ICLR 2023 Papers — Page 10
International Conference on Learning Representations · 1573 papers
Meta-prediction Model for Distillation-Aware NAS on Unseen Datasets
Hayeon Lee (Korea Advanced Institute of Science and Technology), Sung Ju Hwang (Korea Advanced Institute of Science and Technology)
Knowledge DistillationMeta LearningNeural Architecture SearchImage
🎯 What it does: This paper proposes a meta-prediction model named DaSS, which is used to quickly predict the final accuracy of the student network on unseen datasets in the knowledge distillation (KD) scenario, thereby achieving efficient neural architecture search (DaNAS) across datasets and teachers.
Metadata Archaeology: Unearthing Data Subsets by Leveraging Training Dynamics
Shoaib Ahmed Siddiqui (University of Cambridge), Sara Hooker (Cohere for Artificial Intelligence)
ClassificationData-Centric LearningConvolutional Neural NetworkImage
🎯 What it does: This paper studies a unified metadata archaeology method called MAP-D, which infers the metadata labels of samples by utilizing the learning dynamics during the model training process.
MetaGL: Evaluation-Free Selection of Graph Learning Models via Meta-Learning
Namyong Park (Carnegie Mellon University), Christos Faloutsos (Carnegie Mellon University)
Meta LearningGraph Neural NetworkGraph
🎯 What it does: Proposed METAGL, a model selection method for graph learning without evaluation;
MICN: Multi-scale Local and Global Context Modeling for Long-term Series Forecasting
Huiqiang Wang (Sichuan University), Yifei Xiao (University of Electronic Science and Technology of China)
Convolutional Neural NetworkTime Series
🎯 What it does: This paper proposes a Multi-Scale Isometric Convolutional Network (MICN), which extracts short-term features through local convolution and models global correlations using isometric convolution, predicting and fusing after decomposing seasonal and trend components.
Mid-Vision Feedback
Michael Maynord (University of Maryland), Yiannis Aloimonos (University of Maryland)
ClassificationConvolutional Neural NetworkTransformerContrastive LearningImage
🎯 What it does: A mid-vision feedback mechanism based on context is proposed, which achieves adaptive adjustment of high-level category expectations through linear transformations on mid-level features.
MIMT: Masked Image Modeling Transformer for Video Compression
Jinxi Xiang (Tencent AI Lab), Jun Zhang (Tencent AI Lab)
CompressionTransformerOptical FlowVideo
🎯 What it does: This paper studies an autoregressive unordered entropy model based on Masked Image Modeling Transformer (MIMT) for deep video compression.
Min-Max Multi-objective Bilevel Optimization with Applications in Robust Machine Learning
Alex Gu (Massachusetts Institute of Technology), Tsui-Wei Weng (University of California San Diego)
OptimizationRepresentation LearningHyperparameter SearchImage
🎯 What it does: This paper studies and addresses the min-max multi-objective bilevel optimization problem, proposing the single-loop gradient descent-ascent algorithm MORBiT to find robust solutions in robust machine learning tasks such as representation learning and hyperparameter optimization.
Mind the Gap: Offline Policy Optimization for Imperfect Rewards
Jianxiong Li (Tsinghua University), Ya-Qin Zhang (Tsinghua University)
OptimizationReinforcement LearningGenerative Adversarial NetworkTabular
🎯 What it does: A Reward Gap Minimization (RGM) framework is proposed, which can simultaneously correct rewards and learn optimal policies in offline reinforcement learning scenarios with imperfect rewards.
Mind the Pool: Convolutional Neural Networks Can Overfit Input Size
Bilal Alsallakh (Voxel AI), Pamela Bhattacharya (Meta)
ClassificationObject DetectionSegmentationConvolutional Neural NetworkImage
🎯 What it does: This paper studies the problem of overfitting in convolutional neural networks under different input sizes and proposes Spatially Balanced Pooling (SBPool) to alleviate this phenomenon.
Mind's Eye: Grounded Language Model Reasoning through Simulation
Ruibo Liu (Google Research), Andrew M. Dai (Google Research)
TransformerLarge Language ModelPrompt EngineeringTextBenchmarkPhysics RelatedRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Proposes the Mind's Eye framework: automatically converts natural language questions into MuJoCo rendering code, runs physical simulations to obtain results, and injects these results into a large language model (LLM) as prompts, thereby achieving physics-based 'embodied' reasoning.
Mini-batch $k$-means terminates within $O(d/\epsilon)$ iterations
Gregory Schwartzman (Japan Advanced Institute of Science and Technology)
Optimization
🎯 What it does: The study proves that under appropriate batch sizes, mini-batch k-means can terminate within O(d/ε) iterations with high probability when using early stopping conditions.
Minimalistic Unsupervised Representation Learning with the Sparse Manifold Transform
Yubei Chen (Meta), Yann LeCun (Meta)
Representation LearningContrastive LearningImage
🎯 What it does: A minimalistic and interpretable unsupervised representation learning method is proposed, which does not require data augmentation, hyperparameter tuning, or other engineering designs, yet achieves performance close to state-of-the-art self-supervised learning methods.
Minimax Optimal Kernel Operator Learning via Multilevel Training
Jikai Jin (Peking University), Lexing Ying (Stanford University)
🎯 What it does: This study investigates the problem of learning Hilbert–Schmidt operators between infinite-dimensional Sobolev RKHS, provides information-theoretic lower bounds, and designs a multi-layer kernel operator learning algorithm that achieves optimal learning rates.
Minimum Description Length Control
Ted Moskovitz (University College London), Matthew Botvinick
OptimizationReinforcement LearningSequential
🎯 What it does: A multi-task reinforcement learning framework called MDL-Control is proposed based on the Minimum Description Length (MDL) principle, which learns shared structures between tasks through a default policy during training and compresses this policy using a sparse prior.
Minimum Variance Unbiased N:M Sparsity for the Neural Gradients
Brian Chmiel (Habana Labs), Daniel Soudry (Technion)
OptimizationComputational EfficiencyConvolutional Neural NetworkTransformerSupervised Fine-TuningImageText
🎯 What it does: The research utilizes N:M fine-grained sparsity for unbiased minimum variance pruning of neural gradients during the training process, thereby accelerating the GEMM in the gradient update phase.
Mitigating Dataset Bias by Using Per-Sample Gradient
Sumyeong Ahn (Korea Advanced Institute of Science and Technology), Se-Young Yun
ClassificationData-Centric LearningImageText
🎯 What it does: A resampling algorithm based on the gradient norm of each sample (PGD) is proposed, which alleviates dataset bias without human labels by first training a bias model, then calculating the gradient norm and sampling according to its proportion.
Mitigating Gradient Bias in Multi-objective Learning: A Provably Convergent Approach
Heshan Devaka Fernando (Rensselaer Polytechnic Institute), Tianyi Chen (IBM Thomas J. Watson Research Center)
SegmentationDepth EstimationOptimizationReinforcement LearningImage
🎯 What it does: A stochastic multi-objective gradient correction method (MoCo) is proposed to address the bias problem of traditional MGDA and its variants under stochastic gradients, ensuring convergence to Pareto points in the case of non-convex objectives and fixed batch sizes.
Mitigating Memorization of Noisy Labels via Regularization between Representations
Hao Cheng (University of California, Santa Cruz), Yang Liu (University of California, Santa Cruz)
ClassificationRepresentation LearningConvolutional Neural NetworkContrastive LearningImage
🎯 What it does: The paper introduces a regularization based on representation (self-supervised features) in deep networks to suppress the memorization phenomenon when learning from noisy labels.
MixPro: Data Augmentation with MaskMix and Progressive Attention Labeling for Vision Transformer
Qihao Zhao (Beijing University of Chemical Technology), Jun Liu (Singapore University of Technology and Design)
Object DetectionSegmentationTransformerImage
🎯 What it does: This paper proposes MixPro, a data augmentation method for Vision Transformers based on MaskMix (adjustable size mask mixed images) and Progressive Attention Labeling (PAL).
MLPInit: Embarrassingly Simple GNN Training Acceleration with MLP Initialization
Xiaotian Han (Texas A&M University), Neil Shah (Snap Inc)
OptimizationComputational EfficiencyGraph Neural NetworkGraph
🎯 What it does: A weight initialization method for peer MLP based on converged weights (MLPInit) is proposed, which directly uses the trained PeerMLP weights as initialization for GNNs, thereby accelerating GNN training and improving prediction performance.
MMVAE+: Enhancing the Generative Quality of Multimodal VAEs without Compromises
Emanuele Palumbo (ETH Zurich), Julia E Vogt
GenerationData SynthesisMixture of ExpertsAuto EncoderImageTextMultimodality
🎯 What it does: A MMVAE+ model is proposed, which integrates shared and private latent subspaces to improve the generative quality and semantic consistency of multimodal VAE.
MOAT: Alternating Mobile Convolution and Attention Brings Strong Vision Models
Chenglin Yang (Johns Hopkins University), Liang-Chieh Chen (Google Research)
ClassificationObject DetectionSegmentationConvolutional Neural NetworkTransformerImage
🎯 What it does: The MOAT (Mobile Convolution with Attention) module is proposed, which combines the MBConv (inverted bottleneck) of MobileNetV2 with the self-attention of Transformer to construct a series of models ranging from small to large.
MocoSFL: enabling cross-client collaborative self-supervised learning
Jingtao Li (Arizona State University), Michael Spranger (Sony AI)
Federated LearningComputational EfficiencyContrastive LearningImage
🎯 What it does: This paper proposes MocoSFL, a cross-client self-supervised learning framework based on Split Federated Learning, which addresses bottlenecks in computation, memory, and data volume while achieving high accuracy.
Model ensemble instead of prompt fusion: a sample-specific knowledge transfer method for few-shot prompt tuning
XIANGYU PENG, Caiming Xiong (Salesforce Research)
TransformerPrompt EngineeringText
🎯 What it does: In the few-shot prompt tuning scenario, a Sample-Specific Source Model Ensemble (SESoM) is proposed, which utilizes a soft prompt model pre-trained on multiple source tasks to perform sample-weighted fusion on the target task, thereby enhancing few-shot performance.
Model-based Causal Bayesian Optimization
Scott Sussex (ETH Zurich), Andreas Krause (ETH Zurich)
OptimizationGraph Neural NetworkReinforcement LearningGraphBenchmark
🎯 What it does: A Bayesian optimization algorithm MCBO based on a complete Structural Causal Model (SCM) is proposed, utilizing Gaussian processes to model each node and making optimal intervention decisions across the entire model space through the UCB principle, providing a non-asymptotic sublinear upper bound on cumulative regret.
Modeling content creator incentives on algorithm-curated platforms
Jiri Hron (University of Cambridge), Sarah Dean (Cornell University)
Recommendation SystemTabular
🎯 What it does: This paper proposes and studies the exposure game model to characterize the incentive mechanisms for content creators in embedded recommendation systems, and evaluates the impact of creator behavior through a pre-deployment audit assessment algorithm.
Modeling Multimodal Aleatoric Uncertainty in Segmentation with Mixture of Stochastic Experts
Zhitong Gao (ShanghaiTech University), Xuming He (ShanghaiTech University)
SegmentationMixture of ExpertsMultimodalityBiomedical DataComputed Tomography
🎯 What it does: A multi-modal uncertainty modeling framework MoSE is proposed, which captures the inherent stochastic uncertainty of data in semantic segmentation and outputs multiple weighted segmentation results.
Modeling Sequential Sentence Relation to Improve Cross-lingual Dense Retrieval
Shunyu Zhang (Microsoft Research Asia), Nan Duan (Microsoft Research Asia)
RetrievalTransformerContrastive LearningText
🎯 What it does: A multilingual pre-training model for cross-language dense retrieval, called Masked Sentence Model (MSM), is proposed to improve cross-language representation learning by modeling the sequential relationships at the sentence level.
Modeling the Data-Generating Process is Necessary for Out-of-Distribution Generalization
Jivat Neet Kaur (Microsoft Research), Amit Sharma (Microsoft Research)
Domain AdaptationImage
🎯 What it does: Researches the domain generalization problem under multi-attribute distribution shift and proposes the Causal Graph-based Adaptive Constraint Minimization (CACM) algorithm;
Modelling Long Range Dependencies in $N$D: From Task-Specific to a General Purpose CNN
David M Knigge, Jan-jakob Sonke
ClassificationRecognitionOptimizationComputational EfficiencyConvolutional Neural NetworkTransformerImagePoint CloudSequentialBenchmark
🎯 What it does: A Continuous Convolutional Neural Network (CCNN) is proposed, which can handle data of arbitrary resolution, dimension, and length without changing the network structure.
MoDem: Accelerating Visual Model-Based Reinforcement Learning with Demonstrations
Nicklas Hansen (University of California San Diego), Aravind Rajeswaran (Meta AI)
Robotic IntelligenceReinforcement LearningAgentic AIImage
🎯 What it does: This study investigates how to accelerate visual model-based reinforcement learning using a small number of demonstrations, significantly improving sampling efficiency in sparse reward tasks.
Moderate Coreset: A Universal Method of Data Selection for Real-world Data-efficient Deep Learning
Xiaobo Xia (University of Sydney), Tongliang Liu (University of Sydney)
ClassificationData-Centric LearningConvolutional Neural NetworkImage
🎯 What it does: The concept of 'Moderate Coreset' and its implementation method are proposed, using deep representation Euclidean distance and score median to construct a data subset to improve model performance in efficient data learning.
Mole-BERT: Rethinking Pre-training Graph Neural Networks for Molecules
Jun Xia (Westlake University), Stan Z. Li (Westlake University)
Drug DiscoveryGraph Neural NetworkContrastive LearningGraph
🎯 What it does: We propose Mole-BERT, which significantly enhances molecular property prediction and retrieval performance by introducing a context-aware tokenizer based on VQ-VAE and new Masked Atoms Modeling along with Triplet Masked Contrastive Learning for pre-training molecular GNNs.
Molecular Geometry Pretraining with SE(3)-Invariant Denoising Distance Matching
Shengchao Liu (Mila - Quebec AI Institute Universite de Montreal), Jian Tang (Mila - Quebec AI Institute)
Representation LearningDrug DiscoveryScore-based ModelPoint Cloud
🎯 What it does: A SE(3) invariant molecular geometry pre-training framework GeoSSL-DDM is proposed, utilizing self-supervised learning methods for molecular coordinate denoising and distance denoising;
Molecule Generation For Target Protein Binding with Structural Motifs
ZAIXI ZHANG, Qi Liu (University of Science and Technology of China)
GenerationDrug DiscoveryGraph Neural NetworkGraphBiomedical Data
🎯 What it does: A fragment-based 3D molecular generation framework called FLAG is proposed for structure-based drug design.
Momentum Stiefel Optimizer, with Applications to Suitably-Orthogonal Attention, and Optimal Transport
Lingkai Kong (Georgia Institute of Technology), Molei Tao (Georgia Institute of Technology)
OptimizationTransformerImageTextStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: This paper proposes a Stiefel matrix optimizer based on continuous-discrete dynamics, which features natural momentum, structure preservation, and low overhead, and applies it to the orthogonalization of ViT attention heads and low-dimensional projections of high-dimensional optimal transport (OT).
Monocular Scene Reconstruction with 3D SDF Transformers
Weihao Yuan (Alibaba Group), Siyu Zhu (Alibaba Group)
SegmentationPose EstimationTransformerPoint CloudMesh
🎯 What it does: Using a 3D SDF Transformer to achieve scene reconstruction from monocular camera pose images, directly predicting TSDF voxels and generating complete meshes through marching cubes.
More Centralized Training, Still Decentralized Execution: Multi-Agent Conditional Policy Factorization
Jiangxing Wang (Peking University), Zongqing Lu (Peking University)
Reinforcement LearningBenchmark
🎯 What it does: A multi-agent conditional policy decomposition (MACPF) framework is proposed, which explicitly considers inter-agent dependencies under centralized training while achieving decentralized execution.
More ConvNets in the 2020s: Scaling up Kernels Beyond 51x51 using Sparsity
Shiwei Liu (University of Texas at Austin), Zhangyang Wang (University of Texas at Austin)
Object DetectionSegmentationConvolutional Neural NetworkImage
🎯 What it does: This paper proposes a method using sparsification and kernel decomposition, successfully expanding the convolution kernel size from 31×31 to 51×51 and even 61×61, and constructs the SLaK network.
Mosaic Representation Learning for Self-supervised Visual Pre-training
Zhaoqing Wang (Sydney AI Centre, University of Sydney), Tongliang Liu (University of Sydney)
Object DetectionSegmentationRepresentation LearningConvolutional Neural NetworkContrastive LearningImage
🎯 What it does: The Mosaic Representation Learning framework MosRep is proposed, which enriches the background information of small crops by introducing mosaic views in self-supervised visual pre-training, thereby enhancing the quality of visual representations.
Moving Forward by Moving Backward: Embedding Action Impact over Action Semantics
Kuo-Hao Zeng (University of Washington), Ali Farhadi (University of Washington)
Robotic IntelligenceRecurrent Neural NetworkTransformerReinforcement LearningSequential
🎯 What it does: An Action Adaptive Policy (AAP) is proposed, which enables navigation agents to adapt to unknown action effects by learning the embeddings of action impacts in real-time during inference.
MPCFORMER: FAST, PERFORMANT AND PRIVATE TRANSFORMER INFERENCE WITH MPC
Dacheng Li (Carnegie Mellon University), Hao Zhang (University of California, Berkeley)
Safty and PrivacyComputational EfficiencyKnowledge DistillationTransformerLarge Language ModelText
🎯 What it does: A Transformer inference framework called MPCFORMER is proposed, which utilizes MPC and knowledge distillation to significantly accelerate private inference while maintaining performance.
Multi-domain image generation and translation with identifiability guarantees
Shaoan Xie (Carnegie Mellon University), Kun Zhang (Mohamed bin Zayed University of Artificial Intelligence)
Image TranslationGenerationFlow-based ModelGenerative Adversarial NetworkImage
🎯 What it does: This paper proposes a multi-domain image generation and unsupervised image-to-image translation method based on recognizability constraints, which can learn the complete joint distribution from single-domain edge distributions, thereby generating image pairs that are content-consistent and stylistically diverse.
Multi-level Protein Structure Pre-training via Prompt Learning
Zeyuan Wang (Zhejiang University), Huajun Chen (Zhejiang University)
Protein Structure PredictionTransformerPrompt EngineeringBiomedical Data
🎯 What it does: A prompt-based multi-task pre-training framework called PromptProtein is proposed, which utilizes protein sequences to learn multi-layer structural information and achieves flexible utilization in downstream tasks through prompt combinations.
Multi-lingual Evaluation of Code Generation Models
Ben Athiwaratkun (Amazon Web Services AI Labs), Bing Xiang (Amazon Web Services AI Labs)
GenerationAI Code AssistantTransformerLarge Language ModelTextBenchmark
🎯 What it does: Developed a multilingual execution evaluation benchmark MBXP and Multilingual HumanEval, transforming the original Python dataset into multiple languages through a rule-driven conversion framework, and evaluating the functional correctness of code generation models based on this.
Multi-Objective Online Learning
Jiyan Jiang (Tsinghua University), Wenwu Zhu (Tsinghua University)
OptimizationBiomedical Data
🎯 What it does: This paper proposes a multi-objective online convex optimization framework and introduces a new serialized Pareto sub-optimality gap measure along with a corresponding definition of regret.
Multi-objective optimization via equivariant deep hypervolume approximation
Jim Boelrijk (University of Amsterdam), Patrick Forré (University of Amsterdam)
Optimization
🎯 What it does: Proposes DeepHV, which approximates the hypervolume of multi-objective optimization using deep learning.
Multi-Objective Reinforcement Learning: Convexity, Stationarity and Pareto Optimality
Haoye Lu (University of Waterloo), Yaoliang Yu (University of Waterloo)
OptimizationReinforcement LearningSequential
🎯 What it does: This paper rigorously analyzes the convexity and Pareto optimality of the policy-induced value function in multi-objective reinforcement learning, proposes the CAPQL algorithm that enhances concavity in rewards, and verifies its superiority.
Multi-Rate VAE: Train Once, Get the Full Rate-Distortion Curve
Juhan Bae (University of Toronto), Roger Baker Grosse
GenerationCompressionAuto EncoderImageText
🎯 What it does: A variational autoencoder (Multi-Rate VAE) is proposed that can obtain a complete bit rate-distortion curve with a single training session.
Multi-skill Mobile Manipulation for Object Rearrangement
Jiayuan Gu (UC San Diego), Jitendra Malik (UC Berkeley)
Robotic IntelligenceReinforcement LearningTabularBenchmark
🎯 What it does: A modular multi-skill mobile manipulation method M3 is proposed for long-duration object rearrangement tasks;
Multi-task Self-supervised Graph Neural Networks Enable Stronger Task Generalization
Mingxuan Ju (University of Notre Dame), Chuxu Zhang (Brandeis University)
ClassificationRepresentation LearningGraph Neural NetworkContrastive LearningGraph
🎯 What it does: This paper proposes PARETOGNN, a framework for achieving task generalization in graph neural networks through multi-task self-supervised learning.
Multifactor Sequential Disentanglement via Structured Koopman Autoencoders
Nimrod Berman (Ben Gurion University of the Negev), Omri Azencot (Ben Gurion University of the Negev)
Recurrent Neural NetworkAuto EncoderSequentialAudio
🎯 What it does: A multi-factor sequence separation model based on the Koopman autoencoder is proposed.
Multimodal Analogical Reasoning over Knowledge Graphs
Ningyu Zhang (Zhejiang University), Huajun Chen (Zhejiang University)
TransformerVision Language ModelImageTextMultimodalityGraphBenchmark
🎯 What it does: This paper proposes a new multimodal analogy reasoning task that combines images, text, and knowledge graphs to predict missing entities by performing analogy reasoning in the context of knowledge graphs.
Multimodal Federated Learning via Contrastive Representation Ensemble
Qiying Yu (Institute for AI Industry Research Tsinghua University), Jingjing Liu (Tsinghua University)
RetrievalFederated LearningKnowledge DistillationContrastive LearningImageTextMultimodality
🎯 What it does: The CreamFL framework is proposed to achieve multimodal federated learning, allowing the server to train large models and supporting heterogeneous client models.
Multiple sequence alignment as a sequence-to-sequence learning problem
Edo Dotan (Tel Aviv University), Tal Pupko (Tel Aviv University)
TransformerSequentialBiomedical Data
🎯 What it does: A deep learning method based on Transformer, called BetaAlign, is proposed and implemented for multiple sequence alignment.
Multitask Prompt Tuning Enables Parameter-Efficient Transfer Learning
Zhen Wang (Ohio State University), Yoon Kim (Massachusetts Institute of Technology)
OptimizationKnowledge DistillationTransformerPrompt EngineeringText
🎯 What it does: This paper proposes a Multi-Task Prompt Tuning (MPT) method that utilizes multi-source task learning to create a single transferable prompt and efficiently adapts to the target task through low-rank multiplication.
Multivariate Time-series Imputation with Disentangled Temporal Representations
SHUAI LIU (Nanyang Technological University), YUE JIANG (Nanyang Technological University)
Data SynthesisAnomaly DetectionOptimizationTime Series
🎯 What it does: A multivariate time series missing value imputation method called TIDER is proposed, which utilizes low-rank matrix decomposition and constructs interpretable temporal representations.
MultiViz: Towards Visualizing and Understanding Multimodal Models
Paul Pu Liang (Carnegie Mellon University), Ruslan Salakhutdinov (Carnegie Mellon University)
Explainability and InterpretabilityMultimodality
🎯 What it does: A multimodal model interpretability tool named MULTIVIZ is proposed, which visualizes and analyzes the internal mechanisms of the model in four stages (single-modal importance, cross-modal interaction, feature representation, prediction composition);
Mutual Partial Label Learning with Competitive Label Noise
Yan Yan (Carleton University), Yuhong Guo (CIFAR AI Chair, Amii)
ClassificationRecognitionConvolutional Neural NetworkImage
🎯 What it does: A mutual learning-based partial label learning method ML-PLL is proposed, which collaborates learning between the prediction network and class prototype network to correct labels in a competitive label noise environment.
NAGphormer: A Tokenized Graph Transformer for Node Classification in Large Graphs
Jinsong Chen (Huazhong University of Science and Technology), Kun He (Huazhong University of Science and Technology)
ClassificationGraph Neural NetworkTransformerGraph
🎯 What it does: This paper studies the problem of large-scale graph node classification and proposes NAGphormer, which represents each node as a token sequence of multi-hop neighborhood information and uses a Transformer for learning.
NANSY++: Unified Voice Synthesis with Neural Analysis and Synthesis
Hyeong-Seok Choi (Seoul National University), Hyeongju Kim (Supertone, Inc.)
GenerationData SynthesisGenerative Adversarial NetworkContrastive LearningAudio
🎯 What it does: NANSY++, a unified vocoder analysis-synthesis framework, is proposed, capable of self-supervised learning of decoupled pitch, amplitude, language, and timbre features from unlabeled audio, and achieving various speech synthesis tasks through a shared backend synthesizer.
Near-Optimal Adversarial Reinforcement Learning with Switching Costs
Ming Shi (Ohio State University), Ness Shroff (Ohio State University)
Recommendation SystemReinforcement Learning
🎯 What it does: This paper explores the issue of considering switching costs in adversarial reinforcement learning for the first time and proposes two new algorithms to achieve low regret.
Near-optimal Coresets for Robust Clustering
Lingxiao Huang (Nanjing University), Xuan Wu (Huawei Technologies)
Anomaly DetectionOptimizationTabular
🎯 What it does: This paper proposes a near-optimal ε-coreset construction method for handling robust clustering problems such as k-MEDIAN and k-MEANS with m outliers;
Near-Optimal Deployment Efficiency in Reward-Free Reinforcement Learning with Linear Function Approximation
Dan Qiao (University of California Santa Barbara), Yu-Xiang Wang (University of California Santa Barbara)
Recommendation SystemReinforcement Learning
🎯 What it does: This study investigates the deployment efficiency problem in reinforcement learning (RL) with linear function approximation under a no-reward exploration setting, proposing a new algorithm that can collect at most O(dH^2/5ϵ^2) trajectories in H deployments to identify any (potentially data-dependent) ϵ-optimal policy.
Near-optimal Policy Identification in Active Reinforcement Learning
Xiang Li (ETH Zurich), Ilija Bogunovic (University College London)
Reinforcement LearningTabular
🎯 What it does: An active exploration algorithm based on Kernelized Least Squares Value Iteration (AE‑LSVI) is proposed for identifying approximately optimal policies in generative model environments.
Nearly Minimax Optimal Offline Reinforcement Learning with Linear Function Approximation: Single-Agent MDP and Markov Game
Wei Xiong (Hong Kong University of Science and Technology), Tong Zhang (Hong Kong University of Science and Technology)
Reinforcement Learning
🎯 What it does: In offline reinforcement learning, the authors designed the lazy value iteration algorithms LinPEVI-ADV and its variant LinPEVI-ADV+ based on reference-advantage decomposition for linear MDPs and two-player zero-sum Markov games, and proved the approximate optimal sampling complexity under the coverage condition.
NERDS: A General Framework to Train Camera Denoisers from Raw-RGB Noisy Image Pairs
Heewon Kim (Soongsil University), Kyoung Mu Lee (Seoul National University)
RestorationConvolutional Neural NetworkImage
🎯 What it does: Proposes the NERDS framework, which trains a general CNN color image denoising network using only the available raw RAW-RGB noisy images through noise estimation, ISP (RAW2RGB) estimation, and low-resolution pseudo-noise-clean image pairing synthesis.
NeRF-SOS: Any-View Self-supervised Object Segmentation on Complex Scenes
Zhiwen Fan (University of Texas at Austin), Zhangyang Wang (University of Texas at Austin)
Object DetectionSegmentationNeural Radiance FieldContrastive LearningImage
🎯 What it does: A self-supervised object segmentation framework based on NeRF, called NeRF-SOS, is proposed, which can segment objects in complex real scenes from arbitrary viewpoints.
NeRN: Learning Neural Representations for Neural Networks
Maor Ashkenazi (Ben-Gurion University of the Negev), Eran Treister (Technion - Israel Institute of Technology)
Knowledge DistillationRepresentation LearningConvolutional Neural NetworkImage
🎯 What it does: This paper proposes NeRN (Neural Representation for Neural Networks), which uses an MLP to predict the weights of a pre-trained CNN through implicit reconstruction.
Neural Agents Struggle to Take Turns in Bidirectional Emergent Communication
Valentin Taillandier (École Normale Supérieure Paris-Saclay), Paul Michel (École Normale Supérieure Paris-Saclay)
Recurrent Neural NetworkReinforcement LearningTabular
🎯 What it does: This study investigates a bidirectional communication game to explore whether neural network agents can spontaneously generate a turn-taking agreement and assess its impact on task completion.
Neural Architecture Design and Robustness: A Dataset
Steffen Jung (Max Planck Institute for Informatics, University of Siegen), Margret Keuper (Max Planck Institute for Informatics, University of Siegen)
Adversarial AttackNeural Architecture SearchImageBenchmark
🎯 What it does: A robust dataset covering the complete search space of NAS-Bench-201 was constructed, evaluating 6,466 networks against various adversarial attacks and common distortions.
Neural Bregman Divergences for Distance Learning
Fred Lu (University of Maryland), Francis Ferraro (University of Maryland)
RetrievalRepresentation LearningConvolutional Neural NetworkContrastive LearningImageTabular
🎯 What it does: This paper proposes a framework that utilizes Input Convex Neural Networks (ICNN) to directly learn Bregman divergence (Neural Bregman Divergence, NBD), which can automatically infer suitable asymmetric metrics in various tasks (clustering, retrieval, regression, embedding learning, etc.) when trained jointly with a feature extraction network.
Neural Causal Models for Counterfactual Identification and Estimation
Kevin Muyuan Xia (Columbia University), Elias Bareinboim (Columbia University)
Data SynthesisOptimizationGenerative Adversarial Network
🎯 What it does: This paper proposes a framework based on Neural Causal Models (NCM) to address the identification and estimation problem of counterfactual reasoning starting from observational (L1) and experimental (L2) data as well as a causal graph G; it also provides feasible training and inference algorithms (Alg.1-3) and implements a GAN-based NCM training method.
Neural Collapse Inspired Feature-Classifier Alignment for Few-Shot Class-Incremental Learning
Yibo Yang (JD Explore Academy), Dacheng Tao (JD Explore Academy)
ClassificationRepresentation LearningImageBenchmark
🎯 What it does: This paper proposes a framework for Few-Shot Class Incremental Learning (FSCIL) based on the concept of neural collapse, where the classifier weights for all categories are pre-set to a simplex equiangular tight frame (simplex ETF). During training, a dot-regression loss drives the feature vectors to approximate these fixed prototypes, thereby avoiding catastrophic forgetting caused by the misalignment of features and classifiers.
Neural Compositional Rule Learning for Knowledge Graph Reasoning
Kewei Cheng (University of California Los Angeles), Yizhou Sun (University of California Los Angeles)
Recurrent Neural NetworkGraph
🎯 What it does: An end-to-end neural model NCRL is proposed for learning composable logical rules in knowledge graphs.
Neural DAG Scheduling via One-Shot Priority Sampling
Wonseok Jeon (Qualcomm AI Research), Christopher Lott (Qualcomm AI Research)
OptimizationGraph Neural NetworkReinforcement LearningGraph
🎯 What it does: For the directed acyclic graph (DAG) scheduling problem, a one-shot neural network encoder is proposed that directly samples node priorities and generates an optimal scheduling plan through list scheduling.
Neural Design for Genetic Perturbation Experiments
Aldo Pacchiano (Microsoft Research), Luis Voloch (Immunai)
OptimizationDrug DiscoveryTabularBiomedical Data
🎯 What it does: A noise-free batch bandit optimization method based on the Optimistic Arm Elimination (OAE) principle is proposed, specifically designed to quickly discover gene perturbations that maximize cellular phenotypes under a limited experimental budget.
Neural ePDOs: Spatially Adaptive Equivariant Partial Differential Operator Based Networks
Lingshen He (Peking University), Zhouchen Lin (Peking University)
ClassificationRecognitionConvolutional Neural NetworkImage
🎯 What it does: A spatially adaptive and translation equivariant nonlinear partial differential operator (Neural ePDOs) is proposed to replace the traditional linear shared coefficient PDO networks;
Neural Episodic Control with State Abstraction
Zhuo Li (Kyushu University), Jianjun Zhao (Kyushu University)
Reinforcement LearningSequential
🎯 What it does: A neuron temporary control framework NECSA based on grid state abstraction is proposed to enhance the sample efficiency of deep reinforcement learning.
Neural Groundplans: Persistent Neural Scene Representations from a Single Image
Prafull Sharma (Massachusetts Institute of Technology), Vincent Sitzmann (Massachusetts Institute of Technology)
Object DetectionSegmentationGenerationConvolutional Neural NetworkContrastive LearningImageVideo
🎯 What it does: This work proposes a self-supervised conditional Neural Groundplan method that can recover a complete 3D scene representation from a single image and automatically distinguish between static and movable objects, enabling novel view synthesis, instance segmentation, 3D bounding box prediction, and scene editing.
Neural Image-based Avatars: Generalizable Radiance Fields for Human Avatar Modeling
YoungJoong Kwon, Henry Fuchs (University of North Carolina)
GenerationPose EstimationNeural Radiance FieldImage
🎯 What it does: A method is proposed to quickly generate full-body 3D avatars from sparse views, enabling new perspective synthesis and pose animation.
Neural Implicit Shape Editing using Boundary Sensitivity
Arturs Berzins (SINTEF), Leif Kobbelt (RWTH Aachen University)
Point CloudMesh
🎯 What it does: A framework based on Boundary Sensitivity is proposed, which allows for geometric and semantic editing of neural implicit shapes without altering the network architecture or training process.
Neural Lagrangian Schr\"{o}dinger Bridge: Diffusion Modeling for Population Dynamics
Takeshi Koshizuka (University of Tokyo), Issei Sato (University of Tokyo)
Diffusion modelBiomedical DataPhysics RelatedStochastic Differential Equation
🎯 What it does: A framework based on the Neural Lagrangian Schrödinger Bridge (NLSB) is proposed, which uses neural SDE to learn the drift and diffusion processes of population dynamics, capable of reconstructing individual trajectories and capturing stochastic diffusion behavior with only cross-sectional samples.
Neural Networks and the Chomsky Hierarchy
Gregoire Deletang, Pedro A Ortega
Recurrent Neural NetworkTransformerSequential
🎯 What it does: Conducted large-scale experiments on the generalization ability of different neural networks in sequence prediction tasks, using 20,910 models and 15 tasks, exploring their correspondence with the Chomsky hierarchy.
Neural Networks Efficiently Learn Low-Dimensional Representations with SGD
Alireza Mousavi-Hosseini (University of Toronto), Murat A Erdogdu
OptimizationRepresentation LearningTabularStochastic Differential Equation
🎯 What it does: In high-dimensional wide two-layer neural networks, the authors prove that the weights of the first layer converge to the principal subspace of the teacher model through online SGD with weight decay training; further theoretical results are provided regarding generalization error, single exponential target learning, and compressibility.
Neural Optimal Transport
Alexander Korotin (Skolkovo Institute of Science and Technology), Evgeny Burnaev (Skolkovo Institute of Science and Technology)
Image TranslationOptimizationImage
🎯 What it does: A neural network-based algorithm is proposed for calculating the optimal transport graph and plan for both strong and weak transport costs.
Neural Radiance Field Codebooks
Matthew Wallingford (University of Washington), Ali Farhadi (Allen Institute for AI)
Object DetectionSegmentationRetrievalConvolutional Neural NetworkNeural Radiance FieldImage
🎯 What it does: The Neural Radiance Field Codebooks (NRC) method is proposed, which utilizes multi-view reconstruction to learn shareable object codebooks for unsupervised object segmentation and geometric understanding.
Neural Systematic Binder
Gautam Singh (Rutgers University), Sungjin Ahn (KAIST)
SegmentationRepresentation LearningTransformerImage
🎯 What it does: This paper studies an unsupervised object-centric learning framework called SysBinder, which can adaptively decompose images into a 'block-slot' structure.
Neural-based classification rule learning for sequential data
Marine Collery (IBM France Lab), Remy Kusters (IBM Research)
ClassificationExplainability and InterpretabilityConvolutional Neural NetworkTabularSequential
🎯 What it does: A differentiable binary neural network CR2N is proposed, which can simultaneously learn local and global interpretable classification rules in sequential data.
Neuro-Symbolic Procedural Planning with Commonsense Prompting
Yujie Lu (University of California), William Yang Wang (University of California)
Robotic IntelligenceTransformerLarge Language ModelPrompt EngineeringTextMultimodality
🎯 What it does: This study proposes a neural-symbolic procedural planning framework called PLAN, which utilizes large language models to construct intermediary prompts through structural causal models and external knowledge graphs, achieving zero-shot high-order task decomposition.
Neuroevolution is a Competitive Alternative to Reinforcement Learning for Skill Discovery
Felix Chalumeau (InstaDeep), Thomas PIERROT
Robotic IntelligenceReinforcement LearningBenchmark
🎯 What it does: This paper experimentally compares information-theoretic multi-skill reinforcement learning methods (DIAYN, DADS, SMERL, etc.) with quality-diversity (QD) evolutionary methods (MAP-ELITES, PGA-MAP-ELITES, AURORA, PGA-AURORA) in terms of skill discovery across various continuous control and exploration environments, and proposes a unified benchmarking framework.
Neuromechanical Autoencoders: Learning to Couple Elastic and Neural Network Nonlinearity
Deniz Oktay (Princeton University), Ryan P Adams
OptimizationRobotic IntelligenceAuto EncoderImage
🎯 What it does: By jointly learning the geometric morphology of mechanical deformation and neural network control, tasks such as shape matching, translation, and rotation are achieved.
New Insights for the Stability-Plasticity Dilemma in Online Continual Learning
Dahuin Jung (Seoul National University), Sungroh Yoon (Seoul National University)
Knowledge DistillationRepresentation LearningImageBenchmark
🎯 What it does: A framework for online continual learning, MuFAN, is proposed, aiming to enhance both the stability and plasticity of the model, addressing the stability-plasticity dilemma in online CL.
No Reason for No Supervision: Improved Generalization in Supervised Models
Mert Bülent Sarıyıldız (NAVER LABS Europe), Diane Larlus (NAVER LABS Europe)
ClassificationDomain AdaptationConvolutional Neural NetworkSupervised Fine-TuningContrastive LearningImage
🎯 What it does: This paper proposes a supervised learning framework that integrates techniques such as multi-scale cropping, an expendable projector, and online class means, aiming to enhance the training performance on ImageNet-1K while further improving transfer learning performance.
Noise Injection Node Regularization for Robust Learning
Noam Itzhak Levi (Tel Aviv University), Tomer Volansky (Tel Aviv University)
ClassificationDomain AdaptationConvolutional Neural NetworkImage
🎯 What it does: A regularization method called NINR was designed and evaluated, which implements noise injection nodes (NIN) in the network to enhance the robustness of DNNs against input perturbations and domain shifts.
Noise Is Not the Main Factor Behind the Gap Between Sgd and Adam on Transformers, But Sign Descent Might Be
Frederik Kunstner (University of British Columbia), Mark Schmidt (University of British Columbia)
OptimizationTransformerImageText
🎯 What it does: Compare the performance differences between SGD and Adam on Transformers, explore noise and deterministic factors, and propose sign descent and normalized gradients as simplified models.
Noise-Robust De-Duplication at Scale
Emily Silcock (Harvard University), Melissa Dell (Harvard University)
RetrievalOptimizationTransformerContrastive LearningText
🎯 What it does: A deduplicated dataset of historical news texts (NEWS-COPY) was constructed, and a neural network-based deduplication method (bi-encoder and re-ranking) was proposed to efficiently deduplicate large-scale noisy texts.
Non-parametric Outlier Synthesis
Leitian Tao (Wuhan University), Yixuan Li (University of Wisconsin - Madison)
Data SynthesisAnomaly DetectionConvolutional Neural NetworkContrastive LearningImage
🎯 What it does: A non-parametric external sample synthesis (NPOS) framework is proposed, which automatically generates artificial OOD data during training and learns the decision boundary between ID and OOD through uncertainty loss.
Nonlinear Reconstruction for Operator Learning of PDEs with Discontinuities
Samuel Lanthaler (California Institute of Technology), Siddhartha Mishra (ETH Zurich)
Convolutional Neural NetworkPhysics Related
🎯 What it does: This paper studies the problem of approximating solution operators in partial differential equations (such as linear transport equations, inviscid Burgers' equations, and compressible Euler equations) that have discontinuous solutions using operator learning methods.
NORM: Knowledge Distillation via N-to-One Representation Matching
Xiaolong Liu (Intel Labs China), Anbang Yao (Intel Labs China)
Knowledge DistillationConvolutional Neural NetworkContrastive LearningImage
🎯 What it does: A representative matching knowledge distillation method based on linear feature transformation from N-to-1 is proposed;