Quantizing Signals for Linear Classification

In many machine learning applications, once we have learned a classifier, in order to apply it, we may still need to gather features from distributed sensors over communication constrained channels. In this paper, we propose a polynomial complexity algorithm for feature quantization tailored to minimizing the classification error of a linear classifier. Our scheme produces scalar quantizers that are well-tailored to delay-sensitive applications, operates on the same training data used to learn the classifier, and allows each distributed sensor to operate independently of each other. Numerical evaluation indicates up to 65% benefits over alternative approaches. Additionally, we provide an example where, jointly designing the linear classifier and the quantization scheme, can outperform sequential designs.