Linear Classification on Noisy Hardware

Abstract

Motivated by the growing interest in machine learning on nanoscale edge devices, we study the effect of hardware noise and quantization errors on the accuracy of inference by linear classifiers. Our experiments on synthetic and real data sets using well accepted models for hardware noise and errors show that they have a significant impact on the accuracy. For mitigating those effects, we propose an easily implementable strategy by combining insights from linear classification, convex analysis and concentration of measure. Evaluations on synthetic and real data sets show that our simple strategy improves the performance significantly. We end with a brief discussion on a few avenues for further explorations.