Predicting the receptive fields of simple cells from the statistical properties of the visual world requires, as a first step, to determine the relevant regularities of the visual stimuli. A regularity that pervades most natural images is the persistency of contrast structures as the image is seen in finer detail. Suppression of this scale redundancy leads to an efficient processing consisting in a linear stage followed by a non-linear, multiplicative process. The linear piece can be obtained by a simple learning rule and gives a prediction for the receptive fields of simple cells. There is a direct relationship between elementary image features and receptive fields: for each type of elementary image feature the learning rule predicts a new cell. The distribution of the predicted receptive fields are in agreement with the observed properties of simple neurons.

Registration required:

No

Free:

Yes