講演情報
[8p-N304-12]Voice Classification Task with Using a Physical Reservoir Computing Device with Posit Extraction
〇Ahmet KARACALI1, Muzhen Xu1,2, Yuki Usami1,2, Hirofumi Tanaka1,2 (1.LSSE, Kyutech, 2.Nemorph Center, Kyutech)
キーワード：
Physical Reservoir Computing、Posit Number Extraction、Material AI
The rapid growth of machine learning has been accompanied by increasing hardware demands, with performance gains constrained by transistor scaling limits and rising energy consumption. Reservoir computing, offers a promising alternative computational framework by confining training to a simple readout layer, thereby reducing both computational overhead and power requirements. In this work, we propose a low-power digital platform that exploits a nanomaterial-based material random network as a physical reservoir computing (PRC) device within an material architecture. Also, the posit extraction method, it efficiently provides processing power and high accuracy utilising FPGA hardware resources, and in combination with the PRC device, enables a very compact, low-power single-loop system. The reservoir functionality is realized using Yttrium Manganese Oxide (YMnO3), a distinctive ferroelectric material whose network of semiconductive domains and domain walls provides rich nonlinear dynamics in reservoir layers. In the previous study, this PRC device was used to perform speech classification[1]. Current studies have explored digital implementations to facilitate deployment on a unified hardware platform and enable real-time processing. Firstly, raw audio signals, which had undergone reflection and down sampling, were fed into the PRC device to obtain 31 different non-linear outputs. These outputs were then trained using a linear regression algorithm to derive the weight values. Finally, the device outputs and weight values were converted to 16-bit posit representation, and classification was performed by executing MAC operations as shown in Figure 1. As a result, as shown in Figure 2, we can see that the posit system successfully performs the classification process and that the bit system does not cause any data loss. In the next stage, we aim to implement real-time classification by programming the posit classification system on an FPGA.