AAAI Conference on Artificial Intelligence Β· 1014 papers
SciEval: A Multi-Level Large Language Model Evaluation Benchmark for Scientific Research
Liangtai Sun (Shanghai Jiao Tong University), Kai Yu (Shanghai Jiao Tong University)
CodeTransformerLarge Language ModelTextBenchmarkPhysics RelatedChain-of-Thought
π― What it does: SciEval is proposed, a multi-disciplinary, hierarchical scientific research assessment benchmark that covers four dimensions (fundamental knowledge, knowledge application, scientific computation, research capability) in chemistry, physics, and biology, while also providing static, dynamic, and experimental question banks.
π― What it does: A semantic decoupling GAN (SDGAN) is proposed, achieving precise control of facial attribute editing through an attribute-specific editing module and semantic masks.
SECap: Speech Emotion Captioning with Large Language Model
Yaoxun Xu (Tsinghua University), Rongzhi Gu (Chinese University of Hong Kong)
CodeGenerationTransformerLarge Language ModelContrastive LearningTextAudio
π― What it does: Proposes a speech emotion description (caption) task, utilizing large language models to generate natural language emotion descriptions;
Seed-Guided Fine-Grained Entity Typing in Science and Engineering Domains
Yu Zhang (University of Illinois), Jiawei Han (University of Illinois)
CodeClassificationTransformerLarge Language ModelSupervised Fine-TuningContrastive LearningText
π― What it does: A seed-guided fine-grained entity type classification framework SETYPE is proposed, which can be trained using only type names and a small number of seed entities, while simultaneously predicting both seen and unseen type entities.
SeGA: Preference-Aware Self-Contrastive Learning with Prompts for Anomalous User Detection on Twitter
Ying-Ying Chang (National Yang Ming Chiao Tung University), Wen-Chih Peng (National Yang Ming Chiao Tung University)
CodeAnomaly DetectionGraph Neural NetworkLarge Language ModelPrompt EngineeringContrastive LearningText
π― What it does: The SeGA framework is proposed for detecting anomalous users (including bots and trolls) on Twitter, achieving multi-class classification through heterogeneous information networks (HIN) and learning user content preferences.
Selective Deep Autoencoder for Unsupervised Feature Selection
Wael Hassanieh (University of Michigan), Abdallah Chehade (University of Michigan)
CodeRepresentation LearningAuto EncoderTabular
π― What it does: This paper proposes an unsupervised feature selection framework called Selective Deep AutoEncoder (SDAE), which utilizes deep autoencoders and a custom Selective Layer to automatically learn the minimal feature subset that can reconstruct the original feature space.
Selective Focus: Investigating Semantics Sensitivity in Post-training Quantization for Lane Detection
Yunqian Fan (ShanghaiTech University), Xianglong Liu (Beihang University)
CodeAutonomous DrivingOptimizationImage
π― What it does: A 'Selective Focus' framework is proposed for post-training quantization (PTQ) of lane detection models, utilizing semantic sensitivity to guide the quantization process.
CodeExplainability and InterpretabilityRepresentation LearningGraph Neural NetworkContrastive LearningGraph
π― What it does: The SUNNY-GNN framework is proposed, combining self-explanatory graph learning and contrastive learning to generate both sufficient and necessary explanatory subgraphs to enhance GNN prediction performance.
π― What it does: Proposes a Self-Prompt Mechanism (SPM) that generates self-prompts through spatial and channel selection on deep features of Vision Transformer, and injects these prompts into self-attention, enhancing few-shot image recognition performance with only 2% of parameters for efficient fine-tuning.
π― What it does: A self-training few-shot node classification method based on knowledge distillation (KD-FSNC) is proposed, which enhances the performance of the student model through representation distillation (local + global) and pseudo-label distillation after pre-training on the teacher model.
π― What it does: A new RoI feature extractor called Semantic RoI Align (SRA) is proposed, which can extract RoI features that are invariant to various geometric transformations in two-stage detectors.
π― What it does: A new semantic-guided novel category discovery method (Semantic-guided Novel Category Discovery, SNCD) is proposed, which simultaneously completes clustering of unlabeled images and semantic recognition under the name information of existing labeled categories and unknown categories.
π― What it does: This paper proposes two key technologies, PatchTeacher and PillarMix, for semi-supervised 3D object detection. It utilizes high-resolution voxelization of partial scenes to generate high-quality pseudo-labels and enhances data diversity through pillar mixing, ultimately achieving better detection performance.
π― What it does: This study investigates the application of semi-supervised learning in predicting the motion of unlabeled targets, proposing pseudo-label regeneration and BEVMix enhancement.
Sahal Shaji Mullappilly (Mohamed bin Zayed University of Artificial Intelligence), Hisham Cholakkal (Mohamed bin Zayed University of Artificial Intelligence)
π― What it does: A semi-supervised open-world object detection framework SS-OWFormer is proposed to address the problem of unknown object detection and incremental learning in the absence of complete labels.
Separate the Wheat from the Chaff: Model Deficiency Unlearning via Parameter-Efficient Module Operation
Xinshuo Hu (Harbin Institute of Technology), Min Zhang (Harbin Institute of Technology)
CodeTransformerLarge Language ModelSupervised Fine-TuningText
π― What it does: By extracting and subtracting the defect capability in the Anti-Expert Parameter Efficient Module (anti-expert PEM), the authenticity and detoxification of large language models are enhanced while maintaining their foundational capabilities.
CodeClassificationTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextSequential
π― What it does: SeqGPT achieves zero-shot reasoning across tasks by unifying natural language understanding tasks into two types of atomic tasks, constructing a consistent input-output format.
Huankang Guan (City University of Hong Kong), Rynson W.H. Lau (City University of Hong Kong)
CodeObject DetectionSegmentationTransformerImage
π― What it does: Proposes the SeqRank model, which predicts the attention order of salient objects in an image based on the eye movement mechanism of human vision.
π― What it does: A Shared Feature Calibration (SFC) method is proposed, which improves the pseudo-labels generated by CAM in weakly supervised semantic segmentation through image bank resampling and multi-scale distribution weighted consistency loss, enhancing the final segmentation performance.
SGFormer: Semantic Graph Transformer for Point Cloud-Based 3D Scene Graph Generation
Changsheng Lv (Beijing University of Posts and Telecommunications), Huadong Ma (Beijing University of Posts and Telecommunications)
CodeObject DetectionGenerationGraph Neural NetworkTransformerLarge Language ModelPoint Cloud
π― What it does: This paper proposes a Transformer structure named SGFormer for 3D scene graph generation based on point clouds, which can accurately predict object categories and their relationships through global information transmission.
π― What it does: A structure-guided network SGNet is proposed, which uses gradient domain and frequency domain information to enhance the super-resolution quality of low-resolution depth maps.
Shaping Up SHAP: Enhancing Stability through Layer-Wise Neighbor Selection
Gwladys Kelodjou (University of Rennes), Alexandre Termier (University of Rennes)
CodeExplainability and InterpretabilityTabular
π― What it does: This paper studies the instability of Kernel SHAP and proposes a ST-SHAP method based on hierarchical neighbor selection to enhance the stability of explanations. It further demonstrates that using only the first layer of neighbors can yield a fast interpretable solution that is highly consistent with SHAP values.
ShareBERT: Embeddings Are Capable of Learning Hidden Layers
Jia Cheng Hu (University of Modena and Reggio Emilia), Alessandro Capotondi (University of Modena and Reggio Emilia)
CodeCompressionOptimizationTransformerLarge Language ModelSupervised Fine-TuningText
π― What it does: Proposes the ShareBERT architecture, which constructs a near-zero parameter Transformer encoder using embedding parameter sharing (EPS and VEPS), thereby compressing the BERT model without using knowledge distillation;
π― What it does: This paper proposes a combination of smoothness-enhanced adversarial training and maximum squared loss in multi-source domain generalization, aiming to improve the model's performance on unseen domains under long-tailed distributions.
SHoP: A Deep Learning Framework for Solving High-Order Partial Differential Equations
Tingxiong Xiao (Tsinghua University), Jinli Suo (Tsinghua University)
CodeExplainability and InterpretabilityComputational EfficiencyPhysics RelatedOrdinary Differential Equation
π― What it does: This paper proposes the SHoP framework, which utilizes higher-order derivative rules to train MLPs for solving higher-order partial differential equations, and provides explicit solutions through Taylor expansion.
π― What it does: The SiMA-Hand framework is proposed, which enhances the accuracy of 3D hand mesh reconstruction from single-view RGB images by utilizing multi-view information during the training phase.
π― What it does: The SimDistill method is proposed, which achieves knowledge distillation for BEV 3D object detection through a simulated multimodal teacher-student architecture.
Ruohuan Fang (Beihang University), Xiao Bai (Beihang University)
CodeClassificationObject DetectionTransformerVision Language ModelContrastive LearningImage
π― What it does: A context-aware detection framework SIC-CADS based on CLIP image-level multi-label recognition is proposed, which enhances the detection accuracy of open vocabulary object detection by leveraging image-level global knowledge.
SimPSI: A Simple Strategy to Preserve Spectral Information in Time Series Data Augmentation
Hyun Ryu (Korea Advanced Institute of Science and Technology), Chang D. Yoo (Korea Advanced Institute of Science and Technology)
CodeData SynthesisAnomaly DetectionTransformerContrastive LearningTime Series
π― What it does: The SimPSI framework is proposed, which retains core spectral information during the time series data augmentation process through spectral mixing and retention mapping.
π― What it does: This paper proposes a walking recognition framework based on Skeleton Map, named SkeletonGait and SkeletonGait++, and validates the importance of skeletal structural information in walking recognition through alignment experiments with the traditional contour method DeepGaitV2.
π― What it does: A Sketch-and-Refine framework is proposed, which first uses a local direction map to quickly draw rough lane candidates, and then refines the candidates through the Lane Segment Association Module (LSAM) to achieve real-time and high-precision lane detection.
SkyScript: A Large and Semantically Diverse Vision-Language Dataset for Remote Sensing
Zhecheng Wang (Stanford University), Ram Rajagopal (Stanford University)
CodeClassificationRetrievalTransformerVision Language ModelContrastive LearningImageTextMultimodality
π― What it does: SkyScript has been constructedβa large-scale diversified visual-language dataset containing 2.6 million remote sensing image-text pairs, covering 29,000 semantic labels. Continuous pre-training on this dataset resulted in SkyCLIP, enabling zero-shot scene classification, fine-grained attribute classification, and cross-modal retrieval in the remote sensing field.
π― What it does: Proposes the SlowTrack framework, which generates low-intrusion delayed adversarial samples for the entire camera perception system (object detection + multi-object tracking).
Haixia Han (East China Normal University), Yanghua Xiao (Fudan University)
CodeGenerationTransformerLarge Language ModelSupervised Fine-TuningTextChain-of-Thought
π― What it does: An Intrinsic Self-Correction (ISC) mechanism is proposed, enabling small language models (6-13 billion parameters) to automatically verify and correct errors after generating answers.
SoftCLIP: Softer Cross-Modal Alignment Makes CLIP Stronger
Yuting Gao (Tencent Youtu Lab), Xing Sun (Tencent Youtu Lab)
CodeClassificationRetrievalTransformerVision Language ModelContrastive LearningImageTextMultimodality
π― What it does: This paper proposes SoftCLIP, which relaxes the one-to-one alignment constraint of CLIP, utilizing the fine-grained same-modal similarity between image ROIs and text labels as soft targets, and decoupling negative samples to achieve more flexible and accurate cross-modal alignment.
π― What it does: For multi-view 3D object detection, a semantic occupancy branch (3D Semantic Occupancy) is added in the BEV space to simultaneously predict object detection and environmental semantic occupancy, enhancing the perception of physical context.
Solving Satisfiability Modulo Counting for Symbolic and Statistical AI Integration with Provable Guarantees
Jinzhao Li (Purdue University), Yexiang Xue (Purdue University)
CodeOptimizationGraphTabularAgriculture Related
π― What it does: A XOR SMC algorithm is proposed to transform countable problems into SAT, utilizing random XOR constraints to achieve approximate model counting with constant approximation guarantees;
Spanning the Spectrum of Hatred Detection: A Persian Multi-Label Hate Speech Dataset with Annotator Rationales
Zahra Delbari, Mohammad Taher Pilehvar (Cardiff University)
CodeClassificationExplainability and InterpretabilityTransformerLarge Language ModelSupervised Fine-TuningText
π― What it does: This paper constructs a new multi-label, annotated rationale Persian hate speech dataset called PHATE, which contains approximately 7,000 manually annotated tweets.
π― What it does: A SparseEnNet based on adversarial generation is proposed for data augmentation in sequential recommendation, generating more robust augmented sequences.
Spatial Transform Decoupling for Oriented Object Detection
Hongtian Yu (University of Chinese Academy of Sciences), Yunfan Liu (University of Chinese Academy of Sciences)
CodeObject DetectionTransformerImage
π― What it does: A Spatial Transform Decoupling (STD) framework is designed, utilizing the multi-branch structure of Vision Transformer to predict position, size, and angle separately, and enhancing foreground features layer by layer through cascading activation masks, thereby achieving directed object detection.
π― What it does: A GAN inversion method named SDIC is proposed, which first generates a difference map through a Spatial-Contextual Information Prediction Network (DIPN), and then uses this difference map in a Difference Information Compensation Network (DICN) to compensate both the latent code and the early features of the generator, ultimately achieving high-quality reconstruction and editable image output.
SPD-DDPM: Denoising Diffusion Probabilistic Models in the Symmetric Positive Definite Space
Yunchen Li (East China Normal University), Shaohui Lin (Tencent)
CodeGenerationData SynthesisConvolutional Neural NetworkDiffusion modelGraphTime Series
π― What it does: This paper proposes a denoising diffusion probabilistic model (SPD-DDPM) that operates in the space of symmetric positive definite matrices (SPD) for unconditional and conditional generation of SPD matrices, and introduces a deeper SPD-U-Net to enhance the network's fitting capability.
Spear and Shield: Adversarial Attacks and Defense Methods for Model-Based Link Prediction on Continuous-Time Dynamic Graphs
Dongjin Lee (Korea Advanced Institute of Science and Technology), Kijung Shin (Korea Advanced Institute of Science and Technology)
CodeRecommendation SystemAdversarial AttackGraph Neural NetworkGraphTime Series
π― What it does: Designed and implemented an adversarial attack method T-SPEAR for continuous-time dynamic graph link prediction and the corresponding robust training defense method T-SHIELD.
π― What it does: A spectral domain-based graph neural network SComGNN is proposed, which uses low-pass and band-pass filters to extract the correlation and diversity features of complementary goods, respectively, and adaptively fuses them through a dual-stage attention mechanism to address the balance between correlation and diversity in complementary recommendations.
π― What it does: A spectral rendering framework called SpectralNeRF based on NeRF is proposed, which can generate high-quality white light RGB images from multi-wavelength perspectives.
π― What it does: This paper proposes SpFormer, a Transformer-based scanpath modeling framework that can simultaneously capture location, time, and dwell time information.
π― What it does: This paper proposes a Spiking NeRF using a hybrid ANN-SNN framework, which implements a discontinuous density field with discrete spiking neurons, thereby enabling more accurate reconstruction of 3D geometric structures.
SpikingBERT: Distilling BERT to Train Spiking Language Models Using Implicit Differentiation
Malyaban Bal, Abhronil Sengupta (Pennsylvania State University)
CodeComputational EfficiencyKnowledge DistillationSpiking Neural NetworkTransformerLarge Language ModelSupervised Fine-TuningText
π― What it does: A BERT model based on spiking neural networks, SpikingBERT, is proposed and trained through implicit differentiation and knowledge distillation.
Spot the Error: Non-autoregressive Graphic Layout Generation with Wireframe Locator
Jieru Lin (Microsoft Research), Chin-Yew Lin (Microsoft Research)
CodeObject DetectionGenerationTransformerImage
π― What it does: This paper studies non-autoregressive (NAR) graphic layout generation and proposes a learning-based locator that corrects generation errors through iterative mask prediction using rendered wireframe images.
π― What it does: The CoNet framework is proposed, which achieves one-shot open-set object/text detection through visual prototypes, utilizing self-correlation and cross-correlation modules for dense correspondence and spatial alignment.
π― What it does: A self-supervised monocular depth estimation method based on Self Query Layer (SQL) called SQLdepth is proposed, which can learn fine-grained scene geometry through self-cost volume and directly predict depth maps from single-frame images.
STAS: Spatial-Temporal Return Decomposition for Solving Sparse Rewards Problems in Multi-agent Reinforcement Learning
Sirui Chen (Renmin University of China), Yali Du (King's College London)
CodeTransformerReinforcement LearningSequential
π― What it does: This paper proposes STAS, a multi-agent reward splitting method based on spatial-temporal attention and Shapley values, addressing the credit assignment problem under sparse rewards.
StegFormer: Rebuilding the Glory of Autoencoder-Based Steganography
Xiao Ke, Wenzhong Guo (Fuzhou University)
CodeData SynthesisSafty and PrivacyTransformerAuto EncoderImage
π― What it does: A StegFormer model based on autoencoders is proposed, capable of embedding one or more secret images into carrier images of the same resolution, enhancing reliability in real-world scenarios through a normalization training strategy and constrained loss.
STEM: Unleashing the Power of Embeddings for Multi-Task Recommendation
Liangcai Su (Tsinghua University), Jie Jiang (Tsinghua University)
CodeRecommendation SystemMixture of ExpertsVideo
π― What it does: This paper proposes the Shared and Task-Specific Embedding (STEM) paradigm and implements the STEm-Net model, specifically addressing negative transfer in multi-task recommendation, achieving positive transfer especially on comparable samples with a balanced positive-negative sample ratio; it conducts fine-grained partitioning of comparable samples, revealing that traditional shared embedding methods are prone to negative transfer on these samples.
π― What it does: This paper proposes an adversarial attack method for conditional diffusion models called Mean Fluctuation Attack (MFA). It exploits the sensitivity of diffusion models to the mean of noise by inducing a mean shift during the reverse sampling process, thereby degrading the generation quality. Furthermore, it investigates the vulnerabilities of different reverse steps and introduces two attack variants guided by vulnerability: MFA-VT and MFA-MVS.
π― What it does: This paper proposes a Time-domain Multi-Plane Image (TMPI) model that generates high-quality stereo or light field videos from planar videos.
π― What it does: The research focuses on Bayesian optimization under unknown continuous contextual distributions and proposes two algorithms: SBO-KDE and DRBO-KDE.
Structural Information Enhanced Graph Representation for Link Prediction
Lei Shi (Ant Group), Jun Zhou (Ant Group)
CodeGraph Neural NetworkTransformerGraph
π― What it does: In the task of link prediction in graphs, a structure information enhanced graph representation framework (SIEG) is proposed, which improves prediction performance by removing neighbor node features, utilizing GNN to encode neighborhood structures, and introducing a Binary Structural Transformer (BST) to encode the structural relationships of target node pairs.
π― What it does: This paper proposes VehicleMAE, a multimodal pre-training framework for vehicle perception, which utilizes structural priors (vehicle contours) and semantic priors (natural language descriptions) to guide MAE in vehicle image reconstruction.
Dou Hu (Institute of Information Engineering Chinese Academy of Sciences), Songlin Hu (Institute of Information Engineering Chinese Academy of Sciences)
π― What it does: A supervised representation learning framework named Structured Probabilistic Coding (SPC) is proposed, which utilizes an encoder to simultaneously perform probabilistic coding and task prediction, extracting compact and information-rich representations from the input.
π― What it does: Proposes the SSD method, which stores more informative samples summarized from streaming data in memory to enhance the replay effect of online continual learning.
π― What it does: A learning-based framework for vascular junction detection and description, called SuperJunction, is proposed for retinal image registration.
π― What it does: In the surgical tool segmentation task, an end-to-end method was achieved through efficient fine-tuning of the Segment Anything Model (SAM), enabling segmentation with only category prompts and no explicit point/frame prompts.
Symbolic Cognitive Diagnosis via Hybrid Optimization for Intelligent Education Systems
Junhao Shen (Shanghai Institute of AI for Education), Aimin Zhou (Shanghai Institute of AI for Education)
CodeOptimizationExplainability and InterpretabilityTabular
π― What it does: A Symbolic Cognitive Diagnosis (SCD) framework is proposed, utilizing a combination of symbolic trees and gradient optimization to diagnose students' knowledge attributes.
Symmetric Self-Paced Learning for Domain Generalization
Di Zhao (University of Auckland), Philippe Fournier-Viger (Shenzhen University)
CodeDomain AdaptationImage
π― What it does: This paper proposes Symmetric Self-Paced Learning (SSPL) for domain generalization, combining a Symmetric Self-Paced training scheduler and Gradient Difficulty Measurement (GDM);
T-SciQ: Teaching Multimodal Chain-of-Thought Reasoning via Large Language Model Signals for Science Question Answering
Lei Wang (Beijing Forestry University), Heng Tao Shen (Beijing Rongda Technology)
CodeTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextMultimodalityChain-of-Thought
π― What it does: The T-SciQ method is proposed, which generates high-quality chain-of-thought (CoT) and plan-based CoT (PCoT) as teaching signals through large language models, training smaller multimodal models to complete the ScienceQA task.
π― What it does: A low-cost T2I-Adapter module is proposed to align external control signals (such as sketches, depth, color, key points, and semantic segmentation) with the internal knowledge of a pre-trained text-to-image diffusion model, enabling fine control over structure and color.
TA&AT: Enhancing Task-Oriented Dialog with Turn-Level Auxiliary Tasks and Action-Tree Based Scheduled Sampling
Longxiang Liu (Chinese Academy of Sciences), Yang Feng (Chinese Academy of Sciences)
CodeGenerationTransformerLarge Language ModelSupervised Fine-TuningTextBenchmark
π― What it does: This study proposes a dialogue turn-based auxiliary task and an action tree-based scheduling sampling method, improving the understanding and generation of end-to-end task-oriented dialogue systems;
π― What it does: Proposes the TACIT framework for cross-domain text classification with unknown target domains, enhancing cross-domain generalization performance by decoupling features using only source domain data.
Tackling Vision Language Tasks through Learning Inner Monologues
Diji Yang (University of California), Yi Zhang (Mineral)
CodeOptimizationExplainability and InterpretabilityReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningVision Language ModelTextMultimodality
π― What it does: This paper proposes a multimodal optimization framework based on 'Inner Monologue' (IMMO), which utilizes a visual language model (Observer) and a large language model (Reasoner) to generate questions and answers and construct internal dialogues through natural language multi-turn conversations, addressing complex visual language reasoning tasks such as visual question answering and visual entailment.
π― What it does: Without the need for any training, we propose the TagCLIP local-to-global three-step framework (patch-level classification β Dual-Mask Attention Refinement DMAR β Class-level Re-identification CWR), achieving open-vocabulary multi-label classification with CLIP and using the generated labels as pseudo-labels for weakly supervised semantic segmentation.
TagFog: Textual Anchor Guidance and Fake Outlier Generation for Visual Out-of-Distribution Detection
Jiankang Chen (Sun Yat-sen University), Ruixuan Wang (Sun Yat-sen University)
CodeAnomaly DetectionRepresentation LearningLarge Language ModelContrastive LearningImageText
π― What it does: This paper proposes a new learning framework called TagFog, which generates pseudo-discrete samples using Jigsaw techniques and obtains semantic anchors through ID class descriptions generated by ChatGPT via the CLIP text encoder, guiding the visual encoder to learn more compact and semantically rich features, thereby improving visual OOD detection performance.
π― What it does: In response to the poor performance of graph neural networks for tail users (low degree) in friend recommendation tasks, the Tail-STEAK framework is proposed to achieve better representation learning for tail users.
Taming the Sigmoid Bottleneck: Provably Argmaxable Sparse Multi-Label Classification
Andreas Grivas (University of Edinburgh), Adam Lopez (University of Edinburgh)
CodeClassificationBiomedical DataElectronic Health Records
π― What it does: This study investigates the bottleneck of low-rank sigmoid output layers in multi-label classification and proposes an output layer based on Discrete Fourier Transform (DFT) to ensure that all k-sparse label combinations can be correctly predicted.
π― What it does: A Targeted Activation Penalty (TAP) method is proposed for convolutional neural networks to suppress the model's reliance on spurious signals and enhance generalization performance.
π― What it does: This paper proposes a task-adaptive prompt transformer (MetaPrompt) model for cross-domain few-shot learning (CD-FSL), utilizing attention to generate task prompts for rapid task adaptation.
π― What it does: A task-agnostic continual learning framework for generation and representation learning (Continual Variational Autoencoder, CAA) is proposed, which achieves task boundary-free generation and reconstruction of continuously changing data streams through a dual memory system (temporary memory and evolving memory) and a two-step optimization strategy.
π― What it does: This paper proposes a Task-Free Dynamic Sparse Vision Transformer (TFDSViT) for training visual Transformers in continuous learning scenarios without task boundaries.
π― What it does: This paper proposes a Fast Adversarial Training based on Adversarial Sample Classification (TDAT), which systematically alleviates catastrophic overfitting in single-step adversarial training and significantly enhances model robustness through joint improvements in initialization, dynamic label relaxation, and classification-driven loss functions.
π― What it does: A dual-chamber leakage integral firing (TC-LIF) synaptic neuron model is proposed to address the long-term temporal dependency learning problem.
π― What it does: A lightweight table detection method TDeLTA based on text block arrangement learning is proposed, which locates tables using the positions of text blocks rather than image features.
Teaching Large Language Models to Translate with Comparison
Jiali Zeng (Tencent Inc), Jie Zhou (Tencent Inc)
CodeTransformerLarge Language ModelSupervised Fine-TuningContrastive LearningText
π― What it does: By introducing output comparison and preference comparison within the framework of contrastive learning, we fine-tune open-source large language models to enhance machine translation performance.
Jing Li (Xidian University), Wei Xia (Xidian University)
CodeOptimizationGraph
π― What it does: A tensor label learning method based on anchor graphs (TLL-AG) is proposed, which directly obtains soft labels from anchor graphs through orthogonal non-negative matrix factorization without the need for post-processing;
Ternary Spike: Learning Ternary Spikes for Spiking Neural Networks
Yufei Guo (Intelligent Science and Technology Academy of CASIC), Zhe Ma (Intelligent Science and Technology Academy of CASIC)
CodeClassificationSpiking Neural NetworkImage
π― What it does: This paper proposes a three-valued spiking neuron (-1, 0, 1) and learnable three-valued spikes to enhance the representational capacity of spiking neural networks while retaining the advantages of event-driven and additive operations.
Terrain Diffusion Network: Climatic-Aware Terrain Generation with Geological Sketch Guidance
Zexin Hu (University of Sydney), Zhiyong Wang (University of Sydney)
CodeGenerationDiffusion modelAuto EncoderImage
π― What it does: The paper proposes a multi-level terrain generation network (TDN) based on diffusion models, capable of generating climate-aware high-quality terrain maps guided by user-provided sketches of rivers, ridges, basins, and peaks.
π― What it does: A testing time domain adaptation (TT-DA) framework is proposed, which utilizes the affine parameters of the batch normalization (BN) layers for domain knowledge learning, updating only the Ξ³ and Ξ² of BN; at the same time, a self-supervised branch is introduced to provide domain-related supervision for unlabeled data, and a meta-learning dual-layer optimization is employed to enable the affine parameters to quickly adapt to new domains; ultimately improving cross-domain performance without increasing inference costs.
π― What it does: This paper proposes the task of text image inpainting, constructs two real and synthetic datasets for scene text and handwritten text, and introduces a global structure-guided diffusion model (GSDM) to achieve complete restoration of text images.