Conference on Neural Information Processing Systems Β· 1376 papers
Quasi-Monte Carlo Graph Random Features
Isaac Reid (University of Cambridge), Adrian Weller (University of Cambridge)
CodeGraph Neural NetworkGraph
π― What it does: This paper proposes the q-GRF (quasi-graph random features) with resistance termination, which reduces the variance of kernel estimation through negatively correlated random walk lengths while maintaining unbiased estimation, improving upon the previous GRFs method.
π― What it does: A query-based temporal fusion network QTNet is proposed, which achieves temporal information fusion for 3D object detection using sparse query features.
QuIP: 2-Bit Quantization of Large Language Models With Guarantees
Jerry Chee (Cornell University), Christopher De Sa (Cornell University)
CodeCompressionOptimizationTransformerLarge Language ModelText
π― What it does: A post-training quantization method named QuIP is proposed, which can compress large language models to just 2 bits and has theoretical guarantees.
CodeTransformerLarge Language ModelSupervised Fine-TuningTextRetrieval-Augmented Generation
π― What it does: A landmark token-based attention mechanism is designed and implemented, allowing the Transformer to achieve infinite context length without increasing training costs while maintaining random access flexibility.
π― What it does: A sequential time-step training method based on random sparse updates, called RSNG, is proposed to maintain causality and reduce computational costs when solving partial differential equations.
Yifan Pu (Tsinghua University), Gao Huang (Tsinghua University)
CodeObject DetectionTransformerImage
π― What it does: In response to DETR-based object detection, Rank-DETR is proposed, achieving high-quality detection through a series of ranking-based architectures and optimization designs.
π― What it does: Proposes the Rank-N-Contrast (RNC) framework to learn continuous, sequentially aware feature representations, thereby enhancing the performance of deep regression models.
π― What it does: Utilizing features from pre-trained models, combined with a frozen random projection layer and class prototype accumulation, to achieve replay-free continual learning;
π― What it does: Learned and implemented a multi-view consistency ray-surface distance field to represent 3D shapes, achieving efficient point cloud and view generation.
π― What it does: A Continuous Changing Interference (CCC) benchmark is proposed, demonstrating the collapse of existing Test-Time Adaptation (TTA) methods in long-term continuous adaptation, and a simple baseline RDumb is introduced, which only requires periodic resetting of pre-trained weights.
Real-Time Motion Prediction via Heterogeneous Polyline Transformer with Relative Pose Encoding
Zhejun Zhang (ETH Zurich), Luc Van Gool (ETH Zurich)
CodeAutonomous DrivingTransformerPoint Cloud
π― What it does: A real-time motion prediction framework HPTR based on aligned polygons is proposed, which efficiently integrates various traffic elements by combining KNARPE attention;
Real-World Image Super-Resolution as Multi-Task Learning
Wenlong Zhang (Hong Kong Polytechnic University), Chao Dong (Shanghai AI Laboratory)
CodeRestorationSuper ResolutionImage
π― What it does: This paper re-examines the real-world image super-resolution (real-SR) problem from the perspective of multi-task learning and proposes a task grouping method to address the issue of task competition, fine-tuning the real-SR model through this method.
π― What it does: Reformulate Continual Learning as a sequence modeling problem, and utilize Transformers and their efficient variants for Meta-Continual Learning within this framework.
Shashank Rajput (University of Wisconsin-Madison), Maheswaran Sathiamoorthy (Google)
CodeRetrievalRecommendation SystemTransformerText
π― What it does: A recommendation framework called TIGER based on generative retrieval is proposed, which represents products as semantic IDs generated by hierarchical quantization, and directly predicts the next item's semantic ID using a Transformer sequence-to-sequence model.
ReContrast: Domain-Specific Anomaly Detection via Contrastive Reconstruction
Jia Guo (Beijing Institute of Technology), Huiqi Li (Beijing Institute of Technology)
CodeAnomaly DetectionAuto EncoderContrastive LearningImageBiomedical Data
π― What it does: This study investigates unsupervised anomaly detection for specific domains and proposes the ReContrast method, optimizing the entire network to address the domain discrepancy issue of ImageNet pre-trained encoders.
Recovering from Out-of-sample States via Inverse Dynamics in Offline Reinforcement Learning
Ke Jiang (Nanjing University of Aeronautics and Astronautics), Xiaoyang Tan (Peking University)
CodeReinforcement LearningTabular
π― What it does: Proposes an Offline State Recovery (OSR) method in offline reinforcement learning, utilizing an inverse dynamics model to guide the policy in recovering states within the offline data distribution, alleviating the issue of state distribution shift.
π― What it does: This study investigates the use of recursive hypernetworks (RNN+HN) in Meta-RL to achieve improvements in sample efficiency and performance.
π― What it does: Proposed the 'Recursion in Recursion' (RIR) framework, which nests the outer k-ary balanced tree recursion with the inner Efficient Beam-Tree RvNN (EBT-RvNN), achieving logarithmic scalability for long sequences;
Red Teaming Deep Neural Networks with Feature Synthesis Tools
Stephen Casper (Massachusetts Institute of Technology), Dylan Hadfield-Menell (Massachusetts Institute of Technology)
CodeExplainability and InterpretabilityAdversarial AttackConvolutional Neural NetworkImageBenchmark
π― What it does: This paper proposes a new evaluation framework for model debugging using interpretability toolsβassessing the effectiveness of feature synthesis tools by rediscovering implanted network interpretability backdoors (trojans).
Reduced Policy Optimization for Continuous Control with Hard Constraints
Shutong Ding (ShanghaiTech University), Ye Shi (ShanghaiTech University)
CodeOptimizationReinforcement LearningBenchmark
π― What it does: A Reduced Policy Optimization (RPO) algorithm is proposed for reinforcement learning that strictly satisfies equality and inequality hard constraints in continuous control tasks.
π― What it does: Given a density field and training images with known camera poses, this paper proposes a closed-form solution to estimate the color field of the scene, thereby achieving the separation of shape fields and radiance fields during rendering.
Reflexion: language agents with verbal reinforcement learning
Noah Shinn (Northeastern University), Shunyu Yao (Princeton University)
CodeReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningTextChain-of-Thought
π― What it does: Proposes the Reflexion framework, which uses self-reflective text from language models as a 'semantic gradient' in the experience buffer to guide subsequent decisions and generation.
Regularization properties of adversarially-trained linear regression
Antonio H. Ribeiro, Thomas B. SchΓΆn (Uppsala University)
CodeOptimizationAdversarial AttackTabular
π― What it does: This paper studies adversarial training in linear regression and analyzes its regularization properties, revealing the equivalence relationships between adversarial training and minimum norm interpolators, Lasso, Ridge, and square root Lasso in different parameter ranges.
π― What it does: Proposes the Fast and Forgetful Memory (FFM) model as a pluggable alternative to RNNs, achieving faster training and higher rewards in reinforcement learning through a memory mechanism;
Reinforcement Learning with Simple Sequence Priors
Tankred Saanum (Max Planck Institute for Biological Cybernetics), Eric Schulz (Max Planck Institute for Biological Cybernetics)
CodeTransformerReinforcement LearningSequential
π― What it does: Introducing sequence compression priors in reinforcement learning encourages agents to produce compressible and predictable action sequences, thereby accelerating learning, enhancing returns, and improving robustness to noise.
Relative Entropic Optimal Transport: a (Prior-aware) Matching Perspective to (Unbalanced) Classification
Liangliang Shi (Shanghai Jiao Tong University), Junchi Yan (Shanghai Jiao Tong University)
CodeClassificationRepresentation LearningImage
π― What it does: This paper views classification as a matching problem and proposes Relative Entropy Regularized Optimal Transport (RE-OT) and its inverse variant for long-tail classification and representation learning.
Reliable Off-Policy Learning for Dosage Combinations
Jonas Schweisthal (Ludwig Maximilian University of Munich), Stefan Feuerriegel (Ludwig Maximilian University of Munich)
CodeOptimizationDrug DiscoveryReinforcement LearningBiomedical DataElectronic Health Records
π― What it does: A reliable off-policy learning framework is proposed to address the dose combination problem in personalized medicine, capable of estimating joint dose effects, detecting limited overlap regions, and optimizing individual dose schemes under these constraints.
Repetition In Repetition Out: Towards Understanding Neural Text Degeneration from the Data Perspective
Huayang Li (Nara Institute of Science and Technology), Yixuan Su (University of Cambridge)
CodeGenerationData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningTextBiomedical Data
π― What it does: This study investigates the fundamental causes of the degeneration phenomenon in neural text generation, finding a high correlation between the proportion of repeated words in the training data and the rate of generated repetitions. It proposes a method to suppress the model's reliance on repetitions by randomly masking repeated n-grams in the attention mechanism (Repetition Dropout).
π― What it does: A multi-choice learning variant named rMCL is proposed, which addresses the issues of overconfidence and hypothesis collapse in regression tasks using a learnable scoring mechanism.
π― What it does: This paper proposes a physics-driven unrolled deep learning framework that utilizes synthetic noise images, random field maps, fat-water ratios, and coil sensitivity data to train the network for deblurring frequency shifts and fat-water separation in non-Cartesian spiral MRI acquisitions.
Response Length Perception and Sequence Scheduling: An LLM-Empowered LLM Inference Pipeline
Zangwei Zheng (National University of Singapore), Yang You (National University of Singapore)
CodeTransformerLarge Language ModelSupervised Fine-TuningText
π― What it does: A sequence scheduling scheme based on LLM for predicting response length and grouping by length is proposed to significantly improve LLM inference throughput.
π― What it does: A unified 'Responsible AI (RAI) Game' framework is proposed, which reduces various responsible AI objectives such as fairness, robustness, and subgroup balance to a minimax loss problem over a set of predefined distributions. Two types of solving algorithms are provided: game-based online learning methods and boosting-based greedy methods.
π― What it does: An efficient diffusion model called ResShift is designed for image super-resolution tasks, compressing the sampling steps from the common thousands to just 15 steps.
π― What it does: This paper reveals the impact of model architecture and hyperparameters on the bias of face recognition systems through large-scale experiments, and proposes an automatic method using Neural Architecture Search (NAS) + Hyperparameter Optimization (HPO) to find network structures that are both highly accurate and fair, significantly reducing gender and racial bias.
Rethinking Conditional Diffusion Sampling with Progressive Guidance
Anh-Dung Dinh (University of Sydney), Chang Xu (University of Sydney)
CodeGenerationData SynthesisLarge Language ModelDiffusion modelImage
π― What it does: An improved classifier guidance method called Progressive Guidance is proposed for the sampling process of diffusion models, aimed at alleviating the issues of diversity compression and adversarial feature construction in traditional classifier guidance.
π― What it does: A theoretical framework based on bias-variance decomposition is proposed to address the issue of class imbalance in node classification within graph neural networks, utilizing graph data augmentation to estimate model variance and incorporating a regularization term;
Rethinking the Backward Propagation for Adversarial Transferability
Xiaosen Wang (Huawei), Kun He (Huazhong University of Science and Technology)
CodeAdversarial AttackImage
π― What it does: The attack modifies the backpropagation process to reduce the truncation of gradients caused by nonlinear layers (ReLU, max-pooling), thereby enhancing the transferability of the generated adversarial samples across different models.
π― What it does: This paper explores the key components in Masked Graph Modeling (MGM) for molecular graphsβgraph tokenizers and decodersβand proposes the SimSGT framework.
ReTR: Modeling Rendering Via Transformer for Generalizable Neural Surface Reconstruction
Yixun Liang (Hong Kong University of Science and Technology), Ying-Cong Chen (Hong Kong University of Science and Technology)
CodeTransformerPoint Cloud
π― What it does: A Transformer-based rendering framework called ReTR is proposed, improving the traditional volume rendering process to achieve generalizable neural surface reconstruction.
π― What it does: This paper proposes an optimal transport-based retrieval-enhanced multi-instance learning framework, RAM-MIL, aimed at improving weakly supervised classification of pathological images.
CodeClassificationComputational EfficiencyRepresentation LearningAdversarial AttackRecurrent Neural NetworkTransformerLarge Language ModelText
π― What it does: A new multilingual text vectorizer, RETVec, has been developed, utilizing a custom UTF-8 binary character encoding and an optional lightweight model to embed words into a 256-dimensional space, achieving robustness against misspellings and adversarial attacks.
Reversible and irreversible bracket-based dynamics for deep graph neural networks
Anthony Gruber (Sandia National Laboratories), Nathaniel Trask (University of Pennsylvania)
CodeGraph Neural NetworkAuto EncoderGraphPhysics Related
π― What it does: A deep graph neural network architecture based on structure-preserving bracket dynamics is proposed, utilizing external calculus to map graph attention to energy-conserving or dissipative dynamics;
π― What it does: This paper re-evaluates the performance of the classic knowledge distillation method (vanilla KD) on large-scale datasets and points out that evaluations on small-scale datasets underestimate its capabilities.
Revisit Weakly-Supervised Audio-Visual Video Parsing from the Language Perspective
Yingying Fan (Wuhan University), Yutian Lin (Wuhan University)
CodeRecognitionSegmentationData-Centric LearningPrompt EngineeringVision Language ModelContrastive LearningVideoMultimodalityAudio
π― What it does: A weakly supervised audio-visual video parsing method based on language prompts is proposed, which significantly reduces segment-level noise by constructing various event occurrence prompts and utilizing CLIP/CLAP for segment-level label inference and dynamic weighting.
π― What it does: This paper conducts an in-depth study on the robust fairness issue of student models in adversarial robust distillation (ARD) and proposes a class difficulty-based reweighting strategy called Fair-ARD to enhance the worst-case robustness of student models across different classes.
π― What it does: The paper re-evaluates the effectiveness of adversarial training on ImageNet, systematically comparing two modern architectures: ViT and ConvNeXt, and enhancing adversarial robustness through improvements in model stem, pre-training, data augmentation, and training schemes.
Revisiting Implicit Differentiation for Learning Problems in Optimal Control
Ming Xu (Australian National University), Stephen Gould (Australian National University)
CodeOptimizationTime Series
π― What it does: A new IDOC method is proposed, which directly solves the trajectory derivatives of constrained discrete-time optimal control problems using the implicit function theorem, avoiding traditional Riccati recursion.
Revisiting Logistic-softmax Likelihood in Bayesian Meta-Learning for Few-Shot Classification
Tianjun Ke (Renmin University of China), Feng Zhou (Renmin University of China)
CodeClassificationMeta LearningImage
π― What it does: This paper redesigns the logistic-softmax likelihood function with a temperature parameter and applies it to a deep kernel-based Gaussian process meta-learning framework, proposing a data augmentation-based analytical mean-field approximate inference method.
π― What it does: In this study, the authors built upon the TD3+BC minimal offline reinforcement learning algorithm by incorporating a series of implementation detail improvements that have emerged in recent years, resulting in a new offline RL method called ReBRAC.
Reward Scale Robustness for Proximal Policy Optimization via DreamerV3 Tricks
Ryan Sullivan (University of Maryland), Joseph Suarez (Massachusetts Institute of Technology)
CodeOptimizationReinforcement LearningTabular
π― What it does: This study transfers various stability techniques from DreamerV3 to PPO and evaluates their impact on the performance and robustness of PPO.
Reward-Directed Conditional Diffusion: Provable Distribution Estimation and Reward Improvement
Hui Yuan (Princeton University), Mengdi Wang (Princeton University)
CodeGenerationReinforcement Learning from Human FeedbackReinforcement LearningDiffusion modelScore-based ModelImage
π― What it does: This paper proposes a semi-supervised learning framework: first, it learns the reward function using a small amount of data with noisy reward labels, then it generates pseudo-labels for a large amount of unlabeled data, and subsequently trains a reward-conditioned diffusion model with the pseudo-labels to achieve 'reward-guided' generation; theoretical guarantees for this method in reward distribution estimation and reward improvement are also provided.
π― What it does: In continual reinforcement learning, a method is proposed to reconnect the network through learnable permutations of hidden layer neurons, achieving a synergy between structural plasticity and weight learning.
π― What it does: A concept curation pipeline is proposed, utilizing CLIP to bridge the semantic gap between images and text through vision-driven expansion, text-to-vision guided relevance ranking, and clustering sampling, in order to enhance the zero-shot performance of language-supervised semantic segmentation.
π― What it does: A regression-guided multi-instance learning network (RGMIL) is proposed, utilizing a new aggregator called Regressor-Guided Pooling (RGP) to achieve high-quality instance-level representation and bag-level classification.
π― What it does: The RH-BrainFS model is proposed, which extracts regional features through brain graph networks and utilizes a Transformer combined with a fusion bottleneck to integrate structural (SC) and functional (FC) brain networks, addressing the issue of regional heterogeneity between the two modalities.
Risk-Averse Model Uncertainty for Distributionally Robust Safe Reinforcement Learning
James Queeney (Boston University), Mouhacine Benosman (Mitsubishi Electric Research Laboratories)
CodeSafty and PrivacyReinforcement Learning
π― What it does: This paper proposes a safe reinforcement learning framework under model uncertainty (RAMU), which processes model distribution with risk aversion, ensuring that the learned policy maintains safety constraints while achieving robust performance during deployment.
RiskQ: Risk-sensitive Multi-Agent Reinforcement Learning Value Factorization
Siqi Shen (Xiamen University), Cheng Wang (Xiamen University)
CodeReinforcement LearningSequential
π― What it does: This paper proposes a risk-sensitive multi-agent reinforcement learning value decomposition framework called RiskQ, and addresses the issue of traditional value decomposition methods being unable to meet the requirements of collaborative decision-making in the face of risk measures (such as VaR and distorted risk measures) by defining the Risk-Sensitive Individual-Global Maximization (RIGM) principle.
π― What it does: An adaptive sampling and state space encoding framework based on reinforcement learning is proposed, capable of achieving real-time path tracking at low sampling rates.
Artun Saday (Bilkent University), Cem Tekin (Bilkent University)
CodeOptimizationReinforcement Learning from Human FeedbackReinforcement LearningTabular
π― What it does: A Robust Bayesian Satisficing (RBS) framework and RoBOS algorithm are proposed, achieving robust satisfaction against unknown distribution shifts using a threshold Ο;
π― What it does: A new target KRaM based on distance metric learning is proposed to remove specific concepts from distributed representations while preserving as much other information as possible.
Robust Contrastive Language-Image Pretraining against Data Poisoning and Backdoor Attacks
Wenhan Yang (University of California Los Angeles), Baharan Mirzasoleiman (University of California Los Angeles)
CodeRepresentation LearningAdversarial AttackData-Centric LearningTransformerVision Language ModelContrastive LearningImageTextMultimodality
π― What it does: In response to the vulnerability of multimodal vision-language models to target data poisoning and backdoor attacks during the pre-training phase, the ROCLIP method is proposed to achieve robust pre-training;
π― What it does: The study designs and analyzes an unbiased covariance estimation method in the presence of missing values and cell-level contamination, combining it with various detection filtering and imputation strategies to achieve robust covariance estimation in high dimensions.
π― What it does: A robust method for assessing the importance of training data is proposedβthe weighted Banzhaf value, along with its theoretical and empirical analysis.
Robust Knowledge Transfer in Tiered Reinforcement Learning
Jiawei Huang (ETH Zurich), Niao He (ETH Zurich)
CodeReinforcement LearningTabular
π― What it does: A Tiered RL framework is proposed that can learn in parallel and achieve robust knowledge transfer when the source task and target task are not completely identical.
π― What it does: The Progressive Data Expansion (PDE) algorithm is proposed, which studies the learning process of deep learning models in the presence of spurious correlated features and enhances the model's robustness to spurious correlations through a two-stage training approach.
π― What it does: A robust low-rank training algorithm is proposed that maintains near-orthogonal constraints through singular value constraints during low-rank network training.
Robust Model Reasoning and Fitting via Dual Sparsity Pursuit
Xingyu Jiang (Huazhong University of Science and Technology), Jiayi Ma (Wuhan University)
CodeAnomaly DetectionOptimizationImage
π― What it does: A unified Dual Sparse Tracking (DSP) method is proposed for robust inference and fitting of geometric models in the presence of a large number of outliers and unknown model types.
π― What it does: The ERNIE framework is proposed, utilizing adversarial regularization to control the Lipschitz constant of the policy, thereby enhancing robustness in multi-agent reinforcement learning, with its effectiveness proven both theoretically and experimentally.
CodeReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningText
π― What it does: A new learning paradigm called RRHF is proposed, which aligns the output of language models with human preferences by learning to rank responses generated from multiple sources.
RS-Del: Edit Distance Robustness Certificates for Sequence Classifiers via Randomized Deletion
Zhuoqun Huang (University of Melbourne), Benjamin I. P. Rubinstein (University of Melbourne)
CodeAnomaly DetectionSequential
π― What it does: A random deletion (RS-Del) random smoothing method for discrete sequence classifiers is proposed, with a proof of robustness in terms of edit distance.
π― What it does: A high-order channel interaction convolution operator, Rubik's Cube Convolution, is proposed as a lightweight alternative to standard convolution in image restoration tasks.
SAMoSSA: Multivariate Singular Spectrum Analysis with Stochastic Autoregressive Noise
Abdullah Omar Alomar, Devavrat Shah (Massachusetts Institute of Technology)
CodeTime Series
π― What it does: A two-stage algorithm SAMoSSA is proposed for simultaneously estimating time-varying deterministic components and AR correlated noise in multivariate time series data.
π― What it does: A sample-efficient multi-objective molecular optimization framework is proposed, utilizing a hypernetwork-based GFlowNet (HN-GFN) as a sampler in Bayesian optimization (MOBO) to sample diversified batches of molecules from the approximate Pareto front.
Sampling from Gaussian Process Posteriors using Stochastic Gradient Descent
Jihao Andreas Lin (University of Cambridge), Alexander Terenin (Cornell University)
CodeOptimizationTabular
π― What it does: This paper proposes a method for approximately solving the posterior mean and sampling of Gaussian processes using stochastic gradient descent (SGD), and extends it to the setting of inducing points.
SatLM: Satisfiability-Aided Language Models Using Declarative Prompting
Xi Ye (University of Texas at Austin), Greg Durrett (University of Texas at Austin)
CodeLarge Language ModelPrompt EngineeringTextChain-of-Thought
π― What it does: Proposes the SATLM framework, which transforms natural language inference tasks into SAT problems and utilizes SAT solvers to complete the reasoning.
Saving 100x Storage: Prototype Replay for Reconstructing Training Sample Distribution in Class-Incremental Semantic Segmentation
Jinpeng Chen (City University of Hong Kong), Sam Kwong (Lingnan University)
CodeSegmentationConvolutional Neural NetworkImage
π― What it does: The STAR method is proposed, which utilizes the compact prototypes of each old foreground class and the repetition of background pixels to reconstruct the complete class distribution of the single-step training set, thereby addressing the issues of class distribution bias and background shift, significantly improving the performance of class incremental semantic segmentation.
Scalable Primal-Dual Actor-Critic Method for Safe Multi-Agent RL with General Utilities
Donghao Ying (University of California Berkeley), Javad Lavaei (University of California Berkeley)
CodeSafty and PrivacyReinforcement Learning
π― What it does: A scalable Primal-Dual Actor-Critic method is proposed to address the Safe Multi-Agent Reinforcement Learning (Safe MARL) problem under the condition of no global observation, utilizing shadow rewards and ΞΊ-hop policies, and supporting general utility forms of objectives and constraints.
Zijie Li (Carnegie Mellon University), Amir Barati Farimani (Carnegie Mellon University)
CodeTransformerMeshPhysics Related
π― What it does: A scalable Transformer structure (FactFormer) is proposed for proxy modeling of PDEs, which maps high-dimensional functions to one-dimensional sub-functions through learned projections, and employs axial factorization kernel integrals to achieve efficient attention computation.
Scale-teaching: Robust Multi-scale Training for Time Series Classification with Noisy Labels
Zhen Liu (South China University of Technology), Qianli Ma (South China University of Technology)
CodeClassificationConvolutional Neural NetworkTime Series
π― What it does: Proposes the Scale-Teaching framework, which utilizes multi-scale temporal data and a cross-teacher mechanism to handle time series classification tasks with noisy labels.
Niklas Muennighoff (Hugging Face), Colin Raffel (Hugging Face)
CodeComputational EfficiencyData-Centric LearningTransformerLarge Language ModelText
π― What it does: In scenarios with limited unique data, we systematically explore the scale of language models and computational allocation, proposing and validating an extended Chinchilla scaling law for data duplication, and conducting over 400 experiments to quantify the benefits and diminishing returns of multi-round training.
π― What it does: A Bayesian optimization method for multi-order precision hyperparameter optimization based on a set of deep power law functions (Deep Power Law, DPL) is proposed.
CodeTransformerLarge Language ModelMultimodalityMagnetic Resonance ImagingAudio
π― What it does: This study investigates the scaling laws of large language models and audio models in fMRI brain encoding, assessing the impact of model parameters and training data volume on encoding performance.
Gregor Bachmann (ETH Zurich), Thomas Hofmann (ETH Zurich)
CodeClassificationRecognitionImage
π― What it does: This study systematically evaluates the performance of Multi-Layer Perceptrons (MLP) in computer vision tasks, including training from scratch, transfer learning, large-scale pre-training, and scale law analysis.
π― What it does: This paper improves the performance of open vocabulary object detection through large-scale self-supervised training, proposing the OWLv2 structure and the OWL-ST training process.
π― What it does: A numerical improvement to the Riemannian diffusion model is proposed, enabling efficient computation of the heat kernel on symmetric spaces, thereby achieving diffusion generation on high-dimensional manifolds.
π― What it does: A score-based data assimilation method (SDA) is proposed, which approximates the score of an entire long trajectory by learning the scores of short segments, enabling one-time parallel generation/inference of state trajectories of arbitrary lengths; at the same time, the observation model is decoupled from the score network, using observation information only during inference, thus supporting zero-shot observation scenarios.
π― What it does: Developed a SE(3) equivariant augmented coupling flow model for efficiently generating molecular conformations in Cartesian coordinates and learning the complete Boltzmann distribution.
π― What it does: The SEENN method is proposed, which dynamically adjusts the time steps for each input sample during SNN inference to achieve early exit.
π― What it does: This paper proposes the Seal framework, which utilizes semantic superpixels generated by 2D vision foundation models (such as SAM, X-Decoder, etc.) to perform unsupervised cross-modal contrastive learning on automotive point cloud sequences, addressing the issues of high annotation costs, cross-modal consistency, and generalization.
π― What it does: Using pre-trained NeRF as a 3D prior, combined with SAM for single-view manual click prompts, we propose the SA3D framework. Through an iterative process of 'mask inverse rendering' and 'cross-view self-prompting', the 2D segmentation results are projected onto a 3D voxel grid to achieve segmentation of 3D objects.
π― What it does: An improved Segment Anything Model (SAM) is proposed, named HQ-SAM, which retains the original prompt and zero-shot capabilities of SAM while generating higher quality and more refined boundary segmentation masks.
π― What it does: A selective forgetting framework based on continual learning, called Selective Amnesia (SA), is proposed, which can delete or remap specified concepts in pre-trained conditional variational generative models (such as VAE, DDPM, Stable Diffusion) without retraining the model or accessing the original data.