Event Information

HOME > event > ACCMS seminar “Application of machine learning for natural and computational science”

ACCMS seminar “Application of machine learning for natural and computational science”

Post date:2017/08/23

Academic Center for Computing and Media Studies (ACCMS) holds a seminar once in a month. We inviting active researchers from a variety of areas, and ask them to give talks about their research activity or issues in performing research, hoping to provide an opportunity for fruitful discussion among participants.
On an ACCMS seminar at 17 Oct. 2017, we invite Dr. Tokunaga at Kyushu Institute of Technology and Prof. Takami at Oita University. We hope to have a lot of participants.

Date 2017/10/17 (Tuesday)16:30~18:30

South Building of ACCMS 2nd floor, Room 202

Fee Free
Application Not necessary

Academic Center for Computing and Media Studies, Kyoto University


Keiichiro FUKAZAWA, Academic Center for Computing and Media Studies, Kyoto University
Phone 075-753-7430
E-mail : fukazawa * media.kyoto-u.ac.jp (Please replace “*” with “@”.)


Speaker Terumasa Tokunaga (Department of Systems Design and Informatics, Faculty of Computer Science & Systems Engineering, Kyushu Institute of TechnologyAssociate Professor
Title Applications of Bayesian/Sparse inferences for spatiotemporal pattern understanding
Abstract Recent frameworks of deep learning, typified by a 3D convolution neural network (3D-CCN) model, have been succeeded in representing motions features. They have led to growing interest in supervised machine learning approaches for spatiotemporal pattern recognition tasks, including motion predictions and action recognitions. In and around the fields of earth science and biology, vast numbers of observation data have been stored as sequential multimodal images. A successful use of these big imaging data for driving scientific discovery is one of key roles of data science. The representation learning frameworks might be a powerful tool for classifying phenomena of interest into several meaningful groups. However, supervised techniques for machine learning require a large amount of training data, and it consequently poses unrealistic workloads for human experts. A Bayesian approach is expected to be a most promising approach to understand the spatiotemporal patterns. In past five years, we have been developing unsupervised or semi-supervised techniques for analyzing the sequential images based on Bayesian inferences and sparse estimations motivated by some collaborative researches. In this seminar, we introduce recent applications on neuroscience. We will talk about the project scope and discuss some technological issues towards spatiotemporal pattern understanding. (The talk will be given in Japanese.)

Speaker Toshiya Takami (Oita University, Faculty of Science and TechnologyProfessor
Title Numerical Calculation of PDE using Neural Networks
Abstract More and more application areas of deep neural networks are explored after remarkable achievements in the field of image recognition.  There are many fascinating features in neural networks: network weights are automatically determined by supervised learning with a huge number of labeled training data; only small computing resources are necessary in the applied phase with trained weights of networks.  Because of these attractive points, several trials have been done even in numerical computing areas.  In this talk, after introducing these studies, possible methods to compute partial differential equations are considered. (The talk will be given in Japanese.)


We have an elevator for physically handicapped persons. Please contact the office.

Copyright © Academic Center for Computing and Media Studies, Kyoto University, All Rights Reserved.