Learning in evolutive neural architectures: An ill-posed problem?

Abstract

Basically, evolutive architectures are networks able to be modified by adding or pruning neurons or connections. In the paper, by a synthesis a various works, we point out that evolutive architectures involve a lot of tricks because of indeterminacy of solutions and suboptimality, which are characteristic of ill-posed problems. We also emphasize on interest of stopping criteria, essential to control adding as well as pruning procedure and to avoid overfitting. Finally, we suggest another formulation of learning in evolutive architectures based on more realistic “hardware” constraints.