- Adversarial Robustness in Machine Learning
- Domain Adaptation and Few-Shot Learning
- Anomaly Detection Techniques and Applications
- Prosthetics and Rehabilitation Robotics
- Stroke Rehabilitation and Recovery
- Advanced Neural Network Applications
- Bacillus and Francisella bacterial research
- Muscle activation and electromyography studies
- Multimodal Machine Learning Applications
- Vehicle License Plate Recognition
- COVID-19 diagnosis using AI
- Human Pose and Action Recognition
- Signaling Pathways in Disease
- Vibration Control and Rheological Fluids
- Natural Language Processing Techniques
- Network Security and Intrusion Detection
- Gait Recognition and Analysis
- Speech and dialogue systems
- Text and Document Classification Technologies
- Biometric Identification and Security
- Retinal Imaging and Analysis
- Infectious Encephalopathies and Encephalitis
- Advanced Image and Video Retrieval Techniques
- Control and Dynamics of Mobile Robots
- Vibration and Dynamic Analysis
Beibu Gulf University
2022-2025
Nagasaki Institute of Applied Science
2023-2025
Shanghai Jiao Tong University
2019-2024
Second Military Medical University
2024
Henan University
2023-2024
Tencent (China)
2024
Henan Provincial People's Hospital
2023-2024
Zhengzhou University
2023-2024
Gulf University
2023
We propose a variational Bayesian scheme for pruning convolutional neural networks in channel level. This idea is motivated by the fact that deterministic value based methods are inherently improper and unstable. In nutshell, technique introduced to estimate distribution of newly proposed parameter, called saliency, on this, redundant channels can be removed from model via simple criterion. The advantages two-fold: 1) Our method conducts without desire re-training stage, thus improving...
We propose a variational Bayesian framework for enhancing few-shot learning performance. This idea is motivated by the fact that single point based metric approaches are inherently noise-vulnerable and easy-to-be-biased. In nutshell, stochastic inference invoked to approximate bias-eliminated class specific sample distributions. meantime, classifier-free prediction attained leveraging distribution statistics on novel samples. Extensive experimental results several benchmarks well demonstrate...
Lower limb exoskeleton rehabilitation robots have become an important direction for development in today’s society. These can provide support and power to assist patients walking movement. In order achieve better interaction between humans machines the goal of flexible driving, this paper addresses shortcomings traditional elastic actuators designs a series elastic–damping actuator (SEDA). The SEDA combines damping components parallel, feasibility design material selection is demonstrated...
This paper addresses the challenging black-box adversarial attack problem, where only classification confidence of a victim model is available. Inspired by consistency visual saliency between different vision models, surrogate expected to improve performance via transferability. By combining transferability-based and query-based attack, we propose surprisingly simple baseline approach (named SimBA++) using model, which significantly outperforms several state-of-the-art methods. Moreover,...
Population aging is an inevitable trend in contemporary society, and the application of technologies such as human-machine interaction, assistive healthcare, robotics daily service sectors continues to increase. The lower limb exoskeleton rehabilitation robot has great potential areas enhancing human physical functions, training, assisting elderly disabled. This paper integrates structural characteristics limb, motion mechanics, gait features design a biomimetic structure proposes integrated...
Palm vein recognition is an emerging biometric technology that offers enhanced security and privacy. However, acquiring sufficient palm data for training deep learning-based models challenging due to the high costs of collection privacy protection constraints. This has led a growing interest in generating pseudo-palm using generative models. Existing methods, however, often produce unrealistic patterns or struggle with controlling identity style attributes. To address these issues, we...
To address the limitations of existing pruning methods in practical applications, such as necessity training from scratch with sparsity regularization or complex data-driven optimization, we set out a novel perspective to explore parameter redundancy and accelerate deep CNNs. Precisely, argue that channels revealing similar feature information have functional overlap each similarity group can be reduced few representatives little impact on representational power model. After deriving an...
The lack of large-scale data seriously hinders the development palmprint recognition. Recent approaches address this issue by generating realistic pseudo palmprints from Bézier curves. However, significant difference between curves and real limits their effectiveness. In paper, we divide Bézier-Real into creases texture differences, thus reducing generation difficulty. We introduce a new palm crease energy (PCE) domain as bridge to propose two-stage model. first stage generates PCE images...
In this work, we study the ambiguity problem in task of unsupervised 3D human pose estimation from 2D counterpart. On one hand, without explicit annotation, scale is difficult to be accurately captured (scale ambiguity). other might correspond multiple gestures, where lifting procedure inherently ambiguous (pose Previous methods generally use temporal constraints (e.g., constant bone length and motion smoothness) alleviate above issues. However, these commonly enforce outputs fulfill...
In this paper, we decouple unsupervised human mesh recovery into the well-studied problems of 3D pose estimation, and from estimated skeletons, focusing on latter task. The challenges task are two folds: (1) failure (i.e., mismatching – different skeleton definitions in dataset SMPL , ambiguity endpoints have arbitrary joint angle configurations for same coordinates). (2) shape lack constraints body configuration). To address these issues, propose Skeleton2Mesh, a novel lightweight framework...
Prior research has indicated that bisphosphonates (BPs) can improve periodontal disease because of their anti-osteoporosis properties. In vitro studies have shown BPs induce cytotoxicity, inhibit wound healing, and thus affect disease. Denosumab alternative indications. BP denosumab are not known to correlate with gingival disorders. We assessed such a relationship by applying Bayesian nonproportional analyses data in the US FDA Adverse Event Reporting System (FAERS) database. The study...
Predicting the time series of 10.7-cm solar radio flux is a challenging task because its daily variability. This paper proposed non-linear method, convolutional and recurrent neural network combined model to achieve end-to-end F10.7 forecasts. The consists one-dimensional long short-term memory network. CNN extracted features from original data, then trained feature signals in network, outputted predicted values. data during 2003–2014 are used for testing set. mean absolute percentage error...
During the walking process of lower limb exoskeleton rehabilitation robots, inevitable collision impacts will occur when swinging leg lands on ground. The impact reaction force from ground induce vibrations in entire robot’s body bottom to top. To address this phenomenon, considering limitations traditional active compliance and passive methods, a variable stiffness damping actuator (VSDA) structure using magnetorheological damper (MRD) is proposed. Firstly, experimental methods are used...
In this paper, we identify symmetry property in adversarial scenario by viewing attack a fine-grained manner. A newly designed metric called proportion, is thus proposed to count the proportion of examples misclassified between classes. We observe that distribution unbalanced as each class shows vulnerability particular Further, some pairs correlate strongly and have same degree for other. call intriguing phenomenon property. empirically prove widespread then analyze reason behind existence...
Various methods have been proposed to defend against adversarial attacks. However, there is a lack of enough theoretical guarantee the performance, thus leading two problems: First, deficiency necessary training samples might attenuate normal gradient's back-propagation, which leads overfitting and gradient masking potentially. Second, point-wise sampling offers an insufficient support region for data cannot form robust decision-boundary. To solve these issues, we provide analysis reveal...
To address the limitations of existing magnitude-based pruning algorithms in cases where model weights or activations are large and similar magnitude, we propose a novel perspective to discover parameter redundancy among channels accelerate deep CNNs via channel pruning. Precisely, argue that revealing feature information have functional overlap most within each such similarity group can be removed without compromising model's representational power. After deriving an effective metric for...
In this paper, we give a new definition for sample complexity, and further develop theoretical analysis to bridge the gap between complexity model capacity. contrast previous works which study on some toy samples, conduct our more general data space, build qualitative relationship from capacity required achieve comparable performance. Besides, introduce simple indicator evaluate based continuous mapping. Moreover, distribution, paves way understand present representation learning. Extensive...
Abstract The aging population is a necessary trend in today's society, and there are increasing applications everyday services through human-computer interaction, paramedicine robotics. Lower limb exoskeleton rehabilitation robots have broad application prospects the fields of enhancing human functions, training, helping elderly disabled. In this paper, we combine structural characteristics, motion mechanism, gait characteristics lower limbs, bionic design mechanical structure, propose...
Recently adversarial examples have been reported to reveal the fragility of deep learning models. However, most attacks focus on classification task and less attention has paid retrieval task. In this paper, we are first investigate video system in both non-targeted targeted attack terms for copyright protection. Specifically, a triplet scheme is developed take query-relevant query-target pair-wise relationships together enhance performance. We evaluated proposed method commonly used dataset...