Designing Neural Networks from Statistical Models: A New Approach to Data Exploration

Antonio Ciampi and Yves Lechevallier

We develop, in the context of discriminant analysis, a general approach to the design of neural architectures. It consists in building a neural net "around" a statistical model family; larger networks, made up of such elementary networks, are then constructed. It is shown that, on the one hand, the statistical modeling approach provides a systematic way to obtaining good approximations in the neural network context, while, on the other, neural networks offer a powerful expansion to classical model families. A novel integrated approach emerges, which stresses both flexibility (contribution of neural nets) and interpretability (contribution of statistical modeling). A well known data set on birth weight is analyzed by this new approach. The results are rather promising and open the way to many potential applications.

This page is copyrighted by AAAI. All rights reserved. Your use of this site constitutes acceptance of all of AAAI's terms and conditions and privacy policy.