A Casual View of Multi-Layer Perceptrons as Probability Models

Abstract

The purpose of this chapter is to introduce the reader to a type of artificial neural network called a multi-layer perceptron. The intention is not to present a detailed, comprehensive treatise on the subject; instead, we provide a brief, informal, tutorial in the spirit of Chapter 1. We start with the historical background and then introduce the concept of regression. This concept holds for both continuous-valued and binary-valued response variables; however, when applied to the latter, probabilistic classification models are created. In the context of medicine, probabilistic classification models are usually obtained using logistic regression analysis, but if logistic functions are nested to produce multi-layer perceptrons, the high flexibility of the resulting models enables them to handle complex classification tasks. We discuss some important points that should be kept in mind when applying multi-layer perceptrons to classification problems. Finally, we end the tutorial with some recommended reading.