[5p-A202-1~18] 6.3 Oxide electronics

[5p-A202-18] Character feature extraction system composed of polymer neural network

〇Wataru Hikita¹, Tetsuya Asai², Yuji Kuwahara¹, Megumi Akai-Kasaya^1,3 (1.Osaka Univ., 2.Hokkaido Univ., 3.JST PRESTO)

Keywords:PEDOT:PSS, electropolymerization, conductive wire

Analog hardware neural networks constructed by nano-scale nonvolatile memristive (synaptic) devices, such as RRAM crossbars, are one of the emerging technology for hardware neural networks. We present a prototype of molecular neural networks consisting of conducting polymer wires in liquid state, where memristive devices exist at every cross point of polymer wires, forms all-to-all connections between bottom and top polymer wire layers. Through experiments, we demonstrate that the molecular networks constructed by conducting polymer PEDOT/PSS [(poly(3,4-ethylenedioxythiophene) doped with poly (styrene sulfonate) anions] growing between 200μm gap Au electrodes recognizes 3 symbols as a result of the auto encoder algorithm. During the training phase, square growth voltage (=8 Vp-p, 10 kHz) was applied to one of the electrode, and the other electrode was grounded, whereas during the evaluation phase, DC voltage (= 1.0 V) was applied to one of the electrode, and the other electrode was virtually grounded by opamp-based current-to-voltage converters, to measure currents between each electrodes.
This result implies that they expand variety of present neuromorphic computing architectures designed mainly for solid-state CMOS devices.