F.C.D.B.C. Filho, D.L. Bisset and M.C. Fairhurst
Synopsis
This paper explores different architectures for Goal-Seeking Neurons (GSN) (feed-forward, feedback and self-organising) together with different learning rules, and investigates a range of alternative configurations within these three architectures. Practical results are demonstrated in the context of a character recognition problem.
Key words:
Neural Networks, Architectures, RAM-based Neurons, Fast learning, Boolean neural networks, GSN
1. Introduction
A neural network consists of three fundamental components, the computational unit, the architecture employed to connect these units together, and the learning algorithm that is used to modify the unit functions to achieve some high level task. Clearly, different types of unit, architecture, and learning algorhythm will endow the network with different overall properties, and limitations.
This paper is concerned with different architectures and learning methods for networks composed of Goal-Seeking Neurons (Filho et al., 1990). The Goal-Seeking neuron is a particular type of Boolean neuron that uses local low level goals to govern its behaviour, and is capable of one-shot learning. The main aim of this paper is to present learning algorithms and architectures that demonstrate the different styles of computation possible with GSN nodes operating in networks that use either externally supplied training targets or which self organise their responses.