NeurIPS 2023 Papers — Page 6
Conference on Neural Information Processing Systems · 3218 papers
ClusterFomer: Clustering As A Universal Visual Learner
James Chenhao Liang, Dongfang Liu (Rochester Institute of Technology)
ClassificationObject DetectionSegmentationTransformerImage
🎯 What it does: A clustering-based Transformer model called CLUSTERFORMER is proposed, which can uniformly handle image classification, detection, and segmentation tasks.
Clustering the Sketch: Dynamic Compression for Embedding Tables
Henry Tsang, Thomas Dybdahl Ahle (Meta)
CompressionRecommendation SystemTabular
🎯 What it does: A dynamic compression method that combines hashing and clustering is proposed—Clustered Compositional Embeddings (CCE), which can compress large embedding tables in real-time during the training process.
Cocktail: Mixing Multi-Modality Control for Text-Conditional Image Generation
Minghui Hu (South China University of Technology), Tat-Jen Cham (Nanyang Technological University)
GenerationData SynthesisDiffusion modelImageMultimodality
🎯 What it does: A pipeline named Cocktail has been designed to integrate multimodal control signals (such as edge maps, poses, segmentation maps, etc.) into a text-conditioned diffusion model, achieving fine control of multiple modalities with a single model.
CoDA: Collaborative Novel Box Discovery and Cross-modal Alignment for Open-vocabulary 3D Object Detection
Yang Cao (Hong Kong University of Science and Technology), Dan Xu (Huawei Noah's Ark Lab)
Object DetectionTransformerVision Language ModelContrastive LearningPoint Cloud
🎯 What it does: This paper proposes the CoDA framework, which achieves simultaneous localization and classification of unknown category objects in 3D point cloud scenes with only a small number of base category annotations.
CODA: Generalizing to Open and Unseen Domains with Compaction and Disambiguation
Chaoqi Chen (University of Hong Kong), Yizhou Yu (University of Hong Kong)
Domain AdaptationConvolutional Neural NetworkContrastive LearningImage
🎯 What it does: The CODA framework is proposed to address the issues of open testing domain transfer and the coexistence of unknown classes.
CoDet: Co-occurrence Guided Region-Word Alignment for Open-Vocabulary Object Detection
Chuofan Ma (University of Hong Kong), XIAOJUAN QI
Object DetectionVision Language ModelImageText
🎯 What it does: The CoDet framework is proposed, which utilizes image clustering of the same concept in image-text pairs to discover co-occurring objects through cross-image region similarity, thereby achieving automatic generation of unannotated region-word pairs and training an open vocabulary object detector.
CoDrug: Conformal Drug Property Prediction with Density Estimation under Covariate Shift
Siddhartha Laghuvarapu (University of Illinois), Jimeng Sun (University of Illinois)
Drug DiscoveryGraph Neural NetworkTabularBiomedical Data
🎯 What it does: The CoDrug method is proposed to achieve adaptive confidence sets that conform to coverage in drug property prediction under covariate shift through energy models and kernel density estimation.
Cognitive Model Discovery via Disentangled RNNs
Kevin J Miller, Zeb Kurth-Nelson (Google DeepMind)
Explainability and InterpretabilityReinforcement Learning from Human FeedbackRecurrent Neural NetworkAuto EncoderTime SeriesSequential
🎯 What it does: A Disentangled RNN model (DisRNN) is proposed, which automatically learns interpretable cognitive models from behavioral data by introducing an information bottleneck and separate update rules in RNNs.
Cognitive Steering in Deep Neural Networks via Long-Range Modulatory Feedback Connections
Talia Konkle (Harvard University), George A. Alvarez
ClassificationRecognitionAdversarial AttackConvolutional Neural NetworkSupervised Fine-TuningImageMagnetic Resonance Imaging
🎯 What it does: This paper proposes and implements a deep visual network based on Long-Range Modulation Feedback (LRM), utilizing multi-scale learnable multiplicative modulation to achieve cognitive steering, thereby enhancing classification performance, attack robustness, and brain alignment under default operation and target guidance.
Coherent Soft Imitation Learning
Joe Watson (TU Darmstadt), Nicolas Heess (Google DeepMind)
Robotic IntelligenceReinforcement LearningSequential
🎯 What it does: A method combining Behavior Cloning (BC) and Inverse Reinforcement Learning (IRL) is proposed—Coherent Soft Imitation Learning (CSIL). It derives a 'coherent' reward through entropy-regularized reinforcement learning's soft policy iteration, making the BC policy the optimal policy for the reward, which can be further fine-tuned through reinforcement learning.
CoLA: Exploiting Compositional Structure for Automatic and Efficient Numerical Linear Algebra
Andres Potapczynski (New York University), Andrew Gordon Wilson (New York University)
OptimizationComputational EfficiencyTabular
🎯 What it does: Designed and implemented the CoLA framework, which automates large-scale structured linear algebra operations through a combination of linear operator abstraction and multiple dispatch rules, supporting automatic differentiation, low precision, and GPU acceleration.
Cold Diffusion: Inverting Arbitrary Image Transforms Without Noise
Arpit Bansal (University of Maryland), Tom Goldstein (University of Maryland)
RestorationGenerationSuper ResolutionDiffusion modelImage
🎯 What it does: A diffusion framework without the need for random noise (Cold Diffusion) is proposed, which trains a recovery network through arbitrary image degradation (blurring, occlusion, downsampling, etc.) and designs a new sampling algorithm (TACoS) to achieve generation and reverse recovery.
Collaborative Alignment of NLP Models
Fereshte Khani (Microsoft), Marco Tulio Ribeiro (Google DeepMind)
ClassificationTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: A collaborative model alignment framework called CoAlign is proposed, which improves the compliance of NLP models with business rules and values by allowing multiple users to jointly define concepts and using the controversial areas between local concept models and global models to label data.
Collaborative Learning via Prediction Consensus
Dongyang Fan (École Polytechnique Fédérale de Lausanne), Martin Jaggi (École Polytechnique Fédérale de Lausanne)
Federated LearningKnowledge DistillationImage
🎯 What it does: A collaborative learning algorithm based on predictive consensus is proposed, utilizing shared unlabeled data and adaptive trust weights to achieve pseudo-label collaborative training among multiple models.
Collaborative Score Distillation for Consistent Visual Editing
Subin Kim (KAIST), Jinwoo Shin (KAIST)
Image TranslationGenerationData SynthesisKnowledge DistillationDiffusion modelScore-based ModelImageVideoMultimodality
🎯 What it does: A Collaborative Score Distillation (CSD) method is proposed for different visual modalities (panoramic images, videos, and 3D scenes), utilizing a pre-trained text-to-image diffusion model to achieve consistent visual editing.
Collaboratively Learning Linear Models with Structured Missing Data
Chen Cheng (Stanford University), John Duchi (Stanford University)
OptimizationFederated LearningTabular
🎯 What it does: This study designs a communication-efficient distributed algorithm named COLLAB, aimed at collaboratively learning linear models and improving prediction accuracy in multi-agent environments with structured missing features.
Collapsed Inference for Bayesian Deep Learning
Zhe Zeng (University of California), Guy Van den Broeck (University of California)
Computational EfficiencyConvolutional Neural NetworkImageTabular
🎯 What it does: A Bayesian deep learning inference framework called CIBER is proposed, which is based on folded sampling and weighted model integration (WMI) to efficiently approximate Bayesian model averaging (BMA).
CoLLAT: On Adding Fine-grained Audio Understanding to Language Models using Token-Level Locked-Language Tuning
Amila Silva (University of Melbourne), Hugh James Leather (Meta AI)
ClassificationRetrievalTransformerContrastive LearningMultimodalityAudio
🎯 What it does: This paper proposes CoLLAT, a framework for high-fidelity audio understanding and multimodal applications that achieves token-level alignment between audio and text while keeping the pre-trained text model locked.
Color Equivariant Convolutional Networks
Attila Lengyel (Delft University of Technology), Jan van Gemert (Delft University of Technology)
ClassificationRecognitionConvolutional Neural NetworkImage
🎯 What it does: A color equivariant convolution (CEConv) is proposed, achieving equivariance to color shifts by sharing parameters in the hue space while preserving color information.
Combating Bilateral Edge Noise for Robust Link Prediction
Zhanke Zhou (Hong Kong Baptist University), Bo Han (Hong Kong Baptist University)
Recommendation SystemAnomaly DetectionOptimizationGraph Neural NetworkContrastive LearningGraph
🎯 What it does: A robust graph link prediction framework for bidirectional edge noise (which simultaneously affects input topology and target labels) is proposed—Robust Graph Information Bottleneck (RGIB). Two implementable instances are provided: RGIB-SSL based on self-supervised learning and RGIB-REP based on data reparameterization, aimed at enhancing the link prediction performance of graph neural networks on noisy graphs.
Combating Representation Learning Disparity with Geometric Harmonization
Zhihan Zhou (Shanghai Jiao Tong University), Yanfeng Wang (Shanghai Jiao Tong University)
Object DetectionSegmentationRepresentation LearningContrastive LearningImage
🎯 What it does: This paper proposes a self-supervised learning method called Geometric Harmonization (GH) to alleviate representation learning discrepancies under long-tailed data distributions.
Combinatorial Group Testing with Selfish Agents
Giorgos Chionas, Piotr Krysta (Augusta University)
🎯 What it does: This paper proposes and analyzes the combinatorial group testing problem (CGT) in the presence of selfish agents, providing feasible algorithms (BS_Jumps and BB_GR) applicable to both known and unknown hidden set sizes k, while proving the corresponding lower bounds and constructing an adaptive equilibrium (AE) framework.
Combinatorial Optimization with Policy Adaptation using Latent Space Search
Felix Chalumeau (InstaDeep), Thomas D Barrett
OptimizationTransformerReinforcement LearningTabular
🎯 What it does: This paper proposes COMPASS, a method for finding optimal solutions in combinatorial optimization by conditioning strategies on the latent space and using evolutionary search during inference.
Combining Behaviors with the Successor Features Keyboard
Wilka Carvalho (Harvard University), Daniel Zoran (Google DeepMind)
Robotic IntelligenceReinforcement LearningSequential
🎯 What it does: A dynamic query keyboard based on Successor Features (Successor Features Keyboard, SFK) and its training method CSFA is proposed, which can automatically discover accumulators and task encodings in a 3D Playroom environment and achieve rapid transfer across tasks.
Common Ground in Cooperative Communication
Xiaoran Hao (Rutgers University), Patrick Shafto (Institute for Advanced Study)
OptimizationAuto EncoderTabular
🎯 What it does: This paper proposes a unified scalable theory to describe the two-person cooperative communication problem under varying degrees of common ground, modeling it as a constrained minimization problem. It then frames the teacher (data selection) and learner (hypothesis reasoning) as an encoding/decoding process through the variational autoencoder framework, demonstrating its relationship with previous models.
CommonScenes: Generating Commonsense 3D Indoor Scenes with Scene Graph Diffusion
Guangyao Zhai (Technical University of Munich), Benjamin Busam (Technical University of Munich)
GenerationData SynthesisGraph Neural NetworkDiffusion modelAuto EncoderMeshGraph
🎯 What it does: Generate 3D indoor scenes that conform to common sense based on scene graphs, taking into account both layout and shape generation;
Communication-Efficient Federated Bilevel Optimization with Global and Local Lower Level Problems
Junyi Li (University of Maryland), Heng Huang (University of Maryland)
OptimizationFederated LearningReinforcement LearningImage
🎯 What it does: Proposes two federated bi-level optimization algorithms, FedBiOAcc and FedBiOAcc-Local, which utilize local SGD and momentum variance reduction for efficient estimation of hypergradients, achieving communication complexity of O(1/ε) and linear acceleration;
Compact Neural Volumetric Video Representations with Dynamic Codebooks
Haoyu Guo (Zhejiang University), Xiaowei Zhou (Zhejiang University)
Data SynthesisCompressionConvolutional Neural NetworkNeural Radiance FieldVideo
🎯 What it does: A dynamic codebook is proposed to compress and accelerate the learning and rendering of neural voxel videos;
Comparing Apples to Oranges: Learning Similarity Functions for Data Produced by Different Distributions
Leonidas Tsepenekas (JPMorganChase), Daniele Magazzeni (JPMorganChase)
TabularFinance Related
🎯 What it does: An efficient sampling framework is proposed, which learns the similarity function between different distributions (populations) using limited expert feedback, and provides theoretical guarantees for PAC and query counts.
Comparing Causal Frameworks: Potential Outcomes, Structural Models, Graphs, and Abstractions
Duligur Ibeling (Stanford University), Thomas Icard (Stanford University)
🎯 What it does: This paper explores the theoretical relationship between the Rubin Causal Model (RCM) and the Structural Causal Model (SCM), and proposes the conditions under which RCM can be represented by SCM, as well as the abstract representation mechanism of RCM.
Complementary Benefits of Contrastive Learning and Self-Training Under Distribution Shift
Saurabh Garg (Carnegie Mellon University), Aditi Raghunathan (Carnegie Mellon University)
Domain AdaptationConvolutional Neural NetworkContrastive LearningImage
🎯 What it does: This paper studies the effects of combining self-training and contrastive learning under unsupervised domain adaptation (UDA) and demonstrates through experiments and theoretical analysis that the two methods have complementary advantages in scenarios with distribution shifts; however, the combination shows no significant benefits in semi-supervised learning (SSL) scenarios without distribution shifts.
Complex Query Answering on Eventuality Knowledge Graph with Implicit Logical Constraints
Jiaxin Bai (Hong Kong University of Science and Technology), Yangqiu Song (Hong Kong University of Science and Technology)
Graph
🎯 What it does: This paper proposes a Complex Eventuality Query Answering (CEQA) framework for executing complex logical queries with implicit logical constraints (occurrence and temporal order) on Event Knowledge Graphs (EVKG).
Complex-valued Neurons Can Learn More but Slower than Real-valued Neurons via Gradient Descent
Jin-Hui Wu (Nanjing University), Zhi-Hua Zhou (Nanjing University)
Optimization
🎯 What it does: This paper analyzes the learning problem of a single neuron through gradient descent, theoretically comparing the learnability and convergence speed of complex-valued neurons with real-valued neurons.
Complexity Matters: Rethinking the Latent Space for Generative Modeling
Tianyang Hu (Huawei Noah's Ark Lab), Zhenguo Li (Huawei Noah's Ark Lab)
GenerationData SynthesisDiffusion modelAuto EncoderGenerative Adversarial NetworkImage
🎯 What it does: The paper redefines the latent space from the perspective of model complexity, introduces the concepts of GAN-induced distance and optimal latent distribution, and proposes a two-stage Decoupled Autoencoder training method.
Complexity of Derivative-Free Policy Optimization for Structured $\mathcal{H}_\infty$ Control
Xingang Guo (University of Illinois at Urbana-Champaign), Bin Hu (University of Illinois at Urbana-Champaign)
OptimizationReinforcement LearningTime Series
🎯 What it does: This paper proposes and analyzes the feasibility and complexity of using zero-order (only accessing closed-loop H∞ norm) strategy optimization methods in structured H∞ design (fixed static output feedback controllers), providing upper bounds on sample complexity under both precise and imprecise zero-order operators.
Composable Coresets for Determinant Maximization: Greedy is Almost Optimal
Siddharth Gollapudi (Microsoft Research), Varun Sivashankar (Microsoft Research)
OptimizationTabular
🎯 What it does: This paper studies the problem of maximizing determinants in a large-scale data environment and proves that the greedy algorithm has near-optimal approximation guarantees in the setting of composable coresets.
Composing Parameter-Efficient Modules with Arithmetic Operation
Jinghan Zhang (Hong Kong University of Science and Technology), Junxian He (Hong Kong University of Science and Technology)
Domain AdaptationOptimizationTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: The study efficiently performs addition and negation operations on trained parameter-efficient modules (LoRA, IA3) in the parameter space, allowing for the synthesis of new modules without additional training, supporting various scenarios such as distribution generalization, multi-tasking, de-learning, domain transfer, and de-biasing.
Compositional Abilities Emerge Multiplicatively: Exploring Diffusion Models on a Synthetic Task
Maya Okawa (Harvard University), Hidenori Tanaka (Harvard University)
GenerationData SynthesisDiffusion modelImage
🎯 What it does: This study investigates the combinatorial generation capability of conditional diffusion models in synthesis tasks and systematically evaluates their combinatorial generalization performance out of training distribution.
Compositional Foundation Models for Hierarchical Planning
Anurag Ajay (Massachusetts Institute of Technology), Pulkit Agrawal (Massachusetts Institute of Technology)
Robotic IntelligenceTransformerLarge Language ModelDiffusion modelVideoMultimodality
🎯 What it does: A hierarchical decision framework named HiP is proposed, which combines large language models, video diffusion models, and inverse dynamics models to achieve long-term robotic planning tasks; consistency of outputs from different models is ensured through iterative refinement.
Compositional Generalization from First Principles
Thaddäus Wiedemer (University of Tübingen), Wieland Brendel (Max-Planck-Institute for Intelligent Systems)
GenerationData SynthesisRepresentation LearningImage
🎯 What it does: This paper redefines compositionality from the ground up, proposing a compositional data generation model that satisfies identifiability conditions during the generation process, and proves that under certain support conditions, the model can achieve consistent generalization on compositional samples outside the training distribution.
Compositional Policy Learning in Stochastic Control Systems with Formal Guarantees
Đorđe Žikelić (Institute of Science and Technology Austria), Thomas A Henzinger
OptimizationReinforcement LearningSequential
🎯 What it does: A composable reinforcement learning framework CLAPS is proposed for learning neural network policies in stochastic control systems, providing probabilistic guarantees that satisfy the SPECTRL specification.
Compositional Sculpting of Iterative Generative Processes
Timur Garipov (Massachusetts Institute of Technology), Tommi S. Jaakkola (Massachusetts Institute of Technology)
GenerationData SynthesisDiffusion modelFlow-based ModelImageGraph
🎯 What it does: A general 'Compositional Sculpting' framework is proposed, which enables the combination and reshaping of the base distribution through observed variables in iterative generative models (such as GFlowNets and diffusion models), supporting binary operations like harmonic mean and contrast, and allowing direct sampling guided by classifiers.
Compressed Video Prompt Tuning
Bing Li (Beihang University), Di Huang (Beihang University)
RecognitionCompressionComputational EfficiencyTransformerPrompt EngineeringVideo
🎯 What it does: Utilizing a pre-trained raw video model, efficient action recognition is performed on compressed videos through a prompt tuning method (CVPT).
Compression with Bayesian Implicit Neural Representations
Zongyu Guo (University of Science and Technology of China), José Miguel Hernández-Lobato (University of Cambridge)
CompressionImageAudio
🎯 What it does: A compression framework based on Variational Bayesian Implicit Neural Representation (INR) is proposed—COMBINER, which compresses the posterior samples of INR weights using relative entropy coding to directly achieve rate-distortion optimization;
Computational Complexity of Learning Neural Networks: Smoothness and Degeneracy
Amit Daniely (Hebrew University), Gal Vardi (TTI-Chicago)
🎯 What it does: It is proven that under Gaussian input distribution, ReLU networks with a depth of 3 (as well as depth 2 networks with perturbed input distribution and parameters) cannot be learned in polynomial time within the smoothed-analysis framework, and this difficulty persists even when the weight matrix is non-degenerate.
Computational Guarantees for Doubly Entropic Wasserstein Barycenters
Tomas Vaskevicius, Lénaïc Chizat (Institute of Mathematics École Polytechnique Fédérale de Lausanne)
OptimizationPoint Cloud
🎯 What it does: A damped Sinkhorn iterative algorithm for computing the dual entropy regularized Wasserstein barycenter ((λ,τ)-barycenter) is proposed, along with convergence guarantees for any regularization parameters; an approximate version is also provided, supporting free-support scenarios.
Computing a human-like reaction time metric from stable recurrent vision models
Lore Goetschalckx (Brown University), Thomas Serre (Brown University)
Convolutional Neural NetworkRecurrent Neural NetworkImage
🎯 What it does: This study proposes a general methodology to construct a computational metric for human reaction time through a stable recursive visual model, aiming to capture the temporal characteristics of visual decision-making.
Computing Approximate $\ell_p$ Sensitivities
Swati Padmanabhan (Massachusetts Institute of Technology), Qiuyi Zhang
OptimizationComputational EfficiencyTabular
🎯 What it does: This paper studies how to quickly approximate the ℓp sensitivity, total sensitivity, and maximum sensitivity of matrices, and provides corresponding algorithms.
Computing Optimal Equilibria and Mechanisms via Learning in Zero-Sum Extensive-Form Games
Brian Hu Zhang (Carnegie Mellon University), Tuomas Sandholm (Carnegie Mellon University)
OptimizationReinforcement LearningTabularSequential
🎯 What it does: A Lagrangian relaxation framework is proposed to transform the optimal equilibrium of multi-player extensive-form games and mechanism design problems into zero-sum games, and solved through learning algorithms.
Computing Optimal Nash Equilibria in Multiplayer Games
Youzhi Zhang (Hong Kong Institute of Science and Innovation), Venkatramanan Siva Subrahmanian
Optimization
🎯 What it does: This paper proposes a new algorithm called CRM based on correlation plans and their relational constraints, aimed at solving the optimal Nash Equilibrium (NE) in multi-player games, specifically optimizing a given objective function over the entire NE space.
ComSL: A Composite Speech-Language Model for End-to-End Speech-to-Text Translation
Chenyang Le (Shanghai Jiao Tong University), Xuedong Huang (Microsoft Cloud and AI)
RecognitionGenerationTransformerSupervised Fine-TuningTextAudio
🎯 What it does: A Composite Speech-Language (ComSL) end-to-end speech-to-text translation model is constructed, combining the Whisper speech encoder, mBART language decoder, and adapters, utilizing multi-task learning and cross-modal learning (CML) for efficient model fine-tuning.
Concept Algebra for (Score-Based) Text-Controlled Generative Models
Zihao Wang (University of Chicago), Victor Veitch (Google Research)
GenerationDiffusion modelScore-based ModelImageTextStochastic Differential Equation
🎯 What it does: Proposes the concept subspace theory and implements concept algebra in the score model for concept-level editing of text-controlled generation models.
Concept Distillation: Leveraging Human-Centered Explanations for Model Improvement
Avani Gupta (International Institute of Information Technology Hyderabad), P J Narayanan (International Institute of Information Technology Hyderabad)
ClassificationKnowledge DistillationTransformerAuto EncoderImage
🎯 What it does: A concept distillation framework is proposed, utilizing human-centered concept explanations (CAV) and knowledge from a teacher model to pre-train the model for debiasing, enhancement, or introduction of prior knowledge.
ConDaFormer: Disassembled Transformer with Local Structure Enhancement for 3D Point Cloud Understanding
Lunhao Duan (Wuhan University), Dacheng Tao (University of Sydney)
Object DetectionSegmentationTransformerPoint Cloud
🎯 What it does: A Transformer block named ConDaFormer is proposed, specifically designed for 3D point cloud understanding tasks;
Conditional Adapters: Parameter-efficient Transfer Learning with Fast Inference
Tao Lei (Google), Ming-Wei Chang (Google)
Computational EfficiencyTransformerSupervised Fine-TuningTextAudio
🎯 What it does: Introducing Conditional Adapter (CODA) based on pre-trained Transformers, which adds small adapters and learnable routers at each layer to select a subset of tokens for high-cost Transformer computations, performing deep inference only on important tokens to enhance inference efficiency.
Conditional independence testing under misspecified inductive biases
Felipe Maia Polo (University of Michigan), Moulinath Banerjee (University of Michigan)
Tabular
🎯 What it does: This paper studies the performance of regression-based conditional independence (CI) tests under model misspecification and proposes a new robust testing method called the Rao-Blackwellized Predictor Test (RBPT).
Conditional Matrix Flows for Gaussian Graphical Models
Marcello Massimo Negri (University of Basel), Volker Roth (University of Basel)
Flow-based ModelGraphTabularBiomedical Data
🎯 What it does: A Conditional Matrix Flow (CMF) model is constructed, achieving variational inference of the precision matrix in high-dimensional Gaussian Graphical Models (GGM) by constructing conditional normalizing flows in the space of symmetric positive definite matrices. It can simultaneously obtain Bayesian posterior, frequentist solution paths, and marginal likelihood for model selection.
Conditional Mutual Information for Disentangled Representations in Reinforcement Learning
Mhairi Dunion (University of Edinburgh), Stefano V Albrecht
Reinforcement LearningSequential
🎯 What it does: Proposes an auxiliary task named CMID, which utilizes conditional mutual information to learn separable representations of RL observations with correlated features;
Conditional Score Guidance for Text-Driven Image-to-Image Translation
Hyunsoo Lee (Seoul National University), Bohyung Han (Seoul National University)
Image TranslationGenerationDiffusion modelImageText
🎯 What it does: Using a pre-trained text-to-image diffusion model, we design a training-free text-driven image-to-image translation method that selectively edits the target image based on the source image and source text while keeping the background unchanged.
Conditional score-based diffusion models for Bayesian inference in infinite dimensions
Lorenzo Baldassari (University of Basel), Maarten V. de Hoop (Rice University)
Diffusion modelScore-based ModelStochastic Differential Equation
🎯 What it does: A conditional sampling method based on unconditional score diffusion models in infinite-dimensional Hilbert space is proposed, achieving full probabilistic inference for Bayesian inverse problems.
Coneheads: Hierarchy Aware Attention
Albert Tseng (Cornell University), Christopher De Sa (Cornell University)
ClassificationRecognitionGraph Neural NetworkTransformerTextGraph
🎯 What it does: A hyperbolic entailment cone-based attention mechanism called Cone Attention is proposed to explicitly model hierarchical structures in attention calculations.
Conformal Meta-learners for Predictive Inference of Individual Treatment Effects
Ahmed Alaa, Mark van der Laan (University of California Berkeley)
Meta LearningTabular
🎯 What it does: A conformal meta-learner framework based on two-stage pseudo-observation regression has been developed to provide prediction intervals for individual treatment effects (ITE).
Conformal PID Control for Time Series Prediction
Anastasios Nikolas Angelopoulos, Ryan Tibshirani
OptimizationTime SeriesFinance Related
🎯 What it does: An adaptive uncertainty quantification method suitable for time series forecasting, called conformal PID control, is proposed.
Conformal Prediction for Time Series with Modern Hopfield Networks
Andreas Auer (Johannes Kepler University Linz), Sepp Hochreiter (Johannes Kepler University Linz)
Anomaly DetectionOptimizationRecurrent Neural NetworkTime Series
🎯 What it does: A synthetic forecasting method for time series, HopCPT, is proposed, which generates narrower prediction intervals while maintaining the coverage of confidence intervals.
Conformal Prediction for Uncertainty-Aware Planning with Diffusion Dynamics Model
Jiankai Sun (Stanford University), Mac Schwager (Stanford University)
Robotic IntelligenceTransformerReinforcement LearningDiffusion modelTabular
🎯 What it does: This paper combines conformal prediction with diffusion dynamics models to construct a framework for uncertainty-aware planning and offline reinforcement learning.
Conformal Prediction Sets for Ordinal Classification
PRASENJIT DEY, Sivaramakrishnan R Kaveri
ClassificationTabular
🎯 What it does: Design a confidence prediction method for ordinal classification (COPOC) that can generate a minimal and continuous prediction set while ensuring coverage.
Conformalized matrix completion
Yu Gui (University of Chicago), Cong Ma (University of Chicago)
Tabular
🎯 What it does: A distribution-free matrix completion confidence interval method is proposed—Conformalized Matrix Completion (CMC), which provides confidence intervals for missing entries without relying on any low-rank or noise assumptions.
Connected Superlevel Set in (Deep) Reinforcement Learning and its Application to Minimax Theorems
Sihan Zeng (Georgia Institute of Technology), Justin Romberg (Georgia Institute of Technology)
OptimizationReinforcement Learning
🎯 What it does: In the policy optimization of reinforcement learning, it is proven that the upper level set of the objective function is always connected, and this property is extended to policies represented by over-parameterized neural networks; based on this result, a minimax theorem for robust reinforcement learning problems is derived, proving the existence of a Nash equilibrium;
Connecting Certified and Adversarial Training
Yuhao Mao (ETH Zurich), Martin Vechev (ETH Zurich)
OptimizationAdversarial AttackConvolutional Neural NetworkImage
🎯 What it does: A new certificate-free robust network training method called TAPS is proposed, which combines IBP and PGD and achieves joint training through a gradient connector, thereby reducing over-regularization while maintaining verifiability and improving both standard and certified accuracy.
Connecting Multi-modal Contrastive Representations
Zehan Wang (Zhejiang University), Zhou Zhao (Shanghai AI Laboratory)
RetrievalRepresentation LearningContrastive LearningImageTextMultimodalityPoint CloudAudio
🎯 What it does: A training-efficient multimodal contrastive representation connection method C-MCR was developed, which utilizes existing contrastive models (such as CLIP, CLAP, ULIP) to generate broader contrastive representations by reprojecting and aligning overlapping modalities without paired data.
Connecting Pre-trained Language Model and Downstream Task via Properties of Representation
Chenwei Wu (Duke University), Rong Ge (Duke University)
ClassificationRepresentation LearningTransformerLarge Language ModelText
🎯 What it does: This study investigates the transfer mechanism of pre-trained language model representations in downstream tasks, providing theoretical analysis and empirical validation.
ConRad: Image Constrained Radiance Fields for 3D Generation from a Single Image
Senthil Purushwalkam (Salesforce AI Research), Nikhil Naik (Salesforce AI Research)
GenerationData SynthesisDiffusion modelScore-based ModelNeural Radiance FieldImagePoint Cloud
🎯 What it does: A method for generating complete 3D models from a single image is proposed, realized through a novel image-constrained radiance field (ConRad).
Conservative Offline Policy Adaptation in Multi-Agent Games
Chengjie Wu (Tsinghua University), Chongjie Zhang (Washington University in St. Louis)
OptimizationReinforcement Learning
🎯 What it does: This paper proposes a method for offline policy adaptation in multi-agent games, utilizing the behavior data of target agents to find risk-safe opponent exploitation or cooperation strategies through environmental interaction during the training phase.
Conservative State Value Estimation for Offline Reinforcement Learning
Liting Chen (McGill University), Dongmei Zhang (Microsoft)
Reinforcement LearningTabularBenchmark
🎯 What it does: Designed and implemented the CSVE algorithm, which achieves conservative state value estimation by penalizing outlier states in the Bellman iteration and embedding it into the actor-critic framework.
Consistent Aggregation of Objectives with Diverse Time Preferences Requires Non-Markovian Rewards
Silviu Pitis (University of Toronto)
OptimizationReinforcement Learning
🎯 What it does: The paper proves that in multi-objective reinforcement learning, when different objectives have different time preferences (discount factors), simple Markovian reward aggregation is not feasible. It is necessary to introduce non-Markovian rewards or to achieve dynamically consistent multi-objective policies through state space expansion. It also proposes a historical discounting method to address the 'first-generation dictatorship' problem in intergenerational decision-making.
Consistent Diffusion Models: Mitigating Sampling Drift by Learning to be Consistent
Giannis Daras (University of Texas), Constantinos Costis Daskalakis
GenerationData SynthesisDiffusion modelScore-based ModelImageOrdinary Differential Equation
🎯 What it does: A consistency constraint (Consistency Property, CP) is proposed, which forces the model to maintain self-consistency in the generated trajectories during training, thereby reducing sampling drift in diffusion models.
Constant Approximation for Individual Preference Stable Clustering
Anders Aamand (Massachusetts Institute of Technology), Fred Zhang (University of California Berkeley)
OptimizationTabular
🎯 What it does: A polynomial-time algorithm is proposed to achieve constant approximation of individual preference stable (IP) clustering in arbitrary metric spaces, and it is extended to a more general f-IP stability, including maximum distance and minimum distance versions.
Constrained Policy Optimization with Explicit Behavior Density For Offline Reinforcement Learning
Jing Zhang (Hong Kong University of Science and Technology), Bingyi Jing
OptimizationReinforcement LearningGenerative Adversarial NetworkTabularBenchmark
🎯 What it does: A constraint policy optimization method for offline reinforcement learning, CPED, is proposed. It uses Flow-GAN for explicit density estimation of the behavior policy and incorporates a safety region constraint based on that density into policy learning, reducing excessive conservativeness towards OOD points.
Constraint-Conditioned Policy Optimization for Versatile Safe Reinforcement Learning
Yihang Yao (Carnegie Mellon University), Ding Zhao (Google DeepMind)
OptimizationSafty and PrivacyReinforcement LearningTabular
🎯 What it does: This paper proposes a safety reinforcement learning framework called CCPO that achieves zero-shot adaptation under different safety thresholds, addressing the issue that traditional safe RL can only be trained for preset thresholds.
Constructing Non-isotropic Gaussian Diffusion Model Using Isotropic Gaussian Diffusion Model for Image Editing
Xi Yu (Xi'an Jiaotong University), Jian Sun (Xi'an Jiaotong University)
GenerationDiffusion modelImage
🎯 What it does: A non-isotropic Gaussian diffusion model (NGDM) is proposed for image editing.
Construction of Hierarchical Neural Architecture Search Spaces based on Context-free Grammars
Simon Schrodi (University of Freiburg), Frank Hutter (University of Freiburg)
Neural Architecture SearchTransformerImage
🎯 What it does: A unified framework based on Context-Free Grammar (CFG) is proposed to construct a scalable hierarchical neural architecture search space.
Content-based Unrestricted Adversarial Attack
Zhaoyu Chen (Fudan University), Wenqiang Zhang (Tencent)
OptimizationAdversarial AttackConvolutional Neural NetworkTransformerDiffusion modelImage
🎯 What it does: Utilizing diffusion models to map images onto a low-dimensional natural image manifold, followed by latent space optimization on that manifold to generate naturalistic and model-agnostic unrestricted adversarial samples.
Context Shift Reduction for Offline Meta-Reinforcement Learning
Yunkai Gao (University of Science and Technology of China), Yunji Chen (Intelligent Software Research Center)
Meta LearningReinforcement LearningTabular
🎯 What it does: An offline meta reinforcement learning method named CSRO is proposed to address the context shift problem caused by different behavior policies during the training and testing phases; it learns task representations by maximizing task information and minimizing the mutual information of policy information, and employs a context collection strategy without prior knowledge during the testing phase.
Context-guided Embedding Adaptation for Effective Topic Modeling in Low-Resource Regimes
Yishi Xu (Xidian University), Mingyuan Zhou (University of Texas at Austin)
Meta LearningGraph Neural NetworkAuto EncoderText
🎯 What it does: This paper proposes a meta-learning topic model called Meta-CETM, which is based on task-specific semantic graphs and a Gaussian mixture prior. It can quickly adapt to generate contextually relevant word embeddings and discover high-quality topics in low-resource scenarios.
Context-lumpable stochastic bandits
Chung-Wei Lee (University of Southern California), Csaba Szepesvari
Reinforcement Learning
🎯 What it does: This study investigates the contextual bandit problem with S contexts and K actions, assuming the existence of r lumpable context groups (r ≤ min{S, K}), and provides algorithms for achieving ε-optimal strategies and cumulative risk under PAC and online learning frameworks.
Context-PIPs: Persistent Independent Particles Demands Spatial Context Features
Weikang BIAN, Hongsheng Li (Chinese University of Hong Kong)
Object TrackingOptical FlowVideo
🎯 What it does: Proposes the Context-PIPs framework, which enhances video particle (PIP) tracking accuracy by adding Source Feature Enhancement (SOFE) and Target Feature Aggregation (TAFA) modules based on PIPs, utilizing spatial context features.
Contextual Bandits and Imitation Learning with Preference-Based Active Queries
Ayush Sekhari (Massachusetts Institute of Technology), Runzhe Wu (Cornell University)
OptimizationReinforcement Learning from Human FeedbackReinforcement Learning
🎯 What it does: This paper proposes two new algorithms, AURORA for contextual dueling bandits and AURORAE for preference-based active querying through imitation learning.
Contextual Gaussian Process Bandits with Neural Networks
Haoting Zhang (University of California), Zeyu Zheng (University of California)
Recommendation SystemOptimizationReinforcement Learning from Human FeedbackGraphTime Series
🎯 What it does: A neural network accompanied Gaussian process (NN-AGP) model is proposed to address the contextual Gaussian process gambling problem, particularly in cases where the reward function has complex dependencies on intricate contextual variables.
Contextual Stochastic Bilevel Optimization
Yifan Hu (École Polytechnique Fédérale de Lausanne), Daniel Kuhn (École Polytechnique Fédérale de Lausanne)
OptimizationMeta LearningImage
🎯 What it does: A Contextual Stochastic Bi-level Optimization (CSBO) framework is proposed, along with two gradient-based solving algorithms (DL-SGD and RT-MLMC).
Contextually Affinitive Neighborhood Refinery for Deep Clustering
Chunlin Yu (ShanghaiTech University), Jingya Wang (ShanghaiTech University)
Representation LearningGraph Neural NetworkContrastive LearningImage
🎯 What it does: This paper proposes the Contextually Affinitive Neighborhood Refinery (CoNR), which enhances deep clustering performance by mining richer neighbors through online reordering and introducing progressive boundary filtering.
ContiFormer: Continuous-Time Transformer for Irregular Time Series Modeling
Yuqi Chen (Fudan University), Dongsheng Li (Microsoft Research Asia)
ClassificationOptimizationTransformerTime SeriesSequentialOrdinary Differential Equation
🎯 What it does: A continuous time Transformer (ContiFormer) is proposed, which can directly handle irregular time series by integrating the continuous dynamics of Neural ODE with the attention mechanism of Transformer.
Continual Learning for Instruction Following from Realtime Feedback
Alane Suhr (University of California), Yoav Artzi (Cornell University)
Robotic IntelligenceReinforcement Learning from Human FeedbackReinforcement LearningSequential
🎯 What it does: A continuous learning framework based on real-time binary feedback has been designed and implemented, allowing the agent to improve continuously through interactions with human users.
ContinuAR: Continuous Autoregression For Infinite-Fidelity Fusion
WEI W. XING, Zheng Xing (Rockchip Electronics Co., Ltd)
Time SeriesPhysics RelatedStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: A continuous autoregressive model, ContinuAR, has been designed and implemented to handle the infinite-order multi-fidelity fusion problem.
Continuous Parametric Optical Flow
Jianqin Luo (Northwestern Polytechnical University), Yuchao Dai (Northwestern Polytechnical University)
Object TrackingRecurrent Neural NetworkOptical FlowVideoOrdinary Differential Equation
🎯 What it does: A continuous parameterized optical flow model (CPFlow) is proposed, which continuously represents pixel motion in the form of B-spline curves;
Continuous-time Analysis of Anchor Acceleration
Jaewook J. Suh (Seoul National University), Ernest K. Ryu (Seoul National University)
OptimizationTabularOrdinary Differential Equation
🎯 What it does: Analyzed a continuous-time model of anchor acceleration (different from Nesterov acceleration), provided an upper bound on the convergence rate, and proposed an adaptive acceleration method based on this analysis, followed by experimental validation on distributed compressed sensing tasks.
Continuous-Time Functional Diffusion Processes
Giulio Franzese (EURECOM), Pietro Michiardi (EURECOM)
GenerationData SynthesisTransformerDiffusion modelImageStochastic Differential Equation
🎯 What it does: This paper proposes the Functional Diffusion Processes (FDPs), extending score-based diffusion models to infinite-dimensional function spaces and constructing generative models that do not require specialized network architectures.
Contrast Everything: A Hierarchical Contrastive Framework for Medical Time-Series
Yihe Wang (University of North Carolina), Xiang Zhang (University of North Carolina)
Anomaly DetectionRepresentation LearningConvolutional Neural NetworkContrastive LearningTime SeriesBiomedical DataAlzheimer's DiseaseElectrocardiogram
🎯 What it does: This paper proposes a hierarchical contrastive learning framework named COMET, designed to extract general representations of medical time series without relying on a large amount of labeled data.
Contrast, Attend and Diffuse to Decode High-Resolution Images from Brain Activities
Jingyuan Sun (KU Leuven), Marie-Francine Moens (KU Leuven)
GenerationRepresentation LearningTransformerDiffusion modelAuto EncoderImageMultimodalityMagnetic Resonance Imaging
🎯 What it does: High-resolution visual image decoding of brain activity is achieved through dual-stage fMRI representation learning and latent diffusion models.
Contrastive Lift: 3D Object Instance Segmentation by Slow-Fast Contrastive Fusion
Yash Sanjay Bhalgat (Visual Geometry Group University of Oxford), Andrew Zisserman (Visual Geometry Group University of Oxford)
Object DetectionSegmentationNeural Radiance FieldContrastive LearningPoint Cloud
🎯 What it does: Using a 2D pre-trained instance segmentation model, 2D segmentation information is integrated into a 3D neural field to generate 3D object instance segmentation.
Contrastive Modules with Temporal Attention for Multi-Task Reinforcement Learning
Siming Lan (University of Science and Technology of China), Yunji Chen (Intelligent Software Research Center Institute of Software CAS)
Robotic IntelligenceMeta LearningReinforcement LearningMixture of ExpertsContrastive LearningSequential
🎯 What it does: This paper proposes the CMTA method, which achieves module differentiation in multi-task reinforcement learning through the use of contrastive learning and dynamically combines modules at each time step using temporal attention, significantly improving multi-task learning performance.