“On Intelligence,” written by Jeff Hawkins with Sandra Blakeslee, is a great read, full of provocative ideas about how our brains work.
Hawkins founded Palm Computing and Handspring, but he says that his true lifelong passion has been trying to understand our brains. In 2002, he founded the Redwood Neuroscience Institute (now the Redwood Center for Theoretical Neuroscience at Berkeley).
At Redwood, he developed a theory of the brain, that he expounds in this book. The book is a popular science book; you will not find any equations or explicit algorithms. It reads very smoothly, undoubtably due to the fact that Hawkins was helped in writing the book by Sandra Blakeslee, a science writer for the New York Times.
Hawkins argues that the cortex consists of modules that are all performing the same algorithm. The purpose of that algorithm is to learn to complete and predict the spatio-temporal patterns coming into a module from “lower” modules in the brain’s hierarchy, using feedback from “higher” modules.
I find this hypothesis for how the brain works extremely attractive, and “On Intelligence” argues for the hypothesis almost too well, in that the difficulties of converting the hypothesis into a concrete and useful algorithm will slip by the reader, as if by sleight of hand. The problem, of course, is that it is easy to use words to talk about how a brain might work; the hard part is making a machine do the same thing.
After reading the book, I actually spent a few weeks trying to make the hypothesis into a concrete algorithm, without much success. But Hawkins is certainly putting his money where his mouth is: he has founded a company, called Numenta, and Numenta has released software (the “Numenta Platform for Intelligent Computing” or “NuPIC”) that implements part of his theory. NuPIC appears to be based on software originally written by Numenta co-founder and Stanford graduate student Dileep George, but transformed by a team of software developers into a professional-quality product.
However, it is very disappointing that the NUPIC software does not include feedback in its hierarchies, nor does it let you learn and infer temporal patterns (only spatial ones). To be honest, I am not so surprised, because it was these elements (that are obviously central in Hawkins’ theory) that I found it so difficult to integrate into a real algorithm.
Numenta promises that future versions of NuPIC will fill these gaps. I sincerely wish them the best of luck! I should say that I think it is wonderful that an impatient guy like Hawkins pushes the field, using a different approach than the standard academic one.
So in summary, you should definitely read Hawkins’ book, and if you’re intrigued, you should by all means check out the free Numenta software. Whether this line of research will lead to anything significant though, I’m not really sure…