Feature deactivation using partial inhibitory networks during multiple object recognition

Discusses the activation and deactivation phenomenon in a backpropagation network. The failure of an excitatory network to distinguish certain input patterns is explained by using the 'IT' example. In a character recognition problem, the characters 'I' and 'T' have a very special characteristic; the letter 'I' is embodied in the letter 'T'. The partial inhibitory network is introduced; it can perform feature deactivation and multiple object recognition, and is applicable to this type of problem. The deductive factor governs the inhibitory effect in the network. A small deductive factor increases the performance in multiple object recognition, but also increases the training time.<<ETX>>