Sunday, March 14, 2010

Jeff Hawkins - Can a New Theory of the Neocortex Lead to Truly Intelligent Machines?

I don't buy Paul Hawkins' theories, and I do not believe in "truly intelligent machines." But this is interesting in many ways.

Hawkins' position that "our brains take sensory inputs from the world and build a set of beliefs around the causes of those inputs" is valid for child brains, but as adults, the brain actually takes in sense data and other information than constructs a story based on our beliefs to explain the data. There are many deeper layers of activity that are WAY out of our consciousness, including George Lakoff's "cognitive unconscious" (Philosophy in the Flesh), which in itself is a controversial assertion. However, Antonio Damasio has proposed the idea of a "core self," which seems to me to be highly semantic and which assembles in some ways the autobiographical self, and acts as an intermediary between the proto-self (neural/body-based self) and the autobiographical self.

Anyway, the point is that by the time we are adults, we tend to understand the world based on less-than-conscious beliefs that you use to explain the world (and our own choices and preferences) to ourselves. So for Hawkins to say that we build beliefs around sensory input is to have it backwards.
Can a New Theory of the Neocortex Lead to Truly Intelligent Machines?

Jeff Hawkins
is the architect of many computer products, including the PalmPilot and Treo handheld computers. At Numenta he is developing technology derived from the brain model described in his book On Intelligence. Numenta's technology is a new type of memory architecture modeled after the mammalian cortex that can solve problems in pattern recognition and machine learning.

In addition to his positions at Numenta and Palm, Mr. Hawkins is a member of the scientific board of directors at Cold Spring Harbor Laboratory. Hawkins was elected to the National Academy of Engineering in 2003. He earned his B.S. in electrical engineering from Cornell University in 1979.

Talk about intelligent designs:

Jeff Hawkins
says he’s mapped out the way the human brain works, and has begun to fashion thinking machines to emulate the process. It comes down to Hierarchical Temporal Memory (HTM). Basically, he says, our brains take sensory inputs from the world and build a set of beliefs around the causes of those inputs. “Discovering causes is the pinnacle of what brains do,” says Hawkins. But getting good at this kind of “fancy pattern recognition” is something developing humans seem to do effortlessly, and computers only with immense labor. Learning to differentiate a cat and a dog, for instance, doesn’t come naturally to a computer. Hawkins layers his machine brains with nodes that make inferences about outside sensory data, and then pass these hunches on up a hierarchy of nodes until a consensus -- a belief -- evolves about the source of the data. The use of “belief propagation techniques”, says Hawkins, enables an entire system to reach the best overall consensus swiftly. As the thinking machine develops common representations of objects or ideas, it can generalize about new data coming at it, and learn to attend only to information that matters.

When Hawkins presented an HTM vision system with primitive line drawings of a helicopter and a mug, the system learned to identify them, even when their orientations changed dramatically, and when the lines were blurred. But the program also correctly rejected chopped-up versions of the same drawings as nonsense. “Stable beliefs at the top lead to changing predictions and behavior at the bottom,” says Hawkins. Where does this lead? Possibly to “machines that are much smarter than humans,” says Hawkins, computers whose abilities extend beyond sense biology and provide a means to expand such complex fields as weather, cosmology and genetics.


No comments: