Learning Association Fields from Natural Images

Previous studies have shown that it is possible to learn certain properties of the responses of the neurons of the visual cortex, as for example the receptive fields of complex and simple cells, through the analysis of the statistics of natural images and by employing principles of efficient signal encoding from information theory. Here we want to go further and consider how the output signals of ‘complex cells’ are correlated and which information is likely to be grouped together. We want to learn ‘association fields’, which are a mechanism to integrate the output of filters with different preferred orientation, in particular to link together and enhance contours. We used static natural images as training set and the tensor notation to express the learned fields. Finally we tested these association fields in a computer model to measure their performance.