[eightwings]

I applaud Mr. Hawkins for bringing the temporal aspect of intelligence to the forefront of AI research. Others, like myself, have been saying it for years but only someone of Mr. Hawkins’ stature could do it and I am grateful. If I had to choose one word to characterize intelligence, if would have to be 'timing'. Unfortunately, the Marvin Minskys, Doug Lenats and Hans Moravecs of the AI community continue to retard progress in the field by clinging to the symbolic approach. That’s too bad. Unless they heed the writing on the wall, they will fall by the wayside, in my opinion.

Having said that, I have a bone to pick with Mr. Hawkins's approach to cortical learning. Although I agree that the sensory learning mechanism is hierarchical in nature, all the evidence (e.g., Markram, Seijnowski) I have seen so far indicates that the sensory cortex uses a fixed time scale for learning, about 10 ms. The hierarchical aspect is beyond doubt: specialized areas of the visual and auditory sensory cortices that show a gradation of recognition capabilities can be explained using a hierarchical system heavily dependent on feedback. However, you are not going to find any neuron in the sensory cortex that encodes for a dog or a cat. High-level recognition of the type that allows us to recognize a familiar face occurs somewhere else. For that, you need to go to the hippocampus, a very specialized cell assembly that bears little resemblance to the rest of the cortex.

Nothing in the sensory cortex indicates that individual neurons encodes for the probability that a particular signal or sequence will happen, a requirement of Bayesian learning. Cortical neurons are all or nothing neurons. They either fire or they don't. Certainly, synapses are strengthened or weakened according to some probabilistic rule but the evidence shows that it is a fixed probability.

That's not all. The cocktail party problem comes to mind: the brain has the ability to focus on one or more sequences of related events while ignoring all others. Mr. Hawkins and others seem to be under the impression that the Bayesian approach only needs to be scaled up in order to achieve the high level of recognition found in the brain. The history of Bayesian networks does not support this. Obviously, the brain is using a mechanism that has so far escaped our notice and eluded our imagination. Something is missing in our theories. I have my own ideas but this is neither the time nor the place to discuss them. I just hope that Mr. Hawkins does not close his eyes to all criticisms and is willing to retrace his steps every once in a while.
_________________
Louis Savain
Temporal Intelligence

[wlorenz65]