On Intelligence is a worthwhile book in spite of its flaws. Although there is little here that is truly new, the material is very well presented and is nearly convincing. Hawkins purports to be presenting a "comprehensive theory of how the brain works". In truth, the book is a bunch of autobiography, a slanderous attack on the fields of AI and Speech Recognition, twenty pages of dense, impenetrable description of wiring in the neocortex, a few pages of genuine but not terribly original insight, and a few rounds of self congratulation.

The core idea is that the ability to make predictions about the future is the crux of intelligence. The neocortex specializes in this, and a large neocortex is what separates humans from other animals, allowing us to learn complex patterns and make far-ranging predictions. I agree completely with this viewpoint. Back in the mid-90s, my grad-school research group came to the same conclusion. DAn Ellis's 1996 thesis, Prediction-driven computational auditory scene analysis, showed how such an approach could solve difficult low-level perceptual problems in audition. My own 1999 thesis, Sound-source recognition: a theory and computational model, showed how hierarchical models of sound-sources predicted (and matched) human performance in a recognition task.

What Hawkins presents is not a recipe for intelligence; it is descriptive but not prescriptive. Many of the elements he describes are likely to be necessary components of intelligent systems, but they aren't the whole story. Even if we can build hardware that is capable of intelligence, we still won't know how to teach it. Our human cultures and DNA programming are highly evolved to teach children the skills they need to survive, from the way that a mother's tone of voice captures the attention of an infant onward. Emotions likely play a crucial role in intelligence (one that Hawkins dismisses), helping us determine where to focus our intellectual efforts and breaking us out of unproductive cycles (among other things). And surely there are other prerequisites for intelligence that I can't think of off the top of my head.

On Intelligence is good in that it may point neuroscientists in new directions to look for answers to how the brain works. I hope that it inspires the Pattern Recognition and Machine Learning research communities to develop general hierarchical learning algorithms (I struggled to no end with this subject when I was working on my thesis). It seems unlikely that directly copying biological "circuits" is the only way to build machine intelligence, but it's a great place to look for inspiration.

I'm personally inspired to take another look at Stephen Grossberg's research and to flip through Minsky's Society of Mind again. This is heady stuff (pun unintentional), and Hawkins is to be commended, if only for getting more people to think about the topic.