Current address: HNC Software Inc., San Diego, CA 92121-3728, and Sloan Center For Theoretical Neurobiology, The Salk Institute, La Jolla, CA 92037; sirosh@hnc.com.

In Hebbian learning, synaptic efficacies are adjusted based on coincident pre- and postsynaptic activity [19,21]

However, Xing and Gerstein [45] have formulated a different computational explanation for dynamic RFs based on predetermined and static afferent RFs in a simple network.

RF-LISSOM is short for Receptive Field Laterally Interconnected Synergetically Self-Organizing Map.

The similarity was measured by comparing Fourier transforms, autocorrelation functions, and correlation angle histograms of experimental and model maps. See [10] for a discussion of these methods.

Note that if the lateral connection patterns are observed on the cortex directly, it is very difficult to determine their orientation because of the log-polar mapping from the retina to the cortex. The cortical patterns would first have to be mapped back to the visual space. The model bypasses the log-polar transformation for simplicity, and the lateral connection patterns are directly observable.

The inhibitory lateral connections were pruned if their strength remained below . The input spot widths a and b were chosen randomly within . The lateral excitation strength was while decreased gradually from to . The learning rate decreased from to , from to and was a constant . The lower and upper thresholds of the sigmoid increased from to and from to . Small variations of these parameters produce roughly equivalent results