This work is expected to appear in DAC'23 Ph.D Forum
Emerging Non-Volatile Memories (eNVMs) have been demonstrated as promising candidates for the deployment of Deep Neural Networks (DNNs) owing to their superior scalability, non-volatility, and performance. However, current eNVM devices exhibit various non-idealities, which hinder their utilization in highly reliable applications. In this paper, by exploiting the intrinsic fault-tolerability of DNNs, we propose several low-cost and efficient techniques to enhance the robustness of eNVMs-based DNNs.