We picked up a book from the University of Toronto bookstore over the holidays, "On Intelligence", by Jeff Hawkins. I sat down today and read it (it's only around 200 pages). Basically it echoes a number of the theories from my thesis way back in 1998, but wraps it all up in a discussion of how to replicate the structures (i.e. create thinking machines), instead of using the understanding to help design for humans.
Key points of similarity:
* perceptual fields -- all perception happens with the same basic mechanism, different presentation, but the same idea underneath (I'd thought this was just universally understood when I wrote about it, Hawkins says its new)
* perceptual cycle -- predictions control expectations, the "model" is constantly checked against the perceptions looking for new, unexplained experience at all levels
* multiple levels/hierarchy -- recursive structures where the same basic model-formation and checking is applied to the patterns seen at that level
* temporal -- sequences in time are part of perception
Points of differentiation:
* Hawkins implies all time relations are simply linear. My contention is that time relations are actually multi-periodic, likely implemented by having different time-sense neurons with different delay periods
* Hawkins view of "downward" propagating signals seems too simple (at least, as presented here), I think it's just a simplification introduced for the book (page 152 describes the "I didn't fire when I expected to" behaviour, which implies that his model should include the priming stage, though he doesn't explicitly mention the idea of a primed/expecting neuron)
* Hawkins tells us there's a limit to the number of levels. I'd thought of the levels as continuously expanding as we added levels of abstraction. His idea seems to mesh better with the physical constraints of the brain, though I can't see where in his argument he describes how we still retain the ability to "see" with young eyes (e.g. to be able to focus on just a letter in a word)
Hawkins presentation is much more consistent. He's building a ground-up model of how the system works, rather than trying to teach designers enough of it to allow them to better manipulate the system. Still, all in all it seems reasonably compatible.
Pingbacks are closed.