[memolazarte]

How can HTM will be able to tell the difference between say Bush and Clinton faces? Will HTM be a good method for performing this kind of biometric tasks or will be yet another pattern classification system?

As far as I understand from the theory, HTM generates an invariant representation of the world and I suspect details of the actual identity of objects are discarded in favour of robustnest to handle subject variability.

Am I missing something?

[_Big_Mac_]

We humans have huge parts of the brain devoted almost entirely to face recognition, the process is not only cortical AFAIK. But it theoretically should be possible to build face recognition only on this model.

Details build generalized representations. Different sets of details build different sets of representations. Practically there's some balancing to be done between either quick learning with coarse representations and extensive learning with a wide set of causes. You can either have your AI do "okay: eye, eye, nose, eye - it's a face!" which is relatively easy to teach, robust and fast, or "twitching eye, nose a bit fat on the bottom, thin lip.... - it's John Smith". The latter would be bigger (memory wise), slower and harder to teach (LOTS of training sets) but theoretically doable.

Our brain can do both types - we can do fast/coarse modeling (like for example when no distinct orientation points are visible, a bunch of trees is quickly represented as a generic forest) or detailed ones (like faces or when we pay extra attention to things, they become less generic). Those cases are a bit more complicated because they either need Attention or specialized perceptive modules included. Or both.