BP learning of visual motion detection, and analysis of the internal NN structure that emerges

The backpropagation (BP)-type neural network (NN) was extended to deal with spatio-temporal input patterns, and such an NN model was trained to detect computer-simulated visual motion of light/dark boundaries (edges). The internal NN structure organized as a result of this learning was analyzed to find that, (1) one hidden unit is sufficient for uni-directional motion detection, (2) the spatio-temporal receptive field of the hidden unit exhibits stripes of excitatory and inhibitory areas parallel to a preferred direction, and this stripe-pattern receptive field works essentially as a 2D derivative filter, (3) after passing through this filter, the motion signal is processed with a threshold element followed by a low-pass filtering for smoothing. Thus, the present motion detecting NN is a nonlinear system with a sandwich structure of differentiation, threshold and low-pass filtering, as it is comprised of neurons having center-surround opponent receptive fields and a threshold unit with time-delayed input.