I've been longing for a more directed goal for the intellectual stuff I do in my own free time. I used to play a lot of poker, which didn't require much mental energy and didn't take me anywhere. It was fun (though all too often extremely frustrating) at times, sure, but kind of pointless. I'm interested in a lot of things and I enjoy learning about novel ideas, however unrelated to anything else they seem to be.

But one area of interest that has kept me fascinated is the ideas that Jeff Hawkins talked about in his book, On Intelligence. He basically claimed that he has come up with new a theory of intelligence. Somewhat preposterous claim, but the book is compelling. He isn't saying that he has thrown everything out. Almost all of the parts of his theory have been represented in some form or another. But his claim is that his theory puts them all together in a new way.

This stuff is what I want to concentrate on. I want to understand the theory better, I want to understand the neurological background, and I want to get my hands into the implementation of the theory.

How do I know that this is a fruitful approach to the problem? Especially given that I don't know much about the problem space beforehand. How do I know that Hawkins's approach makes sense? How do I know that this approach will have any practical relevance in the future? How do I know that it's worth investing time on this stuff?

I don't know.

His theory has a lot of intuitive appeal for me and once I understood the basics of his ideas, it just seems to make a lot of sense. I'm pretty sure that I first heard of his ideas from Aaron Swartz's weblog entry and Aaron's enthusiasm really caught on me. I went on to see some lectures by Hawkins on the net and to read the book. It all makes a lot of sense.

Now, none of this proves anything. To give you a concrete example, there's this demonstration that Hawkins brings to his lectures, which is basically their first prototype of the concept. It builds a model of 64x64 pictures shown to it. After it has learned these pictures, you can show it new slightly modified versions of the pictures and it can tell what they are.

If you describe the prototype to someone, they will probably not be that impressed. If you describe it to someone who has background in neuroscience or computer science, they will probably tell you that there's nothing new or particularly hard in what the prototype does. Or if it's someone who's working on the similar subjects, they'll tell you that they've already done more impressive stuff.

So it's really hard to get any real reassurance. The only thing that I can go by right now is the feeling I get from his book and his lectures. The really positive part of this is that they have a concrete non-vaporware implementation of the concept, which is downloadable from Numenta's website. The fun thing about it is that from the last I saw, it had a Python interface.

It will take some time until I can really say what impact it will have on me or whether I can make something useful about it. But that's what I will pursue for now. I don't have much free time, but it's all the more reason to direct my efforts. My target for now is to understand the mechanics of what they call the HTM (hierarchical temporal memory system) and understanding (neo-cortical) intelligence in general.