Pulse program for improving learning accuracy and reducing programming energy consumption of ferroelectric synaptic transistor

Abstract

Neuromorphic computing is a next‐generation computing technology featured by parallel data processing and adaptive learning. Two significant factors that improve learning accuracy are the ‘dynamic range’ and ‘linearity’ of the weight update. In a ferroelectric synaptic transistor, the weight update can be modulated by adjusting the applied voltage. The voltage pulse train should be carefully optimized to improve the learning accuracy and reduce programming energy consumption. In this study, we investigated the learning accuracy of neuromorphic computing based on the characteristics of synaptic devices and the program energy consumption according to pulse programs. We demonstrated changes in the analog conductance characteristics of ferroelectric thin‐film transistors by varying the pulse program for synaptic plasticity, discussed the characteristics for improving learning accuracy, and compared the programming energy consumption according to the pulse programs. We proposed a logarithmic‐incremental‐step pulse program that reduces programming energy consumption and improves learning accuracy.