Analyzing the Effects of Noise and Variation on the Accuracy of Analog Neural Networks

Traditional machine learning algorithms and neural networks implemented using digital architectures such as GPUs, TPUs, and FPGAs demonstrate high performance, but the power required to train and predict is too high to be implemented in energy-constrained systems such as implants and edge devices. Although analog classifiers offer the possibility to significantly reduce power consumption by two or three orders of magnitude, the nonidealities inherent to analog circuits such as noise, drift, and process variations make it very difficult to implement accurate analog neural networks. This paper explores the effects of these nonidealities on classification performance.