Abstract
Nearly all neural networks for pattern recognition being dealt with today are associative classifier or identifier networks that are not cognitive. To be cognitive a net must be able to distinguish, on its own, between familiar and unfamiliar or novel sensory signals present at its input, and this can not be done by associative classifiers. It will be argued and shown that, to be truly cognitive, a network must be nonlinear and dynamic and able to manifest bifurcation. This means it should be able to carry out phase space computations with more than one type of attractor and to switch between these depending on whether the sensory input is familiar or novel. Cognition implies, therefore, bifurcation and computing with diverse attractors. Our reasons for adopting this view, which stemmed from known biophysical observations and from our neuromorphic target identification work, are discussed. An example of a cognitive network that computes with both stationary (limit point) and dynamic (periodic) attractors is given to illustrate our thesis. The elements of a neuromorphic radar target identification system which employs these concepts and is capable of distortion invariant recognition of three targets with perfect score is presented. The work presented elucidates the role of periodic attractors in feature binding and cognition and the significance of cognition in autonomous systems.
© 1991 Optical Society of America
PDF ArticleMore Like This
Nabil H. Farhat
IS10 Nonlinear Dynamics in Optical Systems (NLDOS) 1990
Philip D. Henshaw and Steven A. Lis
ThB2 Persistent Spectral Hole Burning: Science and Applications (SHBL) 1991
Alex V. Huynh, John F. Walkup, and Thomas F. Krile
MII8 OSA Annual Meeting (FIO) 1991