On Intelligence by Jeff Hawkins
Publisher: St. Martin's Griffin
On Intelligence is Jeff Hawkins radically different approach to machine intelligence. While traditional AI has leads to several useful products, it becomes clear early on that Hawkins is out for something radically different. Or, as he puts it himself, traditional AI scientists "left out the most important part of building intelligent machines, the intelligence". I do think Hawkins is spot on here. He seeks the failure of AI in its focus on behavior (a tradition with roots dating back to behaviorism, the direction that dominated psychology as AI was born). Even if AI will continue to deliver value it simply won't lead to truly intelligent machine. Rather, intelligent machines have to be based on how the human brain works. In this excellent work, Jeff Hawkins defines such ideas in his unifying memory-prediction theory.
Hawkins starts out by explaining the different approaches in artificial intelligence and why they failed to model intelligence. Of particular interest is the failure of neural networks. Even though modeled on a basic biological structure of the brain, neural networks missed interesting parts such as the brain's feedback loops or its ability to interpret sequences of patterns. In fact, exactly these omissions form the core of Jeff Hawkins memory-prediction framework. According to this theory, the brain not only remembers sequences of events, but also their relationships. Thus, the brain is able to form and maintain a biological model resembling the structure of the physical world. Based on this model, the brain makes probabilistic predictions by identifying invariants in the constantly changing stream of inputs that make up our environment. And Hawkins believes that making predictions is actually the primary function of the cortex.
The memory-prediction theory is presented in the context of modern neuroscience. There's a certain elegance and attractive simplicity to Hawkins theory. One such fascinating example is how the brain remembers patterns without differentiating between sensations and behaviors. Not only is it an appealing simplification; it's one example of how memory guides predictions and thereby intelligent behavior. When we find ourselves in a situation similar to one we remember, this mechanism not only allows the brain to see into a possible future. It also allows recalls of the behaviors that led to such a future. Another example is how Hawkins builds on John Hopkins ideas. That is, the observation that since all regions in the cortex look the same, perhaps it is because they perform the same basic task? Given that hypothesis, it's actually the similarities between different brain regions that are interesting, not their differences.
Obviously, Hawkins goes into much more depth in his book. And like any good theory, Hawkins includes testable predictions. I won't try to cover them here since On Intelligence is such a readable and accessible work, yet with a lot of depth. On Intelligence presents a clear explanation of how conscious processes (that is, with our attention directed at the task) might work. While such a model explains how the brain can beat computers that are orders of magnitude faster at tasks like image processing, I was left wondering about automatic processes. After all, they do make up a large part of the activity in the brain. That said, I do believe Jeff Hawkins is on to something immensely interesting. On Intelligence is a brilliant book and highly entertaining read.
Reviewed April 2011