Then this sensory input is absorbed by tens of thousands of cortical columns, each with a partial picture of the world. They compete and team up through some sort of voting system to build a global point of view. This is the idea of the thousand brains. In an AI system, this could involve a machine controlling different sensors (vision, touch, radar, etc.) to get a more complete model of the world. However, there will usually be many cortical columns for each sense, such as vision.
Then there is continuous learning, where you learn new things without forgetting previous things. Today’s AI systems cannot do this. And finally, we structure knowledge using repositories, which means that our knowledge of the world is relative to our point of view. If I run my finger over the edge of my coffee cup, I can predict that I will feel its edge, because I know where my hand is in relation to the cup.
Your lab recently transitioned from neuroscience to AI. Does this correspond to your theory of a thousand brains?
Rather. Until two years ago, if you walked into our office, it was all neuroscience. Then we made the transition. We felt we had learned enough about the brain to start applying it to AI.
What types of AI work do you do?
One of the first things we looked at was rarity. At any time, only 2% of our neurons are firing; activity is rare. We’ve applied this idea to deep learning networks and are seeing dramatic results, like 50x speedups over existing networks. Parsimony also gives you more robust networks, lower power consumption. Now we are working on continuous learning.
It is interesting that you include movement as the basis of intelligence. Does that mean an AI needs a body? Does it have to be a robot?
In the future, I think the distinction between AI and robotics will disappear. But for now, I prefer the word ’embodied’, because when you talk about robots, it conjures up images of human-like robots, which is not what I’m talking about. The bottom line is that the AI will need to have sensors and be able to move them relative to itself and the things it is modeling. But you could also have a virtual AI that roams the internet.
This idea is quite different from many popular ideas about intelligence, a disembodied brain.
The movement is really interesting. The brain uses the same mechanisms to move my finger over a cup of coffee, or move my eyes, or even when thinking about a conceptual problem. Your brain goes through frames of reference to remember facts it has stored in different places.
The bottom line is that any intelligent system, regardless of its physical form, learns a model of the world by sensing different parts of it, moving through it. This is the foundation; you can’t do without it. Whether it looks like a humanoid robot, a snake robot, a car, an airplane, or, you know, just a computer sitting on your desk browsing the internet, they’re all the same.
What do most AI researchers think of these ideas?
The vast majority of AI researchers don’t really buy into the idea that the brain matters. I mean, yeah, people discovered neural networks a while ago, and they’re sort of inspired by the brain. But most people don’t try to replicate the brain. It’s just whatever works, works. And today’s neural networks work quite well.