Summary
International Symposium on Nonlinear Theory and Its Applications
2016
Session Number:A4L-D
Session:
Number:A4L-D-7
Encoding Multi-Dimensional Time Series Data with Reservoir Computing
Yuichi Katori,
pp.-
Publication Date:2016/11/27
Online ISSN:2188-5079
Summary:
Encoding of given time series data can be basis of various information processing and a key to understanding properties of the dynamical system that generate the time series. Efficient encoding requires to extracts features of the data and decompose the given data with the features. Features on time series can be represented as a set of kernel functions, and a given data can be decomposed using the kernel functions and sparsification of the representation. Reservoir computing paradigm provides a strategy to model multi-dimensional time series data with randomly coupled nonlinear elements. Here we propose to combine the shiftable kernel method and the reservoir computing paradigm. We use the echo state network, one of implementation of the reservoir computing, and propose that the reservoir computing can be utilized to dynamically organize the kernel functions and to efficiently encode multi-dimensional time series. We demonstrate that a complex multi-dimensional time series data can be encoded into a few points in the point process with the proposed method.