Jiacheng Chen
Jiacheng Chen has a total of 14 co-authors. Here are the top co-authors by number of shared publications:
- Bin-Bin Gao
Tencent Youtu Lab, Shenzhen, China
1 shared publication - Chengjie Wang
YouTu Lab, Tencent, Shanghai, China
1 shared publication - Chentao Xu
School of Cyber Science and Engineering, Southeast University, Nanjing 210096, China
1 shared publication - Deyuan Liu
Beihang University
1 shared publication Haoda Qu
School of Mathematics, Southeast University,Nanjing,China,210096
1 shared publication- Hichem Snoussi
University of Technology of Troyes
1 shared publication
Latest Publications
- 17 November 2023
- Published by Institute of Electrical and Electronics Engineers (IEEE)
- p. 5549-5554
Inertial Zeroing Neural Network for Robot Manipulator Control Described by Different-Level Dynamic Linear System
Enhancing the stability and precision of robot manipulator control represents a pivotal focus in engineering. In this paper, we focus on studying how to optimize the stability and accuracy of robot manipulator control. The proposed approach merges an inertial neural network with a discrete zeroing neural network, resulting in a remarkable reduction of shaking and error. The discrete zeroing neural network model is proposed in the basis of the different-level dynamic linear system. However, this model has the disadvantage of significant oscillations in solving different-level dynamic linear systems with great joint angle jitter and derivative of the change in joint values, which may result in hardware damage and control errors. In order to solve this problem, we put forward an inertial zeroing neural network. The tracking error of the inertial zeroing neural network is smaller than the tracking error of the general zeroing neural network. In addition, the joint angle jitter of the inertial zeroing neural network is smaller and the robot manipulator changes more smoothly at the initial phase. The results of the simulation confirm the superiority of the initial zeroing neural network.
- 17 November 2023
- Published by Institute of Electrical and Electronics Engineers (IEEE)
- p. 5549-5554
Inertial Zeroing Neural Network for Robot Manipulator Control Described by Different-Level Dynamic Linear System
Enhancing the stability and precision of robot manipulator control represents a pivotal focus in engineering. In this paper, we focus on studying how to optimize the stability and accuracy of robot manipulator control. The proposed approach merges an inertial neural network with a discrete zeroing neural network, resulting in a remarkable reduction of shaking and error. The discrete zeroing neural network model is proposed in the basis of the different-level dynamic linear system. However, this model has the disadvantage of significant oscillations in solving different-level dynamic linear systems with great joint angle jitter and derivative of the change in joint values, which may result in hardware damage and control errors. In order to solve this problem, we put forward an inertial zeroing neural network. The tracking error of the inertial zeroing neural network is smaller than the tracking error of the general zeroing neural network. In addition, the joint angle jitter of the inertial zeroing neural network is smaller and the robot manipulator changes more smoothly at the initial phase. The results of the simulation confirm the superiority of the initial zeroing neural network.
- 17 November 2023
- Published by Institute of Electrical and Electronics Engineers (IEEE)
- p. 8311-8316
Few-Shot Learning Based on Information Geometry
Few-shot classification, as the name suggests, is a very challenging task where only a few training examples are available for each category. Extensive research has been done to solve the problem of the high cost of labeling a large number of samples. A recent research strategy is to find similarity measures between query sets and support sets for judgment. Recently, information geometry techniques have been applied in deep learning research, which explores the relationships and properties between probability distributions by treating them as geometric objects and using geometric tools. Based on information geometry, this paper characterizes differences in the geometric properties of images as differences in probability distributions. Assume that the data satisfies a multivariate Gaussian distribution and that different images are considered as different points on the manifold according to the parameters that characterize the probability distribution of the images. We measure the similarity between different images by the geodesic and validate three methods for estimating the geodesic on a neural network model with 64 neurons at four levels and improve the model performance by Firth bias. The GM, GMF, KM, and KMF methods are validated in the resnet-12 and resnet-18 frameworks using a meta-learning approach on the miniImageNet and CUB datasets, respectively. The experiments show that the models trained by the above four algorithms have good performance and have advantages over the classical algorithms.
- 17 November 2023
- Published by Institute of Electrical and Electronics Engineers (IEEE)
- p. 8311-8316
Few-Shot Learning Based on Information Geometry
Few-shot classification, as the name suggests, is a very challenging task where only a few training examples are available for each category. Extensive research has been done to solve the problem of the high cost of labeling a large number of samples. A recent research strategy is to find similarity measures between query sets and support sets for judgment. Recently, information geometry techniques have been applied in deep learning research, which explores the relationships and properties between probability distributions by treating them as geometric objects and using geometric tools. Based on information geometry, this paper characterizes differences in the geometric properties of images as differences in probability distributions. Assume that the data satisfies a multivariate Gaussian distribution and that different images are considered as different points on the manifold according to the parameters that characterize the probability distribution of the images. We measure the similarity between different images by the geodesic and validate three methods for estimating the geodesic on a neural network model with 64 neurons at four levels and improve the model performance by Firth bias. The GM, GMF, KM, and KMF methods are validated in the resnet-12 and resnet-18 frameworks using a meta-learning approach on the miniImageNet and CUB datasets, respectively. The experiments show that the models trained by the above four algorithms have good performance and have advantages over the classical algorithms.
- 11 May 2022
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Multimedia
- Vol. 25 (15209210) , 4361-4373
APANet: Adaptive Prototypes Alignment Network for Few-Shot Semantic Segmentation
Few-shot semantic segmentation aims to segment novel-class objects in a given query image with only a few labeled support images. Most advanced solutions exploit a metric learning framework that performs segmentation through matching each query feature to a learned class-specific prototype. However, this framework suffers from biased classification due to incomplete feature comparisons. To address this issue, we present an adaptive prototype representation by introducing class-specific and class-agnostic prototypes and thus construct complete sample pairs for learning semantic alignment with query features. The complementary features learning manner effectively enriches feature comparison and helps yield an unbiased segmentation model in the few-shot setting. It is implemented with a two-branch end-to-end network (\ie, a class-specific branch and a class-agnostic branch), which generates prototypes and then combines query features to perform comparisons. In addition, the proposed class-agnostic branch is simple yet effective. In practice, it can adaptively generate multiple class-agnostic prototypes for query images and learn feature alignment in a self-contrastive manner. Extensive experiments on PASCAL-5 $^i$ and COCO-20 $^i$ demonstrate the superiority of our method. At no expense of inference efficiency, our model achieves state-of-the-art results in both 1-shot and 5-shot settings for few-shot semantic segmentation.- 11 May 2022
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Multimedia
- Vol. 25 (15209210) , 4361-4373
APANet: Adaptive Prototypes Alignment Network for Few-Shot Semantic Segmentation
Few-shot semantic segmentation aims to segment novel-class objects in a given query image with only a few labeled support images. Most advanced solutions exploit a metric learning framework that performs segmentation through matching each query feature to a learned class-specific prototype. However, this framework suffers from biased classification due to incomplete feature comparisons. To address this issue, we present an adaptive prototype representation by introducing class-specific and class-agnostic prototypes and thus construct complete sample pairs for learning semantic alignment with query features. The complementary features learning manner effectively enriches feature comparison and helps yield an unbiased segmentation model in the few-shot setting. It is implemented with a two-branch end-to-end network (\ie, a class-specific branch and a class-agnostic branch), which generates prototypes and then combines query features to perform comparisons. In addition, the proposed class-agnostic branch is simple yet effective. In practice, it can adaptively generate multiple class-agnostic prototypes for query images and learn feature alignment in a self-contrastive manner. Extensive experiments on PASCAL-5 $^i$ and COCO-20 $^i$ demonstrate the superiority of our method. At no expense of inference efficiency, our model achieves state-of-the-art results in both 1-shot and 5-shot settings for few-shot semantic segmentation.