NeurIPS 2023 Papers — Page 5
Conference on Neural Information Processing Systems · 3218 papers
Bounding the Invertibility of Privacy-preserving Instance Encoding using Fisher Information
Kiwan Maeng (Penn State University), G. Edward Suh (Meta)
Recommendation SystemSafty and PrivacyConvolutional Neural NetworkTransformerScore-based ModelImageText
🎯 What it does: This paper studies a measure of instance encoding reversibility based on Fisher Information Leakage (dFIL) and uses this measure to design a privacy-enhanced split inference and encoding training scheme.
Bounding training data reconstruction in DP-SGD
Jamie Hayes (Google DeepMind), Saeed Mahloujifar (Meta AI)
Safty and PrivacyAdversarial AttackConvolutional Neural NetworkSupervised Fine-TuningImage
🎯 What it does: This paper studies the training data reconstruction attack under DP-SGD training and provides a theoretical upper bound and an empirical lower bound on the attack success rate.
BQ-NCO: Bisimulation Quotienting for Efficient Neural Combinatorial Optimization
Darko Drakulic (Naver Labs Europe), Jean-Marc Andreoli (Naver Labs Europe)
OptimizationTransformerReinforcement LearningGraph
🎯 What it does: This paper proposes a general framework that formalizes any combinatorial optimization problem (COP) as a Markov Decision Process (MDP) and further compresses the state space through Bisimulation Quotienting (BQ), enhancing solving efficiency and generalization capability.
Brain Diffusion for Visual Exploration: Cortical Discovery using Large Scale Generative Models
Andrew Luo, Michael J. Tarr (Carnegie Mellon University)
GenerationData SynthesisDiffusion modelImageMagnetic Resonance Imaging
🎯 What it does: Using brain signals to guide diffusion models to synthesize images that can activate specific regions of the visual cortex, thereby exploring the fine-grained functional organization of higher-order visual cortex.
Brain Dissection: fMRI-trained Networks Reveal Spatial Selectivity in the Processing of Natural Images
Gabriel Herbert Sarch, Leila Wehbe (Carnegie Mellon University)
Explainability and InterpretabilityConvolutional Neural NetworkImageMagnetic Resonance Imaging
🎯 What it does: Trained an fMRI response-optimized convolutional network targeting different subregions of the visual cortex, and conducted an interpretability analysis of image feature selectivity for each voxel through network dissection, ultimately mapping spatial feature preferences in 3D spatial attributes such as depth, surface normals, curvature, shadows, as well as object relationships, attributes, and categories.
Brain encoding models based on multimodal transformers can transfer across language and vision
Jerry Tang (University of Texas at Austin), Alexander Huth
TransformerVision Language ModelVideoTextMultimodalityMagnetic Resonance Imaging
🎯 What it does: Using the BridgeTower multimodal Transformer to extract features from stories and movies, and training voxel-wise encoding models on subjects' fMRI data to test the cross-modal transfer performance of these models between language and visual stimuli.
Brain-like Flexible Visual Inference by Harnessing Feedback Feedforward Alignment
Tahereh Toosi (Columbia University), Elias Issa
ClassificationRestorationGenerationConvolutional Neural NetworkAuto EncoderImage
🎯 What it does: The Feedback-Feedforward Alignment (FFA) algorithm is proposed, which jointly trains feedforward and feedback networks to achieve co-optimization of classification and reconstruction within the same architecture, and implements various visual reasoning functions such as denoising, completion, hallucination, and imagination through closed-loop inference.
Brant: Foundation Model for Intracranial Neural Signal
Daoze Zhang (Zhejiang University), Yafeng Li (Nuozhu Technology Co., Ltd.)
Anomaly DetectionRepresentation LearningTransformerSupervised Fine-TuningTime SeriesBiomedical Data
🎯 What it does: A foundational model for intracranial recordings, Brant, is proposed, leveraging large-scale self-supervised pre-training to learn powerful representations of brain signals.
Breadcrumbs to the Goal: Goal-Conditioned Exploration from Human-in-the-Loop Feedback
Marcel Torne Villasevil (Massachusetts Institute of Technology), Abhishek Gupta (University of Washington)
Robotic IntelligenceReinforcement Learning from Human FeedbackReinforcement Learning
🎯 What it does: The HuGE method is proposed, which utilizes low-quality, asynchronous, sparse non-expert human comparative feedback to guide exploration, and independently learns goal-conditioned policies through self-supervised hindsight relabeling.
Break It Down: Evidence for Structural Compositionality in Neural Networks
Michael A. Lepori (Brown University), Ellie Pavlick (Brown University)
TransformerImageText
🎯 What it does: This paper explores whether neural networks can automatically decompose complex tasks into subprograms without explicit symbolic mechanisms, and implement these subprograms within sub-networks, referred to as structural compositionality.
Breaking the Communication-Privacy-Accuracy Tradeoff with $f$-Differential Privacy
Richeng Jin (Zhejiang University), Huaiyu Dai (North Carolina State University)
Federated LearningSafty and PrivacyTabular
🎯 What it does: The paper studies the privacy guarantees of discrete mechanisms with a finite output space under the f-DP (f-differential privacy) framework, and based on this, proposes a sparse ternary compressor that leverages the privacy amplification effect brought by sparsification, breaking the three-way trade-off between communication, privacy, and accuracy.
Bridging Discrete and Backpropagation: Straight-Through and Beyond
Liyuan Liu (Microsoft Research), Jianfeng Gao (Microsoft Research)
OptimizationReinforcement LearningImageOrdinary Differential Equation
🎯 What it does: Theoretical research on gradient estimation methods for discrete latent variables is conducted, proving that Straight-Through (ST) is a first-order Euler approximation. Based on this, a new second-order accurate estimator, ReinMax, is proposed and its effectiveness is validated across various tasks.
Bridging RL Theory and Practice with the Effective Horizon
Cassidy Laidlaw (University of California), Anca Dragan (University of California)
Reinforcement LearningTabular
🎯 What it does: This study constructs the BRIDGE dataset (155 deterministic MDPs) and proposes a new complexity measure—effective horizon—to explain and predict the sample complexity of deep reinforcement learning algorithms (such as PPO and DQN); it also designs a GREedy Over Random Policy (GORP) algorithm based on stochastic rollout as a theoretical tool.
Bridging the Domain Gap: Self-Supervised 3D Scene Understanding with Foundation Models
Zhimin Chen (Clemson University), Bing Li (Clemson University)
Object DetectionSegmentationKnowledge DistillationRepresentation LearningTransformerAuto EncoderContrastive LearningImagePoint Cloud
🎯 What it does: By utilizing various foundational models (such as SAM, Grounding DINO, BLIP, CLIP, etc.) to generate semantic masks, image captions, and textual information, a self-supervised pre-training of 3D point clouds is conducted, forming the Bridge3D framework.
Bringing regularized optimal transport to lightspeed: a splitting method adapted for GPUs
Jacob Lindbäck (KTH), Mikael Johansson (KTH)
Domain AdaptationOptimizationImage
🎯 What it does: A GPU-accelerated regularized optimal transport solver based on Douglas-Rachford splitting (RDROT) is proposed, capable of handling sparse or structured regularization (such as quadratic, group-Lasso, etc.) and achieving parallelization on the GPU.
Bucks for Buckets (B4B): Active Defenses Against Stealing Encoders
Jan Dubiński (Warsaw University of Technology), Adam Dziedzic (CISPA Helmholtz Center for Information Security)
Safty and PrivacyRepresentation LearningAdversarial AttackContrastive LearningImage
🎯 What it does: A proactive defense framework named B4B is proposed, which can real-time prevent model theft attacks on the public Encoder API, with almost no impact on the representation quality for legitimate users.
Budgeting Counterfactual for Offline RL
Yao Liu (Amazon Web Services), Rasool Fakoor (Amazon Web Services)
Reinforcement LearningTabularBenchmark
🎯 What it does: An algorithm for controlling the number of 'counterfactual' decisions through budget management in offline reinforcement learning (BCOL) is proposed. It enhances policy performance by allocating key decisions from a limited set of offline samples using dynamic programming while maintaining a close approximation of the behavior policy.
Bypass Exponential Time Preprocessing: Fast Neural Network Training via Weight-Data Correlation Preprocessing
Josh Alman (Columbia University), Danyang Zhuo (Duke University)
Optimization
🎯 What it does: A preprocessing method based on a weight-data correlation tree (Correlation Tree) is proposed, significantly reducing the iteration time for training large-scale two-layer ReLU networks.
Bypassing spike sorting: Density-based decoding using spike localization from dense multielectrode probes
Yizi Zhang (Columbia University), Liam Paninski (Columbia University)
Reinforcement LearningGaussian SplattingTime Series
🎯 What it does: A neural decoding method that does not rely on spike sorting is proposed, which directly utilizes spike localization and waveform features extracted from high-density multi-electrode probes for behavioral decoding.
Bypassing the Simulator: Near-Optimal Adversarial Linear Contextual Bandits
Haolin Liu (University of Virginia), Julian Zimmert (Google Research)
OptimizationReinforcement Learning
🎯 What it does: An algorithm is proposed for the adversarial linear contextual bandit problem without a simulator, achieving a near-optimal expected regret of ˜O(d^2√T);
Byzantine-Tolerant Methods for Distributed Variational Inequalities
Nazarii Tupitsa (Mohammed Bin Zayed University of Artificial Intelligence), Eduard Gorbunov (Mohammed Bin Zayed University of Artificial Intelligence)
OptimizationFederated LearningImage
🎯 What it does: This paper proposes several Byzantine fault-tolerant methods for distributed variational inequality (VI) problems and provides theoretical convergence analysis under different assumptions.
C-Disentanglement: Discovering Causally-Independent Generative Factors under an Inductive Bias of Confounder
Xiaoyu Liu (University of Maryland), Furong Huang (University of Maryland)
GenerationData SynthesisDomain AdaptationAuto EncoderImage
🎯 What it does: This paper proposes the C-Disentanglement framework and implements the cdVAE model, using weakly supervised labels as an inductive bias to discover causally independent generative factors in the presence of confounding factors.
CADet: Fully Self-Supervised Out-Of-Distribution Detection With Contrastive Learning
Charles Guille-Escuret (ServiceNow Research), Joao Monteiro (ServiceNow Research)
Anomaly DetectionConvolutional Neural NetworkContrastive LearningImage
🎯 What it does: This paper presents CADet, a fully self-supervised contrastive learning method for single-sample OOD and adversarial attack detection.
Cal-DETR: Calibrated Detection Transformer
Muhammad Akhtar Munir (Information Technology University), Fahad Khan
Object DetectionDomain AdaptationTransformerImage
🎯 What it does: This paper proposes a training-time calibration method for Transformer-based object detectors called Cal-DETR. It estimates uncertainty by calculating variance between decoder layers, modulates the class logits using this uncertainty, and performs mixup-style blending in the logit space as a regularization technique to enhance the confidence calibration and detection performance of the detector.
Cal-QL: Calibrated Offline RL Pre-Training for Efficient Online Fine-Tuning
Mitsuhiko Nakamoto (University of California Berkeley), Sergey Levine (University of California Berkeley)
Reinforcement LearningImageTabular
🎯 What it does: This paper proposes an algorithm for calibrating the value function in offline reinforcement learning (Cal-QL) to achieve more efficient online fine-tuning.
Calibrate and Boost Logical Expressiveness of GNN Over Multi-Relational and Temporal Graphs
Dingmin Wang (University of Oxford), Yeyuan Chen (University of Michigan)
ClassificationGraph Neural NetworkGraphTime Series
🎯 What it does: This study investigates the logical expressiveness of R2-GNN on multi-relational graphs and temporal graphs, proving that its original form cannot cover all FOC2 formulas, and proposes a linear-time graph transformation F to enhance expressiveness; it also establishes a hierarchy of expressiveness for multi-relational and temporal graphs, conducting node classification experiments on synthetic and real data.
Calibrated Stackelberg Games: Learning Optimal Commitments Against Calibrated Agents
Nika Haghtalab (University of California), Kunhe Yang (University of California)
OptimizationReinforcement Learning
🎯 What it does: This paper proposes and studies the Calibrated Stackelberg Game (CSG), where agents make optimal responses based solely on adaptively calibrated predictions.
Calibrating “Cheap Signals” in Peer Review without a Prior
Yuxuan Lu (Peking University), Yuqing Kong (Peking University)
🎯 What it does: A one-time, non-prior 'calibration' method is proposed, which allows reviewers to predict others' scores to denoise noisy peer reviews and achieve more robust paper rankings.
Calibrating Neural Simulation-Based Inference with Differentiable Coverage Probability
Maciej Falkiewicz (University of Geneva), Alexandros Kalousis (University of Geneva)
Neural Radiance FieldGenerative Adversarial NetworkBenchmark
🎯 What it does: This paper proposes a differentiable coverage probability regularization method for calibrating the posterior distribution of neural networks in simulation-based inference (SBI) to avoid overconfidence.
Calibration by Distribution Matching: Trainable Kernel Calibration Metrics
Charles Thomas Marx (Stanford University), Stefano Ermon (Stanford University)
OptimizationTabularAgriculture Related
🎯 What it does: A trainable calibration metric based on Maximum Mean Discrepancy (MMD) is proposed, unifying various calibration forms and using it as a regularization term to train probability prediction models.
CAMEL: Communicative Agents for "Mind" Exploration of Large Language Model Society
Guohao Li, Bernard Ghanem (King Abdullah University of Science and Technology)
TransformerLarge Language ModelReinforcement LearningPrompt EngineeringTextChain-of-Thought
🎯 What it does: A role-playing (Role-Playing) framework is proposed, utilizing autonomous communication agents to complete tasks with minimal human input, and generating large-scale dialogue data for studying multi-agent collaboration and LLM capabilities through this framework.
CamoPatch: An Evolutionary Strategy for Generating Camoflauged Adversarial Patches
Phoenix Neale Williams (University of Exeter), Ke Li (University of Exeter)
OptimizationAdversarial AttackConvolutional Neural NetworkTransformerImage
🎯 What it does: A method called 'CamoPatch' based on evolutionary strategies and simulated annealing is proposed, which constructs visually difficult-to-detect adversarial patches using translucent RGB circles;
CaMP: Causal Multi-policy Planning for Interactive Navigation in Multi-room Scenes
Xiaohan Wang (Xi'an Jiaotong University), Shuqiang Jiang (Chinese Academy of Sciences)
Robotic IntelligenceReinforcement LearningAgentic AITabular
🎯 What it does: This paper proposes the CaMP (Causal Multi-Policy Planning) framework, which combines multi-policy and causal inference to address the interactive navigation problem in multi-room scenarios.
Can Language Models Solve Graph Problems in Natural Language?
Heng Wang (Xi'an Jiaotong University), Yulia Tsvetkov (University of Washington)
Graph Neural NetworkLarge Language ModelPrompt EngineeringTextGraphBenchmarkChain-of-Thought
🎯 What it does: The NLGraph benchmark is proposed to evaluate the reasoning ability of large language models on graph problems described in natural language, and two instruction-based prompting methods, Build-a-Graph and Algorithmic Prompting, are introduced for graph reasoning.
Can Language Models Teach? Teacher Explanations Improve Student Performance via Personalization
Swarnadeep Saha (University of North Carolina), Mohit Bansal (University of North Carolina)
TransformerLarge Language ModelPrompt EngineeringTextChain-of-Thought
🎯 What it does: This paper studies how to enable a large language model (Teacher) to teach a weaker language model (Student) to complete reasoning tasks through natural language explanations, exploring when, how, and whether it can improve student performance, and verifying that misleading teachers may lead to performance degradation.
Can Pre-Trained Text-to-Image Models Generate Visual Goals for Reinforcement Learning?
Jialu Gao (Tsinghua University), Huazhe Xu (Tsinghua University)
Robotic IntelligenceReinforcement LearningDiffusion modelImage
🎯 What it does: This paper proposes a zero-shot visual target generation and reinforcement learning method called LfVoid, which utilizes a pre-trained text-to-image diffusion model and image editing techniques to automatically generate target images based on natural language instructions, driving robots to complete manipulation tasks.
Can semi-supervised learning use all the data effectively? A lower bound perspective
Alexandru Tifrea (ETH Zurich), Fanny Yang (ETH Zurich)
ClassificationOptimizationTabular
🎯 What it does: This paper studies the theoretical limits of semi-supervised learning under the 2-Gaussian Mixture Model (2-GMM) on a two-dimensional symmetric sphere. It proves that under any signal-to-noise ratio and sample size, semi-supervised algorithms cannot simultaneously outperform the error rates of optimal supervised learning and optimal unsupervised learning, and demonstrates in experiments that weighted ensemble or self-training can surpass the optimal SSL-S algorithm.
Can You Rely on Your Model Evaluation? Improving Model Evaluation with Synthetic Test Data
Boris van Breugel (University of Cambridge), Mihaela van der Schaar (University of Cambridge)
Data SynthesisDomain AdaptationAnomaly DetectionOptimizationGenerative Adversarial NetworkTabularSequentialBiomedical DataFinance Related
🎯 What it does: Proposes the 3S Testing framework, which evaluates the performance of models under small subgroups and distribution shifts by generating synthetic test sets.
Canonical normalizing flows for manifold learning
Kyriakos Flouris (ETH Zürich), Ender Konukoglu (ETH Zürich)
GenerationRepresentation LearningFlow-based ModelRectified FlowImageTabular
🎯 What it does: We propose Canonical Manifold Flow (CMF), which incorporates L1 regularization on the off-diagonal elements of the metric tensor in flow models to enforce the learning of sparse and approximately orthogonal eigenbases, thereby achieving more compact latent representations.
CAP: Correlation-Aware Pruning for Highly-Accurate Sparse Vision Models
Denis Kuznedelev (Skolkovo Institute of Science and Technology), Dan Alistarh (IST Austria)
ClassificationOptimizationConvolutional Neural NetworkTransformerSupervised Fine-TuningImage
🎯 What it does: A weight correlation-based unstructured pruning method called CAP is proposed for high-precision vision models (such as ViT, ConvNext, etc.), along with an efficient fine-tuning process.
Cappy: Outperforming and Boosting Large Multi-Task LMs with a Small Scorer
Bowen Tan (Carnegie Mellon University), Jindong Chen (Google Research)
GenerationOptimizationComputational EfficiencyTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: A lightweight pre-trained scorer called Cappy is proposed to enhance the performance and efficiency of multi-task LLMs.
CAPro: Webly Supervised Learning with Cross-modality Aligned Prototypes
Yulei Qin (Tencent), Rongrong Ji (Xiamen University)
Domain AdaptationRepresentation LearningTransformerLarge Language ModelContrastive LearningImageTextMultimodality
🎯 What it does: This paper proposes the CAPro framework, which utilizes cross-modal (image and text) aligned prototypes for unsupervised visual representation learning of web data, and enhances model robustness through noise removal and collective bootstrapping.
CARE: Modeling Interacting Dynamics Under Temporal Environmental Variation
Xiao Luo (University of California), Yizhou Sun (University of California)
Graph Neural NetworkGraphTime SeriesOrdinary Differential Equation
🎯 What it does: A Context-attended Graph ODE (CARE) model is proposed to model and predict the dynamics of interacting systems in time-varying environments.
Cascading Bandits: Optimizing Recommendation Frequency in Delayed Feedback Environments
Dairui Wang (Tsinghua University), Wei Qi (Tsinghua University)
Recommendation SystemOptimizationReinforcement LearningTabular
🎯 What it does: A cascade multi-armed bandit model with delayed feedback, variable rewards, and recommendation frequency control is proposed, along with offline polynomial time optimal sequences, online UCB, and contextual UCB algorithms.
Cascading Contextual Assortment Bandits
Hyunjun Choi, Min-hwan Oh (Seoul National University)
Recommendation SystemOptimizationReinforcement LearningTabular
🎯 What it does: A new combination band model, called the Cascaded Contextual Combination Band, is proposed as an extension of the existing cascaded band and combination band, expanding its applicability in practice.
CAST: Cross-Attention in Space and Time for Video Action Recognition
Dongho Lee (Kyung Hee University), Jinwoo Choi (Kyung Hee University)
RecognitionTransformerVideo
🎯 What it does: This paper proposes a two-stream video action recognition framework named CAST (Cross-Attention in Space and Time), which utilizes frozen spatial experts (CLIP) and temporal experts (VideoMAE) to exchange information through cross-attention in a bottleneck adapter, achieving a balanced spatiotemporal representation of videos and making collaborative predictions.
CAT-Walk: Inductive Hypergraph Learning via Set Walks
Ali Behrouz (University of British Columbia), Margo Seltzer (University of British Columbia)
Graph Neural NetworkGraphBenchmark
🎯 What it does: This paper studies a model called CA-Walk that can perform inductive learning in time-varying higher-order graphs by extracting temporal higher-order causal patterns through set-based random walks.
Category-Extensible Out-of-Distribution Detection via Hierarchical Context Descriptions
Kai Liu (Zhejiang University), Jieping Ye (Alibaba Cloud)
ClassificationAnomaly DetectionTransformerPrompt EngineeringVision Language ModelImage
🎯 What it does: In the framework of visual language models, a perceptual and spurious hierarchical context is proposed to learn the precise boundaries of each category, thereby achieving detection and classification extension for samples from unknown distributions.
Causal Component Analysis
Wendong Liang (Max Planck Institute for Intelligent Systems), Bernhard Schölkopf (Max Planck Institute for Intelligent Systems)
Flow-based Model
🎯 What it does: This paper studies the Causal Component Analysis (CauCA) framework, exploring the identifiability of latent variables and mixed functions using multiple sets of intervention data under the premise of a known causal graph.
Causal Context Connects Counterfactual Fairness to Robust Prediction and Group Fairness
Jacy Reese Anthis (University of Chicago), Victor Veitch (University of Chicago)
Tabular
🎯 What it does: This paper establishes a bridge between counterfactual fairness and robust prediction, group fairness through causal context, proposing that under specific causal structures, counterfactual fair predictors can achieve the best accuracy for an unbiased target distribution while corresponding to different group fairness metrics.
Causal de Finetti: On the Identification of Invariant Causal Structure in Exchangeable Data
Siyuan Guo (University of Cambridge), Ferenc Huszár (University of Cambridge)
Tabular
🎯 What it does: A method for causal structure identification based on exchangeable data is proposed, and the statistical validity of the ICM (Independent Causal Mechanism) assumption is proven through the Causal de Finetti theorem.
Causal discovery from observational and interventional data across multiple environments
Adam Li (Columbia University), Elias Bareinboim (Columbia University)
Biomedical Data
🎯 What it does: This paper proposes a method for learning causal structures by simultaneously utilizing observational and interventional data in a multi-domain environment, defining concepts such as S-Markov properties and S-PAG, and providing an implementable S-FCI algorithm.
Causal Discovery from Subsampled Time Series with Proxy Variables
Mingzhou Liu (Peking University), Yizhou Wang (Peking University)
Time SeriesBiomedical DataMagnetic Resonance ImagingAlzheimer's Disease
🎯 What it does: A non-parametric causal discovery algorithm based on proxy variables is proposed, capable of fully identifying causal structures under subsampled time series (where the observation frequency is lower than the causal influence frequency).
Causal Discovery in Semi-Stationary Time Series
Shanyun Gao (Purdue University), Murat Kocaoglu (Purdue University)
Time Series
🎯 What it does: A non-parametric constrained causal discovery algorithm PCMCI Ω is proposed for semi-stationary time series data, capable of identifying causal graphs under periodically changing causal mechanisms.
Causal Effect Identification in Uncertain Causal Networks
Sina Akbari (École Polytechnique Fédérale de Lausanne), Negar Kiyavash (École Polytechnique Fédérale de Lausanne)
OptimizationGraph Neural NetworkGraph
🎯 What it does: This paper studies the problem of causal effect identification in probabilistic ADMGs with edge uncertainty, proposing two optimization problems to find the most likely/credible identifiable subgraphs.
Causal Effect Regularization: Automated Detection and Removal of Spurious Correlations
Abhinav Kumar (Microsoft Research), Amit Sharma (Microsoft Research)
Text
🎯 What it does: A framework called AutoACER is proposed to automatically detect and eliminate spurious correlations between causal attributes and task labels in machine learning models.
Causal Fairness for Outcome Control
Drago Plecko (Columbia University), Elias Bareinboim (Columbia University)
OptimizationBiomedical DataElectronic Health Records
🎯 What it does: A causal fairness framework for the task of 'Outcome Control' (i.e., optimizing outcomes through decision control) is proposed, which includes three types of algorithms: ① an optimal decision algorithm that ensures Benefit Fairness (BF); ② a method to identify the sources of benefit disparities through causal decomposition; ③ causal and utilitarian methods (Causal Benefit Fairness, CBF) to eliminate discrimination along specified unfair causal paths.
Causal Imitability Under Context-Specific Independence Relations
Fateme Jamshidi (École Polytechnique Fédérale de Lausanne), Negar Kiyavash (École Polytechnique Fédérale de Lausanne)
Graph
🎯 What it does: This paper studies the problem of causal imitation learning considering context-specific independence (CSI) relationships, providing criteria and algorithms for imitatability.
Causal Interpretation of Self-Attention in Pre-Trained Transformers
Raanan Yehezkel Rohekar, Shami Nisimov (Intel Labs)
Recommendation SystemExplainability and InterpretabilityTransformerText
🎯 What it does: This paper views the self-attention mechanism of the Transformer as the total effect matrix of a linear Gaussian structural causal model (SCM), proposing an ABCD method based on the attention matrix to achieve zero-shot causal structure learning for a single input sequence, and further developing the CLEANN algorithm to provide causal explanations for model predictions from the learned causal graph.
Causal normalizing flows: from theory to practice
Adrián Javaloy (Saarland University), Isabel Valera (Max Planck Institute for Software Systems)
ClassificationOptimizationFlow-based ModelTabularFinance Related
🎯 What it does: This study proposes a causal inference method based on normalizing flows, introducing causal normalizing flows to achieve the identification of causal structures and interventions/counterfactual predictions.
Cause-Effect Inference in Location-Scale Noise Models: Maximum Likelihood vs. Independence Testing
Xiangyu Sun (Simon Fraser University), Oliver Schulte (Simon Fraser University)
Flow-based ModelTabular
🎯 What it does: This study investigates the robustness of the maximum likelihood (ML) method and the independence test (IT) method in causal inference under the location-scale noise model (LSNMs). It proposes an IT method based on affine flow and validates its superior performance on various synthetic and real datasets under noise distribution misjudgment and conditional variance misguidance.
Causes and Effects of Unanticipated Numerical Deviations in Neural Network Inference Frameworks
Alexander Schlögl (Universität Innsbruck), Rainer Böhme (Universität Innsbruck)
Computational EfficiencyConvolutional Neural NetworkImage
🎯 What it does: This study investigates the numerical discrepancies in inference results when using the same trained CNN model and identical inputs across multiple platforms (75 CPUs and GPUs) and their causes;
CBD: A Certified Backdoor Detector Based on Local Dominant Probability
Zhen Xiang (University of Illinois Urbana-Champaign), Bo Li (University of Illinois Urbana-Champaign)
ClassificationAnomaly DetectionImage
🎯 What it does: This paper proposes a certification-based backdoor detector (CBD) based on adjustable conformal prediction, which determines whether the model has been attacked by a backdoor using the statistic 'Local Dominant Probability' (LDP), and provides computable detection confidence and false positive upper bounds.
CEIL: Generalized Contextual Imitation Learning
Jinxin Liu (Westlake University), Huazhe Xu (Tsinghua University)
Robotic IntelligenceReinforcement LearningSequential
🎯 What it does: A general context imitation learning framework CEIL is proposed, which can approximate expert behavior in various imitation learning scenarios.
CELLE-2: Translating Proteins to Pictures and Back with a Bidirectional Text-to-Image Transformer
Emaad Khwaja (University of California Berkeley), Bo Huang (University of California San Francisco)
GenerationData SynthesisTransformerLarge Language ModelSupervised Fine-TuningImageTextMultimodality
🎯 What it does: The CELL-E 2 model is proposed, a bidirectional non-autoregressive Transformer that can generate cellular localization images from protein amino acid sequences and predict sequences from images.
Censored Sampling of Diffusion Models Using 3 Minutes of Human Feedback
TaeHo Yoon (Seoul National University), Ernest K. Ryu (Seoul National University)
GenerationData SynthesisReinforcement Learning from Human FeedbackDiffusion modelImage
🎯 What it does: Conducting review sampling on pre-trained diffusion models, utilizing minimal human feedback to train a reward model for suppressing image generation.
Certifiably Robust Graph Contrastive Learning
Minhua Lin (Pennsylvania State University), Suhang Wang (Pennsylvania State University)
ClassificationRepresentation LearningAdversarial AttackGraph Neural NetworkContrastive LearningGraph
🎯 What it does: This paper proposes the first framework for achieving provable robustness in Graph Contrastive Learning (GCL) - Randomized Edgedrop Smoothing (RES), which obtains provable robustness against graph structure attacks by randomly dropping edges during training and inference.
Certification of Distributional Individual Fairness
Matthew Robert Wicker (Alan Turing Institute), Adrian Weller (University of Cambridge)
OptimizationTabular
🎯 What it does: This study investigates the Distributed Individual Fairness (DIF) of neural networks, proposing computable convex approximations and upper/lower bounds, along with training regularization.
Certified Minimax Unlearning with Generalization Rates and Deletion Capacity
Jiaqi Liu (Zhejiang University), Kui Ren (Zhejiang University)
Optimization
🎯 What it does: A (ε,δ)-certified machine forgetting algorithm for min-max models is proposed, along with corresponding theoretical bounds on generalization error and deletion capacity.
Certified Robustness via Dynamic Margin Maximization and Improved Lipschitz Regularization
Mahyar Fazlyab (Johns Hopkins University), Rama Chellappa (Johns Hopkins University)
ClassificationOptimizationAdversarial AttackConvolutional Neural NetworkSupervised Fine-TuningImage
🎯 What it does: This paper proposes a differentiable regularization method to explicitly maximize the safety margin of deep classifiers in the input space, thereby enhancing robustness against adversarial perturbations.
Chameleon: Plug-and-Play Compositional Reasoning with Large Language Models
Pan Lu (University of California), Jianfeng Gao (Microsoft Research)
TransformerLarge Language ModelMultimodalityTabular
🎯 What it does: The Chameleon framework is proposed, utilizing large language models as natural language planners to automatically synthesize and sequentially invoke various external tools (LLMs, visual models, web searches, Python programs, rule modules) to achieve pluggable composite reasoning.
Chanakya: Learning Runtime Decisions for Adaptive Real-Time Perception
Anurag Ghosh (Carnegie Mellon University), Tanuja Ganu (Microsoft Research India)
Object DetectionAutonomous DrivingReinforcement LearningImageVideo
🎯 What it does: This work proposes Chanakya, a learning-based execution framework that automatically determines runtime parameters such as input resolution, model, and inference step length in a real-time perception pipeline, maximizing detection accuracy and minimizing latency while maintaining real-time performance (streaming perception paradigm).
Change point detection and inference in multivariate non-parametric models under mixing conditions
Carlos Misael Madrid Padilla (University of Notre Dame), Yi Yu (University of Warwick)
Anomaly DetectionTime Series
🎯 What it does: This paper proposes a multi-point change point detection and inference method for multivariate short-range correlated non-parametric time series.
Characteristic Circuits
Zhongjie Yu (TU Darmstadt), Kristian Kersting (TU Darmstadt)
Tabular
🎯 What it does: This paper proposes and implements a new probabilistic circuit model—Characteristic Circuits (CCs)—to unify the modeling of the joint distribution of discrete and continuous heterogeneous data in the spectral domain (characteristic function) and provides learnable structures and parameters.
Characterization and Learning of Causal Graphs with Small Conditioning Sets
Murat Kocaoglu (Purdue University)
GraphTabular
🎯 What it does: In the case of limited samples, this study investigates the use of conditional independence tests that only include condition sets of size no greater than k to learn causal graphs, and proposes k-Markov equivalence, k-closure graphs, and the corresponding k-PC learning algorithm.
Characterization of Overfitting in Robust Multiclass Classification
Jingyuan Xu (Wuhan University), Weiwei Liu (Wuhan University)
ClassificationAdversarial AttackTabular
🎯 What it does: This paper studies the degree of overfitting of adaptive algorithms in adversarial settings for multi-class classification problems, providing upper and lower bounds for robust overfitting bias.
Characterizing Graph Datasets for Node Classification: Homophily-Heterophily Dichotomy and Beyond
Oleg Platonov (Higher School of Economics), Liudmila Prokhorenkova (Yandex Research)
ClassificationGraph Neural NetworkGraph
🎯 What it does: This paper studies how to quantify the homogeneity and informativeness of graph datasets and proposes improved metrics.
Characterizing Out-of-Distribution Error via Optimal Transport
Yuzhe Lu (Carnegie Mellon University), Katia P. Sycara
Domain AdaptationOptimizationImage
🎯 What it does: This paper proposes a label-free out-of-distribution (OOD) performance prediction method based on optimal transport, called COT, along with its threshold variant COTT. It provides a conservative and more accurate estimate of the model's error on external distributions under the assumption of unchanged label distribution.
Characterizing the Impacts of Semi-supervised Learning for Weak Supervision
Jeffrey Li (University of Washington), Alexander Ratner (Snorkel AI)
ClassificationTransformerTextTabular
🎯 What it does: This study investigates the role of semi-supervised learning (SSL) in programmatic weak supervision (WS) and constructs a three-axis design space to systematically evaluate the effects of thresholding, SSL techniques, and re-labeling.
Characterizing the Optimal $0-1$ Loss for Multi-class Classification with a Test-time Attacker
Sihui Dai (Princeton University), Prateek Mittal (Princeton University)
ClassificationAdversarial AttackImage
🎯 What it does: This paper constructs a conflict hypergraph and solves the corresponding linear programming to provide the optimal lower bound of 0-1 loss under the presence of an attacker during multi-class testing. It also proposes various efficient methods for calculating lower/upper bounds; further experiments demonstrate a significant gap between existing robust training models and the optimal loss.
Chasing Fairness Under Distribution Shift: A Model Weight Perturbation Approach
Zhimeng Jiang (Texas A&M University), Xia Hu (Texas A&M University)
Domain AdaptationOptimizationTabular
🎯 What it does: This paper proposes a robust fair regularization method (RFR) based on model weight perturbation, achieving fairness transfer under distribution shift conditions.
ChatGPT-Powered Hierarchical Comparisons for Image Classification
Zhiyuan Ren (Michigan State University), Xiaoming Liu (Michigan State University)
ClassificationTransformerLarge Language ModelContrastive LearningImageText
🎯 What it does: In zero-shot open vocabulary image classification, the classification ability of CLIP is enhanced by constructing a hierarchical contrastive description based on LLM, without any additional training.
Chatting Makes Perfect: Chat-based Image Retrieval
Matan Levy (Hebrew University of Jerusalem), Dani Lischinski (Hebrew University of Jerusalem)
RetrievalTransformerLarge Language ModelVision Language ModelContrastive LearningImageTextMultimodality
🎯 What it does: ChatIR is proposed, a system that continuously refines retrieval queries through interaction with users to ultimately retrieve target images.
Cheap and Quick: Efficient Vision-Language Instruction Tuning for Large Language Models
Gen Luo (Xiamen University), Rongrong Ji (Xiamen University)
TransformerLarge Language ModelSupervised Fine-TuningVision Language ModelImageTextMultimodality
🎯 What it does: This paper proposes a lightweight Mixture-of-Modality Adaptation (MMA) framework for efficient visual-language instruction fine-tuning of large language models, resulting in the LaVIN model.
Cheaply Estimating Inference Efficiency Metrics for Autoregressive Transformer Models
Deepak Narayanan (NVIDIA), Percy Liang (Stanford University)
Computational EfficiencyTransformerLarge Language ModelTextBenchmark
🎯 What it does: This paper proposes a cost model and idealized runtime metrics for inference with autoregressive Transformer models, demonstrating how to efficiently estimate inference efficiency and costs through a small number of benchmarks, enabling comparable analysis across different models and APIs.
Cinematic Mindscapes: High-quality Video Reconstruction from Brain Activity
Zijiao Chen (National University of Singapore), Juan Helen Zhou (National University of Singapore)
RestorationGenerationDiffusion modelContrastive LearningVideoMultimodalityMagnetic Resonance Imaging
🎯 What it does: Proposed the MinD-Video system, which reconstructs high-quality continuous videos using fMRI data;
Circuit as Set of Points
Jialv Zou (Horizon Robotics), Chang Huang (Horizon Robotics)
TransformerPoint Cloud
🎯 What it does: Abstracts the components in circuit design as point clouds, directly encoding the original node coordinates using Transformer to output multi-scale grid features for congestion and DRC violation prediction;
CL-NeRF: Continual Learning of Neural Radiance Fields for Evolving Scene Representation
Xiuzhe Wu (University of Hong Kong), XIAOJUAN QI
Knowledge DistillationRepresentation LearningNeural Radiance FieldImageBenchmark
🎯 What it does: This paper studies the continuous updating of NeRF to adapt to scene changes and avoid forgetting with only a small number of new images.
CLadder: Assessing Causal Reasoning in Language Models
Zhijing Jin (ETH Zurich), Bernhard Schölkopf (Max Planck Institute for Intelligent Systems)
TransformerLarge Language ModelPrompt EngineeringTextBenchmarkChain-of-Thought
🎯 What it does: This paper presents the CLADDER dataset and the CAUSALCOT prompting strategy for evaluating the formal causal reasoning capabilities of large language models.
Class-Conditional Conformal Prediction with Many Classes
Tiffany Ding (University of California), Ryan Tibshirani
ClassificationImage
🎯 What it does: This paper proposes a clustering-based conformal prediction method aimed at achieving class conditional coverage in classification tasks with limited sample sizes and numerous categories.
Class-Distribution-Aware Pseudo-Labeling for Semi-Supervised Multi-Label Learning
Ming-Kun Xie (Nanjing University of Aeronautics and Astronautics), Sheng-Jun Huang (Nanjing University of Aeronautics and Astronautics)
ClassificationSegmentationConvolutional Neural NetworkSupervised Fine-TuningImage
🎯 What it does: A class distribution-aware pseudo-labeling method (CAP) is proposed for semi-supervised multi-label learning, which generates pseudo-labels by setting thresholds for each class and estimating class distributions using labeled data, thereby making full use of unlabeled samples.
Classification of Heavy-tailed Features in High Dimensions: a Superstatistical Approach
Urte Adomaityte (King's College London), Pierpaolo Vivo (King's College London)
ClassificationOptimization
🎯 What it does: This study investigates the learning and classification of two types of data point clouds from superstatistical heavy-tailed distributions using generalized linear models (GLM) in high dimensions, and derives precise non-trivial limit results for the empirical risk minimization (ERM) estimator in the large sample limit.
CLeAR: Continual Learning on Algorithmic Reasoning for Human-like Intelligence
Bong Gyun Kang (Seoul National University), Sungroh Yoon (Seoul National University)
Knowledge DistillationRecurrent Neural NetworkSequential
🎯 What it does: A continuous learning method CLeAR is proposed for abstract logical reasoning tasks, addressing issues such as input-task decoupling, dynamic dimensions, and OOD generalization.
Clifford Group Equivariant Neural Networks
David Ruhe (University of Amsterdam), Patrick Forré (University of Amsterdam)
Graph Neural NetworkMesh
🎯 What it does: A type of O(n) and E(n) equivariant neural network (CGENNs) based on Clifford algebra is proposed.
CLIP-OGD: An Experimental Design for Adaptive Neyman Allocation in Sequential Experiments
Jessica Dai (University of California Berkeley), Christopher Harshaw (Massachusetts Institute of Technology)
OptimizationSequential
🎯 What it does: This study proposes an adaptive Neyman allocation experimental design called CLIP-OGD, aimed at improving the accuracy of causal inference in sequential experiments.
CLIP4HOI: Towards Adapting CLIP for Practical Zero-Shot HOI Detection
Yunyao Mao (University of Science and Technology of China), Houqiang Li (University of Science and Technology of China)
ClassificationObject DetectionTransformerPrompt EngineeringVision Language ModelImage
🎯 What it does: We propose CLIP4HOI, a two-stage framework for zero-shot human-object interaction detection, which first independently detects humans and objects, then generates all possible human-object pairs, and uses CLIP for fine-grained interaction discrimination.
Closing the Computational-Statistical Gap in Best Arm Identification for Combinatorial Semi-bandits
Ruo-Chun Tzeng (KTH Royal Institute of Technology), Chi-Jen Lu (Academia Sinica)
Optimization
🎯 What it does: The study investigates the problem of best arm identification in combinatorial semi-bands and proposes an algorithm called Perturbed Frank-Wolfe Sampling (P-FWS).
Closing the gap between the upper bound and lower bound of Adam's iteration complexity
Bohan Wang (University of Science and Technology of China), Wei Chen (Institute of Computing Technology, Chinese Academy of Sciences)
Optimization
🎯 What it does: This paper derives a new convergence guarantee for the Adam algorithm, bridging the gap between the upper and lower bounds of Adam's iterative complexity in the existing literature. It proves that under the L-smooth condition and the assumption of bounded noise variance, the upper and lower bounds of Adam's iterative complexity match.
CluB: Cluster Meets BEV for LiDAR-Based 3D Object Detection
Yingjie Wang (University of Science and Technology of China), Yanyong Zhang (Shanghai AI Laboratory)
Object DetectionAutonomous DrivingKnowledge DistillationTransformerSupervised Fine-TuningPoint Cloud
🎯 What it does: A 3D object detection framework called CluB is designed, which combines BEV and clustering features. An auxiliary clustering branch is added to the BEV backbone, and the detection performance is enhanced through feature-level fusion and query-level expansion.
Cluster-aware Semi-supervised Learning: Relational Knowledge Distillation Provably Learns Clustering
Yijun Dong (New York University), Rachel Ward (University of Texas at Austin)
ClassificationKnowledge DistillationImage
🎯 What it does: This paper proves through theoretical analysis of Relational Knowledge Distillation (RKD) that it is equivalent to spectral clustering of the group induction graph revealed by the teacher model, thereby achieving low clustering error in semi-supervised classification.