These 1014 AAAI 2024 papers come with a code repository. Each shows an AI one-line summary below β get the verified repo link + the full 6-part summary (innovation, method, data, results, limitations) and search every AAAI 2024 paper, free trial on arXivSub.
3D Visibility-Aware Generalizable Neural Radiance Fields for Interacting Hands
Xuan Huang (Shenzhen Campus of Sun Yat-sen University), Xiaodan Liang (Tencent)
π― What it does: A visible perception NeRF model that can generalize from a single image is proposed for generating high-quality images of interactive two-hand scenes.
π― What it does: An end-to-end 3D Referring Expression Segmentation framework, 3D-STMN, is proposed, which directly performs semantic matching at the superpoint level, avoiding error propagation and redundant computation in a two-stage process.
A Brain-Inspired Way of Reducing the Network Complexity via Concept-Regularized Coding for Emotion Recognition
Han Lu (Fudan University), Qiang Luo (Fudan University)
CodeRecognitionExplainability and InterpretabilityComputational EfficiencyConvolutional Neural NetworkContrastive LearningImage
π― What it does: This paper proposes a brain-inspired emotional recognition framework that employs a dual pathway (conceptual pathway and perceptual pathway) to reduce network complexity and enhance interpretability through emotional concept regularization.
A Compiler for Weak Decomposable Negation Normal Form
Petr Illner (Charles University), Petr KuΔera (Charles University)
Code
π― What it does: This paper introduces weakly decomposable negation normal form (wDNNF) circuits into the knowledge compilation map and proves that they share the same properties as DNNF circuits in terms of queries and transformations, while being more concise than DNNF; it then proposes and implements a compiler named Bella that can compile CNF formulas into wDNNF circuits.
A Convolutional Neural Network Interpretable Framework for Human Ventral Visual Pathway Representation
Mufan Xue (Beijing Institute of Technology), Guoyuan Yang (Beijing Institute of Technology)
CodeExplainability and InterpretabilityRepresentation LearningConvolutional Neural NetworkImageMagnetic Resonance Imaging
π― What it does: This study designs and validates an interpretable convolutional neural network framework (CNNIF) that encodes voxel responses of the human dorsal visual pathway by combining feature-weighted receptive fields (FWRF) with spatial pooling fields, and reveals the hierarchical representation of the concept of 'human' across layers such as V1βV2βV3βhV4βFFA through network dissection.
A Cross-View Hierarchical Graph Learning Hypernetwork for Skill Demand-Supply Joint Prediction
Wenshuo Chao (Hong Kong University of Science and Technology), Hao Liu (Hong Kong University of Science and Technology)
CodeGraph Neural NetworkGraphTime SeriesFinance Related
π― What it does: This paper proposes a Cross-View Hierarchical Graph Learning Hypernetwork (CHGH) for jointly predicting changes in the demand and supply of skills in the labor market. The model encodes the graph structures of demand and supply, clusters hierarchical relationships, and constructs a conditional hyper-decoder using historical supply-demand gaps to achieve accurate predictions of future skill demand and supply.
A Dual Stealthy Backdoor: From Both Spatial and Frequency Perspectives
Yudong Gao (China University of Petroleum), Weifeng Liu (China University of Petroleum)
CodeAdversarial AttackImage
π― What it does: A dual hidden backdoor attack method (DUBA) that is concealed in both spatial and frequency domains is designed, utilizing wavelet transform, mixed smoothing with FFT and DCT, and a weak trigger training + strong attack strategy to achieve a high success rate and low detectability.
π― What it does: Proposed and implemented a Dual-Way Enhanced (DWE) framework, treating multimodal information (text and images) as a neural text matching task to improve the entity linking problem.
A Fast Exact Solver with Theoretical Analysis for the Maximum Edge-Weighted Clique Problem
Lu Liu (University of Electronic Science and Technology of China), Yi Zhou (University of Electronic Science and Technology of China)
CodeOptimizationGraphBenchmark
π― What it does: This paper systematically studies the Maximum Edge Weight Clique Problem (MEWCP). It first proves that the problem remains NP-hard even when the minimum degree of the graph is $n-2$. Then, it proposes a branch-and-bound algorithm called MEWCat based on a new, tighter upper bound, and provides a theoretical time complexity upper bound of $O^*(1.4423^n)$. Finally, experiments are conducted on various benchmarks.
A Goal Interaction Graph Planning Framework for Conversational Recommendation
Xiaotong Zhang (Dalian University of Technology), Xianchao Zhang (Dalian University of Technology)
CodeRecommendation SystemGraph Neural NetworkTransformerLarge Language ModelTextGraph
π― What it does: A target interaction graph planning framework is proposed for multi-target conversation recommendation, which can dynamically plan the target sequence in the dialogue and guide users to achieve the final recommendation.
Liming Pan (University of Science and Technology of China), Ivan Dokmanic (Universitat Basel)
CodeGraph Neural NetworkGraphTime Series
π― What it does: A Graph Dynamics Prior (GDP) framework is proposed, utilizing multi-order polynomial graph filtering and shallow single-step GNN to reconstruct interaction graphs from observed dynamics by sharing graph structures in parallel.
π― What it does: This paper proposes a multi-modal document-level relation extraction (MDocRE) task, utilizing both text and corresponding video modalities to enhance the handling of long-distance dependencies and multi-mention selection issues.
A Label Disambiguation-Based Multimodal Massive Multiple Instance Learning Approach for Immune Repertoire Classification
Fan Xu (ShanghaiTech University), Jianhua Yao (Shanghai Jiao Tong University)
CodeClassificationTransformerSupervised Fine-TuningMultimodalityBiomedical Data
π― What it does: A multi-modal large-scale multi-instance learning model based on label disambiguation (LaDMΒ³IL) is proposed for immune receptor library classification and related receptor identification.
CodeGenerationDrug DiscoveryTransformerDiffusion modelContrastive LearningMultimodalityBiomedical Data
π― What it does: A multi-modal contrastive diffusion model (MMCD) is proposed for simultaneously generating sequences and structures of therapeutic peptides.
π― What it does: A new dual-branch network is proposed, capable of detecting image tampering under difficult conditions such as small-scale tampering and double compression with the same quality factor.
π― What it does: This paper proposes the Embedding-Ignoring Conflict (EIC) in information contrastive learning and designs the PiGCL method to dynamically capture and ignore certain negative samples to enhance graph contrastive learning performance.
A Novel Energy Based Model Mechanism for Multi-Modal Aspect-Based Sentiment Analysis
Tianshuo Peng (Wuhan University), Hai Zhao (Shanghai Jiao Tong University)
CodeClassificationRecognitionData-Centric LearningTransformerLarge Language ModelPrompt EngineeringImageTextMultimodality
π― What it does: A unified multimodal fine-grained sentiment analysis framework DQPSA is proposed to address the interrelationship issues of visual information attention differences, modality gaps, and span boundaries in multimodal sentiment analysis.
A Novel Skip Orthogonal List for Dynamic Optimal Transport Problem
Xiaoyang Xu (University of Science and Technology of China), Hu Ding (University of Science and Technology of China)
CodeOptimizationTabular
π― What it does: A dynamic optimal transport algorithm is proposed, which can quickly update the optimal transport plan after changes in the weights or positions of data points;
A Plug-and-Play Quaternion Message-Passing Module for Molecular Conformation Representation
Angxiao Yue (Renmin University of China), Hongteng Xu (Renmin University of China)
CodeDrug DiscoveryGraph Neural NetworkGraph
π― What it does: A pluggable quaternion information transmission module (QMP) is proposed to enhance the representation of molecular conformations in 3D molecular graph neural networks.
A Positive-Unlabeled Metric Learning Framework for Document-Level Relation Extraction with Incomplete Labeling
Ye Wang (East China Normal University), Wenxin Hu (East China Normal University)
CodeTransformerTextBiomedical Data
π― What it does: To address the issue of incomplete annotations in document-level relation extraction, a positive-negative sample unlabeled metric learning framework (PΒ³M) is proposed. This framework enhances the model's generalization and robustness by applying dropout augmentation on positive samples and using unlabeled relations as pseudo-negative samples for mixup.
A Provably Accurate Randomized Sampling Algorithm for Logistic Regression
Agniva Chowdhury (Oak Ridge National Laboratory), Pradeep Ramuhalli (Oak Ridge National Laboratory)
CodeClassificationOptimizationComputational EfficiencyTabularFinance Related
π― What it does: A randomized sampling-based logistic regression algorithm is proposed, which obtains an approximate maximum likelihood estimate by randomly sampling and rescaling the observed values, thus quickly solving large-scale logistic regression problems.
π― What it does: A robust 3D multimodal medical image fusion framework is proposed, utilizing visual-semantic consistency to achieve complementary enhancement of visual fusion and lesion segmentation, significantly improving the accuracy and robustness against degradation of both tasks.
A Sequentially Fair Mechanism for Multiple Sensitive Attributes
Francois Hu (University of Montreal), Arthur Charpentier (University of Quebec at Montreal)
CodeTabular
π― What it does: This paper proposes a sequential fairness mechanism that utilizes multiple Wasserstein barycenters for post-processing predictions with multiple sensitive attributes to achieve group fairness.
π― What it does: The TURRET method is proposed to address the issue of mismatched robot state-action spaces in cross-domain multi-source transfer learning, utilizing graph neural networks to achieve unified state embedding and adaptive weighted multi-source strategies to accelerate target task learning.
A Twist for Graph Classification: Optimizing Causal Information Flow in Graph Neural Networks
Zhe Zhao (University of Science and Technology of China), Yang Wang (University of Science and Technology of China)
CodeClassificationOptimizationExplainability and InterpretabilityGraph Neural NetworkGraph
π― What it does: This paper proposes an ICL framework that combines information theory and causal learning, specifically designed for graph classification tasks. It can decompose graph features into causal and non-causal parts during the training process, maximizing causal information and minimizing non-causal information through multi-objective optimization, thereby enhancing the model's robustness and interpretability against out-of-distribution data.
π― What it does: A unified Knowledge Transfer Network (KTN) is proposed, which utilizes labeled known categories and unlabeled mixed data to achieve knowledge transfer through entropy soft distinction and prototype weighting, enhancing the performance of general category discovery.
A User-Friendly Framework for Generating Model-Preferred Prompts in Text-to-Image Synthesis
Nailei Hei (Fudan University), Wenqiang Zhang (Fudan University)
CodeGenerationData SynthesisTransformerLarge Language ModelPrompt EngineeringDiffusion modelImageText
π― What it does: A framework is proposed for automatically converting user-input rough prompts into fine-grained prompts preferred by the model, along with the construction of a corresponding training dataset.
Abstract and Explore: A Novel Behavioral Metric with Cyclic Dynamics in Reinforcement Learning
Anjie Zhu (University of Electronic Science and Technology of China), Jie Shao (University of Electronic Science and Technology of China)
CodeReinforcement LearningSequential
π― What it does: The BCD (Behavioral Metric with Cyclic Dynamics) method is proposed to address the issues of state representation collapse and insufficient dynamics correlation in environments with program-generated, sparse rewards and high noise, thereby improving exploration efficiency.
Accelerating the Global Aggregation of Local Explanations
Alon Mor (Technion Israel Institute of Technology), Benny Kimelfeld (Technion Israel Institute of Technology)
CodeExplainability and InterpretabilityComputational EfficiencyText
π― What it does: A global aggregation acceleration method for Anchor local explanations is proposed, along with a new probability aggregation function and various runtime optimizations.
π― What it does: An adaptive multi-round retrieval paradigm named Ada-Retrieval is proposed for sequence recommendation systems, which better captures potential candidates by iteratively optimizing user representations.
π― What it does: This paper proposes AdaCCD, a cross-language adaptive code clone detection method that utilizes a pre-trained multilingual programming language model to discover semantically similar/dissimilar pairs in unlabeled data of the target language through adaptive contrastive learning, and improves performance through iterative training.
π― What it does: This paper proposes a text-conditioned image editing algorithm called AdapEdit, based on time-space adaptive guidance, to achieve soft editing tasks (such as fine-grained changes in posture, actions, adjectives, etc.) without the need for additional training or optimization.
AdapterGNN: Parameter-Efficient Fine-Tuning Improves Generalization in GNNs
Shengrui Li (Tsinghua University), Jing Bai (Microsoft Research Asia)
CodeDrug DiscoveryGraph Neural NetworkSupervised Fine-TuningGraphBiomedical Data
π― What it does: A parameter-efficient fine-tuning framework designed specifically for graph neural networks, called AdapterGNN, is proposed and evaluated.
Adaptive FSS: A Novel Few-Shot Segmentation Framework via Prototype Enhancement
Jing Wang (University of Science and Technology Beijing), Tianxiang Zhang (University of Science and Technology Beijing)
CodeSegmentationTransformerImage
π― What it does: Proposes an adapter-based few-shot semantic segmentation framework called Adaptive FSS, which utilizes a Prototype Adaptive Module to enhance prototypes for new categories, achieving rapid adaptation.
π― What it does: This paper proposes an Adaptive Interactive Graph Network (AdaIGN) that models multimodal dialogue emotions through a learnable node and edge selection strategy.
Adaptive Hardness Negative Sampling for Collaborative Filtering
Riwei Lai (Harbin Engineering University), Li Chen (Hong Kong Baptist University)
CodeRecommendation SystemTabular
π― What it does: A new negative sampling paradigm called Adaptive Hardness Negative Sampling (AHNS) is proposed, aimed at alleviating false positive and false negative issues by adaptively selecting negative samples of varying hardness during the training process.
Adaptive Meta-Learning Probabilistic Inference Framework for Long Sequence Prediction
Jianping Zhu (Dalian University of Technology), Fei Wu (Zhejiang University)
CodeOptimizationMeta LearningTransformerAuto EncoderTime SeriesSequentialFinance Related
π― What it does: This paper proposes an adaptive meta-learning probabilistic inference framework based on sequence decomposition, called AMPIF, for long sequence prediction.
Adaptive Prompt Routing for Arbitrary Text Style Transfer with Pre-trained Language Models
Qingyi Liu (Sun Yat-sen University), Keze Wang (Datastory)
CodeGenerationTransformerLarge Language ModelPrompt EngineeringText
π― What it does: An Adaptive Prompt Routing (APR) framework is proposed, which automatically selects the optimal prompt from a set of readable prompts for each input sentence for arbitrary text style conversion, and generates stylized text through LLM.
Adaptive Shortcut Debiasing for Online Continual Learning
Doyoung Kim (KAIST), Jae-Gil Lee (KAIST)
CodeClassificationImage
π― What it does: The DropTop framework is proposed to enhance the model's transferability and stability in online continual learning by suppressing shortcut feature bias.
π― What it does: An Adaptive Uncertainty-based Learning (AUL) framework is proposed from the perspective of uncertainty to address the issues of matching ambiguity and unidirectional cross-modal alignment in text retrieval-based pedestrian retrieval.
π― What it does: This paper studies a method for generating imperceptible adversarial facial identity attacks using a latent diffusion model in latent space.
Advancing Spatial Reasoning in Large Language Models: An In-Depth Evaluation and Enhancement Using the StepGame Benchmark
Fangjun Li (University of Leeds), Anthony G. Cohn (University of Leeds)
CodeTransformerLarge Language ModelPrompt EngineeringTextBenchmarkChain-of-Thought
π― What it does: This paper addresses the template errors in the StepGame benchmark and conducts an in-depth evaluation and enhancement of the performance of large language models in spatial reasoning tasks.
π― What it does: A manifold hypothesis-based adversarial purification framework is proposed, achieving a robust model without adversarial training during testing through variational inference.
π― What it does: A framework called SIASM based on the principle of structural information is proposed to actively model the attack behavior of social robots, maximizing network influence and enhancing concealment.
Adversarially Balanced Representation for Continuous Treatment Effect Estimation
Amirreza Kazemi (Simon Fraser University), Martin Ester (Simon Fraser University)
CodeRepresentation LearningAdversarial AttackGenerative Adversarial NetworkTextBiomedical Data
π― What it does: A framework for adversarial balanced representation learning (ACFR) is proposed for estimating continuous treatment effects. By minimizing the KL divergence of the representation layer and using a cross-attention network to predict outcomes, it can achieve both covariate balance and make full use of treatment value information.
Xinyao Li (University of Electronic Science and Technology of China), Ke Lu (University of Electronic Science and Technology of China)
CodeDomain AdaptationImage
π― What it does: A lightweight Bi-ATEN module is proposed to achieve multi-source unsupervised domain adaptation, enabling adaptation to the target domain using a small number of trainable parameters without extensive fine-tuning of the source model.
π― What it does: Developed a self-supervised alternative model training framework (AGS) that trains the alternative model using only unlabeled data, thereby achieving transferable black-box adversarial attacks.
π― What it does: A new budget online kernel learning model, Ahpatron, is proposed, which significantly improves the error bounds of previous work and addresses an open problem related to the upper bound of hypothesis space constraints.
π― What it does: Train NeRF for unsupervised learning under low light and overexposure conditions, and generate new views with normal lighting by removing/adding the Concealing Field during inference.
ALISON: Fast and Effective Stylometric Authorship Obfuscation
Eric Xing (Washington University in St. Louis), Dongwon Lee (Pennsylvania State University)
CodeAdversarial AttackTransformerLarge Language ModelText
π― What it does: A method for author identity obfuscation based on grammatical features, ALISON, is proposed, which utilizes a lightweight internal author discriminator and a masked language model to perform semantic-preserving style replacement of multiple word segments at once.
π― What it does: The K+K scheme is proposed, using Target-Aware Universum (TAU) to generate K classes of pseudo-unknown samples, and treating known and unknown categories equally through Dual Contrastive Loss, improving open-set recognition.
π― What it does: A multilingual text-to-image diffusion model called AltDiffusion has been proposed and trained, supporting 18 languages, directly processing multilingual prompts to generate high-quality images.
π― What it does: The VENUS method is proposed to address the knowledge fusion problem under heterogeneous architectures of multi-task teacher models, training a unified multi-task student with unlabeled data.
π― What it does: This paper proposes a new framework for panoramic occlusion relationship reasoning and generative occlusion completion (HORI+GMC) for efficient inference of invisible shapes and determination of occlusion relationships.
π― What it does: A novel single-stage network for underwater object detection, AMSP-UOD, is proposed to specifically address noise and attenuation issues in underwater images.
Arijit Shaw (Chennai Mathematical Institute), Kuldeep S. Meel (University of Toronto)
CodeTabularBenchmark
π― What it does: A novel approximate Skolem function counting algorithm, SkolemFC, is studied, which can estimate the number of Skolem functions without enumerating them.
An Attentive Inductive Bias for Sequential Recommendation beyond the Self-Attention
Yehjin Shin (Yonsei University), Noseong Park (Yonsei University)
CodeRecommendation SystemTransformerSequential
π― What it does: This paper proposes a Fourier Transform-based attention mechanism called BSARec, aimed at addressing the low-frequency filtering and over-smoothing issues caused by self-attention in Transformers for sequential recommendation.
π― What it does: A self-regressive text-to-graph (ATG) framework is proposed, which jointly extracts entities and relationships from unstructured text and achieves this through generating linearized graph structures.
An Eager Satisfiability Modulo Theories Solver for Algebraic Datatypes
Amar Shah (University of California), Sanjit A. Seshia
CodeBenchmark
π― What it does: This paper proposes a method for transforming quantifier-free ADT queries into an eager solving approach that only involves UF, and implements a prototype solver called Algaroba.
π― What it does: An efficient knowledge transfer strategy from static images to event domains is proposed to help spiking neural networks (SNNs) improve generalization ability on small sample event data.
π― What it does: The SubInfer framework is proposed, which achieves efficient learning and inference of large-scale heterogeneous graphs through meta-path subgraph partitioning, global vertex completion, and distributed subgraph training/inference.
An Embedding-Unleashing Video Polyp Segmentation Framework via Region Linking and Scale Alignment
Zhixue Fang (Shenzhen University), Jing Qin (Hong Kong Polytechnic University)
CodeSegmentationTransformerVideo
π― What it does: A framework for embedded release is proposed, modeling video segmentation as appearance-level semantic embedding. It utilizes a Proposal Generation Network (PGN) to generate mask proposals and generates background and dynamic semantics through Cross-Scale Region Linking (CRL) and Center-Aware Scale Alignment (CSA) modules in the Appearance Embedding Network (AEN), followed by segmentation through non-parametric semantic interaction.
An Empirical Study of CLIP for Text-Based Person Search
Min Cao (Soochow University), Min Zhang (Harbin Institute of Technology)
CodeRetrievalCompressionTransformerVision Language ModelContrastive LearningImageText
π― What it does: This paper conducts systematic experiments on CLIP to explore its potential in the text-driven face search (TBPS) task, proposing a baseline TBPS-CLIP that combines data augmentation, loss functions, and training techniques without using complex modules.
An Interpretable Approach to the Solutions of High-Dimensional Partial Differential Equations
Lulu Cao (Xiamen University), Min Jiang (Xiamen University)
CodeOptimizationExplainability and InterpretabilityComputational EfficiencyTabularPhysics Related
π― What it does: A method called HD-TLGP is proposed, which is based on genetic programming symbolic regression, structural transfer, automatic differentiation, and pruning operators, to solve the analytical solutions of high-dimensional partial differential equations.
Analyzing Generalization in Policy Networks: A Case Study with the Double-Integrator System
Ruining Zhang (University of Electronic Science and Technology of China), Jian Cheng (Air Force Engineering University)
CodeReinforcement LearningSequential
π― What it does: An analysis and experimental validation of the generalization failure of deep reinforcement learning policy networks for double integrator systems when expanding the state space.
Angle Robustness Unmanned Aerial Vehicle Navigation in GNSS-Denied Scenarios
Yuxin Wang (Zhejiang University), Mingli Song (Zhejiang University)
CodeAutonomous DrivingTransformerSimultaneous Localization and MappingImage
π― What it does: A robust angle navigation model for drone point-to-point navigation is proposed, which directly predicts the direction angle to address flight deviations in GNSS-denied environments.
π― What it does: AnomalyDiffusion is proposed, a few-shot anomaly image generation method based on diffusion models, capable of synthesizing realistic and diverse anomaly-image-mask pairs on normal images according to a given anomaly mask.
π― What it does: A two-stage text-to-image synthesis pipeline named Any-Size-Diffusion (ASD) is proposed. Stage-I (ARAD) generates well-synthesized low-resolution images within a limited aspect ratio range through multi-bit ratio training; Stage-II (FSTD) quickly enlarges the images generated in Stage-I to high-definition images of any size using implicit overlapping tile sampling, avoiding seams and significantly reducing memory usage.
π― What it does: The AnyStereo module is proposed, which can accurately upsample disparity maps at arbitrary scales in the iterative stereo matching process, addressing the shortcomings of traditional convolutional upsampling in recovering high-frequency details.
JunHoo Lee, Nojun Kwak (Seoul National University)
CodeMeta LearningImage
π― What it does: A label-equivalence-based arbitrary category meta-learning framework is proposed, and a semantic classifier and Mixup technique are introduced to enhance generalization ability.
π― What it does: A voxel grid-based framework for arbitrary scale point cloud upsampling, PU-VoxelNet, is proposed, which maps sparse point clouds to a predefined grid space and generates high-quality point clouds through density distribution.
Are You Concerned about Limited Function Evaluations: Data-Augmented Pareto Set Learning for Expensive Multi-Objective Optimization
Yongfan Lu (East China Normal University), Aimin Zhou (East China Normal University)
CodeOptimizationGenerative Adversarial Network
π― What it does: This paper proposes a Data Augmentation Pareto Set Learning (DA-PSL) method to address expensive multi-objective optimization (EMOP) problems, achieving a more accurate approximation of the Pareto front under a limited function evaluation budget.
π― What it does: Proposes the ArtBank framework, which achieves artistic style transfer by combining pre-trained large models with an implicit style prompt bank.
Wenxuan Tu (National University of Defense Technology), Xinwang Liu (National University of Defense Technology)
CodeGraph Neural NetworkGraph
π― What it does: A unified Attribute-Missing Graph Clustering Network (AMGC) is proposed, which enhances the clustering performance of graphs with missing attributes through alternating optimization of clustering and attribute inference.
Audio Generation with Multiple Conditional Diffusion Model
Zhifang Guo (Institute of Computing Technology, Chinese Academy of Sciences), Xiangdong Wang (Institute of Computing Technology, Chinese Academy of Sciences)
CodeGenerationData SynthesisDiffusion modelAudio
π― What it does: A multi-condition controlled text-to-audio generation model is proposed, which achieves fine-grained controllable generation of audio content and style by incorporating control conditions such as timestamps, pitch curves, and energy curves.
Augmented Commonsense Knowledge for Remote Object Grounding
Bahram Mohammadi (Australian Institute for Machine Learning), Javen Qinfeng Shi (Australian Institute for Machine Learning)
CodeRecognitionObject DetectionTransformerVision Language ModelMultimodality
π― What it does: This paper proposes a model called ACK based on augmented common sense knowledge to enhance visual-language navigation in the REVERIE task.
π― What it does: This paper proposes Auto-Prox, a framework for automatically discovering zero-cost proxies for training-free Vision Transformer (ViT) architecture search.
π― What it does: AOGNet is proposed, a self-regressive, global-local conditional 360-degree image generation network that can gradually complete the filling of panoramic images by simultaneously utilizing NFoV images and open vocabulary text prompts.
π― What it does: A Transformer-based framework called AVSegFormer is proposed for audio-visual segmentation tasks (single sound source, multiple sound sources, and semantic segmentation), achieving pixel-level localization and segmentation of sound source targets.
Gagan Biradar (University of Massachusetts), Yair Zick (University of Massachusetts)
CodeExplainability and InterpretabilityAdversarial AttackTabularFinance Related
π― What it does: Design and evaluate three axiomatically-based feature importance metrics for aggregating attributable explanations (AXp) to generate more robust and interpretable model explanations.
π― What it does: A multi-stage deep learning pipeline is proposed, which first performs spinal segmentation on low-quality X-ray images and utilizes CycleGAN for mask refinement. Then, based on the segmentation mask, it predicts the B-spline curve of the spinal centerline, and finally combines curve slope analysis with a regression model to obtain Cobb angle estimation.
Backward Responsibility in Transition Systems Using General Power Indices
Christel Baier (TU Dresden), Jakob Piribauer (TU Dresden)
Code
π― What it does: A backward responsibility measure based on general weight indices in cooperative games (such as Shapley value and Banzhaf value) is proposed, distinguishing between optimistic and pessimistic definitions of responsibility, and implementing both exact calculation and random approximation algorithms.
π― What it does: A sparse and targeted backdoor injection attack called BadRL is designed to achieve target actions and weaken cumulative rewards in reinforcement learning through triggers.
π― What it does: Developed the BAIT framework to unify the experimental environment, data, and models of AI-ITP, systematically comparing formula embedding architectures.
π― What it does: A statistical framework is proposed to decompose the accuracy of human-machine ensemble into individual accuracy and diversity, and to study the relationship between ensemble scale and performance.
BAT: Behavior-Aware Human-Like Trajectory Prediction for Autonomous Driving
Haicheng Liao (University of Macau), Chengzhong Xu (Tsinghua University)
CodeAutonomous DrivingRecurrent Neural NetworkTransformerMultimodalityTime Series
π― What it does: A behavior-aware multimodal trajectory prediction model (BAT) is proposed, which combines four modules: behavior, interaction, priority, and location, enabling high-accuracy predictions of surrounding vehicle trajectories.
π― What it does: This study investigates the use of complex logical evidence for incremental Bayesian inference on knowledge graphs and proposes the BIKG framework.
BBScore: A Brownian Bridge Based Metric for Assessing Text Coherence
Zhecheng Sheng (University of Minnesota), Dongyeop Kang (University of Minnesota)
CodeTransformerLarge Language ModelContrastive LearningText
π― What it does: This paper proposes a no-reference text coherence metric called BBScore based on Brownian Bridge theory, and trains a contrastive learning sentence encoder (GPT-2 + MLP) to obtain the trajectory of the text in the latent space of the Brownian Bridge, thereby calculating the coherence score.
π― What it does: A BDIQA video question-answering dataset has been constructed, focusing on cognitive reasoning of beliefs, desires, and intentions in videos.
π― What it does: A lightweight skeleton action recognition model FRF-GCN is proposed, which significantly improves model performance and efficiency by combining bidirectional fusion, targeted adjacency matrices, and a parallel three-dimensional attention mechanism.
BeliefFlow: A Framework for Logic-Based Belief Diffusion via Iterated Belief Change
Nicolas Schwind (National Institute of Advanced Industrial Science and Technology), Pierre Marquis (University of Artois)
CodeGraph
π― What it does: This paper proposes and analyzes the Belief Flow Networks (BFNs) framework for modeling logic-based belief propagation in agent networks, proving that consensus can be reached in strongly connected networks and providing a polynomial-time optimal 'buy-sell' scheme.
Benchmarking Large Language Models in Retrieval-Augmented Generation
Jiawei Chen (Chinese Information Processing Laboratory, Institute of Software, Chinese Academy of Sciences), Le Sun (Chinese Information Processing Laboratory, Institute of Software, Chinese Academy of Sciences)
CodeGenerationRetrievalTransformerLarge Language ModelTextMultimodalityBenchmarkRetrieval-Augmented Generation
π― What it does: This paper constructs a benchmark called RGB for retrieval-augmented generation (RAG) to systematically evaluate the performance of large language models (LLMs) on four core capabilities: noise robustness, negation rejection, information integration, and counterfactual robustness.
Benchmarking Large Language Models on Controllable Generation under Diversified Instructions
Yihan Chen (University of Science and Technology of China), Zhendong Mao (University of Science and Technology of China)
CodeGenerationTransformerLarge Language ModelPrompt EngineeringTextBenchmark
π― What it does: A benchmark called CoDI-Eval is proposed to evaluate the controllable text generation capabilities of large language models under natural language instructions.
Better than Random: Reliable NLG Human Evaluation with Constrained Active Sampling
Jie Ruan (Peking University), Yuesheng Zhu (Peking University)
CodeGenerationTextMultimodality
π― What it does: This paper studies the sampling problem in natural language generation (NLG) evaluation and proposes a Constrained Active Sampling Framework (CASF). Through the collaboration of a learner, a system sampler, and a constrained controller, it selects representative samples for manual evaluation, thereby obtaining more reliable system rankings.