FPGA Implementation of Reservoir Computing Based on Pseudo-Billiard Dynamics in Hypercube

Abstract

Reservoir computing (RC) is one of the frameworks of the recurrent neural network (RNN) and is applied to processing of time series data. The implementation of machine learning and neural networks generally demands large computational resources (or circuit resource) and power. Katori et al. proposed a reservoir computing model based on pseudo-billiard dynamics on a hypercube. This hypercube-based reservoir computing (HRC) can be implemented with less circuit resource and with low power consumption. In this study, we improve the HRC model based on the hardware-oriented algorithm that reduces the circuit resource consumption in digital circuit implementation with Field Programmable Gate Array (FPGA). We evaluate the proposed model with time series generation tasks and confirm that the accuracy of the time-series generation is comparable with the previously proposed model. This research may enhance the ability of RC and contribute to establish a new platform for artificial intelligence.