Implementation of neural constructivism with programmable hardware

Abstract

Most neural network models base their "learning" capability on changing the strengths of interconnection between computational elements. However, according to "neural constructivism", an environmentally-guided neural circuit building offers powerful learning capabilities while minimizing the need for domain-specific structure prespecification. This paper presents a field programmable hardware implementation of an unsupervised constructive neural network with online size adaptation, a form of neural constructivism, and presents a color learning and recognition application.