Stephen Egelen, Jim Stone, Harry Barrow
A novel unsupervised learning method for extracting spatio-temporal invariances has been developed in Stone(1995). The learning method works by trying to jointly minimise the short term variance of a unit's output, whilst maximising the long term variance of the output. The learning method has been applied to extracting disparity from a temporal sequence of random dot stereograms. This paper reports on developing and applying the learning method to a spatial task, both in one and two dimensions. Random dot stereograms were used as input to a three layer feedforward network. After learning, output units in the network became selective for disparity. This confirms the usefulness of the learning method, and leads the way to creating a full spatio-temporal model, using both temporal and spatial information.
Download compressed postscript file