Daisuke Kimura
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Daisuke Kimura has a total of 22 co-authors. Here are the top co-authors by number of shared publications:
- Naoyuki Matsumoto
Technology Research Association for Single Wall Carbon Nanotubes, National Institute of Advanced Industrial Science and Technology Ibaraki, AIST Tsukuba Central 5, 1-1-1 Higashi, Tsukuba, 305-8565, JAPAN
2 shared publications Ayaka Namekawa
Furukawa Electric Co., Ltd., 4-3-2, Okano, Nishi, Yokohama, Kanagawa 220-0073, Japan
1 shared publication- Hiroaki Horiyama
National Institute of Advanced Industrial Science and Technology (AIST), Multi-Material Research Institute , Nagoya , Aichi , 463-8560 , Japan
1 shared publication - Hirokazu Nosato
National Institute of Advanced Industrial Science and Technology (AIST)
1 shared publication Kai Ikoma
Department of Mechanical and Control Engineering, Kyushu Institute of Technology, 1-1 Sensui-cho, Tobata-ku, Kitakyushu, Fukuoka 804-8550, Japan
1 shared publication- Kazufumi Kobashi
National Institute of Advanced Industrial Science and Technology
1 shared publication
Latest Publications
- Open Access
- 26 August 2026
- Published by Elsevier in SSRN Electronic Journal
Design Conditions of Convolutional Neural Networks for Theoretically Justified Visualization of Image Regions Contributing to Predictions
Many gradient-based class activation mapping (CAM) methods have been proposed to visualize image regions that contribute to CNN predictions. However, these methods often involve heuristic approximations, and their visualizations are not theoretically guaranteed to accurately reflect the actual contributions to model outputs. We derive CNN design conditions for theoretically justified visualization based on a theoretical framework exploiting the piecewise linearity of CNNs with ReLU activation functions. In this framework, each output can be explicitly decomposed into contributions from the intermediate feature and biases. This decomposition reveals that biases in convolutional layers and implicit biases introduced by batch normalization induce spatially structured contributions and therefore may be non-negligible in explaining the basis of CNN predictions. To address this issue, biases with spatially structured contributions are excluded by design. This enables the theoretically justified visualization of image regions that contribute positively and negatively to predictions without introducing approximations. To validate the effectiveness of the proposed method, we conduct experiments using synthetic geometric datasets with explicitly defined contributions in both regression and classification tasks. The results demonstrate that excluding biases with spatially structured contributions substantially improves the agreement between the visualized and predefined contributions.
- 26 August 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessDesign Conditions of Convolutional Neural Networks for Theoretically Justified Visualization of Image Regions Contributing to Predictions
Many gradient-based class activation mapping (CAM) methods have been proposed to visualize image regions that contribute to CNN predictions. However, these methods often involve heuristic approximations, and their visualizations are not theoretically guaranteed to accurately reflect the actual contributions to model outputs. We derive CNN design conditions for theoretically justified visualization based on a theoretical framework exploiting the piecewise linearity of CNNs with ReLU activation functions. In this framework, each output can be explicitly decomposed into contributions from the intermediate feature and biases. This decomposition reveals that biases in convolutional layers and implicit biases introduced by batch normalization induce spatially structured contributions and therefore may be non-negligible in explaining the basis of CNN predictions. To address this issue, biases with spatially structured contributions are excluded by design. This enables the theoretically justified visualization of image regions that contribute positively and negatively to predictions without introducing approximations. To validate the effectiveness of the proposed method, we conduct experiments using synthetic geometric datasets with explicitly defined contributions in both regression and classification tasks. The results demonstrate that excluding biases with spatially structured contributions substantially improves the agreement between the visualized and predefined contributions.
- Open Access
- 14 May 2026
- Published by Elsevier in Journal of Environmental Chemical Engineering
- Vol. 14 (3)
Multidimensional correlation analysis between charcoal structure and microbial colonization toward functional biochar design
- 14 May 2026
- Published by Elsevier in Journal of Environmental Chemical Engineering
- Vol. 14 (3)
Open AccessMultidimensional correlation analysis between charcoal structure and microbial colonization toward functional biochar design
- 2 March 2026
- Published by IOP Publishing in Japanese Journal of Applied Physics
- Vol. 65 (5) , 050901
High-conductivity carbon nanotube fibers via wet spinning of long carbon nanotubes
Carbon nanotube (CNT) fibers are promising lightweight conductors, yet their electrical conductivity remains limited compared with metallic benchmarks. Here, we demonstrate the fabrication of highly aligned and densely packed CNT fibers by wet spinning CNTs with long effective lengths. Wet-spun fibers prepared from chlorosulfonic acid dispersions exhibit an as-spun conductivity of 11.4 MS m −1 . Subsequent doping with fuming nitric acid increases the conductivity to 15.9 MS m −1 , representing one of the highest values reported for wet-spun CNT fibers. These results establish effective nanotube length and high packing density as key design parameters for high-performance lightweight conductors.
- 2 March 2026
- Published by IOP Publishing in Japanese Journal of Applied Physics
- Vol. 65 (5) , 050901
High-conductivity carbon nanotube fibers via wet spinning of long carbon nanotubes
Carbon nanotube (CNT) fibers are promising lightweight conductors, yet their electrical conductivity remains limited compared with metallic benchmarks. Here, we demonstrate the fabrication of highly aligned and densely packed CNT fibers by wet spinning CNTs with long effective lengths. Wet-spun fibers prepared from chlorosulfonic acid dispersions exhibit an as-spun conductivity of 11.4 MS m −1 . Subsequent doping with fuming nitric acid increases the conductivity to 15.9 MS m −1 , representing one of the highest values reported for wet-spun CNT fibers. These results establish effective nanotube length and high packing density as key design parameters for high-performance lightweight conductors.