NeurIPS 2023 Papers — Page 3
Conference on Neural Information Processing Systems · 3218 papers
Aging with GRACE: Lifelong Model Editing with Discrete Key-Value Adaptors
Thomas Hartvigsen (University of Virginia), Marzyeh Ghassemi (Massachusetts Institute of Technology)
TransformerSupervised Fine-TuningText
🎯 What it does: This paper addresses the issue of continuous and immediate local correction of deployed large models when errors occur, proposing a lifelong model editing method called GRACE. GRACE inserts discrete key-value adapters at a certain layer of the pre-trained model, achieving rapid and interpretable model behavior modification by storing the hidden representations (keys) of erroneous samples and their corresponding correction outputs (values), without altering the original weights.
Agnostic Multi-Group Active Learning
Nicholas Rittler, Kamalika Chaudhuri (University of California San Diego)
Optimization
🎯 What it does: This paper proposes an active learning framework suitable for multi-group learning, providing algorithms for both agnostic and group realizable scenarios, along with consistency proofs and upper bounds on label complexity.
Agnostically Learning Single-Index Models using Omnipredictors
Aravind Gollakota (Apple), Konstantinos Stavropoulos (University of Texas at Austin)
🎯 What it does: An efficient algorithm for agnostic learning of single-index models (SIM) under any monotonic Lipschitz activation function is proposed, along with a theoretical error upper bound.
AI for Interpretable Chemistry: Predicting Radical Mechanistic Pathways via Contrastive Learning
Mohammadamin Tavakoli (University of California), David Van Vranken (University of California)
Explainability and InterpretabilityDrug DiscoveryGraph Neural NetworkTransformerContrastive LearningGraphBenchmark
🎯 What it does: A free radical reaction prediction system RMechRP based on symmetric learning and mechanism chains has been developed, which can predict free radical mechanism steps at the single-step or pathway level and provide interpretable outputs such as arrows and SMIRKS.
AiluRus: A Scalable ViT Framework for Dense Prediction
Jin Li (Shanghai Jiao Tong University), Qi Tian (Huawei Cloud)
Object DetectionSegmentationTransformerImage
🎯 What it does: A pluggable adaptive resolution strategy AiluRus is proposed, which dynamically merges tokens using a spatially aware density clustering algorithm in the intermediate layers of ViT to accelerate dense prediction tasks.
Aiming towards the minimizers: fast convergence of SGD for overparametrized problems
Chaoyue Liu (Cornell University), Yian Ma (Cornell University)
OptimizationImage
🎯 What it does: This paper proposes a new 'aiming' regularization condition and proves that in over-parameterized problems, using this condition, stochastic gradient descent (SGD) can achieve the same linear convergence rate as deterministic gradient descent while maintaining a large step size.
AIMS: All-Inclusive Multi-Level Segmentation for Anything
Lu Qi (University of California Merced), Ming-Hsuan Yang (Google)
Object DetectionSegmentationTransformerImage
🎯 What it does: The AIMS (All-Inclusive Multi-Level Segmentation) task and unified model are proposed, capable of segmenting images into three levels: parts, entities, and relationships, and supporting mask-based interactive segmentation.
AlberDICE: Addressing Out-Of-Distribution Joint Actions in Offline Multi-Agent RL via Alternating Stationary Distribution Correction Estimation
Daiki E. Matsunaga (Korea Advanced Institute of Science and Technology), Kee-Eung Kim (Korea Advanced Institute of Science and Technology)
Reinforcement Learning
🎯 What it does: We propose AlberDICE, an algorithm for offline multi-agent reinforcement learning that utilizes alternating best responses and static distribution correction to avoid the out-of-distribution (OOD) problem of joint actions, overcoming the challenge of exponential explosion in the joint action space.
Aleatoric and Epistemic Discrimination: Fundamental Limits of Fairness Interventions
Hao Wang (Massachusetts Institute of Technology IBM Watson Artificial Intelligence Lab), Flavio Calmon
ClassificationOptimizationTabular
🎯 What it does: This paper proposes the concept of a fair Pareto frontier, characterizing the achievable set of predictive distributions C using information theory and Blackwell comparison experiments, and provides a greedy iterative algorithm to approximate this frontier, thereby quantifying the inherent aleatoric discrimination in the data and the epistemic discrimination caused by model selection.
ALGO: Synthesizing Algorithmic Programs with Generated Oracle Verifiers
Kexun Zhang (University of California Santa Barbara), Lei Li (Carnegie Mellon University)
OptimizationAI Code AssistantTransformerLarge Language ModelText
🎯 What it does: The ALGO framework is proposed, which uses brute-force enumeration reference oracles generated by LLMs to guide and verify the automatic synthesis of algorithm programs.
Algorithm Selection for Deep Active Learning with Imbalanced Datasets
Jifan Zhang (University of Wisconsin - Madison), Robert D Nowak
ClassificationData-Centric LearningMeta LearningImage
🎯 What it does: An adaptive algorithm selection framework TAILOR suitable for deep active learning is proposed, which can dynamically select the optimal active learning algorithm across different datasets.
Algorithmic Regularization in Tensor Optimization: Towards a Lifted Approach in Matrix Sensing
Ziye Ma (University of California Berkeley), Somayeh Sojoudi (University of California Berkeley)
Optimization
🎯 What it does: Under the lifted matrix sensing framework, this study investigates the implicit regularization effect of gradient descent (GD) on tensor optimization for approximate rank-1 representations and proves that it can lead to escape directions at constrained points.
Align Your Prompts: Test-Time Prompting with Distribution Alignment for Zero-Shot Generalization
Jameel Hassan Abdul Samadh (Mohamed Bin Zayed University of AI), Salman Khan (Australian National University)
Domain AdaptationTransformerPrompt EngineeringContrastive LearningImageMultimodality
🎯 What it does: This paper proposes a testing-time prompt tuning method called PromptAlign, which enhances the zero-shot generalization ability of CLIP by aligning the token statistics of the visual channel with the statistics of pre-trained data using a single test sample.
Aligning Gradient and Hessian for Neural Signed Distance Function
Ruian Wang (Shandong University), Wenping Wang (Texas A&M University)
GenerationOptimizationPoint Cloud
🎯 What it does: A gradient-Hessian alignment loss method is proposed for surface reconstruction in undirected point clouds where normal vectors cannot be obtained;
Aligning Language Models with Human Preferences via a Bayesian Approach
Jiashuo WANG (Hong Kong Polytechnic University), Wenjie Li
GenerationRecommendation SystemReinforcement LearningContrastive LearningText
🎯 What it does: A Bayesian framework d-PM is proposed to model inconsistencies in human preferences and to calibrate existing NLG models using contrastive learning, in order to generate text that is more universally acceptable and less controversial.
Aligning Optimization Trajectories with Diffusion Models for Constrained Design Generation
Giorgio Giannone (Massachusetts Institute of Technology), Faez Ahmed (Massachusetts Institute of Technology)
GenerationOptimizationDiffusion modelMesh
🎯 What it does: This paper proposes Diffusion Optimization Models (DOM), which achieve efficient generation of constrained designs by aligning the sampling trajectories of diffusion models with physics-based iterative optimization trajectories during the training phase.
Aligning Synthetic Medical Images with Clinical Knowledge using Human Feedback
Shenghuan Sun (University of California, San Francisco), Ahmed Alaa
GenerationData SynthesisReinforcement LearningDiffusion modelImageBiomedical Data
🎯 What it does: A pathologist-in-the-loop framework was constructed by incorporating binary feasibility feedback provided by pathologists during the generation of bone marrow cell images, to fine-tune a conditional diffusion model for generating more clinically feasible synthetic medical images.
Alignment with human representations supports robust few-shot learning
Ilia Sucholutsky (Princeton University), Thomas L. Griffiths (Princeton University)
Domain AdaptationRepresentation LearningAdversarial AttackMeta LearningImage
🎯 What it does: This paper explores and verifies how the alignment degree of models with human representations in the embedding space affects few-shot learning, adversarial robustness, and domain transfer performance. An information-theoretic framework is established through the evaluation of 491 pre-trained models and over 425,000 human similarity assessments, confirming a U-shaped relationship.
ALIM: Adjusting Label Importance Mechanism for Noisy Partial Label Learning
Mingyu Xu (Chinese Academy of Sciences), Jianhua Tao (Tsinghua University)
ClassificationContrastive LearningImage
🎯 What it does: To address the problem of partial label learning with noise, a framework is proposed that adjusts label importance (ALIM) to balance the candidate label set with model predictions, compatible with existing methods and capable of automatically adjusting weights.
All Points Matter: Entropy-Regularized Distribution Alignment for Weakly-supervised 3D Segmentation
Liyao Tang (University of Sydney), Dacheng Tao (La Trobe University)
SegmentationDomain AdaptationTransformerImagePoint Cloud
🎯 What it does: This study focuses on weakly supervised 3D point cloud semantic segmentation and proposes the ERDA learning strategy, which utilizes entropy regularization and distribution alignment to jointly generate more reliable pseudo-labels, fully leveraging all unlabeled points for training.
Alleviating the Semantic Gap for Generalized fMRI-to-Image Reconstruction
Tao Fang (Zhejiang University), Gang Pan (Zhejiang University)
GenerationData SynthesisMixture of ExpertsDiffusion modelImageBiomedical DataMagnetic Resonance Imaging
🎯 What it does: This paper proposes a general fMRI-to-image reconstruction framework called GESS, aimed at eliminating the semantic gap between training and testing, generating semantically stable and structurally consistent images.
AlpacaFarm: A Simulation Framework for Methods that Learn from Human Feedback
Yann Dubois (Stanford University), Tatsunori Hashimoto
Reinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: This paper studies a low-cost simulation framework called AlpacaFarm, designed for rapid experimentation and iteration in learning human feedback for instruction-following models.
Alternating Gradient Descent and Mixture-of-Experts for Integrated Multimodal Perception
Hassan Akbari (Google Research), Hartwig Adam (Google Research)
OptimizationTransformerMixture of ExpertsImageVideoTextMultimodalityAudio
🎯 What it does: This paper proposes an Integrated Multi-Modal Perception (IMP) framework that utilizes a single Transformer encoder to simultaneously process images, videos, text, and audio.
Alternating Updates for Efficient Transformers
Cenk Baykal (Google Research), Xin Wang (Google Research)
Computational EfficiencyTransformerMixture of ExpertsText
🎯 What it does: The Alternating Updates (AltUp) method is proposed, which expands the token representation dimension through a block-wise prediction-computation-correction mechanism while maintaining the same computational cost, thereby enhancing the expressive power of the Transformer.
Alternation makes the adversary weaker in two-player games
Volkan Cevher (École Polytechnique Fédérale de Lausanne), Luca Viano (École Polytechnique Fédérale de Lausanne)
OptimizationReinforcement LearningTabular
🎯 What it does: This study investigates the alternating gameplay in a two-player game and proposes a variant of Alternating Online Linear Optimization (Alternating OLO), analyzing the regret bounds of learners in this setting.
AmadeusGPT: a natural language interface for interactive animal behavioral analysis
Shaokai Ye (École Polytechnique Fédérale de Lausanne), Mackenzie W Mathis
Object DetectionPose EstimationAI Code AssistantTransformerLarge Language ModelPrompt EngineeringVideo
🎯 What it does: Developed AmadeusGPT, a natural language interface that can automatically convert natural language descriptions of animal behavior into executable Python code, enabling no-code behavior analysis;
AMAG: Additive, Multiplicative and Adaptive Graph Neural Network For Forecasting Neuron Activity
Jingyuan Li (University of Washington), Eli Shlizerman (University of Washington)
Graph Neural NetworkTransformerTime SeriesBiomedical Data
🎯 What it does: A graph neural network model called AMAG is proposed for predicting neural activity, which can accurately generate future neural signals given past recordings.
Ambient Diffusion: Learning Clean Distributions from Corrupted Data
Giannis Daras (University of Texas at Austin), Adam Klivans (University of Texas at Austin)
RestorationGenerationData SynthesisDiffusion modelImageBiomedical DataMagnetic Resonance ImagingStochastic Differential Equation
🎯 What it does: This paper proposes Ambient Diffusion, a diffusion model framework that can learn clean distributions even when only highly damaged samples are available.
AMDP: An Adaptive Detection Procedure for False Discovery Rate Control in High-Dimensional Mediation Analysis
Jiarong Ding (Xi'an Jiaotong University), Xuehu Zhu (Xi'an Jiaotong University)
Biomedical Data
🎯 What it does: An adaptive detection procedure for high-dimensional mediation analysis (AMDP) is proposed, which can efficiently identify significant mediating variables while controlling the FDR.
Amortized Reparametrization: Efficient and Scalable Variational Inference for Latent SDEs
Kevin Course (University of Toronto), Prasanth B. Nair (University of Toronto)
OptimizationComputational EfficiencyVideoTime SeriesStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: A solver-free inference method based on amortized reparameterization is proposed for modeling latent stochastic differential equations (Latent SDE) in high-dimensional time series.
An $\varepsilon$-Best-Arm Identification Algorithm for Fixed-Confidence and Beyond
Marc Jourdan (Inria), Emilie Kaufmann (University of Lille)
OptimizationTabular
🎯 What it does: A new ε-Best-Arm Identification algorithm EB-TC ε₀ is proposed, and its theoretical and experimental performance is provided under fixed confidence, fixed budget, and anytime settings.
An active learning framework for multi-group mean estimation
Abdellah Aznag (Columbia University), Adam N. Elmachtoub (Columbia University)
OptimizationReinforcement LearningTabular
🎯 What it does: An active learning framework for multi-group mean estimation is proposed, along with the optimal Variance-UCB algorithm.
An Adaptive Algorithm for Learning with Unknown Distribution Drift
Alessio Mazzetto (Brown University), Eli Upfal (Brown University)
Sequential
🎯 What it does: An adaptive algorithm is proposed and analyzed for addressing the learning problem of unknown distribution drift. This algorithm can learn a family of functions based on the most recent observation sequence without any prior knowledge.
An Alternating Optimization Method for Bilevel Problems under the Polyak-Łojasiewicz Condition
Quan Xiao (Rensselaer Polytechnic Institute), Tianyi Chen (Rensselaer Polytechnic Institute)
OptimizationImage
🎯 What it does: This paper studies a bilevel optimization problem where the lower-level objective satisfies the Polyak-Łojasiewicz (PL) condition, and proposes a General Alternating Method (GALET) that can solve ε-stationary points without additional constraints.
An Alternative to Variance: Gini Deviation for Risk-averse Policy Gradient
Yudong Luo (University of Waterloo), Yangchen Pan (University of Oxford)
Reinforcement Learning
🎯 What it does: This paper proposes the use of Gini Deviation as an alternative risk measure to variance, deriving and implementing the gradient estimation of Gini Deviation in mean-risk reinforcement learning based on policy gradients, and validating its feasibility on various discrete and continuous control tasks.
An Efficient and Robust Framework for Approximate Nearest Neighbor Search with Attribute Constraint
Mengzhao Wang (Hangzhou Dianzi University), Jiongkang Ni (Hangzhou Dianzi University)
RetrievalOptimizationGraph Neural NetworkImageTextMultimodalityAudio
🎯 What it does: A Natural Hybrid Query (NHQ) framework is proposed, capable of simultaneously handling Approximate Nearest Neighbor Search (ANNS) and attribute constraints within the same graph index, allowing for results that meet both feature similarity and attribute matching in a single query.
An Efficient Dataset Condensation Plugin and Its Application to Continual Learning
Enneng Yang (Northeastern University), Guibing Guo (Northeastern University)
ClassificationSafty and PrivacyKnowledge DistillationData-Centric LearningImage
🎯 What it does: A low-rank data distillation plugin is proposed, which compresses large-scale image data onto a low-dimensional manifold and integrates it with existing methods such as DC/DM/DSA to achieve efficient distillation of datasets, and this plugin is applied to continual learning tasks.
An Efficient Doubly-Robust Test for the Kernel Treatment Effect
Diego Martinez-Taboada (Carnegie Mellon University), Edward Kennedy
Computational EfficiencyReinforcement LearningTabularElectronic Health Records
🎯 What it does: A computationally efficient, non-permutation dual robust kernel distribution effect testing method AIPW-xKTE is proposed for causal inference with observational data.
An Efficient End-to-End Training Approach for Zero-Shot Human-AI Coordination
Xue Yan (Institute of Automation, Chinese Academy of Sciences), Yali Du (King's College London)
Robotic IntelligenceReinforcement Learning
🎯 What it does: An end-to-end training framework E3T is proposed, utilizing a hybrid partner strategy (ego strategy + random strategy) and a partner modeling module to achieve zero-shot human-machine collaboration.
An Empirical Study Towards Prompt-Tuning for Graph Contrastive Pre-Training in Recommendations
Haoran Yang (University of Technology Sydney), Guandong Xu (University of Technology Sydney)
Recommendation SystemGraph Neural NetworkPrompt EngineeringContrastive LearningGraph
🎯 What it does: This paper proposes the CPTPP framework, which combines graph contrastive learning pre-training with prompt tuning. In recommendation tasks, user/item embeddings are first pre-trained using GCL, then personalized soft prompts are automatically generated based on graph interactions, and finally, downstream recommendations are completed through prompt fusion.
An Exploration-by-Optimization Approach to Best of Both Worlds in Linear Bandits
Shinji Ito (NEC Corporation), Kei Takemura (NEC Corporation)
Optimization
🎯 What it does: This paper proposes a linear Bandit algorithm based on the Exploration-Optimization (EXO) framework, which can achieve near-optimal cumulative loss upper bounds in both adversarial and stochastic environments.
An Improved Relaxation for Oracle-Efficient Adversarial Contextual Bandits
Kiarash Banihashem (University of Maryland), Max Springer (University of Maryland)
OptimizationReinforcement Learning from Human Feedback
🎯 What it does: An oracle-efficient algorithm is proposed to solve the contextual gambling problem with i.i.d. contexts and adversarial costs, providing an upper bound on expected loss.
An Inductive Bias for Tabular Deep Learning
Ege Beyazit (Amazon), Bilal H Fadlallah
ClassificationOptimizationTabular
🎯 What it does: This study investigates the spectral bias of neural networks on tabular data, proposing frequency reduction as a prior. It achieves feature scaling and ranking through learnable layers to reduce the irregularity of the objective function, validated on 14 benchmark datasets.
An Information Theory Perspective on Variance-Invariance-Covariance Regularization
Ravid Shwartz-Ziv (New York University), Yann LeCun (Meta AI)
ClassificationRepresentation LearningConvolutional Neural NetworkTransformerContrastive LearningImage
🎯 What it does: This paper provides an in-depth analysis of Variance-Invariance-Covariance Regularization (VICReg) from an information theory perspective, deriving its relationship with mutual information maximization, presenting generalization bounds for unsupervised pre-training, and proposing an improved SSL method along with a more accurate entropy estimator based on this analysis.
An Information-Theoretic Evaluation of Generative Models in Learning Multi-modal Distributions
Mohammad Jalali (Isfahan University of Technology), Farzan Farnia (Chinese University of Hong Kong)
GenerationData SynthesisGenerative Adversarial NetworkImageMultimodality
🎯 What it does: A matrix entropy evaluation method based on Rènyi kernel entropy is proposed to measure the diversity of generative models under multimodal distributions.
An information-theoretic quantification of the content of communication between brain regions
Marco Celotto (University Medical Center Hamburg Eppendorf), Stefano Panzeri (University Medical Center Hamburg Eppendorf)
Time SeriesSequential
🎯 What it does: A measure of information theory called Feature-specific Information Transfer (FIT) is proposed to quantify the directed information flow about specific features between brain regions.
An Inverse Scaling Law for CLIP Training
Xianhang Li (University of California Santa Cruz), Cihang Xie (University of California Santa Cruz)
ClassificationRetrievalComputational EfficiencyTransformerVision Language ModelContrastive LearningImageText
🎯 What it does: The study investigates and verifies the inverse scaling law of CLIP training, proposing that fewer image/text tokens can be used for efficient training under large models, and based on this, the CLIPA training framework is designed.
An Iterative Self-Learning Framework for Medical Domain Generalization
Zhenbang Wu (University of Illinois Urbana Champaign), Jimeng Sun (University of Illinois Urbana Champaign)
Domain AdaptationTransformerTabularBiomedical DataElectronic Health Records
🎯 What it does: A self-learning framework SLDG is proposed to automatically discover potential subdomains in medical EHR data and train dedicated classifiers for each subdomain to address the domain generalization problem.
An Optimal and Scalable Matrix Mechanism for Noisy Marginals under Convex Loss Functions
Yingtai Xiao (Penn State University), Daniel Kifer (Penn State University)
OptimizationSafty and PrivacyTabularFinance Related
🎯 What it does: A matrix mechanism named ResidualPlanner is proposed, which can provide unbiased, optimal, and scalable privacy-preserving answers for multidimensional marginal queries under Gaussian random noise.
An Optimal Structured Zeroth-order Algorithm for Non-smooth Optimization
Marco Rando (University of Genova), Silvia Villa (University of Genova)
Optimization
🎯 What it does: A structured zero-order (gradient-free) optimization algorithm O-ZD is proposed and analyzed, suitable for non-smooth black-box optimization problems.
An Optimization-based Approach To Node Role Discovery in Networks: Approximating Equitable Partitions
Michael Scholkemper (RWTH Aachen University), Michael T Schaub
OptimizationGraph Neural NetworkGraphBiomedical Data
🎯 What it does: This paper proposes a node role discovery method based on Equitable Partition (EP) and an optimization framework, and provides the corresponding cost function and algorithm.
Analysis of Variance of Multiple Causal Networks
Zhongli Jiang (Purdue University), Dabao Zhang (University of California)
Biomedical Data
🎯 What it does: This paper studies an algorithm called NetANOVA for the parallel construction and comparison of multiple Directed Cyclic Graphs (DCGs) to identify common and differing causal relationships in multiple datasets.
Analyzing Generalization of Neural Networks through Loss Path Kernels
Yilan Chen (University of California San Diego), Tsui-Wei Weng (University of California San Diego)
OptimizationNeural Architecture SearchImage
🎯 What it does: This study investigates the generalization ability of neural networks trained via gradient flow and equivalently constructs a loss path kernel to relate it to universal kernel machines, thereby providing a general and tight generalization upper bound.
Analyzing the Sample Complexity of Self-Supervised Image Reconstruction Methods
Tobit Klug (Technical University of Munich), Reinhard Heckel (Technical University of Munich)
RestorationSuper ResolutionConvolutional Neural NetworkImageMagnetic Resonance Imaging
🎯 What it does: Theoretical analysis and large-scale experimental validation of the sample complexity of self-supervised methods based on unbiased gradient estimation in deep learning image reconstruction.
Analyzing Vision Transformers for Image Classification in Class Embedding Space
Martina G. Vilas (Goethe University Frankfurt), Gemma Roig (Goethe University Frankfurt)
ClassificationRepresentation LearningTransformerImage
🎯 What it does: This study investigates how to project the internal representations of Vision Transformers into class embedding space, thereby reverse engineering image classification networks.
Anchor Data Augmentation
Nora Schneider (ETH Zurich), Fernando Perez-Cruz (Swiss Data Science Center)
OptimizationData-Centric LearningTabularTime Series
🎯 What it does: A new regression data augmentation method called Anchor Data Augmentation (ADA) is proposed, which generates diverse training samples through linear interpolation/extrapolation of samples in the anchor point space obtained from clustering at different γ values, thereby enhancing the generalization and robustness of regression models.
AND: Adversarial Neural Degradation for Learning Blind Image Super-Resolution
Fangzhou Luo (McMaster University), Yanhui Guo (McMaster University)
RestorationSuper ResolutionConvolutional Neural NetworkGenerative Adversarial NetworkImage
🎯 What it does: The paper proposes an Adversarial Neural Degradation (AND) model that generates diverse nonlinear degradations through adversarial training in an unsupervised environment, enhancing the robustness of blind super-resolution networks to unknown degradations.
Annotator: A Generic Active Learning Baseline for LiDAR Semantic Segmentation
Binhui Xie (Beijing Institute of Technology), Xinjing Cheng (Tsinghua University)
SegmentationDomain AdaptationPoint Cloud
🎯 What it does: This paper presents Annotator, a general active learning baseline for LiDAR semantic segmentation that efficiently selects voxels for labeling and improves model performance under three labeling efficiency scenarios (AL, ASFDA, ADA).
Anonymous and Copy-Robust Delegations for Liquid Democracy
Markus Utke (Eindhoven University of Technology), Ulrike Schmidt-Kraepelin (Simons Laufer Mathematical Sciences Institute)
🎯 What it does: A liquid democratic delegation rule with assignable weights is proposed, along with a polynomial time computation method, proving that it satisfies properties such as anonymity, replication robustness, and consistency.
Anonymous Learning via Look-Alike Clustering: A Precise Analysis of Model Generalization
Adel Javanmard (University of Southern California), Vahab Mirrokni (Google Research)
OptimizationSafty and PrivacyTabular
🎯 What it does: This study investigates the generalization error of the minimum norm linear regression model after anonymization processing (look-alike clustering) and provides high-dimensional asymptotic analysis.
ANPL: Towards Natural Programming with Interactive Decomposition
Di Huang (Institute of Computing Technology, Chinese Academy of Sciences), Yunji Chen (Institute of Computing Technology, Chinese Academy of Sciences)
AI Code AssistantTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: The ANPL system is proposed, allowing users to control/data flow by drawing a 'sketch' and describing 'hole' modules in natural language for the LLM to implement automatically; it supports interactive decomposition, debugging, and recursive subprograms;
ANTN: Bridging Autoregressive Neural Networks and Tensor Networks for Quantum Many-Body Simulation
Zhuo Chen (Massachusetts Institute of Technology), Marin Soljacic (Massachusetts Institute of Technology)
TransformerSequentialPhysics Related
🎯 What it does: A novel quantum wave function representation method called Autoregressive Neural TensorNet (ANTN) is proposed and implemented, which combines tensor networks with autoregressive neural networks for efficient simulation of quantum many-body systems.
Any-to-Any Generation via Composable Diffusion
Zineng Tang (University of North Carolina), Mohit Bansal (University of North Carolina)
GenerationData SynthesisDiffusion modelContrastive LearningImageVideoTextMultimodalityAudio
🎯 What it does: A diffusion model called CoDi is trained and proposed, which can handle arbitrary combinations of input and output modalities, supporting the generation of any combination of text, images, videos, and audio.
Anytime Model Selection in Linear Bandits
Parnian Kassraie (ETH Zurich), Aldo Pacchiano (Boston University)
OptimizationTabular
🎯 What it does: An online model selection and linear bandit optimization algorithm named ALEXP is proposed, which can adaptively explore and exploit without prior knowledge of the environment.
Anytime-Competitive Reinforcement Learning with Policy Prior
Jianyi Yang (University of California Riverside), Shaolei Ren (University of California Riverside)
Reinforcement LearningTime Series
🎯 What it does: This paper proposes the Anytime-Competitive Markov Decision Process (A-CMDP) framework and designs a two-stage algorithm ACD (Safe Action Set Projection) and ACRL (Model-based RL) to achieve optimal learning under cost constraints and rewards at any time.
Approximate Allocation Matching for Structural Causal Bandits with Unobserved Confounders
Lai Wei (Life Sciences Institute University of Michigan), Murat Kocaoglu (Purdue University)
Reinforcement LearningGraph
🎯 What it does: This paper proposes a more efficient decision-making strategy for the structural causal bandit problem with unobserved confounding variables by utilizing a known causal graph.
Approximate Heavy Tails in Offline (Multi-Pass) Stochastic Gradient Descent
Krunoslav Lehman Pavasovic (Inria Paris), Umut Simsekli (Inria Paris)
OptimizationImageTabular
🎯 What it does: This paper studies the emergence of heavy-tailed phenomena in offline stochastic gradient descent (SGD), particularly how to understand the mechanism of such heavy-tailed behavior when training data is limited.
Approximate inference of marginals using the IBIA framework
Shivani Bathla (Indian Institute of Technology Madras), Vinita Vasudevan (Indian Institute of Technology Madras)
🎯 What it does: An approximate marginal inference algorithm based on the IBIA framework is proposed, which first transforms the PGM into a serialized linkage potential tree forest (SLCTF), and then performs backward propagation through heuristic belief-update between adjacent potential trees to obtain the posterior marginal probabilities of the variables.
Approximately Equivariant Graph Networks
Ningyuan Teresa Huang, Soledad Villar (Johns Hopkins University)
Pose EstimationGraph Neural NetworkImageGraph
🎯 What it does: This study investigates active symmetry on fixed graphs, proposing a covariant graph network obtained through graph coarsening to approximate symmetry groups, and provides a risk-bias-variance balance formula.
Approximation-Generalization Trade-offs under (Approximate) Group Equivariance
Mircea Petrache (Universidad de Chile), Shubhendu Trivedi (Massachusetts Institute of Technology)
🎯 What it does: A quantitative theoretical analysis of the generalization error and approximation error of equivariant models under constraints or approximate symmetry is proposed.
AR-Diffusion: Auto-Regressive Diffusion Model for Text Generation
Tong Wu (Tsinghua University), Weizhu Chen (Microsoft Azure AI)
GenerationTransformerDiffusion modelText
🎯 What it does: The AR-Diffusion model is proposed, which combines autoregressive language generation with diffusion models, constructing a dual diffusion strategy at the sentence and word levels, and introducing a jump step mechanism for parallel decoding.
Arbitrarily Scalable Environment Generators via Neural Cellular Automata
Yulun Zhang (Carnegie Mellon University), Jiaoyang Li (Carnegie Mellon University)
OptimizationRobotic IntelligenceReinforcement Learning
🎯 What it does: This paper combines Quality Diversity (QD) algorithms with Neural Cellular Automata (NCA) to train multi-robot system environments that can generate regular layouts in environments of any scale, significantly improving system throughput.
Architecture Matters: Uncovering Implicit Mechanisms in Graph Contrastive Learning
Xiaojun Guo (Peking University), Yisen Wang (Peking University)
Representation LearningGraph Neural NetworkContrastive LearningGraph
🎯 What it does: This paper systematically evaluates and reveals the different characteristics of Graph Contrastive Learning (GCL) and Visual Contrastive Learning (VCL).
Are aligned neural networks adversarially aligned?
Nicholas Carlini (Google DeepMind), Ludwig Schmidt (University of Washington)
Adversarial AttackTransformerLarge Language ModelVision Language ModelTextMultimodality
🎯 What it does: This paper studies the robustness of aligned language models when faced with adversarial inputs, proposing the concept of 'adversarial alignment' and systematically evaluating the behavior of text models and multimodal models under adversarial input attacks.
Are Diffusion Models Vision-And-Language Reasoners?
Benno Krojer (Mila and McGill University), Siva Reddy (Mila and McGill University)
GenerationRetrievalDiffusion modelImageTextMultimodalityBenchmark
🎯 What it does: This paper proposes a framework called DiffusionITM that transforms diffusion models (such as Stable Diffusion) into image-text matching (ITM) models, and constructs the GDBench benchmark to evaluate the reasoning capabilities of diffusion models across various visual-language tasks. Additionally, it enhances the model's discriminative performance by fine-tuning with hard negative samples on MS-COCO.
Are Emergent Abilities of Large Language Models a Mirage?
Rylan Schaeffer (Stanford University), Sanmi Koyejo (Stanford University)
TransformerLarge Language ModelText
🎯 What it does: This paper questions the phenomenon of 'emergent abilities' in large language models, suggesting that these performances are artifacts caused by measurement methods, and that the emergence can be eliminated through different metrics or larger samples.
Are GATs Out of Balance?
Nimrah Mustafa (CISPA Helmholtz Center for Information Security), Rebekka Burkholz (CISPA Helmholtz Center for Information Security)
ClassificationOptimizationGraph Neural NetworkGraph
🎯 What it does: This study investigates the gradient flow dynamics of Graph Attention Networks (GAT), deriving a gradient conservation law that reveals the issue of untrainability in deep GATs caused by standard initialization, and proposes a balanced initialization scheme.
Are Vision Transformers More Data Hungry Than Newborn Visual Systems?
Lalit Pandey (Indiana University), Justin Newell Wood
RecognitionRepresentation LearningTransformerContrastive LearningImageVideo
🎯 What it does: This study explores whether Vision Transformers (ViTs) require more data than biological visual systems by comparing the learning capabilities of ViTs and newly hatched chicks. Through controlled rearing experiments in a virtual environment, it was found that ViTs can learn object recognition abilities similar to those of newly hatched chicks in impoverished visual environments.
ARTIC3D: Learning Robust Articulated 3D Shapes from Noisy Web Image Collections
Chun-Han Yao (University of California Merced), Varun Jampani (Google Research)
GenerationPose EstimationOptimizationDiffusion modelImage
🎯 What it does: Using sparse and noisy network images, the three-dimensional jointed shape and texture of animals are restored through a self-supervised approach, supporting animation.
ARTree: A Deep Autoregressive Model for Phylogenetic Inference
Tianyu Xie (Peking University), Cheng Zhang (Peking University)
Graph Neural NetworkBiomedical Data
🎯 What it does: A self-regressive tree topology generation model ARTree based on graph neural networks is proposed for learning the probability distribution of evolutionary trees and variational Bayesian inference.
ASIF: Coupled Data Turns Unimodal Models to Multimodal without Training
Antonio Norelli (Sapienza Università di Roma), Francesco Locatello (Institute of Science and Technology Austria)
ClassificationRetrievalRepresentation LearningTransformerVision Language ModelContrastive LearningImageTextMultimodality
🎯 What it does: This paper proposes the ASIF method, which aligns pre-trained unimodal image and text encoders with a small number of image-text pairs into a multimodal model without requiring any additional training.
ASPEN: Breaking Operator Barriers for Efficient Parallelization of Deep Neural Networks
Jongseok Park (Seoul National University), Kyunghan Lee (Seoul National University)
OptimizationComputational EfficiencyConvolutional Neural NetworkTransformer
🎯 What it does: ASPEN is proposed, a DNN inference framework based on fine-grained tile-level dynamic scheduling, eliminating traditional operator synchronization barriers to achieve opportunistic parallelism.
Assessor360: Multi-sequence Network for Blind Omnidirectional Image Quality Assessment
Tianhe Wu (Shenzhen International Graduate School Tsinghua University), Yujiu Yang (Shenzhen International Graduate School Tsinghua University)
TransformerImage
🎯 What it does: This paper proposes a multi-sequence network called Assessor360 for blind quality assessment of panoramic images under no-reference conditions.
Assumption violations in causal discovery and the robustness of score matching
Francesco Montagna (University of Genoa), Francesco Locatello (Institute of Science and Technology Austria)
Score-based ModelTabular
🎯 What it does: A systematic evaluation of eleven causal discovery methods under observational IID data was conducted, with a particular focus on scenarios of assumption violations;
Asymmetric Certified Robustness via Feature-Convex Neural Networks
Samuel Pfrommer (University of California), Somayeh Sojoudi (University of California)
ClassificationAdversarial AttackImage
🎯 What it does: A provably robust classification method against asymmetric attacks (which only concern misclassifying sensitive classes as non-sensitive classes) is proposed, defining a feature-convex neural network structure that can provide a closed-form, computable safety radius.
Asymptotically Optimal Quantile Pure Exploration for Infinite-Armed Bandits
Xiao-Yue Gong (Carnegie Mellon University), Mark Sellke (Harvard University)
OptimizationReinforcement Learning from Human Feedback
🎯 What it does: This paper studies the pure exploration problem under the infinite-armed Bandit model, proposing near-optimal sample complexity algorithms under two settings: fixed confidence and fixed budget, and provides corresponding lower bounds, proving that there is an essential difference in sample complexity between the two settings.
Asymptotics of Bayesian Uncertainty Estimation in Random Features Regression
Youngsoo Baek (Duke University), Sayan Mukherjee (Max Planck Institute for Mathematics in the Sciences)
Tabular
🎯 What it does: This paper compares the asymptotic behavior of the Bayesian posterior predictive variance (PPV) and the risk of the MAP estimator under over-parameterization conditions in random feature regression models.
Asynchronous Proportional Response Dynamics: Convergence in Markets with Adversarial Scheduling
Yoav Kolumbus (Cornell University), Noam Nisan (Hebrew University of Jerusalem)
OptimizationTabular
🎯 What it does: This paper studies the proportion response dynamics (PRD) of participants updating their bids asynchronously in a linear Fisher market, and proves that PRD still converges to market equilibrium under updates from any activated subset of opponents; it further proves the uniqueness of equilibrium bids under generic parameters; and maps PRD to a related game with a potential function, using this potential function to demonstrate asynchronous convergence; it also explores the convergence of optimal response and no-swap regret dynamics; finally, it provides simulation validation for random markets.
Asynchrony-Robust Collaborative Perception via Bird's Eye View Flow
Sizhe Wei (Shanghai Jiao Tong University), Ya Zhang (Shanghai Jiao Tong University)
Object DetectionAutonomous DrivingOptical FlowImagePoint Cloud
🎯 What it does: This paper proposes an asynchronous collaborative perception framework called CoBEVFlow based on Bird's Eye View (BEV) flow, which can dynamically compensate for asynchronous information caused by communication delays and clock discrepancies in multi-agent systems such as vehicles and robots, thereby improving perception accuracy.
ATMAN: Understanding Transformer Predictions Through Memory Efficient Attention Manipulation
Björn Deiseroth (Aleph Alpha), Kristian Kersting (Technical University Darmstadt)
Explainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelVision Language ModelTextMultimodality
🎯 What it does: This paper proposes ATMAN, a memory-efficient explanation method for generative Transformers, which generates a relevance map of input to output by utilizing scalar perturbations of attention scores.
ATTA: Anomaly-aware Test-Time Adaptation for Out-of-Distribution Detection in Segmentation
Zhitong Gao (ShanghaiTech University), Xuming He (ShanghaiTech University)
SegmentationDomain AdaptationAnomaly DetectionImage
🎯 What it does: This paper proposes a dual-layer adaptive framework ATTA for dense OOD detection in semantic segmentation under domain shift conditions.
Attacks on Online Learners: a Teacher-Student Analysis
Riccardo Giuseppe Margiotta, Guido Sanguinetti (International School for Advanced Studies)
OptimizationAdversarial AttackReinforcement LearningImage
🎯 What it does: This study investigates the impact of online data poisoning attacks on machine learning models, analyzing the dynamics of the attacks from both theoretical and experimental perspectives using a teacher-student framework, and proposes a greedy attack strategy.
Attention as Implicit Structural Inference
Ryan Singh (University of Sussex), Christopher Buckley
TransformerSequential
🎯 What it does: This paper proposes a probabilistic framework that treats attention in Transformers as structural inference, and based on this, designs two novel attention mechanisms (multi-hop attention and scalable attention), validating their effectiveness through theoretical analysis and simulation experiments.
Attentive Transfer Entropy to Exploit Transient Emergence of Coupling Effect
Xiaolei Ru (Tongji University), Gang Yan (Tongji University)
ClassificationAnomaly DetectionConvolutional Neural NetworkTime Series
🎯 What it does: This paper proposes a trainable transfer entropy method based on attention mechanisms, utilizing the instantaneous coupling effects in time series to enhance the reconstruction of coupling networks;
AttrSeg: Open-Vocabulary Semantic Segmentation via Attribute Decomposition-Aggregation
Chaofan Ma (Shanghai Jiao Tong University), Yanfeng Wang (Shanghai Jiao Tong University)
SegmentationTransformerLarge Language ModelImage
🎯 What it does: This paper proposes an attribute decomposition-aggregation framework (AttrSeg) that achieves open vocabulary semantic segmentation by decomposing category names into multiple attribute descriptions and then aggregating them into a single global description.
AUDIT: Audio Editing by Following Instructions with Latent Diffusion Models
Yuancheng Wang (Microsoft), sheng zhao
GenerationData SynthesisDiffusion modelAudio
🎯 What it does: This paper presents the AUDIT model, which enables audio editing based on natural language instructions.
Auditing Fairness by Betting
Ben Chugg (Carnegie Mellon University), Aaditya Ramdas (Carnegie Mellon University)
TabularFinance Related
🎯 What it does: A non-parametric sequential hypothesis testing method based on game theory has been developed for continuously monitoring the group fairness of deployed classification and regression models in real-time auditing.
Auditing for Human Expertise
Rohan Alur (Massachusetts Institute of Technology), Dennis Shung (Yale University)
Biomedical DataElectronic Health Records
🎯 What it does: A statistical testing method based on conditional independence (ExpertTest) has been developed to detect whether human experts incorporate valuable information beyond observable features;
Augmentation-Aware Self-Supervision for Data-Efficient GAN Training
Liang Hou (Chinese Academy of Sciences), Xueqi Cheng (Chinese Academy of Sciences)
GenerationData SynthesisGenerative Adversarial NetworkImage
🎯 What it does: This paper proposes an augmented perception self-supervised discriminator that can predict augmentation parameters while simultaneously distinguishing between real and fake data, thereby improving the training efficiency of GANs in scenarios with limited data.
Augmentation-Free Dense Contrastive Knowledge Distillation for Efficient Semantic Segmentation
Jiawei Fan (Intel Labs China), Anbang Yao (Intel Labs China)
SegmentationKnowledge DistillationContrastive LearningImage
🎯 What it does: A dense contrastive knowledge distillation method Af-DCD has been developed for efficient semantic segmentation network distillation, without data augmentation and without memory buffering.