These 241 IJCAI 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 IJCAI 2023 paper, free trial on arXivSub.
A Bitwise GAC Algorithm for Alldifferent Constraints
Zhe Li (National University of Defense Technology), Zhanshan Li (Jilin University)
CodeOptimizationComputational EfficiencyBenchmark
π― What it does: Proposed a GAC algorithm called Alldiff bit based on bitwise operations for efficiently handling all-different constraints.
π― What it does: Propose a two-stage framework named CEKFA to enhance link prediction performance in open knowledge graphs (OpenKG) through the normalization of relation phrases (RP) and triplets.
A Diffusion Model with Contrastive Learning for ICU False Arrhythmia Alarm Reduction
Feng Wu (Xi'an Jiaotong University), Li-wei H. Lehman (Massachusetts Institute of Technology)
CodeAnomaly DetectionTransformerDiffusion modelContrastive LearningTime SeriesBiomedical DataElectronic Health RecordsElectrocardiogram
π― What it does: Propose a generative method based on conditional diffusion models and contrastive learning for detecting false alarms in ICU arrhythmia alerts.
π― What it does: Proposes a fast adaptive randomized PCA algorithm called farPCA, which can automatically determine the rank of low-rank decomposition under a given error tolerance.
π― What it does: Propose an algorithm based on hierarchical population training (HiPT), enabling deep reinforcement learning agents to learn multiple optimal response strategies and dynamically switch between them when facing diverse partners, thereby enhancing collaboration with unknown partners.
A Large-Scale Film Style Dataset for Learning Multi-frequency Driven Film Enhancement
Zinuo Li (Shenzhen Institute of Advanced Technology, Chinese Academy of Sciences), Chi-Man Pun (Shenzhen Institute of Advanced Technology, Chinese Academy of Sciences)
π― What it does: Proposed the FilmSet large-scale movie-style image dataset and developed the FilmNet framework based on the Laplacian pyramid, achieving multi-band image style transfer;
A Logic-based Approach to Contrastive Explainability for Neurosymbolic Visual Question Answering
Thomas Eiter (TU Wien), Johannes Oetsch (TU Wien)
CodeObject DetectionExplainability and InterpretabilityConvolutional Neural NetworkRecurrent Neural NetworkVision Language ModelContrastive LearningImageTextBenchmark
π― What it does: Proposed a logic-based contrastive explanation framework that generates contrastive explanations (CE) using a neuro-symbolic VQA architecture (perception module + LSTM + ASP reasoning)
A Mathematical Runtime Analysis of the Non-dominated Sorting Genetic Algorithm III (NSGA-III)
Simon Wietheger (University of Potsdam), Benjamin Doerr (Institut Polytechnique de Paris)
CodeOptimizationBenchmark
π― What it does: This paper conducts the first analysis of the mathematical runtime of NSGA-III on the three-objective ONEMINMAX (3-OMM) benchmark, proving that the algorithm does not lose Pareto optimal solutions and can cover the complete Pareto front within an expected O(n log n) iterations.
π― What it does: This paper proposes a new branch-and-bound algorithm (DiseMKP), which introduces an improved upper bound (DisePUB) and incremental pruning (inprocessing) strategies into solving the maximum k-plex problem (MKP). These strategies dynamically prune the graph during the search process, significantly reducing the search space.
A Unifying Formal Approach to Importance Values in Boolean Functions
Hans Harder (University of Paderborn), Clemens Dubslaff (Eindhoven University of Technology)
CodeExplainability and InterpretabilityComputational EfficiencyBenchmark
π― What it does: Proposes a unified IVF framework, formally defining the importance of variables in Boolean functions, and provides multiple instances and symbolic computation methods.
Action Recognition with Multi-stream Motion Modeling and Mutual Information Maximization
Yuheng Yang (Jilin University), Kui Ren (Zhejiang University)
CodeRecognitionGraph Neural NetworkVideo
π― What it does: This paper proposes a multi-stream action recognition framework that integrates low-level (coordinates, bone length, velocity, etc.) and high-level (angular acceleration) motion features, and supervises the features through mutual information maximization.
Harsha Kokel (IBM T.J. Watson Research Center), Shirin Sohrabi (IBM T.J. Watson Research Center)
CodeOptimizationReinforcement Learning
π― What it does: Automate the reduction of action labels in classical planning tasks, introduce the concept of actionable mutex groups, and achieve label space compression through parameter seed set problems (solved using delete-free planning).
π― What it does: This paper proposes an active visual exploration method based on Transformer attention entropy (Attention-Map Entropy, AME), which directly selects the next frame perspective (glimpse) by leveraging the attention uncertainty within the MAE autoencoder, without requiring additional sampling decision heads or auxiliary losses;
π― What it does: Proposed AMCRNet, which improves video action detection through multi-scale context and bidirectional high-order interaction relationships;
Adaptive Estimation Q-learning with Uncertainty and Familiarity
Xiaoyu Gong (Jilin University), Zongze Li (Jilin University)
CodeReinforcement Learning
π― What it does: Propose the Adaptive Estimation Q-learning (AEQ) method, which dynamically adjusts Q-value estimation by combining uncertainty and familiarity, thereby mitigating overestimation and underestimation biases in offline reinforcement learning.
Adaptive Path-Memory Network for Temporal Knowledge Graph Reasoning
Hao Dong (Chinese Academy of Sciences), Yanjie Fu (University of Central Florida)
CodeRepresentation LearningGraph Neural NetworkGraphTime Series
π― What it does: Propose a relation-based adaptable path memory network (DaeMon), which performs knowledge graph reasoning at future time points by leveraging relational chain (path) information between query entities and candidates in historical time.
Advancing Post-Hoc Case-Based Explanation with Feature Highlighting
Eoin M. Kenny (Massachusetts Institute of Technology), Mark T. Keane (University College Dublin)
CodeClassificationExplainability and InterpretabilityConvolutional Neural NetworkImage
π― What it does: Propose two post-hoc case-based reasoning (CBR) methods combined with feature highlighting techniques, which can extract and display multiple salient features from image classification models for each prediction, and associate these features with corresponding cases in the training data to generate interpretable prediction explanations.
π― What it does: This study proposes the AdvAmd method, which transforms adversarial examples into a 'friendly force' that improves model accuracy on clean data through fine-grained attacks, generating intermediate samples, incorporating auxiliary batch normalization (BN), and using adaptive loss, while maintaining adversarial robustness.
CodeExplainability and InterpretabilityAdversarial AttackReinforcement LearningSequentialBenchmark
π― What it does: Proposed a task-agnostic safety protection framework called AdvEx-RL, which trains an adversary policy to identify and eliminate all actions violating safety constraints, then learns a safety policy that acts as a safety valve to block dangerous actions during runtime.
π― What it does: This paper proposes the APD framework, which enhances RSI change detection through alignment, perturbation, and decoupling operations.
An Effective and Efficient Time-aware Entity Alignment Framework via Two-aspect Three-view Label Propagation
Li Cai (East China Normal University), Man Lan (East China Normal University)
CodeComputational EfficiencyRepresentation LearningGraphTime Series
π― What it does: Propose a non-neural network time-aware entity alignment framework called LightTEA, which achieves efficient alignment of temporal knowledge graphs using label propagation, sparse similarity, Sinkhorn operations, and iterative learning.
π― What it does: This paper studies a nominal-invariant graph neural network representation and ensemble strategy for automated theorem proving, achieving higher solving rates on multiple domain datasets.
π― What it does: Designed a lower bound based on 2-hop adjacency, and proposed the EMOS exact branch-and-bound algorithm to solve the minimum dominating set problem.
Analyzing and Combating Attribute Bias for Face Restoration
Zelin Li (Southern University of Science and Technology), Bo Tang (Southern University of Science and Technology)
CodeRestorationSuper ResolutionConvolutional Neural NetworkVision Language ModelAuto EncoderGenerative Adversarial NetworkImage
π― What it does: This paper proposes the DebiasFR framework, aiming to address the attribute bias problem that occurs during facial restoration, i.e., the restored facial attributes (such as gender and age) significantly differ from those in the original low-resolution image or real high-resolution image.
π― What it does: Propose a quantification method based on Markov Decision Processes (MDP), using a probabilistic model to check evidence of autonomous agents achieving intentional events in uncertain environments
π― What it does: Proposed an APR framework based on autoencoders for online long-range point cloud registration, leveraging aggregated point cloud reconstruction to enhance feature representation;
Asynchronous Communication Aware Multi-Agent Task Allocation
Ben Rachmut (Ben Gurion University of the Negev), Roie Zivan (Ben Gurion University of the Negev)
CodeOptimizationTabular
π― What it does: Proposes an asynchronous communication-aware multi-agent task allocation algorithm called FMC ATA, aimed at solving task allocation problems with spatial and temporal constraints in physical environments, particularly suitable for scenarios with unreliable communication and dynamically changing environments.
Augmenting Automated Spectrum Based Fault Localization for Multiple Faults
Prantik Chatterjee (Indian Institute of Technology Kanpur), Subhajit Roy (Indian Institute of Technology Kanpur)
CodeTabularBenchmark
π― What it does: Proposed ARTEMIS, an SBFL enhancement method based on multi-universe analysis, which improves the effectiveness of multi-defect localization by simulating the repair of dominant defects.
Automatic Verification for Soundness of Bounded QNP Abstractions for Generalized Planning
Zhenhe Cui (Sun Yat-sen University), Yongmei Liu (Sun Yat-sen University)
Code
π― What it does: Provide a proof-theoretic characterization of the soundness of general planning (GP) abstraction, propose automatically verifiable sufficient conditions, and subsequently implement a Bounded QNP abstraction soundness verification system based on an SMT solver.
Beyond Homophily: Robust Graph Anomaly Detection via Neural Sparsification
Zheng Gong (University of Science and Technology of China), Jingyu Peng (University of Science and Technology of China)
CodeAnomaly DetectionGraph Neural NetworkGraph
π― What it does: Propose a graph anomaly detection framework called SparseGAD based on neural sparsification, which removes structural noise and combines homogeneity and heterogeneity information for robust anomaly detection.
Bipolar Abstract Dialectical Frameworks Are Covered by Kleeneβs Three-valued Logic
Ringo Baumann (Leipzig University), Maximilian Heinrich (Bauhaus University Weimar)
CodeComputational Efficiency
π― What it does: This paper demonstrates that Kleene's three-valued logic can be directly applied in Bipolar Abstract Dialogue Frameworks (BADFs) to compute the Ξ-operator, thereby avoiding the enumeration of all two-valued completions.
Pengpeng Chen (China's Aviation System Engineering Research Institute), Peng Lin (China's Aviation System Engineering Research Institute)
CodeAdversarial AttackText
π― What it does: This paper proposes a SubPac framework for black-box attacks, which can achieve data destruction by optimizing instance selection and label poisoning under unknown label aggregation models.
Black-box Prompt Tuning for Vision-Language Model as a Service
Lang Yu (East China Normal University), Liang He (East China Normal University)
CodeClassificationOptimizationPrompt EngineeringVision Language ModelImageTextMultimodality
π― What it does: Propose a black-box prompt tuning framework BPT-VLM for vision-language models in the MaaS scenario, which can learn task-related visual and linguistic prompts without accessing gradients or model structures;
π― What it does: This paper proposes a Multi-Relationship Margin Loss (MRM) to significantly enhance the detection capability of unknown samples in few-shot open set recognition (FSOSR) tasks while maintaining the classification accuracy of known classes.
π― What it does: Proposed a constrained version of TPE, called c-TPE, based on the tree-structured Parzen estimator (TPE), for addressing inequality constraints in hyperparameter optimization.
Calibrating a Deep Neural Network with Its Predecessors
Linwei Tao (University of Sydney), Chang Xu (University of Sydney)
CodeClassificationExplainability and InterpretabilityImage
π― What it does: Proposed a predecessor combination search (PCS) method that achieves better model calibration by selecting the optimal predecessor weight combinations for each network block.
Can You Improve My Code? Optimizing Programs with Local Search
Fatemeh Abdollahi (University of Alberta), Levi H. S. Lelis (University of Alberta)
CodeOptimizationText
π― What it does: This paper proposes a program optimization method called POLIS, which uses local search and existing enumerative synthesizers to incrementally improve existing programs line by line, aiming to enhance quantifiable objective functions (e.g., game scores)
Case-Based Reasoning with Language Models for Classification of Logical Fallacies
Zhivar Sourati (University of Southern California), Alain Mermoud (armasuisse Science and Technology)
CodeClassificationTransformerLarge Language ModelPrompt EngineeringTextRetrieval-Augmented Generation
π― What it does: The study proposes a method that combines case-based reasoning (CBR) with language models to classify logical fallacies in natural language arguments.
Choose your Data Wisely: A Framework for Semantic Counterfactuals
Edmund Dervakos (National Technical University of Athens), Giorgos Stamou (National Technical University of Athens)
CodeExplainability and InterpretabilityData-Centric LearningImageAudio
π― What it does: Proposed a semantic inverse causal explanation framework based on knowledge graphs, along with a corresponding algorithm for computing the minimal edit set.
CiT-Net: Convolutional Neural Networks Hand in Hand with Vision Transformers for Medical Image Segmentation
Tao Lei (Shaanxi University of Science and Technology), Asoke Nandi (Brunel University London)
CodeSegmentationConvolutional Neural NetworkTransformerImageBiomedical Data
π― What it does: Proposed a dual-branch parallel CiT-Net architecture, combining dynamic deformable convolution with deformable window adaptive complementary attention mechanism for medical image segmentation.
CodeClassificationTransformerContrastive LearningImageAgriculture Related
π― What it does: Propose the CLE-ViT model, combining self-supervised instance-level contrastive learning with Vision Transformer for ultra-fine-grained visual classification.
Co-Certificate Learning with SAT Modulo Symmetries
Markus Kirchweger (TU Wien), Stefan Szeider (TU Wien)
CodeOptimizationComputational EfficiencyGraphPhysics Related
π― What it does: Proposes a collaborative co-certificate learning method based on SAT modular symmetry (SMS) to search for graphs satisfying co-NP properties in graph space, thereby improving the lower bound of Kochen-Specker vector system sizes.
π― What it does: Propose a new semi-supervised medical image segmentation framework UCMT, combining collaborative average teacher and uncertainty-based regional mixing to enhance pseudo-label quality while maintaining model differences.
π― What it does: This paper proposes two new algorithms for computing the twin-width of a graph: a more compact SAT encoding (SAT-CPT), and another algorithm based on branch-and-bound (BB-CCH).
π― What it does: This paper proposes a heterogeneous curvature space model based on Ricci curvature to accomplish end-to-end unsupervised graph clustering tasks.
Contrastive Learning and Reward Smoothing for Deep Portfolio Management
Yun-Hsuan Lien (National Yang Ming Chiao Tung University), Yu-Shuen Wang (National Yang Ming Chiao Tung University)
CodeGraph Neural NetworkReinforcement LearningContrastive LearningTime SeriesFinance Related
π― What it does: Utilizing deep reinforcement learning (DRL) combined with contrastive learning and reward smoothing to train agents for asset allocation and trading in financial markets.
π― What it does: Proposes ContrastMotion, a self-supervised LiDAR scene motion estimator based on pillar representations, which predicts pillar correspondences through contrastive learning of pillar features to obtain scene motion.
π― What it does: Proposes the Context Outlooker (COOL) local attention mechanism, which can be inserted into any Transformer model, utilizing multi-window local attention and convolutional blocks to enhance local context information, thereby improving performance on NLP tasks.
CROP: Towards Distributional-Shift Robust Reinforcement Learning Using Compact Reshaped Observation Processing
Philipp Altmann (LMU Munich), Thomy Phan (LMU Munich)
CodeDomain AdaptationReinforcement Learning
π― What it does: Design and evaluate Compact Reshaped Observation Processing (CROP), which reduces observational information in reinforcement learning through three manual compression methods (radius, action, object) to improve training speed and robustness to distribution drift.
π― What it does: Proposed a cross-modal global interaction and local alignment (GILA) framework for audio-visual speech recognition (AVSR), achieving deep complementarity and temporal consistency between audio and visual features.
CTW: Confident Time-Warping for Time-Series Label-Noise Learning
Peitian Ma (South China University of Technology), Qianli Ma (South China University of Technology)
CodeClassificationConvolutional Neural NetworkAuto EncoderTime Series
π― What it does: Proposes a CTW method for learning with time series label noise, which expands the clean sample distribution by applying time warping on confident samples and enhances model robustness by eliminating class bias through category-normalized loss.
π― What it does: Propose a cross-perspective trajectory prediction model (XVTP3D) based on shared 3D queries, which can generate multi-modal predictions while maintaining 3D perspective consistency.
π― What it does: Propose a distributed anomaly detection framework that uses RNN to predict the action distribution of peer agents and calculates normality scores to identify attacked agents;
π― What it does: Proposed a multi-contrast MRI super-resolution and reconstruction model (MC-CDic) based on deep unfolded convolutional dictionaries, achieving precise reference image transfer by explicitly separating common and unique features;
π― What it does: Propose DeepPSL, an end-to-end trainable framework that combines deep learning with Probabilistic Soft Logic (PSL), achieving a unified system that integrates perception (deep network predicate prediction) and reasoning (HL-MRF).
DeLELSTM: Decomposition-based Linear Explainable LSTM to Capture Instantaneous and Long-term Effects in Time Series
Chaoqun Wang (City University of Hong Kong), Zhixiang Huang (JD Digits)
CodeExplainability and InterpretabilityRecurrent Neural NetworkTime Series
π― What it does: Proposed a DeLELSTM model that decomposes LSTM hidden states into a linear combination of past and current information, capturing the long-term and immediate effects of each variable to achieve interpretability in time series prediction.
π― What it does: Propose a denoising self-incremental learning framework DSL for social recommendation, which can perform adaptive cross-view alignment between user-item interactions and social graphs, thereby improving recommendation quality under sparse data.
π― What it does: Construct a detector that does not require pre-generated adversarial samples or access to the target face recognition system, using training data containing only real human faces and their self-perturbed samples. The detector cascades self-perturbation generation, decision boundary regularization, and max pooling classifier on input images, ultimately achieving detection of adversarial faces generated by unknown attack methods.
Diagnose Like a Pathologist: Transformer-Enabled Hierarchical Attention-Guided Multiple Instance Learning for Whole Slide Image Classification
Conghao Xiong (Chinese University of Hong Kong), Irwin King (Chinese University of Hong Kong)
CodeClassificationTransformerImageBiomedical Data
π― What it does: Proposed a hierarchical attention-guided multi-instance learning (HAG-MIL) framework that mimics pathologists' layer-by-layer focus and classification on multi-resolution whole slide images (WSI).
DiffAR: Adaptive Conditional Diffusion Model for Temporal-augmented Human Activity Recognition
Shuokang Huang (Imperial College London), Julie McCann (Imperial College London)
CodeRecognitionConvolutional Neural NetworkTransformerDiffusion modelTime Series
π― What it does: Propose the DiffAR method, which utilizes an Adaptive Conditional Diffusion Model (ACDM) to perform time augmentation on WiFi CSI, including predicting future time periods and imputing missing values. The enhanced CSI is combined with the original CSI to train an ensemble classifier, achieving more robust human activity recognition.
π― What it does: This paper proposes a Discriminative-Invariant Representation Learning (DIRL) framework to address the selection bias problem in recommendation systems.
Disentanglement of Latent Representations via Causal Interventions
GaΓ«l Gendron (University of Auckland), Gillian Dobbie (University of Auckland)
CodeGenerationExplainability and InterpretabilityRepresentation LearningGraph Neural NetworkAuto EncoderImage
π― What it does: Propose a causality-intervention-based quantized variational autoencoder (CT-VAE) to achieve interpretable disentanglement of image causal factors and single-factor interventions.
Distilling Universal and Joint Knowledge for Cross-Domain Model Compression on Time Series Data
Qing Xu (Institute for Infocomm Research, A*STAR), Zhenghua Chen (Institute for Infocomm Research, A*STAR)
CodeCompressionDomain AdaptationKnowledge DistillationGenerative Adversarial NetworkTime Series
π― What it does: Propose an end-to-end cross-domain knowledge distillation framework called UNI-KD for compressing deep models in time series tasks, enabling knowledge transfer between source and target domains.
Chunxu Zhang (Key Laboratory of Symbolic Computation and Knowledge Engineering of Ministry of Education), Bo Yang (Key Laboratory of Symbolic Computation and Knowledge Engineering of Ministry of Education)
π― What it does: Proposed the PFedRec framework to learn lightweight, edge-deployable personalized models in federated recommendation environments, achieving fine-grained personalization for scoring functions and item embeddings through a dual personalization mechanism.
π― What it does: Propose a hybrid differentiable rendering method to simultaneously reconstruct triangle mesh geometry and physically-based PBR materials from real images captured with handheld cameras from multiple perspectives under uncontrolled lighting conditions.
π― What it does: Propose a multi-stage data filtering method for NLP model fine-tuning, which dynamically skips unimportant training samples during training based on real-time training loss, significantly reducing forward and backward computational costs.
π― What it does: This paper proposes a lightweight index structure called Grid Tree for efficiently performing nearest neighbor search of text keywords in dynamic game maps.
π― What it does: Proposed a completely non-autoregressive decoder (CND) based on curriculum learning, achieving generation of sign language translation results in a single decoding step, significantly reducing inference latency.
π― What it does: Proposed a new general arc consistency (GAC) algorithm for all-different constraints that completely eliminates the reliance on strong connected component (SCC) computation in traditional algorithms;
π― What it does: Propose the ABL-KG framework, which can automatically mine and memorize logical rules from large-scale knowledge graphs without manual annotation of knowledge bases, and then use these rules in abductive learning (ABL) to improve machine learning models.
π― What it does: Propose a dynamic structure development spiking neural network (DSD-SNN) that achieves continual learning for multi-task scenarios through random growth of new neurons, adaptive pruning, and freezing mechanisms.
Ensemble Reinforcement Learning in Continuous Spaces -- A Hierarchical Multi-Step Approach for Policy Training
Gang Chen (Victoria University of Wellington), Victoria Huang (National Institute of Water and Atmospheric Research)
CodeReinforcement Learning
π― What it does: This paper proposes a hierarchical multi-step ensemble deep deterministic policy gradient algorithm (HED) for reinforcement learning in continuous action spaces;
Explainable Multi-Agent Reinforcement Learning for Temporal Queries
Kayla Boggess (University of Virginia), Lu Feng (University of Virginia)
CodeExplainability and InterpretabilityReinforcement LearningTime SeriesSequential
π― What it does: This paper proposes a method that generates strategy-level contrastive explanations for time series queries using an abstract model of multi-agent reinforcement learning (MARL) policies, and determines whether the query is feasible; if not feasible, it generates complete and correct explanations to clarify the reasons for failure.
Explainable Reinforcement Learning via a Causal World Model
Zhongwei Yu (Institute of Automation, Chinese Academy of Sciences), Dengpeng Xing (Institute of Automation, Chinese Academy of Sciences)
CodeExplainability and InterpretabilityRecurrent Neural NetworkReinforcement LearningWorld Model
π― What it does: Propose an interpretable world model in reinforcement learning without prior causal structure, which learns environment dynamics through causal discovery and an attention inference network, and generates interpretable explanations for action decisions via causal chains.
Exploiting Non-Interactive Exercises in Cognitive Diagnosis
Fangzhou Yao (University of Science and Technology of China), Shijin Wang (University of Science and Technology of China)
CodeData-Centric LearningTabularSequential
π― What it does: In cognitive diagnosis, the authors propose an EIRS (Exercise-aware Informative Response Sampling) framework, which alleviates the long-tail problem caused by sparse student interactions by leveraging the potential sorting information of uninteracted exercises.
Exploring Leximin Principle for Fair Core-Selecting Combinatorial Auctions: Payment Rule Design and Implementation
Hao Cheng (Nanjing University), Chongjun Wang (Nanjing University)
CodeOptimizationTabularBenchmark
π― What it does: Studied core selection in combinatorial auctions, proposed a fair payment rule based on the Leximin principle (BLO), and provided an implementation algorithm.
Faster Exact MPE and Constrained Optimization with Deterministic Finite State Automata
Filippo Bistaffa (Artificial Intelligence Institute, Spanish National Research Council)
CodeOptimizationComputational EfficiencyBenchmark
π― What it does: Propose a function representation based on Deterministic Acyclic Finite State Automaton (DAFSA) for performing exact Most Probable Explanation (MPE) and constraint optimization tasks in Bucket Elimination (BE), forming the FABE algorithm.
Feature Staleness Aware Incremental Learning for CTR Prediction
Zhikai Wang (Shanghai Jiao Tong University), Kangyi Lin (Tencent)
CodeRecommendation SystemTabular
π― What it does: Propose a solution called FeSAIL for the feature obsolescence problem in incremental learning of CTR prediction models, which can adaptively replay samples with obsolete features and adjust feature embedding updates.
FedBFPT: An Efficient Federated Learning Framework for Bert Further Pre-training
Xin'ao Wang (Zhejiang University), Lidan Shou (Zhejiang University)
CodeFederated LearningComputational EfficiencyRepresentation LearningTransformerLarge Language ModelText
π― What it does: Propose the FEDBFPT framework, which utilizes federated learning to further pre-train only partial layers of the BERT model on clients, and progressively advances layer by layer through the PL-SDL strategy, ultimately generating a global BERT model applicable to downstream tasks.
π― What it does: This paper proposes the FedNoRo two-phase federated learning framework to address class imbalance and label noise heterogeneity in medical image data.
π― What it does: Propose FEDOBD, which utilizes semantic block-level importance for adaptive block dropping combined with NNADQ quantization to efficiently train large neural networks in federated learning.
FedSampling: A Better Sampling Strategy for Federated Learning
Tao Qi (Tsinghua University), Xing Xie (Microsoft Research Asia)
CodeFederated LearningImageText
π― What it does: Propose FedSampling, which performs uniform sampling of each local sample in federated learning and estimates the global total sample count through a local differential privacy mechanism, thereby achieving data utilization at the data level comparable to centralized training.
π― What it does: Propose an Ξ΅-close minimum change parameter tuning algorithm based on regional verification for multi-parameter, multi-CPT Bayesian networks (pBN), minimizing parameter changes while satisfying given quantitative constraints.
π― What it does: This paper generates supervised sentence representations by fully fine-tuning BERT and performing Prompt-tuning on 10 NLU tasks, then maps them to fMRI brain activation data using a regression neural decoding method to evaluate their decoding performance in the brain language network.
CodeOptimizationExplainability and InterpretabilitySequentialBenchmark
π― What it does: Studied how to generate formal, interpretable decision sequence explanations for deep learning-based planning strategies, proposing a linearly decomposable algorithm.
G2Pxy: Generative Open-Set Node Classification on Graphs with Proxy Unknowns
Qin Zhang (Shenzhen University), Shirui Pan (Griffith University)
CodeClassificationGraph Neural NetworkGraph
π― What it does: Proposes a generative open-set node classification method called G2Pxy, which achieves identification of unknown classes and classification of known classes by generating proxy unknown nodes in graph neural networks.
Gapformer: Graph Transformer with Graph Pooling for Node Classification
Chuang Liu (Wuhan University), Wenbin Hu (Wuhan University)
CodeClassificationTransformerGraph
π― What it does: Propose Gapformer, a node classification model that combines graph Transformer with graph pooling, reducing computational complexity and irrelevant information by first pooling the graph before computing attention.
π― What it does: This paper proposes using the flatness of the local loss surface as a proxy metric for NAS search, constructing the GeNAS framework;
Generalization Guarantees of Self-Training of Halfspaces under Label Noise Corruption
Lies Hadjadj (Universit e Grenoble Alpes), Sana Louhichi (Universit e Grenoble Alpes)
CodeClassificationImageText
π― What it does: Proposed a self-training algorithm that iteratively learns from labeled and unlabeled data using half-spaces (linear threshold functions), combining two steps: exploration (searching for half-spaces with large cosine distances and pseudo-labeling) and pruning (removing low-confidence samples), ultimately obtaining a series of half-space predictors.
Generalized Discriminative Deep Non-Negative Matrix Factorization Based on Latent Feature and Basis Learning
Zijian Yang (ShanghaiTech University), Lu Sun (ShanghaiTech University)
CodeClassificationRepresentation LearningImage
π― What it does: Proposed a Generalized Deep Non-Negative Matrix Factorization (GDNMF) that simultaneously performs deep decomposition on features and bases, integrates shallow linear and deep nonlinear models, and implements a semi-supervised version.
Genetic Prompt Search via Exploiting Language Model Probabilities
Jiangjiang Zhao (Beijing University of Posts and Telecommunications), Fangchun Yang (Beijing University of Posts and Telecommunications)
CodeOptimizationTransformerLarge Language ModelPrompt EngineeringText
π― What it does: This paper proposes a genetic algorithm called GAP3, guided by language model probabilities, for automatically searching discrete prompts on black-box pre-trained language models.
π― What it does: Proposes GCFGAE, a graph autoencoder combining federated learning and split learning for unsupervised federated graph learning, addressing accuracy degradation caused by non-IID graphs.