[mpthompson]

Hi Everyone,

I read "On Intelligence" multiple times about 18 months ago and have since been both mulling over the ideas in my mind as well as doing some background research to better understand how the algorithms might be implemented in software and hardware. In my research I came across the work of Stephan Grossberg and Adaptive Resonance Theory (ART). Jeff only briefly mentions Dr. Grossberg's work with modeling cortical columns in his book, but the ideas and concepts underlaying ART seem very similar to HTM. If anything, HTM is quite a bit more expansive than ART, but ART seems to have a more firmly established mathematical and theoretical foundation that lends itself well to implementation.

While looking deeper into ART what struck me is how similar the LGN, V1 and V2 cortical schematics and in Dr. Grossberg's paper "How Does The Cerebral Cortex Work?" look very similar to the drawings in Chapter 6 - How the Cortex Works of "On Intelligence". Both theories seem to share a lot of common characteristics including attention/prediction, the importance of intra-layer feedback, and inhibitory as well as excitatory interactions between neurons.

The question I have is whether HTM nodes could be implemented using the establish techniques borrowed from Adaptive Resonance Theory? In particular, FuzzyART nodes that would work well with analog inputs. I'm probably going to try this myself, but I wanted to see if this is ground that has already been broken by others and if there are any pitfalls in such an approach I should avoid.

For those interested in learning more about ART, Dr. Grossberg maintains an archive of his papers at the following URL:

http://cns.bu.edu/Profiles/Grossberg/

In any case, I look forward to sharing ideas with other people on the forum interested in creating their own HTM-like networks. My particular interest is in sensory-motor control of robots and implementing HTM-like networks in environments with minimal computational resources or in hardware.

Mike Thompson