Synchronization opponent systems: Attractor basin transient statistics as a population code for object representation

摘要

A network of recurrent logistic units with local, orientation-sensitive coupling is used to study the transformation of spatial forms in a computational framework called synchronization opponent cooperative action networks. Evolutionary search finds dynamics for which curves, rank-ordered by a generating parameter, are transformed to produced a similarity space with dimensions defined by partitions of the phase space. The network also recognizes objects by normalizing several views of 3-D objects to a characteristic distribution. A spatio-temporal population code where both active and quiet units convey information is proposed, with the phase space interpreted as ensemble average frequency of coupled populations.