Hey guys,
I don't think we've talked about it much, but Jeff Hawkins, founder of
Palm, turned cognitive neuroscientist a few years ago, publishing "On
Intelligence" and founding a company, Numenta, and the Redwood
Neuroscience Institute (eventually renamed to Redwood Center for
Theoretical Neuroscience), instantiated as a virtual institute in
collaboration with UCB's HWNI. (Kind of like the Kavli Institute for Brain
and Mind, locally.)
He read a bunch of the neuroscience literature and, like Hecht-Nielsen,
decided you don't need all of the brain to do biologically-inspired AI,
just neocortex. He came up with a model he calls Hierarchical Temporal
Memory, lucidly argued for in his book. After 18 months of hacking away at
it, backed by a team of software developers and a research institute, he's
recently decided to release the code as a "research release."
If you find the time, check it out. Could provide us with some insight,
especially where they seem to be stuck. (They haven't gotten predictive
behavior yet.)
HTM White Paper:
http://www.numenta.com/Numenta_HTM_Concepts.pdf
(This is the kind of thing I'm looking for when I talk about the project
needing a design document.)
Code:
http://www.numenta.com/for-developers/software.php
Programmer's Guide:
http://www.numenta.com/for-developers/software/programmer-guide/TOC.html
Alex
|