Julian Francis Miller

pdf

Abstract: We construct artificial neural networks (ANNs) by executing evolved programs inside two neural components: the body (soma) and the dendrite. The programs decide whether neu- rons and their dendrites move, change, die or replicate. When the programs are executed they build a neural structure from which multiple conventional ANNs can be extracted each of which can solve a different computational problem. In bi- ological brain development electrical activity strongly influ- ences developmental changes. This is known as activity de- pendence. In this paper, we describe various ways activity de- pendence could be implemented and investigated. We show that allowing the soma bias to be activity dependent improves performance.

Part of the 2020 Workshop on Developmental Neural Networks