π― What it does: The SP-gra2seq method is proposed, which generates robust graphical sketch representations by learning semantic similarity (synonym proximity) to dynamically connect sketch patches and perform graph convolution message passing, supplemented by clustering constraints.
LIQUID: A Framework for List Question Answering Dataset Generation
Seongyun Lee (Korea University), Jaewoo Kang (Korea University)
CodeGenerationData SynthesisTransformerLarge Language ModelSupervised Fine-TuningTextBiomedical Data
π― What it does: Proposes the LIQUID framework, which can automatically generate multi-answer list question-answer datasets from unlabeled texts (such as Wikipedia and PubMed), significantly reducing the cost of manual annotation.
Rohan Ghosh (National University of Singapore), Mehul Motani (National University of Singapore)
CodeConvolutional Neural NetworkImage
π― What it does: A continuous space entropy measure based on local intrinsic dimension, called ID-Entropy, is proposed, and it is proven to satisfy the main properties of discrete entropy, thereby constructing a new information bottleneck criterion and a generalized error upper bound.
Peiyu Yang (University of Western Australia), Ajmal Mian (University of Western Australia)
CodeExplainability and InterpretabilityConvolutional Neural NetworkImage
π― What it does: This study proposes the Local Path Integration (LPI) method, which improves traditional path attribution methods by integrating gradients using reference samples from the local distribution of the input, addressing shortcomings in weak dependence and reference selection.
Logic and Commonsense-Guided Temporal Knowledge Graph Completion
Guanglin Niu (Beihang University), Bo Li (Beihang University)
CodeGraphTime Series
π― What it does: This paper proposes a temporal knowledge graph completion model LCGE that simultaneously considers event timeliness, causality, and common sense.
π― What it does: This paper proposes LONE SAMPLER, a method for generating discrete node embeddings through coordinated local neighborhood sampling, utilizing the local structure of the graph and sampling theory to construct interpretable and scalable node representations.
π― What it does: A weakly supervised learning framework (Losses over Labels, LoL) is proposed that directly converts weak labelers into loss functions, eliminating the need for generating pseudo-labels.
Lottery Pools: Winning More by Interpolating Tickets without Increasing Training or Inference Cost
Lu Yin (Eindhoven University of Technology), Mykola Pechenizkiy (University of Liverpool)
CodeOptimizationConvolutional Neural NetworkImage
π― What it does: By performing linear interpolation on Lottery Ticket subnetworks obtained from different iterations, stronger sparse subnetworks are constructed, and these interpolated subnetworks are then interpolated back to the dense network, resulting in performance improvements.
Low Resource Quantitative Information Extraction via Structure Searching and Prefix-Based Text Generation
Tongliang Li (Beihang University), Zhoujun Li (Beihang University)
CodeGenerationData SynthesisTransformerLarge Language ModelPrompt EngineeringText
π― What it does: In low-resource environments, this paper proposes to automatically generate training samples through constituency parse tree search and use a prefix text generation model to extract quantitative information.
Low-Resource Personal Attribute Prediction from Conversations
Yinan Liu (Nankai University), Jiaoyan Chen (University of Manchester)
CodeClassificationRecommendation SystemTransformerLarge Language ModelText
π― What it does: Under low-resource conditions, the PEARL framework is proposed to predict user personal attributes using unannotated dialogue data in conversations.
LWSIS: LiDAR-Guided Weakly Supervised Instance Segmentation for Autonomous Driving
Xiang Li (Beijing Institute of Technology), Jianbing Shen (University of Macau)
CodeSegmentationAutonomous DrivingPoint Cloud
π― What it does: Using LiDAR point clouds and 3D boxes as weak supervision to train a 2D instance segmentation model, reducing the need for 2D mask annotations.
π― What it does: A Reddit personal narrative dataset named STORIES was constructed, and the M-sense model was proposed to automatically identify the climax and conclusion in narratives.
π― What it does: A full-process framework based on a multi-modal masked autoencoder (MΒ³AE) is proposed for brain tumor segmentation with missing modalities, including self-supervised pre-training, model inversion to complete missing modalities, and memory-efficient self-distillation.
π― What it does: Introducing Model Augmentation (MA-GCL) in graph contrastive learning, which generates more diverse contrastive views by altering the architecture of the view encoder, enhancing the effectiveness of unsupervised node representation learning.
MAPS-KB: A Million-Scale Probabilistic Simile Knowledge Base
Qianyu He (Fudan University), Yanghua Xiao (Fudan University)
CodeGenerationRetrievalTransformerLarge Language ModelSupervised Fine-TuningText
π― What it does: The first million-level probabilistic personification (simile) knowledge base MAPS-KB has been constructed, covering 4.3M (topic, property, vehicle) triples, providing multi-dimensional probabilistic information (credibility, typicality), and achieving SOTA in three types of downstream tasks (personification explanation, generation, and text polishing).
Materialisation-Based Reasoning in DatalogMTL with Bounded Intervals
PrzemysΕaw A. WaΕΔga (University of Oxford), Bernardo Cuenca Grau (University of Oxford)
CodeTabularTime SeriesBenchmark
π― What it does: For finite interval programs in DatalogMTL, a materialized and terminating reasoning algorithm is proposed, which achieves model unfolding and fact reasoning by detecting saturation states.
π― What it does: This paper proposes a Multi-Granularity Contrastive Learning (MCL) framework that integrates dictionary information into the character-word grid structure for Chinese named entity recognition tasks, achieving mutual calibration of character and word representations and highlighting keywords through Cross-Granularity Contrastive Learning (CCL) and Bi-Granularity Contrastive Learning (BCL).
π― What it does: A diffusion model for 3D molecular generation, MDM, has been developed, capable of generating high-quality 3D molecules from scratch.
π― What it does: Fine-tune normal samples using the new Mean-Shifted Contrastive Loss on pre-trained ImageNet features to enhance single-class anomaly detection performance.
MEID: Mixture-of-Experts with Internal Distillation for Long-Tailed Video Recognition
Xinjie Li (Pennsylvania State University), Huijuan Xu (Pennsylvania State University)
CodeRecognitionKnowledge DistillationMixture of ExpertsVideoBenchmark
π― What it does: A two-expert mixture framework based on internal distillation has been designed and implemented to address the frame-level imbalance problem in multi-label long-tail video recognition using frame-level attention and complementary frame selection.
π― What it does: This paper proposes a memory contrast-based co-salient object detection framework called MCCL, which achieves real-time high-precision detection using a Transformer encoder along with three main modules: GCAM, MCM, and AIL.
Meta-Sketch: A Neural Data Structure for Estimating Item Frequencies of Data Streams
Yukun Cao (University of Science and Technology of China), Xike Xie (University of Science and Technology of China)
CodeMeta LearningTime Series
π― What it does: A 'Meta-Sketch' data structure based on meta-learning and memory-augmented neural networks is proposed for estimating the frequency of items in data streams within limited space.
π― What it does: The multi-language code representation learning method MetaTPTrans, based on meta-learning, utilizes a meta-learner to dynamically generate feature extractor parameters for each programming language, achieving joint learning of language-independent and language-specific information.
Wenye Li (Chinese University of Hong Kong), Zichen Ma (Chinese University of Hong Kong)
CodeOptimizationComputational EfficiencyTabular
π― What it does: This paper proposes a two-stage method for solving the metric approximation problem of noisy distance matrices, first quickly obtaining an approximate solution within a subset of isometric embeddings, and then refining it to a global optimum through alternating projections.
Metric Residual Network for Sample Efficient Goal-Conditioned Reinforcement Learning
Bo Liu (University of Texas at Austin), Peter Stone (University of Texas at Austin)
CodeReinforcement Learning
π― What it does: This paper proposes the Metric Residual Network (MRN), a neural network architecture based on the triangle inequality, aimed at improving sample efficiency in goal-conditioned reinforcement learning.
π― What it does: This paper proposes a new 3D object detection network called MGTANet, which can encode continuous LiDAR point cloud sequences and utilize temporal information to improve detection accuracy.
MGTCF: Multi-Generator Tropical Cyclone Forecasting with Heterogeneous Meteorological Data
Cheng Huang (Zhejiang University of Technology), YuQuan Wu (Chinese Academy of Sciences)
CodeGenerationData SynthesisRecurrent Neural NetworkGenerative Adversarial NetworkMultimodalityTime Series
π― What it does: A multi-generator multi-modal tropical cyclone prediction framework (MGTCF) is proposed, utilizing heterogeneous meteorological data and environmental information to predict trajectory and intensity.
CodeAnomaly DetectionRepresentation LearningContrastive LearningTime Series
π― What it does: A mask-based hierarchical clustering contrastive learning framework (MHCCL) is proposed, which enhances unsupervised time series representation learning by removing outliers with an upward mask and filtering positive and negative samples with a downward mask.
π― What it does: MicroAST is proposed, a lightweight and ultra-fast super-resolution arbitrary style transfer model that can complete style transfer on 4K images in about 0.5 seconds;
Mind the Gap: Polishing Pseudo Labels for Accurate Semi-supervised Object Detection
Lei Zhang (Northwestern Polytechnical University), Wei Wei (Northwestern Polytechnical University)
CodeObject DetectionImage
π― What it does: This paper proposes a dual pseudo-label refinement framework that utilizes two structurally different refinement networks (classification refinement network and box regression refinement network) to correct the pseudo-labels generated by the teacher detector, achieving end-to-end joint training within a semi-supervised object detection framework.
Minimax AUC Fairness: Efficient Algorithm with Provable Convergence
Zhenhuan Yang (Etsy), Yiming Ying (University at Albany)
CodeOptimizationTabular
π― What it does: A framework for minimizing and maximizing AUC fairness by considering both intra-group and inter-group AUC in AUC optimization is proposed, aiming to eliminate the model's differential ranking bias towards different sensitive attribute groups.
π― What it does: A music-to-dance generation framework based on style embeddings is proposed, which achieves precise control and diverse generation of dance styles by learning dance prototypes and forming controllable style embeddings through linear combinations.
π― What it does: This paper proposes two core technologies for long-tail video recognition (VLTR): a learnable feature aggregator and minority-oriented neighborhood expansion (MOVE).
π― What it does: Proposes the Hidden State Attention (HSA) module, which directly corrects the hidden states in the recursive video super-resolution model to reduce artifacts in real scenes.
π― What it does: A latent variable encoding-decoding model HiS-Dialog based on contrastive learning is designed and implemented to decouple and control the dialogue style in mixed dialogue systems.
π― What it does: This paper proposes the Mixture Manifold Network (MMN) to solve inverse problems by mixing multiple backward models with a single forward model, using the forward model to generate data for training, thereby improving the accuracy and speed of inverse modeling.
π― What it does: In model-based multi-agent reinforcement learning, the MAG framework is proposed, treating local models as multi-step decision-makers and combining the current policy as the environment to optimize multi-step prediction errors, thereby reducing error propagation and improving model quality;
π― What it does: The Mixture of Expert Clusters (MoEC) method is proposed, which addresses the issues of sparse data allocation and overfitting in large-scale MoE models by clustering experts and introducing variance constraints and cluster-level expert dropout.
Molformer: Motif-Based Transformer on 3D Heterogeneous Molecular Graphs
Fang Wu (Westlake University), Stan Z. Li (Westlake University)
CodeDrug DiscoveryGraph Neural NetworkTransformerReinforcement LearningGraphBiomedical Data
π― What it does: A Motif-based Transformer model called Molformer is proposed, which represents molecules using heterogeneous molecular graphs (including atom layers and Motif layers) and learns their 3D geometric information.
MSDC: Exploiting Multi-State Power Consumption in Non-intrusive Load Monitoring Based on a Dual-CNN Model
Jialing He (Chongqing University), Liehuang Zhu (Beijing Institute of Technology)
CodeConvolutional Neural NetworkTime Series
π― What it does: This paper proposes a multi-state dual CNN model (MSDC) for non-intrusive load monitoring tasks, which decomposes aggregated power signals into the power consumption of individual devices.
Multi-Action Dialog Policy Learning from Logged User Feedback
Shuo Zhang (Xi'an Jiaotong University), Junlan Feng (China Mobile Research)
CodeRecommendation SystemReinforcement Learning from Human FeedbackSupervised Fine-TuningReinforcement LearningText
π― What it does: A multi-action dialogue strategy learning framework based on BanditMatch is proposed, which enhances multi-action dialogue strategies by utilizing bandit feedback from historical dialogues.
Multi-Label Few-Shot ICD Coding as Autoregressive Generation with Prompt
Zhichao Yang (University of Massachusetts), Hong Yu (University of Massachusetts)
CodeClassificationGenerationTransformerSupervised Fine-TuningPrompt EngineeringTextBiomedical DataElectronic Health Records
π― What it does: For the ICD coding task, the authors propose a framework based on autoregressive generation and prompts: first, pre-training is conducted to generate assessment and plan texts using SOAP-structured clinical records, and then the model is guided by prompts to sequentially generate disease descriptions and map them back to ICD codes, achieving multi-label classification, especially with significant results in low-sample scenarios.
π― What it does: A multi-level combinatorial reasoning robot, MCR-Agent, is proposed, which decomposes instructions into observable sub-goals and learns navigation and interaction sub-policies separately.
CodeDrug DiscoveryGraph Neural NetworkContrastive LearningGraphBiomedical Data
π― What it does: This paper proposes a Multi-Relation Contrastive Learning Graph Neural Network (MRCGNN) for predicting drug-drug interaction events (DDI events).
Ammar Shaker (NEC Laboratories Europe), Carolin Lawrence (NEC Laboratories Europe)
CodeDomain AdaptationBiomedical Data
π― What it does: A method for multi-source survival domain adaptation is proposed, aimed at effectively adapting from multiple survival source domains to a new survival target domain, especially in cases of data scarcity and partial information censorship.
π― What it does: Designed and implemented a multi-stream representation learning module to learn the spatiotemporal and interaction features of pedestrian trajectories under the CVAE framework, achieving multimodal prediction.
MultiAct: Long-Term 3D Human Motion Generation from Multiple Action Labels
Taeryung Lee (Seoul National University), Kyoung Mu Lee (Seoul National University)
CodeGenerationPose EstimationTransformerVideo
π― What it does: Proposes the MultiAct framework, which achieves long-term 3D human motion generation through recursive generation of multi-action labels;
MultiSpider: Towards Benchmarking Multilingual Text-to-SQL Semantic Parsing
Longxu Dou (Harbin Institute of Technology), Jian-Guang Lou (Microsoft Research Asia)
CodeTransformerLarge Language ModelSupervised Fine-TuningTextBenchmark
π― What it does: A high-quality text-to-SQL semantic parsing dataset covering seven languages, MULTISPIDER, has been constructed, and a schema-based augmentation framework, SAVE, has been proposed to enhance the performance of cross-language text-to-SQL.
π― What it does: A framework for detecting diabetic retinopathy based on multi-view retinal images is proposed, which utilizes a cross-interaction self-attention module to fuse convolutional and Transformer features, and achieves inter-view information interaction through multi-view stitching.
NAS-LID: Efficient Neural Architecture Search with Local Intrinsic Dimension
Xin He (Hong Kong Baptist University), Xiaowen Chu (Hong Kong Baptist University)
CodeNeural Architecture SearchImage
π― What it does: A lightweight single network architecture search method called NAS-LID is proposed, which utilizes the local intrinsic dimension (LID) for subnet segmentation, thereby reducing interference between different sub-networks and enhancing performance ranking correlation.
Nearest-Neighbor Sampling Based Conditional Independence Testing
Shuai Li (East China Normal University), Wang Wen (New York University Shanghai)
CodeTabular
π― What it does: A conditional independence test method based on 1-nearest neighbor sampling (NNSCIT) is proposed, which can be implemented without knowing the conditional distribution.
π― What it does: This paper proposes an end-to-end unsupervised graph contrastive learning framework named NCLA, which can automatically learn graph augmentations and node representations.
Neighborhood-Regularized Self-Training for Learning with Few Labels
Ran Xu (Emory University), Carl Yang (Emory University)
CodeClassificationOptimizationGraph Neural NetworkTransformerLarge Language ModelTextGraph
π― What it does: The NeST method is proposed, which reduces pseudo-label noise and enhances training stability in semi-supervised self-training through neighborhood regularization for sample selection and multi-round prediction aggregation.
Nested Named Entity Recognition as Building Local Hypergraphs
Yukun Yan (Tsinghua University), Sen Song (Tsinghua University)
CodeRecognitionRecurrent Neural NetworkGraph Neural NetworkTextBiomedical Data
π― What it does: Proposes the Local Hypergraph Builder Network (LHBN), which achieves nested named entity recognition by first predicting entity boundaries and then constructing local hypergraphs for each boundary.
Networked Anti-coordination Games Meet Graphical Dynamical Systems: Equilibria and Convergence
Zirou Qiu (University of Virginia), Anil Vullikanti (University of Virginia)
CodeOptimizationGraph Neural NetworkGraph
π― What it does: This paper studies the existence and search of pure Nash equilibria in networked evolutionary anti-cooperative games, as well as the convergence time of synchronous and sequential updating schemes under self-correlated and self-independent decision-making modes.
π― What it does: Proposes the Networked Restless Bandits model to study the positive externality problem of resource allocation in community structures;
π― What it does: Finding network models (WsrNets) that exhibit good robustness under various attack intensities through integrated neural architecture search.
Seongmin Hong (Seoul National University), Se Young Chun (Seoul National University)
CodeFlow-based ModelBiomedical DataPhysics Related
π― What it does: A reversible and at least second-order continuously differentiable non-uniform B-spline flow is proposed for distribution modeling and efficient sampling of physical systems.
π― What it does: A neural dynamic focusing topic model (NDF-TM) is proposed, which can explicitly decouple topic proportions and topic activity in document sequences, thereby better capturing the time-varying topic structure.
π― What it does: By constructing a class-conditional probability distribution model, this study investigates how residual convolutional networks fit feature representations of different categories;
π― What it does: A sequence-based dynamic neighbor graph and multi-step dependency prediction model (SNPM) is proposed for next location (POI) recommendation.
NHITS: Neural Hierarchical Interpolation for Time Series Forecasting
Cristian Challu (Carnegie Mellon University), Artur Dubrawski (Carnegie Mellon University)
CodeTime Series
π― What it does: A new long-period time series forecasting model, NHITS, is proposed, which achieves specialized predictions for different frequency components through multi-rate input sampling and hierarchical interpolation, thereby maintaining low variance and reducing prediction fluctuations over long periods.
Jun Wu (University of Illinois), Elizabeth Ainsworth (University of Illinois)
CodeDomain AdaptationRecommendation SystemGraph Neural NetworkGraphAgriculture Related
π― What it does: The paper proposes a cross-network transfer learning framework that utilizes the Graph Subtree Discrepancy metric based on Weisfeiler-Leman subtrees to achieve knowledge transfer from the source graph to the target graph.
Normalizing Flow Ensembles for Rich Aleatoric and Epistemic Uncertainty Modeling
Lucas Berry (McGill University), David Meger (McGill University)
CodeOptimizationComputational EfficiencyFlow-based ModelTabularTime Series
π― What it does: This paper proposes a Normalizing Flow (NF) ensemble model constructed through a fixed dropout mask, balancing rich aleatoric uncertainty and model epistemic uncertainty, and evaluates its performance within an active learning framework.
π― What it does: A neighborhood voxel pruning method based on spatial point distribution is proposed for accelerating and optimizing energy consumption in sparse 3D convolutional networks.
Now Weβre Talking: Better Deliberation Groups through Submodular Optimization
Jake Barrett (University of Edinburgh), Ariel D. Procaccia (Harvard University)
CodeOptimizationTabular
π― What it does: This paper proposes a framework for topic discussion group allocation based on submodular optimization, aimed at enhancing interaction and diversity in citizen assemblies.
NQE: N-ary Query Embedding for Complex Query Answering over Hyper-Relational Knowledge Graphs
Haoran Luo (Beijing University of Posts and Telecommunications), Kaiyang Wan (Beijing Institute of Computer Technology and Application)
CodeTransformerGraph
π― What it does: Proposes the NQE model, which supports complex query answering for arbitrary n-ary first-order logic queries on hyper-relational knowledge graphs;
NuWLS: Improving Local Search for (Weighted) Partial MaxSAT by New Weighting Techniques
Yi Chu (Institute of Software, Chinese Academy of Sciences), Chuan Luo (School of Software, Beihang University)
CodeOptimizationBenchmark
π― What it does: In this paper, the authors propose two new weighting techniques (soft clause weight initialization and Dist-Weighting) and based on this, design a new Stochastic Local Search (SLS) solver NuWLS to solve the (Weighted) Partial MaxSAT (W)PMS problem.
π― What it does: This paper proposes an occupancy plane (OPlanes) representation method based on perspective frustum forward plane slicing, used to reconstruct 3D human bodies from a single RGB-D image.
π― What it does: This study investigates how to achieve system identification and prediction in input-output systems with irregular sampling using a continuous-time random state space model (ODE-RSSM).
π― What it does: A variant of PPO based on offline data (Off-Policy PPO) is proposed, which significantly improves sample efficiency by designing a new clipped approximate objective for the safe utilization of offline experiences.
On the Calibration and Uncertainty with PΓ³lya-Gamma Augmentation for Dialog Retrieval Models
Tong Ye (University of Science and Technology of China), Jing Xiao (Ping An Technology)
CodeRetrievalTransformerLarge Language ModelText
π― What it does: This paper proposes an efficient dialogue retrieval model uncertainty and calibration framework PG-DRR, which uses a PΓ³lya-Gamma enhanced Gaussian process layer to replace traditional dense layers, achieving more reliable confidence and calibration.
π― What it does: A Domain-wise Adversarial Training (DAT) method is proposed, which suppresses domain-related noise features by introducing domain-specific perturbations in each training domain, thereby enhancing the model's generalization ability on out-of-distribution (OOD) data.
π― What it does: This paper conducts a systematic analysis of Parameter-Efficient Fine-Tuning (PEFT) methods, proposes a unified sparse fine-tuning model framework, and provides theoretical upper bounds on the stability and generalization error of sparsity. It also identifies the issue of projection discontinuity in projection-based PEFT methods and proposes the Second-Order Approximation Method (SAM) to select adjustable parameters. Finally, large-scale experiments were conducted on the GLUE and SuperGLUE tasks using RoBERTa-base.
On the Sample Complexity of Representation Learning in Multi-Task Bandits with Global and Local Structure
Alessio Russo (KTH Royal Institute of Technology), Alexandre Proutiere (KTH Royal Institute of Technology)
CodeOptimizationRepresentation Learning
π― What it does: This paper studies the sample complexity of optimal arm identification in multi-task weighted multi-armed bandits (MAB) when all tasks share the same optimal representation while the predictor is task-specific.
π― What it does: A Progressive Volume Distillation (PVD) method is proposed, enabling arbitrary one-to-one conversions between different NeRF architectures (such as MLP, sparse tensors, low-rank tensors, hash tables, and their combinations), allowing for quick adaptation to different task requirements in later stages.
Xincheng Yao (Shanghai Jiao Tong University), Zhenyu Liu (Shanghai Jiao Tong University)
CodeAnomaly DetectionTransformerAuto EncoderImage
π― What it does: A cross-category and multi-category anomaly detection method PMAD based on patch-level reconstruction and prototype-guided proposal masking is proposed.
Online Platforms and the Fair Exposure Problem under Homophily
Jakob Schoeffer (Karlsruhe Institute of Technology), Marc Juarez (University of Edinburgh)
CodeRecommendation SystemOptimizationText
π― What it does: This study proposes the 'fair exposure' problem, constructs a temporal model considering group homogeneity in dissemination, and analyzes how platforms can maximize user clicks and likes under limited intervention while satisfying fair exposure constraints. It also provides theoretical solutions for fair-unrelated and fair-constrained optimization and evaluates the 'fairness cost' caused by fair constraints.
π― What it does: A framework for online semi-supervised learning with mixed-type streaming features is proposed, and an online algorithm OSLMF based on Gaussian Copula and density peak clustering is designed.
π― What it does: This paper proposes a global black-box optimizer OPT-GAN based on GAN, which can find the global optimal solution in multi-dimensional multi-peak functions by learning and continuously updating the target distribution.
Optimistic Whittle Index Policy: Online Learning for Restless Bandits
Kai Wang (Harvard University), Milind Tambe (Google Research)
CodeOptimizationReinforcement LearningBiomedical Data
π― What it does: An online learning algorithm UCWhittle based on the Whittle index is proposed, which can simultaneously learn unknown transition probabilities and execute approximately optimal scheduling strategies in RMAB (Random Multi-Armed Bandit).
π― What it does: A dynamic depth graph convolutional network (D-DGCN) is proposed, which dynamically integrates information from multiple unordered posts on social media to generate user personality profiles and perform multi-label personality prediction.
Out-of-Distribution Generalization by Neural-Symbolic Joint Training
Anji Liu (Peking University), Yitao Liang (Peking University)
CodeDomain AdaptationExplainability and InterpretabilityConvolutional Neural NetworkRecurrent Neural NetworkImageSequential
π― What it does: A neural-symbolic joint training framework NTOC is proposed, which can learn generalizable neural features and symbolic rules simultaneously without prior symbolic knowledge, addressing the OOD generalization problem.
Overcoming Concept Shift in Domain-Aware Settings through Consolidated Internal Distributions
Mohammad Rostami (Information Sciences Institute, University of Southern California), Aram Galstyan (Information Sciences Institute, University of Southern California)
π― What it does: A sequential model adaptation method based on internal distribution, SMAUI, is proposed to address the problem of concept drift in the absence of available source data.
Parameter-Efficient Model Adaptation for Vision Transformers
Xuehai He (University of California Santa Cruz), Xin Eric Wang (Microsoft Research)
CodeClassificationTransformerImageBenchmark
π― What it does: This paper studies how to efficiently adapt large pre-trained Vision Transformer (ViT) models for image classification tasks with very few trainable parameters.
π― What it does: A point cloud upsampling network based on parametric surface constraints is proposed, capable of generating high-density, smooth point clouds from sparse point clouds, and can also be transferred to point cloud completion tasks.
Participatory Budgeting Designs for the Real World
Roy Fairstein (Ben-Gurion University of the Negev), Kobi Gal (Ben-Gurion University of the Negev)
CodeTabular
π― What it does: Experimental evaluation of voting formats and aggregation rules in participatory budgeting (PB), analyzing the impact of different voting methods on user experience, outcome stability, and social welfare.
π― What it does: This paper proposes an automatic mining method based on structural patterns, called PaTeCon, for automatically generating temporal constraints and detecting temporal conflicts from knowledge graphs.
PDFormer: Propagation Delay-Aware Dynamic Long-Range Transformer for Traffic Flow Prediction
Jiawei Jiang (Beihang University), Jingyuan Wang (Renmin University of China)
CodeTransformerTime Series
π― What it does: The PDFormer model is proposed, which utilizes spatial self-attention to capture dynamic, long-range spatial dependencies and explicitly models the time delay of traffic information propagation to achieve accurate traffic flow prediction.
π― What it does: A three-layer sparsification framework called Tri-Level E-ViT is proposed, which removes redundant data at the sample, token, and attention connection levels to significantly accelerate the training and inference of Vision Transformers.
π― What it does: This paper proposes a Temporal Relation Mining (TRM) framework based on phrase-level temporal relations, modeling the temporal localization of video sentences using the temporal relations between sentences and phrases, and enhancing phrase-level prediction performance through consistency and exclusivity constraints.
PiCor: Multi-Task Deep Reinforcement Learning with Policy Correction
Fengshuo Bai (University of Chinese Academy of Sciences), Bo Xu (Chinese Academy of Sciences)
CodeRobotic IntelligenceReinforcement Learning
π― What it does: This paper proposes the PiCor framework, which enhances sample efficiency and task generalization ability in multi-task deep reinforcement learning by separating policy optimization and policy correction into two stages.
π― What it does: A NAS performance predictor named PINAT is proposed, which achieves efficient representation and prediction of network structures by incorporating partial permutation invariant embedding layers (PITE) and self-attention layers (PISA) into the Transformer architecture, using the Laplacian matrix as positional encoding.
π― What it does: A new PDE solver called PIXEL is proposed, which combines a trainable grid representation with a small MLP, utilizing automatic differentiation and traditional optimization algorithms to solve forward and inverse PDE problems.
π― What it does: A weakly semi-supervised object detection framework called Point-Teaching is proposed, which effectively utilizes point annotations through point matching, MIL, and copy-paste augmentation.
CodeFederated LearningAdversarial AttackImageTabularFinance Related
π― What it does: This paper proposes a distributed backdoor attack method called Cerberus Poisoning, which utilizes colluding malicious participants to achieve covert attacks against various defense mechanisms in federated learning through trigger fine-tuning and model bias regularization.