Retinal architecture in CNN

Abstract

There is a remarkable and compelling similarity between the architecture of the retina and that of cellular neural nets (CNN): Both are massively parallel analog array processors where the strength and form of connections between neighboring elements determines the characteristics of the image processing operation. This close relationship allows us to transfer algorithms from one platform (retinal wetware) to the other (analogic software). The full complement of retinal algorithms, organized in separate interactive sheets of activity, for a complete retinal subroutine that operates in real time. Retinal algorithms can be modified in a variety of ways to form "what if" functions that are testable in the physiological preparation. These algorithms can also be implemented in CNN then applied to real-world problems. The author describes here some of his recent advances in implementing retinal function in CNN.