[10p-W810-1~9] Science of the Material Intelligence: Bringing out the Intrinsic Learning and Optimization Capabilities of Materials

[10p-W810-5] Learning Material -Building Neural Network-

〇Megumi Akai-Kasaya^1,2 (1.Osaka Univ., 2.JST PRESTO)

Keywords:conductive polymer, neural network, machine learning

Custom-designed neuromorphic hardware specialized for an artificial neural network using emerging material memory devices, i.e., crossbar array consisting of two-terminal metal oxide memristors, is now actively developed. A future of which various stage of neuromorphic hardware is widely used in edge technologies is a clear and present demand of society. Appearance of new synaptic devices consisting of organic material with unique benefits has recently been reported, though no neural network learning were performed yet. We here present a prototype of molecular neural network consisting of conducting polymer wires. The wire grows and bridges between electrodes immersed in monomer solution increasing their conductance that can be stuffily kept as a memristive change memory. The polymer wires can be taught to perform data encoding function realizing feature extraction of three 3×3 binary letters through un autoencoder that is un-supervised learning. This demonstration broadens variety of material and framework to realize physical neural network.