I recently finished On Intelligence, a book on the underlying mechanism of cognition by Jeff Hawkins and Sandra Blakeslee. I very highly recommend it to anyone interested in brains and cognition, it is a very accessible read, with excellent content.
I’d been slowly, slowly working through the book, which should have taken me about 3 hours in two sittings, over the course of several weeks, due to lack of free time, and finally got a block of time on the bus on to the way to SoutheastCon to finish. The cohesion and detail of my understanding probably suffered from reading half of it in 10 minute sittings over the course of several weeks, and the other half on a single shot later, but it was still excellent.
The important thing is that the book has a wonderful main argument: Basically, they argue that the neocortex is running a single, simple hierarchical memory-prediction model everywhere, for all the senses, and this algorithm is intelligence. It is a beautiful, simple model, and like most such models is largely untestable with current technology. Unlike most such untestable models, the end of the book includes a list of “just out of reach” testable predictions, which shows welcome understanding and acknowledgment of the issue.
I only had a few objections to the ideas in the book. Chiefly, I object to the degree to which he rejects behavioral equivalence. I pretty firmly do believe that any system which perfectly emulates intelligence over all sets of inputs and outputs in a given domain is intelligent in that domain, and tend toward the “Virtual Mind” argument on such things. In particular feel that if there IS a single, simple algorithm for intelligence, there should be a (probably unbounded) number of “intelligencally equivalent” algorithms which yield intelligence, just as there are an infinite number of computationally equivalent mechanisms for computation. In general, it seems unlikely to me that there is only a single mechanism by which intelligence (which may be sufficiently different than our own to be difficult to recognize) can arise. This fits well with the idea of domain-specific intelligences he suggests in the latter portion of the book.
The authors themselves are neat as well; Jeff Hawkins was the founder of Palm and Handspring, and is roughly the father of handheld/ubiquitous/mobile computing. He was initially trained as an electrical engineer, then, like many other interesting EEs, decided he was more inclined to pursue his interest in intelligent machines, which has resulted in the Redwood Center for Theoretical Neuroscience and Numenta, to understand the brain and build brain-like machines. He has a TED Talk on the same topic.
I’d like to find a book (or other large body of relatively accessible text) on the “Emergent property of parallel systems” or the similar “Society of Mind” theory of intelligence, it’s the only other one I’m aware of that seems both reasonable and testable.