NeurIPS 2023 Papers — Page 18
Conference on Neural Information Processing Systems · 3218 papers
LEPARD: Learning Explicit Part Discovery for 3D Articulated Shape Reconstruction
Di Liu (Rutgers University), Dimitris N. Metaxas (Rutgers University)
SegmentationPose EstimationContrastive LearningImage
🎯 What it does: Unsupervised reconstruction of animal 3D articulated shapes from a single image, automatically discovering semantically consistent 3D parts.
Let the Flows Tell: Solving Graph Combinatorial Problems with GFlowNets
Dinghuai Zhang (Mila), Ling Pan (Mila)
OptimizationGraph Neural NetworkReinforcement LearningGraph
🎯 What it does: Designed an unsupervised learning framework based on GFlowNet to solve NP-hard combinatorial optimization problems in graph theory (maximum independent set, maximum clique, minimum dominating set, maximum cut).
Leveraging Early-Stage Robustness in Diffusion Models for Efficient and High-Quality Image Synthesis
Yulhwa Kim (Seoul National University), jae-joon kim
GenerationData SynthesisComputational EfficiencyDiffusion modelImage
🎯 What it does: This paper proposes a sparse quantization method based on early-stage robustness for the reverse diffusion process of diffusion models, significantly reducing the activation bit width.
Leveraging Locality and Robustness to Achieve Massively Scalable Gaussian Process Regression
Robert F Allison, Edward Pyzer-Knapp (IBM Research)
Gaussian SplattingTabular
🎯 What it does: An efficient Gaussian process regression algorithm that decouples parameter estimation from prediction is proposed, utilizing nearest neighbor prediction to achieve low computational cost and good prediction accuracy.
Leveraging Pre-trained Large Language Models to Construct and Utilize World Models for Model-based Task Planning
Lin Guan (Arizona State University), Subbarao Kambhampati (Arizona State University)
TransformerLarge Language ModelWorld ModelText
🎯 What it does: This study proposes a new method that utilizes pre-trained large language models (LLMs) for model-based task planning by constructing a clear world model (PDDL).
Leveraging sparse and shared feature activations for disentangled representation learning
Marco Fumero (Sapienza University of Rome), Francesco Locatello
Domain AdaptationRepresentation LearningMeta LearningImage
🎯 What it does: By utilizing sparse and shared feature activation in multi-task supervised learning, we learn disentangled representations to enhance generalization capabilities on out-of-distribution (OOD) data.
Leveraging the two-timescale regime to demonstrate convergence of neural networks
Pierre Marion (Sorbonne Université), Raphaël Berthier (EPFL)
OptimizationTabular
🎯 What it does: This paper studies the training dynamics of shallow neural networks under a two-time scale (where the internal layer step size is much smaller than the external layer) and proves that the gradient flow can converge to a global optimal solution with a finite number of neurons, applicable to piecewise constant objective functions.
Leveraging Vision-Centric Multi-Modal Expertise for 3D Object Detection
Linyan Huang (Shanghai AI Lab), Hongyang Li (Shanghai AI Lab)
Object DetectionAutonomous DrivingKnowledge DistillationTransformerMultimodalityPoint Cloud
🎯 What it does: This paper proposes the VCD framework, which includes the Visual Center Expert VCD-E and the Camera-Only Apprentice VCD-A. It utilizes LiDAR depth as a prior, long-term sequence fusion, trajectory-based distillation, and occupancy reconstruction to enhance the 3D detection performance of cameras.
Lexinvariant Language Models
Qian Huang (Stanford University), Percy Liang (Stanford University)
TransformerLarge Language ModelText
🎯 What it does: Proposes and studies a lexinvariant language model without stable word embeddings, proving that it can approximate standard LM in sufficiently long contexts;
LICO: Explainable Models with Language-Image COnsistency
Yiming Lei (Fudan University), Hongming Shan (Fudan University)
ClassificationExplainability and InterpretabilityPrompt EngineeringVision Language ModelContrastive LearningImage
🎯 What it does: The LICO framework is proposed, which enhances the interpretability and classification performance of visual models by aligning image feature space with CLIP text features.
Lie Point Symmetry and Physics-Informed Networks
Tara Akhound-Sadegh (McGill University), Siamak Ravanbakhsh (McGill University)
Supervised Fine-TuningReinforcement LearningTime SeriesPhysics Related
🎯 What it does: This paper proposes the incorporation of a Lie point symmetry loss function into Physics-Informed Neural Networks (PINNs), allowing the network to learn solutions while obtaining constraints from neighboring solutions, thereby enhancing the performance of PDE solving under low sample conditions.
Lift Yourself Up: Retrieval-augmented Text Generation with Self-Memory
Xin Cheng (Peking University), Rui Yan (Remin University of China)
GenerationRetrievalTransformerLarge Language ModelTextRetrieval-Augmented Generation
🎯 What it does: Proposes the Selfmem framework, which utilizes text generated by the model itself as infinite memory to iteratively enhance the quality of retrieval-augmented text generation.
LightSpeed: Light and Fast Neural Light Fields on Mobile Devices
Aarush Gupta, Laszlo Attila Jeni
GenerationData SynthesisComputational EfficiencyNeural Radiance FieldImage
🎯 What it does: We propose LightSpeed, a neural light field model based on classic two-plane smooth board ray parameterization and multi-dimensional feature grids, achieving real-time rendering on mobile devices.
Lightweight Vision Transformer with Bidirectional Interaction
Qihang Fan (Institute of Automation, Chinese Academy of Sciences), Ran He (Institute of Automation, Chinese Academy of Sciences)
ClassificationObject DetectionSegmentationTransformerImage
🎯 What it does: A Fully Adaptive Self-Attention (FASA) module is proposed, which combines convolution to extract local features and downsampled self-attention to obtain global features, and achieves the fusion of local and global information through bidirectional adaptive interaction; based on this, a lightweight visual Transformer FAT is constructed.
Likelihood Ratio Confidence Sets for Sequential Decision Making
Nicolas Emmenegger (ETH Zürich), Andreas Krause (ETH Zürich)
OptimizationReinforcement LearningSequential
🎯 What it does: A likelihood ratio-based anytime valid confidence sequence is proposed for adaptive uncertainty estimation in sequential decision-making (such as Bandit, reinforcement learning, active learning).
Likelihood-Based Diffusion Language Models
Ishaan Gulrajani (Stanford University), Tatsunori Hashimoto
GenerationTransformerLarge Language ModelDiffusion modelText
🎯 What it does: A likelihood-based diffusion language model called Plaid has been researched and implemented, resulting in Plaid1B, which outperforms the small autoregressive model GPT-2 124M in zero-shot log likelihood;
LIMA: Less Is More for Alignment
Chunting Zhou (Meta AI), Omer Levy (Meta AI)
TransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Fine-tuned the 65B LLaMA model using only 1000 carefully selected instruction alignment samples on a pre-trained model.
Limits, approximation and size transferability for GNNs on sparse graphs via graphops
Thien Le (Massachusetts Institute of Technology), Stefanie Jegelka (Technical University of Munich)
Graph Neural NetworkGraph
🎯 What it does: This study investigates the generalization (transfer) performance of graph neural networks on sparse graphs of different scales and provides theoretical proofs for the upper bounds of approximation and transfer errors.
Linear Time Algorithms for k-means with Multi-Swap Local Search
Junyu Huang (Central South University), Jianxin Wang (Central South University)
OptimizationTabular
🎯 What it does: A linear time algorithm for k-means with multi-exchange local search is proposed, along with a feasible sampling acceleration implementation scheme.
LinGCN: Structural Linearized Graph Convolutional Network for Homomorphically Encrypted Inference
Hongwu Peng (University of Connecticut), Caiwen Ding (University of Connecticut)
RecognitionComputational EfficiencyKnowledge DistillationNeural Architecture SearchGraph Neural NetworkGraph
🎯 What it does: The LinGCN framework is designed to reduce the multiplication depth and latency of graph convolutional network inference under CKKS homomorphic encryption by utilizing differentiable structured linearization and learnable polynomial substitutes.
Linguistic Binding in Diffusion Models: Enhancing Attribute Correspondence through Attention Map Alignment
Royi Rassin (Bar-Ilan University), Gal Chechik (NVIDIA)
GenerationData SynthesisTransformerDiffusion modelText
🎯 What it does: By introducing syntactic information during the inference process of the diffusion model and utilizing the loss function of the cross-attention map, the attention maps of modifiers and entity nouns overlap while separating from other words, thereby improving the attribute-object binding errors in text-to-image generation.
LinkerNet: Fragment Poses and Linker Co-Design with 3D Equivariant Diffusion
Jiaqi Guan (University of Illinois Urbana-Champaign), Jianzhu Ma (Tsinghua University)
Drug DiscoveryDiffusion modelGraph
🎯 What it does: This paper proposes the LinkerNet model, which jointly generates fragment poses and 3D linkers without knowing the relative positions of the fragments, achieving the co-design of PROTACs and general fragment linkers.
List and Certificate Complexities in Replicable Learning
Peter Dixon, N V Vinodchandran
🎯 What it does: This paper studies replicable learning algorithms and proposes two metrics: list replicability and certificate replicability. It designs algorithms that achieve optimal list complexity of d+1, minimize sample and certificate complexity in estimating multiple coin biases and in non-adaptive statistical query learnable classes, and provides corresponding lower bound proofs.
LLM-Pruner: On the Structural Pruning of Large Language Models
Xinyin Ma (National University of Singapore), Xinchao Wang (National University of Singapore)
CompressionTransformerLarge Language ModelText
🎯 What it does: A task-agnostic structured pruning framework called LLM-Pruner has been designed and implemented for compressing large language models (LLaMA, Vicuna, ChatGLM), retaining only a small amount of public data and completing compression and recovery within three hours.
LLMScore: Unveiling the Power of Large Language Models in Text-to-Image Synthesis Evaluation
Yujie Lu (University of California), William Yang Wang (University of California)
GenerationData SynthesisExplainability and InterpretabilityTransformerLarge Language ModelVision Language ModelImageTextMultimodality
🎯 What it does: A text and image alignment evaluation framework called LLMScore based on large language models is proposed, which can generate multi-granularity (whole image and object-level) visual descriptions and provide scores and reasons.
LMC: Large Model Collaboration with Cross-assessment for Training-Free Open-Set Object Recognition
Haoxuan Qu (Singapore University of Technology and Design), Jun Liu (Singapore University of Technology and Design)
RecognitionObject DetectionTransformerLarge Language ModelImageTextChain-of-Thought
🎯 What it does: An untrained open-set object recognition framework LMC is proposed, which utilizes ChatGPT, DALL‑E, CLIP, and DINO to collaboratively generate virtual open-set categories and synthesize diverse images for each category. During the inference phase, it combines text and image alignment to reduce the impact of pseudo-visual features.
Local Convergence of Gradient Methods for Min-Max Games: Partial Curvature Generically Suffices
Guillaume Wang (École polytechnique fédérale de Lausanne), Lénaïc Chizat (École polytechnique fédérale de Lausanne)
Optimization
🎯 What it does: This paper studies the convergence properties of gradient methods (including continuous-time gradient flow, Sim-GDA, Alt-GDA, Extra-Gradient, etc.) near local Nash equilibria in two-player zero-sum games.
Locality Sensitive Hashing in Fourier Frequency Domain For Soft Set Containment Search
Indradyumna Roy (Indian Institute of Technology Bombay), Abir De (Indian Institute of Technology Gandhinagar)
RetrievalOptimizationTabular
🎯 What it does: Designed and implemented an asymmetric local sensitive hash based on the Fourier frequency domain (FOURIERHASHNET) for soft set inclusion retrieval.
Locality-Aware Generalizable Implicit Neural Representation
Doyup Lee (Kakao Brain), Wook-Shin Han (POSTECH)
RestorationGenerationData SynthesisTransformerDiffusion modelImage
🎯 What it does: This paper proposes a general implicit neural representation framework that utilizes a Transformer encoder and a locally aware decoder for high-quality image reconstruction and few-shot viewpoint synthesis.
Localized Symbolic Knowledge Distillation for Visual Commonsense Models
Jae Sung Park (University of Washington), Yejin Choi (University of Washington)
GenerationKnowledge DistillationTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringVision Language ModelImageTextMultimodality
🎯 What it does: This paper generates visual commonsense question answering by allowing a large language model (ChatGPT) to produce diverse textual descriptions of images and local regions. It then uses a supervised critique model to filter high-quality samples, constructing a corpus of 1 million local visual commonsense data, and fine-tunes visual-language models such as BLIP-2 on this corpus, enabling them to answer questions in a 'pointing' manner under zero-shot conditions.
Locally Invariant Explanations: Towards Stable and Unidirectional Explanations through Local Invariant Learning
Amit Dhurandhar (IBM Research), Vijay Arya (IBM Research)
Explainability and InterpretabilityImageTextTabular
🎯 What it does: A local explanation method called LINEX, based on invariant risk minimization (IRM) and game theory, is proposed, which can generate feature importance explanations with high fidelity, stability, and unidirectionality (consistency in the same direction).
Lockdown: Backdoor Defense for Federated Learning with Isolated Subspace Training
Tiansheng Huang (Georgia Institute of Technology), Ling Liu (Georgia Institute of Technology)
Federated LearningImage
🎯 What it does: In the federated learning environment, Lockdown is proposed, utilizing isolated subspace training and dynamic subspace search to counteract backdoor attacks, addressing the issue of toxic coupling in traditional sparsification defenses within federated learning.
LoCoOp: Few-Shot Out-of-Distribution Detection via Prompt Learning
Atsuyuki Miyai (University of Tokyo), Kiyoharu Aizawa (University of Tokyo)
Anomaly DetectionTransformerPrompt EngineeringContrastive LearningImage
🎯 What it does: This paper proposes LoCoOp, a prompt learning method for few-shot OOD detection using CLIP local features.
Logarithmic-Regret Quantum Learning Algorithms for Zero-Sum Games
Minbo Gao (Chinese Academy of Sciences), Qisheng Wang (Nagoya University)
Reinforcement LearningPhysics Related
🎯 What it does: An online quantum algorithm based on quantum sampling is proposed for quickly computing approximate Nash equilibria in zero-sum games.
LogSpecT: Feasible Graph Learning Model from Stationary Signals with Recovery Guarantees
Shangyuan Liu (Chinese University of Hong Kong), Anthony Man-Cho So (Chinese University of Hong Kong)
Graph Neural NetworkGraph
🎯 What it does: This paper proposes LogSpecT and its robust version rLogSpecT for learning graph structures from stationary graph signals, addressing the infeasibility issues of traditional SpecT and rSpecT.
Long Sequence Hopfield Memory
Hamza Tahir Chaudhry (Harvard University), Cengiz Pehlevan (Harvard University)
Time SeriesSequential
🎯 What it does: A sequence memory network based on Dense Associative Memory (DenseNet and its variants) is proposed, capable of storing and accurately recalling long sequences;
Long-Term Fairness with Unknown Dynamics
Tongxin Yin (University of Michigan), Yang Liu (University of California, Santa Cruz)
OptimizationReinforcement LearningTabular
🎯 What it does: This paper models the long-term fairness problem as an online reinforcement learning problem influenced by unknown dynamics, with the goal of minimizing cumulative loss while constraining cumulative fairness violations.
Look Beneath the Surface: Exploiting Fundamental Symmetry for Sample-Efficient Offline RL
Peng Cheng (Beijing Jiaotong University), Li Jiang (Tsinghua University)
Reinforcement LearningTabularBenchmarkOrdinary Differential Equation
🎯 What it does: A dynamical model TDM based on time-reversible symmetry is proposed, and an efficient offline reinforcement learning algorithm TSRL is designed based on this model.
Look Ma, No Hands! Agent-Environment Factorization of Egocentric Videos
Matthew Chang (University of Illinois), Saurabh Gupta (University of Illinois)
Image TranslationObject DetectionSegmentationRobotic IntelligenceDiffusion modelVideo
🎯 What it does: This paper proposes to separate the human hands and the environment in first-person perspective videos into two representations: agents and the environment, and enhances the performance of downstream robotic tasks through a video hand-removal technique.
Lookaround Optimizer: $k$ steps around, 1 step average
Jiangtao Zhang (Zhejiang University), Mingli Song (Zhejiang University)
OptimizationConvolutional Neural NetworkTransformerImage
🎯 What it does: Proposes the Lookaround optimizer, which alternates between multi-data augmentation training and weight averaging to continuously explore the interior of loss basins and achieve flatter minima.
Lookup Table meets Local Laplacian Filter: Pyramid Reconstruction Network for Tone Mapping
Feng Zhang (Huazhong University of Science and Technology), Nong Sang (Huazhong University of Science and Technology)
RestorationTransformerImage
🎯 What it does: An end-to-end HDR image tone mapping framework is proposed, combining 3D LUT global operations with learnable local Laplacian filters, and achieving high-resolution efficient processing through Laplacian pyramid decomposition.
Loss Decoupling for Task-Agnostic Continual Learning
Yan-Shuo Liang (Nanjing University), Wu-Jun Li (Nanjing University)
ClassificationOptimizationSequential
🎯 What it does: This paper proposes a method called Loss Decoupling (LODE) to separate the two learning objectives of new tasks in task-agnostic continual learning, thereby enhancing the balance between stability and plasticity of the model between old and new tasks.
Loss Dynamics of Temporal Difference Reinforcement Learning
Blake Bordelon (Harvard University), Cengiz Pehlevan (Harvard University)
Reinforcement LearningSequentialPhysics Related
🎯 What it does: This study investigates the learning dynamics of value function linear approximation in TD learning and derives typical learning curves using statistical physics mean field theory, revealing the effects of learning platforms, reward shaping, and learning rate annealing.
Lossy Image Compression with Conditional Diffusion Models
Ruihan Yang (University of California), Stephan Mandt (University of California)
CompressionDiffusion modelAuto EncoderImageVideo
🎯 What it does: An end-to-end lossy image compression framework is proposed, combining traditional transform coding with conditional diffusion models, achieving reverse diffusion reconstruction controlled by content latent variables.
Lovász Principle for Unsupervised Graph Representation Learning
Ziheng Sun (Chinese University of Hong Kong), Jicong Fan (Chinese University of Hong Kong)
OptimizationRepresentation LearningGraph Neural NetworkGraph
🎯 What it does: An unsupervised graph-level representation learning framework based on the Lovász number (Lovász principle) is proposed, enhancing its expressive power through subgraph Lovász numbers.
Low Tensor Rank Learning of Neural Dynamics
Arthur Pellegrino (University of Edinburgh), Angus Chadwick (University of Edinburgh)
Recurrent Neural NetworkTime SeriesSequentialOrdinary Differential Equation
🎯 What it does: This study investigates the low tensor rank structure of the weight tensor in recurrent neural networks (RNNs) during the learning process and applies this framework to large-scale neural recordings and task training of RNNs, revealing that the weight updates resulting from learning maintain a low tensor rank.
Lower Bounds on Adaptive Sensing for Matrix Recovery
Praneeth Kacham (Carnegie Mellon University), David Woodruff
🎯 What it does: This study investigates the lower bounds on the number of measurements and adaptive rounds required by adaptive sensing algorithms using linear measurements in low-rank matrix recovery.
LuminAIRe: Illumination-Aware Conditional Image Repainting for Lighting-Realistic Generation
Jiajun Tang (Peking University), Boxin Shi (Peking University)
Image TranslationGenerationData SynthesisImage
🎯 What it does: The LuminAIRe task is proposed, utilizing 3D lighting and geometric information to achieve lighting realism in Conditional Image Repainting (CIR), and providing the CAR-LUMINAIRE dataset.
LVM-Med: Learning Large-Scale Self-Supervised Vision Models for Medical Imaging via Second-order Graph Matching
Duy Minh Ho Nguyen, Mathias Niepert (University of Stuttgart)
ClassificationObject DetectionSegmentationGraph Neural NetworkTransformerContrastive LearningImageBiomedical DataMagnetic Resonance ImagingComputed TomographyUltrasound
🎯 What it does: A self-supervised visual model LVM-Med based on second-order graph matching was trained, undergoing large-scale pre-training on approximately 1.3M medical images and fine-tuning on 15 downstream tasks (segmentation, classification, detection, etc.);
Machine learning detects terminal singularities
Tom Coates (Imperial College London), Sara Veneziale (Imperial College London)
ClassificationComputational EfficiencyTabularPhysics Related
🎯 What it does: This paper utilizes machine learning methods to identify the ultimate singularities of toric Fano varieties with Picard rank two, and based on this, it draws the 'morphological map' of Q-Fano varieties.
Macro Placement by Wire-Mask-Guided Black-Box Optimization
Yunqi Shi (Nanjing University), Chao Qian (Nanjing University)
OptimizationTabularBenchmark
🎯 What it does: A new macro block placement framework called WireMask-BBO is proposed, which achieves high-quality non-overlapping placement by directly treating macro block positions as gene expressions and using wire mask-based greedy improvement evaluation.
MADG: Margin-based Adversarial Learning for Domain Generalization
Aveen Dayal (Indian Institute of Technology Hyderabad), Vineeth N. Balasubramanian
Domain AdaptationGenerative Adversarial NetworkImageBenchmark
🎯 What it does: The MADG algorithm is proposed, which uses margin-based adversarial learning to achieve domain-invariant features and provides a corresponding theoretical error upper bound.
MAG-GNN: Reinforcement Learning Boosted Graph Neural Network
Lecheng Kong (Washington University in St. Louis), Muhan Zhang (Peking University)
OptimizationGraph Neural NetworkReinforcement LearningGraph
🎯 What it does: This paper proposes a graph neural network model named MAG-GNN, which enhances the expressive power of GNNs by using reinforcement learning to select a small number of the most discriminative subgraphs from all possible subgraphs, while significantly reducing the computational overhead of subgraph enumeration.
Make Pre-trained Model Reversible: From Parameter to Memory Efficient Fine-Tuning
Baohao Liao (University of Amsterdam), Christof Monz (University of Amsterdam)
TransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This paper proposes three reversible parameter and memory-efficient fine-tuning methods (MEFT) that transform pre-trained language models (PLMs) into reversible networks without additional pre-training, thereby significantly reducing activation memory usage by only caching the final output during fine-tuning.
Make the U in UDA Matter: Invariant Consistency Learning for Unsupervised Domain Adaptation
Zhongqi Yue (Nanyang Technological University), Hanwang Zhang (Nanyang Technological University)
Domain AdaptationConvolutional Neural NetworkTransformerImageText
🎯 What it does: A new unsupervised domain adaptation method called ICON is proposed, which eliminates source domain-specific associations by learning to balance the consistency between source domain labels and target domain clustering, thereby improving the model's domain invariance.
Making Scalable Meta Learning Practical
Sang Keun Choe (Carnegie Mellon University), Eric Xing (Carnegie Mellon University)
Meta LearningSupervised Fine-TuningImageText
🎯 What it does: A scalable meta-learning framework named SAMA is proposed, enabling efficient training of meta-learning in large-scale models and multi-GPU environments.
Managing Temporal Resolution in Continuous Value Estimation: A Fundamental Trade-off
Zichen Zhang (University of Alberta), Dale Schuurmans (University of Alberta)
Reinforcement LearningTime Series
🎯 What it does: This paper studies the impact of time discretization on the Monte Carlo policy evaluation error in continuous-time reinforcement learning, providing a closed-form expression for the error and deriving the optimal sampling step size.
Many-body Approximation for Non-negative Tensors
Kazu Ghalamkari (RIKEN), Yoshinobu Kawahara (Osaka University)
OptimizationImageTime Series
🎯 What it does: A convex optimization framework based on many-body approximation and energy models is proposed for non-negative tensor decomposition and missing value imputation.
Marginal Density Ratio for Off-Policy Evaluation in Contextual Bandits
Muhammad Faaiz Taufiq (University of Oxford), Jean-Francois Ton
Tabular
🎯 What it does: A new offline policy evaluation (OPE) method called the Marginal Ratio (MR) estimator is proposed, which estimates the value of the target policy by focusing on the marginal distribution shift of the outcome Y.
Marich: A Query-efficient Distributionally Equivalent Model Extraction Attack
Pratik Karmakar (National University of Singapore), Debabrota Basu (Univ. Lille)
OptimizationAdversarial AttackConvolutional Neural NetworkTransformerReinforcement LearningImageText
🎯 What it does: A black-box model extraction attack method based on public data, MARICH, is proposed, which can construct a replica model that is distributionally equivalent to the target model with a very small number of queries.
MarioGPT: Open-Ended Text2Level Generation through Large Language Models
Shyam Sudhakaran, Sebastian Risi
GenerationTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
🎯 What it does: This paper presents MarioGPT, a large language model capable of generating playable Super Mario Bros levels based on natural language prompts.
Markovian Sliced Wasserstein Distances: Beyond Independent Projections
Khai Nguyen (University of Texas at Austin), Nhat Ho (University of Texas at Austin)
GenerationData SynthesisOptimizationComputational EfficiencyImagePoint Cloud
🎯 What it does: This paper proposes the Markovian sliced Wasserstein (MSW) distance, which improves the redundancy projection problem of traditional Sliced Wasserstein by introducing a first-order Markov structure in the projection direction.
Mask Propagation for Efficient Video Semantic Segmentation
Yuetian Weng (Monash University), Bohan Zhuang (Monash University)
SegmentationComputational EfficiencyTransformerOptical FlowVideo
🎯 What it does: An efficient video semantic segmentation framework based on mask propagation, MPVSS, is proposed. It uses a powerful query-based image segmenter to generate binary masks and class predictions on key frames, and then employs a query-based flow estimation module to generate segment-level flows for each mask. Finally, the masks are projected onto non-key frames through bilinear projection along the flow direction, achieving fast and accurate video segmentation.
Masked Image Residual Learning for Scaling Deeper Vision Transformers
Guoxi Huang (Baidu Inc), Adrian G. Bors (University of York)
Object DetectionSegmentationTransformerContrastive LearningImage
🎯 What it does: A self-supervised framework called Masked Image Residual Learning (MIRL) is proposed to address the performance degradation issue of deep Vision Transformers (ViT) during pre-training with Masked Image Modeling (MIM) and to enable efficient training of deeper ViTs.
Masked Space-Time Hash Encoding for Efficient Dynamic Scene Reconstruction
Feng Wang (Tsinghua University), Huaping Liu (Tsinghua University)
GenerationComputational EfficiencyNeural Radiance FieldVideo
🎯 What it does: A Masked Space-Time Hash encoding (MSTH) method is proposed, which decomposes dynamic scenes into a weighted combination of 3D hash encoding and 4D hash encoding, and achieves automatic separation of static and dynamic areas through a learnable mask, significantly reducing hash collisions and training costs.
Masked Two-channel Decoupling Framework for Incomplete Multi-view Weak Multi-label Learning
Chengliang Liu (Harbin Institute of Technology), Yong Xu (Harbin Institute of Technology)
ClassificationRepresentation LearningContrastive LearningImage
🎯 What it does: A Masked Two-Channel Decoupling Framework (MTD) is proposed to specifically address the issue of incomplete multi-view weak multi-label learning (iMvWMLC).
Mass-Producing Failures of Multimodal Systems with Language Models
Shengbang Tong (University of California Berkeley), Jacob Steinhardt (University of California Berkeley)
GenerationAnomaly DetectionTransformerLarge Language ModelPrompt EngineeringTextMultimodality
🎯 What it does: MULTIMON has been developed, an automated evaluation framework based on error consistency and large language models, capable of mining single errors from text corpora, generalizing them into systematic failures, and automatically generating new failure instances;
MathNAS: If Blocks Have a Role in Mathematical Architecture Design
Wang Qinsi, Sihai Zhang (University of Science and Technology of China)
OptimizationComputational EfficiencyNeural Architecture SearchImage
🎯 What it does: A general framework called MathNAS is proposed to transform NAS into Integer Linear Programming (ILP), utilizing modular search space to split into block performance evaluation and network performance prediction, thereby reducing the network search complexity from exponential to polynomial level.
Matrix Compression via Randomized Low Rank and Low Precision Factorization
Rajarshi Saha (Stanford University), Mert Pilanci (Stanford University)
CompressionImageText
🎯 What it does: A matrix compression algorithm LPLR is proposed, which simultaneously achieves low-rank decomposition and low-precision quantization, and provides a theoretical error upper bound.
MAViL: Masked Audio-Video Learners
Po-Yao Huang (Meta), Christoph Feichtenhofer (Meta)
ClassificationRetrievalRepresentation LearningTransformerAuto EncoderContrastive LearningVideoMultimodalityAudio
🎯 What it does: A self-supervised audio-video learning framework called MAViL has been developed, utilizing three objectives: masked reconstruction, cross-modal and intra-modal contrastive learning, and self-training to learn unified audio-video representations.
Max-Margin Token Selection in Attention Mechanism
Davoud Ataee Tarzanagh (University of Pennsylvania), Samet Oymak (University of Michigan)
OptimizationTransformerImage
🎯 What it does: This paper studies the optimization dynamics of the Transformer attention mechanism, proving that gradient descent converges to the maximum margin token selection on attention weights, and explores the joint convergence of attention heads and prediction heads.
Max-Sliced Mutual Information
Dor Tsur (Ben-Gurion University), Kristjan Greenewald (Massachusetts Institute of Technology IBM Watson AI Lab)
Computational EfficiencyRepresentation LearningImageTabular
🎯 What it does: This paper proposes the Maximum Sliced Mutual Information (mSMI), a scalable metric that combines CCA linear projection with mutual information to capture nonlinear dependencies between high-dimensional random variables and achieve scalable estimation.
Maximization of Average Precision for Deep Learning with Adversarial Ranking Robustness
Gang Li (Texas A&M University), Tianbao Yang (Texas A&M University)
OptimizationAdversarial AttackImage
🎯 What it does: A framework for adversarial robustness optimization targeting average precision (AP) is proposed, along with a stochastic optimization algorithm that can be efficiently solved without the need for large batch training.
Maximize to Explore: One Objective Function Fusing Estimation, Planning, and Exploration
Zhihan Liu (Northwestern University), Zhaoran Wang (Northwestern University)
OptimizationRobotic IntelligenceReinforcement Learning
🎯 What it does: A unified RL framework called MEX (Maximize to Explore) is proposed, which automatically balances exploration and exploitation by simultaneously maximizing the estimated error penalty and the hypothesis-based expected return within a single unconstrained objective.
Maximum Average Randomly Sampled: A Scale Free and Non-parametric Algorithm for Stochastic Bandits
Masoud Moravej Khorasani (University of Melbourne), Erik Weyer (University of Melbourne)
Reinforcement LearningTabular
🎯 What it does: A non-parametric, scale-free UCB algorithm called MARS is proposed, based on random subsampling, to address the exploration-exploitation trade-off in multi-armed bandits.
Maximum Independent Set: Self-Training through Dynamic Programming
Lorenzo Brusca (Ecole Polytechnique Federale de Lausanne), Volkan Cevher (University of Wisconsin Madison)
OptimizationGraph Neural NetworkGraph
🎯 What it does: A graph neural network framework based on dynamic programming is proposed, which solves the maximum independent set (MIS) by recursively comparing subgraphs.
Maximum State Entropy Exploration using Predecessor and Successor Representations
Arnav Kumar Jain (Mila-Quebec Artificial Intelligence Institute), Glen Berseth (Mila-Quebec Artificial Intelligence Institute)
Recurrent Neural NetworkReinforcement LearningSequential
🎯 What it does: This paper proposes ηψ-Learning, a method for learning efficient exploration strategies by maximizing the state access entropy of a single finite-length trajectory, and achieves non-Markovian, deterministic exploration by combining predecessor representation and successor representation.
May the Force be with You: Unified Force-Centric Pre-Training for 3D Molecular Conformations
Rui Feng (Georgia Institute of Technology), Chao Zhang (Georgia Institute of Technology)
Drug DiscoveryTransformerGraph
🎯 What it does: A unified force-based pre-training model ET-OREO is proposed, which learns representations of 3D molecules in both equilibrium and non-equilibrium conformations.
MCUFormer: Deploying Vision Tranformers on Microcontrollers with Limited Memory
Yinan Liang (Tsinghua University), Jiwen Lu (Tsinghua University)
ClassificationOptimizationNeural Architecture SearchTransformerImage
🎯 What it does: Deploying visual Transformers on microcontrollers with extremely low memory, the MCUFormer framework is proposed to achieve co-optimization of hardware and algorithms.
Mechanic: A Learning Rate Tuner
Ashok Cutkosky (Boston University), Harsh Mehta (Google Research)
OptimizationTransformerText
🎯 What it does: A learning rate regulator named MECHANIC is proposed, which can automatically adjust the learning rate scaling factor dynamically for any base optimization algorithm (such as SGD, AdamW, Lion, etc.).
Mechanism Design for Collaborative Normal Mean Estimation
Yiding Chen (University of Wisconsin Madison), Kirthevasan Kandasamy (University of Wisconsin Madison)
Optimization
🎯 What it does: A mechanism (C3D) is designed to prevent agent falsification and free-riding in collaborative normal mean estimation.
MeCo: Zero-Shot NAS with One Data and Single Forward Pass via Minimum Eigenvalue of Correlation
Tangyu Jiang (Beijing Normal University), Rongfang Bie (Beijing Normal University)
OptimizationNeural Architecture SearchConvolutional Neural NetworkImage
🎯 What it does: A novel zero-cost proxy called MeCo (and its optimized version MeCo opt) is proposed, which evaluates network architecture by calculating the minimum eigenvalue of the feature map from a single random input;
Med-UniC: Unifying Cross-Lingual Medical Vision-Language Pre-Training by Diminishing Bias
Zhongwei Wan (Ohio State University), Rossella Arcucci (Imperial College London)
ClassificationObject DetectionSegmentationConvolutional Neural NetworkTransformerVision Language ModelContrastive LearningImageTextBiomedical DataMagnetic Resonance ImagingComputed Tomography
🎯 What it does: This paper proposes a cross-lingual medical vision-language pre-training framework Med-UniC, which combines chest X-ray images and radiology reports in English and Spanish to achieve cross-lingual visual-language alignment and self-supervised visual alignment, and unifies text representation through cross-lingual text alignment regularization (CTR).
Meek Separators and Their Applications in Targeted Causal Discovery
Kirankumar Shiragur (Broad Institute of MIT and Harvard), Caroline Uhler (Broad Institute of MIT and Harvard)
Graph
🎯 What it does: Proposed the Meek separator and utilized it to construct a random algorithm, achieving logarithmic-level intervention count approximation in two types of causal discovery problems: subset search and causal mean matching;
Meet in the Middle: A New Pre-training Paradigm
Anh Tuan Nguyen (Microsoft Azure AI), Weizhu Chen (Microsoft Azure AI)
GenerationData-Centric LearningAI Code AssistantTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Proposes the Meet-in-the-Middle (MIM) pre-training paradigm, which simultaneously trains forward (left-to-right) and backward (right-to-left) language models within the same shared decoder, and encourages consistency between the two models' word distributions at each position through consistency regularization.
MEGABYTE: Predicting Million-byte Sequences with Multiscale Transformers
LILI YU, Mike Lewis (Meta AI)
GenerationCompressionComputational EfficiencyTransformerImageTextMultimodalityAudio
🎯 What it does: A multi-scale decoder called MEGABYTE is proposed for autoregressive modeling of million-byte sequences.
MeGraph: Capturing Long-Range Interactions by Alternating Local and Hierarchical Aggregation on Multi-Scaled Graph Hierarchy
Honghua Dong (Tencent Robotics X), Lei Han (Tencent Robotics X)
Graph Neural NetworkGraphBenchmark
🎯 What it does: The MeGraph model is proposed, constructing a mega graph to capture long-range interactions in the graph through alternating local (intra) and hierarchical (inter) aggregation.
Memory Efficient Optimizers with 4-bit States
Bingrui Li (Tsinghua University), Jun Zhu (Tsinghua University)
OptimizationTransformerSupervised Fine-TuningImageText
🎯 What it does: A compressed low-precision optimizer that compresses the state of optimizers like Adam/AdamW to 4-bit integers, balancing accuracy and memory efficiency.
Memory-Constrained Algorithms for Convex Optimization
Moise Blanchard, Patrick Jaillet (Massachusetts Institute of Technology)
Optimization
🎯 What it does: A class of recursive cutting plane algorithms is designed to solve convex feasibility problems with memory constraints (which can be extended to first-order convex optimization).
Memory-Efficient Fine-Tuning of Compressed Large Language Models via sub-4-bit Integer Quantization
Jeonghoon Kim (NAVER Cloud), Dongsoo Lee (NAVER Cloud)
CompressionOptimizationComputational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: A parameter-efficient, quantization-aware fine-tuning method called PEQA has been developed for low-bit quantization of large language models, achieving task adaptation while maintaining a small model size.
MEMTO: Memory-guided Transformer for Multivariate Time Series Anomaly Detection
Junho Song (Seoul National University), Sungzoon Cho (Seoul National University)
Anomaly DetectionTransformerTime Series
🎯 What it does: A memory-guided reconstruction model based on Transformer (MEMTO) is proposed for unsupervised multivariate time series anomaly detection.
Meta-AdaM: An Meta-Learned Adaptive Optimizer with Momentum for Few-Shot Learning
Siyuan Sun (Iowa State University), Hongyang Gao (Iowa State University)
OptimizationMeta LearningRecurrent Neural NetworkImage
🎯 What it does: A novel meta-learning optimizer, Meta-AdaM, is proposed for few-shot learning scenarios to achieve rapid convergence and better utilize weight update history.
Meta-Adapter: An Online Few-shot Learner for Vision-Language Model
Cheng Cheng (Xi'an JiaoTong University), Ying Shan (Tencent AI Lab)
ClassificationObject DetectionSegmentationMeta LearningTransformerPrompt EngineeringVision Language ModelContrastive LearningImage
🎯 What it does: This paper proposes Meta-Adapter, a lightweight, online-learning residual adapter that achieves efficient few-shot classification, detection, and segmentation on the CLIP model using a small number of samples.
Meta-in-context learning in large language models
Julian Coda-Forno (Max Planck Institute for Biological Cybernetics), Eric Schulz (Max Planck Institute for Biological Cybernetics)
Meta LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelPrompt EngineeringTextTabular
🎯 What it does: This study investigates how large language models enhance their learning capabilities through meta-in-context learning via contextual learning, validated on one-dimensional regression, two-armed bandits, real regression datasets, and NLP benchmarks.
Meta-Learning Adversarial Bandit Algorithms
Mikhail Khodak (Carnegie Mellon University), Steven Wu
OptimizationMeta Learning
🎯 What it does: An online meta-learning framework is proposed, capable of jointly tuning the initialization, step size, and regularization parameters of internal algorithms when encountering multiple similar tasks, in order to reduce the average regret.
Meta-learning families of plasticity rules in recurrent spiking networks using simulation-based inference
Basile Confavreux (Institute of Science and Technology Austria), Tim P. Vogels (Institute of Science and Technology Austria)
Meta LearningSpiking Neural NetworkTime Series
🎯 What it does: In large-scale plastic synaptic variable recursive spiking networks, the filtering simulation inference (fSBI) method is used to gradually filter out synaptic plasticity rules that can produce biologically acceptable dynamics.
Meta-Learning with Neural Bandit Scheduler
Yunzhe Qi (University of Illinois at Urbana-Champaign), Jingrui He (University of Illinois at Urbana-Champaign)
Meta LearningDrug DiscoveryImage
🎯 What it does: This paper proposes a task scheduling framework BASS based on contextual Bandit, which dynamically selects training tasks in meta-learning to enhance the generalization performance of the meta-model.
Metis: Understanding and Enhancing In-Network Regular Expressions
Zhengxin Zhang (Tsinghua University), Hengyang Xu (Tsinghua University)
Computational EfficiencyKnowledge DistillationRecurrent Neural NetworkSequential
🎯 What it does: Converts regular expression rules into byte-level recurrent neural networks, and then transforms them into a pooling soft random forest that can be implemented on switches through semi-supervised knowledge distillation, achieving high zero-shot accuracy, better performance after training, and significantly improved throughput after deployment.
Metropolis Sampling for Constrained Diffusion Models
Nic Fishman (University of Oxford), Valentin De Bortoli (École Normale Supérieure Ulm)
GenerationData SynthesisOptimizationDiffusion modelScore-based ModelTabularTime Series
🎯 What it does: A Metropolis-based discretization method for reflected Brownian motion is proposed for training diffusion models in constrained Riemannian multi-dimensional spaces.
MG-ViT: A Multi-Granularity Method for Compact and Efficient Vision Transformers
Yu Zhang (Tongji University), Liang Hu (Tongji University)
ClassificationObject DetectionSegmentationCompressionComputational EfficiencyTransformerImage
🎯 What it does: A compressible visual Transformer framework MG-ViT based on multi-granularity segmentation is proposed, which can significantly reduce computational cost while maintaining or improving classification accuracy.