Low Power Analog Integrated Classifier Circuits Based on Gaussian Mixture Model for Biomedical Applications
Georgios Gennis
Abstract
The purpose of this thesis is the development of analog classifiers based on the Gaussian fuction. Specifically, two general architectures will be discussed· one of a Bayesian classifier with a Gaussian probability density function and one of a Gaussian Mixture Model-based classifier. These fully analog architectures are adaptable to a wide range of applications. The training, the operation and the adaptation, to different applications, procedures are described in detail. Additionally, the basic building circuits used for these architectures are presented. In this thesis 4 different Bump circuit architectures and 2 Winner-Take-All circuit architectures will be analyzed. A toy and 4 real-world classification problems have been used to confirm the proposed architectures as well as the sensitivity of the circuits.