Fault tolerance of oscillatory neural network: device oscillator based small network to digital oscillator based large networks

Abstract

Oscillatory Neural Networks (ONN) are inevitable when it comes to solving combinatorial optimization problems. ONNs are also extremely energy efficient for AI workloads compared to conventional Deep Neural Networks (DNNs). Analysis of fault and failure tolerance of ONNs is crucial for understanding the reliability of the networks. This work illustrates the fault tolerance of the ONN in solving constraint optimization problems such as vertex coloring and digit recognition problems. For vertex coloring, a 4-node network across various configurations and different component failure levels has been analyzed using a device oscillator. The findings confirm that the network is highly robust to failures, demonstrating tolerance to variations in resistance of up to 40% and in capacitance of up to 60%. The analysis was then extended to bigger networks varying from 16-node network to 784-node network, using a digital oscillator for digi recognition of digits 0, 1, and 7. The results suggest that the tolerance shoots up rapidly as the network size increases, enhancing the stability of the ONN, making it highly robust. A saturation point exists beyond which the law of diminishing returns is observed. A tolerance of up to 99.9% in frequency fault and up to 59% in stuck-at fault is observed for extremely large networks of size 784 neurons.