At the HTM workshop earlier this week, Numenta demoed their nupic 1.6 release and showed some pretty impressive advancements in the learning algorithms they have been developing. Their prototype HTM is a vision system that does recognition on objects like cows, ducks, and cell phones–they’ve trained it extensively and it can recognize these objects from pretty much any angle, and it recognizes objects both in photos and in crayon type sketches. This system should be commercial grade within a year, I expect.
Qualia labs presented some motion capture demos were astounding–their HTM was able to distinguish between 4 distinct action performed by lab actors with 98% accuracy. One of my main goals in attending the event was to assess the state of the current nupic platform; I was expecting to see a tool ready for experimentation, but I think now that within the next year the platform will be solving some real problems.
Some of the most interesting demos to me were those involving “unsupervised training.” In unsupervised training, you feed data into an HTM and tell it to identify n states. The HTM then identifies patterns in the data and categorizes the data into n-groups. A motion capture demo, for example, instructed an HTM to identify the motions of actors into 4 states: the HTM ended up dividing movements into walking, sitting, jumping, and running. Although this is a somewhat trivial example, one can imagine how unsupervised training could reveal interesting patterns in large data sets–for example, feed in data about people’s blood sugar levels or some piece of information about their genetic sequence. Tell an HTM to split the subjects into a few groups, and then correlate the results with data like presence of certain diseases.
When I arrived in Palo Alto, I had quite a casual interest in Numenta and expected that nupic would be a technology to incorporate into whatever else I’m working on in about two years. Perhaps because of the impressive demos, perhaps because of the energy you get just from going to an event like this where there are a bunch of people excited about something, or, perhaps by chance, I am definitely now considering diving more fully into nupic sooner than later.
