CodeClassificationAnomaly DetectionTransformerTime SeriesSequentialBiomedical DataElectronic Health RecordsFinance Related
π― What it does: The TCMBN model is proposed, which combines Transformer and Conditional Mixture Bernoulli Network to predict concurrent multi-labels in event streams, and learns a real-time label graph structure through least squares sparse precision matrix learning.
π― What it does: A sample confidence adaptive random smoothing training method (CAT-RS) is proposed, which enhances confidence calibration and adversarial robustness by dynamically controlling the robustness target of each sample.
Context-Aware Safe Medication Recommendations with Molecular Graph and DDI Graph Embedding
Qianyu Chen (Beijing Institute of Technology), Mingzhong Wang (University of the Sunshine Coast)
CodeRecommendation SystemDrug DiscoveryGraph Neural NetworkGraphBiomedical DataElectronic Health Records
π― What it does: A drug recommendation framework CARMEN based on context-aware graph neural networks is proposed, integrating patient history, molecular graphs, and drug-drug interaction graphs.
ConTextual Masked Auto-Encoder for Dense Passage Retrieval
Xing Wu (Institute of Information Engineering Chinese Academy of Sciences), Songlin Hu (Kuaishou Technology)
CodeRetrievalTransformerAuto EncoderText
π― What it does: A new generative pre-training method called CoT-MAE is proposed, which utilizes contextual information across text spans for self-supervised and context-supervised masked autoencoding to enhance the quality of text representations in dense retrieval.
Continual Learning with Scaled Gradient Projection
Gobinda Saha (Purdue University), Kaushik Roy (Purdue University)
CodeClassificationReinforcement LearningImage
π― What it does: A continuous learning method based on gradient projection is proposedβScaled Gradient Projection (SGP), which achieves a balance between learning new tasks and retaining old tasks by scaling gradients in the important gradient space of old tasks according to their importance.
π― What it does: This paper proposes an unsupervised, task-agnostic continuous generative model framework that utilizes short-term memory (STM) and long-term memory (teacher) to achieve continuous learning from infinite data streams through a knowledge incremental assimilation mechanism and continuous generative knowledge distillation.
π― What it does: A metric learning model based on contrastive attention is proposed to automatically identify the similarity of damaged prints in early modern printed books, thereby inferring the printer.
π― What it does: Proposes a Contrastive Identity-Aware (CIA) learning module that utilizes contrastive learning to enhance the distinguishability of credit allocation among different agents in the Value Decomposition (VD) network, thereby promoting diverse behaviors in multi-agent collaboration;
Contrastive Learning Reduces Hallucination in Conversations
Weiwei Sun (Shandong University), Zhaochun Ren (Shandong University)
CodeGenerationKnowledge DistillationTransformerLarge Language ModelContrastive LearningText
π― What it does: A hybrid strategy based on contrastive learning, MixCL, is proposed to reduce the hallucination generation of pre-trained language models in knowledge-driven dialogues.
π― What it does: This paper proposes a method for open set recognition based on supervised contrastive learning called ConOSR, which enhances feature quality using soft target contrastive learning and enables the detection of unknown samples.
π― What it does: This paper proposes a contrastive pre-training based checkpoint sequence representation learning method (CACSR), which generates more challenging positive and negative samples through spatial-temporal augmentation and adversarial perturbations in the latent space, thereby learning mobile trajectory representations with higher-level semantics.
Controlling Class Layout for Deep Ordinal Classification via Constrained Proxies Learning
Cong Wang (Nanjing University), Qing Gu (Nanjing University)
CodeClassificationAgentic AIImage
π― What it does: A framework for deep ordinal classification through constrained proxy layout is proposed, called Constrained Proxies Learning (CPL).
π― What it does: This paper proposes the ConvMatch network, which utilizes convolutional neural networks to learn the correspondence between two views for more accurate outlier rejection and geometric estimation.
π― What it does: This paper proposes a Conversational Neural Topic Model (ConvNTM) for dialogue scenarios, capturing multi-turn structures through hierarchical encoding and modeling the interactions between speakers and audiences using a multi-role graph network to achieve utterance-level topic distributions. Additionally, a word co-occurrence constraint is introduced as an auxiliary objective to enhance topic coherence.
π― What it does: A Cooperative and Adversarial Learning (CALE) framework is proposed, which achieves the mutual cooperation and confrontation of discriminative and transfer modules by generating easy and hard samples, unifying the optimization objectives of both.
π― What it does: By generating spatially adaptive parameters through downsampling high-resolution images, and then using implicit representation of coordinate queries (multi-layer perceptron) to reconstruct missing areas pixel by pixel, efficient high-resolution image inpainting is achieved.
π― What it does: This paper proposes an unsupervised factual consistency detection framework called CoP, which incorporates prompts in an additional reasoning step to control the preferences of the generative model, thereby calculating probability differences to identify factual inconsistencies in summaries.
Correlation Loss: Enforcing Correlation between Classification and Localization
Fehmi Kahraman (Middle East Technical University), Emre Akbas (Middle East Technical University)
CodeObject DetectionImage
π― What it does: A loss function for directly optimizing the correlation between classification and localization (Correlation Loss) is proposed and validated, and it is applied as a plugin to various object detectors based on NMS and NMS-free, significantly improving detection accuracy.
Corruption-Tolerant Algorithms for Generalized Linear Models
Bhaskar Mukhoty (Mohamed Bin Zayed University of Artificial Intelligence), Purushottam Kar (Indian Institute of Technology Kanpur)
CodeClassificationOptimizationTabular
π― What it does: The SVAM (Sequential Variance-Altered MLE) framework is proposed to uniformly address robust learning of generalized linear models (GLM) under adversarially tampered training data labels, covering least squares, logistic regression, gamma regression, and more.
Counterfactual Dynamics Forecasting β a New Setting of Quantitative Reasoning
Yanzhu Liu (Institute for Infocomm Research and Centre for Frontier AI Research, A*STAR), Joo-Hwee Lim (Institute for Infocomm Research and Centre for Frontier AI Research, A*STAR)
π― What it does: This study investigates the problem of counterfactual dynamics prediction, using deep neural networks to simulate multi-agent dynamic systems and predict future trajectories under different counterfactual interventions.
Coupling Artificial Neurons in BERT and Biological Neurons in the Human Brain
Xu Liu (Northwestern Polytechnical University), Xintao Hu
CodeTransformerLarge Language ModelTextTime SeriesMagnetic Resonance Imaging
π― What it does: A framework is proposed that couples fine-grained artificial neurons in the Transformer model (the query/key dimensions of each multi-head self-attention in BERT) with functional neural networks in the human brain (functional brain networks obtained through fMRI), establishing a correspondence between the two using temporal activation synchronization, and providing a neurolinguistic interpretation of the coupling results using semantic information such as part-of-speech tags.
Cross-Domain Adaptative Learning for Online Advertisement Customer Lifetime Value Prediction
Hongzu Su (University of Electronic Science and Technology of China), Ke Lu (Shandong Normal University)
CodeDomain AdaptationRecommendation SystemTabular
π― What it does: A Cross-Domain Adaptation Framework (CDAF) is proposed to alleviate the data scarcity problem in the target domain by leveraging rich data from the source domain, and to enhance the customer lifetime value (LTV) prediction on online advertising platforms using domain-invariant information.
π― What it does: A cross-domain graph anomaly detection method (ACT) is proposed, which achieves efficient anomaly detection on unlabeled target graphs through unsupervised contrastive learning on the target graph and joint optimization of anomaly-aware domain alignment with the representation of the target graph.
π― What it does: This study proposes the CrysGNN pre-training framework, which utilizes 800K unlabeled crystal graph structures to self-supervised learn localized chemical features and global crystal structures at the node and graph levels, respectively, and injects the pre-trained knowledge into existing material property prediction models through knowledge distillation.
π― What it does: By employing multi-level negative sample augmentation (cross-video clip/video-level and self-shuffling masking) combined with multi-stage curriculum learning, we suppress the video localization model's dependence on the distribution bias of temporal annotations and enhance cross-modal semantic matching capabilities.
π― What it does: This paper proposes a Circular Decoupled Feature Translation Network (CDFTN) that generates pseudo-labeled samples and trains a robust classifier by exchanging domain-invariant liveness features with domain-specific content features, thereby achieving cross-scene face anti-spoofing.
DACOM: Learning Delay-Aware Communication for Multi-Agent Reinforcement Learning
Tingting Yuan (University of Gttingen), Xiaoming Fu (Beijing University of Posts and Telecommunications)
CodeReinforcement Learning
π― What it does: This paper proposes a delay-aware communication model called DACOM, aimed at enhancing cooperative performance in multi-agent reinforcement learning by learning appropriate waiting times to balance communication benefits and delay costs.
π― What it does: An end-to-end DarkFeat model has been developed, capable of detecting and describing local features directly from extremely low-light RAW images.
π― What it does: This paper proposes a backward-compatible model upgrade framework called Darwinian Model Upgrades (DMU), which enables continuous model upgrades for large-scale retrieval systems without the need to recalculate old features (backfill).
Data-Efficient Image Quality Assessment with Attention-Panel Decoder
Guanyi Qin (Tsinghua University), Yan Zhang (Xiamen University)
CodeTransformerImage
π― What it does: A blind image quality assessment method based on a Transformer encoder-decoder (DEIQT) has been developed, which re-encodes the CLS token through the decoder and incorporates an attention panel to reduce prediction uncertainty.
DC-Former: Diverse and Compact Transformer for Person Re-identification
Wen Li (Ant Group), Wei Chu (Ant Group)
CodeRecognitionTransformerImage
π― What it does: Introduce multi-class Tokens in Vision Transformer and use self-heterogeneous constraints (SDC) to make the embedding subspaces corresponding to different Tokens orthogonal, resulting in diversified and compact feature representations, thereby improving person re-identification performance.
De-biased Teacher: Rethinking IoU Matching for Semi-supervised Object Detection
Kuo Wang (Sun Yat-sen University), Fan Zhou
CodeObject DetectionKnowledge DistillationImage
π― What it does: This paper proposes a semi-supervised object detection framework called De-biased Teacher, which eliminates the traditional IoU matching process and directly uses soft labels for consistency regularization to reduce training bias.
π― What it does: Decomposes the generative zero-shot learning framework and proposes a simplified method based on attribute generalization and bias correction.
π― What it does: For the inverse problem of video snapshot compression imaging (SCI), a reconstruction framework based on the Deep Equilibrium Model (DEQ) is proposed, which implicitly learns non-expansive operators and analytically solves fixed points, achieving infinite iterations, infinite network depth, while keeping memory consumption constant.
π― What it does: A self-supervised graph node representation learning framework DGSI based on information bottleneck is proposed, which can simultaneously capture structural and semantic information.
π― What it does: Proposed a single end-to-end network DP3DF, achieving super-resolution, denoising, and exposure enhancement for low-light noisy videos;
π― What it does: This study investigates the effectiveness of deep spiking neural networks (SNN) in modeling human and rodent visual cortices, comparing it with mainstream CNNs and ViTs.
π― What it does: A network named DENet is proposed, which separates watermarked and non-watermarked images through contrastive learning of high-order embeddings, and combines a self-attention enhancement module to achieve the removal of visible watermarks.
π― What it does: A training framework named DENEB is proposed to enhance the unbiased generalization performance of models in the presence of dataset bias and noisy labels.
π― What it does: This paper proposes CLQEβa denoising pre-training framework based on curriculum learning for machine translation quality estimation (QE), which enhances model performance by gradually introducing noisy pseudo QE data.
π― What it does: A deep learning architecture is proposed for accelerating the estimation of expected information gain (EIG) in Bayesian optimal experimental design, achieving 'design amortization' for all possible designs.
Detecting Sources of Healthcare Associated Infections
Hankyu Jang (University of Iowa), Sriram V. Pemmaraju (University of Iowa)
CodeOptimizationGraphBiomedical DataElectronic Health Records
π― What it does: In the hospital-acquired infection (HAI) transmission model, a new source detection method is proposed, utilizing a load sharing model to identify the source of infection.
DHGE: Dual-View Hyper-Relational Knowledge Graph Embedding for Link Prediction and Entity Typing
Haoran Luo (Beijing University of Posts and Telecommunications), Kaiyang Wan (Beijing University of Posts and Telecommunications)
CodeKnowledge DistillationRepresentation LearningGraph Neural NetworkGraphBiomedical Data
π― What it does: Proposes a dual-view hyper-relational knowledge graph embedding model DHGE, addressing the issue of traditional single-view hyper-relational KG embeddings neglecting hierarchical structures;
Chunlei Xin (Chinese Information Processing Laboratory, Institute of Software, Chinese Academy of Sciences), Le Sun (Xiaomi AI Lab)
CodeGenerationTransformerLarge Language ModelSupervised Fine-TuningTextBenchmark
π― What it does: This paper proposes a dialogue rewriting framework that can convert incomplete and noisy utterances in multi-turn dialogues into complete, context-independent natural language sentences.
π― What it does: This paper proposes a differentiable feature acquisition method called DiFA, which dynamically selects a small number of features for prediction tasks.
Bao Duong (Deakin University), Thin Nguyen (Deakin University)
CodeFlow-based ModelTabular
π― What it does: This paper proposes the DINE (Diffeomorphic Information Neural Estimator) method for accurately estimating the conditional mutual information (CMI) of continuous random variables, and further applies it to conditional independence testing.
π― What it does: This paper proposes a differentiable search-based heterogeneous information network graph neural network architecture search method called PMMM, which automatically learns multi-graph elements (meta-multigraph) and stable partial message propagation strategies.
Directed Acyclic Graph Structure Learning from Dynamic Graphs
Shaohua Fan (Beijing University of Posts and Telecommunications), Chuan Shi (Peng Cheng Laboratory)
CodeGraph Neural NetworkGraphTime Series
π― What it does: This paper proposes a directed acyclic graph (DAG) structure learning method for dynamic graphs called GraphNOTEARS, which aims to estimate the feature generation mechanisms both within the same time slice (intra-slice) and across time slices (inter-slice);
Discriminability and Transferability Estimation: A Bayesian Source Importance Estimation Approach for Multi-Source-Free Domain Adaptation
Zhongyi Han (Shandong University), Yilong Yin (Shandong University)
CodeDomain AdaptationImage
π― What it does: This paper proposes a source model importance estimation method based on Bayesian inference, called DATE, which can assign weights for multi-source free domain adaptation and improve performance without accessing the source data.
Disentangle and Remerge: Interventional Knowledge Distillation for Few-Shot Object Detection from a Conditional Causal Perspective
Jiangmeng Li (University of Chinese Academy of Sciences), Fuchun Sun (Tsinghua University)
CodeObject DetectionKnowledge DistillationImage
π― What it does: This paper introduces knowledge distillation in few-shot object detection and identifies and eliminates the negative impact of teacher errors on the student model through a structural causal model.
Disentangled Representation for Causal Mediation Analysis
Ziqi Xu (University of South Australia), Ke Wang (Simon Fraser University)
CodeRepresentation LearningAuto EncoderTabular
π― What it does: A new causal mediation analysis method called DMAVAE is proposed, aimed at accurately estimating natural direct effects, natural indirect effects, and total effects.
Benedict Wilkins (Royal Holloway University of London), Kostas Stathis (Royal Holloway University of London)
CodeReinforcement LearningImageSequential
π― What it does: A framework based on causal inference is proposed, which separates the spontaneous effects (reafference) and external effects (exafference) in agent perception by comparing 'doing nothing' (no action) and provides corresponding learning algorithms.
π― What it does: DisGUIDE proposes a data-independent model extraction framework that simultaneously trains two clone models and uses disagreement loss to generate query samples that better force the clone models to diverge, thereby improving extraction efficiency.
Dish-TS: A General Paradigm for Alleviating Distribution Shift in Time Series Forecasting
Wei Fan (University of Central Florida), Yanjie Fu (University of Central Florida)
CodeRecurrent Neural NetworkTransformerTime Series
π― What it does: This paper proposes a general neural network paradigm called Dish-TS, which uses dual coefficient networks (BACKCONET and HORICONET) to normalize and denormalize distribution shifts in time series forecasting, thereby enhancing the model's generalization performance.
Avraham Natan (Ben Gurion University of the Negev), Meir Kalech (Ben Gurion University of the Negev)
CodeSafty and PrivacyTabular
π― What it does: A Distributed Spectrum Fault Localization (DSFL) framework is proposed, and two distributed SFL algorithms (single fault DSFLA-SINGLE and multiple fault DSFLA-MULTI) are designed to achieve the same diagnostic results as centralized SFL while ensuring privacy protection.
π― What it does: This paper proposes a distributionally robust optimization framework based on probabilistic groups (PG-DRO) aimed at addressing the robustness issues of machine learning models against spurious correlations.
Diversity Maximization in the Presence of Outliers
Daichi Amagata (Osaka University)
CodeAnomaly DetectionOptimizationTabular
π― What it does: This paper studies the solution methods for the maximum-minimum diversity problem in the presence of outliers and proposes two approximation algorithms.
π― What it does: A completely decentralized multi-agent reinforcement learning framework (DM2) is proposed, achieving cooperation through independent target distribution matching by each agent.
DMIS: Dynamic Mesh-Based Importance Sampling for Training Physics-Informed Neural Networks
Zijiang Yang (University of Science and Technology Beijing), Dongmei Fu (University of Science and Technology Beijing)
CodeOptimizationComputational EfficiencyMeshBenchmarkPhysics Related
π― What it does: A dynamic mesh importance sampling (DMIS) scheme is proposed to accelerate the training of physics-informed neural networks (PINNs) and improve the accuracy of solutions.
π― What it does: This paper proposes a novel semantic tree structure expansion method called DNG, which constructs representations for each node by explicitly distinguishing between inherited features and supplementary features, and utilizes non-Gaussian constraints to achieve the directionality of is-a relationships.
Baojian Zhou (Fudan University), Steven Skiena (Stony Brook University)
CodeOptimizationTabular
π― What it does: This paper proposes an efficient algorithm AUC-opt for precisely optimizing AUC linear classifiers in two-dimensional space, and provides a recursively extendable implementation to higher dimensions; it also proves that linear AUC optimization is NP-complete when the dimension is not fixed.
π― What it does: A domain decorrelation method based on Potential Energy Ranking (PoER) is proposed, which explicitly separates label and domain information using ranking and clustering losses at both shallow and deep levels.
π― What it does: This paper proposes Domain Generalised Faster R-CNN, which addresses the domain transfer problem in object detection by introducing a new regularization term into Faster R-CNN, particularly enhancing the generalization ability to unknown domains without assuming covariate shift.
Domain-Adapted Dependency Parsing for Cross-Domain Named Entity Recognition
Chenxiao Dou (Nanhu Academy of Electronics and Information Technology), Xiangang Li (Beike)
CodeRecognitionDomain AdaptationRecurrent Neural NetworkLarge Language ModelText
π― What it does: This paper proposes a method that utilizes cross-domain dependency parsing (DP) as an auxiliary task to enhance named entity recognition (NER) in low-resource domains.
π― What it does: This paper studies the problem of domain generalization for crowd counting in unseen scenarios, proposing a meta-learning-based domain general model that implements dynamic subdomain partitioning of source domain data, re-encoding of domain-invariant and domain-specific memory modules, and corresponding loss constraints.
π― What it does: To address the issue of overconfidence in self-supervised learning (SSL) pre-trained speech recognition models, a hierarchical aggregation method is proposed to relax confidence and apply it to beam search decoding, thereby improving recognition accuracy.
Jeremiasz WoΕosiuk (Deepsolver), Jacek MaΕdziuk (Warsaw University of Technology)
CodeReinforcement LearningTabular
π― What it does: This paper studies an improved method for estimating counterfactual values (CFV) in poker games, proposing to obtain CFV by predicting expected value (EV) and then multiplying it by matching probability, instead of the traditional direct prediction of CFV.
π― What it does: A zero-shot hand-drawn sketch to 3D model retrieval method is proposed, and a large-scale Doodle2Object (D2O) dataset is constructed.
DPText-DETR: Towards Better Scene Text Detection with Dynamic Points in Transformer
Maoyuan Ye (Wuhan University), Dacheng Tao (The University of Sydney)
CodeObject DetectionTransformerImage
π― What it does: Proposes the DPText-DETR model, which utilizes dynamic points to achieve more efficient and accurate scene text detection in the Transformer framework;
π― What it does: This paper proposes an end-to-end visual grounding method that jointly extracts phrases from sentences and locates the corresponding image regions, addressing the shortcomings of traditional methods when phrases are unknown.
π― What it does: A foundational reinforcement learning framework for Dr. G zero-shot modeling is proposed, utilizing self-supervised dual contrastive learning and recursive state inverse dynamics to train encoders and world models, achieving strong generalization in visual control tasks under unseen visual disturbances.
π― What it does: A deep graph convolutional network DRGCN is proposed to alleviate the over-smoothing problem through a dynamically evolving initial residual mechanism.
Dropout Is NOT All You Need to Prevent Gradient Leakage
Daniel Scheliga (Technische Universitaet Ilmenau), Marco Seeland (Technische Universitaet Ilmenau)
CodeFederated LearningSafty and PrivacyAdversarial AttackConvolutional Neural NetworkTransformerImage
π― What it does: This study investigates the impact of using dropout on gradient inversion attacks in federated learning and proposes a new Dropout Inversion Attack (DIA) that can approximate the client model without the attacker knowing the specific dropout mask, thereby recovering training data.
π― What it does: A dual-label guided graph improvement framework, DuoLGR, is proposed for multi-view graph clustering, significantly enhancing clustering performance on low-homogeneity graphs.
Dual Memory Aggregation Network for Event-Based Object Detection with Learnable Representation
Dongsheng Wang (Dalian University of Technology), Huchuan Lu (Huawei Technologies Co. Ltd)
CodeObject DetectionConvolutional Neural NetworkTime Series
π― What it does: This paper proposes a learnable event representation method called EventPillars and a Dual Memory Aggregation Network (DMANet) to achieve high-precision object detection on event cameras.
π― What it does: A deep clustering framework DMICC based on dual mutual information constraints is proposed, which minimizes mutual information for redundancy at the feature level and maximizes mutual information for robust representation at the instance level, using K-means to output clustering results.
π― What it does: An end-to-end Transformer-based zero-shot learning framework called DUET is proposed, which achieves fine-grained induction and classification of image attributes through cross-modal semantic alignment and attribute-level contrastive learning.
Dynamic Ensemble of Low-Fidelity Experts: Mitigating NAS βCold-Startβ
Junbo Zhao (Tsinghua University), Yu Wang (Tsinghua University)
CodeOptimizationNeural Architecture SearchMixture of ExpertsImage
π― What it does: This paper proposes a Dynamic Ensemble Low-Fidelity Experts framework to enhance the ranking quality of predictors and accelerate the search during the cold start phase of predictor-based NAS by utilizing various low-cost performance estimations.
CodeRecommendation SystemNeural Architecture SearchGraph Neural NetworkGraphTime Series
π― What it does: A method for automated design of dynamic heterogeneous graph neural network architectures is proposed, aiming to automatically search for the optimal structure for different dynamic heterogeneous graph tasks without the need for human intervention.
π― What it does: A dynamic structure pruning method is proposed, which automatically learns the kernel grouping of different convolutional layers, significantly compressing the network while maintaining accuracy.
π― What it does: A dynamic retrieval-re-ranking-generation (DyRRen) framework is proposed to address numerical reasoning problems involving tables and long texts.
EASAL: Entity-Aware Subsequence-Based Active Learning for Named Entity Recognition
Yang Liu (Shenzhen Research Institute of Big Data), Tsung-Hui Chang (Chinese University of Hong Kong)
CodeRecognitionTransformerLarge Language ModelTextBiomedical Data
π― What it does: This paper proposes an active learning method based on entity-aware subsequences, EASAL, for named entity recognition tasks, significantly reducing annotation costs.
π― What it does: An efficient convolution kernel size learning framework, EffConv, has been designed to predict the optimal kernel size and generate corresponding weights within a small number of training epochs, achieving end-to-end differentiable optimization of CNN kernel sizes.
Effective and Stable Role-Based Multi-Agent Collaboration by Structural Information Principles
Xianghua Zeng (Beihang University), Angsheng Li (Zhongguancun Laboratory)
CodeReinforcement LearningTabularBenchmark
π― What it does: A role discovery method based on the principle of structural information, SIRD, is proposed and embedded in the MARL framework SR-MARL to achieve automatic role decomposition and learning for multi-agent collaboration.
π― What it does: A continuous learning framework based on lightweight adapters is proposed, which constructs compressed snapshots by freezing and saving adapters after each task training, utilizing knowledge distillation to allow the global model to continuously review old task knowledge while learning new tasks, thereby alleviating catastrophic forgetting.
π― What it does: A two-stage open-source intent classification method called CLAB is designed, which first obtains robust and balanced semantic representations through K-center contrastive learning, and then utilizes a tunable spherical decision boundary to achieve precise classification of known intents and effective recognition of unknown intents.
Efficient End-to-End Video Question Answering with Pyramidal Multimodal Transformer
Min Peng (Chongqing Institute of Green and Intelligent Technology, Chinese Academy of Sciences), Xiang-Dong Zhou (Chongqing Institute of Green and Intelligent Technology, Chinese Academy of Sciences)
π― What it does: An end-to-end video question answering model PMT is proposed, which interacts with text through a pyramid multimodal Transformer on multi-scale spatio-temporal features to achieve semantic reasoning of video content.
Marcel WienΓΆbst (University of LΓΌbeck), Maciej Liskiewicz (University of LΓΌbeck)
CodeGraph Neural NetworkGraph
π― What it does: This paper designs and implements a method that can enumerate all DAGs in all Markov equivalence classes with linear time delay, and provides an enumeration sequence that achieves a structural Hamming distance of no more than three.
Efficient Top-K Feature Selection Using Coordinate Descent Method
Lei Xu (Northwestern Polytechnical University), Xuelong Li (Northwestern Polytechnical University)
CodeOptimizationTabular
π― What it does: This paper proposes a non-parametric coordinate descent framework CD-LSR for efficiently solving the feature selection problem with βββ°-norm constraints.
π― What it does: A combinatorial search framework based on graph search (ESIA*) is proposed for non-invasive EEG/MEG source imaging, and provable optimality is provided within this framework.
π― What it does: A S3-CL framework is proposed for self-supervised learning of graph node representations through structural and semantic comparisons.
CodeExplainability and InterpretabilityTabularBenchmark
π― What it does: By automatically extracting high-confidence rules from data and using them as background knowledge, combined with formal reasoning techniques, more concise and credible 'why' (AXp) and 'why not' (CXp) explanations are generated;