15:15 〜 15:30
▼ [12p-A405-8] Reservoir computing with Single-Walled Carbon Nanotube/Polyoxometalate Random Network
キーワード：Reservoir computing, Single-walled carbon nanotube, Polyoxometalate
Reservoir computing has turned out to be an energy efficient way for permorming brain like real time prediction and classification tasks. Recently the focus has been on implementing hardware platforms with varied physical devices that satisfies the basic reservoir properties of non-linearity, fading memory and high dimensionality. We study these properties using a random network of single-walled carbon nanotube (SWNT)/Porphyrin-polyoxometalate (POr-POM) by performing electrical measurements of the thin film on a glass substrate of Al sputtered multi electrode system. I-V studies showed a non-linear characteristic with pinched hysteresis memory effect due to the redox Por-POM. The higher harmonics via FFT analysis and sine waves of different phases generated at various outputs confirmed the high dimensional property of the device. Owing to the fulfillment of the basic reservoir characteristics a rudimentary reservoir task of waveform generation was performed with a custom built probe instrument controlled by Labview software. Different waves of triangle, square and sawtooth was generated from a sine wave by training the output weights of the reservoir and finally linearly combining them. It was fitted against the target wave and mean square error was calculated (MSE) to get the fitting accuracy. Triangle wave of accuracy 99.8% showed the best fitting with accuracy decreasing for square and sawtooth. We believe that infinite harmonics are required in order to have a better trained high dimensional. Further analysis on increasing the output electrodes and NARMA prediction tasks are being carried out for developing a full-fledged reservoir system that can help in speech recognition task in near future.