Jeff Hawkins Announces the Thousand Brains Project for AI
spectrum.ieee.org
spectrum.ieee.org
Often it takes time for the engineering to catch up to the theory.
Related to all this: In light of the recent AI Spring ("Attention is all you need" era), is anyone finding ML/AI slowly converging with our understanding of neurology or are we drifting to a different model where 'True AI' won't be found in emulating our own brains? It seems the latter. Our weird and wonderful brains were brewed under evolutionary constraints that have produced remarkable amounts of redundancy, because, well... a brain is incredibly precious to the body its operating within. I don't know if building an AI computer needs to be burdened with concepts born of those constraints?
It seems like there could be a loose relationship between having "many brains" and having many experts in MoE. That's why I thought something like the Arctic Snowflake LLM or whatever they called it, or a model with LOTs of experts, could have an advantage. Although that one hasn't taken over any leaderboard, obviously.
The new DeepMind paper about transformers with NARs is really interesting also.
It seems like there are eventually going to be several really powerful but different approaches to general purpose AI.
Referencing a 7 year old paper.
> To overcome these challenges, the Thousand Brains Project aims to develop a new AI platform by reverse engineering the neocortex
Oh just going to reverse-engineer the brain, why didn't anyone think of that before
> Today’s neural networks are based on basic neuroscience from 80 years ago. We’ve learned a lot about neuroscience since then, and we want to use that knowledge to advance AI
Huh? I wasn't aware any neural networks were based on biology at all other than the loose activate and fire idea.
> “Once we learn how to build one cortical column, we can build as many as we want,” Hawkins says.
Easy, just have to learn how to do it
> When you look at deep networks, they don’t fundamentally understand the world, which is why if you just change one tiny feature of an image, they often don’t recognize it
Not really true unless you develop something specifically to do this. Also, "understand" needs to be defined
Fluff piece that reads as an ad hoping for funding. Sounds like a money pit to me.
Sounds like you are aware then.
Neural networks significantly diverge from actual brain mechanisms. But they are originally based on the original concept as you seem to know. And this person is, perhaps unjustifiably, claiming that they want to bring it closer.