講演情報
[16p-K306-3]FPGA Based Physical Reservoir Computing with Ag2Se Atomic Switching Network Device
〇(D)Ahmet KARACALI1, Yuki Usami1,2, Hirofumi Tanaka1,2 (1.LSSE, Kyutech, 2.Neumorph Center, Kyutech)
キーワード：
Reservoir Computing、Atomic Switching Network、Material AI
Although machine learning has advanced significantly, its performance is physically limited by hardware constraints like transistor size and power consumption. Reservoir Computing (RC) offers a more efficient alternative, requiring only the readout layer to be trained and enabling faster, low-power computations. This work defines a low-power digital application using a nanomaterial device, namely a material random network, as a physical reservoir—a technique known as in-material reservoir computing. Here, we focused on the Ag2Se Atomic Switching Network (ASN) device, which exhibits nonlinear switching and memory characteristics with its atomic switch junctions and is tested with a waveform generation task. An 11 Hz sinusoidal wave was applied to the 16-electrodes Ag2Se nanowire device, producing 15 nonlinear outputs and converted digital bits using MCP3008 ADC. These outputs were trained using linear regression to create various waveforms. The resulting weight values were then implemented in the following hardware. Each weight was converted into a 16-bit digital signal and summed with the device's digital outputs to generate a 24-bit waveform using FPGA in Sequential blocks for the waveform generation task, as illustrated in Figure 1. Simulated Results in the test bench Verilog HDL, were compared with the analog weight values. The analog weight results for waveform generation were highly accurate, with 99.25% for cosine and 80.33% for sawtooth. The results acknowledged by digital signals were also impressive, with 98.98% for cosine and 69.39% for sawtooth using only one digital signal processing unit in FPGA. This study demonstrates the real-time use of a physical RC device while consuming low power as 1.709 W with using only one digital signal processing unit in FPGA, and the potential to produce low-power-consuming application specific integrated circuits (ASICs) using physical RC devices is promising.