講演情報
[19p-P06-4]Delaying and multiplying Analog Input Signal at In-Materio Reservoir Computing
〇(M1)Takuya Kawabata1, Yuki Usami1, Hirofumi Tanaka1 (1.Kyutech)
キーワード：
reservoir computing、in-materio reservoir、delay
In-materio reservoir computing (IMRC) is a prospective non-von Neuman approach in machine learning. Reservoir computing (RC) can effectively process time series data using a low number of parameters and a simple learning algorithm. IMRC can further reduce energy consumption by replacing the reservoir layer with nanomaterial. However, convenient adjusting of the internal state of the reservoir is required for high IMRC performance. Because in most IMRCs, the internal condition, such as the number of nodes, node weight, and node function, is clamped after material reservoirs are fabricated, most IMRCs keep memory capacity (MC) low. One of the candidates to increase the MC is using a signal processing system to delay the input signals. To avoid using analog-digital and digital-analog converters for the delay system from the view of power consumption, we demand a more efficient method for IMRC to decrease power consumption to be an alternative to von Neuman's architecture. In the present work, we aim to suggest a power-friendly way for IMRC by replicating and delaying analog input signals. We could successfully increase the MC sufficiently by using all-pass filters to implement the pre-processing. Moreover, the accuracy of 32.1 % was significantly raised to 85.3 % by using all-pass filters at ten nonlinear autoregressive moving average (NARMA) tasks. The method without converters could work not only as short-term memory but also as reservoir parameters and will improve material reservoir performance and power consumption.