NeurIPS 2023 Papers — Page 4
Conference on Neural Information Processing Systems · 3218 papers
Augmented Memory Replay-based Continual Learning Approaches for Network Intrusion Detection
Suresh kumar Amalapuram, Bheemarjuna Tamma
Anomaly DetectionConvolutional Neural NetworkTabularTime Series
🎯 What it does: This paper addresses the problem of continual learning in network intrusion detection and proposes two enhancement methods: Extended Class-Balanced Random Sampling (ECBRS) and Perturbation-Assisted Parameter Approximation for Virtual SGD Updates (PAPA);
Augmenting Language Models with Long-Term Memory
Weizhi Wang (University of California Santa Barbara), Furu Wei (Microsoft Research)
TransformerLarge Language ModelTextRetrieval-Augmented Generation
🎯 What it does: This paper proposes a framework called LONGMEM that combines large language models with long-term memory, utilizing a frozen main model as a memory encoder and a lightweight side network (SideNet) as a memory retriever and reader to achieve retrieval and integration of long contexts.
Autodecoding Latent 3D Diffusion Models
Evangelos Ntavelis (ETH Zurich), Sergey Tulyakov (Snap Inc)
GenerationData SynthesisPose EstimationConvolutional Neural NetworkDiffusion modelAuto EncoderImageVideo
🎯 What it does: This paper proposes a framework based on a 3D self-decoder and a 3D diffusion model for learning and generating static and dynamic 3D assets from 2D images or monocular videos.
AutoGO: Automated Computation Graph Optimization for Neural Network Evolution
Mohammad Salameh (Huawei Technologies Canada), Di Niu (University of Alberta)
Super ResolutionOptimizationNeural Architecture SearchConvolutional Neural NetworkReinforcement LearningImage
🎯 What it does: In this paper, the authors propose an automated computation graph optimization framework called AutoGO, which achieves the goal of significantly reducing FLOPs and hardware latency while maintaining or improving task performance by segmenting, mutating, and evaluating hardware friendliness of the low-level computation graph of neural networks.
Automated Classification of Model Errors on ImageNet
Momchil Peychev (ETH Zurich), Martin Vechev (ETH Zurich)
ClassificationRecognitionConvolutional Neural NetworkContrastive LearningImage
🎯 What it does: An automated error classification framework is proposed for fine-grained classification of model errors on ImageNet and evaluation of error distribution across different models.
Automatic Clipping: Differentially Private Deep Learning Made Easier and Stronger
Zhiqi Bu (Amazon Web Services), George Karypis (Amazon Web Services)
OptimizationSafty and PrivacyImageTextTabular
🎯 What it does: A technique called Automatic Clipping is proposed, which eliminates the need to adjust the clipping threshold R in DP training. It directly implements gradient clipping using a normalization method for each sample's gradient (with a stability constant γ), compatible with various DP optimizers.
Automatic Grouping for Efficient Cooperative Multi-Agent Reinforcement Learning
Yifan Zang (Institute of Automation, Chinese Academy of Sciences), Jian Cheng (Institute of Automation, Chinese Academy of Sciences)
Reinforcement LearningTabular
🎯 What it does: This paper proposes an automatic grouping multi-agent reinforcement learning framework called GoMARL, which utilizes subgroups as a bridge to achieve efficient collaboration.
Automatic Integration for Spatiotemporal Neural Point Processes
Zihao Zhou (University of California San Diego), Rose Yu (University of California San Diego)
Point CloudTime Series
🎯 What it does: Proposes the AutoSTPP automatic integration framework, which utilizes a dual network to achieve precise likelihood estimation of three-dimensional space-time point processes.
Autonomous Capability Assessment of Sequential Decision-Making Systems in Stochastic Settings
Pulkit Verma (Arizona State University), Siddharth Srivastava (Arizona State University)
OptimizationExplainability and InterpretabilityRobotic IntelligenceReinforcement LearningAgentic AISequential
🎯 What it does: This paper proposes an interactive query-based algorithm QACE for learning interpretable probability transition models in black-box sequential decision systems, helping users assess the capabilities of AI systems.
Auxiliary Losses for Learning Generalizable Concept-based Models
Ivaxi Sheth (CISPA Helmholtz Center for Information Security), Samira Ebrahimi Kahou (École de technologie supérieure)
ClassificationRepresentation LearningImage
🎯 What it does: A collaborative concept bottleneck model (Coop-CBM) is proposed, which incorporates auxiliary task label prediction into the concept bottleneck model, and enhances the separability of concept representations and task accuracy through concept orthogonal loss (COL).
AV-NeRF: Learning Neural Fields for Real-World Audio-Visual Scene Synthesis
Susan Liang (University of Rochester), Chenliang Xu (University of Rochester)
GenerationData SynthesisNeural Radiance FieldVideoMultimodalityAudio
🎯 What it does: This paper proposes a NeRF-based system called AV-NeRF, which can synthesize visual videos and corresponding spatial stereo audio under any camera trajectory, achieving synchronized audio-video synthesis of real scenes.
AVIS: Autonomous Visual Information Seeking with Large Language Model Agent
Ziniu Hu (Google Research), Alireza Fathi (Google Research)
Object DetectionRetrievalTransformerLarge Language ModelAgentic AIPrompt EngineeringImageText
🎯 What it does: This paper proposes the AVIS framework, which utilizes large language models to dynamically call various tools (such as image captioning, VQA, object detection, OCR, image search, web search, etc.) through tree search to answer knowledge-intensive visual questions.
Back-Modality: Leveraging Modal Transformation for Data Augmentation
Zhi Li (Zhejiang University), Yin Zhang (Zhejiang University)
ClassificationData SynthesisTransformerLarge Language ModelDiffusion modelImageTextMultimodalityAudio
🎯 What it does: The Back-Modality framework is proposed, utilizing cross-modal transformations (image↔text, text↔speech) to achieve data augmentation, and transferring single-modal augmentation methods to another modality.
BadTrack: A Poison-Only Backdoor Attack on Visual Object Tracking
Bin Huang (Tsinghua University), Zhi Wang (Tsinghua University)
Object TrackingAdversarial AttackVideo
🎯 What it does: This paper proposes a backdoor attack (BadTrack) for visual object tracking (VOT) models that only poisons the data during the training phase. By embedding a trigger in the background of video frames, the tracker maintains performance under normal inputs but loses tracking when the trigger appears.
Balance, Imbalance, and Rebalance: Understanding Robust Overfitting from a Minimax Game Perspective
Yifei Wang (Peking University), Yisen Wang (Peking University)
OptimizationAdversarial AttackConvolutional Neural NetworkImage
🎯 What it does: This study investigates the phenomenon of robust overfitting in adversarial training and explains its mechanism from the perspective of minimax games. It proposes a new adversarial training method called ReBAT, which suppresses robust overfitting through three techniques for rebalancing the trainer and the attacker (BoAT loss, smaller learning rate decay factor, and increased attack strength during training).
Balanced Training for Sparse GANs
Yite Wang (University of Illinois Urbana-Champaign), Ruoyu Sun (Chinese University of Hong Kong)
GenerationData SynthesisGenerative Adversarial NetworkImage
🎯 What it does: This paper studies the application of Dynamic Sparse Training (DST) in Generative Adversarial Networks (GANs) and proposes the Balance Ratio (BR) metric to measure the balance between the generator and discriminator. Based on this, the ADAPT algorithm is designed to achieve balanced training of sparse GANs.
Balancing memorization and generalization in RNNs for high performance brain-machine Interfaces
Joseph T Costello, Cynthia Chestek
Hyperparameter SearchConvolutional Neural NetworkRecurrent Neural NetworkTransformerSupervised Fine-TuningTime SeriesSequentialBiomedical Data
🎯 What it does: This paper compares the real-time closed-loop decoding performance of LSTM, GRU, convolutional networks, Transformer, and Kalman filter in a multi-degree-of-freedom fingertip brain-machine interface experiment, and explores the impact of RNN's memorization strategy on functional recovery.
Balancing Risk and Reward: A Batched-Bandit Strategy for Automated Phased Release
Yufan Li (Harvard University), Iavor Bojinov (Harvard Business School)
🎯 What it does: A risk-budget-based batch Bandit algorithm is proposed, which automatically determines the allocation ratio of phased release at each stage, achieving a balance between risk control and rapid scaling.
Banana: Banach Fixed-Point Network for Pointcloud Segmentation with Inter-Part Equivariance
Congyue Deng (Stanford University), Leonidas Guibas (Stanford University)
SegmentationGraph Neural NetworkPoint Cloud
🎯 What it does: This paper proposes a Banana network that achieves cross-component equivariance during point cloud segmentation through fixed-point iteration, enabling robust segmentation under pose transformations between components in multi-body systems (such as articulated objects or multi-object scenes).
Bandit Social Learning under Myopic Behavior
Kiarash Banihashem (University of Maryland), Aleksandrs Slivkins (Microsoft Research)
🎯 What it does: This paper studies the multi-armed bandit (Bandit Social Learning) problem within the framework of social learning, assuming that users (agents) arrive sequentially and make greedy or biased decisions based solely on historical information. It explores how this 'profit-driven' behavior can lead to exploration failures and high scheduling losses in the system.
Bandit Task Assignment with Unknown Processing Time
Shinji Ito (NEC Corporation), Ken-Ichi Kawarabayashi (National Institute of Informatics)
OptimizationReinforcement Learning from Human FeedbackReinforcement LearningTabular
🎯 What it does: A new model of 'Bandit Task Allocation' is proposed, integrating combinatorial constraints, unknown processing times, and rewards within a multi-armed bandit framework, along with a corresponding definition of the optimal reward strategy.
BanditPAM++: Faster $k$-medoids Clustering
Mo Tiwari (Stanford University), Martin Jinye Zhang (Carnegie Mellon University)
OptimizationComputational EfficiencyImageText
🎯 What it does: This paper presents BanditPAM++, an improved version of the k-medoids clustering algorithm that significantly enhances running efficiency while maintaining clustering quality comparable to BanditPAM and PAM.
BasisFormer: Attention-based Time Series Forecasting with Learnable and Interpretable Basis
Zelin Ni (Shanghai Jiao Tong University), Weiyao Lin (Shanghai Jiao Tong University)
TransformerContrastive LearningTime Series
🎯 What it does: Proposes BasisFormer, an end-to-end time series forecasting framework based on learnable and interpretable bases.
Batch Bayesian Optimization For Replicable Experimental Design
Zhongxiang Dai (National University of Singapore), Patrick Jaillet (Massachusetts Institute of Technology)
OptimizationHyperparameter SearchRobotic IntelligenceReinforcement Learning from Human FeedbackReinforcement LearningTabularTime SeriesAgriculture Related
🎯 What it does: A batch Bayesian optimization framework called BTS-RED is proposed, which can balance the evaluation of more unique experimental conditions with multiple replications for each condition under a fixed total budget, and supports risk-averse optimization.
Batchnorm Allows Unsupervised Radial Attacks
Amur Ghose (University of Waterloo), Pascal Poupart (University of Waterloo)
Adversarial AttackConvolutional Neural NetworkTransformerImage
🎯 What it does: A label-free adversarial attack method is proposed, which generates adversarial samples using the angular loss from the intermediate layers of batch normalization networks.
Bayes beats Cross Validation: Efficient and Accurate Ridge Regression via Expectation Maximization
Shu Tew, Daniel F. Schmidt (Monash University)
Computational EfficiencyHyperparameter SearchTabularTime Series
🎯 What it does: A Bayesian ridge regression-based Expectation-Maximization (EM) algorithm is proposed for quickly and accurately tuning the regularization hyperparameter λ without the need for grid search, along with complete theoretical and implementation details.
BayesDAG: Gradient-Based Posterior Inference for Causal Discovery
Yashas Annadani (KTH Royal Institute of Technology), Wenbo Gong (Microsoft Research)
GraphStochastic Differential Equation
🎯 What it does: Proposes the BayesDAG framework, which utilizes the combination of SG-MCMC and VI to achieve Bayesian posterior sampling of causal structures and function parameters.
Bayesian Active Causal Discovery with Multi-Fidelity Experiments
Zeyu Zhang (Renmin University of China), Xing Xie (Microsoft Research Asia)
Graph
🎯 What it does: This paper studies active causal discovery in a multi-fidelity experimental environment and proposes a complete framework based on Bayesian methods.
Bayesian Extensive-Rank Matrix Factorization with Rotational Invariant Priors
Farzad Pourkamali (Ecole Polytechnique Federale de Lausanne), Nicolas Macris (Ecole Polytechnique Federale de Lausanne)
🎯 What it does: In the matrix factorization (MF) problem, the author presents a computable rotation-invariant estimator (RIE) for reconstructing two hidden factor matrices from a noise-polluted observation matrix, specifically addressing the high-rank (extensive-rank) limit.
Bayesian Learning of Optimal Policies in Markov Decision Processes with Countably Infinite State-Space
Saghar Adler (University of Michigan), Vijay Subramanian (University of Michigan)
OptimizationReinforcement LearningSequential
🎯 What it does: The study investigates the unknown parameters of Markov Decision Processes (MDP) in countably infinite state spaces, using Bayesian posterior sampling and dynamic episode Thompson Sampling methods to learn optimal or approximately optimal control policies.
Bayesian Learning via Q-Exponential Process
Shuyi Li (Arizona State University), Shiwei Lan (Arizona State University)
ImageTime SeriesComputed TomographyFinance Related
🎯 What it does: Proposes the q-exponential process (Q-EP) as a Bayesian prior for Lq regularization, addressing the limitations of GP and Besov in high-dimensional function spaces;
Bayesian Metric Learning for Uncertainty Quantification in Image Retrieval
Frederik Rahbæk Warburg (Technical University of Denmark), Søren Hauberg (Technical University of Denmark)
RetrievalContrastive LearningImage
🎯 What it does: A Bayesian metric learning method based on Laplace approximation (LAM) is proposed, treating the contrastive loss as the negative log-likelihood on the sphere and deriving a positive definite Hessian approximation;
Bayesian nonparametric (non-)renewal processes for analyzing neural spike train variability
David Liu (University of Cambridge), Máté Lengyel (Central European University)
Spiking Neural NetworkTime Series
🎯 What it does: A Bayesian nonparametric non-renewal process (NPNR) is proposed for modeling instantaneous fluctuations and dependencies in neural spike time series.
Bayesian Optimisation of Functions on Graphs
Xingchen Wan (University of Oxford), Xiaowen Dong (University of Oxford)
OptimizationGraph Neural NetworkGraph
🎯 What it does: This paper proposes a Bayesian optimization framework (BayesOptG) for optimizing black-box functions defined on nodes in unknown or large graph structures, capable of efficiently finding the global optimum node within a limited number of queries.
Bayesian Optimization with Cost-varying Variable Subsets
Sebastian Shenghong Tay (National University of Singapore), Bryan Kian Hsiang Low (National University of Singapore)
Optimization
🎯 What it does: Proposed and solved the Bayesian optimization problem with controllable variable subsets that can be selected in each round of experiments, where each subset has different costs (BOCVS).
Bayesian Risk-Averse Q-Learning with Streaming Observations
Yuhao Wang (Georgia Institute of Technology), Enlu Zhou (Georgia Institute of Technology)
OptimizationReinforcement LearningSequential
🎯 What it does: This paper proposes a multi-stage Bayesian risk-averse Q-learning algorithm that utilizes real-time streaming observations to update the Bayesian posterior, addressing the optimal policy problem of the infinite-horizon Markov decision process (BRMDP).
Bayesian target optimisation for high-precision holographic optogenetics
Marcus Triplett, Liam Paninski (Columbia University)
OptimizationBiomedical Data
🎯 What it does: This paper proposes a Bayesian inference-based holographic optogenetic target optimization framework called Bataro, aimed at minimizing off-target activation in two-photon holographic optogenetics.
BayesTune: Bayesian Sparse Deep Model Fine-tuning
Minyoung Kim (Samsung AI Center), Timothy Hospedales (University of Edinburgh)
OptimizationTransformerSupervised Fine-TuningImageText
🎯 What it does: We propose and implement BayesTune, an automatic sparse fine-tuning framework based on Bayesian inference, which determines which base model parameters need to be updated using posterior scale values, reducing the need for manual design and additional modules.
BCDiff: Bidirectional Consistent Diffusion for Instantaneous Trajectory Prediction
Rongqing Li (Beijing Institute of Technology), Guoren Wang (Beijing Institute of Technology)
Recurrent Neural NetworkTransformerDiffusion modelTime SeriesSequential
🎯 What it does: A bidirectional consistent diffusion framework BCDiff is proposed for instantaneous trajectory prediction, which simultaneously generates unobserved historical trajectories and future trajectories under limited observations by coupling two diffusion models.
Behavior Alignment via Reward Function Optimization
Dhawal Gupta (University of Massachusetts), Bruno Castro da Silva (University of Massachusetts)
OptimizationReinforcement Learning
🎯 What it does: The BARFI framework is proposed, which automatically learns behavior alignment reward functions through double-layer optimization to integrate auxiliary rewards and enhance the efficiency of reinforcement learning.
Belief Projection-Based Reinforcement Learning for Environments with Delayed Feedback
Jangwon Kim (Pohang University of Science and Technology), Soohee Han (Pohang University of Science and Technology)
Reinforcement LearningTabular
🎯 What it does: The BPQL (Belief Projection-Based Q-learning) algorithm is proposed for training reinforcement learning agents in environments with fixed delay feedback, addressing the state space explosion problem caused by traditional augmented state methods.
BERT Lost Patience Won't Be Robust to Adversarial Slowdown
Zachary Coalson (Oregon State University), Sanghyun Hong (Oregon State University)
Adversarial AttackTransformerLarge Language ModelText
🎯 What it does: Developed the WAFFLE attack method to assess the robustness of multi-output language models against 'slowdown' adversarial attacks.
Best Arm Identification with Fixed Budget: A Large Deviation Perspective
Po-An Wang (KTH Royal Institute of Technology), Alexandre Proutiere (KTH Royal Institute of Technology)
Reinforcement Learning
🎯 What it does: This paper studies how to identify the Best Arm (BAI-FB) through adaptive sampling strategies under a fixed sampling budget.
Beta Diffusion
Mingyuan Zhou (University of Texas at Austin), Huangjie Zheng (University of Texas at Austin)
GenerationData SynthesisDiffusion modelImage
🎯 What it does: Beta Diffusion is proposed, utilizing a multiplicative Beta process for training and sampling generative models within the range of 0 to 1.
Better Correlation and Robustness: A Distribution-Balanced Self-Supervised Learning Framework for Automatic Dialogue Evaluation
Peiwen Yuan (Beijing Institute of Technology), Kan Li (Beijing Institute of Technology)
Recommendation SystemTransformerContrastive LearningText
🎯 What it does: A distribution-balanced self-supervised learning framework BCR is proposed to enhance the automatic evaluation of dialogue systems;
Better Private Linear Regression Through Better Private Feature Selection
Travis Dick (Google Research), Matthew Joseph (Google Research)
OptimizationSafty and PrivacyTabular
🎯 What it does: A differential privacy feature selection algorithm based on Kendall rank correlation, DPKendall, is proposed and combined with existing plug-and-play private linear regression methods (Tukey, Boosted AdaSSP).
Better with Less: A Data-Active Perspective on Pre-Training Graph Neural Networks
Jiarong Xu (Fudan University), Yang Yang (Finvolution Group)
ClassificationData-Centric LearningGraph Neural NetworkContrastive LearningGraph
🎯 What it does: This paper proposes a data-driven graph pre-training framework called APT, which utilizes a graph selector to choose the most informative graph samples based on prediction uncertainty and graph attributes, achieving pre-training with a small amount of high-quality data.
Beyond Average Return in Markov Decision Processes
Alexandre Marthe (Ecole Normale Supérieure de Lyon), Claire Vernade (University of Tuebingen)
OptimizationReinforcement LearningSequential
🎯 What it does: This paper studies which statistical functions of rewards (such as expectation, risk measures, etc.) can be accurately solved and optimized in finite-horizon undiscounted Markov Decision Processes (MDPs), and conducts a theoretical analysis of the role and limitations of Distributed Reinforcement Learning (DistRL).
Beyond Black-Box Advice: Learning-Augmented Algorithms for MDPs with Q-Value Predictions
Tongxin Li (Chinese University of Hong Kong Shenzhen), Adam Wierman (California Institute of Technology)
Reinforcement Learning
🎯 What it does: A projection tracking strategy (PROP) is proposed, which combines untrusted machine learning Q-value suggestions with a robust baseline policy to form an adjustable robust budget;
Beyond Confidence: Reliable Models Should Also Consider Atypicality
Mert Yuksekgonul (Stanford University), Carlos Guestrin
ClassificationAnomaly DetectionLarge Language ModelImageTextBiomedical Data
🎯 What it does: This paper proposes using the atypicality of samples and categories to supplement traditional confidence, in order to more comprehensively measure and enhance the reliability and uncertainty quantification of models.
Beyond Deep Ensembles: A Large-Scale Evaluation of Bayesian Deep Learning under Distribution Shift
Florian Seligmann (Karlsruhe Institute of Technology), Gerhard Neumann (Karlsruhe Institute of Technology)
Domain AdaptationConvolutional Neural NetworkTransformerSupervised Fine-TuningImageTabularBenchmark
🎯 What it does: This paper conducts a large-scale systematic evaluation of various Bayesian Deep Learning (BDL) algorithms on real distribution shift tasks, covering single-mode posterior, deep ensemble, multi-mode ensemble (MultiX), and fine-tuning of pre-trained models, with a focus on generalization, calibration, and posterior approximation quality.
Beyond Exponential Graph: Communication-Efficient Topologies for Decentralized Learning via Finite-time Convergence
Yuki Takezawa (Kyoto University), Makoto Yamada (Kyoto University)
OptimizationFederated LearningConvolutional Neural NetworkImage
🎯 What it does: A new decentralized learning topology, BASE-(k+1) graph, is proposed, which can achieve finite-time convergence under any number of nodes and maximum degree k, while balancing fast consensus rate and low communication cost.
Beyond Geometry: Comparing the Temporal Structure of Computation in Neural Circuits with Dynamical Similarity Analysis
Mitchell Ostrow (Massachusetts Institute of Technology), Ila R Fiete
ClassificationRecognitionOptimizationRecurrent Neural NetworkReinforcement LearningTime SeriesSequential
🎯 What it does: A dynamic similarity analysis (DSA) method is proposed to compare the similarity of neural networks or neural circuits at the level of temporal dynamics.
Beyond Invariance: Test-Time Label-Shift Adaptation for Addressing "Spurious" Correlations
Qingyao Sun (Cornell University), Alexander D'Amour (Google DeepMind)
Domain AdaptationImageBenchmark
🎯 What it does: This paper proposes a method for label-shift correction during testing (TTLSA), which trains a discriminative model with meta-labels (y, z) on the source domain and estimates the meta-label prior using unlabeled samples in the target domain, thereby adaptively addressing distribution shifts caused by 'spurious' correlations.
Beyond MLE: Convex Learning for Text Generation
Chenze Shao (Chinese Academy of Sciences), Yang Feng (Chinese Academy of Sciences)
GenerationOptimizationTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Proposes a training objective based on convex functions, allowing text generation models to learn sharper probability distributions, covering autoregressive, non-autoregressive models, and large models;
Beyond Myopia: Learning from Positive and Unlabeled Data through Holistic Predictive Trends
Wang Xinrui, Songcan Chen (Nanjing University of Aeronautics and Astronautics)
ClassificationAnomaly DetectionImageTabularAlzheimer's DiseaseFinance Related
🎯 What it does: This paper proposes a positive-unlabeled learning method based on the trend of prediction scores for positive and negative samples, utilizing positive sample resampling, time point process analysis, and trend score segmentation to automatically label unlabeled data and complete binary classification.
Beyond Normal: On the Evaluation of Mutual Information Estimators
Paweł Czyż (ETH Zurich), Alexander Marx (ETH Zurich)
MultimodalityBenchmark
🎯 What it does: This paper proposes a method for constructing diverse distributions and provides a language-independent benchmarking platform for mutual information estimators, systematically evaluating the performance of commonly used estimators.
Beyond NTK with Vanilla Gradient Descent: A Mean-Field Analysis of Neural Networks with Polynomial Width, Samples, and Time
Arvind Venkat Mahankali (Stanford University), Tengyu Ma (Stanford University)
Optimization
🎯 What it does: This paper studies the learning capability of two-layer neural networks under gradient descent (projection gradient flow) without any special modifications, providing upper bounds on polynomial width, sample size, and iteration time, and proving that it can converge to a non-zero error in polynomial time;
Beyond Pretrained Features: Noisy Image Modeling Provides Adversarial Defense
Zunzhi You (University of Sydney), Chang Xu (Seoul National University)
Adversarial AttackTransformerSupervised Fine-TuningImage
🎯 What it does: Proposes a Noiseless Image Modeling (NIM) framework and implements a tunable denoising adversarial defense method De3 based on a pre-trained decoder.
Beyond probability partitions: Calibrating neural networks with semantic aware grouping
Jia-Qi Yang (Nanjing University), Le Gan (Nanjing University)
OptimizationConvolutional Neural NetworkImage
🎯 What it does: This study investigates the problem of probability calibration in deep networks, proposing a Partitioned Calibration Error (PCE) framework. It divides the input space into different subsets through a learned semantic-aware grouping function, applying temperature scaling within each subset to achieve better calibration.
Beyond Uniform Sampling: Offline Reinforcement Learning with Imbalanced Datasets
Zhang-Wei Hong (Massachusetts Institute of Technology), Pulkit Agrawal
Reinforcement Learning
🎯 What it does: This study investigates the conservativeness issue of offline reinforcement learning in the context of imbalanced datasets, proposing a resampling method that learns state-action density ratio weights, focusing only on high-reward samples, and integrating it as a plug-in module with existing offline RL algorithms.
Beyond Unimodal: Generalising Neural Processes for Multimodal Uncertainty Estimation
Myong Chol Jung (Monash University), Lan Du (Monash University)
ClassificationAnomaly DetectionContrastive LearningImageMultimodality
🎯 What it does: This paper proposes a framework for uncertainty estimation in multimodal data—Multimodal Neural Processes (MNPs), which is an extension based on the principles of Neural Processes (NP);
Bi-Level Offline Policy Optimization with Limited Exploration
Wenzhuo Zhou (University of California)
OptimizationReinforcement LearningTabularBiomedical Data
🎯 What it does: A dual-layer offline reinforcement learning optimization framework is proposed, utilizing confidence sets and uncertainty control to achieve conservative policy evaluation and optimization.
Bias in Evaluation Processes: An Optimization-Based Model
L. Elisa Celis (Yale University), Nisheeth K Vishnoi
OptimizationGraphTabular
🎯 What it does: This paper proposes an optimization model based on information entropy and risk aversion to characterize the biases that occur during the evaluation process, and derives a closed-form solution for the output distribution and its parameter relationships.
Bicriteria Approximation Algorithms for the Submodular Cover Problem
Wenjing Chen (Texas A and M University), Victoria G. Crawford (Texas A and M University)
OptimizationGraphTabular
🎯 What it does: This paper addresses the Submodular Cover (SCP) problem and proposes various approximation algorithms, along with theoretical analysis and experimental validation across multiple variants (monotonic, non-monotonic, regularized monotonic).
Bicriteria Multidimensional Mechanism Design with Side Information
Siddharth Prasad (Carnegie Mellon University), Tuomas Sandholm (Carnegie Mellon University)
OptimizationReinforcement Learning
🎯 What it does: This paper designs a multidimensional mechanism design meta-mechanism that relies on side information, capable of optimizing social welfare and revenue simultaneously without depending on prior distributions.
Bifurcations and loss jumps in RNN training
Lukas Eisenmann (Central Institute of Mental Health, Medical Faculty Mannheim, Heidelberg University), Daniel Durstewitz (Central Institute of Mental Health, Medical Faculty Mannheim, Heidelberg University)
Recurrent Neural NetworkTime Series
🎯 What it does: This study investigates the bifurcation phenomenon in training processes of ReLU-based recurrent neural networks (RNNs) and the resulting loss jumps. It proposes a search algorithm (SCYFI) that can accurately locate fixed points and periodic solutions, and uses bifurcation theory to mathematically prove gradient explosion/vanishing.
Bilevel Coreset Selection in Continual Learning: A New Formulation and Algorithm
Jie Hao (George Mason University), Mingrui Liu (George Mason University)
OptimizationConvolutional Neural NetworkImage
🎯 What it does: This paper proposes a second-order optimization framework BCSR based on softened TopK regularization, which can efficiently select representative core samples and update the model in continual learning.
BiMatting: Efficient Video Matting via Binarization
Haotong Qin (Beihang University), Fisher Yu (Dartmouth College)
SegmentationCompressionComputational EfficiencySupervised Fine-TuningVideo
🎯 What it does: Real-time video matting is implemented on edge devices, achieving significant compression through model binarization.
Binarized Neural Machine Translation
Yichi Zhang (Cornell University), Orhan Firat (Google DeepMind)
OptimizationComputational EfficiencyKnowledge DistillationTransformerText
🎯 What it does: A novel 1-bit binarization scheme (Binarized Neural Machine Translation, BMT) is implemented in the Transformer machine translation model, binarizing weights and activations while maintaining translation quality.
Binarized Spectral Compressive Imaging
Yuanhao Cai (Tsinghua University), Haoqian Wang
RestorationCompressionConvolutional Neural NetworkImage
🎯 What it does: This paper studies a hyperspectral image recovery method in spectral compressed imaging (SCI) based on Binarized Neural Networks (BNN), proposing the BiSRNet model that can be deployed on resource-constrained devices.
Binary Classification with Confidence Difference
Wei Wang (University of Tokyo), Masashi Sugiyama (University of Tokyo)
ClassificationRecommendation SystemImage
🎯 What it does: This paper proposes a weakly supervised learning framework for binary classification that utilizes confidence difference information from unlabeled sample pairs, and provides an unbiased risk estimator and risk correction method.
Binary Radiance Fields
Seungjoo Shin (POSTECH), Jaesik Park (Seoul National University)
Data SynthesisCompressionNeural Radiance FieldPoint CloudBenchmark
🎯 What it does: This paper proposes a Binary Radiance Field (BiRF), which achieves an extremely compressed representation of radiance fields by quantizing feature grid parameters to +1/-1 and combining 2D-3D mixed feature grids.
BIOT: Biosignal Transformer for Cross-data Learning in the Wild
Chaoqi Yang (University of Illinois Urbana-Champaign), Jimeng Sun (University of Illinois Urbana-Champaign)
TransformerContrastive LearningTime SeriesBiomedical DataElectrocardiogram
🎯 What it does: The BIOT model is proposed to achieve cross-data learning, capable of handling multi-channel, varying lengths, and missing values of biological signals.
BIRD: Generalizable Backdoor Detection and Removal for Deep Reinforcement Learning
Xuan Chen (Purdue University), Dawn Song (University of California Berkeley)
OptimizationReinforcement LearningSequential
🎯 What it does: The BIRD method is proposed, which can detect and remove backdoors in deep reinforcement learning policies without knowing the details of the attacks.
Birder: Communication-Efficient 1-bit Adaptive Optimizer for Practical Distributed DNN Training
Hanyang Peng (Peng Cheng Laboratory), Ge Li (Peking University)
OptimizationConvolutional Neural NetworkTransformerSupervised Fine-TuningImageText
🎯 What it does: Birder is proposed, a 1-bit adaptive optimizer that achieves extremely low communication overhead for distributed DNN training.
Birth of a Transformer: A Memory Viewpoint
Alberto Bietti (Flatiron Institute), Leon Bottou (Facebook AI Research Meta)
TransformerText
🎯 What it does: This paper constructs a synthetic dataset based on a binary model to study how Transformers learn global and contextual knowledge during training, revealing the 'induction head' mechanism formed by two layers of Transformers.
BiSLS/SPS: Auto-tune Step Sizes for Stable Bi-level Optimization
Chen Fan (University of British Columbia), Christos Thrampoulidis (University of British Columbia)
OptimizationHyperparameter SearchTabular
🎯 What it does: This paper proposes a bi-level optimization algorithm based on adaptive step sizes (BiSLS-Adam/SGD) and corresponding SLS/SPS variants (SPSB/SLSB), along with its convergence analysis.
Black-box Backdoor Defense via Zero-shot Image Purification
Yucheng Shi (University of Georgia), Ninghao Liu (University of Georgia)
Adversarial AttackDiffusion modelImage
🎯 What it does: A black-box backdoor defense framework named ZIP (Zero-shot Image Purification) is proposed, which first disrupts the trigger through linear transformations (such as blurring) and then utilizes a pre-trained diffusion model to restore semantic information, resulting in purified images.
Black-Box Differential Privacy for Interactive ML
Haim Kaplan (Tel Aviv University and Google Research), Uri Stemmer (Tel Aviv University and Google Research)
Safty and Privacy
🎯 What it does: A new interactive differential privacy framework (Challenge Differential Privacy) is proposed, along with an efficient transformation method to convert any non-private online learner into a private online learner, achieving polynomial error bounds on observable sequences.
BLIP-Diffusion: Pre-trained Subject Representation for Controllable Text-to-Image Generation and Editing
Dongxu Li (Salesforce AI Research), Steven Hoi
GenerationData SynthesisTransformerVision Language ModelDiffusion modelImageTextMultimodality
🎯 What it does: This paper proposes the BLIP-Diffusion model, which integrates the BLIP-2 visual-language encoder with a latent diffusion model to achieve controllable text-to-image generation and editing based on pre-trained topic representations.
Block Broyden's Methods for Solving Nonlinear Equations
Chengchang Liu (Chinese University of Hong Kong), John C.S. Lui (Chinese University of Hong Kong)
Optimization
🎯 What it does: Two block Broyden methods (block good method and block bad method) are proposed for solving nonlinear equations, and explicit superlinear convergence rates are provided.
Block Coordinate Plug-and-Play Methods for Blind Inverse Problems
Weijie Gan (Washington University in St. Louis), Ulugbek Kamilov
RestorationOptimizationImageBiomedical DataMagnetic Resonance Imaging
🎯 What it does: A block coordinate plug-and-play (PnP) method, BC-PnP, is proposed for the joint estimation of images and measurement operators in blind inverse problems.
Block Low-Rank Preconditioner with Shared Basis for Stochastic Optimization
Jui-Nan Yen (University of California, Los Angeles), Cho-Jui Hsieh (Google)
OptimizationTransformerAuto EncoderImage
🎯 What it does: A shared low-rank preconditioner for base blocks is proposed to accelerate deep network training and reduce storage and computational costs.
Block-Coordinate Methods and Restarting for Solving Extensive-Form Games
Darshan Chakrabarti (Columbia University), Christian Kroer (Columbia University)
OptimizationReinforcement Learning
🎯 What it does: The first cyclic coordinate update method (ECyclicPDA) was developed to solve two-player zero-sum extensive-form games, and a restart heuristic was proposed to significantly accelerate convergence.
Block-State Transformers
Jonathan Pilault (Google DeepMind), Ross Goroshin (Google DeepMind)
TransformerLarge Language ModelTextSequential
🎯 What it does: Proposes the Block-State Transformer (BST), which integrates State Space Models with Block Transformers for long sequence language modeling, achieving a fully parallelized layer structure.
Blocked Collaborative Bandits: Online Collaborative Filtering with Per-Item Budget Constraints
Soumyabrata Pal (Google Research), Prateek Jain (Google Research)
Recommendation SystemTabular
🎯 What it does: A new B-LATTICE algorithm is proposed to solve the collaborative multi-armed bandit problem with limited recommendation counts for each user per item, and a corresponding sublinear return (i.e., cumulative reward) upper bound is provided.
Blockwise Parallel Transformers for Large Context Models
Hao Liu (University of California Berkeley), Pieter Abbeel (University of California Berkeley)
Computational EfficiencyTransformerReinforcement LearningText
🎯 What it does: Proposes Blockwise Parallel Transformers (BPT), which significantly reduces the memory consumption of Transformers by employing block-level parallel computation on self-attention and feedforward networks, supporting longer input sequences;
Blurred-Dilated Method for Adversarial Attacks
Yang Deng (Sun Yat-sen University), Zibin Zheng (Sun Yat-sen University)
Adversarial AttackConvolutional Neural NetworkImage
🎯 What it does: Proposes a method called 'Blurred-Dilated (BD)' that modifies the source model by reducing downsampling, introducing BlurPool, and using dilated convolutions to generate more transferable adversarial examples.
Boosting Adversarial Transferability by Achieving Flat Local Maxima
Zhijin Ge (Xidian University), Yuanyuan Liu (Xidian University)
Adversarial AttackImage
🎯 What it does: This paper studies how to enhance the cross-model transferability of adversarial attacks, proposing the PGN attack method that penalizes the gradient norm to position adversarial samples at flat local maxima.
Boosting Learning for LDPC Codes to Improve the Error-Floor Performance
Hee-Youl Kwak (University of Ulsan), Jong-Seon No (Seoul National University)
Supervised Fine-Tuning
🎯 What it does: This paper addresses the error floor phenomenon of LDPC codes and proposes a three-stage training framework: first, using a benchmark decoder to learn the sponge effect, and then training a post-decoder with uncorrected codewords; it employs block-level training with backpropagation to alleviate gradient vanishing; and designs weight sharing using only two types of weights (satisfying/not satisfying check nodes) to significantly reduce the number of parameters.
Boosting Spectral Clustering on Incomplete Data via Kernel Correction and Affinity Learning
Fangchen Yu (Chinese University of Hong Kong), Wenye Li (Shenzhen Research Institute of Big Data)
ClassificationRecognitionOptimizationImageAudio
🎯 What it does: This paper proposes a non-interpolative framework that utilizes kernel correction and self-expressive affinity learning methods to directly construct a high-quality similarity matrix on missing data, thereby enhancing the performance of spectral clustering.
Boosting Verification of Deep Reinforcement Learning via Piece-Wise Linear Decision Neural Networks
Jiaxu Tian (East China Normal University), Min Zhang (East China Normal University)
Computational EfficiencyReinforcement Learning
🎯 What it does: This paper proposes transforming traditional deep reinforcement learning (DRL) controllers from nonlinear DNNs into directly verifiable piecewise linear decision neural networks (PLDNNs), and optimizes these linear policies through a reverse training method to achieve reachability analysis and safety verification of DRL systems.
Boosting with Tempered Exponential Measures
Richard Nock (Google Research), Manfred K Warmuth
ClassificationOptimizationTabular
🎯 What it does: This paper proposes a new Boosting algorithm called t-ADABOOST, which generalizes the traditional AdaBoost by maintaining a temperature-controlled index measure (TEM) on the samples and minimizing the corresponding temperature-controlled relative entropy.
Bootstrapped Training of Score-Conditioned Generator for Offline Design of Biological Sequences
Minsu Kim (Korea Advanced Institute of Science and Technology), Jinkyoo Park (Korea Advanced Institute of Science and Technology)
GenerationOptimizationDrug DiscoveryRecurrent Neural NetworkReinforcement LearningScore-based ModelSequentialBiomedical Data
🎯 What it does: A score-conditioned generator (BOOTGEN) based on bootstrapped training is proposed for offline optimization of biological sequences (DNA, RNA, proteins) to maximize a given black-box scoring function.
Bootstrapping Vision-Language Learning with Decoupled Language Pre-training
Yiren Jian (Dartmouth), Soroush Vosoughi (Dartmouth)
GenerationRetrievalTransformerLarge Language ModelPrompt EngineeringVision Language ModelContrastive LearningImageTextMultimodality
🎯 What it does: A new training framework is proposed: first, train the Prompt-Former (P-Former) using single-modal text data to generate ideal soft prompts, and then align visual features with these prompts during VL pre-training, thereby improving visual-language pre-training based on frozen large language models.
Bottleneck Structure in Learned Features: Low-Dimension vs Regularity Tradeoff
Arthur Jacot (New York University)
OptimizationRepresentation LearningTabular
🎯 What it does: This paper performs a Taylor expansion of the representation cost of deep regularizable neural networks and proposes three metrics for quantifying network learning characteristics: low-rank bias R(0), regularization term R(1), and second-order regularization R(2). Based on this theory, it proves that as the depth approaches infinity, bottleneck structures will emerge within the network, meaning that hidden layers mostly occupy a low-dimensional subspace implied by R(0). Furthermore, it shows that when the NTK remains on the order of O(L), the hidden representations maintain low dimensionality and finite norms in most layers. Numerical experiments validate that the network can capture summary statistics of the problem in a two-dimensional hidden space by learning rotational symmetry.
Bounce: Reliable High-Dimensional Bayesian Optimization for Combinatorial and Mixed Spaces
Leonard Papenmeier (Lund University), Matthias Poloczek (Amazon)
OptimizationTabular
🎯 What it does: This study investigates the reliability issues of high-dimensional Bayesian optimization in combinatorial and mixed spaces, and proposes the Bounce algorithm, which can adaptively bin, nest embeddings, and dynamically manage trust regions.
Boundary Guided Learning-Free Semantic Control with Diffusion Models
Ye Zhu (Princeton University), Yan Yan (Wuhan University)
GenerationData SynthesisDomain AdaptationDiffusion modelImage
🎯 What it does: Theoretical and empirical analysis of the high-dimensional latent space of pre-trained unsupervised diffusion models is conducted, proposing the BoundaryDiffusion method for learning free single-step semantic control, which directly utilizes frozen DDMs to perform image editing, text-driven editing, and unconditional semantic control.
Bounded rationality in structured density estimation
Tianyuan Teng (Peking University), Hang Zhang (Peking University)
Multimodality
🎯 What it does: Designed and implemented structured density estimation experiments to record the internal structure of high-dimensional probability distributions in human continuous experiences, and modeled behavioral data using the maximum likelihood density estimation framework (DEF).