Technology Review article on Numenta

Thanks to Martin for bringing this article to my attention. It discusses how the new alorithms are sufficiently powerful that commercial applications of the technology are now imminent. I thought it was interesting that Itamar Arel was quoted with positive things to say about Numenta's tech. I have mentioned him before on this blog, and have wondered why he hasn't worked more closely with Numenta since they have similar goals. Arel has a competing deep learning system known as DESTIN that Ben Goertzel wants to use as the sensory perception portion of a child-like robot (if there is such a thing, the idea of a child-like robot seems to be a bit creepy to me as the parent of living, breathing children). Here the link to the article:

http://www.technologyreview.com/business/26811/

Another new Hawkins talk

On December 2, Jeff Hawkins gave a talk at Berkeley. It is similar to the talk from three weeks ago, but with some added tidbits sprinkled throughout. One nugget was Hawkins' statement that there is no existing machine learning model that comes even close to HTM for the depth to which it maps to the real cortical anatomy. This is exactly the point I made in my debate with Michael Anissimov on his blog. Here's the link:

http://www.archive.org/details/Redwood_Center_2010_12_02_vs265_26_Jeff_Hawkins

HTM hardware implementation

I came across an interesting Powerpoint document by Dan Hammerstrom, a Professor in the Electrical and Computer Engineering Department at Portland State University. He has in the past published papers discussing potential hardware versions of HTM, and works with the DARPA SYNAPSE team that is attempting to create brain-like hardware.

In any event, he is collaborating with Numenta to create a hardware implementation of the new learning algorithms (he calls them "HTM3" as opposed to the prior software ("HTM2"). Hammerstrom says that Numenta is running into serious scaling problems with the new algorithms due to the limitations of present-day CPU's, and they are concerned that this will impact the wide adoption of their algorithms. Interestingly, they have tried using GPU's, but it hasn't helped much, so they are looking at more custom hardware tailored specifically to their algorithms. Working with Hammerstrom, they are looking at three possibilities:

1. More optimal use of CPU's and GPU's
2. FPGA's
3. Custom silicon created specifically for Numenta

Now it is even more clear that Numenta is not focusing on computer vision yet because today's computers don't have the horsepower to run the software. In any event, here is a link:

http://web.cecs.pdx.edu/~strom/talks/hh_my_research_web.pdf

Latest Numenta newsletter

A few days ago Numenta sent out a newsletter with a quick update on their work. The newsletter notes that Numenta has now posted the Smith Group lecture on its website (it runs much more smoothly than the version on the University's website). It also announced the first additions/updates to the new learning algorithm documentation. The new additions were helpful, particularly the addition of an appendix that goes into some depth about the neuron model used by the HTM software. It includes some of the graphics used by Numenta in the online lecture.

Sadly, the newsletter noted that Numenta is temporarily deferring its work on computer vision problems in favor of applications that are more focused on temporal patterns, such as web click prediction and credit card fraud prediction. I guess that I can't say that I am too surprised by this. In hindsight, based on the online video and the whitepaper, it is clear that Numenta ran into some problems with its vision experiments with the new algorithms. The current algorithms can model layer three or four of the cortex (layer 3 for variable order time based learning or layer 4 for learning that does not rely on context). The whitepaper hypothesizes that layer four allows the brain to learn spatial invariance while layer three allows the brain to learn temporal invariance but that for vision problems the brain is somehow combining layers 3 and 4 to create spatial and temporal invariance at the same time. Until Numenta figures out how to model both layers at the same time working together like the real brain, computer vision probably isn't going to work terribly well.

Smithgroup Hawkins November 12 lecture available online

Surprisingly, Jeff Hawkins' recent lecture is now available online:

http://www.beckman.illinois.edu/gallery/video.aspx?webSiteID=o2RWiAWUQEKPgd_8QBTYOA&videoID=fYhfoB6NFE2ytFPl7XLnTA

Comparing the new algorithms to the first generation

It is interesting to compare the new algorithms to the original zeta1 algorithms released by Numenta back in 2007. In a May 2007 blog post, Numenta discussed the limitations of those algorithms. Here were the limitations noted at that time (see this link for the blog post):

1. The 2007 algorithms had no time based inference, so inference was based only on, for instance, a single snapshot of a picture to recognize an object. Now, of course, the algorithms fully employ time-based inference, which should make computer vision applications (and other applications) based on HTM much more powerful.

2. In 2007, time was used for learning, but it was only "first-order" time based learning. That meant that when the software was attempting to learn sequences of patterns, it would only account for the current time step and one prior time step. Imagine trying to learn invariant representations for dogs, cars, people, and other complex objects based on only two consecutive still pictures of data. Our brains learn by seeing many "movies" of objects in the world around us, so this was a very significant limitation on the power of HTM. Now, it appears that HTM can learn sequences of essentially unlimited length.

3. The 2007 algorithms had abrupt discrete "nodes" without overlapping boundaries. According to the blog, this diminished the ability of the system to create invariant representations. Now, the levels of the HTM hierarchy are one continuous region (no more nodes). This is a big change that I actually wasn't expecting, which is good, because continuous regions of neurons rather than discrete nodes is the way that the brain works.

4. The 2007 algorithms did not use sparse distributed representations, which also severely limited the scalability of the algorithms due to memory requirements. Now, it goes without saying that sparse distributed representations are the key to making the new algorithms work. Not only does this make the algorithms much, much more scalable, it also facilitates generalization.

In short, every single major listed shortcoming of the original HTM software has now been addressed. I expect to see many commercial applications start to come from Numenta's work. Hopefully my blog will soon be able to focus as much on applications as on the core technology. It will be interesting to see the extent to which this technology takes off over the next few years. Personally, I am particularly interested in robotics, and hope to see HTMs begin to be used to create robots that can intelligently perceive the world and perform useful tasks. Navigation, object recognition and manipulation, and language understanding are all things that could theoretically be done by HTM.

Dileep George's departure from Numenta is now permanent

That's a big loss, as Dileep was the guy who was able to take the ideas in On Intelligence and create a mathematical model and the first working software implementation of HTM.

According to Dileep's blog, he has now started a venture-capital funded company called Vicarious Systems, Inc. Its stated goal is to develop AI applications, starting with computer vision applications. If you go to vicariousinc.com, you can sign up for a corporate email newsletter.

It now makes sense that Numenta is trying to hire a Senior Software Engineer.