Recognition of objects rotated in depth using partial synchronization of chaotic units

A regular array of discrete-time nonlinear oscillators with recurrent connections (coupled map lattice or CML) can perform object recognition by acting as a dynamical recognizer, while simultaneously performing computations to normalize class members to a common representation. Partition cells of the network state space serve as dimensions of the representation space. Occupancy statistics in each such cell after a brief evolution of the system, governed by intrinsic time-varying dynamics of the oscillators and the forms presented as initial conditions, result in a population code measured over all units. Results on recognition of paperclip objects rotated in depth using an ensemble of classifiers are reported. With training on seven views separated by 30 degrees, the system achieves recognition rates of 85% for a set of twenty paperclip objects. Recognition is achieved in an average of 12 iterations of the recurrent system Performance degrades with fewer training views, with angular distance from training views, and with an increasing number of objects. Biological support for this theory of object recognition is examined.