The construction of a Boolean competitive neural network using ideas from immunology

• Published in: Neurocomputing (Amsterdam) 2003, Vol.50, p.51-85
• Main Authors: de Castro, Leandro Nunes, Von Zuben, Fernando J., de Deus Jr, Getúlio A.
• Format: Article
• Language: English
• Subjects: Clonal selection; Competitive neural networks; Immune system; Pattern recognition
• ISSN: 0925-2312; 1872-8286
• DOI: 10.1016/S0925-2312(01)00698-1

The immune system is capable of recognizing and responding to microorganisms and molecules that cannot be perceived by our sensory mechanisms, which send stimuli straight into the brain. It performs an accessory role for nervous cognition. This paper main goals are: (1) to show how some immune principles and theories can be used as sources of inspiration to develop novel neural network learning algorithms; (2) to survey the main works from the literature that employ the immune metaphor for the development of neural network architectures; and (3) to illustrate, with a new network model, how this source of inspiration can be actually used to develop a neural network learning algorithm. The novel learning algorithm proposed has the main features of competitive learning, automatic generation of the network structure and binary representation of the connection strengths (weights). The behavior of the algorithm is primarily described using a benchmark task, and some of its potential applications are illustrated using two simple real-world problems and a binary character recognition task. The results show that the network is a promising tool for solving problems that are inherently binary, and also that the immune system provides a new paradigm to search for neural network learning algorithms.