early-exit Ensembles on Edge Devices

Hong Jia (University of Cambridge), Cecilia Mascolo (University of Cambridge)

CodeDomain AdaptationComputational EfficiencyImage

🎯 What it does: Proposed the TinyTTA framework, achieving efficient test-time adaptation (TTA) on resource-constrained MCUs, and developed the TinyTTA Engine library;

To Err Like Human: Affective Bias-Inspired Measures for Visual Emotion Recognition Evaluation

Chenxi Zhao (Nankai University), Jufeng Yang (Nankai University)

CodeRecognitionConvolutional Neural NetworkImage

🎯 What it does: Proposes an emotional distance based on the Mikel emotion wheel to define misclassification costs, and introduces two new visual emotion recognition evaluation metrics: ECC (Overall Evaluation) and EMC (Misclassification Evaluation Only), while validating their effectiveness in semi-supervised learning.

Token Merging for Training-Free Semantic Binding in Text-to-Image Synthesis

taihang Hu, Yaxing Wang (Nankai University)

CodeGenerationData SynthesisTransformerDiffusion modelImageText

🎯 What it does: Proposes a training-free Token Merging (ToMe) method to address the semantic binding issue in text-to-image generation;

TOPA: Extending Large Language Models for Video Understanding via Text-Only Pre-Alignment

Wei Li (Zhejiang University), Yi Yang (Zhejiang University)

CodeTransformerLarge Language ModelContrastive LearningVideoTextMultimodality

🎯 What it does: A pre-alignment framework called TOPA is constructed, which enables large language models to possess video understanding capabilities using only text.

TopoLogic: An Interpretable Pipeline for Lane Topology Reasoning on Driving Scenes

Yanping Fu (Institute of Computing Technology, Chinese Academy of Sciences University of Chinese Academy of Sciences), Feng Dai (Institute of Computing Technology, Chinese Academy of Sciences)

CodeAutonomous DrivingExplainability and InterpretabilityGraph Neural NetworkTransformerImage

🎯 What it does: An interpretable lane topology reasoning method called TopoLogic is designed, which combines geometric distance and semantic similarity to determine the connectivity between lanes.

Topological Generalization Bounds for Discrete-Time Stochastic Optimization Algorithms

Rayna Andreeva (University of Edinburgh), Umut Simsekli (Ecole Normale Superieure PSL Research University)

CodeOptimizationGraph Neural NetworkTransformerImage

🎯 What it does: A new class of topology-based complexity measures (α-weighted lifespan and forward quantization) is proposed, and its upper bound on the generalization error of discrete-time stochastic optimization algorithms is provided.

Toward a Well-Calibrated Discrimination via Survival Outcome-Aware Contrastive Learning

Dongjoon Lee (Chung Ang University), Changhee Lee (Korea University)

CodeContrastive LearningTime SeriesBiomedical Data

🎯 What it does: A deep survival model based on contrastive learning, ConSurv, is proposed, which constructs contrastive loss using weighted negative sampling based on the similarity of patient event times, thereby enhancing the model's discriminative ability while maintaining calibration.

Toward Dynamic Non-Line-of-Sight Imaging with Mamba Enforced Temporal Consistency

Yue Li (University of Science and Technology of China), Zhiwei Xiong (University of Science and Technology of China)

CodeRestorationVideo

🎯 What it does: A dynamic NLOS reconstruction framework based on ST-Mamba is proposed, utilizing multi-frame transient information to achieve high-resolution hidden object recovery.

Toward Self-Improvement of LLMs via Imagination, Searching, and Criticizing

Ye Tian (Tencent AI Lab), Dong Yu (Tencent AI Lab)

CodeTransformerLarge Language ModelReinforcement LearningText

🎯 What it does: The ALPHALLM framework is proposed, achieving self-improvement of LLMs;

Toward Semantic Gaze Target Detection

Samy Tafasca (Idiap Research Institute), Jean-marc Odobez

CodeRecognitionObject DetectionTransformerContrastive LearningImageBenchmark

🎯 What it does: An end-to-end semantic gaze target detection model is designed, capable of predicting both gaze heatmaps and target categories simultaneously.

Towards a Scalable Reference-Free Evaluation of Generative Models

Azim Ospanov (Chinese University of Hong Kong), Farzan Farnia (Chinese University of Hong Kong)

CodeGenerationComputational EfficiencyImageVideoText

🎯 What it does: A scalable reference-agnostic evaluation method FKEA based on random Fourier features has been developed to efficiently estimate the diversity score of generative models.

Towards a Theoretical Understanding of the 'Reversal Curse' via Training Dynamics

Hanlin Zhu (University of California Berkeley), Stuart Russell (University of California Berkeley)

CodeTransformerLarge Language ModelTextChain-of-Thought

🎯 What it does: This paper investigates the flaws of autoregressive large language models in reverse logical reasoning (the 'reversal curse') through theoretical analysis and experiments, and further explores the impact of chain-of-thought (COT) reasoning.

Towards Accurate and Fair Cognitive Diagnosis via Monotonic Data Augmentation

Zheng Zhang (University of Science and Technology of China), Enhong Chen (University of Science and Technology of China)

CodeTabular

🎯 What it does: To address the issue of data sparsity in cognitive diagnosis, a data augmentation framework based on monotonicity constraints, CMCD, is proposed to enhance the accuracy and fairness of diagnosis.

Towards an Information Theoretic Framework of Context-Based Offline Meta-Reinforcement Learning

Lanqing Li (Zhejiang Lab), Pheng-Ann Heng (Chinese University of Hong Kong)

CodeRobotic IntelligenceMeta LearningReinforcement LearningContrastive LearningSequential

🎯 What it does: A unified information-theoretic framework for context-based offline meta reinforcement learning (COMRL) is proposed, integrating existing COMRL algorithms and revealing their relationships through theoretical analysis.

Towards Croppable Implicit Neural Representations

Maor Ashkenazi (Ben Gurion University of the Negev), Eran Treister (Ben Gurion University of the Negev)

CodeImage TranslationRestorationSuper ResolutionImageVideoAudio

🎯 What it does: Proposes the Local-Global SIREN architecture, which can directly crop or extend the already encoded implicit neural representation without retraining or fine-tuning, and can be further used for various downstream tasks.

Towards Diverse Device Heterogeneous Federated Learning via Task Arithmetic Knowledge Integration

Mahdi Morafah (University of California San Diego), Bill Lin (Czech Technical University in Prague)

CodeFederated LearningKnowledge DistillationConvolutional Neural NetworkTransformerImageText

🎯 What it does: In the context of federated learning with heterogeneous devices, the TAKFL framework is proposed, which utilizes knowledge distillation to treat the model set of each device prototype as an independent task for knowledge transfer, and achieves multi-source knowledge fusion through self-supervised regularization and adaptive task arithmetic.

Towards Flexible 3D Perception: Object-Centric Occupancy Completion Augments 3D Object Detection

Chaoda Zheng (Shenzhen University), Zhen Li (Chinese University of Hong Kong)

CodeObject DetectionAutonomous DrivingTransformerPoint Cloud

🎯 What it does: This paper proposes an Object-Centric Occupancy representation and implements an enhanced 3D object detection and shape completion network based on this representation.

Towards Global Optimal Visual In-Context Learning Prompt Selection

Chengming Xu (Fudan University), Yanwei Fu (Fudan University)

CodeImage TranslationObject DetectionSegmentationTransformerContrastive LearningImage

🎯 What it does: A Transformer-based list-wise ranking model and consistency-aware aggregator are proposed for Visual Context Learning (VICL) to select the optimal contextual examples from multiple candidate samples, thereby enhancing the model's segmentation, detection, and coloring performance.

Towards Harmless Rawlsian Fairness Regardless of Demographic Prior

Xuanqian Wang (Beihang University), Yew-Soon Ong (Nanyang Technological University)

CodeClassificationOptimizationTabular

🎯 What it does: A VFair method based on loss variance minimization is proposed, achieving Rawlsian fairness during training without using any sensitive attributes; and harmless fairness is achieved through dynamic gradient adjustment.

Towards Learning Group-Equivariant Features for Domain Adaptive 3D Detection

Sangyun Shin (University of Oxford), Niki Trigoni (University of Oxford)

CodeObject DetectionDomain AdaptationAutonomous DrivingPoint Cloud

🎯 What it does: A domain-adaptive 3D object detection framework called GroupExp-DA is proposed, which utilizes a grouping strategy to balance training attention and improves detection accuracy through group equivariant spatial features.

Towards Neuron Attributions in Multi-Modal Large Language Models

Junfeng Fang (University of Science and Technology of China), Tat-Seng Chua (National University of Singapore)

CodeSegmentationGenerationTransformerLarge Language ModelDiffusion modelImageTextMultimodality

🎯 What it does: A neuron attribution method named NAM for multimodal large language models (MLLM) is proposed, which can identify key neurons during text and image generation processes and enable image editing.

Towards Next-Level Post-Training Quantization of Hyper-Scale Transformers

Junhan Kim (Samsung Research), Yongkweon Jeon (Samsung Research)

CodeOptimizationComputational EfficiencyTransformerLarge Language ModelText

🎯 What it does: A new post-training quantization algorithm called aespa is proposed, specifically targeting weight quantization for large-scale Transformer models, balancing accuracy and efficiency.

Towards Principled Graph Transformers

Luis MΓΌller (RWTH Aachen University), Christopher Morris (RWTH Aachen University)

CodeGraph Neural NetworkTransformerGraph

🎯 What it does: A graph learning framework based on Edge Transformer is proposed, achieving the expressiveness of 3-WL through triangular attention.

Towards Robust Multimodal Sentiment Analysis with Incomplete Data

Haoyu Zhang (Chinese University of Hong Kong), Tianshu Yu (Wuhan University)

CodeAdversarial AttackTransformerMultimodality

🎯 What it does: A robust model called LNLN is proposed to address data missingness in multimodal emotion analysis.

Towards Scalable and Stable Parallelization of Nonlinear RNNs

Xavier Gonzalez (Stanford University), Scott Linderman

CodeRecurrent Neural NetworkTime SeriesSequential

🎯 What it does: This paper proposes a scalable and stable parallelization method for nonlinear RNNs, overcoming the computational complexity and numerical instability issues of the original DEER method.

Towards the Transferability of Rewards Recovered via Regularized Inverse Reinforcement Learning

Andreas Schlaginhaufen (Γ‰cole Polytechnique FΓ©dΓ©rale de Lausanne), Maryam Kamgarpour (Γ‰cole Polytechnique FΓ©dΓ©rale de Lausanne)

CodeReinforcement Learning

🎯 What it does: This study investigates the transferability of reward functions learned from limited expert demonstrations within the framework of regularized Inverse Reinforcement Learning (IRL) across different environments (with different transition dynamics), proposing new theoretical conditions and algorithms.

Towards training digitally-tied analog blocks via hybrid gradient computation

Timothy Nest (Montreal Institute of Learning Algorithms), Maxence Ernoult (Rain AI)

CodeOptimizationImage

🎯 What it does: A mixed digital-analog feedforward coupling energy model (ff-EBMs) is proposed, along with an end-to-end BP-EP gradient chain algorithm for training deep networks on hardware mixed architectures.