Hierachical learning of visual invariances via spatio-temporal constraints

An unsupervised learning algorithm is presented for the acquisition of low-level vision tasks, such as the extraction of surface depth. A key assumption is that perceptually salient parameters in the image of a moving object vary smoothly over time. This assumption is consistent with a learning rule which maximises the long-term variance of a unit´s outputs, whilst simultaneously minimising its short-term variance. This learning rule involves a linear combination of anti-Hebbian and Hebbian weight changes, over short and long time scales respectively. It can be shown that the learning rule maximises the mutual information between a unit´s output and the short-term exponentially weighted mean of that unit´s output. The model is demonstrated on a hyper-acuity task: estimating sub-pixel stereo disparity from a temporal sequence of stereograms. The model can discriminate disparities as small as one-tenth of the width of an input receptor, which compares with human performance on stereo hyper-acuity tasks. Furthermore, the algorithm generalises, without additional learning, to previously unseen image sequences. The results of applying the learning method independently to successive layers of a hierarchical network are presented