Expectation based feedback in a neural network which recognises hand-drawn characters and symbols

Concerns the development of a neural net system to recognise input characters and symbols. The symbols are such as are found on engineering drawings and suchlike, the purpose of the system being input to a CAD system. A flexible, general-purpose, connectionist system for recognising distorted patterns is described. The problem of uncertainty in the featural inputs which contribute to a pattern is particularly addressed. A feedback or resonance process is used to attempt to match these inputs to a member of the expected pattern set. The system is described in computational, rather than mathematical, terms. The feature detectors are too complex to characterise mathematically. However, mathematical analysis shows that the system can perform particular types of classification which are not possible for a linear system, even if the individual feature detectors themselves were linear. The system has been applied to the recognition of hand-drawn characters and symbols, but would appear to be applicable to other pattern classification tasks as well