M.A. Kerin and T.J. Stonharn
Synopsis
A new architecture for unsupervised learning based on digital neural networks is proposed. The system lends itself easily to hardware implementation and its ability to successfully cluster real world data in real time is demonstrated. Performance levels are contrasted with those of existing methods to illustrate the new systems improved characteristics.
Key words:
Neural networks, self-organisation, digital neural models.
1. Introduction
The degree to which any intelligent machine can cope with an unpredictable and constantly changing environment, determines the level of autonomy it can realistically achieve. This in turn determines the range of real world tasks for which it will be suitable. For neural networks to fulfil their potential as a basis for autonomous systems, they must possess the ability to deal with novel experiences in a logical manner; extracting and learning relevant information from such occurrences so that they will be better equipped to deal with similar events in the future. Yet this process must be achieved without the aid of an external teacher and in such a manner that prior knowledge is complemented, not lost. Systems capable of such behaviour are said to be 'self-organising', or to perform 'unsupervised learning', and their development is a well established area of neural network research. Several architectures exist (e.g., Kohonen (1982), Carpenter & Grossberg (1987)) which will automatically categorise patterns presented to them. However the majority of such models use analogue nodes, which are difficult to implement in hardware. This paper introduces a new self-organising system based on digital nodes. Such nodes are easily realised using standard components and their fundamentally different methods of operation can result in useful characteristics.