Regularization in Vision

In the 1985 Nature review article “Computational vision and regularization theory” (http://www.nature.com/nature/journal/v317/n6035/pdf/317314a0.pdf) by Tomaso Poggio, Vincent Torre and Christof Koch, which outlines the role of regularization methods in finding plausible solutions to computer vision problems, the authors have this very interesting speculation about biological vision:

 

“One of the mysteries of biological vision is its speed. Parallel processing has often been advocated as the answer to this problem. The model of computation provided by digital processes is, however, unsatisfactory, especially given the increasing evidence that neurones (sic) are complex devices, very different from simple digital switches. It is, therefore, interesting to consider whether the regularization approach to early vision may lead to a different type of parallel computation. We have recently suggested that linear, analog networks (either electrical or chemical) are, in fact, a natural way of solving the variational principles dictated by standard regularization theory.”

 

Not only do the authors provide a plausible reason for the computational speed for biological vision, but they also provide a compelling argument against the AI philosophy of treating intelligence and cognitive processes as separate from their biological/physical substrates.