[4Rin1] Interactive 2

[4Rin1-83] Image recognition model based on convolutional reservoir computing

〇Yoshihiro Yonemura¹, Yuichi Katori^1,2 (1.Future University Hakodate, 2.The Institute of Industrial Science, The University of Tokyo)

Keywords:reservoir computing, image recognition

The convolutional neural network (CNN) achieves high accuracy score in image recognition tasks by extracting spatial features of input images and can be utilized to construct a classifier by combining with other machine learning techniques. However, the computational cost of CNN is not small because its learning rule requires to configures large number of parameter values. Reservoir computing (RC) generates complex time series with simple learning rule. Recently, the mixture model of CNN and RC has been proposed. Computational costs and the number of learning parameters can be reduced with RC on the mixture model. Some RC based image recognition models have been proposed in previous studies, but properties of those models are unclear. In this research, we construct a model composed of CNN and RC and analyze its performance. We confirmed high recognition accuracy of the model and the relationships between parameter values and the recognition accuracy.