Yan-Jen Huang
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Yan-Jen Huang has a total of 2 co-authors. Here are the top co-authors by number of shared publications:
- Hsin-Lung Wu
Department of Computer Science and Information Engineering, National Taipei University, New Taipei City, Taiwan
2 shared publications Ching- Chen
Department of Computer Science and Information Engineering, National Taipei University, New Taipei City, Taiwan
1 shared publication
Latest Publications
- Open Access
- 25 January 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 12, 21559-21568
Block-Wise Separable Convolutions: An Alternative Way to Factorize Standard Convolutions
In this paper, we introduce block-wise separable convolutions (BlkSConv) to replace the standard convolutions for compressing deep CNN models. First, BlkSConv expresses the standard convolutional kernel as an ordered set of block vectors each of which is a linear combination of fixed basis block vectors. Then it eliminates most basis block vectors and their corresponding coefficients to obtain an approximated convolutional kernel. Moreover, the proposed BlkSConv operation can be efficiently realized via a combination of pointwise and group-wise convolutions. Thus the constructed networks have smaller model size and fewer multiply-adds operations while keeping comparable prediction accuracy. We also develop a hyperparameter search framework based on principal component analysis (PCA) to determine a qualified hyperparameter setting of the block depth and number of basis block vectors. By this search framework, we construct networks which achieve nice prediction performance while simultaneously satisfying the constraints of model size and model efficiency. Our code, data, and models are available at https://github.com/yanjenhuang/blksconv.
- 25 January 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 12, 21559-21568
Open AccessBlock-Wise Separable Convolutions: An Alternative Way to Factorize Standard Convolutions
In this paper, we introduce block-wise separable convolutions (BlkSConv) to replace the standard convolutions for compressing deep CNN models. First, BlkSConv expresses the standard convolutional kernel as an ordered set of block vectors each of which is a linear combination of fixed basis block vectors. Then it eliminates most basis block vectors and their corresponding coefficients to obtain an approximated convolutional kernel. Moreover, the proposed BlkSConv operation can be efficiently realized via a combination of pointwise and group-wise convolutions. Thus the constructed networks have smaller model size and fewer multiply-adds operations while keeping comparable prediction accuracy. We also develop a hyperparameter search framework based on principal component analysis (PCA) to determine a qualified hyperparameter setting of the block depth and number of basis block vectors. By this search framework, we construct networks which achieve nice prediction performance while simultaneously satisfying the constraints of model size and model efficiency. Our code, data, and models are available at https://github.com/yanjenhuang/blksconv.
- Open Access
- 16 November 2023
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 11, 130306-130318
New Coarse-to-Fine Approaches for Age Estimation Based on Separable Convolutions
In this paper, we study lightweight age estimation methods based on a coarse-to-fine approach in which the network performs age prediction with multiple stages. In each stage, the network only focuses on refining the coarse age prediction generated from the previous stage. The final age prediction is the combination of all staged prediction values. We observe that these stages have a causal relationship, that is, the output of each stage is highly correlated with outputs of its former stages. Thus, each stage should share the information of its previous stage before making a refined prediction. Based on this observation, we construct a new compact CNN model called Homologous Stagewise Regression Network (HSR-Net). In HSR-Net, each stage shares the information of the last convolutional layer and then generates its own refined value. In addition, HSR-Net also addresses the age group ambiguity problem by utilizing an easy dynamic range construction. In order to enhance the prediction performance of HSR-Nets, it is naive to increase the number of kernels in each convolutional layer of HSR-Nets. However, the constructed HSR-Net has extremely large parameter size. To address this problem, we propose the separable HSR-Nets (SepHSR-Nets) where standard convolutions are replaced by depth-wise separable convolutions in the convolutional layers of HSR-Nets. In general, the parameter size of SepHSR-Nets ranges from 10K to 75K without sacrificing prediction performance. Experimental results show that SepHSR-Nets achieve competitive performance compared with the state-of-the-art compact models. Our code, data, and models are available at https://github.com/yanjenhuang/hsr-net.
- 16 November 2023
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 11, 130306-130318
Open AccessNew Coarse-to-Fine Approaches for Age Estimation Based on Separable Convolutions
In this paper, we study lightweight age estimation methods based on a coarse-to-fine approach in which the network performs age prediction with multiple stages. In each stage, the network only focuses on refining the coarse age prediction generated from the previous stage. The final age prediction is the combination of all staged prediction values. We observe that these stages have a causal relationship, that is, the output of each stage is highly correlated with outputs of its former stages. Thus, each stage should share the information of its previous stage before making a refined prediction. Based on this observation, we construct a new compact CNN model called Homologous Stagewise Regression Network (HSR-Net). In HSR-Net, each stage shares the information of the last convolutional layer and then generates its own refined value. In addition, HSR-Net also addresses the age group ambiguity problem by utilizing an easy dynamic range construction. In order to enhance the prediction performance of HSR-Nets, it is naive to increase the number of kernels in each convolutional layer of HSR-Nets. However, the constructed HSR-Net has extremely large parameter size. To address this problem, we propose the separable HSR-Nets (SepHSR-Nets) where standard convolutions are replaced by depth-wise separable convolutions in the convolutional layers of HSR-Nets. In general, the parameter size of SepHSR-Nets ranges from 10K to 75K without sacrificing prediction performance. Experimental results show that SepHSR-Nets achieve competitive performance compared with the state-of-the-art compact models. Our code, data, and models are available at https://github.com/yanjenhuang/hsr-net.