Re:

web.mit.edu/mcgovern/html/News.../0702_poggio.shtml - Three years old.

Hi,

Yes, that work is quite well-known (and quite good).... It's based on a system that emulates only the feedforward connections of the visual cortex; and it's not actually the highest-accuracy video or image classifier currently available.... The novel thing about Poggio's approach isn't so much its classification accuracy (which is not the best) but the fact that it achieves GOOD (even if not the best) classification accuracy via an architecture with so much biological realism.

Itamar's architecture, like that of Jeff Hawkins (but unlike Poggio's), does not aspire to detailed biological realism. This allows it to use recurrent connections (which Poggio doesn't use in his architecture yet so far as I know, probably because we don't yet know how recurrent connections work in the visual cortex, and Poggio is committed to biological realism).

Actually, I think the "secret sauce" of AGI lies in cognition not vision ... but I also think that it's great to have a sensorimotor system to hook up to one's cognition system. And Poggio's vision system is NOT suitable to be hooked up to a cognition system due to the lack of recurrent connections. Jeff Hawkins' vision system is not suitable either because Hawkins has made some very poor architectural choices. Itamar's system is suitable to hook up to a cognition system, because it has both feedforward and feedback connections and his architecture and algorithms make sense..

ben