Inherent Stochasticity of Ovonic Threshold Switch for Neuronal Dropout of Edge-AI Hardware

For efficient training of embedded Edge-AI, we demonstrate a compact neuronal dropout by exploiting the inherent stochasticity of a highly scalable ovonic threshold switch (OTS). To implement reliable probabilistic operations, the random nature of OTS is intensively investigated, and a pre-set pulse scheme is presented. We verify successful dropout implementation by a mere 1T3R structure that controls the probability without requiring external circuits. Experimental results reveal its potential for edge applications (i.e., arrhythmia detection) by improving accuracies (up to 11.4 %) with limited training data (100 signals).