FURUKAWA Tetsuo

写真a

Title

Professor

Laboratory

2-4 Hibikino, Wakamatsu-ku, Kitakyushu-shi, Fukuoka

Research Fields, Keywords

E-mail

E-mail address

Phone

+81-93-695-6124

Post Graduate Education 【 display / non-display

  • 1989.03  Osaka University  Graduate School, Division of Engineering Science  Master's Course  Completed  JAPAN

Degree 【 display / non-display

  • Osaka University -  Doctor of Engineering  1998.10

Biography in Kyutech 【 display / non-display

  • 2014.04
    -
    Now

    Kyushu Institute of TechnologyGraduate School of Life Science and Systems Engineering   Department of Human Intelligence Systems   Professor  

  • 2006.04
    -
    2014.03

    Kyushu Institute of TechnologyGraduate School of Life Science and Systems Engineering   Department of Brain Science and Engineering   Professor  

 

Publications (Article) 【 display / non-display

  • Space-And-Cost-Efficient Neural Control /Sensory Element Using an Analog FPGA

    Kobayashi F., Furukawa T.

    Proceedings of 2019 International Conference on System Science and Engineering, ICSSE 2019      71 - 74   2019.07  [Refereed]

     View Summary

    © 2019 IEEE. As neural networks are applied to control and sensory, software neuron models cannot sometimes fulfill speed requirement as well as simple add-And-sigmoid is not enough for functionality. This paper proposes a small and inexpensive hardware neuron based on FitzHugh-Nagumo model. By making full use of chip property and maximum circuit packing through placement search, it surpasses the previous implementation by factors of 4 for space and 15 for cost.

    DOI Scopus

  • Tensor self-organizing map for kansei analysis

    Proceedings of Joint 10th International Conference on Soft Computing and Intelligent Systems and 19th International Symposium on Advanced Intelligent Systems in conjunction with Intelligent Systems Workshop 2018      796 - 801   2019.05  [Refereed]

    Japan  Toyama  2018.12  -  2018.12

     View Summary

    © 2018 IEEE. In Kansei analysis, impressions of various objects are commonly measured using evaluation words. When using this approach, it is necessary to examine all combinations of three elements: subjects, objects, and evaluation words. However, the exhaustive analysis required is not an easy task because of the enormous number of combinations. Additionally, if it is necessary to reveal the relationship between the impressions and physical features of objects such as colors or shapes, the number of combinations increases enormously and the task becomes unrealistic. In this paper, we introduce a method called the tensor self-organizing map (TSOM) that visualizes the relationships between the elements. We applied the TSOM to Kansei analysis of landscape images and studied how the impressions were dependent on the subjects. We also investigated the relationships between these subject-dependent impressions and the physical features. Through these experiments, we demonstrate that the TSOM can be a useful tool for Kansei analysis.

    DOI Scopus

  • Distance metric learning for the self-organizing map using a co-training approach

    Yoneda K., Furukawa T.

    Proceeding of ICICIC2018      CD-ROM   2018.12  [Refereed]

    China 

     View Summary

    The aim of this work is to develop a method of distance metric learningfor self-organizing maps. We rst conducted an investigation in a multi-view learningsetting, in which Mahalanobis metrics were determined so that two (or more) viewsreached a consensus in latent variable estimation. We examined two approaches of multi-view learning: co-training and ensemble. Although both approaches worked as expected,our results suggested that the co-training approach performed better. We further extendedthe method to a single-view learning setting by introducing the concept of pseudo multi-view learning.

  • Distance metric learning for the self-organizing map using a co-training approach

    Yoneda K., Furukawa T.

    International Journal of Innovative Computing, Information and Control    14 ( 6 ) 2343 - 2351   2018.12  [Refereed]

     View Summary

    The aim of this work is to develop a method of distance metric learningfor self-organizing maps. We rst conducted an investigation in a multi-view learningsetting, in which Mahalanobis metrics were determined so that two (or more) viewsreached a consensus in latent variable estimation. We examined two approaches of multi-view learning: co-training and ensemble. Although both approaches worked as expected,our results suggested that the co-training approach performed better. We further extendedthe method to a single-view learning setting by introducing the concept of pseudo multi-view learning.

    DOI Scopus

  • Simultaneous analysis of subjective and objective data using coupled tensor self-organizing maps: Wine aroma analysis with sensory and chemical data

    International Conference on Neural Information Processing    11306 LNCS   24 - 35   2018.12  [Refereed]

    Cambodia  Siem Reap  2018.12  -  2018.12

     View Summary

    © Springer Nature Switzerland AG 2018. In this paper, we propose a method for simultaneous analysis of subjective and objective data. The method, named coupled tensor self-organizing map (SOM), consists of two tensor SOMs, one of which learns the subjective data while the other learns the objective data. The coupled tensor SOM visualizes the dataset as three maps, namely, one target object map, and two survey item maps corresponding to the subjective and objective data. This method can be further extended to generate extra maps such as a map of attributes. In addition, the coupled tensor SOM also provides an interactive visualization of the relationship between the target objects and the survey items by coloring these three maps. We applied our proposed method to the wine aroma dataset. Our results indicate that this method facilitates an intuitive overview of the dataset.

    DOI Scopus

display all >>

Conference Prsentations (Oral, Poster) 【 display / non-display

  • Visualization of document-word relation by modeling their joint probability on latent spaces

    Takuro Ishida, Keisuke Yoneda, Hajime Hatano, Tetsuo Furukawa

    (Miyakojima Marine Terminal )  2020.01  -  2020.01 

     View Summary

    The aim of this work is to visualize the relation of documents and words by embedding them to the product space of latent spaces. To achieve the task, the proposed method models the simultaneous probability of documents and words as the one on the product latent spaces. We applied the method to visualize the conference papers of NeurIPS and showed how it visualize the document set.

  • Visualization of children's interactions in the group discussion by Tensor SOM

    Keisuke Kusumoto, Keiichi Horio, Tetsuo Furukawa

    (Miyakojima Marine Terminal )  2020.01  -  2020.01 

     View Summary

    Our aim is to visualize the relationship between children and their social developmental states in a kindergarten. In this work, we focused on the imitation behaviors among the children during the group discussion sessions. We visualized the leading--following frequencies of imitation behaviors by using the tensor self-organizing map.

  • Optimal Transport based Autoencoder for class and style Disentanglement

    Florian Tambon, Tetsuo Furukawa

    (Miyakojima Marine Terminal )  2020.01  -  2020.01 

  • Visualization tool for basketball team performance by multi-level SOM

    Kanta Senoura, Hideaki Ishibashi, Tetsuo Furukawa

    (Miyakojima Marine Terminal )  2020.01  -  2020.01 

     View Summary

    The purpose of this work is to develop a method to visualize the relation between the team performance and the member composition. The proposed method employs Multi-Level Self-Organizing Map (SOM) to analyze both at the member level and at the team level, and then their relation is modeled by regression. The proposed method was applied to the game data of National Basketball Association.

  • Visualization of Relational data by Embedding to Direct Product Space

    Kazuki Miyazaki, Ryuji Watanabe, Tetsuo Furukawa

    (Miyakojima Marine Terminal )  2020.01  -  2020.01 

     View Summary

    The aim of this work is to develop a modeling method of relational data. Relational data is a dataset observed obtained from object pairs belonging to several domains. The task of the proposed method is to embed the objects into the direct product latent spaces, each of which is corresponding to the domain. In this work, we employed the kernel soother to represent the product manifold, and it estimates the latent variables by a non-parametric unsupervised manner.

display all >>

Honors and Awards 【 display / non-display

  • SCIS&ISIS2018 with ISWS2018 Poster Session Award

    2018.12.12     JAPAN

    Winner: Kaido Iwamoto, Tohru Iwasaki, Tetsuo Furukawa

  • IEEE-young award

    2016.09.01     JAPAN

    Winner: Tohru Iwasaki, Tetsuo Furukawa

  • IEEE Computational Intelligence Society Japan Chapter Young Researcher Award

    2015.03.16     JAPAN

    Winner: Yuki Toshima, Tetsuo Furukawa

     View Summary

    Multimode data (relational data) is generally expressed as a tensor. In the analysis of tensor data, not only analysis of individual modes but also relations between modes need to be analyzed. In actual data analysis, it is often the case that multiple tensor datasets are combined by sharing some modes to form a larger data complexe. In this case, inter-mode analysis across multiple tensor datasets is also required, making analysis more difficult. The purpose of this research is to develop a visualization method to grasp the overall image of complex tensor data. In this research, we constructed a Tensor SOM Network that combines multiple Tensor SOMs, and further we developed a information propagation method for the inter-mode visualization belonging to different tensor data.