An investigation of trained neural networks from a neurophysiological perspective

The application of theoretical neural networks to preprocessed images was investigated with the aim of developing a computational recognition system. The neural networks were trained by means of a back-propagation algorithm, to respond selectively to computer-generated bars and edges. The receptive fields of the trained networks were then mapped, in terms of both their synaptic weights and their responses to spot stimuli. There was a direct relationship between the pattern of weights on the inputs to the hidden units (the units in the intermediate layer between the input and the output units), and their receptive field as mapped by spot stimuli. This relationship was not sustained at the level of the output units in that their spot-mapped responses failed to correspond either with the weights of the connections from the hidden units to the output units, or with a qualitative analysis of the networks. Part of this discrepancy may be ascribed to the output function used in the back-propagation algorithm.