Impact of RTN on Pattern Recognition Accuracy of RRAM-Based Synaptic Neural Network

Abstract

Resistive switching memory devices can be categorized into either filamentary or non-filamentary ones depending on the switching mechanisms. Both types have been investigated as novel synaptic devices in hardware neural networks, but there is a lack of comparative study between them, especially in random telegraph noise (RTN) which could induce large resistance fluctuations. In this letter, we analyze the amplitude and occurrence rate of RTN in both Ta <sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sub> O <sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">5</sub> filamentary and TiO <sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sub> /a-Si (a-VMCO) non-filamentary resistive switching memory (RRAM) devices and evaluate its impact on the pattern recognition accuracy of neural networks. It is revealed that the non-filamentary RRAM has a tighter RTN amplitude distribution and much lower RTN occurrence rate than its filamentary counterpart, which leads to negligible RTN impact on recognition accuracy, making it a promising candidate in synaptic application.