Hybrid CNN-Transformer Feature Fusion for Single Image Deraining
Xiang Chen (Nanjing University of Science and Technology), Hao Li (Shenyang Aerospace University)
CodeRestorationConvolutional Neural NetworkTransformerMixture of ExpertsImage
π― What it does: A lightweight hybrid CNN-Transformer feature fusion network HCT-FFN is proposed, achieving single image deraining in a staged progressive manner.
π― What it does: A lightweight image super-resolution network called Hybrid Pixel-Unshuffled Network (HPUN) is proposed, which enhances feature representation through self-residual depthwise separable convolutions and pixel-unshuffle downsampling.
HybridCap: Inertia-Aid Monocular Capture of Challenging Human Motions
Han Liang (ShanghaiTech University), Lan Xu (ShanghaiTech University)
CodePose EstimationOptimizationVideoMultimodality
π― What it does: This paper proposes HybridCap, which achieves real-time high-quality 3D capture of challenging human actions in a lightweight setup with a single camera and four IMUs.
π― What it does: Proposes the HybridPrompt framework, which combines language models with human priors to fine-tune visual question answering models through a mix of cloze and verify-style prompts;
π― What it does: This paper proposes HyperJump, a hyperparameter optimization method that accelerates the process through risk modeling based on HyperBand.
π― What it does: This paper proposes a three-way contrastive learning framework, TriCL, for unsupervised hypergraph representation learning, which can simultaneously capture structural information of nodes, hyperedges, and node-hyperedge relationships.
π― What it does: This paper studies the task of event causal relationship identification and proposes the KADE framework, which utilizes dual enhancement through external knowledge and internal analogy.
π― What it does: This paper proposes an Inverse Kinematics Optimization Layer (IKOL) that combines optimization and regression to achieve 3D human pose and shape estimation within an end-to-end framework.
π― What it does: This paper reveals through experiments and analysis that ImageNet pre-training transfers non-robust features to downstream models in transfer learning, resulting in poor performance of downstream models under adversarial attacks.
π― What it does: This paper studies the imbalance problem in label distribution learning and proposes a model that aligns the distribution of feature representations and label representations to address the distribution shift between the training set and the test set.
Imperceptible Adversarial Attack via Invertible Neural Networks
Zihan Chen (National University of Defense Technology), Dejian Guan (National University of Defense Technology)
CodeAdversarial AttackFlow-based ModelImage
π― What it does: A framework for adversarial attacks called AdvINN is proposed, which utilizes Invertible Neural Networks (INN) to generate adversarial samples with a high attack success rate without significantly altering the visual quality of the images.
Improved Algorithm for Regret Ratio Minimization in Multi-Objective Submodular Maximization
Yanhao Wang (East China Normal University), Fanxu Meng (Nanjing University)
CodeOptimizationGraph
π― What it does: This paper studies the return ratio minimization (RRM) problem in multi-objective submodular maximization and proposes the HS-RRM algorithm based on Ξ΅-kernel, Ξ΄-net, and HITTINGSET transformation.
Improved Kernel Alignment Regret Bound for Online Kernel Learning
Junfan Li (Tianjin University), Shizhong Liao (Tianjin University)
CodeOptimizationTabular
π― What it does: A new online kernel learning algorithm POMDR is proposed, which improves the kernel alignment regret bound under hinge loss and reduces computational complexity.
Improving Biomedical Entity Linking with Cross-Entity Interaction
Zhenran Xu (Harbin Institute of Technology), Baotian Hu (Harbin Institute of Technology)
CodeDrug DiscoveryTransformerPrompt EngineeringBiomedical Data
π― What it does: This paper proposes a cross-entity interactive re-ranking model based on prompt tuning to address the ambiguity problem in biomedical entity linking.
Improving Distantly Supervised Relation Extraction by Natural Language Inference
Kang Zhou (Iowa State University), Qi Li (Iowa State University)
CodeTransformerLarge Language ModelText
π― What it does: This paper proposes a relationship extraction framework that combines distant supervision and natural language inference (DSRE-NLI). It automatically mines and filters templates through semi-automatic relation representation (SARV), significantly improving data quality and reducing labor costs.
Improving Dynamic HDR Imaging with Fusion Transformer
Rufeng Chen (Hangzhou Dianzi University), Shanxin Yuan (Queen Mary University of London)
CodeRestorationTransformerImage
π― What it does: This paper proposes a Transformer-based HDR fusion framework called HFT, which achieves end-to-end reconstruction from multiple exposed LDR images to HDR images.
Improving Long-Horizon Imitation through Instruction Prediction
Joey Hejna (Stanford University), Lerrel Pinto (New York University)
CodeRobotic IntelligenceReinforcement Learning from Human FeedbackTransformerSupervised Fine-TuningReinforcement LearningText
π― What it does: By having agents additionally predict task instructions during the imitation learning process, their performance in long-term planning tasks is improved.
π― What it does: A pluggable Dual Prior Modulation Network (DPMN) is proposed, which utilizes two types of image-level priorsβtext masks and graphic recognition resultsβthrough a dual-branch design to progressively improve and integrate super-resolution results, thereby enhancing the clarity of scene text images and downstream recognition performance.
π― What it does: This paper studies the improvement of the quality of synchronous machine translation (SiMT) models by first generating pseudo-targets from monolingual data using sequence-level knowledge distillation, and then merging them with bilingual data for training. It also proposes a monolingual sampling strategy for SiMT that considers chunk length and monotonicity to reduce hallucinations and enhance robustness.
Improving the Cross-Lingual Generalisation in Visual Question Answering
Farhad Nooralahzadeh (University of Zurich), Rico Sennrich (University of Zurich)
CodeTransformerSupervised Fine-TuningVision Language ModelMultimodalityBenchmark
π― What it does: Improving the performance of pre-trained multilingual audiovisual models in cross-lingual visual question answering tasks, three strategies are proposed: language prior loss, sparse fine-tuning (SFT), and code mixing (CDM), and evaluated on the xGQA benchmark.
π― What it does: The paper proposes the Neighborhood Conformal Prediction (NCP) algorithm to improve the uncertainty quantification of deep classifiers.
Inconsistent Cores for ASP: The Perks and Perils of Non-monotonicity
Johannes K. Fichte (TU Wien), Stefan Szeider (TU Wien)
CodeGraph
π― What it does: This paper studies the concept of Inconsistent Core (IC) in Answer Set Programming (ASP), systematically analyzes the complexity of finding IC, Minimal Inconsistent Core (MIC), and Minimal Strongly Minimal Inconsistent Core (SMIC), and provides an implementation based on ASP-Core-2.
π― What it does: This paper proposes the Incremental Reinforcement Learning (Incremental RL) problem, where the state and action space of the environment continuously expand during the training process. It designs a Dual-Adaptive Ξ΅-greedy Exploration algorithm (DAE) that dynamically adjusts the Ξ΅ value through a Meta Policy and estimates the least tried actions to achieve efficient exploration. Additionally, two benchmark environments, Expanding World and Incremental Atari, are released to validate the method.
π― What it does: This paper studies an incremental few-shot object detection method called Incremental-DETR, which achieves incremental learning under the condition of having only a few new class samples and no base class samples.
π― What it does: A labeled speech corpus of 1,684 hours in 12 Hindi languages (Kathbath) was constructed, and based on this, a multi-task SLU evaluation benchmark (IndicSUPERB) was released, covering ASR, speaker verification/recognition, language recognition, query retrieval, and keyword spotting.
π― What it does: A cross-language topic model based on information maximization, InfoCTM, is proposed, which can automatically discover aligned cross-language topics and generate high-quality topic distributions.
Infusing Definiteness into Randomness: Rethinking Composition Styles for Deep Image Matting
Zixuan Ye (Huazhong University of Science and Technology), Hao Lu (Huazhong University of Science and Technology)
CodeSegmentationData SynthesisImage
π― What it does: This study investigates the training data generation process for depth image matting and proposes a new reasonable foreground combination operator RCF and ternary/quaternary combination styles to enhance the diversity of training samples and model generalization.
π― What it does: By constructing a dynamic memory buffer to retrieve instances similar in semantic meaning to sentences, and using self-attention aggregation to obtain smoothed positive samples, the generalization ability of sentence representations is enhanced within an unsupervised contrastive learning framework.
π― What it does: This paper proposes a Global Convolutional Attention block (GCA) and Intensity-Aware Loss (IAL) to enhance the performance of dynamic facial expression recognition in wild videos.
Interpreting Unfairness in Graph Neural Networks via Training Node Attribution
Yushun Dong (University of Virginia), Jundong Li (University of Georgia)
CodeClassificationExplainability and InterpretabilityComputational EfficiencyGraph Neural NetworkGraph
π― What it does: This paper studies the fairness bias that arises in node classification tasks using Graph Neural Networks (GNNs) and proposes the BIND framework, which combines the influence measurement of training nodes with Probabilistic Distribution Disparity (PDD) to explain and quantify model bias.
Intersection Coordination with Priority-Based Search for Autonomous Vehicles
Jiaoyang Li (Carnegie Mellon University), Sven Koenig (University of Southern California)
CodeAutonomous DrivingOptimization
π― What it does: This paper proposes a three-layer algorithm PSL (Priority-Based Search + Safe Interval Path Planning + Linear Programming) to address the coordination problem of unmanned vehicles at signal-less intersections, aiming to minimize the total time for vehicles to leave the intersection.
Interventional SHAP Values and Interaction Values for Piecewise Linear Regression Trees
Artjom Zern (SCHUFA Holding AG), Gjergji Kasneci (University of Tuebingen)
CodeExplainability and InterpretabilityComputational EfficiencyTabular
π― What it does: This paper proposes an efficient exclusive SHAP computation method, extending the TreeSHAP algorithm to piecewise linear regression trees and implementing background data aggregation.
Intriguing Findings of Frequency Selection for Image Deblurring
Xintian Mao (East China Normal University), Yan Wang (East China Normal University)
CodeRestorationConvolutional Neural NetworkImage
π― What it does: A residual block for frequency-selective ReLU in the frequency domain (Res FFT-ReLU Block) is proposed, which is integrated into existing deblurring networks to fuse frequency domain and spatial domain features for restoring clear images.
π― What it does: A non-replay phase-isolated incremental learning framework ESN is designed, utilizing a pre-trained ViT backbone and independent classifiers, and achieving cross-phase knowledge aggregation through energy self-normalization and voting inference, eliminating phase interference and performance imbalance.
π― What it does: The IterDE framework is proposed, which gradually compresses knowledge graph embedding models through iterative knowledge distillation, maintaining high performance at low dimensions.
Joint Multimodal Entity-Relation Extraction Based on Edge-Enhanced Graph Alignment Network and Word-Pair Relation Tagging
Li Yuan (South China University of Technology), Qing Li (Hong Kong Polytechnic University)
CodeRecognitionObject DetectionGraph Neural NetworkTransformerVision Language ModelTextMultimodality
π― What it does: This paper proposes a joint task of multimodal named entity recognition and relation extraction, called JMERE, and designs a word pair relation labeling and edge-enhanced graph alignment network to achieve joint extraction.
KerPrint: Local-Global Knowledge Graph Enhanced Diagnosis Prediction for Retrospective and Prospective Interpretations
Kai Yang (Zhongguancun Laboratory), Bing Xie (Peking University)
CodeClassificationExplainability and InterpretabilityGraph Neural NetworkTransformerSupervised Fine-TuningTabularBiomedical DataElectronic Health Records
π― What it does: The KerPrint model is proposed, which combines local and global knowledge graphs, and achieves retrospective and prospective explanations for diagnostic predictions through a time-aware attention mechanism and element-level attention.
π― What it does: This paper proposes an auxiliary OOD data generation method called KIRBY, based on key feature replacement and patching, to train a rejection network for OOD detection without modifying the original classifier.
π― What it does: Knowledge fusion for multi-label classification (MLC) problems: In the absence of labeled data, multiple pre-trained teacher models (each covering different sets of labels) are fused into a single student model, enabling the student to predict the union of all teacher label sets.
Changyi Xiao (University of Science and Technology of China), Yixin Cao (Singapore Management University)
CodeFlow-based ModelGraphBenchmark
π― What it does: This paper proposes embedding knowledge graph entities/relations into the permutation group and introduces uncertainty through normalizing flow, forming a new knowledge graph embedding model called NFE.
Knowledge-Bridged Causal Interaction Network for Causal Emotion Entailment
Weixiang Zhao (Harbin Institute of Technology), Bing Qin (Harbin Institute of Technology)
CodeTransformerText
π― What it does: Proposes the Knowledge-Bridged Causal Interaction Network (KBCIN) to identify the causal sentences that trigger non-neutral emotions in dialogues.
π― What it does: A KT-Net based on a teacher-assistant-student network is proposed, utilizing knowledge transfer to achieve unpaired 3D point cloud completion.
π― What it does: To address the long-tail label problem in multi-label text classification, this paper proposes a dual-layer framework based on Label-Specific Feature Augmentation (LSFA), which first learns decoupled label-specific representations and then generates positive samples for tail labels in the feature space using a prototype-driven VAE.
π― What it does: A layout generation method based on action sequences is proposed, which directly predicts actions such as copy and margin instead of border coordinates;
Layout-Aware Dreamer for Embodied Visual Referring Expression Grounding
Mingxiao Li (KU Leuven), Marie-Francine Moens (KU Leuven)
CodeRobotic IntelligenceTransformerReinforcement LearningVision Language ModelImageText
π― What it does: Designed and implemented an agent named Layout-aware Dreamer (LAD) that can navigate unknown indoor environments and locate remote targets by utilizing layout learning and goal imagination modules upon receiving high-level natural language instructions.
π― What it does: In the supervised limited text-to-SQL task flow, two approaches that combine semi-supervised learning (SSL) with continual learning (CL) are proposed, namely VANILLA and the improved SFNET.
Learning by Applying: A General Framework for Mathematical Reasoning via Enhancing Explicit Knowledge Learning
Jiayu Liu (University of Science and Technology of China), Qi Liu (University of Illinois at Urbana-Champaign)
CodeGraph Neural NetworkAuto EncoderText
π― What it does: A general learning-application framework (LeAp) is proposed, which allows existing mathematical word problem (MWP) solvers to learn and utilize the relationships between words, words and operators through an explicit knowledge graph, thereby enhancing reasoning capabilities.
π― What it does: A multi-task network architecture search method based on conflict detection, CoNAL, is proposed, which automatically switches between shared modules and purely exclusive modules to alleviate gradient conflicts, and incorporates conflict detection operations during the search process.
π― What it does: This study proposes a Context-Aware Classifier (CAC) that dynamically generates classifier weights for each image to adapt to the potential distribution of different images, thereby improving semantic segmentation accuracy.
π― What it does: This paper proposes a continuous depth representation-based RGB-guided depth super-resolution network, GeoDSR, which can perform super-resolution on low-resolution depth maps at any scaling factor and any shape.
π― What it does: This paper proposes a framework that simultaneously learns control strategies and formal reach-avoid guarantees, utilizing neural networks to learn controllers and generate verifiable Reach-Avoid Stochastic Markov (RASM) certificates, completing a closed-loop process from learning to verification.
Learning Deep Hierarchical Features with Spatial Regularization for One-Class Facial Expression Recognition
Bingjun Luo (Tsinghua University), Yue Gao (Tsinghua University)
CodeRecognitionAnomaly DetectionAuto EncoderImage
π― What it does: This paper proposes HS-OCFER, a single-class facial expression recognition network trained using only a single normal expression category, aimed at detecting anomalous expressions that were not present during training.
Learning Dynamic Latent Spaces for Lifelong Generative Modelling
Fei Ye (University of York), Adrian G. Bors (University of York)
CodeGenerationData SynthesisAuto EncoderImage
π― What it does: This paper studies a generative model suitable for lifelong learning without task labelsβOnline Recursive Variational Autoencoder (ORVAE), which achieves knowledge transfer and prevents forgetting through recursive expansion and attention mechanisms.
π― What it does: This paper proposes a gradient descent-based iterative function system (IFS) parameter learning method that can generate visually similar fractal images for any target image (not necessarily fractal images).
π― What it does: A noise detector based on LSTM is trained using the probability sequences during the training process to identify mislabeled samples in the training set.
Learning Instrumental Variable from Data Fusion for Treatment Effect Estimation
Anpeng Wu (Zhejiang University), Fei Wu (Shanghai AI Laboratory)
CodeRepresentation LearningTabular
π― What it does: The Meta-EM algorithm is proposed, which automatically reconstructs source labels through representation learning and EM iteration in multi-source data fusion, using group instrumental variables (GIV) to achieve causal effect estimation without prior instrumental variables.
π― What it does: This paper proposes a method to automatically discover and utilize constraints from background knowledge (BK) to limit the hypothesis space of Inductive Logic Programming (ILP), thereby 'discovering areas that do not need to be searched' before the search, significantly accelerating the learning process.
Learning Markov Random Fields for Combinatorial Structures via Sampling through LovΓ‘sz Local Lemma
Nan Jiang (Purdue University), Yexiang Xue (Purdue University)
CodeGenerationOptimizationGraphTabular
π― What it does: This paper proposes a differentiable sampler NELSON based on the LovΓ‘sz Local Lemma (LLL) and embeds it into a Contrastive Divergence (CD) learning framework to train constrained Markov Random Fields (MRF), thereby achieving high-quality generation of combinatorial structures that satisfy hard constraints (such as K-SAT, acyclic directed graphs, and vehicle routing paths).
Learning Motion-Robust Remote Photoplethysmography through Arbitrary Resolution Videos
Jianwei Li (Xi'an Jiaotong University), Jingang Shi (Xi'an Jiaotong University)
CodeConvolutional Neural NetworkOptical FlowVideo
π― What it does: Two pluggable modules, PFE and TFA, are proposed to enhance the robustness of remote photoacoustic blood pressure measurement under arbitrary resolution video and head motion.
Moulik Choraria (University of Illinois at Urbana-Champaign), Lav R. Varshney (University of Illinois at Urbana-Champaign)
CodeClassificationDomain AdaptationText
π― What it does: This paper proposes a 'Partial Invariance' framework (P-IRM) in the context of concept drift, which relaxes the complete invariance constraint of traditional IRM by partitioning or conditioning the training environment, thereby enhancing the model's generalization ability in external environments.
π― What it does: Through self-learning cycles, programs are generated unsupervised from the integer sequences in OEIS using tree search and tree neural networks, ultimately discovering implementations for 27,987 sequences without any human annotations.
Learning Progressive Modality-Shared Transformers for Effective Visible-Infrared Person Re-identification
Hu Lu (Jiangsu University), Pingping Zhang (Dalian University of Technology)
CodeRecognitionRetrievalTransformerImage
π― What it does: A progressive modality sharing framework PMT based on Transformer is proposed and implemented, using grayscale images as auxiliary for progressive learning, combined with modality sharing enhanced loss and discriminative center loss to improve visible-infrared person re-identification performance.
Learning Relational Causal Models with Cycles through Relational Acyclification
Ragib Ahsan (University of Illinois at Chicago), Elena Zheleva (University of Illinois at Chicago)
CodeGraphTabular
π― What it does: This paper proposes the 'relation acyclic operation' and proves that under this condition, the existing relational causal discovery algorithm RCD can still maintain measurability and completeness in the presence of cyclic relational causal models, along with experimental validation.
π― What it does: This paper proposes the concept of a Bi-level Knowledge Graph (Bi-level KG), constructing a novel knowledge graph that includes base-level triples and hierarchical triples, and introduces the BiVE model for embedding learning on this graph; it also designs a random walk-based incremental data augmentation strategy; and proposes two new tasks: Triplet Prediction and Conditional Link Prediction.
π― What it does: A safe action model learning algorithm for numerical planning, N-SAM, is proposed, which can learn action models that comply with safety constraints based only on observed trajectories and can be used for planning.
Learning Second-Order Attentive Context for Efficient Correspondence Pruning
Xinyi Ye (Huazhong University of Science and Technology), Zhiguo Cao (Huazhong University of Science and Technology)
CodeComputational EfficiencyTransformerImage
π― What it does: A correspondence pruning method based on second-order attention (Attention in Attention) is proposed, which achieves efficient and accurate outlier removal by integrating first-order feature-consistent context and second-order attention-consistent context.
Learning Semantic Degradation-Aware Guidance for Recognition-Driven Unsupervised Low-Light Image Enhancement
Naishan Zheng (University of Science and Technology of China), Feng Zhao (University of Science and Technology of China)
CodeClassificationObject DetectionImage
π― What it does: This paper proposes a semantic degradation-aware guidance (SDAG) method for unsupervised low-light image enhancement, which learns the impact of low-light degradation on semantic features through self-supervised reconstruction and embeds it into the existing ULLIE network to improve downstream visual recognition performance.
Learning Similarity Metrics for Volumetric Simulations with Multiscale CNNs
Georg Kohl (Technical University of Munich), Nils Thuerey (Technical University of Munich)
CodeOptimizationConvolutional Neural NetworkMultimodalityPhysics Related
π― What it does: A similarity model based on entropy is proposed, and a multi-scale convolutional network VolSiM is trained to evaluate the similarity of three-dimensional physical simulation data.
Learning Single Image Defocus Deblurring with Misaligned Training Pairs
Yu Li (Harbin Institute of Technology), Wangmeng Zuo (Peng Cheng Laboratory)
CodeRestorationOptical FlowImage
π― What it does: This paper proposes a Joint De-blurring and Re-blurring Learning framework (JDRL) that addresses the deformation artifacts caused by training pairing errors in the single image defocus deblurring task through optical flow-based deformation and a spatially invariant re-blurring module.
CodeRestorationSegmentationConvolutional Neural NetworkSpiking Neural NetworkSupervised Fine-TuningTime Series
π― What it does: A waveform-guided spike enhancement (WGSE) module based on discrete wavelet transform (DWT) and lightweight CNN is proposed to learn and generate more effective spike flow representations.
Learning the Finer Things: Bayesian Structure Learning at the Instantiation Level
Chase Yakaboski (Dartmouth College), Jr
CodeTabularBiomedical Data
π― What it does: Proposes an instance hierarchy learning method based on Bayesian Knowledge Base (BKB), constructing the BKBSL algorithm using minimum entropy inference and MDL scoring;
Learning to Break Symmetries for Efficient Optimization in Answer Set Programming
Alice Tarzariol (University of Klagenfurt), Mark Law (ILASP Limited)
CodeOptimizationTabular
π― What it does: A new method is proposed to enhance the symmetry breaking constraints in Answer Set Programming (ASP) for optimization problems through Inductive Logic Programming (ILP), aiming to improve optimization efficiency.
Patrick Hemmer (Karlsruhe Institute of Technology), Niklas KΓΌhl (Karlsruhe Institute of Technology)
CodeClassificationData-Centric LearningImageBiomedical Data
π― What it does: This paper proposes a three-step method that utilizes a small number of expert predictions and semi-supervised learning to generate artificial expert predictions, supporting the learning-to-defer algorithm to work efficiently even when expert labels are scarce.
π― What it does: A framework for generating unbiased scene graphs is proposed, utilizing attribute-guided predicate features to generate balanced data and enhance the fairness of predicate classification.
Learning to Imagine: Distillation-Based Interactive Context Exploitation for Dialogue State Tracking
Jinyu Guo (Beijing University of Posts and Telecommunications), Zihan Wang (University of Tokyo)
CodeKnowledge DistillationTransformerLarge Language ModelText
π― What it does: A general module DICE-DST is proposed, which utilizes a teacher encoder to learn dialogue context knowledge and enhances the encoder of historical DST models through transcription attention alignment distillation, addressing the issue of missing key context when only part of the dialogue history is used.
Learning to Know Myself: A Coarse-to-Fine Persona-Aware Training Framework for Personalized Dialogue Generation
Yunpeng Li (Institute of Information Engineering), Wei Peng (Institute of Information Engineering)
CodeGenerationTransformerLarge Language ModelContrastive LearningText
π― What it does: A two-stage coarse-to-fine persona-aware training framework is proposed, which first enhances the model's sensitivity to persona by constructing persona-aware questions, and then improves fine-grained consistency through contrastive learning to generate negative samples.
Learning to Learn Better for Video Object Segmentation
Meng Lan (Wuhan University), Dacheng Tao (JD Explore Academy)
CodeObject TrackingSegmentationTransformerVideo
π― What it does: A new joint learning framework LLB is proposed, combining target transfer and matching two branches, and a Discriminative Label Generation Module (DLGM) and an Adaptive Fusion Module (AFM) are designed to improve target feature representation.
π― What it does: A method is proposed for generating character-consistent dialogues by utilizing memory entailment relationships and discourse-level relationships.
Learning to Select Pivotal Samples for Meta Re-weighting
Yinjun Wu (University of Pennsylvania), Mayur Naik (University of Pennsylvania)
CodeClassificationMeta LearningImage
π― What it does: A learning framework is proposed to automatically select high-quality meta samples (key samples) from a large-scale defect-laden training set to enhance the performance of the meta reweighting algorithm.
Learning to Select Prototypical Parts for Interpretable Sequential Data Modeling
Yifei Zhang (Institute of Information Engineering Chinese Academy of Sciences), Cunqing Ma (Institute of Information Engineering Chinese Academy of Sciences)
CodeClassificationExplainability and InterpretabilityConvolutional Neural NetworkRecurrent Neural NetworkTextSequentialBiomedical DataElectrocardiogram
π― What it does: This paper proposes a Self-Explaining Selection Model (SESM), which encodes each subsequence as a concept by using a learnable subsequence selection mechanism on the sequence, and then aggregates these concepts with linear weighting to make predictions, thereby achieving interpretability for sequence models.
Learning Topology-Specific Experts for Molecular Property Prediction
Suyeon Kim (Pohang University of Science and Technology), Hwanjo Yu (Pohang University of Science and Technology)
CodeClassificationDrug DiscoveryGraph Neural NetworkMixture of ExpertsGraphTabular
π― What it does: We propose TopExpert, which utilizes a clustering gating module to divide molecules into several expert groups based on topological similarity, allowing each expert to learn features within the corresponding topological subset for molecular property prediction.
π― What it does: Transforming partial label learning problems into semi-supervised learning tasks, using high-confidence pseudo-labels and consistency regularization to train the model.
π― What it does: A lightweight learnable noise module, LeNo, has been designed and implemented to enhance the robustness of salient object detection models against adversarial attacks while maintaining high accuracy on clean images.
π― What it does: This study investigates node injection attacks on black-box graph neural networks and proposes a gradient-free reinforcement learning attack framework called G2A2C.
Leveraging Structure for Improved Classification of Grouped Biased Data
Daniel Zeiberg (Northeastern University), Predrag Radivojac (Northeastern University)
CodeClassificationTabular
π― What it does: This paper studies the semi-supervised binary classification problem under the condition of naturally grouped data and biased labeled data, proposing an algorithm that utilizes data structure to improve classification performance.
π― What it does: A two-stage method is proposed for zero-shot learning through subclass discrimination, which first aligns synthetic embeddings using contrastive learning at the feature level, and then implements dynamic prototype updates using prototype modulation at the classifier level to enhance the recognition of unseen attribute-object combinations.
Lifelong Compression Mixture Model via Knowledge Relationship Graph
Fei Ye (University of York), Adrian G. Bors (University of York)
CodeCompressionOptimizationAuto EncoderImage
π― What it does: This paper proposes a lifelong generative mixture model (LGMM) based on dynamic expansion and compression, addressing the issues of network forgetting and model size inflation in task-agnostic continual learning (TFCL).
π― What it does: This paper proposes a lifelong knowledge graph embedding model named LKGE, which can quickly learn new facts, transfer existing knowledge, and retain old knowledge as the knowledge graph continues to grow.
π― What it does: A lifecycle person re-identification method KRC is proposed, which combines knowledge replay, knowledge refresh, and knowledge consolidation in three stages, allowing the model to maintain old knowledge while improving the performance of both old and new tasks as new tasks are continuously added.
π― What it does: This paper proposes an Online Adversarial Expansion Strategy (OAES) that dynamically increases the capacity of the VAE model to mitigate catastrophic forgetting in the scenario of Task-Free Continual Learning (TFCL).
Lifting (D)QBF Preprocessing and Solving Techniques to (D)SSAT
Che Cheng (National Taiwan University), Jie-Hong R. Jiang (National Taiwan University)
CodeTabularBenchmark
π― What it does: The first DSSAT solver has been implemented (by transforming DSSAT into SSAT and using dependency elimination) and an independent DSSATpre preprocessor has been developed, which simplifies SSAT/DSSAT formulas based on HQSpre.
π― What it does: The study introduces a lifelong learning framework in Mixed Integer Programming (MIP) and proposes the LIMIP method to continuously learn and make branching decisions in multi-task sequences.