A NONLINEAR HEBBIAN NETWORK THAT LEARNS TO DETECT DISPARITY IN RANDOM-DOT STEREOGRAMS

An intrinsic limitation of linear, Hebbian networks is that they are c apable of learning only from the linear pairwise correlations within a n input stream. To explore what higher forms of structure could be lea rned with a nonlinear Hebbian network, we constructed a model network containing a simple form of nonlinearity and we applied it to the prob lem of learning to detect the disparities present in random-dot stereo grams. The network consists of three layers, with nonlinear sigmoidal activation functions in the second-layer units. The nonlinearities all ow the second layer to transform the pixel-based representation in the input layer into a new representation based on coupled pairs of left- right inputs. The third layer of the network then clusters patterns oc curring on the second-layer outputs according to their disparity via a standard competitive learning rule. Analysis of the network dynamics shows that the second-layer units' nonlinearities interact with the He bbian learning rule to expand the region over which pairs of left-righ t inputs are stable. The learning rule is neurobiologically inspired a nd plausible, and the model may shed light on how the nervous system l earns to use coincidence detection in general.