A three level hierarchical nonlinear model of cortical function for perception

Brains are characterized by every property that engineers and computer scientists detest and avoid. They are chaotic, unstable, nonlinear, nonstationary, non-Gaussian, asynchronous, noisy, and unpredictable in fine grain, yet undeniably they are among the most successful devices that a billion years of evolution has produced. No one can justifiably claim that he or she has modeled brains, but they are a flowing spring of new concepts, and they provide a gold standard of what we can aspire to accomplish in developing more intelligent machines. The most fertile source of ideas with which to challenge and break the restrictions that characterize modern engineering practice is the electroencephalogram (EEG). It was the action potential of single neurons that provided the foundation of neurobiology for the 20th century, and in its time it supported the development of digital computers, neural networks, and computational neuroscience. Now in the 21st century, the EEG will lead us in a remarkably different direction of growth for the computing industry, which will be dominated by highly parallel, hierarchically organized, distributed analog machines. These devices now exist in prototype form. They feed on noise in support of chaotic attractor landscapes, which are shaped by reinforcement learning through self-governed experience, not training by `teachers´, and they may solve many of the problems of interfacing between finite state automata and the infinite complexity of the real world