Jeff Hawkins is one of those jaw dropping people who seem to have an amazing capacity to make things happen. He's the pioneer behind "pen based computing" and the Palm and Handspring companies.
More recently he's used his money and inventiveness to found the Redwood Centre for Theoretical Neuroscience, based at the Helen Wills Neuroscience Institute at Berkeley. From this, along with long term collaborator Donna Dubinsky, he's spun off a company called Numenta.
Why the business update? Well, Hawkins is of the view that current research in AI will never succeed in emulating "holistic" human intelligence. In that he's right. In the AI community, we're very good at solving particular problems (such as how to win at chess) and developing particular techniques (like the neural net or genetic algorithm). We have computers that are a lot smarter than computers "ought" to be - this area of research is known as "weak AI". We're also pretty good at using some of these intelligent modes of computing to better understand living intelligences, like human perception for instance.
Strong AI - making a computer that thinks like a human - has been a dirty word for some time now. The reins on this area of research were tugged in pretty sharply at the onset of the AI winter and have never been slackened since.
Hawkins may be about to change all that with his model of Hierarchical Temporal Memory (HTM). The details of his group's work can be found here and [PDF] here.
Hence the business update. I get my jollies from evolutionary computation and logic programming, so this is not my area. From the (very) little I do know about neurally inspired computing, this doesn't look like "neural nets 2.0". But I don't know: is this old news, a pipe dream, or the shot in the arm that brings "hard AI" back to life?
If I read something like this from any other entrepreneur I'd be cynical as all getout. On Hawkins' past form however, I'm downgrading that to merely sceptical. With an option on secretly hoping to be convinced.
And if all that was just too dry for a Friday - take a look at this "tutorial" on neural architecture:
(Spotted on Mind Hacks
2 comments:
Way cool stuff! Personally, I'm more interested in modelling intelligence rather than the brain itself. The brain is clearly one good working model of intelligence, but I don't think a cortex is required, which is why massively simplified models of neural networks can still be useful. My worry about HTMs is that from a (admittedly cursory) look through the Spectrum article and their website, I can't see any actual working demonstrations, apart from a line-drawing recognition system, which doesn't sound especially ground-breaking. Definitely one to keep an eye on though...
A network of artificial nerves is growing in a Swiss supercomputer, meant to simulate a natural brain, cell-for-cell. See Neurotechnology: Growing a Brain in Switzerland.
Post a Comment