On-Chip Adaptation for Reducing Mismatch in Analog Non-Volatile Device Based Neural Networks

Analog non-volatile devices are gaining prominence in computational applications due to their potential for enhanced energy efficiency and higher density compared to conventional memory technologies. This inherent advantage positions them as a promising solution for edge devices. However, these devices are susceptible to intrinsic mismatch and variation, which can severely compromise the overall accuracy of neural networks implemented with them. Our on-chip measurements reveal that variations and mismatch can lead to a significant reduction in neural network accuracy (> 10% accuracy drop). In response to these challenges, this work introduces an innovative on-chip adaptation mechanism leveraging hot-electron injection to address the issue of mismatch and variation. Through extensive experimentation, we demonstrate that our proposed method achieves a substantial reduction in overall mismatch (standard deviation of 1.23% ). This reduction in mismatch, in turn, results in a noteworthy enhancement in the accuracy of analog neural networks. By addressing the critical issue of mismatch through on-chip adaptation, our research contributes to the development of more robust and accurate analog non-volatile device-based neural networks, paving the way for their effective deployment in edge computing and other resource-constrained applications.