所報 107

8/46

1) Node-perturbation Learning Applied for Soft-committee Machine, IPSJ Transaction on Mathematical Modeling and Its Applications, Vol. 13, No. 2, 61-68, Aug. 2020.2) Performance of pre-learned convolution neural networks applied to recognition of overlapping digits, 2020 IEEE International Conference on Big Data and Smart Computing, Feb. 2020.3) Statistical Mechanics of Node-perturbation Learning with Noisy Baseline, Journal of the Physical Society of Japan, 86, 024002 2017.4) Analysis of Function of Rectified Linear Unit Used in Deep learning, International joint conference on neural networks, July 2015.Our research approach is a part of artificial Intelligence by using statistical mechanics. We use online learning which is supervised learning that uses a piece of data with its target in a learning iteration. There is no correlation between data in online learning, so we can apply the statistical mechanics method to analyse online learning behaviour. This analysis shows us how the network approaches the target at speed or error without simulations. Differential equations of order parameter are derived from the learning equation and are used to depict the behaviour of a network. In this case, the learning equation is a microscopic behaviour of the network, and the error of the network is a macroscopic behaviour. To derive the macroscopic equations from microscopic equation, we use the statistical mechanics. Derivation of the macroscopic equations is similar to the derivation of Boyle-Charles’ law from a dynamic equation of the electron.From this approach, we can get an insight into learning, then we can find out why and how the learning system is be-having and how accurate the network is.Kazuyuki Hara received his BEng and MEng degrees from Nihon University in 1979 and 1981, respectively, and his PhD degree from Kanazawa University in 1997. He was engaged at NEC Home Electronics Corporation from 1981 until 1987, after which he joined the Toyama Polytechnic College as a lecturer. In 1998, he joined the Tokyo Metropolitan College of Technology as an associate professor and he became a professor in 2005. He served as a professor at Nihon University in 2010. Dr Hara is a director of the Research Center for Artificial Intelligence, the College of Industrial Technology, and a member of the Nihon University AI Society (NUAIS). His current research interests include the statistical mechanics of on-line learning and the applications of artificial intelligence in visual systems. He is a senior member of the IPSJ and a member of the JPS, IEEE, and IEICE.－　　－6Research AchievementsResearch approachKazuyuki HARAProfessor, Department of Electrical and Electric Engineering