Reliability Enhancements in Memristive Neural Network Architectures

• Published in: IEEE transactions on nanotechnology Vol. 18; pp. 866 - 878
• Main Authors: Gnawali, Krishna Prasad, Paudel, Bijay Raj, Mozaffari, Seyed Nima, Tragoudas, Spyros
• Format: Journal Article
• Language: English
• Published: New York IEEE 2019; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: Algorithms; Back propagation; Biological neural networks; Computer architecture; Learning algorithm; Machine learning; Memristive Crossbar Array; Memristors; Microprocessors; Multi-layer neural network; Network reliability; Neural networks; Neuromorphic Computing; Neurons; Non-Volatile Memory; Semiconductor devices; Stochastic Gradient Descent; System-on-chip; Transistor Aging; Transistors
• ISSN: 1536-125X; 1941-0085
• DOI: 10.1109/TNANO.2019.2933806

Memristive Crossbar Arrays (MCAs) are widely used in designing fast and compact neuromorphic systems. However, such systems require on-chip implementation of the backpropagation algorithm to accommodate process variations. This paper proposes a low hardware overhead on-chip implementation of the backpropagation algorithm that utilizes effectively the very dense MCAs. On-chip learning using the proposed architecture increases the reliability of the neuromorphic system in the presence of process variations in the neural component. The second contribution of this paper is an architectural enhancement to cope with another reliability consideration, namely the aging transistors in the MCA. Experimental results show the impact of reliability enhancement.