A5100324286- Domain Adaptation and Few-Shot Learning
- Multimodal Machine Learning Applications
- Anomaly Detection Techniques and Applications
- COVID-19 diagnosis using AI
- Machine Learning and Data Classification
- Advanced Graph Neural Networks
- Robotics and Sensor-Based Localization
- Text and Document Classification Technologies
- Sparse and Compressive Sensing Techniques
- Advanced Computational Techniques and Applications
- Topic Modeling
- Robot Manipulation and Learning
- Industrial Vision Systems and Defect Detection
- Music and Audio Processing
- Cryptography and Data Security
- Human Pose and Action Recognition
- Advanced Authentication Protocols Security
- Advanced Control Systems Optimization
- Security in Wireless Sensor Networks
- Soil and Unsaturated Flow
- Reinforcement Learning in Robotics
- Advanced Neural Network Applications
- Image and Object Detection Techniques
- Generative Adversarial Networks and Image Synthesis
- Constraint Satisfaction and Optimization
The University of Sydney
2020-2024
Beihang University
2014-2024
VERSES (United States)
2022-2024
National University of Defense Technology
2017-2022
Guangxi University
2022
The University of Texas at Austin
2021
Chinese University of Hong Kong
2021
University of Technology Sydney
2017-2020
Tianjin University
2019
University of Kentucky
2019
Continual learning methods aim at training a neural network from sequential data with streaming labels, relieving catastrophic forgetting. However, existing are based on and designed for convolutional networks (CNNs), which have not utilized the full potential of newly emerged powerful vision transformers. In this paper, we propose novel attention-based framework Lifelong Vision Transformer (LVT), to achieve better stability-plasticity trade-off continual learning. Specifically, an...
With the development of hardware and software in camera computing units, visual inspection system (VIS) plays an increasing significant role fault task. This paper proposes a VIS to inspect missing bogie block key (BBK) used on freight trains. BBK is important component keep wheel sets from separating out bogies. The one most common faults threatening running safety. first acquires image by acquisition system, then hierarchical framework containing bearing cap (BC) detection, region...
Scene Graph Generation (SGG) aims to build a structured representation of scene using objects and pairwise relationships, which benefits downstream tasks. However, current SGG methods usually suffer from sub-optimal graph generation because the long-tailed distribution training data. To address this problem, we propose Resistance Training Prior Bias (RTPB) for generation. Specifically, RTPB uses distributed-based prior bias improve models' detecting ability on less frequent relationships...
Graph Convolutional Networks (GCNs) suffer from performance degradation when models go deeper. However, earlier works only attributed the degeneration to over-smoothing. In this paper, we conduct theoretical and experimental analysis explore fundamental causes of in deep GCNs: over-smoothing gradient vanishing have a mutually reinforcing effect that deteriorate more quickly GCNs. On other hand, existing anti-over-smoothing methods all perform full convolutions up model depth. They could not...
"A picture is worth a thousand words", significantly beyond mere categorization. Accompanied by that, many patches of the image could have completely irrelevant meanings with categorization if they were independently observed. This reduce efficiency large family few-shot learning algorithms, which limited data and highly rely on comparison patches. To address this issue, we propose Class-aware Patch Embedding Adaptation (CPEA) method to learn "class-aware embeddings" The key idea CPEA...
GoDec is an efficient low-rank matrix decomposition algorithm. However, optimal performance depends on sparse errors and Gaussian noise. This paper aims to address the problem that a composed of component unknown corruptions. We introduce robust local similarity measure called correntropy describe corruptions and, in doing so, obtain more faster algorithm: GoDec+. Based half-quadratic optimization greedy bilateral paradigm, we deliver solution maximum criterion (MCC)-based problem....
Few-shot learning suffers from the scarcity of labeled training data. Regarding local descriptors an image as representations for could greatly augment existing Existing descriptor based few-shot methods have taken advantage this fact but ignore that semantics exhibited by may not be relevant to semantic. In paper, we deal with issue a new perspective imposing semantic consistency image. Our proposed method consists three modules. The first one is extractor module, which can extract large...
Continual learning is an intellectual ability of artificial agents to learn new streaming labels from sequential data. The main impediment continual catastrophic forgetting, a severe performance degradation on previously learned tasks. Although simply replaying all previous data or continuously adding the model parameters could alleviate issue, it impractical in real-world applications due limited available resources. Inspired by mechanism human brain deepen its past impression, we propose...
Neural networks tend to suffer performance deterioration on previous tasks when they are applied multiple sequentially without access data. The problem is commonly known as catastrophic forgetting, a significant challenge in continual learning (CL). To overcome the regularization-based CL methods construct term, which can be considered approximation loss function of tasks, penalize update parameters. However, rigorous theoretical analysis limited. Therefore, we theoretically analyze...
Inspired by the human learning principle that easier concepts first and then gradually paying more attention to harder ones, curriculum uses nonuniform sampling of mini-batches according order examples' difficulty. Just as a teacher adjusts progress each student, proper should be adapted current state model. Therefore, in contrast recent works using fixed curriculum, we devise new method, Adaptive Curriculum Learning (Adaptive CL), adapting difficulty examples Specifically, make use loss...
CLIP has the ability to align texts and images is nearly most frequently used foundation model in cross-modal zero-shot learning. However, our experimental findings reveal that suffers from a bias text-to-image retrieval, resulting decrease CLIP's learning performance. We analytically discover partly arises imbalanced range of similarity scores obtained by CLIP. Accordingly, we propose Balanced Similarity with Auxiliary Prompts (BSAP) mitigate retrieval Specifically, BSAP designs auxiliary...
With the development of both hardware and software technologies in camera computer, automated visual inspection system is being used more intelligent transportation for its high efficiency. For safety operation, it necessary to perform fault train mechanical components. As one most widely small components freight trains, bogie block key (BBK) keep wheel sets from separating out bogies, likely cause terrible accidents. This study proposes a vision‐based inspect missing BBK automatically. To...
It is difficult to implement three-dimensional (3D) measurement in small field of view (FOV) or confined space with traditional sensors, for they cannot be put into operated flexibly such circumstances. To solve the problem, a sensor constructed by an electronic endoscope and pair mirrors designed, combining flexible characteristics transmission wire advantages stereo technology. The calibration two corresponding points matching methods are described. For applications as diameter 3-D circle,...
Flight safety is of vital importance to the aviation industry. As one most typical security events, hard landing extremely concerned by airlines and related studies have received extensive attention in recent years. However, existing regression or risk based models either suffers from low prediction accuracy, cannot provide good interpretability, making themselves impractical real applications. To solve these problems, this paper we propose CurveCluster: a curve clustering-based approach...
This work studies the task of glossification, which aim is to em transcribe natural spoken language sentences for Deaf (hard-of-hearing) community ordered sign glosses. Previous sequence-to-sequence models trained with paired sentence-gloss data often fail capture rich connections between two distinct languages, leading unsatisfactory transcriptions. We observe that despite different grammars, glosses effectively simplify ease deaf communication, while sharing a large portion vocabulary...
Streaming label learning aims to model newly emerged labels for multilabel classification systems, which requires plenty of new data training. However, in changing environments, only a small amount can practically be collected. In this work, we formulate and study few-shot streaming (FSLL), models emerging with few annotated examples by utilizing the knowledge learned from past labels. We propose meta-learning framework, semantic inference network (SIN), learn infer correlation between adapt...
The lower limb power-assist exoskeletons are expected to help paraplegic people walk again in daily life. However, most of these deal with walking the scene that has been seen or an external vision sensor, rather than unknown environment. It is a great challenge understand wear's intention and plan footstep sequence scene. Moreover, traditional visual planning dominated by robot, which can lead awkward trajectory plan. Therefore, we construct system propose onboard algorithm based on Bezier...