Effect of Parameter Drift on the Accuracy of Analog CNN Accelerators

Challenges and opportunities in the design of high-precision analog machine learning accelerators will be studied. An analytical approach has been proposed to evaluate the effect of process variations and mismatch on the performance of Convolutional Neural Networks (CNNs). A sample CNN based on the VGG11 architecture, modified to classify the CIFAR-10 dataset, was developed and used as part of this study. The analysis carried out in this work reveals that despite the inherent sensitivity of analog circuits to process variations and mismatch, certain computational layers in a CNN show high sensitivity to drift. As a result, proper calibration techniques can be used to target the most sensitive part of the computing system, and mitigate the effects of process variations and mismatch, thus maintaining a high overall system performance and accuracy. Such an approach will enable the development of machine learning accelerators based on analog topologies that are energy-efficient and high-speed without compromising their accuracy.