Self-Curable Synaptor With Tri-Node Charge-Trap FinFET for Semi-Supervised Learning

Abstract

Semi-supervised learning (SSL) with pseudo-labeling is applied to the non-volatile computing-in-memory (nvCIM) architecture through weight updates of a synaptic transistor (synaptor). The synaptor is a tri-node FinFET enclosing a charge-trap layer. For on-chip training over extended periods, self-curing induced by electrothermal annealing (ETA) is utilized to raise the tunneling oxide temperature of the synaptor until it exceeds 500 ◦C. As a result, a classification accuracy of 86.4% is achieved by training only 1, 000 labeled datasets with self-curing operations. This accuracy level is comparable to that of supervised learning (SL) with 10, 000 labeled training datasets. Not only the MNIST but also the CIFAR-10 dataset was verified whether it yields similar results when using SSL.