Performance of Crossbar based Long Short Term Memory with Aging Memristors

The Long Short Term Memory (LSTM) neural networks find a wide range of applications in time series prediction problems. The long-term accuracy and reliability of LSTM memristor crossbar array are subjected to the memristor device's endurance and failures. Memristor aging and its impact on such LSTM's performance is an open problem. This paper analyzes the effects of different types of aging typically exhibited in memristor devices on the crossbar performance. The performance results are analyzed on two datasets, (1) SMS Spam and (2) IMDB movie review. Our analysis indicated that the different aging type shows different performance deterioration levels in the crossbar based LSTM system. Here, the aging analysis for oxide-based memristor implementation are primarily considered when used in CMOS-Memristor hybrid crossbars.