These 1376 NeurIPS 2023 papers come with a code repository. Each shows an AI one-line summary below — get the verified repo link + the full 6-part summary (innovation, method, data, results, limitations) and search every NeurIPS 2023 paper, free trial on arXivSub.
“Why Not Looking backward?” A Robust Two-Step Method to Automatically Terminate Bayesian Optimization
Shuang Li (Harbin Institute of Technology), Wei Li (Harbin Institute of Technology)
CodeOptimizationReinforcement LearningTabular
🎯 What it does: A two-step automatic termination Bayesian optimization method is proposed, which is based on detecting the search entering a local convex region and using a local reward threshold as a criterion.
🎯 What it does: A continuous momentum-based asynchronous decentralized deep learning training algorithm A CiD 2 2 was researched and implemented, significantly accelerating communication between nodes;
🎯 What it does: A temporal contrastive learning framework named TACO is proposed, which can simultaneously learn latent representations of states and actions and apply them to visual continuous control in reinforcement learning.
$p$-Poisson surface reconstruction in curl-free flow from point clouds
Yesom Park (Seoul National University), Myungjoo Kang (Seoul National University)
CodePoint CloudBenchmark
🎯 What it does: Reconstructing closed surfaces from unordered point clouds through implicit neural representation, resulting in smooth and detail-rich surface geometry.
$p$-value Adjustment for Monotonous, Unbiased, and Fast Clustering Comparison
Kai Klede (Friedrich-Alexander Universität Erlangen-Nürnberg), Bjoern Eskofier (Helmholtz Zentrum München)
CodeImageTabular
🎯 What it does: Proposed and implemented the p-value adjusted Rand index (PMI₂), addressing type II bias in clustering comparison and ensuring monotonicity; also provided two efficient computation schemes: standardized approximation and Monte Carlo approximation, and validated its superiority on various datasets.
🎯 What it does: The first open-world indoor 3D instance segmentation framework, 3D-OWIS, is proposed, capable of recognizing known categories and unknown objects during inference and achieving incremental learning upon obtaining labels.
🎯 What it does: A molecular generation method based on 3D voxel grids, VoxMol, has been developed, using score-based single-step denoising and Langevin walk-jump sampling to generate 3D molecules.
🎯 What it does: Proposes the 3D-aware VQA task, constructs the Super-CLEVR-3D dataset, and designs the PO3D-VQA model for 3D structure reasoning and answer generation.
3D-IntPhys: Towards More Generalized 3D-grounded Visual Intuitive Physics under Challenging Scenes
Haotian Xue (Georgia Institute of Technology), Hsiao-Yu Tung (Massachusetts Institute of Technology)
CodeObject DetectionSegmentationData SynthesisGraph Neural NetworkNeural Radiance FieldVideoPoint CloudPhysics Related
🎯 What it does: The 3D-IntPhys framework is proposed to unsupervisedly reconstruct 3D point clouds from multi-view RGB images and instance masks, and to predict the physical dynamics of point clouds using graph neural networks, achieving long-term predictions for complex materials such as fluids, particles, and rigid bodies.
3D-LLM: Injecting the 3D World into Large Language Models
Yining Hong (University of California), Chuang Gan (Massachusetts Institute of Technology)
CodeObject DetectionGenerationRetrievalTransformerLarge Language ModelVision Language ModelTextPoint Cloud
🎯 What it does: A novel 3D-LLM is proposed, capable of inputting 3D point clouds and their features along with language prompts to perform various 3D-related tasks;
Jingkang Yang (Nanyang Technological University), Ziwei Liu (Hong Kong Baptist University)
CodeSegmentationGenerationRobotic IntelligenceTransformerLarge Language ModelVideoPoint Cloud
🎯 What it does: This paper proposes the 4D Panoramic Scene Graph (PSG-4D) task and constructs a panoramic segmentation and dynamic relationship joint generation framework called PSG4DFormer, which includes RGB-D videos and point cloud sequences, and demonstrates collaborative decision-making with large language models in real service robots.
A Bayesian Approach To Analysing Training Data Attribution In Deep Learning
Elisa Nguyen (Tübingen AI Center University of Tübingen), Seong Joon Oh (Tübingen AI Center University of Tübingen)
CodeConvolutional Neural NetworkTransformerImage
🎯 What it does: A reliability assessment of Training Data Attribution (TDA) is conducted from a Bayesian perspective, quantifying its noise and variance, and comparing various approximation methods.
Enrico Giudice (University of Basel), Giusi Moffa (University College London)
CodeGaussian SplattingTabular
🎯 What it does: This paper proposes a method for complete Bayesian structure learning on Gaussian Process Networks, utilizing MCMC sampling, Laplace approximation, and importance sampling to achieve computable posterior distributions.
🎯 What it does: A sparse online Newton method (SONew) is proposed, achieving linear time/space second-order preconditioners through LogDet divergence and structured sparse graphs;
A Cross-Moment Approach for Causal Effect Estimation
Yaroslav Kivva (Ecole Polytechnique Federale de Lausanne), Negar Kiyavash (Ecole Polytechnique Federale de Lausanne)
CodeTabular
🎯 What it does: In a linear structural causal model (SCM) with only a single proxy variable and potential confounding, a method based on three-variable cross-moments is proposed to estimate the causal effect of the treatment variable on the outcome variable.
🎯 What it does: The G2MILP framework is proposed, which uses deep generative models to generate new mixed-integer linear programming instances under limited data conditions.
🎯 What it does: A dual-stream neural network inspired by the dual pathways (dorsal/ventral) of the human visual system is proposed and implemented. It utilizes retinal transformations to generate different field inputs, with WhereCNN learning spatial attention and eye movement control, and WhatCNN learning object recognition, recursively constructing scene representations through multiple fixations.
A Framework for Fast and Stable Representations of Multiparameter Persistent Homology Decompositions
David Loiseaux (Inria d'Université Côte d'Azur), Andrew Blumberg
CodeClassificationComputational EfficiencyRepresentation LearningPoint CloudGraphTime Series
🎯 What it does: This paper proposes a general framework T-CDR for representing multi-parameter persistent homology decompositions as vectors, and based on this, provides a provably stable S-CDR representation.
A General Framework for Equivariant Neural Networks on Reductive Lie Groups
Ilyes Batatia (University of Cambridge), Christoph Ortner (University of British Columbia)
CodeRecognitionPoint Cloud
🎯 What it does: A general G-Equivariant Cluster Expansion and G-MACE architecture is proposed, which can construct equivariant neural networks for finite-dimensional representations of any simple Lie group, and the lie-nn library implementing this framework has been released.
🎯 What it does: This paper proposes a convolutional neural network based on Group Triple Correlation (G-TC) layers to achieve group invariance (G-invariance);
🎯 What it does: This paper presents a general theory for analyzing the error limits when there is an imperfect match between real functions and equivariant network symmetry, and introduces the concepts of point-to-point correctness, incorrectness, and external equivariance; it then provides lower bounds on errors for classification, balanced regression, and equivariant regression tasks, and validates the theory through experiments.
🎯 What it does: A biologically interpretable hippocampal generative model based on the Helmholtz machine has been established, utilizing theta rhythm to switch between forward (wake) and backward (sleep) information flow, and employing local synaptic rules for learning to infer and generate networks. This ultimately achieves autoencoding of random high-dimensional perceptual inputs, path integration of 1D trajectories, and structural transfer across environments.
🎯 What it does: A graph theory-based open-world semi-supervised learning framework is proposed, and the Spectral Open-world Representation Learning (SORL) method is designed;
A Heat Diffusion Perspective on Geodesic Preserving Dimensionality Reduction
Guillaume Huguet (Université de Montréal), Smita Krishnaswamy (Yale University)
CodeDiffusion modelTabularBiomedical Data
🎯 What it does: This paper proposes a direct association between thermal diffusion and geodesic distance on manifolds through the Varadhan formula, defining thermal geodesic similarity, and based on this, constructs a Heat Geo dimensionality reduction method;
A Neural Collapse Perspective on Feature Evolution in Graph Neural Networks
Vignesh Kothapalli (New York University), Joan Bruna (New York University)
CodeGraph Neural NetworkGraph
🎯 What it does: This study investigates the feature evolution of Graph Neural Networks (GNN) in node-level classification tasks and quantifies it from the perspective of Neural Collapse (NC).
🎯 What it does: This paper proposes a new local-global hierarchical 3D graph isomorphism framework to evaluate and enhance the expressive power of 3D equivariant graph neural networks, and based on this, designs two main modules: LSE (Local Substructure Encoding) and FTE (Frame Transition Encoding), ultimately realizing the LEFTNet model.
A One-Size-Fits-All Approach to Improving Randomness in Paper Assignment
Yixuan Even Xu (Tsinghua University), Fei Fang (Carnegie Mellon University)
CodeOptimizationText
🎯 What it does: A general random paper allocation method (Perturbed Maximization, PM) is proposed, which significantly enhances randomness while maintaining high matching quality through a concave perturbation function.
🎯 What it does: A regularized conditional GAN (rcGAN) is proposed for quickly sampling the posterior distribution of image recovery problems from constrained measurements with high quality;
A Riemannian Exponential Augmented Lagrangian Method for Computing the Projection Robust Wasserstein Distance
Bo Jiang (Nanjing Normal University), Ya-Feng Liu (Chinese Academy of Sciences)
CodeOptimization
🎯 What it does: This paper proposes a method for calculating the projection robust Wasserstein (PRW) distance optimization on the Stiefel manifold and Euclidean space, which mainly includes the Riemannian Exponential Augmented Lagrangian Method (REALM) and an approximate Riemannian Barzilai-Borwein and Sinkhorn iteration (iRBBS) algorithm used in the subproblems.
🎯 What it does: This paper elevates the common non-convex optimization problems in deep learning to the probability measure space, constructing a strict convex optimization framework, and derives a new class of infinite-dimensional variational inference (ID-GVI) algorithms using Wasserstein gradient flows, unifying various uncertainty quantification methods such as Bayesian, variational Bayesian, and deep ensemble.
A Robust Exact Algorithm for the Euclidean Bipartite Matching Problem
Akshaykumar G Gattani, Pouyan Shirzadian (Virginia Tech)
CodeOptimizationPoint Cloud
🎯 What it does: A divide-and-conquer Hungarian algorithm based on random translation quadtrees is proposed, which can solve the Euclidean bipartite matching (i.e., 1-Wasserstein distance) on a two-dimensional random point set in expected time ˜(n^{7/4} log Δ), and is generalized to arbitrary dimensions.
A Scalable Neural Network for DSIC Affine Maximizer Auction Design
Zhijian Duan (Peking University), Xiaotie Deng (Peking University)
CodeTransformer
🎯 What it does: A scalable neural network, AMenuNet, is proposed for designing multi-item auction mechanisms that satisfy dominant strategy incentive compatibility (DSIC) and individual rationality (IR).
A Simple Solution for Offline Imitation from Observations and Examples with Possibly Incomplete Trajectories
Kai Yan (University of Illinois Urbana-Champaign), Yu-Xiong Wang (University of Illinois Urbana-Champaign)
CodeRobotic IntelligenceReinforcement Learning from Human FeedbackReinforcement LearningAgentic AISequential
🎯 What it does: A simple and effective offline imitation learning framework called TAILO is proposed, which utilizes a discriminator to identify expert states and thresholds the discounted sum of their scores along future trajectories, directly using it as the weights for weighted behavior cloning.
A Smooth Binary Mechanism for Efficient Private Continual Observation
Joel Daniel Andersson (Copenhagen University), Rasmus Pagh (Copenhagen University)
CodeSafty and PrivacyComputational EfficiencyGaussian SplattingTime Series
🎯 What it does: A 'smooth binary mechanism' is proposed, which utilizes balanced binary tree leaves to reduce noise variance in the problem of differentially private counting under continuous observation, while maintaining a consistent error distribution at each time step.
A Theory of Link Prediction via Relational Weisfeiler-Leman on Knowledge Graphs
Xingyue Huang (University of Oxford), Pablo Barcelo
CodeRecommendation SystemGraph Neural NetworkGraphBiomedical Data
🎯 What it does: This paper proposes a Conditional Message Passing Neural Network (C-MPNN) framework for knowledge graph link prediction and conducts a theoretical analysis of its expressive power. It also provides a unified perspective with traditional models such as R-MPNN and NBFNet, and experimentally validates the effectiveness of the design choices.
A Unified Approach to Count-Based Weakly Supervised Learning
Vinay Shukla (University of California), Guy Van den Broeck (University of California)
CodeClassificationOptimizationTabular
🎯 What it does: A unified counting constraint method (Count Loss) is proposed, which constrains weakly supervised labels (such as label proportions, multiple instance learning, and positive-negative unlabeled learning) by calculating the counting distribution of model-predicted labels, directly optimizing probabilities rather than approximations.
A Unified Approach to Domain Incremental Learning with Memory: Theory and Algorithm
Haizhou Shi (Rutgers University), Hao Wang (Rutgers University)
CodeDomain AdaptationKnowledge DistillationImage
🎯 What it does: A unified framework called UDIL is proposed to address the problem of domain incremental learning with limited memory, theoretically unifying existing methods under the same error upper bound and achieving the tightest error bounds through adaptive coefficients.
🎯 What it does: A unified theoretical framework for U-Net is proposed, and based on this framework, various improved U-Net architectures (such as Multi-ResNet, U-Net based on boundary conditions and geometric structures) are designed and applied to tasks such as image segmentation, PDE surrogate modeling, and diffusion models.
A Unified Solution for Privacy and Communication Efficiency in Vertical Federated Learning
Ganyu Wang (Western University), Charles Ling
CodeOptimizationFederated LearningSafty and PrivacyImage
🎯 What it does: A hybrid optimization framework is proposed for vertical federated learning, using zero-order optimization for the output layer and first-order optimization for the remaining layers, combined with compression to achieve a unified solution for privacy protection and communication efficiency.
A-NeSI: A Scalable Approximate Method for Probabilistic Neurosymbolic Inference
Emile van Krieken (University of Edinburgh Vrije Universiteit Amsterdam), Annette Ten Teije (Vrije Universiteit Amsterdam)
CodeExplainability and InterpretabilityComputational EfficiencyImage
🎯 What it does: A scalable approximate reasoning framework named A-NESI is proposed, which combines neural networks and symbolic reasoning to achieve probabilistic neural-symbolic learning.
🎯 What it does: This paper proposes an A3FL backdoor attack for federated learning, utilizing adversarial adaptive triggers to maintain efficiency and persistence of the backdoor in the global training dynamics.
Accelerated Zeroth-order Method for Non-Smooth Stochastic Convex Optimization Problem with Infinite Variance
Nikita Kornilov (Moscow Institute of Physics and Technology), Samuel Horváth (Mohammed Bin Zayed University of Artificial Intelligence)
CodeOptimizationTabular
🎯 What it does: A zero-order acceleration method for non-smooth stochastic convex optimization with infinite variance noise (ZO-clipped-SSTM and R-ZO-clipped-SSTM) is proposed, along with high-probability convergence and optimal operator complexity.
🎯 What it does: This paper enhances the inference speed and prediction accuracy of Molecular Graph Neural Networks (Molecular GNN) through the Knowledge Distillation (KD) method, without altering the student model structure, thus maintaining the high throughput of the lightweight model.
🎯 What it does: A Probability Tree State Abstraction (PTSA) algorithm is proposed to accelerate the search process by aggregating similar paths in the MCTS search tree.
Accessing Higher Dimensions for Unsupervised Word Translation
Sida Wang
CodeText
🎯 What it does: A new unsupervised word translation method called coocmap is proposed, which utilizes high-dimensional signals instead of low-dimensional word vectors for translation.
🎯 What it does: Proposes the Cross Modal Generalization (CMG) task and constructs the Uni-Code framework to learn a unified discrete representation across modalities, achieving zero-shot generalization to other modalities with only one modality labeled.
🎯 What it does: A text-driven human motion generation model called GraphMotion is proposed, which is based on a hierarchical semantic graph. By breaking down action descriptions into three layers of nodes: movement, action, and details, it achieves fine-grained control over motion details.
Active Learning for Semantic Segmentation with Multi-class Label Query
Sehyun Hwang (Pohang University of Science and Technology), Suha Kwak (Pohang University of Science and Technology)
CodeSegmentationConvolutional Neural NetworkImage
🎯 What it does: This paper proposes an active learning framework for semantic segmentation, with the core idea being multi-class label querying: within a selected local area, all occurring categories are manually annotated, and then a two-stage training strategy is employed to address label uncertainty.
Christian Lange (University of Edinburgh), Oisin Mac Aodha (University of Edinburgh)
CodeClassificationOptimizationTabularAgriculture Related
🎯 What it does: This paper proposes an active learning method that efficiently estimates the distribution range of unknown species using fewer field observation points by selecting the most uncertain geographical locations for sampling after online updates. This is achieved through weighted averaging in the trained multi-species range model space and using spatial implicit features obtained from transfer learning.
Active Negative Loss Functions for Learning with Noisy Labels
Xichen Ye (Shanghai University), Weiqin Tong (Shanghai University)
CodeClassificationOptimizationImage
🎯 What it does: A new noise-robust learning framework based on 'negative' loss (ANL) is proposed, and various new noise-robust loss functions are constructed based on it.
Samuel Holt (University of Cambridge), Mihaela van der Schaar (University of Cambridge)
CodeOptimizationRobotic IntelligenceReinforcement LearningTime SeriesBiomedical Data
🎯 What it does: The study actively decides when to conduct expensive observations in continuous-time control tasks and proposes an active observation control method based on uncertainty thresholds.
Active representation learning for general task space with applications in robotics
Yifang Chen (University of Washington), Guanya Shi (Carnegie Mellon University)
CodeRepresentation LearningRobotic IntelligenceTabularTime Series
🎯 What it does: This paper proposes a general active representation learning framework that can adaptively select the most informative source tasks in any discrete or continuous source task space to enhance the few-shot learning performance of the target task.
🎯 What it does: A framework named AdANNS is proposed, which utilizes adaptive multi-dimensional embeddings represented by Matryoshka. It employs different dimensional representations at various stages of approximate nearest neighbor search (ANNS), such as clustering, linear search, and quantization, significantly reducing computational costs while maintaining accuracy.
AdaPlanner: Adaptive Planning from Feedback with Language Models
Haotian Sun (Georgia Institute of Technology), Chao Zhang (Georgia Institute of Technology)
CodeRobotic IntelligenceReinforcement Learning from Human FeedbackTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: Proposes AdaPlanner, a framework that adaptively generates and refines action plans using large language models (LLMs) under closed-loop environmental feedback;
🎯 What it does: Predicting answers to complex logical queries on knowledge graphs and achieving more accurate reasoning through score calibration of a pre-trained neural link predictor.
Adaptive Contextual Perception: How To Generalize To New Backgrounds and Ambiguous Objects
Zhuofan Ying (Columbia University), Mohit Bansal (University of North Carolina at Chapel Hill)
CodeObject DetectionDomain AdaptationImage
🎯 What it does: This paper studies how visual models adaptively utilize background information to generalize to new backgrounds and ambiguous objects, and proposes two types of OOD settings (BACKGROUND-INVARIANCE and OBJECT-DISAMBIGUATION), analyzing the performance differences of the model in these two contexts.
Adaptive Normalization for Non-stationary Time Series Forecasting: A Temporal Slice Perspective
Zhiding Liu (University of Science and Technology of China), Enhong Chen (University of Science and Technology of China)
CodeTransformerTime Series
🎯 What it does: An adaptive normalization framework based on time slicing, SAN, is proposed as a model-agnostic plugin to normalize/denormalize various time series forecasting models, significantly improving prediction accuracy.
Adaptive Privacy Composition for Accuracy-first Mechanisms
Ryan Rogers, Aaditya Ramdas (Carnegie Mellon University)
CodeSafty and PrivacyTextStochastic Differential Equation
🎯 What it does: A privacy filter is proposed that can adaptively combine ex-post privacy mechanisms (such as the Brownian noise reduction mechanism) with zCDP mechanisms, thereby achieving more flexible interactive privacy analysis while maintaining overall (ε,δ)-DP.
Adaptive Test-Time Personalization for Federated Learning
Wenxuan Bao (University of Illinois Urbana-Champaign), Jingrui He (University of Illinois Urbana-Champaign)
CodeDomain AdaptationFederated LearningImage
🎯 What it does: This paper proposes an algorithm called ATP (Adaptive Test-Time Personalization) aimed at addressing the model personalization problem for unlabeled test clients in a federated learning environment.
🎯 What it does: A learnable weighted filter network architecture is proposed, which adaptively generates weights for point clouds through neural networks, calculates persistent homology, and vectorizes it for classification.
🎯 What it does: A framework for posterior uncertainty estimation based on high-dimensional hypothesis testing is proposed, utilizing a Bayesian neural network encoder to extract latent representations, and then performing OOD detection on test samples using an adjustable regularized Hotelling T² (ARHT) statistic.
Adaptive whitening with fast gain modulation and slow synaptic plasticity
Lyndon Duong, David Lipshutz (Flatiron Institute)
CodeImage
🎯 What it does: Design and implement a multi-timescale adaptive whitening neural network model that combines synaptic plasticity and rapid gain modulation to achieve fast whitening for different contexts.
🎯 What it does: A self-supervised ranking-based pre-training task (AdaptSSR) is proposed to learn representations of user behavior sequences, reducing reliance on semantically consistent data augmentation.
Addressing the speed-accuracy simulation trade-off for adaptive spiking neurons
Luke Taylor (University of Oxford), Nicol Spencer Harper
CodeOptimizationComputational EfficiencySpiking Neural NetworkTime Series
🎯 What it does: A block-based algorithm reconstruction is proposed, utilizing the Absolute Refractory Period (ARP) of neurons to convert the stepwise simulation of the ALIF model into O(1) parallel computation per block, significantly accelerating simulation and training on GPUs.
🎯 What it does: Proposed the Adversarial Weighted Risk Minimization (AWRM) framework and implemented the GALILEO algorithm for offline environment model learning to enhance adversarial robustness and generalization ability.
🎯 What it does: This paper evaluates the true usability and robustness of robust and non-robust features within a broader learning paradigm (supervised learning, contrastive learning, masked image modeling, diffusion models).
🎯 What it does: A framework called DataFix based on tree models and adversarial learning is proposed to locate and correct feature distribution drift in datasets.
🎯 What it does: This paper proposes and evaluates a class of Hamiltonian-based graph neural flows (HANG and HANG-quad) aimed at enhancing the robustness of graph neural networks against adversarial attacks.
🎯 What it does: This paper proposes applying Adversarial Self-Training (AST) to Gradual Domain Adaptation (GDA) to enhance the adversarial robustness and clean accuracy of the target domain.
🎯 What it does: The Successive Layer Training (SLT) method is proposed to implement federated learning training on resource-constrained edge devices, gradually freezing the front layers and expanding the back head to significantly reduce memory requirements while maintaining parameter cooperation.
AiluRus: A Scalable ViT Framework for Dense Prediction
Jin Li (Shanghai Jiao Tong University), Qi Tian (Huawei Cloud)
CodeObject DetectionSegmentationTransformerImage
🎯 What it does: A pluggable adaptive resolution strategy AiluRus is proposed, which dynamically merges tokens using a spatially aware density clustering algorithm in the intermediate layers of ViT to accelerate dense prediction tasks.
AlberDICE: Addressing Out-Of-Distribution Joint Actions in Offline Multi-Agent RL via Alternating Stationary Distribution Correction Estimation
Daiki E. Matsunaga (Korea Advanced Institute of Science and Technology), Kee-Eung Kim (Korea Advanced Institute of Science and Technology)
CodeReinforcement Learning
🎯 What it does: We propose AlberDICE, an algorithm for offline multi-agent reinforcement learning that utilizes alternating best responses and static distribution correction to avoid the out-of-distribution (OOD) problem of joint actions, overcoming the challenge of exponential explosion in the joint action space.
ALGO: Synthesizing Algorithmic Programs with Generated Oracle Verifiers
Kexun Zhang (University of California Santa Barbara), Lei Li (Carnegie Mellon University)
CodeOptimizationAI Code AssistantTransformerLarge Language ModelText
🎯 What it does: The ALGO framework is proposed, which uses brute-force enumeration reference oracles generated by LLMs to guide and verify the automatic synthesis of algorithm programs.
🎯 What it does: An adaptive algorithm selection framework TAILOR suitable for deep active learning is proposed, which can dynamically select the optimal active learning algorithm across different datasets.
Algorithmic Regularization in Tensor Optimization: Towards a Lifted Approach in Matrix Sensing
Ziye Ma (University of California Berkeley), Somayeh Sojoudi (University of California Berkeley)
CodeOptimization
🎯 What it does: Under the lifted matrix sensing framework, this study investigates the implicit regularization effect of gradient descent (GD) on tensor optimization for approximate rank-1 representations and proves that it can lead to escape directions at constrained points.
🎯 What it does: A Bayesian framework d-PM is proposed to model inconsistencies in human preferences and to calibrate existing NLG models using contrastive learning, in order to generate text that is more universally acceptable and less controversial.
ALIM: Adjusting Label Importance Mechanism for Noisy Partial Label Learning
Mingyu Xu (Chinese Academy of Sciences), Jianhua Tao (Tsinghua University)
CodeClassificationContrastive LearningImage
🎯 What it does: To address the problem of partial label learning with noise, a framework is proposed that adjusts label importance (ALIM) to balance the candidate label set with model predictions, compatible with existing methods and capable of automatically adjusting weights.
🎯 What it does: This study focuses on weakly supervised 3D point cloud semantic segmentation and proposes the ERDA learning strategy, which utilizes entropy regularization and distribution alignment to jointly generate more reliable pseudo-labels, fully leveraging all unlabeled points for training.
Alleviating the Semantic Gap for Generalized fMRI-to-Image Reconstruction
Tao Fang (Zhejiang University), Gang Pan (Zhejiang University)
CodeGenerationData SynthesisMixture of ExpertsDiffusion modelImageBiomedical DataMagnetic Resonance Imaging
🎯 What it does: This paper proposes a general fMRI-to-image reconstruction framework called GESS, aimed at eliminating the semantic gap between training and testing, generating semantically stable and structurally consistent images.
CodeReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: This paper studies a low-cost simulation framework called AlpacaFarm, designed for rapid experimentation and iteration in learning human feedback for instruction-following models.
AmadeusGPT: a natural language interface for interactive animal behavioral analysis
Shaokai Ye (École Polytechnique Fédérale de Lausanne), Mackenzie W Mathis
CodeObject DetectionPose EstimationAI Code AssistantTransformerLarge Language ModelPrompt EngineeringVideo
🎯 What it does: Developed AmadeusGPT, a natural language interface that can automatically convert natural language descriptions of animal behavior into executable Python code, enabling no-code behavior analysis;
🎯 What it does: This paper proposes Ambient Diffusion, a diffusion model framework that can learn clean distributions even when only highly damaged samples are available.
🎯 What it does: An adaptive detection procedure for high-dimensional mediation analysis (AMDP) is proposed, which can efficiently identify significant mediating variables while controlling the FDR.
An Alternative to Variance: Gini Deviation for Risk-averse Policy Gradient
Yudong Luo (University of Waterloo), Yangchen Pan (University of Oxford)
CodeReinforcement Learning
🎯 What it does: This paper proposes the use of Gini Deviation as an alternative risk measure to variance, deriving and implementing the gradient estimation of Gini Deviation in mean-risk reinforcement learning based on policy gradients, and validating its feasibility on various discrete and continuous control tasks.
🎯 What it does: This paper proposes the CPTPP framework, which combines graph contrastive learning pre-training with prompt tuning. In recommendation tasks, user/item embeddings are first pre-trained using GCL, then personalized soft prompts are automatically generated based on graph interactions, and finally, downstream recommendations are completed through prompt fusion.
🎯 What it does: A matrix entropy evaluation method based on Rènyi kernel entropy is proposed to measure the diversity of generative models under multimodal distributions.
Xianhang Li (University of California Santa Cruz), Cihang Xie (University of California Santa Cruz)
CodeClassificationRetrievalComputational EfficiencyTransformerVision Language ModelContrastive LearningImageText
🎯 What it does: The study investigates and verifies the inverse scaling law of CLIP training, proposing that fewer image/text tokens can be used for efficient training under large models, and based on this, the CLIPA training framework is designed.
An Optimal and Scalable Matrix Mechanism for Noisy Marginals under Convex Loss Functions
Yingtai Xiao (Penn State University), Daniel Kifer (Penn State University)
CodeOptimizationSafty and PrivacyTabularFinance Related
🎯 What it does: A matrix mechanism named ResidualPlanner is proposed, which can provide unbiased, optimal, and scalable privacy-preserving answers for multidimensional marginal queries under Gaussian random noise.
🎯 What it does: Theoretical analysis and large-scale experimental validation of the sample complexity of self-supervised methods based on unbiased gradient estimation in deep learning image reconstruction.
🎯 What it does: This study investigates how to project the internal representations of Vision Transformers into class embedding space, thereby reverse engineering image classification networks.
Approximate Allocation Matching for Structural Causal Bandits with Unobserved Confounders
Lai Wei (Life Sciences Institute University of Michigan), Murat Kocaoglu (Purdue University)
CodeReinforcement LearningGraph
🎯 What it does: This paper proposes a more efficient decision-making strategy for the structural causal bandit problem with unobserved confounding variables by utilizing a known causal graph.
Ningyuan Teresa Huang, Soledad Villar (Johns Hopkins University)
CodePose EstimationGraph Neural NetworkImageGraph
🎯 What it does: This study investigates active symmetry on fixed graphs, proposing a covariant graph network obtained through graph coarsening to approximate symmetry groups, and provides a risk-bias-variance balance formula.
🎯 What it does: This paper combines Quality Diversity (QD) algorithms with Neural Cellular Automata (NCA) to train multi-robot system environments that can generate regular layouts in environments of any scale, significantly improving system throughput.
🎯 What it does: This paper proposes a framework called DiffusionITM that transforms diffusion models (such as Stable Diffusion) into image-text matching (ITM) models, and constructs the GDBench benchmark to evaluate the reasoning capabilities of diffusion models across various visual-language tasks. Additionally, it enhances the model's discriminative performance by fine-tuning with hard negative samples on MS-COCO.
🎯 What it does: This study investigates the gradient flow dynamics of Graph Attention Networks (GAT), deriving a gradient conservation law that reveals the issue of untrainability in deep GATs caused by standard initialization, and proposes a balanced initialization scheme.
🎯 What it does: A self-regressive tree topology generation model ARTree based on graph neural networks is proposed for learning the probability distribution of evolutionary trees and variational Bayesian inference.
ASIF: Coupled Data Turns Unimodal Models to Multimodal without Training
Antonio Norelli (Sapienza Università di Roma), Francesco Locatello (Institute of Science and Technology Austria)
CodeClassificationRetrievalRepresentation LearningTransformerVision Language ModelContrastive LearningImageTextMultimodality
🎯 What it does: This paper proposes the ASIF method, which aligns pre-trained unimodal image and text encoders with a small number of image-text pairs into a multimodal model without requiring any additional training.