2T Neuromorphic Device based on oxide semiconductor with High Linearity and Symmetry for High-Precision Training.

Artificial intelligence technology based on deep neural networks performs a tremendous number of matrix-vector multiplications in the training process. In the current computing architecture, where memory and processing units are completely separated, power consumption and delay during training are enormous. This work proposes a neuromorphic device consisting of two metal oxide semiconductor transistors optimized for neural network training to address this problem. The proposed neuromorphic device controls a conductance of a read transistor by adjusting gate potential with a write transistor. The neuromorphic device can store weights for several minutes and exhibits near-ideal potential/depression characteristics. As the proposed neuromorphic device can provide very high levels of linear/symmetric weight update, it can be applied to AI systems with high training accuracy.