Presentation Information
[18a-PA2-14]Real-Time Hardware System Using Ag2Se Atomic Switching Network Reservoir Computing Device
〇(D)Ahmet KARACALI1, Takumi Kotooka1, Yuki Usami1,2, Hirofumi Tanaka1,2 (1.LSSE, Kyutech, 2.Nemorph Center, Kyutech)
Keywords:
Reservoir Computing,Real-Time Processing,Material AI
Traditional neural networks consume significant power due to the complex computations involved in processing signals across multiple layers. An appealing computational framework called reservoir computing (RC), which transforms the input signal into a higher-dimensional environment, provides a notable benefit compared to artificial neural network-based systems by substantially simplifying the learning process, as training is only needed for the readout layer. The utilization of physical implementations of RC has been noted for its ability to perform fast computations while consuming minimal power. In a previous study, we processed analogue signals in real time using simple OPAMP circuits with an Ag2Se atomic switching device. [1] In this study, we designed a real-time circuit that processes the same nanomaterial digitally with the help of an FPGA, achieving both lower power consumption and better performance. Fig. 1 shows the printed circuit board containing two MCP3008 ADC modules, two MX7545 DAC modules, and two regulator circuits that determine the reference voltage for these modules connected to the Cyclone IV FPGA module. The FPGA module multiplies and accumulates the 10-bit Ag2Se outputs fed by the ADCs using weight parameters previously trained by linear regression, thereby generating 12-bit outputs for the DAC module. A waveform generation task was performed to test the system, using a sine waveform to generate different waveforms. An 11 Hz sinusoidal wave was fed into the BNC component at the top of Fig. 1, and the DAC outputs were read from the BNC module below. Figure 2 shows the results of generating sawtooth and triangle waveforms using the waveform generation task. As expected, in both waveforms higher accuracy rates are demonstrated with the digital system. Furthermore, by performing the waveform generation task with a power consumption of 26 mW, the digital system shows promise for performing more challenging tasks.