MVLSC Home · Issue Contents · Forthcoming Papers
Classification Functions for Machine Learning and Data Mining: Recent Results
Tsutomu Sasao
A classification function is a function that assigns integer values to inputs based on a training set. For instance, a training set consisting of images of airplanes and automobiles can be used to infer a function that distinguishes between unseen images of airplanes and automobiles. While neural networks are commonly used for this purpose, they require significant hardware resources and consume high power. This paper presents recent advancements in classification functions that address these issues. We use logic synthesis techniques to reduce the number of variables and simplify sum-of-products expressions (SOPs). This simplification improves training and hardware efficiency and reduces power dissipation, albeit at the cost of lower accuracy compared to neural networks. Experimental results for MNIST handwritten character recognition circuits and fashion-MNIST recognition circuits are presented to illustrate these techniques.
Keywords: logic minimization, linear transformation, MNIST, generalization ability, dimension reduction