NeurIPS 2023 Papers — Page 16
Conference on Neural Information Processing Systems · 3218 papers
Intervention Generalization: A View from Factor Graph Models
Gecia Bravo-Hermsdorff (University College London), Ricardo Silva (University College London)
Graph Neural NetworkGraph
🎯 What it does: This paper proposes an Interventional Factor Model (IFM) based on factor graphs, providing identifiable conditions through a minimally assumed factorization structure, and implementing inference of expected outcomes under unobserved interventions using message passing and algebraic methods.
Intra-Modal Proxy Learning for Zero-Shot Visual Categorization with CLIP
Qi Qian (Alibaba Group), Juhua Hu (University of Washington)
ClassificationRepresentation LearningTransformerContrastive LearningImage
🎯 What it does: By directly learning category proxies in the visual space, improvements in zero-shot visual classification are achieved.
Intriguing Properties of Quantization at Scale
Arash Ahmadian (Cohere For AI), Sara Hooker (Cohere For AI)
TransformerLarge Language ModelText
🎯 What it does: Researching the sensitivity of quantization in large-scale language models, systematically evaluating the impact of different training hyperparameters on post-training quantization to INT8.
Intrinsic Dimension Estimation for Robust Detection of AI-Generated Texts
Eduard Tulchinskii (AI Foundation and Algorithm Lab), Irina Piontkovskaya (Artificial Intelligence Research Institute)
ClassificationAnomaly DetectionTransformerLarge Language ModelText
🎯 What it does: This study investigates the intrinsic dimension of text samples as a feature to distinguish between human-written text and AI-generated text, and proposes a detector based on Persistent Homology Dimension (PHD).
Invariant Anomaly Detection under Distribution Shifts: A Causal Perspective
João B. S. Carvalho (ETH Zürich), Joachim M. Buhmann (ETH Zürich)
Domain AdaptationAnomaly DetectionImageBiomedical DataMagnetic Resonance Imaging
🎯 What it does: Under distribution shift, a Partial Condition Invariant Regularization (PCIR) is proposed and implemented from a causal perspective to enhance the robustness of anomaly detection models against domain and covariate shifts.
Invariant Learning via Probability of Sufficient and Necessary Causes
Mengyue Yang (University College London), Jun Wang (University College London)
Domain AdaptationRepresentation LearningImage
🎯 What it does: This paper proposes a risk function based on the Probabilistic Necessary and Sufficient (PNS) causality to learn representations that incorporate both necessary and sufficient causal information during source domain training, thereby achieving robust generalization to unseen domains.
Inverse Dynamics Pretraining Learns Good Representations for Multitask Imitation
David Brandfonbrener (New York University), Joan Bruna (New York University)
Representation LearningRobotic IntelligenceContrastive LearningImage
🎯 What it does: This paper studies the use of Inverse Dynamics Pretraining in multi-task imitation learning to learn low-dimensional representations of visual inputs and fine-tune them for new tasks.
Inverse Preference Learning: Preference-based RL without a Reward Function
Joey Hejna (Stanford University), Dorsa Sadigh (Stanford University)
Reinforcement LearningSequential
🎯 What it does: This paper proposes Inverse Preference Learning (IPL), a method that directly uses the Q-function to learn implicit rewards in offline preference reinforcement learning, eliminating the need for an explicit reward network.
Inverse Reinforcement Learning with the Average Reward Criterion
Feiyang Wu (Georgia Institute of Technology), Anqi Wu (Georgia Institute of Technology)
Reinforcement LearningSequential
🎯 What it does: This paper proposes an inverse reinforcement learning (IRL) method under the average reward (AMDP) framework, combining Stochastic Policy Mirror Descent (SPMD) to solve the average reward MDP and Inverse Policy Mirror Descent (IPMD) to solve the maximum entropy IRL dual problem.
Investigating how ReLU-networks encode symmetries
Georg Bökman (Chalmers University of Technology), Fredrik Kahl (Chalmers University of Technology)
ClassificationRecognitionOptimizationConvolutional Neural NetworkImage
🎯 What it does: The study investigates whether the layers of ReLU networks maintain equivariance after training for group equivariance, and proposes the GCNN barrier metric to measure the network's proximity to GCNN.
IPMix: Label-Preserving Data Augmentation Method for Training Robust Classifiers
Zhenglin Huang (Zhejiang Sci-Tech University), Xi Yang (University of Science and Technology of China)
ClassificationData SynthesisAnomaly DetectionConvolutional Neural NetworkSupervised Fine-TuningImage
🎯 What it does: A label-preserving data augmentation method named IPMix is proposed, which integrates image-level, patch-level, and pixel-level augmentations, and incorporates synthetic images such as fractals to enhance structural diversity, thereby training a more robust classifier.
Is Distance Matrix Enough for Geometric Deep Learning?
Zian Li (Peking University), Muhan Zhang (Peking University)
Graph Neural NetworkGraph
🎯 What it does: This paper studies the use of distance matrices for deep learning on geometric graphs and proposes the k-DisGNN model based on k-WL to address the expressiveness incompleteness of the traditional Vanilla DisGNN, further demonstrating its learnability of high-order geometric information and unifying DimeNet and GemNet.
Is Heterogeneity Notorious? Taming Heterogeneity to Handle Test-Time Shift in Federated Learning
Yue Tan (University of Technology Sydney), Guodong Long (University of Technology Sydney)
Domain AdaptationFederated LearningContrastive LearningImage
🎯 What it does: This paper proposes a federated learning framework named FedICON, which utilizes contrastive learning to extract and share invariant features from heterogeneous clients, and subsequently implements unsupervised contrastive learning locally for adaptation during testing.
Is Learning in Games Good for the Learners?
William Brown (Columbia University), Kiran Vodrahalli (Google Research)
Reinforcement Learning
🎯 What it does: This paper studies the trade-off between rewards and regrets in learning algorithms within two-player repeated games, proposing an asymmetric Φ-regret generalized equilibrium. It analyzes the maximum achievable rewards when facing learners with no exchange regret and no external regret, and explores how to query simulate learning Stackelberg strategies in unknown games.
Is RLHF More Difficult than Standard RL? A Theoretical Perspective
Yuanhao Wang (Princeton University), Chi Jin (Princeton University)
Reinforcement Learning from Human FeedbackReinforcement Learning
🎯 What it does: This paper theoretically proves that RLHF (Reinforcement Learning from Human Feedback) is not more difficult than traditional RL (Reward-based Learning). It proposes an interface P2R for converting preferences into rewards and two methods for transforming general preferences into two-player zero-sum Markov games to find Nash equilibria.
Is This Loss Informative? Faster Text-to-Image Customization by Tracking Objective Dynamics
Anton Voronov (Moscow Institute of Physics and Technology), Max Ryabinin (Yandex)
GenerationOptimizationComputational EfficiencyDiffusion modelImage
🎯 What it does: This paper studies the rapid personalization training process of text-to-image models on small datasets and proposes an early stopping discriminator based on training loss variance (DVAR), which can significantly accelerate model adaptation.
Is Your Code Generated by ChatGPT Really Correct? Rigorous Evaluation of Large Language Models for Code Generation
Jiawei Liu, LINGMING ZHANG
AI Code AssistantTransformerLarge Language ModelText
🎯 What it does: The EvalPlus framework is proposed, which automatically generates a large number of test cases through LLM and mutation techniques, expanding and evaluating the functional correctness of LLM code generation; based on this, HUMANEVAL is upgraded to generate HUMANEVAL+ and its compressed version HUMANEVAL+-MINI.
iSCAN: Identifying Causal Mechanism Shifts among Nonlinear Additive Noise Models
Tianyu Chen (University of Texas at Austin), Pradeep Kumar Ravikumar
Score-based ModelGraphTabularBiomedical Data
🎯 What it does: This study proposes an unsupervised method called iSCAN, which utilizes the Jacobian of the score matrix of mixed distributions to identify changes in causal mechanisms between different environments in nonlinear additive noise models (ANM). It can locate varying variables and further estimate structural changes.
Isometric Quotient Variational Auto-Encoders for Structure-Preserving Representation Learning
In Huh (Samsung Electronics), Dae Sin Kim (UNIST)
Representation LearningAuto EncoderImage
🎯 What it does: An Isometric Quotient Variational Auto-Encoders (IQVAE) is proposed, which learns the quotient space of the data manifold under unsupervised conditions while maintaining Riemannian isometry, thereby obtaining a low-dimensional representation that preserves structure.
ISP: Multi-Layered Garment Draping with Implicit Sewing Patterns
Ren Li (École Polytechnique Fédérale de Lausanne), Pascal Fua (École Polytechnique Fédérale de Lausanne)
GenerationData SynthesisOptimizationConvolutional Neural NetworkAuto EncoderMesh
🎯 What it does: Proposes a multi-layer garment draping method based on implicit sewing patterns;
Iterative Reachability Estimation for Safe Reinforcement Learning
Milan Ganai (University of California San Diego), Sicun Gao (University of California San Diego)
OptimizationSafty and PrivacyReinforcement Learning
🎯 What it does: This paper proposes a safety reinforcement learning framework called RESPO based on reachability estimation, which optimizes rewards within the feasible set while maintaining continuous safety, minimizes cumulative violations within the infeasible set, and attempts to return to the feasible set; it also provides a corresponding Actor-Critic algorithm and proves its convergence.
Iteratively Learn Diverse Strategies with State Distance Information
Wei Fu (Tsinghua University), Yi Wu (Tsinghua University)
Robotic IntelligenceReinforcement LearningSequential
🎯 What it does: This paper studies a diversification strategy learning method based on state space distance and proposes an iterative learning framework called SIPO.
Jaccard Metric Losses: Optimizing the Jaccard Index with Soft Labels
Zifu Wang (KU Leuven), Matthew B. Blaschko (KU Leuven)
SegmentationKnowledge DistillationConvolutional Neural NetworkTransformerImage
🎯 What it does: Designed and implemented a Jaccard Metric Loss (JML) compatible with soft labels, combining it with label smoothing, knowledge distillation, and semi-supervised learning to enhance the accuracy and calibration performance of semantic segmentation models.
Jailbroken: How Does LLM Safety Training Fail?
Alexander Wei (University of California Berkeley), Jacob Steinhardt (University of California Berkeley)
Safty and PrivacyAdversarial AttackTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: This paper studies the reasons why large language models (LLMs) remain vulnerable to 'jailbreak' attacks even after secure training. It proposes two failure modes (competitive objectives and generalization mismatch) and designs and evaluates various black-box attack methods based on these modes, demonstrating that models are still highly susceptible to attacks even after the latest secure training and red team evaluations.
Jigsaw: Learning to Assemble Multiple Fractured Objects
Jiaxin Lu (University of Texas at Austin), Qixing Huang (University of Texas at Austin)
Object DetectionSegmentationPose EstimationGraph Neural NetworkPoint Cloud
🎯 What it does: An end-to-end Jigsaw framework is proposed for the automatic reassembly of 3D shattered fragments of multiple objects, restoring their global pose.
Joint Attribute and Model Generalization Learning for Privacy-Preserving Action Recognition
Duo Peng (Singapore University of Technology and Design), Jun Liu (Singapore University of Technology and Design)
RecognitionSafty and PrivacyMeta LearningVideo
🎯 What it does: This paper proposes a unified Meta-learning framework (MPPAR) for training video anonymization models to achieve dual objectives of privacy protection and action recognition under unknown privacy attributes and attack models.
Joint Bayesian Inference of Graphical Structure and Parameters with a Single Generative Flow Network
Tristan Deleu (Mila Quebec AI Institute), Yoshua Bengio (Mila Quebec AI Institute)
Graph Neural NetworkFlow-based ModelBiomedical Data
🎯 What it does: A single Generative Flow Network (JSP-GFN) is constructed to approximate the joint posterior distribution of the structure and conditional probability distribution parameters of Bayesian networks.
Joint Data-Task Generation for Auxiliary Learning
Hong Chen (Tsinghua University), Wenwu Zhu (Tsinghua University)
Recommendation SystemOptimizationImage
🎯 What it does: A framework for jointly generating auxiliary data and tasks (DTG‑AuxL) is proposed, allowing the main task to benefit even when the original auxiliary data/tasks are unfavorable;
Joint Feature and Differentiable $ k $-NN Graph Learning using Dirichlet Energy
Lei Xu (Northwestern Polytechnical University), Xuelong Li (Northwestern Polytechnical University)
ClassificationOptimizationGraph Neural NetworkAuto EncoderTextGraph
🎯 What it does: This paper proposes a deep unsupervised feature selection network based on Dirichlet energy, while constructing an adaptive graph structure in the feature subspace through a differentiable k-NN graph learner (DGL), achieving joint optimization of feature selection and graph learning.
Joint Learning of Label and Environment Causal Independence for Graph Out-of-Distribution Generalization
Shurui Gui (Texas A&M University), Shuiwang Ji (Texas A&M University)
Domain AdaptationAnomaly DetectionOptimizationGraph Neural NetworkGenerative Adversarial NetworkGraph
🎯 What it does: A graph OOV generalization method utilizing Label and Environmental Causal Independence (LECI) is proposed, aiming to discover causal subgraphs for robust prediction.
Joint processing of linguistic properties in brains and language models
SUBBA REDDY OOTA, Mariya Toneva (Microsoft)
TransformerLarge Language ModelTextMultimodalityMagnetic Resonance Imaging
🎯 What it does: This study investigates the impact of different syntactic/semantic attributes in language models (BERT, GPT-2) on the alignment with human brain fMRI recordings, using a linear removal method to directly assess the contribution of each attribute to brain signal prediction performance.
Joint Prompt Optimization of Stacked LLMs using Variational Inference
Alessandro Sordoni (Microsoft Research), Nicolas Le Roux (Microsoft Research)
TransformerLarge Language ModelPrompt EngineeringTextChain-of-Thought
🎯 What it does: Design and train a Deep Language Network (DLN) by stacking LLM layers and jointly optimizing prompts to improve the performance of small LLMs on reasoning and understanding tasks.
Joint Training of Deep Ensembles Fails Due to Learner Collusion
Alan Jeffares (University of Cambridge), Mihaela van der Schaar (University of Cambridge)
Convolutional Neural NetworkImageTabular
🎯 What it does: This paper studies why the joint training of deep ensemble models performs poorly and reveals the phenomenon of 'learner collusion'.
k-Median Clustering via Metric Embedding: Towards Better Initialization with Differential Privacy
Chenglin Fan (Cognitive Computing Lab Baidu Research), Xiaoyun Li (Cognitive Computing Lab Baidu Research)
OptimizationSafty and PrivacyImage
🎯 What it does: Proposes a k-median initialization method based on Hierarchical Well-Separated Trees (HST) and provides a differentially private (DP) version, followed by clustering implementation using local search.
K-Nearest-Neighbor Local Sampling Based Conditional Independence Testing
Shuai Li (East China Normal University), Yanfeng Yang (East China Normal University)
ClassificationComputational EfficiencyTabular
🎯 What it does: A conditional independence testing method based on k-nearest neighbor local sampling and classifier estimation is proposed, which can control the first type error and maintain high test power in high-dimensional conditional variable and small sample situations.
KAKURENBO: Adaptively Hiding Samples in Deep Neural Network Training
Thao Nguyen Truong, Mohamed Wahib (RIKEN Center for Computational Science)
Computational EfficiencyConvolutional Neural NetworkImage
🎯 What it does: Designed and implemented KAKURENBO, a method for dynamically hiding the least important samples during the training process of deep neural networks, significantly reducing training time while maintaining accuracy.
KD-Zero: Evolving Knowledge Distiller for Any Teacher-Student Pairs
Lujun Li (Hong Kong University of Science and Technology), Yang Ya
ClassificationObject DetectionSegmentationCompressionKnowledge DistillationConvolutional Neural NetworkTransformerImage
🎯 What it does: A KD-Zero framework is proposed for any teacher-student network pair, which automatically evolves a knowledge distiller to achieve model compression and performance improvement.
Keep Various Trajectories: Promoting Exploration of Ensemble Policies in Continuous Control
Chao Li (Institute of Automation Chinese Academy of Sciences), Xinwen Hou (Institute of Automation Chinese Academy of Sciences)
Robotic IntelligenceReinforcement LearningSequential
🎯 What it does: The TEEN algorithm is designed and implemented, which encourages trajectory diversity among multiple sub-policies through trajectory distribution differences and mutual information regularization, thereby improving sample efficiency and final performance in continuous control tasks.
Kernel Quadrature with Randomly Pivoted Cholesky
Ethan Nicholas Epperly, Elvira Moreno Ferreira
OptimizationComputational EfficiencyTabular
🎯 What it does: This paper proposes a continuous kernel quadrature method based on Random Pivot Cholesky (RP-Cholesky) for generating efficient quadrature nodes and weights in arbitrary spaces, measures, and kernels.
Kernel Stein Discrepancy thinning: a theoretical perspective of pathologies and a practical fix with regularization
Clement Benard, Sébastien Da Veiga (ENS Angers)
Gaussian SplattingTabular
🎯 What it does: This paper conducts an in-depth theoretical analysis of the Stein thinning algorithm, identifies its two major pathologies in multimodal targets, and proposes a regularized Stein thinning based on this analysis, which improves sample quality by combining entropy regularization and Laplace correction.
Kernel-Based Tests for Likelihood-Free Hypothesis Testing
Patrik Robert Gerber, Rui Sun (Massachusetts Institute of Technology)
ImagePhysics Related
🎯 What it does: A mixed likelihood-free hypothesis testing (mLFHT) model is proposed, and a learnable kernel test statistic is constructed based on MMD;
Kernelized Cumulants: Beyond Kernel Mean Embeddings
Patric Bonnier (Mathematical Institute University of Oxford), Zoltán Szabó (Department of Statistics London School of Economics)
TabularTime Series
🎯 What it does: Cumulants (kernelized cumulants) are defined in RKHS, and their applications in two-sample tests and independence tests lead to the proposal of higher-order statistics compared to MMD/HSIC.
Kernelized Reinforcement Learning with Order Optimal Regret Bounds
Sattar Vakili (MediaTek Research), Julia Olkhovskaya (TU Delft)
Reinforcement Learning
🎯 What it does: The π-KRVI algorithm is proposed, which constructs confidence upper bounds using domain partitioning kernel ridge regression and optimistic value iteration (LSVI) for the state-action value function represented in the Reproducing Kernel Hilbert Space (RKHS), and provides a sublinear regret upper bound under polynomial spectral decay kernels.
Keypoint-Augmented Self-Supervised Learning for Medical Image Segmentation with Limited Annotation
Zhangsihao Yang (Arizona State University), Yalin Wang (Arizona State University)
SegmentationConvolutional Neural NetworkTransformerContrastive LearningImageBiomedical DataMagnetic Resonance ImagingComputed Tomography
🎯 What it does: Under low-label conditions, a Keypoint Augmentation Fusion layer (KAF) was designed to inject long-range self-attention into the UNet backbone, and a model was pre-trained through global and local self-supervised learning (SSL), followed by fine-tuning on a small amount of labeled data to achieve medical image segmentation.
Kiki or Bouba? Sound Symbolism in Vision-and-Language Models
Morris Alper (Tel Aviv University), Hadar Averbuch-Elor (Tel Aviv University)
ClassificationGenerationTransformerVision Language ModelDiffusion modelContrastive LearningImageTextMultimodality
🎯 What it does: This study investigates the performance of visual-language models (CLIP and Stable Diffusion) in the association between phonemes and shapes, examining whether the kiki-bouba phenomenon naturally occurs in these models.
Kissing to Find a Match: Efficient Low-Rank Permutation Representation
Hannah Dröge (University of Siegen), Michael Moeller
OptimizationComputational EfficiencyPoint Cloud
🎯 What it does: A method is proposed to represent permutation matrices through low-rank matrix decomposition and nonlinear transformations to address the storage and computation issues of large-scale permutation matrices.
Knowledge Diffusion for Distillation
Tao Huang (University of Sydney), Chang Xu (University of Sydney)
ClassificationObject DetectionSegmentationKnowledge DistillationDiffusion modelAuto EncoderImage
🎯 What it does: This paper proposes DiffKD, a method that utilizes diffusion models to denoise student model features before knowledge distillation; by treating student features as a noisy version of teacher features, a trained diffusion model is used to progressively denoise them, resulting in cleaner features that better align with the teacher distribution, which enhances the distillation effect.
Knowledge Distillation for High Dimensional Search Index
Zepu Lu (University of Science and Technology of China), Enhong Chen (University of Science and Technology of China)
RetrievalCompressionKnowledge DistillationGraph
🎯 What it does: This study proposes a framework called KDindex, which transfers the top-k retrieval results from high-precision search indexes (such as graph-structured indexes) to lightweight quantized indexes through knowledge distillation, improving the accuracy of high-dimensional retrieval.
Knowledge Distillation Performs Partial Variance Reduction
Mher Safaryan (Institute of Science and Technology Austria), Dan Alistarh (Institute of Science and Technology Austria)
OptimizationKnowledge DistillationConvolutional Neural NetworkSupervised Fine-TuningImage
🎯 What it does: This paper explores the impact of knowledge distillation on the optimization of student models, proving that it can be viewed as a partial variance reduction mechanism and providing convergence theory.
Knowledge-Augmented Reasoning Distillation for Small Language Models in Knowledge-Intensive Tasks
Minki Kang (KRAFTON), Sung Ju Hwang (KRAFTON)
Knowledge DistillationTransformerSupervised Fine-TuningTextBiomedical DataRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: A knowledge-enhanced reasoning distillation method KARD is proposed, which combines reasoning generated by large models with external knowledge retrieval to fine-tune small language models, enabling them to generate reasonable justifications and provide correct answers in knowledge-intensive reasoning tasks.
Koopa: Learning Non-stationary Time Series Dynamics with Koopman Predictors
Yong Liu (Tsinghua University), Mingsheng Long (Tsinghua University)
OptimizationAuto EncoderTime Series
🎯 What it does: The Koopa model is designed and implemented, which decomposes non-stationary time series into time-invariant and time-variant components using the Koopman theory, employing stacked Koopman Predictor blocks for hierarchical forecasting.
Koopman Kernel Regression
Petar Bevanda (TU Munich), Sandra Hirche (TU Munich)
Time Series
🎯 What it does: This paper proposes and implements Koopman Kernel Regression (KKR), which utilizes RKHS to learn linear time-invariant (LTI) predictors, achieving multi-step predictions from system trajectories to future states.
Kronecker-Factored Approximate Curvature for Modern Neural Network Architectures
Runa Eschenhagen (University of Cambridge), Philipp Hennig (University of Tübingen)
OptimizationGraph Neural NetworkTransformerImageGraph
🎯 What it does: Extend the K-FAC method to modern networks with linear weight sharing layers, and propose two variants: K-FAC-expand and K-FAC-reduce.
Kullback-Leibler Maillard Sampling for Multi-armed Bandits with Bounded Rewards
Hao Qin (University of Arizona), Chicheng Zhang (University of Arizona)
Reinforcement LearningTabular
🎯 What it does: This paper proposes the Kullback-Leibler Maillard Sampling (KL-MS) algorithm for addressing the multi-armed bandit problem with reward distributions constrained to [0,1], and provides a closed-form action probability along with a complete finite-time and asymptotic optimal theoretical analysis.
L-C2ST: Local Diagnostics for Posterior Approximations in Simulation-Based Inference
Julia Linhart (Universite Paris-Saclay), Pedro L. C. Rodrigues (Univ. Grenoble Alpes)
Flow-based ModelTabularBenchmark
🎯 What it does: A local diagnostic method ℓ-C2ST is proposed to assess the local consistency of posterior approximations in Simulation-Based Inference (SBI). This method only requires samples from the joint distribution, can be tested under any given observation, and provides interpretable graphical diagnostics.
L-CAD: Language-based Colorization with Any-level Descriptions using Diffusion Priors
Zheng Chang (Beijing University of Posts and Telecommunications), Boxin Shi (Peking University)
Image TranslationGenerationData SynthesisDiffusion modelImageText
🎯 What it does: A unified language-driven image coloring model L-CAD is proposed, which can accept text descriptions at any level (complete, partial, sparse) and generate colored images that match the descriptions.
L2T-DLN: Learning to Teach with Dynamic Loss Network
Zhaoyang Hai (Beijing Institute of Technology), Mirna Yunita (Beijing Institute of Technology)
ClassificationObject DetectionSegmentationKnowledge DistillationRecurrent Neural NetworkImage
🎯 What it does: A learning-to-teach (L2T-DLN) framework based on Dynamic Loss Networks (DLN) and an LSTM teacher model is proposed, achieving adaptive adjustment of the dynamic loss function through a three-stage asynchronous optimization (student, DLN, teacher) to enhance the performance of the student model across various tasks.
Label Correction of Crowdsourced Noisy Annotations with an Instance-Dependent Noise Transition Model
Hui Guo (University of Western Ontario), Grace Yi
ClassificationImage
🎯 What it does: A Bayesian framework-based instance-dependent noise transfer matrix model is proposed, which is used to correct crowd-sourced noisy labels, ultimately training a more accurate classifier.
Label Poisoning is All You Need
Rishi Dev Jha (University of Washington), Sewoong Oh (University of Washington)
Knowledge DistillationAdversarial AttackConvolutional Neural NetworkTransformerImage
🎯 What it does: A method for backdoor attacks called FLIP is proposed, which achieves backdoor attacks solely through label poisoning, and is extended to softFLIP in the context of knowledge distillation.
Label Robust and Differentially Private Linear Regression: Computational and Statistical Efficiency
Xiyang Liu (University of Washington), Arun Suggala
Safty and PrivacyComputational EfficiencyGaussian SplattingTabular
🎯 What it does: This paper proposes a robust algorithm for linear regression that satisfies (ε,δ)-differential privacy (Alg.1), capable of ensuring low sample complexity and linear time complexity even in the presence of label tampering by an attacker.
Label-efficient Segmentation via Affinity Propagation
Wentong Li (Zhejiang University), Lei Zhang (Ant Group)
SegmentationGraph Neural NetworkImage
🎯 What it does: A pluggable Affinity Propagation (APro) module is proposed, which iteratively generates soft pseudo-labels using global and local pixel mutual information, thereby enhancing weakly supervised segmentation performance.
Label-Only Model Inversion Attacks via Knowledge Transfer
Ngoc-Bao Nguyen (Singapore University of Technology and Design), Ngai-man Cheung
RecognitionGenerationAdversarial AttackGenerative Adversarial NetworkImage
🎯 What it does: This paper proposes a Label-Only model inversion attack method based on knowledge transfer, called LOKT. It first trains a generative adversarial network T-ACGAN using the hard labels from the target model to obtain a proxy model, and then performs a white-box attack on the proxy model to reconstruct private data.
Label-Retrieval-Augmented Diffusion Models for Learning from Noisy Labels
Jian Chen (University at Buffalo), Changyou Chen (University at Buffalo)
ClassificationData-Centric LearningDiffusion modelContrastive LearningImage
🎯 What it does: Addressed the problem of learning from noisy labels
Labeling Neural Representations with Inverse Recognition
Kirill Bykov (ATB Potsdam), Marina MC Höhne
Explainability and InterpretabilityComputational EfficiencyConvolutional Neural NetworkTransformerImage
🎯 What it does: The INVERT method is proposed, which uses inverse recognition technology to associate the internal representations of neural networks with interpretable concepts, thereby providing global explanations for individual neurons or sub-networks.
LaFTer: Label-Free Tuning of Zero-shot Classifier using Language and Unlabeled Image Collections
Muhammad Jehanzeb Mirza (Institute of Computer Graphics and Vision), Horst Bischof (Institute of Computer Graphics and Vision)
ClassificationRepresentation LearningTransformerLarge Language ModelVision Language ModelContrastive LearningImageTextMultimodality
🎯 What it does: This paper proposes a completely label-free and parameter-efficient visual-language model fine-tuning method called LaFTer, which utilizes text descriptions generated by LLMs for self-supervised training of text, and then conducts pseudo-label self-training on unlabeled images, significantly improving zero-shot classification performance.
LambdaBeam: Neural Program Search with Higher-Order Functions and Lambdas
Kensen Shi (Google DeepMind), Charles Sutton (Google DeepMind)
Large Language ModelTabular
🎯 What it does: A neural search method called LAMBDABEAM is proposed, which can dynamically generate and use λ-functions in program synthesis.
LANCE: Stress-testing Visual Models by Generating Language-guided Counterfactual Images
Viraj Uday Prabhu (Georgia Institute of Technology), Judy Hoffman (Georgia Institute of Technology)
GenerationAdversarial AttackTransformerLarge Language ModelPrompt EngineeringDiffusion modelImageText
🎯 What it does: An automated method called LANCE is proposed, which uses language-guided adversarial image generation to stress test trained visual models and generate diverse, realistic, and challenging adversarial samples.
Landscape Surrogate: Learning Decision Losses for Mathematical Optimization Under Partial Information
Arman Zharmagambetov (Meta), Yuandong Tian (Meta)
OptimizationTabularFinance Related
🎯 What it does: A framework named LANCER is proposed for learning decision loss under partial information to accelerate the mathematical optimization process.
Langevin Quasi-Monte Carlo
Sifan Liu (Stanford University)
OptimizationStochastic Differential Equation
🎯 What it does: A new Langevin quasi-Monte Carlo (LQMC) algorithm is proposed, which improves the sampling quality of the Langevin Monte Carlo (LMC) algorithm by using low-discrepancy complete uniform distribution (CUD) sequences.
Language Is Not All You Need: Aligning Perception with Language Models
Shaohan Huang (Microsoft), Furu Wei (Microsoft)
TransformerLarge Language ModelVision Language ModelImageTextMultimodality
🎯 What it does: KOSMOS-1 is proposed, a multimodal large language model trained from scratch, capable of perceiving, following instructions, and learning context across general modalities such as text and images;
Language Model Alignment with Elastic Reset
Michael Noukhovitch (Mila, Université de Montréal), Aaron Courville (Mila, Université de Montréal)
Reinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: An algorithm called Elastic Reset is proposed to reduce language drift and enhance rewards during RLHF fine-tuning of language models through a periodic reset strategy and exponential moving average models.
Language Model Tokenizers Introduce Unfairness Between Languages
Aleksandar Petrov (University of Oxford), Adel Bibi (University of Oxford)
Large Language ModelText
🎯 What it does: This paper studies the unfairness of language model tokenizers across different languages, quantifying the differences in token lengths and discussing their impact on cost, latency, and context capacity.
Language Models are Weak Learners
Hariharan Manikandan (Bosch Center for Artificial Intelligence), J Zico Kolter (Carnegie Mellon University)
TransformerLarge Language ModelPrompt EngineeringTabular
🎯 What it does: By converting tabular data into natural language descriptions, large language models generate weak learners through summarization, which are then integrated into AdaBoost for boosting learning.
Language Models Can Improve Event Prediction by Few-Shot Abductive Reasoning
Xiaoming Shi (Ant Group), Hongyuan Mei (TTIC)
TransformerLarge Language ModelPrompt EngineeringTextTime SeriesSequentialRetrieval-Augmented Generation
🎯 What it does: The LAMP framework is proposed, utilizing large language models for spontaneous reasoning to provide causal evidence for event sequence models, thereby enhancing the predictive performance of future events.
Language Models can Solve Computer Tasks
Geunwoo Kim (University of California), Stephen Marcus McAleer
TransformerLarge Language ModelReinforcement LearningPrompt EngineeringTextChain-of-Thought
🎯 What it does: A recursive critique and improvement (RCI) prompting method based on large language models is proposed to perform computer tasks and enhance reasoning capabilities.
Language Models Don't Always Say What They Think: Unfaithful Explanations in Chain-of-Thought Prompting
Miles Turpin (New York University), Samuel R. Bowman (New York University)
Explainability and InterpretabilityAdversarial AttackTransformerLarge Language ModelPrompt EngineeringTextBenchmarkChain-of-Thought
🎯 What it does: This paper explores the trustworthiness of Chain of Thought (CoT) explanations through experiments with large language models (GPT-3.5 and Claude-1.0) under CoT prompting. It finds that the models are susceptible to input biases (such as always suggesting answer A) which lead to unfaithful CoT explanations, resulting in decreased accuracy.
Language Models Meet World Models: Embodied Experiences Enhance Language Models
Jiannan Xiang (University of California San Diego), Zhiting Hu (University of California San Diego)
TransformerLarge Language ModelSupervised Fine-TuningWorld ModelText
🎯 What it does: Fine-tune pre-trained language models using embodied experiences collected from a virtual home environment (VirtualHome) to enhance their reasoning and planning capabilities in physical environments.
Language Quantized AutoEncoders: Towards Unsupervised Text-Image Alignment
Hao Liu (University of California Berkeley), Pieter Abbeel (University of California Berkeley)
GenerationRepresentation LearningTransformerLarge Language ModelAuto EncoderImageText
🎯 What it does: Proposes the Language Quantized AutoEncoder (LQAE), a variant of VQ-VAE that unsupervisedly maps images to text space, utilizing a pre-trained language denoising model to generate text encodings and perform image reconstruction;
Language Semantic Graph Guided Data-Efficient Learning
Wenxuan Ma (Beijing Institute of Technology), Jingxuan Kang (University of Liverpool)
ClassificationData-Centric LearningGraph Neural NetworkLarge Language ModelSupervised Fine-TuningImageVideoAudio
🎯 What it does: This paper proposes a framework that utilizes label semantic information through a Language Semantic Graph (LSG) to enhance data-efficient learning;
Language-based Action Concept Spaces Improve Video Self-Supervised Learning
Kanchana Ranasinghe (Stony Brook University), Michael S Ryoo
ClassificationRepresentation LearningTransformerLarge Language ModelContrastive LearningVideoText
🎯 What it does: A language-based self-supervised learning framework LSS is proposed, utilizing the visual-text alignment features of CLIP and the action concept space for video representation learning, achieving label-free zero-shot video classification.
Language-driven Scene Synthesis using Multi-conditional Diffusion Model
An Dinh Vuong (FSOFT AI Center), Anh Nguyen (University of Liverpool)
GenerationData SynthesisDiffusion modelTextPoint Cloud
🎯 What it does: A language-driven scene synthesis task is proposed, which jointly generates new objects from text prompts, human actions, and existing objects in three modalities.
Laplacian Canonization: A Minimalist Approach to Sign and Basis Invariant Spectral Embedding
George Ma, Yisen Wang (Peking University)
Graph Neural NetworkGraphBiomedical Data
🎯 What it does: A normalization method for the eigenvectors of the Laplacian operator (Laplacian Canonization) is proposed to eliminate sign and basis ambiguities in spectral embedding;
Large Language Models Are Latent Variable Models: Explaining and Finding Good Demonstrations for In-Context Learning
Xinyi Wang (University of California), William Yang Wang (University of California)
ClassificationTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: A latent variable model for LLM is proposed from a Bayesian perspective, and a two-step demonstration selection algorithm is designed: first, a small LLM learns the task concept words, and then examples that best infer the concept are selected for in-context learning in the large model.
Large Language Models Are Semi-Parametric Reinforcement Learning Agents
Danyang Zhang (Shanghai Jiao Tong University), Kai Yu (Shanghai Jiao Tong University)
TransformerLarge Language ModelReinforcement LearningPrompt EngineeringText
🎯 What it does: A semi-parametric reinforcement learning framework called REMEMBERER based on LLM is proposed, which utilizes external sustainable experience memory to enhance decision-making performance.
Large Language Models are Visual Reasoning Coordinators
Liangyu Chen (Nanyang Technological University), Ziwei Liu (Nanyang Technological University)
TransformerLarge Language ModelPrompt EngineeringVision Language ModelMultimodality
🎯 What it does: This paper proposes Cola, which utilizes large language models as coordinators to integrate descriptions and answers generated by multiple visual language models (such as OFA and BLIP) through natural language prompts to complete visual reasoning tasks.
Large Language Models Are Zero-Shot Time Series Forecasters
Nate Gruver (New York University), Andrew Gordon Wilson (New York University)
TransformerLarge Language ModelTime Series
🎯 What it does: This paper proposes a framework called LLMTIME that transforms time series into text strings and then utilizes pre-trained large language models (LLMs) for zero-shot prediction.
Large Language Models as Commonsense Knowledge for Large-Scale Task Planning
Zirui Zhao (National University of Singapore), David Hsu (National University of Singapore)
TransformerLarge Language ModelReinforcement LearningWorld ModelText
🎯 What it does: This paper proposes the LLM-MCTS method, which uses large language models (LLMs) as both a world model and a search heuristic to solve large-scale task planning problems.
Large Language Models can Implement Policy Iteration
Ethan Brooks (University of Michigan), Satinder Singh (University of Michigan)
TransformerLarge Language ModelReinforcement LearningPrompt EngineeringText
🎯 What it does: Implementing policy iteration using large language models, leveraging context learning to infer world models and policies, completing Q-value estimation and greedy improvement without the need for gradients or expert demonstrations.
Large Language Models for Automated Data Science: Introducing CAAFE for Context-Aware Automated Feature Engineering
Noah Hollmann (University of Freiburg), Frank Hutter (University of Freiburg)
Explainability and InterpretabilityData-Centric LearningTransformerLarge Language ModelPrompt EngineeringTabularChain-of-Thought
🎯 What it does: Using large language models to achieve context-aware automatic feature engineering (CAAFE), generating interpretable Python feature engineering code to enhance the predictive performance of tabular data.
Large language models implicitly learn to straighten neural sentence trajectories to construct a predictive representation of natural language.
Eghbal A. Hosseini (Massachusetts Institute of Technology), Evelina Fedorenko (Massachusetts Institute of Technology)
TransformerLarge Language ModelText
🎯 What it does: This study investigates the geometric structure of internal representations in large-scale language models, proposing and validating the hypothesis that sentence representation trajectories tend to become 'straightened' at different layers of the model, explaining the mechanism behind prediction tasks.
Large Language Models of Code Fail at Completing Code with Potential Bugs
Tuan Dinh (University of Wisconsin-Madison), George Karypis (Amazon Web Services)
AI Code AssistantTransformerLarge Language ModelText
🎯 What it does: This study investigates the ability of large language models to complete code in the presence of potentially defective code contexts, and proposes two new datasets and three post-correction methods.
Large language models transition from integrating across position-yoked, exponential windows to structure-yoked, power-law windows
David Skrill (University of Rochester Medical Center), Samuel Victor Norman-Haignere
TransformerLarge Language ModelText
🎯 What it does: This paper measures the time integration window of large language models (LLMs) through a black-box method based on word replacement and further quantifies how these windows change with sentence structure.
Large-Scale Distributed Learning via Private On-Device LSH
Tahseen Rabbani (University of Maryland), Furong Huang (University of Maryland)
Recommendation SystemFederated LearningSafty and PrivacyComputational EfficiencyTabular
🎯 What it does: This paper proposes the PGHash framework, which achieves privatized, personalized, and memory-efficient LSH pruning on the device side for distributed/federated training of large-scale recommendation networks.
LART: Neural Correspondence Learning with Latent Regularization Transformer for 3D Motion Transfer
Haoyu Chen (University of Oulu), Guoying Zhao (University of Oulu)
GenerationData SynthesisPose EstimationTransformerMesh
🎯 What it does: This paper proposes a Transformer-based 3D motion transfer framework called LART, which can transfer motion from a driving sequence to unseen static 3D meshes while achieving high-fidelity motion generation and preserving the geometric details of the meshes.
Last-Iterate Convergent Policy Gradient Primal-Dual Methods for Constrained MDPs
Dongsheng Ding (University of Pennsylvania), Alejandro Ribeiro (University of Pennsylvania)
OptimizationReinforcement Learning
🎯 What it does: Two single-time-scale policy gradient dual optimization methods (RPG-PD and OPG-PD) are proposed to solve the optimal policy for infinite discount constrained Markov decision processes, and it is proven that the last iteration of its policy iteration can converge to the constrained optimal solution in a globally non-asymptotic manner.
Latent Diffusion for Language Generation
Justin Lovelace (Cornell University), Kilian Q Weinberger
GenerationData SynthesisCompressionTransformerLarge Language ModelDiffusion modelAuto EncoderText
🎯 What it does: A self-encoder is constructed using a pre-trained encoder-decoder language model to compress variable-length text into a fixed-length low-dimensional latent representation, and then a continuous diffusion model is trained on this latent space to generate natural language.
Latent exploration for Reinforcement Learning
Alberto Silvio Chiappa (École Polytechnique Fédérale de Lausanne), Alexander Mathis (École Polytechnique Fédérale de Lausanne)
Reinforcement LearningSequentialStochastic Differential Equation
🎯 What it does: This paper proposes an exploration method called Lattice, which injects time-related and actuator-related noise into the last layer of the policy network's latent state to enhance learning efficiency in high-dimensional continuous control tasks.
Latent Field Discovery in Interacting Dynamical Systems with Neural Fields
Miltiadis Kofinas (University of Amsterdam), Efstratios Gavves (University of Amsterdam)
Graph Neural NetworkAuto EncoderTime SeriesSequentialPhysics Related
🎯 What it does: The Aether framework is proposed, which utilizes neural fields and equivariant graph networks to unsupervisedly discover global field effects in the system from observed trajectories and applies them to trajectory prediction.
Latent Graph Inference with Limited Supervision
Jianglin Lu (Northeastern University), Yun Fu (Northeastern University)
OptimizationGraph Neural NetworkGraph
🎯 What it does: To address the supervision starvation problem in latent graph inference, this paper identifies and recovers key edges that have been damaged by sparsification through CUR decomposition to supplement supervision.
Latent SDEs on Homogeneous Spaces
Sebastian Zeng (University of Salzburg), Roland Kwitt (University of Salzburg)
Time SeriesStochastic Differential Equation
🎯 What it does: This paper proposes a variational Bayesian method for learning latent stochastic differential equations (latent SDEs) on a homeomorphic space (using the unit sphere as an example), constructing solvable and structure-preserving SDEs using the action of Lie groups;