What are the criticisms regarding the performance of HTM?

Question

I just recently learned about the existence of this hierarchical temporal memory (HTM). I already read the document Hierarchical Temporal Memory: Concepts, Theory and Terminology (by Jeff Hawkins and Dileep George), which seems rather easy to understand, but one red flag is that the document is neither peer-reviewed nor attempting to explain why it should work in details.

I tried to look around for some independent sources. I found a few papers that compare its performance against others, but none explain why it performs well (or not). I noticed some comments claiming that it was looked down by mainstream expert, but I was unable to find any actual criticisms.

What are the criticisms regarding the performance of HTM? Since HTM is meant to be generic, any domain-specific criticism should be related to a more fundamental problem.

Furthermore, there are a huge amount of training data to use, enough even for multiple months training session. Basically, any criticisms regarding size or length of training is not relevant.

Answer 1

Criticisms against Jeff Hawkins are well summarized in the following essay taken from http://www.theregister.co.uk/2014/03/29/hawkins_ai_feature/

I myself believe that the HTM theory has a huge potential and will be a foundation of true machine intelligence. IBM recently announced to back up the HTM theory and started the Cortical Learning Center including some hundred members.

http://www.technologyreview.com/news/536326/ibm-tests-mobile-computing-pioneers-controversial-brain-algorithms/

As is pointed out in the essay, Dillep George, a cofounder of Numenta, made the startup Vicarious, which attracted a huge amount of fund, of which fact suggests a potential of the HTM theory.

Source: Criticisms against Jeff Hawkins, The Register

His media-savvy, confident approach appears to have stirred up some ill feeling among other academics who point out, correctly, that Hawkins hasn't published widely, nor has he invented many ideas on his own.

Numenta has also had troubles, partly due to Hawkins' idiosyncratic view on how the brain works.

In 2010, for example, Numenta cofounder Dileep George left to found his own company, Vicarious, to pick some of the more low-hanging fruit in the promising field of AI. From what we understand, this amicable separation stemmed from a difference of opinion between George and Hawkins, as George tended towards a more mathematical approach, and Hawkins to a more biological one.

Hawkins has also come in for a bit of a drubbing from the intelligentsia, with NYU psychology professor Gary Marcus dismissing Numenta's approach in a New Yorker article headlined Steamrolled by Big Data.

Other academics El Reg interviewed for this article did not want to be quoted, as they felt Hawkins' lack of peer-reviewed papers combined with his entrepreneurial persona reduced the credibility of his entire approach.

Hawkins brushes off these criticisms and believes they come down to a difference of opinion between him and the AI intelligentsia.

"These are complex biological systems that were not designed by mathematical principles [that are] very difficult to formalize completely," he told us.

Answer 2

I've been studying HTM for a while. It's kinda cool. The default behavior of an HTM is to analyze temporal data. On the other hand, nowadays, you need a "fancy" neural network (e.g. a recurrent neural network) to model a sequence-to-sequence problem (e.g. a chatbot). But HTM can be naturally applied to this type of problems!

I'm planning on making some kind of interactive TV with it, with control data sent up with the visual data, then constrain the visual data to the control data (that is possible, but don't expect it to suddenly dream new video, that doesn't happen). But, I must say, the coolest thing about it is the video you teach it with won't come out on playback, rather it'll show the most typical version of it, which is its form of understanding, and actually making it really cool. So that's kinda like the dream you get out of it.

If you're using HTM with text,

1. it stores letters,
2. then it produces syllables.
3. then it takes these syllables and produces commonness out of these,
4. then it forms words out of the common syllables,
5. then it takes these words and determines what they have in common,
6. then it forms possibly larger words,
7. then it forms groups of words,
8. then it forms groups of sentences

So, each time, it goes up a level, it "forgets" a little more, just to maybe ingrain the groups together more firmly. These more closely ingrained groups will playback a little differently than the record playing into it. And this playback hasn't been seen by a lot of people.

I wonder if it comes up with its own sentences completely.

It makes the sentence grow a little slower, much better. It's much harder work than not forgetting anything and simply having a flat playback of the record. I'd say the slower your sentence grows, the better you did it, so if there is some more way to find typicality, you should do it.