These 696 AAAI 2023 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 2023 paper, free trial on arXivSub.
3D Human Pose Lifting with Grid Convolution
Yangyuxuan Kang (Institute of Software, Chinese Academy of Sciences), Enhua Wu (University of Macau)
π― What it does: A 3D human pose enhancement network based on grid convolution is proposed, which significantly improves the regression effect from 2D to 3D poses by transforming irregular skeletal graphs into regular grids for convolutional learning.
A Benchmark and Asymmetrical-Similarity Learning for Practical Image Copy Detection
Wenhao Wang (University of Technology Sydney), Yi Yang (Zhejiang University)
CodeRetrievalContrastive LearningImageBenchmark
π― What it does: This paper first constructs an image copy detection benchmark NDEC that includes a large number of hard negative samples, and proposes an Asymmetric Similarity Learning (ASL) method based on feature norm ratios to distinguish edited copies from hard-to-differentiate hard negative queries.
π― What it does: This paper studies a meta-reasoning model called CoPE that can execute actions in parallel during the planning process and proposes various solving algorithms.
A Generalized Scalarization Method for Evolutionary Multi-Objective Optimization
Ruihao Zheng (Southern University of Science and Technology), Zhenkun Wang (Southern University of Science and Technology)
CodeOptimization
π― What it does: A general Lp scalarization method (GLp) is proposed and embedded into the global replacement strategy of MOEA/D to address the mismatch between subproblems and solutions caused by different Lp scalarizations.
A Generative Approach for Script Event Prediction via Contrastive Fine-Tuning
Fangqi Zhu (Harbin Institute of Technology), Ruifeng Xu (Peng Cheng Laboratory)
CodeGenerationTransformerLarge Language ModelSupervised Fine-TuningContrastive LearningText
π― What it does: A two-stage generative model is proposed, using event-centered pre-training and contrastive fine-tuning to predict the next event in a script.
π― What it does: A Graph Fusion Model (GFMRC) is proposed, which constructs a graph of cross-language alignment and monolingual syntactic information, and learns the attention matrix in a Transformer to enhance cross-language reading comprehension performance.
Karolina StaΕczak (University of Copenhagen), Isabelle Augenstein (ETH ZΓΌrich)
CodeTransformerReinforcement LearningText
π― What it does: A latent variable-based intrinsic detection model is proposed to locate a subset of neurons encoding linguistic properties in pre-trained contextual representations.
A Model-Agnostic Heuristics for Selective Classification
Andrea Pugnana (Scuola Normale Superiore), Salvatore Ruggieri (University of Pisa)
CodeClassificationImageTabular
π― What it does: This paper proposes a model-agnostic selective classification method SCROSS, which implements a rejection strategy for probabilistic binary classifiers through cross-fitting and sample-specific quantile estimation.
π― What it does: A span-based continual learning named entity recognition model, SpanKL, is proposed to address the conflict between forward compatibility and backward compatibility in traditional sequence labeling methods.
π― What it does: A method based on pair-approximation is proposed, constructing a partial differential equation model that describes the learning dynamics of infinitely large multi-agent systems in stochastic games based on Q-learning, capable of predicting the evolution of both strategy distribution and environmental state distribution over time.
A Provable Framework of Learning Graph Embeddings via Summarization
Houquan Zhou (Institute of Computing Technology, Chinese Academy of Sciences), Xueqi Cheng (Institute of Computing Technology, Chinese Academy of Sciences)
π― What it does: This paper proposes a node embedding learning framework based on graph summarization called GELSUMM, provides theoretical proofs, and presents closed-form recovery formulas for DeepWalk, LINE, and GCN, validating its effectiveness through experiments.
A Semi-parametric Model for Decision Making in High-Dimensional Sensory Discrimination Tasks
Stephen Keeley (Fordham University), Michael Shvartsman (Meta)
CodeTabular
π― What it does: A semi-parametric psychometric model is proposed, combining traditional unidimensional sigmoid parameterization with a Gaussian process (GP) prior for contextual dimensions beyond intensity;
π― What it does: A multi-codebook vector quantization TTS system (MQTTS) for spontaneous speech in the real world is proposed and implemented, enhancing synthesis quality through multi-codebook discrete representation, single-head cross-attention, monotonic alignment, and silent audio prompts.
AC-Band: A Combinatorial Bandit-Based Approach to Algorithm Configuration
Jasmin Brandt (Paderborn University), Kevin Tierney (Bielefeld University)
CodeOptimizationHyperparameter SearchTabular
π― What it does: Proposes AC-Band, an algorithm configuration method based on combinatorial multi-armed bandits, which can find near-optimal configurations within a given budget.
π― What it does: Based on CLIP, CLIP4VLA is designed with an audio encoder that is consistent with the structure of the visual encoder, and distinguishes between speech and non-speech information in audio through audio type tokens; during the pre-training phase, cross-modal and single-modal contrastive learning is used to learn the associations between vision, text, and audio; subsequently, fine-tuning is performed on downstream tasks such as video retrieval and video caption generation.
π― What it does: A semi-supervised colorectal polyp segmentation framework called ACL-Net is proposed, which utilizes affinity map alignment between student and teacher networks to continuously optimize pseudo-labels, thereby improving segmentation performance.
π― What it does: An adaptive token-mixing mechanism called Active Token Mixer (ATM) is proposed, and the ATMNet backbone and ATMFPN neck are constructed around it.
Adaptive Dynamic Filtering Network for Image Denoising
Hao Shen (Hefei University of Technology), Wandi Zhang (Hefei University of Technology)
CodeRestorationConvolutional Neural NetworkImage
π― What it does: An image denoising network ADFNet based on adaptive dynamic filtering is proposed, aiming to efficiently preserve details while removing noise.
π― What it does: An Adaptive Hierarchical-Branch Fusion Framework (AHBF-OKD) is proposed, which constructs a deep incremental hierarchical branch and recursively uses a Teacher Assistant and attention mechanism for online knowledge distillation to address the homogenization problem of traditional OKD.
Adaptive Perturbation-Based Gradient Estimation for Discrete Latent Variable Models
Pasquale Minervini (University of Edinburgh), Mathias Niepert (University of Stuttgart)
CodeOptimizationGraph
π― What it does: An adaptive gradient estimation method called AIMLE is proposed for efficiently solving gradients in deep learning models that include discrete algorithms.
ADMoE: Anomaly Detection with Mixture-of-Experts from Noisy Labels
Yue Zhao (Carnegie Mellon University), Ahmed Awadallah (Microsoft)
CodeAnomaly DetectionMixture of ExpertsTabular
π― What it does: The ADMoE framework is proposed, utilizing a Mixture-of-Experts structure to enable end-to-end learning of anomaly detection models in a weakly supervised environment with multi-source noisy labels; it also provides a model-agnostic plugin approach that can be directly embedded into any neural network-based anomaly detection method.
π― What it does: This paper proposes an Adversarial Self-Attention (ASA) mechanism, which incorporates a learnable adversarial mask into the self-attention layer of the Transformer, thereby suppressing the model's over-reliance on keywords and enhancing the generalization and robustness of pre-training and fine-tuning.
Zhipeng Zhong (Shenzhen University), Guoping Qiu (University of Nottingham)
CodeGenerationRetrievalTransformerLarge Language ModelVision Language ModelImageTextMultimodality
π― What it does: This paper proposes a framework that combines Image Aesthetic Quality Assessment (AQA) and Image Aesthetic Description (IAC), and introduces the Aesthetic Relevance Score (ARS) for the first time to measure the degree to which text describes the aesthetics of images.
π― What it does: A global performance predictor AIO-P has been constructed, which can be pre-trained on multiple tasks and search spaces, and transferred to unseen tasks and architectures.
π― What it does: A new architecture for Alignment Enhanced Fine-tuning of pre-trained document image models (AETNet) is proposed, utilizing additional visual and textual Transformers and alignment loss to improve downstream task performance.
AMOM: Adaptive Masking over Masking for Conditional Masked Language Model
Yisheng Xiao (Soochow University), Min Zhang (Microsoft Research Asia)
CodeGenerationTransformerLarge Language ModelText
π― What it does: This paper proposes an adaptive masking strategy AMOM to improve the Conditional Masked Language Model (CMLM) for enhancing the quality and inference speed of non-autoregressive text generation.
π― What it does: In the customer service dialogue environment lacking large-scale labeled data, the author proposes a two-step semi-supervised fine-tuning methodβTask-Adaptive Fine-Tuning (TAFT). It first uses unlabeled speaker role information (customer/customer service) as a simple task for adaptation, then freezes the original model and only fine-tunes an additional lightweight adaptive layer to complete dual-task predictions of customer satisfaction and problem status.
An Efficient Algorithm for Fair Multi-Agent Multi-Armed Bandit with Low Regret
Matthew Jones (Northeastern University), Thy Nguyen (Northeastern University)
CodeOptimizationReinforcement Learning from Human FeedbackTabular
π― What it does: This paper proposes an efficient algorithm that utilizes UCB and the logarithmic concavity of the Nash social welfare function to achieve low returns in the multi-agent multi-armed bandit problem, reaching an upper bound of approximately O~(β(NKT)+NK), and provides a non-efficient algorithm that achieves this upper bound.
An Efficient Deep Reinforcement Learning Algorithm for Solving Imperfect Information Extensive-Form Games
Linjian Meng (Nanjing University), Yang Gao (Nanjing University)
CodeReinforcement Learning
π― What it does: Proposes the FTRL-based ORD-DGF algorithm and its deep implementation Deep FTRL-ORW for efficiently learning Nash equilibria in large incomplete information extensive-form games.
π― What it does: This paper proposes an unsupervised contrastive learning framework based on multilingual back-translation (RTT) data augmentation to generate more semantically expressive sentence embeddings.
An Extreme-Adaptive Time Series Prediction Model Based on Probability-Enhanced LSTM Neural Networks
Yanhong Li (Santa Clara University), David C. Anastasiu (Santa Clara Valley Water District)
CodeRecurrent Neural NetworkGaussian SplattingMultimodalityTime Series
π― What it does: A highly adaptive time series forecasting model NEC+ is proposed, which learns the prediction functions for both extreme events and normal events simultaneously, and dynamically selects them using a classifier;
π― What it does: This paper proposes a joint optimization algorithm for wage determination and online task allocation on crowdsourcing platforms, aiming to maximize platform revenue.
An Operator Theoretic Approach for Analyzing Sequence Neural Networks
Ilan Naiman (Ben Gurion University), Omri Azencot (Ben Gurion University)
CodeClassificationExplainability and InterpretabilityRecurrent Neural NetworkTextTime SeriesBiomedical DataElectrocardiogramOrdinary Differential Equation
π― What it does: A KANN method based on Koopman theory is proposed for the global analysis of the hidden state dynamics of sequential neural networks and to reveal semantic information.
Anomaly Segmentation for High-Resolution Remote Sensing Images Based on Pixel Descriptors
Jingtao Li (Wuhan University), Yanfei Zhong (Wuhan University)
CodeSegmentationAnomaly DetectionConvolutional Neural NetworkImageAgriculture Related
π― What it does: This paper proposes a high-resolution remote sensing image anomaly segmentation model (ASD) based on pixel descriptors, achieving precise localization of anomalous pixels by learning compact, diverse, and feature-rich pixel descriptors in the feature space.
π― What it does: A deep distillation framework based on attention (ADD) is proposed, utilizing 3D-aware positional encoding for distillation in monocular 3D object detection to enhance 3D localization accuracy.
CodeRecognitionOptimizationTransformerSupervised Fine-TuningTextBiomedical Data
π― What it does: In response to the low-resource and class-imbalanced named entity recognition task, the authors transformed the traditional BIO annotation into two tasks (whether it belongs to an entity and whether it is the first word of the entity) and used an AUC maximization loss function for training.
π― What it does: Reconstructing a 3D ear model from a single view image, constructing a high-quality ear dataset (AudioEar3D and AudioEar2D), and combining the reconstructed ear with a 3D human model to simulate personalized HRTF for accurate spatial audio rendering.
Auto-Weighted Multi-View Clustering for Large-Scale Data
Xinhang Wan (National University of Defense Technology), Lu Zhou (Nanjing University of Aeronautics and Astronautics)
CodeOptimizationComputational EfficiencyImage
π― What it does: A self-weighted multi-view clustering method AWMVC is proposed, which utilizes multi-dimensional matrix decomposition and automatically weighted generation of consensus clustering matrices, suitable for large-scale data.
π― What it does: This paper proposes an automated weight initialization algorithm called AutoInit, which can maintain the network's mean at zero and variance at one by analyzing the propagation of signal mean and variance under different network structures, activation functions, layer types, and hyperparameter settings, thus avoiding gradient explosion or vanishing.
Automated Verification of Propositional Agent Abstraction for Classical Planning via CTLK Model Checking
Kailun Luo (Dongguan University of Technology)
CodeAgentic AIBenchmark
π― What it does: The research addresses the existence problem of propositional agent abstraction given a refined mapping and provides an automatic verification method based on CTLK model checking.
Automated Verification of Social Laws in Numeric Settings
Ronen Nir (Technion Israel Institute of Technology), Erez Karpas (Technion Israel Institute of Technology)
CodeOptimizationTabularBenchmark
π― What it does: The study verifies the robustness of the social method in a numerical planning environment, proposing a scheme to compile robustness verification into a single-agent numerical planning problem, and implements and evaluates it.
π― What it does: The AVCAffe dataset has been constructed and made publicly available, recording 108 hours of audio-video from 106 subjects in a remote work context, divided into 58k segments by task, with self-reported emotional (arousal, valence) and cognitive load (mental demand, time demand, effort, etc.) labels.
π― What it does: A GAN-based speech synthesizer named Avocodo is proposed, aimed at eliminating imaging and aliasing noise caused by upsampling and downsampling in traditional GAN vocoders;
π― What it does: A background mixing-based data augmentation method (BGMix) and an incremental consistency loss with real data are proposed to construct a weakly supervised change detection framework that only requires image-level labels.
π― What it does: A pruning method based on GPU Block Column Pruning (BCBP) is proposed, utilizing the tiling structure of convolutional layers to prune by column blocks while maintaining the same sparsity rate for each block, thereby maximizing the utilization of GPU computing resources.
Balanced Meta Learning and Diverse Sampling for Lifelong Task-Oriented Dialogue Systems
Qiancheng Xu (Georgia Institute of Technology), Ruifeng Xu (Harbin Institute of Technology)
CodeMeta LearningTransformerLarge Language ModelText
π― What it does: A two-stage lifelong task-oriented dialogue system MetaLTDS is proposed, which balances catastrophic forgetting suppression and knowledge transfer;
Kiran Tomlinson (Cornell University), Jon Kleinberg (Cornell University)
CodeTabular
π― What it does: This study investigates how the length of the ballot (i.e., the number of candidates each voter can rank) affects the election results in Instant Runoff Voting (IRV). It proposes that theoretically, up to k-1 different winners can occur and provides a construction for the matching lower bound. It also explores scenarios under preference constraints such as single-peaked and single-crossing. The impact of different ballot lengths on the winners is validated through simulations and actual election data from PrefLib.
π― What it does: A new federated neural matching method (NAFI) is proposed, which compensates for the lack of global information in the matching cost of PFNM by adding a KullbackβLeibler (KL) penalty term to the original PFNM framework.
Jakob Weissteiner (University of Zurich), Sven Seuken (University of Zurich)
CodeOptimizationTabularBenchmark
π― What it does: A combination allocation mechanism based on Bayesian optimization, BOCA, is proposed to efficiently collect agent preference information in combinatorial auctions.
π― What it does: This study investigates behavior estimation under multi-source data, proposing a latent variable model and the LBRAC-v method to learn and allocate multiple behavior strategies.
Bellman Meets Hawkes: Model-Based Reinforcement Learning via Temporal Point Processes
Chao Qu (Ant Group), Hongyuan Mei (Toyota Technological Institute at Chicago)
CodeReinforcement LearningSequential
π― What it does: This paper proposes a modeling reinforcement learning framework based on the Hawkes point process (NHPI) for learning optimal intervention strategies in continuous-time, random event-driven semi-Markov decision problems.
BETA-CD: A Bayesian Meta-Learned Cognitive Diagnosis Framework for Personalized Learning
Haoyang Bi (University of Science and Technology of China), Jinze Wu (iFLYTEK AI Research)
CodeMeta LearningTabular
π― What it does: The BETA-CD framework is proposed, which combines Bayesian hierarchical models and meta-learning for personalized learning cognitive diagnosis, allowing for rapid adaptation to new students and quantifying the uncertainty of diagnostic results.
Better Context Makes Better Code Language Models: A Case Study on Function Call Argument Completion
Hengzhi Pei (University of Illinois Urbana-Champaign), George Karypis (Amazon Web Services)
CodeAI Code AssistantTransformerLarge Language ModelSupervised Fine-TuningText
π― What it does: A reproducible collection of Python project environments (PYENVS) was constructed, and based on this, a function call parameter completion dataset (CALLARGS) was created. A program analyzer was used to extract function implementation and usage information across files and projects, studying its impact on the completion performance of pre-trained code language models.
Better Generalized Few-Shot Learning Even without Base Data
Seong-Woong Kim (Inha University), Dong-Wan Choi (Inha University)
CodeClassificationMeta LearningImage
π― What it does: This study proposes a zero-shot generalized few-shot learning method to address the issue of integrating new category knowledge into pre-trained models without base class samples.
Hedayat Zarkoob (University of British Columbia), Kevin Leyton-Brown (University of British Columbia)
CodeTabular
π― What it does: This study improves the Bayesian inference framework of the peer assessment system by incorporating low-effort scoring modeling, upper truncation handling of discrete rating levels, and interpretable weighted average score output.
π― What it does: This paper proposes BEVDepth, a multi-view BEV 3D object detection framework achieved through explicit depth supervision and a camera-aware depth prediction module.
π― What it does: A multi-view 3D object detection framework named BEVStereo is proposed, which enhances depth estimation through dynamic temporal stereo technology and achieves high-precision detection of targets such as vehicles and pedestrians by combining the BEVDepth structure.
Beyond Graph Convolutional Network: An Interpretable Regularizer-Centered Optimization Framework
Shiping Wang (Fuzhou University), Yong Chen (Beijing University of Posts and Telecommunications)
CodeOptimizationExplainability and InterpretabilityGraph Neural NetworkGraph
π― What it does: An interpretable optimization framework based on regularization is proposed, and under this framework, a dual-regularization graph convolutional network tsGCN is designed.
π― What it does: This paper proposes an unsupervised graph representation learning framework called GREET, which can distinguish between similar edges and dissimilar edges, and generates node representations using both types of edges.
Bi-directional Feature Reconstruction Network for Fine-Grained Few-Shot Image Classification
Jijie Wu (Lanzhou University of Technology), Yi-Zhe Song (University of Surrey)
CodeClassificationMeta LearningTransformerImage
π― What it does: A bidirectional feature reconstruction network (Bi-FRN) is proposed, achieving bidirectional reconstruction from support to query and from query to support, enhancing the discriminative ability of fine-grained few-shot classification.
Haichao Yu (University of Illinois Urbana-Champaign), Humphrey Shi (Wormpex AI Research)
CodeClassificationImage
π― What it does: Designed and trained BoostNet, a framework that treats early exit dynamic neural networks as a gradient boosting additive model to address the issue of data distribution mismatch between the training and inference phases.
π― What it does: An efficient rendering method based on sparse point clouds is proposed, utilizing radiance mapping to achieve single-pixel inference.
Boosting Semi-Supervised Semantic Segmentation with Probabilistic Representations
Haoyu Xie (Northeastern University), Chang Xu (University of Sydney)
CodeSegmentationContrastive LearningImage
π― What it does: This paper proposes a probabilistic pixel representation contrastive learning framework (PRCL), which maps pixels to multivariate Gaussian distributions. It utilizes probabilistic information to suppress the negative impact of inaccurate pseudo-labels on contrastive learning, thereby enhancing the representation quality of semi-supervised semantic segmentation.
Breaking Immutable: Information-Coupled Prototype Elaboration for Few-Shot Object Detection
Xiaonan Lu (Aerospace Information Research Institute, Chinese Academy of Sciences), Kun Fu (Aerospace Information Research Institute, Chinese Academy of Sciences)
π― What it does: A Few-Shot object detection framework ICPE based on conditional information coupling and dynamic aggregation is proposed, which can generate high-quality prototypes with query awareness for each query image.
π― What it does: The Calibrated Teacher framework is proposed, which generates more reliable pseudo-labels by online calibrating the confidence of the teacher network in sparse annotated object detection, and combines Focal IoU Weight to reduce the misguidance of negative samples caused by missing annotations.
π― What it does: This paper proposes the CALIP method, which achieves cross-modal interaction between visual and textual features in CLIP through a non-parametric attention mechanism to enhance zero-shot classification performance.
π― What it does: This paper proposes a label-specific feature-based unsupervised learning method called UCL, which first generates soft pseudo-labels through graph label enhancement, then classifies samples into positive, negative, and uncertain based on confidence levels, utilizes spectral clustering to obtain cluster centers to construct label-specific features, and finally combines these with existing classifiers (such as SURE) to improve performance.
π― What it does: This paper studies the search for Strong Lottery Tickets (SLT) in generative models, which can achieve or exceed the generative performance of the full model in a sparse subnetwork without updating weights at random initialization.
π― What it does: A cascade network named CasFusionNet is proposed, which utilizes dense feature fusion to achieve joint inference of point cloud semantic scene completion and semantic segmentation.
Causal Conditional Hidden Markov Model for Multimodal Traffic Prediction
Yu Zhao (Beihang University), Mulan Wang (Beihang University)
CodeRecurrent Neural NetworkGraph Neural NetworkAuto EncoderMultimodalityTime Series
π― What it does: This paper proposes the Causal Conditional Hidden Markov Model (CCHMM), which views multimodal traffic flow prediction as a conditional Markov process and decouples the core physical concepts of the traffic system (regional attraction factors, traffic demand factors, and speed factors) through causal structures.
Causal Inference with Conditional Instruments Using Deep Generative Models
Debo Cheng (Guangxi Normal University), Thuc Duy Le (University of South Australia)
CodeRepresentation LearningAuto EncoderTabularFinance Related
π― What it does: A conditional instrumental variable (CIV) method based on deep generative models is proposed to discover CIVs and their conditional sets from data with potential confounding factors, in order to estimate average causal effects.
Causal Recurrent Variational Autoencoder for Medical Time Series Generation
Hongming Li (University of Florida), Jose Principe (University of Florida)
CodeGenerationData SynthesisRecurrent Neural NetworkAuto EncoderTime SeriesBiomedical Data
π― What it does: Developed a cyclic variational autoencoder (CR-VAE) integrated with Granger causality graphs for the generation and causal discovery of medical time series.
π― What it does: A cross-domain transferable attack method CDTA is designed, utilizing unlabeled contrastive learning to train a feature extractor and generate attack samples.
Centerless Multi-View K-means Based on the Adjacency Matrix
Han Lu (Xidian University), Wei Xia (University of Evansville)
CodeOptimizationGraph Neural NetworkMultimodality
π― What it does: This paper proposes a centralized multi-view K-Means clustering method that constructs a distance matrix using an adjacency matrix and fully exploits multi-view information through tensor Schatten p-norm regularization, completing clustering without the need for initializing centroids.
Nikil Roashan Selvam (University of California), YooJung Choi (Arizona State University)
CodeOptimizationTabular
π― What it does: This study investigates the fairness issues under partial observation conditions in probabilistic circuits (PC), proposes and tests the definition of discrimination patterns, and provides a framework for summarizing these patterns.
CF-ViT: A General Coarse-to-Fine Method for Vision Transformer
Mengzhao Chen (Xiamen University), Rongrong Ji (Xiamen University)
CodeClassificationTransformerImage
π― What it does: CF-ViT is proposed, a two-stage coarse-fine granularity visual Transformer that first uses coarse granularity partitioning for rapid classification, and if confidence is insufficient, it performs fine-grained re-partitioning on important areas to further improve accuracy.
π― What it does: This paper proposes a Class-Independent Regularization (CIR) for training deep networks under noisy labels, which regularizes the traditional softmax classifier by decomposing K-class multiclass classification into K independent binary classifiers, and implements a heterogeneous adaptive co-teaching mechanism to achieve complementary training of sample selection and classification.
CLIP-ReID: Exploiting Vision-Language Model for Image Re-identification without Concrete Text Labels
Siyuan Li (East China Normal University), Qingli Li (East China Normal University)
CodeRecognitionRetrievalTransformerVision Language ModelContrastive LearningImage
π― What it does: Fine-tune the pre-trained CLIP vision-language model for person/vehicle ReID tasks and propose a two-stage training strategy to introduce learnable text tokens.
π― What it does: Introducing CLIPVG: a text-guided image editing framework based on differentiable vector graphics, which optimizes color and geometric parameters directly in the vector domain using CLIP loss, enabling semantic image editing without the need for additional generative models.
π― What it does: A Cluster-guided Contrastive Graph Clustering network (CCGC) is proposed, which enhances the performance of unsupervised graph node clustering by guiding the construction of positive and negative samples through high-confidence clustering results.
π― What it does: This paper proposes an unsupervised meta-learning model based on causal structures, called CMVAE, to eliminate background bias and improve few-shot classification performance.
π― What it does: This study proposes the CaTHPO framework, which can accelerate the Bayesian Optimization (BO) process in cross-program hyperparameter optimization by transferring knowledge from similar programs.
Cogito Ergo Summ: Abstractive Summarization of Biomedical Papers via Semantic Parsing Graphs and Consistency Rewards
Giacomo Frisoni (University of Bologna), Gianluca Moro (University of Bologna)
CodeGenerationGraph Neural NetworkTransformerLarge Language ModelReinforcement LearningTextBiomedical DataReview/Survey Paper
π― What it does: This paper proposes an abstract medical literature summarization framework COGITOERGOSUMM based on semantic graphs (event extraction and AMR), which can generate more concise and readable summaries while maintaining factual consistency.
π― What it does: This paper proposes a Collaborative Enhancement (COLA) method that combines user-item interaction graphs with retrieval of similar dialogues to simultaneously improve item representation and user preference modeling, thereby enhancing the recommendation effectiveness and generation quality of conversational recommendation systems.
Collective Intelligence in Human-AI Teams: A Bayesian Theory of Mind Approach
Samuel Westby (Northeastern University), Christoph Riedl (Northeastern University)
CodeText
π― What it does: A multi-agent network based on Bayesian inference was constructed to infer and enhance the collective intelligence of human teams in real-time during team chats.
π― What it does: A single-model self-supervised mixed pre-training framework called CoMAE is proposed, which uses a unified Transformer encoder to perform cross-modal contrastive learning and multi-modal masked autoencoder pre-training on RGB and depth data, thereby learning high-quality multi-modal representations.
π― What it does: By extending the discriminator into a feature embedding network and maximizing entropy in its embedding space, MaEM-GAN is proposed to alleviate mode collapse in GANs.
Compact Transformer Tracker with Correlative Masked Modeling
Zikai Song (Huazhong University of Science and Technology), Wei Yang (La Trobe University)
CodeObject TrackingTransformerImageVideo
π― What it does: A compact Transformer tracker (CTTrack) based on raw self-attention is proposed, with a pluggable mutual mask decoder added during the training phase to enhance feature representation capabilities.
π― What it does: This paper proposes a competitive semi-supervised relation extraction framework called AdvSRE, which constructs an extreme minimax game on unlabeled text, utilizing the adversarial interplay between a generator and a discriminator to fully exploit diverse relational expressions.
π― What it does: Proposed a Compositional Prototypical Networks that learns attribute-level component prototypes and reuses these prototypes to construct category prototypes in few-shot classification.
Compressed Heterogeneous Graph for Abstractive Multi-Document Summarization
Miao Li (University of Melbourne), Jey Han Lau (University of Melbourne)
CodeGraph Neural NetworkTransformerTextGraph
π― What it does: A multi-document summarization model based on compressed heterogeneous graphs, HGSUM, is proposed. It constructs a heterogeneous graph using three types of nodes (words, sentences, documents) and various edge types, and encodes it through a graph attention network. The compressed output is then used as input for the text decoder.