NeurIPS 2023 Papers — Page 7
Conference on Neural Information Processing Systems · 3218 papers
Contrastive Moments: Unsupervised Halfspace Learning in Polynomial Time
Xinyuan Cao (Georgia Institute of Technology), Santosh Vempala
OptimizationContrastive Learning
🎯 What it does: A polynomial-time algorithm is proposed for learning margin-based half-spaces in high-dimensional spaces without requiring labels. Under specific distribution assumptions, the algorithm can effectively identify hidden half-spaces.
Contrastive Retrospection: honing in on critical steps for rapid learning and generalization in RL
Chen Sun (Mila), Blake Aaron Richards
Reinforcement LearningContrastive Learning
🎯 What it does: This paper studies a module called ConSpec that can be directly integrated into existing reinforcement learning algorithms. It utilizes offline contrastive learning to learn 'prototypes' from successful and failed experiences, in order to identify key steps in tasks and generate intrinsic rewards, thereby accelerating long-term credit assignment.
Contrastive Sampling Chains in Diffusion Models
Junyu Zhang (Central South University), Chang Xu (University of Sydney)
GenerationData SynthesisDiffusion modelContrastive LearningImageStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: For the pre-trained diffusion model, a contrastive sampling chain is constructed, and InfoNCE contrastive loss is added during the fine-tuning process. This approach combines dynamic weighting and BPTT to reduce numerical discretization errors, thereby improving generation quality and sampling speed.
Contrastive Training of Complex-Valued Autoencoders for Object Discovery
Aleksandar Stanić (IDSIA), Jürgen Schmidhuber (IDSIA)
Object DetectionRepresentation LearningConvolutional Neural NetworkAuto EncoderContrastive LearningImage
🎯 What it does: Improved the complex-valued autoencoder and enhanced object separation capability through contrastive learning, achieving unsupervised object discovery.
Controlling Text-to-Image Diffusion by Orthogonal Finetuning
Zeju Qiu (Max Planck Institute for Intelligent Systems), Bernhard Schölkopf (Max Planck Institute for Intelligent Systems)
GenerationData SynthesisDiffusion modelImageTextMultimodality
🎯 What it does: This paper proposes an Orthogonal Finetuning (OFT) method for fine-tuning large-scale text-to-image diffusion models, enabling them to better perform controlled generation tasks (such as subject-driven generation and multimodal controlled generation).
Convergence analysis of ODE models for accelerated first-order methods via positive semidefinite kernels
Jungbin Kim (Seoul National University), Insoon Yang (Seoul National University)
OptimizationOrdinary Differential Equation
🎯 What it does: A systematic analysis of the convergence of continuous-time accelerated gradient flow models is conducted, transforming the proof of convergence rate into the verification of positive semi-definite kernels, and proposing a new framework for continuous-time performance estimation problems (Continuous-PEP);
Convergence Analysis of Sequential Federated Learning on Heterogeneous Data
Yipeng Li (Beijing University of Posts and Telecommunications), Xinchen Lyu (Beijing University of Posts and Telecommunications)
OptimizationFederated LearningConvolutional Neural NetworkImage
🎯 What it does: This paper studies the convergence of Sequential Federated Learning (SFL) under heterogeneous data and compares it with traditional Parallel Federated Learning (PFL), providing upper bounds for strongly convex, generally convex, and non-convex objective functions.
Convergence of Actor-Critic with Multi-Layer Neural Networks
Haoxing Tian (Boston University), Ioannis Paschalidis
Reinforcement Learning
🎯 What it does: This study proves that under Markov sampling, the deep Actor-Critic method with a single loop and single time scale can converge on a multi-layer neural network with width m, with an error upper bound of O(T^{-1/2}) + O(ε) + ˜O(m^{-1/2}).
Convergence of Adam Under Relaxed Assumptions
Haochuan Li (Massachusetts Institute of Technology), Ali Jadbabaie (Massachusetts Institute of Technology)
OptimizationConvolutional Neural NetworkImage
🎯 What it does: It is proven that Adam and its variance-reduced version can still converge to an ϵ-differentiable solution under weaker assumptions (non-global smoothness, no global gradient bounds, and noise can be sub-Gaussian), with gradient complexities of O(ϵ⁻⁴) and O(ϵ⁻³), respectively.
Convergence of Alternating Gradient Descent for Matrix Factorization
Rachel Ward (University of Texas), Tamara G. Kolda (MathSci.ai)
Optimization
🎯 What it does: Applying Alternating Gradient Descent (AGD) to non-convex matrix decomposition objectives, it is proven that under specific random initialization, linear convergence can be achieved with a certain iterative complexity.
Convergence of mean-field Langevin dynamics: time-space discretization, stochastic gradient, and variance reduction
Taiji Suzuki (University of Tokyo), Atsushi Nitanda (Kyushu Institute of Technology)
OptimizationStochastic Differential Equation
🎯 What it does: A unified theoretical framework is proposed for the non-asymptotic convergence analysis of mean field Langevin dynamics (MFLD) with finite particles, discrete time, and stochastic gradients.
Convergent Bregman Plug-and-Play Image Restoration for Poisson Inverse Problems
Samuel Hurault (University of Bordeaux), Nicolas Papadakis (University of Bordeaux)
RestorationConvolutional Neural NetworkImage
🎯 What it does: A Plug-and-Play image recovery framework based on Bregman distance is proposed, with the design of a Bregman Score denoiser, and two convergent algorithms, B-RED and B-PnP, are introduced for the Poisson inverse problem.
Convex and Non-convex Optimization Under Generalized Smoothness
Haochuan Li (Massachusetts Institute of Technology), Ali Jadbabaie (Massachusetts Institute of Technology)
Optimization
🎯 What it does: This paper studies the convergence of methods such as Gradient Descent (GD), Stochastic Gradient Descent (SGD), and Nesterov Accelerated Gradient (NAG) under a more general 'ℓ-smoothness' condition, proving that classic convergence rates can be achieved in both convex and non-convex settings.
Convex-Concave Zero-Sum Markov Stackelberg Games
Denizalp Goktas (Brown University), Amy Greenwald (Brown University)
OptimizationReinforcement LearningSequential
🎯 What it does: A nested SGDA algorithm based on policy gradient is proposed, which can achieve global optimality in zero-sum Markov Stackelberg games with continuous states and continuous actions by utilizing the noisy gradient of observed trajectories, and it is proven to converge in polynomial time; the reach-avoid problem is modeled as a convex-concave zero-sum Markov Stackelberg game, and experiments verify that the Stackelberg limit strategy outperforms the Nash strategy.
Convolution Monge Mapping Normalization for learning on sleep data
Theo Gnassounou (University Paris-Saclay), Alexandre Gramfort (University Paris-Saclay)
Domain AdaptationConvolutional Neural NetworkTime SeriesBiomedical Data
🎯 What it does: A convolutional Monge mapping normalization method based on Wasserstein barycenter (CMMN) is proposed for unsupervised domain adaptation of sleep EEG data.
Convolutional Neural Operators for robust and accurate learning of PDEs
Bogdan Raonic (ETH Zurich), Emmanuel de Bezenac (ETH Zurich)
Convolutional Neural NetworkBenchmarkPhysics Related
🎯 What it does: A convolutional neural operator learning framework is proposed—Convolutional Neural Operator (CNO), which achieves continuous-discrete equivalence by modifying basic operations such as convolution, upsampling/downsamping, and activation;
Convolutional State Space Models for Long-Range Spatiotemporal Modeling
Jimmy T.H. Smith (NVIDIA), Wonmin Byeon (NVIDIA)
GenerationComputational EfficiencyConvolutional Neural NetworkVideo
🎯 What it does: A new convolutional state space model (ConvSSM) is proposed for long-term spatiotemporal modeling, combining the advantages of convolutional LSTM and state space methods.
Convolutional Visual Prompt for Robust Visual Perception
Yun-Yun Tsai (Columbia University), Junfeng Yang (Columbia University)
ClassificationDomain AdaptationOptimizationConvolutional Neural NetworkPrompt EngineeringContrastive LearningImage
🎯 What it does: A lightweight Convolutional Visual Prompt (CVP) is proposed to perform self-supervised adaptation on out-of-distribution (OOD) samples without labels during testing, enhancing the robustness of visual models.
Convolutions Die Hard: Open-Vocabulary Segmentation with Single Frozen Convolutional CLIP
Qihang Yu (ByteDance), Liang-Chieh Chen (ByteDance)
Object DetectionSegmentationConvolutional Neural NetworkContrastive LearningImage
🎯 What it does: This paper proposes a single-stage open vocabulary segmentation framework called FC-CLIP, which combines a frozen convolutional CLIP backbone with Mask2Former to achieve mask generation and classification.
Cookie Consent Has Disparate Impact on Estimation Accuracy
Erik Miehling (IBM Research), Robert Nelson Redmond
Recommendation SystemTabular
🎯 What it does: This study investigates how users' choices to accept or reject cookies in different populations affect the accuracy of recommendation systems in estimating users' potential attributes, such as demographic information and interest preferences.
Coop: Memory is not a Commodity
Jianhao Zhang (OneFlow Research), Jinhui Yuan (OneFlow Research)
OptimizationComputational EfficiencyConvolutional Neural NetworkRecurrent Neural NetworkTransformer
🎯 What it does: Proposes Coop, a joint optimization of tensor allocation and recomputation to reduce memory fragmentation.
Coordinating Distributed Example Orders for Provably Accelerated Training
A. Feder Cooper (Cornell University), Christopher De Sa (Cornell University)
OptimizationComputational EfficiencyRecurrent Neural NetworkSupervised Fine-TuningTabularTime Series
🎯 What it does: The CD-GraB algorithm is proposed, which accelerates training by coordinating the gradient order of each worker node in distributed training, utilizing provable permutation-based example sorting.
CoPriv: Network/Protocol Co-Optimization for Communication-Efficient Private Inference
Wenxuan Zeng (Peking University), Ru Huang (Peking University)
OptimizationSafty and PrivacyKnowledge DistillationConvolutional Neural NetworkImage
🎯 What it does: Proposes the CoPriv framework, which jointly optimizes the 2PC inference protocol and DNN architecture to significantly reduce the communication overhead of private inference.
Core-sets for Fair and Diverse Data Summarization
Sepideh Mahabadi (Microsoft Research), Stojan Trajanovski (Microsoft)
Text
🎯 What it does: This paper proposes a composable core-set construction algorithm for approximate solutions to diversity maximization (FDM) tasks under fairness/partition constraints, focusing on three commonly used diversity metrics (minimum pairwise distance, total pairwise distance, and total nearest neighbor distance).
CORNN: Convex optimization of recurrent neural networks for rapid inference of neural dynamics
Fatih Dinc (Stanford University), Hidenori Tanaka (Harvard University)
OptimizationComputational EfficiencyRecurrent Neural NetworkReinforcement LearningTime SeriesSequential
🎯 What it does: A convex optimization framework named CORNN is proposed for the rapid training of data-constrained recurrent neural networks (dRNN), enabling real-time network reconstruction of large-scale neural recordings.
Correlation Aware Sparsified Mean Estimation Using Random Projection
Shuli Jiang (Carnegie Mellon University), Gauri Joshi (Carnegie Mellon University)
Federated LearningComputational EfficiencyAuto EncoderImage
🎯 What it does: A distributed mean estimation encoding-decoding framework based on random projection, called Rand-Proj-Spatial, is proposed to improve the accuracy of mean estimation by utilizing cross-client correlations under communication-constrained conditions.
Correlative Information Maximization: A Biologically Plausible Approach to Supervised Deep Neural Networks without Weight Symmetry
Bariscan Bozkurt (University College London), Alper Tunga Erdogan
ClassificationConvolutional Neural NetworkSupervised Fine-TuningImage
🎯 What it does: A supervised deep neural network framework based on CorInfoMax (Correlation Information Maximization) is proposed, which can solve the weight symmetry problem and achieve a multi-chamber neuron model.
CorresNeRF: Image Correspondence Priors for Neural Radiance Fields
Yixing Lao (University of Hong Kong), Hengshuang Zhao (University of Hong Kong)
GenerationData SynthesisDepth EstimationNeural Radiance FieldImage
🎯 What it does: Utilizing image correspondence relationships under sparse views as a prior, we improve NeRF training to enhance view synthesis and surface reconstruction quality.
Corruption-Robust Offline Reinforcement Learning with General Function Approximation
Chenlu Ye (Hong Kong University of Science and Technology), Tong Zhang (Hong Kong University of Science and Technology)
Reinforcement LearningTabular
🎯 What it does: This paper proposes a robust algorithm to combat data corruption in offline reinforcement learning, applicable to general function approximation (including deep networks).
CosNet: A Generalized Spectral Kernel Network
Yanfang Xue (Southeast University), hui xue
ClassificationRepresentation LearningRecurrent Neural NetworkImageTime Series
🎯 What it does: A complex-valued spectral kernel network (CosNet) is proposed, which extends spectral kernel mapping to the complex domain and embeds it into neural networks to capture long-range or periodic features of time series data.
Counterfactual Conservative Q Learning for Offline Multi-agent Reinforcement Learning
Jianzhun Shao (Tsinghua University), Xiangyang Ji (Tsinghua University)
Reinforcement LearningSequential
🎯 What it does: This paper proposes a Counterfactual Conservative Q Learning (CFCQL) algorithm for offline multi-agent reinforcement learning to address the overestimation problem caused by joint action distribution shift and to achieve offline training for team collaboration.
Counterfactual Evaluation of Peer-Review Assignment Policies
Martin Saveski (University of Washington), Johan Ugander (Stanford University)
🎯 What it does: This paper constructs an offline policy evaluation framework using a random assignment strategy introduced in peer review to assess the impact of different allocation policies on review quality.
Counterfactual Generation with Identifiability Guarantees
Hanqi Yan (University of Warwick), Kun Zhang (Carnegie Mellon University)
GenerationData SynthesisFlow-based ModelAuto EncoderText
🎯 What it does: Achieving counterfactual text generation in an unsupervised environment, a MATTE model is proposed that utilizes relative sparsity to achieve distinguishable content and style.
Counterfactual Memorization in Neural Language Models
Chiyuan Zhang (Google Research), Nicholas Carlini (Google DeepMind)
GenerationTransformerLarge Language ModelText
🎯 What it does: The paper proposes and empirically studies the concept of 'counterfactual memory', which systematically identifies the memory of rare information in neural language models by comparing the prediction differences when a model includes or excludes a certain training sample. It further defines and quantifies 'counterfactual influence' to assess the impact of individual training samples on the validation set and generated text predictions.
Counterfactual-Augmented Importance Sampling for Semi-Offline Policy Evaluation
Shengpu Tang (University of Michigan), Jenna Wiens (University of Michigan)
Reinforcement LearningTabularElectronic Health Records
🎯 What it does: A semi-offline policy evaluation framework is proposed, which improves traditional importance sampling offline evaluation using human-provided counterfactual annotations.
Counterfactually Comparing Abstaining Classifiers
Yo Joong Choe (University of Chicago), Aaditya Ramdas (Carnegie Mellon University)
ClassificationImage
🎯 What it does: This study investigates how to evaluate and compare black-box non-rejecting (abandonable) classifiers and proposes a 'counterfactual score' metric;
Counterfactually Fair Representation
Zhiqun Zuo (Ohio State University), Xueru Zhang (Ohio State University)
Representation LearningAdversarial AttackAuto EncoderGenerative Adversarial NetworkTabular
🎯 What it does: This paper proposes a method to generate adversarial samples using all features and constructs representations that satisfy counterfactual fairness through symmetric functions, thereby training a predictive model that meets perfect counterfactual fairness.
Counting Distinct Elements in the Turnstile Model with Differential Privacy under Continual Observation
Palak Jain (Boston University), Adam Smith (Boston University)
Safty and PrivacyComputational EfficiencySequential
🎯 What it does: This paper studies how to count the number of different elements in a turnstile model with differential privacy under continuous observation, particularly in the case where insertion and deletion operations are allowed in the data stream.
Counting Distinct Elements Under Person-Level Differential Privacy
Thomas Steinke (Google DeepMind), Alexander Knop (Google)
Safty and PrivacyText
🎯 What it does: The study investigates methods for counting distinct elements (vocabulary size) under individual-level differential privacy constraints. It proposes constructing queries with controllable sensitivity by limiting the maximum number of elements contributed by each individual, using maximum flow/matching to solve and adding Laplace noise, and then automatically selecting the best k using the Generalized Exponential Mechanism (GEM) to obtain high-confidence lower bound estimates. A linear-time approximation algorithm is also provided.
Coupled Reconstruction of Cortical Surfaces by Diffeomorphic Mesh Deformation
Hao Zheng (University of Pennsylvania), Yong Fan (University of Pennsylvania)
RestorationSegmentationConvolutional Neural NetworkImageBiomedical DataMagnetic Resonance ImagingAlzheimer's Disease
🎯 What it does: A deep learning framework is proposed that simultaneously generates the white matter and cortical surfaces from a mid-thickness surface using three homotopic deformations, and directly estimates the cortical thickness at each vertex.
Covariance-adaptive best arm identification
El Mehdi Saad (Universite Paris Saclay), Nicolas Verzelen
TabularSequential
🎯 What it does: This paper studies the problem of optimal arm identification with fixed confidence in a multi-armed bandit model that allows simultaneous querying of multiple arms, and provides corresponding algorithms and theoretical analysis.
CP-SLAM: Collaborative Neural Point-based SLAM System
Jiarui Hu (Zhejiang University), Zhaopeng Cui (Zhejiang University)
OptimizationFederated LearningRobotic IntelligenceNeural Radiance FieldSimultaneous Localization and MappingPoint Cloud
🎯 What it does: This paper presents CP-SLAM, a collaborative neural point cloud SLAM system that achieves multi-robot cooperative localization and dense mapping.
CQM: Curriculum Reinforcement Learning with a Quantized World Model
Seungjae Lee (Seoul National University), H. Jin Kim (Seoul National University)
Reinforcement LearningWorld ModelMultimodality
🎯 What it does: A semantic goal space is constructed by automatically defining a quantized world model, and calibrated curriculum goals are generated within this space to guide the RL agent in achieving complex tasks without a pre-specified goal space.
Creating a Public Repository for Joining Private Data
James Cook (Independent Researcher), Nina Mishra (Amazon)
OptimizationSafty and PrivacyTabular
🎯 What it does: This paper proposes a non-interactive data repository scheme based on private count sketch, allowing the sender to publish tables of sensitive attributes in a differentially private manner. The receiver can perform approximate joins on their local data using unique identifiers and complete linear queries and optimizations (such as training machine learning models) based on this.
Creating Multi-Level Skill Hierarchies in Reinforcement Learning
Joshua Benjamin Evans, Özgür Şimşek (University of Bath)
Reinforcement Learning
🎯 What it does: This paper proposes a method for automatically generating multi-layer skill hierarchies based on Louvain modularity maximization, which can directly generate skills, initial sets of skills, execution strategies, termination conditions, and hierarchical dependencies from interaction graphs without the need for manual intervention.
Credal Marginal MAP
Radu Marinescu (IBM Research), Alexander G. Gray
OptimizationGraph
🎯 What it does: This paper studies and defines the Marginal MAP inference task in trustworthy networks (CMMAP) and proposes exact and approximate inference algorithms.
Critical Initialization of Wide and Deep Neural Networks using Partial Jacobians: General Theory and Applications
Darshil Doshi (University of Maryland), Andrey Gromov (Meta)
Image
🎯 What it does: This paper analyzes the properties of gradient propagation in deep neural networks at initialization by introducing the Partial Jacobian and its average norm (APJN), and proposes a low-cost, directly applicable criticality diagnostic method.
CROMA: Remote Sensing Representations with Contrastive Radar-Optical Masked Autoencoders
Anthony Fuller (Carleton University), James R Green
ClassificationSegmentationRepresentation LearningTransformerAuto EncoderContrastive LearningImageMultimodality
🎯 What it does: Proposes the CROMA framework, which utilizes contrastive learning and masked autoencoders to jointly pre-train a remote sensing multimodal (optical and radar) Transformer model, learning single-modal and multimodal representations that can be used for multiple tasks;
Cross-Domain Policy Adaptation via Value-Guided Data Filtering
Kang Xu (Fudan University), Wei Li (Fudan University)
Domain AdaptationReinforcement LearningSequential
🎯 What it does: To address the online dynamic adaptation problem with dynamic differences between the source and target domains, a Value-Guided Data Filtering (VGDF) algorithm based on value consistency for data selection is proposed.
Cross-Episodic Curriculum for Transformer Agents
Lucy Xiaoyang Shi (Stanford University), Yuke Zhu (University of Texas at Austin)
Robotic IntelligenceTransformerReinforcement LearningAgentic AISequential
🎯 What it does: This paper proposes and implements a Cross-Episodic Curriculum (CEC) method that enhances the sample efficiency and generalization ability of Transformer agents by utilizing cross-episode experiences.
Cross-links Matter for Link Prediction: Rethinking the Debiased GNN from a Data Perspective
Zihan Luo (Huazhong University of Science and Technology), Hai Jin (Huazhong University of Science and Technology)
Recommendation SystemGraph Neural NetworkGraph
🎯 What it does: A dual-structure framework is proposed, utilizing supervised enhancement, dual GNNs, and embedding fusion to eliminate the prediction bias between cross-community links and internal links.
Cross-modal Active Complementary Learning with Self-refining Correspondence
Yang Qin (Sichuan University), Peng Hu (Sichuan University)
RetrievalContrastive LearningImageTextMultimodality
🎯 What it does: This paper addresses the issue of noisy correspondence in image-text matching and proposes a general robust framework called CRCL to enhance the performance of matching models in noisy environments.
Cross-modal Prompts: Adapting Large Pre-trained Models for Audio-Visual Downstream Tasks
Haoyi Duan (Zhejiang University), Zhou Zhao (Shanghai Artificial Intelligence Laboratory)
ClassificationRecognitionConvolutional Neural NetworkRecurrent Neural NetworkTransformerPrompt EngineeringContrastive LearningVideoMultimodalityAudio
🎯 What it does: By injecting a bidirectional spatial-channel-temporal attention module DG-SCT into a frozen large-scale pre-trained audio-visual encoder, the use of audio and visual as soft prompts dynamically adjusts feature extraction, significantly improving the performance of multimodal tasks.
Cross-Scale MAE: A Tale of Multiscale Exploitation in Remote Sensing
Maofeng Tang (University of Tennessee), Hairong Qi (University of Tennessee)
ClassificationSegmentationRepresentation LearningTransformerAuto EncoderContrastive LearningImage
🎯 What it does: This paper proposes Cross-Scale MAE, a self-supervised multi-scale remote sensing image representation learning framework that utilizes cross-scale consistency to learn robust features.
CRoSS: Diffusion Model Makes Controllable, Robust and Secure Image Steganography
Jiwen Yu (Peking University), Jian Zhang (Peking University)
GenerationSafty and PrivacyDiffusion modelImage
🎯 What it does: A framework for unencapsulated image steganography based on diffusion models, CRoSS, is proposed, enabling information hiding and recovery without the need for training.
CrossGNN: Confronting Noisy Multivariate Time Series Via Cross Interaction Refinement
Qihe Huang (University of Science and Technology of China), Yang Wang (Suzhou Institute for Advanced Research)
Graph Neural NetworkTime Series
🎯 What it does: This paper proposes CrossGNN, a graph neural network with linear complexity for multivariate time series (MTS) prediction, capable of simultaneously refining cross-scale and cross-variable interactions.
Crystal Structure Prediction by Joint Equivariant Diffusion
Rui Jiao (Tsinghua University), Yang Liu (Tsinghua University)
Graph Neural NetworkDiffusion modelPhysics Related
🎯 What it does: A crystal structure prediction method based on joint equivariant diffusion, called DiffCSP, is proposed;
CS-Isolate: Extracting Hard Confident Examples by Content and Style Isolation
Yexiong Lin (University of Sydney), Tongliang Liu (University of Sydney)
ClassificationAuto EncoderImage
🎯 What it does: Proposes the CS-Isolate framework, which utilizes content-style separation to extract confidently labeled samples that are difficult to classify under label noise and improve classification performance.
CS4ML: A general framework for active learning with arbitrary data based on Christoffel functions
Juan M Cardenas, Nick Dexter (Florida State University)
OptimizationMultimodalityBiomedical DataMagnetic Resonance Imaging
🎯 What it does: A general framework CS4ML is proposed for active learning, supporting arbitrary Hilbert spaces, nonlinear approximation spaces, multimodal data, and any linear sampling operator.
CSLP-AE: A Contrastive Split-Latent Permutation Autoencoder Framework for Zero-Shot Electroencephalography Signal Conversion
Anders Vestergaard Nørskov (Technical University of Denmark), Morten Mørup (Technical University of Denmark)
ClassificationDomain AdaptationRepresentation LearningConvolutional Neural NetworkTransformerAuto EncoderContrastive LearningTime SeriesBiomedical Data
🎯 What it does: A contrastive learning + segmentation latent space autoencoder framework for zero-shot transformation of EEG signals (CSLP-AE) is proposed, achieving single-trial EEG transformation and feature extraction under unseen subjects and tasks.
CSOT: Curriculum and Structure-Aware Optimal Transport for Learning with Noisy Labels
Wanxing Chang (ShanghaiTech University), Jingya Wang (ShanghaiTech University)
ClassificationOptimizationContrastive LearningImage
🎯 What it does: This paper proposes a Curriculum Learning and Structure-Aware Optimal Transport (CSOT) method for denoising and re-labeling training samples in the presence of label noise, thereby enhancing the model's robustness and generalization performance.
Curriculum Learning for Graph Neural Networks: Which Edges Should We Learn First
Zheng Zhang (Emory University), Liang Zhao (Emory University)
OptimizationRepresentation LearningGraph Neural NetworkGraph
🎯 What it does: A relationship curriculum learning (RCL) method for graph neural networks is proposed, which improves representation learning by gradually adding edges from easy to difficult.
Curriculum Learning With Infant Egocentric Videos
Saber Sheybani (Indiana University), Zoran Tiganj (Indiana University)
Representation LearningTransformerContrastive LearningVideo
🎯 What it does: The study evaluates the impact of developmental sequences on visual representation learning using first-person videos collected from infants aged 2-12 months wearing head-mounted cameras, employing a self-supervised learning model.
Curvature Filtrations for Graph Generative Model Evaluation
Joshua Southern (Imperial College London), Bastian Rieck (Helmholtz Munich)
GenerationData SynthesisOptimizationGraph Neural NetworkGraphBiomedical Data
🎯 What it does: A curvature filtering method that combines discrete curvature with topological data analysis is proposed to evaluate the distribution distance of graph generative models.
Curve Your Enthusiasm: Concurvity Regularization in Differentiable Generalized Additive Models
Julien Niklas Siems, Martin Genzel (Merantix Momentum)
OptimizationExplainability and InterpretabilityTabularTime Series
🎯 What it does: A differentiable concurvity regularization method is proposed to enhance the interpretability of Generalized Additive Models (GAM).
Customizable Image Synthesis with Multiple Subjects
Zhiheng Liu (University of Science and Technology of China), Yang Cao (University of Science and Technology of China)
GenerationData SynthesisDiffusion modelImage
🎯 What it does: A method for multi-subject customized image synthesis is proposed, which can be implemented without retraining. It utilizes a pre-trained diffusion model, subject residual word vectors, and layout guidance to achieve arbitrary subject combinations.
CWCL: Cross-Modal Transfer with Continuously Weighted Contrastive Loss
Rakshith Sharma Srinivasa (Samsung Research America), Hongxia Jin (Samsung Research America)
RetrievalDomain AdaptationRepresentation LearningTransformerContrastive LearningImageTextMultimodalityAudio
🎯 What it does: A new cross-modal contrastive loss function, CWCL, is proposed, which utilizes the similarity information from a frozen pre-trained model in one modality to achieve representation learning in another modality, thereby enabling zero-shot transfer across modalities.
CycleNet: Rethinking Cycle Consistency in Text-Guided Diffusion for Image Manipulation
Sihan Xu (University of Michigan), Joyce Chai (University of Michigan)
Image TranslationGenerationDiffusion modelImageText
🎯 What it does: CycleNet is proposed, a framework for unpaired image-to-image translation that incorporates cyclic consistency regularization on a pre-trained text-guided latent diffusion model, achieving image + text dual-condition translation through a single ControlNet side network; it also contributes the multi-domain state change dataset ManiCups; additionally, a FastCycleNet variant for rapid training is provided.
D-CIPHER: Discovery of Closed-form Partial Differential Equations
Krzysztof Kacprzyk (University of Cambridge), Mihaela van der Schaar (University of Cambridge)
Time SeriesPhysics RelatedOrdinary Differential Equation
🎯 What it does: A new method called D-CIPHER is proposed for discovering closed-form partial differential equations (PDEs) and their higher-order ordinary differential equations (ODEs) from noisy sparse data.
D-Separation for Causal Self-Explanation
Wei Liu (Huazhong University of Science and Technology), Yang Qiu (Huazhong University of Science and Technology)
ClassificationExplainability and InterpretabilityText
🎯 What it does: A new self-explanatory framework called MCD is proposed, aiming to find causal rationalization by minimizing conditional dependence.
D$^2$CSG: Unsupervised Learning of Compact CSG Trees with Dual Complements and Dropouts
Fenggen Yu (Simon Fraser University), Hao Zhang (Simon Fraser University)
GenerationData SynthesisPoint CloudMesh
🎯 What it does: This paper proposes the D2 CSG model, which achieves unsupervised learning of a single CAD shape through a dual-branch structure and complementary primitives, generating a compact CSG tree.
D4Explainer: In-distribution Explanations of Graph Neural Network via Discrete Denoising Diffusion
Jialin Chen (Yale University), Zhitao Ying
Explainability and InterpretabilityAdversarial AttackGraph Neural NetworkDiffusion modelGraphTabular
🎯 What it does: A unified framework D4Explainer based on discrete denoising diffusion models is proposed for generating counterfactual and model-level explanations for graph neural networks.
DAC-DETR: Divide the Attention Layers and Conquer
Zhengdong Hu (University of Technology Sydney), Yi Yang (Baidu Inc.)
Object DetectionTransformerImage
🎯 What it does: It was found that the effects of cross-attention and self-attention on queries in the DETR decoder are opposing, and an auxiliary decoder without self-attention (DAC-DETR) is proposed to separate and train these two types of attention independently to improve training efficiency and detection accuracy.
DAMEX: Dataset-aware Mixture-of-Experts for visual understanding of mixture-of-datasets
Yash Jain (Georgia Institute of Technology), Vibhav Vineet (Microsoft Research)
Object DetectionDomain AdaptationTransformerMixture of ExpertsImageBenchmark
🎯 What it does: This paper proposes Dataset-aware Mixture-of-Experts (DAMEX), which replaces part of the FFN in the DINO detection framework with MoE and trains a router to focus each expert on the corresponding dataset, thereby achieving the fusion and general detection of multiple datasets within a single model.
DäRF: Boosting Radiance Fields from Sparse Input Views with Monocular Depth Adaptation
Jiuhn Song (Korea University), Seungryong Kim (Korea University)
Data SynthesisDepth EstimationNeural Radiance FieldImage
🎯 What it does: A robust 3D reconstruction framework called DRF a is proposed, which is based on the online complementary training of sparse-view neural radiance fields (NeRF) and a pre-trained monocular depth estimation network (MDE) using a small number of real images.
DASpeech: Directed Acyclic Transformer for Fast and High-quality Speech-to-Speech Translation
Qingkai Fang (Chinese Academy of Sciences), Yang Feng (Chinese Academy of Sciences)
Knowledge DistillationTransformerMultimodalityAudio
🎯 What it does: A non-autoregressive two-step direct speech-to-speech translation model DASpeech is designed, which first generates the target text and then generates the target speech.
Data Augmentations for Improved (Large) Language Model Generalization
Amir Feder (Columbia University), David Blei (Columbia University)
ClassificationDomain AdaptationTransformerLarge Language ModelPrompt EngineeringTextBiomedical DataElectronic Health Records
🎯 What it does: This paper proposes a causal structure-based adversarial data augmentation method called CATO, which utilizes auxiliary data and large language models to generate counterfactual texts, enhancing the out-of-distribution generalization performance of text classification models.
Data Market Design through Deep Learning
Sai Srivatsa Ravindranath (Harvard John A. Paulson School of Engineering and Applied Sciences), David C. Parkes (Harvard John A. Paulson School of Engineering and Applied Sciences)
OptimizationExplainability and Interpretability
🎯 What it does: This paper proposes a mechanism for automating the design of data markets using deep learning (RochetNet and RegretNet), which can maximize the expected revenue of information sellers in both single-buyer and multi-buyer scenarios, and can verify known theoretical solutions, extend to more complex settings, and infer optimal design structures.
Data Minimization at Inference Time
Cuong Tran (University of Virginia), Ferdinando Fioretto (University of Virginia)
Safty and PrivacyData-Centric LearningTabularFinance Related
🎯 What it does: This paper proposes a personalized data minimization method that achieves accurate predictions by disclosing only a small number of sensitive features during inference.
Data Pruning via Moving-one-Sample-out
Haoru Tan (Hong Kong University), XIAOJUAN QI
ClassificationComputational EfficiencyData-Centric LearningImage
🎯 What it does: A novel data pruning method called MoSo is proposed, which evaluates the importance of samples by measuring the impact of removing a single sample on empirical risk, and provides an approximate estimate with linear complexity.
Data Quality in Imitation Learning
Suneel Belkhale (Stanford University), Dorsa Sadigh (Stanford University)
Robotic IntelligenceReinforcement Learning from Human FeedbackTransformerTabular
🎯 What it does: This paper studies the definition and evaluation of data quality in imitation learning, and proposes measuring data quality by minimizing distribution shift.
Data Selection for Language Models via Importance Resampling
Sang Michael Xie (Stanford University), Percy Liang (Stanford University)
Domain AdaptationLarge Language ModelText
🎯 What it does: This paper proposes a data selection framework called DSIR based on importance resampling, aimed at selecting data that conforms to the target distribution from a vast amount of unlabeled text.
Data-Centric Learning from Unlabeled Graphs with Diffusion Model
Gang Liu (University of Notre Dame), Meng Jiang (University of Notre Dame)
ClassificationData-Centric LearningGraph Neural NetworkDiffusion modelGraph
🎯 What it does: Utilizing diffusion models to transfer knowledge from unlabeled graph data to downstream graph attribute prediction tasks through data augmentation, generating task-specific labeled graph samples;
Data-Dependent Bounds for Online Portfolio Selection Without Lipschitzness and Smoothness
Chung-En Tsai (National Taiwan University), Yen-Huan Li (National Taiwan University)
OptimizationFinance Related
🎯 What it does: Two new algorithms for online portfolio selection (OPS) are proposed, and asymptotic bounds on gradual variation and small-loss data-dependent losses are provided under the condition of not satisfying the no-junk-bonds assumption.
Data-driven Optimal Filtering for Linear Systems with Unknown Noise Covariances
Shahriar Talebi (Harvard University), Mehran Mesbahi (University of Washington)
OptimizationReinforcement LearningTime Series
🎯 What it does: Through the stochastic policy optimization based on the estimation-control duality, the steady-state Kalman gain of linear systems is learned directly from observed output data using stochastic gradient descent;
Data-Informed Geometric Space Selection
Shuai Zhang (ETH Zurich), Wenqi Jiang (ETH Zurich)
Recommendation SystemMixture of ExpertsContrastive LearningGraph
🎯 What it does: This paper proposes an adaptive geometric space selection method based on sparse gated MOE, allowing each data point to automatically select the most suitable K geometric spaces and form a Cartesian product space for representation.
DaTaSeg: Taming a Universal Multi-Dataset Multi-Task Segmentation Model
Xiuye Gu (Google Research), David A Ross
SegmentationTransformerContrastive LearningImageBenchmark
🎯 What it does: This paper presents DaTaSeg, a unified image segmentation model for multiple datasets and multiple tasks.
Dataset Diffusion: Diffusion-based Synthetic Data Generation for Pixel-Level Semantic Segmentation
Quang Ho Nguyen (VinAI Research), Khoi Nguyen (VinAI Research)
SegmentationGenerationData SynthesisPrompt EngineeringDiffusion modelImage
🎯 What it does: Utilize Stable Diffusion to generate synthetic images with pixel-level semantic segmentation labels, thereby constructing a dataset that can be directly used to train semantic segmentation models.
DatasetDM: Synthesizing Data with Perception Annotations Using Diffusion Models
Weijia Wu (Zhejiang University), Chunhua Shen (Zhejiang University)
SegmentationGenerationData SynthesisPose EstimationDepth EstimationTransformerLarge Language ModelPrompt EngineeringDiffusion modelImage
🎯 What it does: We propose DatasetDM, a text-to-data generation framework based on pre-trained diffusion models, capable of infinitely generating high-quality synthetic images along with corresponding perceptual annotations (semantic segmentation, instance segmentation, depth maps, human pose, etc.).
DAW: Exploring the Better Weighting Function for Semi-supervised Semantic Segmentation
Rui Sun (University of Science and Technology of China), Feng Wu (University of Science and Technology of China)
SegmentationConvolutional Neural NetworkImage
🎯 What it does: A semi-supervised semantic segmentation method based on Distribution-Aware Weighting (DAW) is proposed, which systematically analyzes the trade-off between the use and discard of pseudo-labels, and derives the optimal weighting function by explicitly modeling the confidence distribution.
DDCoT: Duty-Distinct Chain-of-Thought Prompting for Multimodal Reasoning in Language Models
Ge Zheng (ShanghaiTech University), Sibei Yang (ShanghaiTech University)
Explainability and InterpretabilityTransformerLarge Language ModelPrompt EngineeringVision Language ModelTextMultimodalityChain-of-Thought
🎯 What it does: A Duty-Distinct Chain-of-Thought (DDCoT) prompting framework is proposed, utilizing negative space prompts, visual recognition, and joint reasoning to generate general multimodal explanations, which are applied to zero-shot reasoning and fine-tuning of language models.
DDF-HO: Hand-Held Object Reconstruction via Conditional Directed Distance Field
Chenyangguang Zhang (Tsinghua University), Xiangyang Ji (Tsinghua University)
Object DetectionGenerationPose EstimationImagePoint CloudMesh
🎯 What it does: The DDF-HO method is proposed, which constructs a 3D model of handheld objects using a single RGB image and hand pose information.
De novo Drug Design using Reinforcement Learning with Multiple GPT Agents
Xiuyuan Hu (Tsinghua University), Hao Zhang (Microsoft Research)
Drug DiscoveryTransformerLarge Language ModelReinforcement LearningBiomedical Data
🎯 What it does: A multi-agent reinforcement learning framework MolRL-MGPT is proposed, using the GPT generative model for de novo drug molecule design, encouraging multi-agent collaborative exploration of diverse chemical spaces.
Debias Coarsely, Sample Conditionally: Statistical Downscaling through Optimal Transport and Probabilistic Diffusion Models
Zhong Yi Wan (Google Research), Leonardo Zepeda-Nunez
GenerationData SynthesisDiffusion modelScore-based ModelTime SeriesPhysics RelatedStochastic Differential Equation
🎯 What it does: A two-stage probabilistic framework was developed to perform statistical downsampling using unpaired data. First, optimal transport is used to remove the bias of the low-resolution model, and then a conditional diffusion model is employed for upsampling to generate high-resolution flow field samples.
Debiased and Denoised Entity Recognition from Distant Supervision
Haobo Wang (Zhejiang University), Junbo Zhao (Zhejiang University)
RecognitionTransformerSupervised Fine-TuningText
🎯 What it does: The DesERT framework is proposed to improve remote supervision-based named entity recognition through dual-head decoupled learning and debiased self-training.
Debiasing Conditional Stochastic Optimization
Lie He (École Polytechnique Fédérale de Lausanne), Shiva Kasiviswanathan
OptimizationTabular
🎯 What it does: This paper studies Conditional Stochastic Optimization (CSO) and its finite sum variant (FCCO), and proposes a bias-corrected gradient estimation method based on stochastic extrapolation.
Debiasing Pretrained Generative Models by Uniformly Sampling Semantic Attributes
Walter Gerych (Worcester Polytechnic Institute), Emmanuel Agu (Worcester Polytechnic Institute)
GenerationData SynthesisDiffusion modelGenerative Adversarial NetworkImage
🎯 What it does: This paper proposes a method to debias pre-trained generative models without retraining the generative model and without using real data. It utilizes existing pre-trained generative models and potentially biased attribute classifiers to construct a fair noise distribution and train a distribution mapper, enabling the generative model to produce samples that are uniformly distributed across various attribute categories.
Debiasing Scores and Prompts of 2D Diffusion for View-consistent Text-to-3D Generation
Susung Hong (Korea University), Seungryong Kim (Korea University)
GenerationData SynthesisLarge Language ModelDiffusion modelScore-based ModelText
🎯 What it does: This paper proposes two debiasing methods—score debiasing (dynamic truncation-based gradient clipping) and prompt debiasing (using MLM language models to identify and eliminate conflicting words in prompts)—to address the perspective inconsistency (Janus) problem in text-to-3D generation.
Decentralized Matrix Sensing: Statistical Guarantees and Fast Convergence
Marie Maros (Purdue University), Gesualdo Scutari (Purdue University)
OptimizationGraph
🎯 What it does: The paper proposes a decentralized matrix sensing algorithm implemented in a serverless network, capable of recovering low-rank matrices in a distributed environment and providing statistical and convergence guarantees.
Decentralized Randomly Distributed Multi-agent Multi-armed Bandit with Heterogeneous Rewards
Mengfan Xu (Northwestern University), Diego Klabjan (Northwestern University)
OptimizationFederated LearningGraph Neural NetworkReinforcement LearningGraph
🎯 What it does: The study focuses on decentralized multi-agent multi-armed bandits and designs the DrFed-UCB algorithm under randomly changing graphs and heterogeneous rewards.