TerminalDigit

Right on. I’m glad I’m not the only person who thought this.

The interesting thing is that Jeff is right in so many ways, and most of the skilled neuroscientists agree that mountcastle’s ideas are in fact correct. And Jeff is also right in that most of the way the world works can boil down into a simplistic algorithm.

Thanks for your input, Phillip Shoemaker. When you run around the blogosphere trying to do PR for Jeff, you ought to disclose that you’re the Director of Developer Programs for Numenta, Inc. That information seems relevant for the appropriate interpretation of your remarks.

But what do I know? I’m just a lowly neuroscientist who apparently needs to be told by someone with a vested interest in spinning the public opinion towards HTM what my colleagues think of Jeff’s little treatise. You may wish to read Mountcastle’s paper yourself. He does a decent job describing the functional organization of certain regions of cortex into columnar units, but at no point does he claim that he knows the entire cortex to be organized thusly, nor does he make the gigantic leap of faith required to conclude that if you understand a cortical column, you understand how to build a brain capable of human-like intelligence. Even if 100% of the cortex were comprised of identical columnar units (it isn’t), you won’t find a neurobiologist on the planet willing to state that they know for sure that 100% of language learning or object recognition actually occurs at the cortical level.

In any case, it’s clear that neither Jeff nor you are scientists. Real scientists consider Jeff (and everyone else) to be wrong in every single way until he proves himself right. As of this writing, HTM has failed to do what Jeff claims intelligent computers should be able to do: learn languages from scratch, perform robust visual object recognition, etc. It performs at the level of AI, not at the level of human brains. Just because your friend wrote a book (and signs your paychecks), that doesn’t make his claims accurate. His ideas will be considered right when (if) they produce the grandiose results that he claims, and when these results can be independently reproduced, passing peer-review, repeatedly. I encourage Jeff to generate these results and submit them to a peer-reviewed journal so he can find out exactly how right he is. Thus far, I’ve only seen HTM cited in conference proceedings which demonstrate its ability to perform similarly to other AI algorithms (i.e. substantially inferior to human performance).