NeurIPS 2023 Papers — Page 29
Conference on Neural Information Processing Systems · 3218 papers
SwiFT: Swin 4D fMRI Transformer
Peter Yongho Kim (Seoul National University), Taesup Moon (Seoul National University)
ClassificationRepresentation LearningTransformerContrastive LearningBiomedical DataMagnetic Resonance Imaging
🎯 What it does: A novel 4D Swin Transformer model, SwiFT, has been developed for end-to-end learning of brain spatiotemporal dynamics directly from high-dimensional fMRI volumes, as well as for biological and cognitive predictions.
SwiftSage: A Generative Agent with Fast and Slow Thinking for Complex Interactive Tasks
Bill Yuchen Lin (Allen Institute for Artificial Intelligence), Xiang Ren (University of Southern California)
Anomaly DetectionOptimizationRobotic IntelligenceTransformerLarge Language ModelSupervised Fine-TuningAgentic AIText
🎯 What it does: This paper studies a generative agent SWIFTSAGE based on dual process theory for complex interactive reasoning tasks.
Switching Autoregressive Low-rank Tensor Models
Hyun Dong Lee (Stanford University), Scott Linderman
VideoTime Series
🎯 What it does: The Switching Autoregressive Low-rank Tensor (SALT) model is proposed, which constrains the autoregressive tensor of ARHMM with low-rank tensor decomposition, combining the interpretability of ARHMM with the parameter efficiency of SLDS.
Switching Temporary Teachers for Semi-Supervised Semantic Segmentation
Jaemin Na (Ajou University), Wonjun Hwang (Ajou University)
SegmentationDomain AdaptationImage
🎯 What it does: Proposes a Dual Teacher framework that uses dual temporary teachers to alternately train a single student model, alleviating the teacher-student coupling problem.
Symbol-LLM: Leverage Language Models for Symbolic System in Visual Human Activity Reasoning
Xiaoqian Wu (Shanghai Jiao Tong University), Cewu Lu (Shanghai Jiao Tong University)
RecognitionObject DetectionExplainability and InterpretabilityTransformerLarge Language ModelVision Language ModelImageMultimodality
🎯 What it does: Proposes the Symbol-LLM framework, which uses LLM to generate a wide range of symbols and rules that are logically sound, achieving System-2 reasoning in visual activity understanding.
Symbolic Discovery of Optimization Algorithms
Xiangning Chen (Google), Quoc V Le
OptimizationDiffusion modelImageText
🎯 What it does: This paper formalizes the problem of optimizer discovery as a program search, utilizing techniques such as evolutionary search, abstract execution, funnel-based filtering, and program simplification to automatically discover a new optimizer called Lion, which is validated across various vision, vision-language, diffusion model, and language modeling tasks.
SyncDiffusion: Coherent Montage via Synchronized Joint Diffusions
Yuseung Lee (KAIST), Minhyuk Sung (KAIST)
GenerationData SynthesisDiffusion modelImage
🎯 What it does: A pluggable synchronization module (SYNCDIFFUSION) is proposed, which generates coherent montages (such as panoramas, textures, conditional images, etc.) by performing gradient descent updates on noisy images during the joint diffusion process.
SyncTREE: Fast Timing Analysis for Integrated Circuit Design through a Physics-informed Tree-based Graph Neural Network
Yuting Hu (University at Buffalo), Jinjun Xiong (University at Buffalo)
Graph Neural NetworkContrastive LearningGraph
🎯 What it does: A tree-structured graph neural network called SyncTREE is proposed for fast and accurate prediction of the timing (delay and slope) of integrated circuit interconnect RC trees.
Synthetic Combinations: A Causal Inference Framework for Combinatorial Interventions
Abhineet Agarwal (University of California), Suhas Vijaykumar (Amazon)
Tabular
🎯 What it does: This paper proposes a new causal inference framework aimed at learning the potential outcomes of N heterogeneous units under p intervention combinations, avoiding the high cost of conducting N × 2^p experiments.
Synthetic Experience Replay
Cong Lu (University of Oxford), Jack Parker-Holder (University of Oxford)
Data SynthesisReinforcement LearningDiffusion modelScore-based ModelImageTabular
🎯 What it does: This paper proposes a Synthetic Experience Replay (SYNTHER) method based on diffusion models for synthesizing and upsampling experience data in both offline and online reinforcement learning, aimed at enhancing sample efficiency and learning effectiveness.
Synthetic-to-Real Pose Estimation with Geometric Reconstruction
Qiuxia Lin (National University of Singapore), Angela Yao (National University of Singapore)
Pose EstimationDomain AdaptationConvolutional Neural NetworkImage
🎯 What it does: A framework for pose estimation that adapts from synthetic to real domains based on geometric reconstruction is proposed, combining pseudo-labels and reconstruction supervision to enhance pose accuracy.
Systematic Visual Reasoning through Object-Centric Relational Abstraction
Taylor Whittington Webb, Jonathan Cohen
RecognitionObject DetectionTransformerImage
🎯 What it does: Combining object-centered representation with relational abstraction, the OCRA model is proposed for visual reasoning.
T2T: From Distribution Learning in Training to Gradient Search in Testing for Combinatorial Optimization
Yang Li (Shanghai Jiao Tong University), Junchi Yan (Shanghai Jiao Tong University)
OptimizationGraph Neural NetworkDiffusion modelGraph
🎯 What it does: Proposed the T2T framework: during the training phase, a discrete diffusion model is used to learn the high-quality solution distribution for each instance, while in the testing phase, instance-level gradient search is performed by incorporating the target gradient into the diffusion denoising process to improve the initial solution.
TabMT: Generating tabular data with masked transformers
Manbir S Gulati, Paul F Roysdon
GenerationData SynthesisSafty and PrivacyTransformerTabular
🎯 What it does: This paper proposes TabMT, a table data generation model based on Masked Transformer, which can handle heterogeneous fields and naturally supports missing values.
Tackling Heavy-Tailed Rewards in Reinforcement Learning with Function Approximation: Minimax Optimal and Instance-Dependent Regret Bounds
Jiayi Huang (Peking University), Lin Yang
Reinforcement Learning
🎯 What it does: Two linear reinforcement learning algorithms aimed at large state-action spaces with heavy-tailed rewards are proposed: HEAVY-OFUL (linear bandit) and HEAVY-LSVI-UCB (linear MDP).
Tailoring Self-Attention for Graph via Rooted Subtrees
Siyuan Huang (Shanghai Jiaotong University), Zhouhan Lin (Shanghai Jiaotong University)
ClassificationGraph Neural NetworkTransformerGraph
🎯 What it does: Proposes a Subtree Attention Mechanism (STA) and designs STAGNN based on it for node classification tasks.
Taking the neural sampling code very seriously: A data-driven approach for evaluating generative models of the visual system
Suhas Shrinivasan (Institute for Computer Science and Campus Institute for Data Science University of Göttingen), Fabian H. Sinz (Institute for Computer Science and Campus Institute for Data Science University of Göttingen)
GenerationData SynthesisFlow-based ModelImage
🎯 What it does: Modeling the responses of visual neurons to natural images, formalizing the NSC as a learnable generative model, fitting neural data using deep generative models and classical models, inferring posteriors, and evaluating predictive performance.
Tame a Wild Camera: In-the-Wild Monocular Camera Calibration
Shengjie Zhu (Michigan State University), Xiaoming Liu (Michigan State University)
Object DetectionPose EstimationAutonomous DrivingNeural Radiance FieldSimultaneous Localization and MappingImage
🎯 What it does: A method for unconstrained calibration of monocular camera intrinsic parameters based on pixel-level incident field learning is proposed, which remains invariant after image cropping and scaling.
Taming Local Effects in Graph-based Spatiotemporal Forecasting
Andrea Cini (Università della Svizzera italiana), Cesare Alippi (Politecnico di Milano)
Graph Neural NetworkGraphTime Series
🎯 What it does: A theoretical and experimental framework is proposed to explore the interaction between global models and local effects in Spatio-Temporal Graph Neural Networks (STGNN), using learnable node embeddings as a means of achieving local specialization through 'sparsification'; two embedding space regularization methods (variational and clustering) are also provided to enhance transfer learning effectiveness.
Tanh Works Better with Asymmetry
Dongjin Kim (Korea University), Suhyun Kim (Korea Institute of Science and Technology)
ClassificationConvolutional Neural NetworkImage
🎯 What it does: This study explores whether placing Batch Normalization after the activation function (Swap) when using bounded activation functions can improve model performance. It finds that this approach produces asymmetric saturation and high sparsity, allowing bounded activation functions to approximate ReLU, significantly increasing accuracy.
Tanimoto Random Features for Scalable Molecular Machine Learning
Austin Tripp (University of Cambridge), José Miguel Hernández-Lobato (University of Cambridge)
Drug DiscoveryGraphBenchmark
🎯 What it does: Two random feature methods are proposed to approximate the Tanimoto kernel (TMM) in molecular fingerprints and its new extended kernel TDP. Theoretical upper bounds on the error are provided and evaluated on real molecular data.
TART: A plug-and-play Transformer module for task-agnostic reasoning
Kush Bhatia (Stanford University), Christopher Re
ClassificationTransformerLarge Language ModelSupervised Fine-TuningImageTextAudio
🎯 What it does: This study investigates the performance gap between large language models (LLMs) in context learning and task-specific fine-tuning, and proposes a general reasoning module named TART, which enhances the model's reasoning ability by combining a Transformer trained on synthetic logistic regression tasks with the embeddings of any pre-trained model.
Task Arithmetic in the Tangent Space: Improved Editing of Pre-Trained Models
Guillermo Ortiz-Jimenez (École Polytechnique Fédérale de Lausanne), Pascal Frossard (École Polytechnique Fédérale de Lausanne)
ClassificationRecognitionTransformerVision Language ModelImageMultimodality
🎯 What it does: The research improved the task arithmetic editing method of pre-trained models, enabling the model to achieve multi-task learning and task forgetting through addition and subtraction of task vectors.
Task-aware Distributed Source Coding under Dynamic Bandwidth
Po-han Li (University of Texas at Austin), Hyeji Kim (University of Texas at Austin)
Object DetectionCompressionRobotic IntelligenceAuto EncoderImage
🎯 What it does: This paper proposes a task-aware distributed source coding framework called NDPCA, which can dynamically adapt to any available bandwidth with a single model in multi-sensor networks, while maintaining high performance for downstream tasks such as denoising, robotic arm grasping, and satellite target detection after compression.
Task-aware world model learning with meta weighting via bi-level optimization
Huining Yuan (Beihang University), Yue Deng (Beihang University)
OptimizationMeta LearningReinforcement LearningWorld ModelSequential
🎯 What it does: A task-aware world model learning framework TEMPO based on dual-layer optimization is proposed, which uses a meta-weight network to perform task-aware weighting of training samples;
Task-Robust Pre-Training for Worst-Case Downstream Adaptation
Jianghui Wang (Peking University), Zhouchen Lin (Peking University)
SegmentationDepth EstimationDomain AdaptationOptimizationTransformerAuto EncoderImageText
🎯 What it does: A pre-training method based on minimax loss has been designed and implemented, enabling the base model to achieve balanced worst-case performance across multiple downstream tasks.
TaskMet: Task-driven Metric Learning for Model Learning
Dishank Bansal (Meta), Brandon Amos (Meta)
OptimizationExplainability and InterpretabilityReinforcement LearningTabularFinance Related
🎯 What it does: This paper proposes a task-driven metric learning method called TaskMet, which allows the model to learn interpretable metrics using downstream task losses while maintaining the original prediction space, thereby improving task performance.
Taylor TD-learning
Michele Garibbo (University of Bristol), Laurence Aitchison (University of Bristol)
Reinforcement LearningSequential
🎯 What it does: A model-based TD learning framework called Taylor TD is proposed, which uses first-order Taylor expansion to analytically integrate the uncertainties of actions and states, significantly reducing the variance of TD updates, and integrates it into the TD3 algorithm to form TaTD3.
TD Convergence: An Optimization Perspective
Kavosh Asadi (Amazon), Rasool Fakoor (Amazon)
OptimizationReinforcement Learning
🎯 What it does: This paper explains the convergence of TD learning from an optimization perspective, proving that TD can converge under various function approximations and error metrics;
Team-PSRO for Learning Approximate TMECor in Large Team Games via Cooperative Reinforcement Learning
Stephen Marcus McAleer, Tuomas Sandholm (Carnegie Mellon University)
Reinforcement LearningVideoTabular
🎯 What it does: Two algorithms, Team-PSRO and Team-PSRO-Mix-and-Match, are proposed to extend PSRO for learning approximate TMECor in two-team zero-sum games, achieving scalable training in large-scale team games such as Google Research Football.
Temperature Balancing, Layer-wise Weight Analysis, and Neural Network Training
Yefan Zhou (Dartmouth), Yaoqing Yang (Nanjing University)
OptimizationConvolutional Neural NetworkImage
🎯 What it does: This paper proposes and implements TempBalance, a hierarchical learning rate scheduling method based on the Heavy-Tail Self-Regularization theory, which enhances training effectiveness by dynamically adjusting the 'temperature' of each layer.
Template-free Articulated Neural Point Clouds for Reposable View Synthesis
Lukas Uzolas (Delft University of Technology), Petr Kellnhofer (Delft University of Technology)
GenerationData SynthesisPose EstimationNeural Radiance FieldGaussian SplattingVideoPoint Cloud
🎯 What it does: A forward deformation-based neural point cloud and linear blend skinning (LBS) model is proposed, capable of learning relocatable dynamic NeRF from sparse multi-view videos;
TempME: Towards the Explainability of Temporal Graph Neural Networks via Motif Discovery
Jialin Chen (Yale University), Zhitao Ying
Explainability and InterpretabilityGraph Neural NetworkGraphTime Series
🎯 What it does: This paper proposes TempME, an interpretable framework for temporal graph neural networks based on the information bottleneck, which utilizes temporal motifs to generate compact and consistent explanation subgraphs.
Tempo Adaptation in Non-stationary Reinforcement Learning
Hyunin Lee (University of California Berkeley), Somayeh Sojoudi (University of California Berkeley)
Reinforcement LearningSequential
🎯 What it does: This paper proposes a framework for addressing the issue of asynchronous timing between agents and environments in non-stationary reinforcement learning, called Proactively Synchronizing Tempo (ProST), and provides theoretical and experimental validation.
Temporal Causal Mediation through a Point Process: Direct and Indirect Effects of Healthcare Interventions
Çağlar Hızlı (Aalto University), Pekka Marttinen (Aalto University)
Time SeriesBiomedical DataElectronic Health Records
🎯 What it does: A dynamic causal mediation analysis framework based on point process is proposed to separate the direct and indirect effects of medical interventions on continuous time outcomes.
Temporal Conditioning Spiking Latent Variable Models of the Neural Response to Natural Visual Scenes
Gehua Ma (Zhejiang University), Huajin Tang (Zhejiang University)
Spiking Neural NetworkVideo
🎯 What it does: A time-conditioned impulse latent variable model (TeCoS-LVM) is proposed to simulate the firing of retinal neurons under natural visual stimuli, directly outputting discrete spike sequences.
Temporal Continual Learning with Prior Compensation for Human Motion Prediction
Jianwei Tang (Sun Yat-sen University), Jian-Fang Hu (Sun Yat-sen University)
Pose EstimationRecurrent Neural NetworkGraph Neural NetworkTransformerTime SeriesSequential
🎯 What it does: A multi-stage continual learning framework called Temporal Continual Learning (TCL) has been designed and implemented, introducing a Prior Compensation Factor (PCF) to alleviate knowledge forgetting, aimed at human motion prediction.
Temporal Dynamic Quantization for Diffusion Models
Junhyuk So (POSTECH), Eunhyeok Park (POSTECH)
GenerationCompressionComputational EfficiencyDiffusion modelImage
🎯 What it does: A Temporal Dynamic Quantization (TDQ) method is proposed to quantize the activations of diffusion models, aiming to reduce storage and computational costs, especially suitable for low-precision bit widths.
Temporal Robustness against Data poisoning
Wenxiao Wang (University of Maryland), Soheil Feizi (University of Maryland)
Adversarial AttackData-Centric LearningTransformerLarge Language ModelTextTime Series
🎯 What it does: This study investigates the introduction of timestamps in data to model data poisoning attacks, proposing a time-based attack budget (earliness and duration) and corresponding definitions of temporal robustness.
Temporally Disentangled Representation Learning under Unknown Nonstationarity
Xiangchen Song (Carnegie Mellon University), Kun Zhang (Carnegie Mellon University)
Representation LearningRecurrent Neural NetworkAuto EncoderVideoTime SeriesSequential
🎯 What it does: This study investigates learning identifiable causal representations from unsupervised sequential data using time-delayed causal relationships under unknown non-stationary distributions.
TensorNet: Cartesian Tensor Representations for Efficient Learning of Molecular Potentials
Guillem Simeon (Universitat Pompeu Fabra), Gianni De Fabritiis (Universitat Pompeu Fabra)
Drug DiscoveryGraph Neural NetworkTabularPhysics Related
🎯 What it does: A Cartesian second-order tensor-based O(3)-equivariant message passing network, TensorNet, is proposed for learning molecular potential energy, forces, and other vector/tensor properties.
Test-Time Amendment with a Coarse Classifier for Fine-Grained Classification
Kanishk Jain (International Institute of Information Technology Hyderabad), Vineet Gandhi (University of Tübingen)
ClassificationImage
🎯 What it does: A Hierarchical Ensembles (HiE) post-correction method is proposed: fine-grained and coarse-grained classifiers are trained separately, and during inference, the fine-grained predictions are multiplied by the corresponding parent class probabilities and normalized, thereby reducing error severity and improving Top-1 accuracy.
Test-Time Distribution Normalization for Contrastively Learned Visual-language Models
Yifei Zhou (University of California), Ser-Nam Lim (University of Central Florida)
ClassificationRetrievalVision Language ModelContrastive LearningImageText
🎯 What it does: A technique using Distribution Mean Normalization (DN) during the inference phase is proposed and evaluated to align with the InfoNCE objective of contrastive learning models (such as CLIP) during training, improving the performance of traditional dot product similarity.
Test-time Training for Matching-based Video Object Segmentation
Juliette Bertrand (Czech Technical University in Prague), Giorgos Tolias (Czech Technical University in Prague)
Object DetectionSegmentationSupervised Fine-TuningVideoBenchmark
🎯 What it does: For the task of video object segmentation (VOS) based on matching, a scheme for adaptive fine-tuning of the model during inference is proposed and validated, known as Test-time Training (TTT).
Tester-Learners for Halfspaces: Universal Algorithms
Aravind Gollakota (University of Texas at Austin), Arsen Vasilyan (Massachusetts Institute of Technology)
🎯 What it does: The first universal half-space testing-learning algorithm is proposed for a wide range of structured distributions, including all strongly log-convex distributions and all log-convex distributions under the KLS conjecture;
Testing the General Deductive Reasoning Capacity of Large Language Models Using OOD Examples
Abulhair Saparov (New York University), He He (Google)
TransformerLarge Language ModelPrompt EngineeringTextChain-of-Thought
🎯 What it does: This paper systematically evaluates the general deductive reasoning capabilities of large language models (LLMs) in out-of-distribution (OOD) inference by constructing a programmable synthetic dataset PRONTOQA-OOD, focusing on reasoning rules, proof depth, width, and the generalization of combinatorial reasoning.
TexQ: Zero-shot Network Quantization with Texture Feature Distribution Calibration
Xinrui Chen (Tsinghua University), Yonghong He (Tsinghua University)
CompressionKnowledge DistillationConvolutional Neural NetworkGenerative Adversarial NetworkImage
🎯 What it does: A zero-shot quantization method called TexQ is proposed, which utilizes texture feature distribution to calibrate synthetic samples and quantize the network;
Text Alignment Is An Efficient Unified Model for Massive NLP Tasks
Yuheng Zha (University of California San Diego), Zhiting Hu (University of California San Diego)
ClassificationData SynthesisTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: A text alignment model (ALIGN) is proposed, which unifies various NLP tasks into a task that measures the degree of information alignment between two pieces of text.
Text Promptable Surgical Instrument Segmentation with Vision-Language Models
Zijian Zhou (King's College London), Miaojing Shi (Tongji University)
Object DetectionSegmentationTransformerPrompt EngineeringVision Language ModelImageMultimodality
🎯 What it does: A text prompt-based surgical instrument segmentation method is proposed, transforming the segmentation task into an image + text input, utilizing a vision-language model for fine recognition of different instruments.
Text-to-Image Diffusion Models are Zero Shot Classifiers
Kevin Clark (Google DeepMind), Priyank Jaini (Google DeepMind)
ClassificationGenerationTransformerDiffusion modelImageText
🎯 What it does: This paper proposes a method to directly use text-to-image diffusion models (such as Imagen and Stable Diffusion) as zero-shot classifiers and quantifies the model's discriminative ability through this method.
TextDiffuser: Diffusion Models as Text Painters
Jingye Chen (Hong Kong University of Science and Technology), Furu Wei (Microsoft Research)
RecognitionGenerationData SynthesisTransformerDiffusion modelImageTextBenchmark
🎯 What it does: A two-stage diffusion model called TextDiffuser is proposed, which first uses a Transformer to generate keyword layouts, and then employs a diffusion model combined with character-level segmentation masks to generate images with accurate and coherent text, supporting text filling.
Textually Pretrained Speech Language Models
Michael Hassid (Hebrew University of Jerusalem), Yossi Adi (Hebrew University of Jerusalem)
RecognitionGenerationTransformerLarge Language ModelBenchmarkAudio
🎯 What it does: This paper proposes a warm-init method called TWIST for Speech Language Model (SpeechLM) based on a pre-trained text language model, and trains the largest SpeechLM (7B/13B) model on a large-scale speech dataset, while releasing two versions of the speech-based StoryCloze benchmark.
TFLEX: Temporal Feature-Logic Embedding Framework for Complex Reasoning over Temporal Knowledge Graph
Xueyuan Lin (Beijing University of Posts and Telecommunications), Mingzhi Sun (Beijing University of Posts and Telecommunications)
Graph Neural NetworkGraphTime Series
🎯 What it does: This paper proposes the TFLEX framework, which can perform multi-hop logical reasoning on temporal knowledge graphs (TKG), supporting entity queries and temporal queries, covering all first-order logical operations as well as temporal operations such as After, Before, and Between.
The Adversarial Consistency of Surrogate Risks for Binary Classification
Natalie Frank, Jonathan Niles-Weed (New York University)
ClassificationOptimizationAdversarial Attack
🎯 What it does: This study addresses the consistency of approximating the 0-1 loss with surrogate loss in adversarial robust learning for binary classification problems, specifically proving that the minimization sequence of adversarial risk can also minimize adversarial classification risk.
The Bayesian Stability Zoo
Shay Moran (Technion - Israel Institute of Technology), Jonathan Shafer (University of California Berkeley)
🎯 What it does: This paper explores the interrelationships among various definitions of stability in learning theory and constructs a complete equivalence network divided into two main categories: distribution-dependent and distribution-independent. It also proposes a boosting algorithm that can enhance weak learners under distribution-independent KL stability.
The Behavior and Convergence of Local Bayesian Optimization
Kaiwen Wu (University of Pennsylvania), Jacob R. Gardner (University of Pennsylvania)
Optimization
🎯 What it does: This paper conducts a systematic study of Local Bayesian Optimization (LBO), providing empirical experimental validation that local optimization can achieve very good local solutions on high-dimensional Gaussian process (GP) sample paths, and also presents a rigorous convergence rate proof for the recently proposed GIBO algorithm in both noise-free and noisy environments.
The Benefits of Being Distributional: Small-Loss Bounds for Reinforcement Learning
Kaiwen Wang (Cornell University), Wen Sun (Cornell University)
Reinforcement LearningTabular
🎯 What it does: This paper provides theoretical guarantees for distributed reinforcement learning (DistRL) in the context of small losses, and proposes corresponding online and offline algorithms as well as a distributed contextual bandit (DistCB).
The Best of Both Worlds in Network Population Games: Reaching Consensus and Convergence to Equilibrium
Shuyue Hu (Shanghai Artificial Intelligence Laboratory), Georgios Piliouras (Singapore University of Technology and Design)
🎯 What it does: The paper studies the use of Smooth Fictional Play (SFP) to achieve consistency and equilibrium convergence in network population games.
The CLIP Model is Secretly an Image-to-Prompt Converter
Yuxuan Ding (Xidian University), Lingqiao Liu (Australian Institute for Machine Learning)
GenerationData SynthesisTransformerPrompt EngineeringDiffusion modelImageText
🎯 What it does: By utilizing the inverse matrix of the visual projection layer of CLIP, images are directly mapped to text prompts usable for Stable Diffusion, achieving instant conversion from images to prompts, and providing two extensions based on this: lightweight fine-tuning (SD‑IPC‑FT) and online customization (SD‑IPC‑CT).
The Clock and the Pizza: Two Stories in Mechanistic Explanation of Neural Networks
Ziqian Zhong (Massachusetts Institute of Technology), Jacob Andreas (Massachusetts Institute of Technology)
TransformerTabular
🎯 What it does: The study investigates how neural networks implement the known Clock algorithm and the new Pizza algorithm when training for modular addition, and explores the phase transitions of the algorithms under different hyperparameters.
The Contextual Lasso: Sparse Linear Models via Deep Neural Networks
Ryan Thompson (University of New South Wales), Robert Kohn (University of New South Wales)
Supervised Fine-TuningTabular
🎯 What it does: This paper proposes a context-sparse linear model and the corresponding Context Lasso estimator, utilizing feedforward neural networks to learn context features and sparse coefficient functions, and implementing a projection layer to impose a constraint on the average ℓ₁ norm, balancing interpretability and expressive power.
The Crucial Role of Normalization in Sharpness-Aware Minimization
Yan Dai (Tsinghua University), Suvrit Sra (Massachusetts Institute of Technology)
OptimizationConvolutional Neural NetworkImage
🎯 What it does: This study investigates the impact of the normalization factor in Sharpness-Aware Minimization (SAM) on algorithm stability and the drift behavior along the minima manifold.
The Curious Price of Distributional Robustness in Reinforcement Learning with a Generative Model
Laixi Shi (California Institute of Technology), Yuejie Chi (Carnegie Mellon University)
Reinforcement Learning
🎯 What it does: This paper studies the distributional robustness in reinforcement learning under the framework of generative models, particularly the sample complexity of distributionally robust Markov decision processes (RMDPs).
The Distortion of Binomial Voting Defies Expectation
Yannai Gonczarowski, Shirley Zhang (Harvard University)
🎯 What it does: The research addresses the expected distortion problem based on distribution in voting rules and proposes a new distribution-independent voting rule—Binomial Voting, providing theoretical guarantees for its expected distortion and expected welfare.
The Double-Edged Sword of Implicit Bias: Generalization vs. Robustness in ReLU Networks
Spencer Frei (University of California Davis), Nathan Srebro
Recurrent Neural NetworkTabularOrdinary Differential Equation
🎯 What it does: This paper studies the implicit bias of gradient flow in training clustering data with two-layer ReLU networks, proving its convergence to the KKT points of the maximum margin problem, thereby achieving good generalization, but leading to a lack of robustness, which remains unrobust even in over-parameterized cases.
The emergence of clusters in self-attention dynamics
Borjan Geshkovski (Massachusetts Institute of Technology), Philippe Rigollet (Massachusetts Institute of Technology)
TransformerSequentialOrdinary Differential Equation
🎯 What it does: This paper treats the self-attention layer of the Transformer as an interactive particle system, studying the geometric limit of token evolution through layer iterations when the weights do not change over time, proving that they converge and are distributed around a finite number of 'cluster points' or hyperplanes.
The Emergence of Essential Sparsity in Large Pre-trained Models: The Weights that Matter
AJAY KUMAR JAISWAL, Zhangyang Wang (University of Texas at Austin)
TransformerLarge Language ModelVision Language ModelMultimodality
🎯 What it does: This study investigates the phenomenon of 'essential sparsity' in large pre-trained visual and language Transformers, which maintains high performance even after a single round of direct pruning.
The Equivalence of Dynamic and Strategic Stability under Regularized Learning in Games
Victor Boone (University of Grenoble Alpes), Panayotis Mertikopoulos (University of Grenoble Alpes)
🎯 What it does: The paper conducts a theoretical analysis of the long-term behavior of regularized no-regret learning in finite games, proposing and proving an equivalence between the 'club set' and the stability of learning dynamics, and providing convergence rate estimates for different regularization methods (such as entropy regularization, projection regularization, etc.).
The Exact Sample Complexity Gain from Invariances for Kernel Regression
Behrooz Tahmasebi (Massachusetts Institute of Technology), Stefanie Jegelka (Massachusetts Institute of Technology)
Point Cloud
🎯 What it does: This study investigates the precise improvement of sample complexity when using Kernel Ridge Regression (KRR) on compact manifolds with Lie group invariance; it provides the optimal convergence rate and lower bound for measurable invariance.
The expressive power of pooling in Graph Neural Networks
Filippo Maria Bianchi (UiT Arctic University of Norway), Veronica Lachi (University of Siena)
Graph Neural NetworkGraph
🎯 What it does: This study investigates the expressive power of pooling layers in graph neural networks and provides sufficient conditions for maintaining the expressiveness of MP layers, verifying whether different pooling methods meet these conditions. Additionally, a synthetic dataset EXPWL1 based on the WL test is proposed for empirical testing of the expressiveness of pooling layers.
The Gain from Ordering in Online Learning
Vasilis Kontonis (University of Texas at Austin), Christos Tzamos (University of Wisconsin Madison)
OptimizationPoint Cloud
🎯 What it does: This study investigates linear regression and ReLU regression in self-directed online learning (self-guided), exploring how to reduce cumulative loss by selecting the order of samples.
The geometry of hidden representations of large transformer models
Lucrezia Valeriani (AREA Science Park), Alberto Cazzaniga (AREA Science Park)
Representation LearningProtein Structure PredictionTransformerLarge Language ModelImageBiomedical Data
🎯 What it does: This study investigates the geometric and statistical properties of hidden representations at different levels in large self-supervised Transformers, analyzing their evolution with respect to intrinsic dimension (ID) and neighborhood overlap.
The Geometry of Neural Nets' Parameter Spaces Under Reparametrization
Agustinus Kristiadi (Vector Institute), Philipp Hennig (University of Tübingen)
OptimizationConvolutional Neural NetworkImageOrdinary Differential Equation
🎯 What it does: This paper explores the geometric invariance of neural networks under reparametrization, proposing that explicitly using Riemannian metrics and following the corresponding transformation rules can ensure the invariance of quantities such as the Hessian, gradient descent trajectories, and probability density patterns. This is validated in scenarios such as infinitely wide networks, Laplace marginal likelihood, and preconditioned optimizers.
The Goldilocks of Pragmatic Understanding: Fine-Tuning Strategy Matters for Implicature Resolution by LLMs
Laura Eline Ruis, Edward Grefenstette (University College London)
TransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextChain-of-Thought
🎯 What it does: This study constructs a binary implicature resolution task and systematically evaluates various large language models (baseline, dialogue fine-tuning, benchmark instruction fine-tuning, example instruction fine-tuning) to explore their reasoning abilities regarding conversational implicature.
The Grand Illusion: The Myth of Software Portability and Implications for ML Progress.
Fraser Mince (Cohere for AI Community), Sara Hooker (Cohere for AI)
Benchmark
🎯 What it does: Evaluate the portability and performance differences between TensorFlow, PyTorch, and JAX on GPU and TPU, and release a public test set.
The Graph Pencil Method: Mapping Subgraph Densities to Stochastic Block Models
Lee M. Gunderson (University College London), Peter Orbanz (University College London)
Graph Neural NetworkGraph
🎯 What it does: The Graph Pencil Method is proposed, which accurately maps the densities of finite star and double star subgraphs to the parameters of the Stochastic Block Model (SBM) through root subgraph algebra and matrix pencil techniques, achieving direct inference with no significant computational overhead.
The Impact of Positional Encoding on Length Generalization in Transformers
Amirhossein Kazemnejad (Mila), Siva Reddy (Mila)
TransformerSequentialChain-of-Thought
🎯 What it does: This study systematically evaluates the impact of different positional encoding schemes on the decoder Transformer in length generalization (inferring long sequences from short sequences).
The Learnability of In-Context Learning
Noam Wies (Hebrew University of Jerusalem), Amnon Shashua (Hebrew University of Jerusalem)
Large Language Model
🎯 What it does: A learnability framework for in-context learning based on PAC theory is proposed, and a proof is provided that a frozen model can achieve limited sample complexity for downstream tasks through a small number of context examples when pre-trained on a mixed task distribution.
The Memory-Perturbation Equation: Understanding Model's Sensitivity to Data
Peter Nickl (RIKEN Center for AI Project), Mohammad Emtiyaz Khan (RIKEN Center for AI Project)
OptimizationConvolutional Neural NetworkImage
🎯 What it does: Proposes the Memory-Perturbation Equation (MPE), which quickly estimates the model's sensitivity to training data through natural gradient and uses this estimate to predict generalization performance during training.
The noise level in linear regression with dependent data
Ingvar Ziemann (University of Pennsylvania), Nikolai Matni (University of Pennsylvania)
🎯 What it does: This study investigates the non-asymptotic error upper bound of OLS linear regression under β-mixing dependent data without the assumption of realizability.
The Pick-to-Learn Algorithm: Empowering Compression for Tight Generalization Bounds and Improved Post-training Performance
Dario Paccagnan (Imperial College London), Simone Garatti (Politecnico di Milano)
ClassificationCompressionConvolutional Neural NetworkImageTabular
🎯 What it does: By constructing the Pick-to-Learn (P2L) meta-algorithm, any learning algorithm is transformed into a learning process with compression properties, thereby obtaining compact generalization bounds and improving post-training performance.
The probability flow ODE is provably fast
Sitan Chen (Harvard University), Adil Salim (Microsoft Research)
GenerationData SynthesisOptimizationDiffusion modelScore-based ModelStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: This paper presents, for the first time, a polynomial time convergence guarantee for the score-based generative model (SGM) implemented via probability flow ODE under the OU forward process, and proposes two prediction-correction algorithms.
The Pursuit of Human Labeling: A New Perspective on Unsupervised Learning
Artyom Gadetsky (École Polytechnique Fédérale de Lausanne), Maria Brbic
ClassificationRepresentation LearningMeta LearningContrastive LearningImage
🎯 What it does: A novel unsupervised framework called HUME is designed and implemented to infer the true human labels of datasets by training linear classifiers on different pre-trained representation spaces and evaluating their generalizability to search for optimal labels.
The Quantization Model of Neural Scaling
Eric J Michaud, Max Tegmark (Massachusetts Institute of Technology)
TransformerLarge Language ModelText
🎯 What it does: A quantization model is proposed to explain the power-law decline of neural network scaling and the sudden emergence of new capabilities.
The Rank-Reduced Kalman Filter: Approximate Dynamical-Low-Rank Filtering In High Dimensions
Jonathan Schmidt (University of Tübingen), Filip Tronarp (Lund University)
Time SeriesStochastic Differential Equation
🎯 What it does: A low-rank Kalman filter (RRKF) is proposed, which achieves approximate Gaussian filtering and posterior estimation in high-dimensional state spaces by maintaining a low-rank approximation of the covariance.
The Rashomon Importance Distribution: Getting RID of Unstable, Single Model-based Variable Importance
Jon Donnelly (Duke University), Edward P Browne
Biomedical Data
🎯 What it does: This study investigates how to quantify variable importance through the calculation of Rashomon Importance Distribution (RID) in the presence of the Rashomon effect and data instability.
The Rise of AI Language Pathologists: Exploring Two-level Prompt Learning for Few-shot Weakly-supervised Whole Slide Image Classification
Linhao Qu (Fudan University), Zhijian Song (Fudan University)
ClassificationTransformerLarge Language ModelPrompt EngineeringImageBiomedical Data
🎯 What it does: With a small amount of bag-level annotations, the problem of few-shot weakly supervised classification of whole slide images (WSI) is addressed through dual-layer prompt learning.
The s-value: evaluating stability with respect to distributional shifts
Suyash Gupta (Stanford University), Dominik Rothenhaeusler
Domain AdaptationOptimizationTabular
🎯 What it does: This paper proposes the s-value metric to quantify the stability of statistical parameters under distribution shifts, and constructs overall and directional instability through KL divergence; it also provides estimable formulas for mean and directional s-values and applies them to parameter transfer (transfer learning).
The Shaped Transformer: Attention Models in the Infinite Depth-and-Width Limit
Lorenzo Noci (ETH Zurich), Daniel M. Roy (University of Toronto)
TransformerSequentialStochastic Differential Equation
🎯 What it does: This study investigates the attention mechanism of the Transformer in the limit of infinite depth and width, proposing shaped attention and deriving its neural covariance SDE to address the rank collapse issue.
The Simplicity Bias in Multi-Task RNNs: Shared Attractors, Reuse of Dynamics, and Geometric Representation
Elia Turner (Israel Institute of Technology), Omri Barak (Israel Institute of Technology)
Recurrent Neural NetworkSequential
🎯 What it does: The study systematically investigates the Simplified Bias in multi-task RNNs by designing simplified tasks with a unified input-output structure (such as fixed points, limit cycles, and line/plane attractors) to observe how RNNs share or separate attractors under different training schemes (gated, orthogonal, parallel).
The Surprising Effectiveness of Diffusion Models for Optical Flow and Monocular Depth Estimation
Saurabh Saxena (Google DeepMind), David J. Fleet (University of Toronto)
Image TranslationDepth EstimationDiffusion modelOptical FlowImage
🎯 What it does: This work proposes a general denoising diffusion model (DDVM) that directly treats optical flow and monocular depth estimation as an image-to-image translation task, without the need for specialized network architectures or loss functions.
The Target-Charging Technique for Privacy Analysis across Interactive Computations
Edith Cohen (Google Research and Tel Aviv University), Xin Lyu (University of California Berkeley)
Safty and Privacy
🎯 What it does: This paper proposes the Target Charging Technique (TCT), a unified privacy analysis framework for differential privacy algorithms that repeatedly use the same sensitive dataset in interactive or batch processing environments, and provides a complete set of tools (such as NotPrior, ConditionalRelease, Boundary Wrapper, etc.) to implement privacy costs that are only charged for 'target hits'.
The Transient Nature of Emergent In-Context Learning in Transformers
Aaditya K Singh, Felix Hill (Google DeepMind)
TransformerText
🎯 What it does: This paper studies the emergence and disappearance of in-context learning (ICL) by training on artificially synthesized data (Omniglot and LLaMA token embeddings) using a Transformer. It finds that ICL is often transient and is replaced by instance-weighted learning (IWL);
The Tunnel Effect: Building Data Representations in Deep Neural Networks
Wojciech Masarczyk (Warsaw University of Technology), Tomasz Trzcinski (Warsaw University of Technology)
ClassificationRepresentation LearningConvolutional Neural NetworkImage
🎯 What it does: This study investigates the 'tunnel effect' that occurs during the training of deep neural networks, where the network is divided into an extractor and a compression tunnel.
The Utility of “Even if” Semifactual Explanation to Optimise Positive Outcomes
Eoin M. Kenny (Massachusetts Institute of Technology), Weipeng Fuzzy Huang
Recommendation SystemAnomaly DetectionOptimizationExplainability and InterpretabilityTabularFinance Related
🎯 What it does: A framework based on 'even if...' counterfactual explanations is proposed and implemented to optimize the benefits of positive decisions made by models (such as loan approvals), rather than just explaining negative outcomes.
Theoretical Analysis of the Inductive Biases in Deep Convolutional Networks
Zihao Wang (Peking University), Lei Wu (Peking University)
Convolutional Neural NetworkTabular
🎯 What it does: At the theoretical level, this paper systematically analyzes the prior biases of Convolutional Neural Networks (CNNs), focusing mainly on four structures: multi-channel, down-sampling, weight sharing, and locality, and studies their interaction with network depth.
Theoretical and Practical Perspectives on what Influence Functions Do
Andrea Schioppa (Google DeepMind), Polina Zablotskaia (Google DeepMind)
ClassificationOptimizationConvolutional Neural NetworkTransformerSupervised Fine-TuningImageText
🎯 What it does: Analyze and verify the theoretical assumptions and practical effects of Influence Functions (IF) in deep learning, revealing that they are only effective within a limited number of fine-tuning steps, and propose an IF-based misjudgment correction method.
Theoretically Guaranteed Bidirectional Data Rectification for Robust Sequential Recommendation
yatong sun, Yan Wang (Macquarie University)
Recommendation SystemConvolutional Neural NetworkRecurrent Neural NetworkTransformerSequential
🎯 What it does: A theoretically guaranteed bidirectional data correction framework (BirDRec) is proposed, which enhances the robustness of sequence recommendation systems by correcting unreliable targets and inputs through predicted scores.
Thin and deep Gaussian processes
Daniel Augusto de Souza (University College London), César Lincoln Mattos
Explainability and InterpretabilityComputational EfficiencyTabular
🎯 What it does: A new high-order Gaussian process model, called Thin and Deep GP (TDGP), is proposed to overcome the limitations of existing deep Gaussian process models, maintaining interpretability and learning low-dimensional embeddings.
Thinker: Learning to Plan and Act
Stephen Chung (University of Cambridge), David Krueger (University of Cambridge)
Robotic IntelligenceReinforcement LearningWorld ModelSequential
🎯 What it does: Proposes the Thinker algorithm, allowing RL agents to autonomously interact with the learned world model and perform planning, integrating model inference and decision-making;