Monday, July 03, 2006

Highly Recommended Paper on Hierarchical Temporal Memory

If you're interested in AI, or just more generally, in the ways in which computers can be applied, I'd highly recommend this white paper (PDF, HTML-sans-diagrams) about the Hierachical Temporal Memory (HTM) technology being developed by Jeff Hawkin's research company, Numenta. I think it's a real step forwards.

I think the name 'Hierarchical Temporal Memory' gives the mistaken impression it's just a storage technology. But it's not really. It's really a generic mechanism for learning and predicting causes. Specifically, a hierachy of causes. The reason they refer to it as a memory is that the system encodes a memory that is, in effect, of the sorts of causes it has come across before, which it uses as part of its prediction. The significance of what it does is explained in the paper.

What HTMs actually are is a theory of the general functioning of the neocortex. Neocortex seems to provide such a generic mechanisms for learning and predicting across hierachies of causes.

I've spoken earlier about Steve Grand's excellent book Growing Up With Lucy, which is actually a very similar theory of how the neocortex works. The thing about this paper is that it goes more into concrete details of how the system works.

If you're interested in knowing who the people behind HTMs are:

The founders of Numenta are Jeff Hawkins, Donna Dubinsky and Dileep George. Hawkins and Dubinsky were co-founders of Palm Computing and Handspring and have worked together for a dozen years. Hawkins is known as the architect of mobile computing products such as the PalmPilot and the Treo smartphone. Dubinsky was CEO of Palm and Handspring, and is CEO of Numenta. Dileep George has worked with Hawkins at the Redwood Neuroscience Institute since the summer of 2003 and has extended and formalized Hawkins' theory of the neocortex.[source]

No comments:

Post a Comment