Effect of Stochastic Resonance on Classification Accuracy of Neural Networks Utilizing Inherent Stochasticity in Threshold Voltage of Ovonic Threshold Switching Device

In this study, stochastic resonance (SR) exploits the inherent stochastic characteristics of the OTS threshold voltage to enhance the inference performance of neural networks. First, the threshold switching of the OTS device is characterized, and a signal detection using an OTS device is proposed. Next, we investigate the impact of stochasticity in the threshold voltage on detecting weak signals in the SR system. Finally, by evaluating the inference performance of the artificial neural network, we confirm that the inherent stochasticity can effectively restore the degraded MNIST image in poor visibility conditions in the OTS device. As a result, the recognition accuracy was improved from 10.28% to 95.78% when the stochasticity characteristic was reflected. These results show that stochasticity in the device can improve system performance.