Investigating the recognition of false patterns in backpropagation networks

This paper shows how a simple independent set of units working in parallel with a conventional multilayer perceptron network architecture enhance the reliability of the network with respect to the rejection of misshapen patterns. It is shown that rejection of invalid characters can be significantly improved without any degradation of performance with respect to the recognition of valid patterns. Other advantages of this approach are that its applicability is independent of the classes of patterns present in a domain of application and that the auxiliary network embodying the guard units can also be used in conjunction with other kinds of network which also lack the ability effectively to reject unknown patterns