Jeff Hawkins and Hierarchical Temporal Memory
Tuesday, June 19th, 2007I finally got around to looking at some of this stuff on Jeff Hawkins and his Hierarchical Temporal Memory (HTM) that people have been sending me. Hawkins is trying to build a biologically plausible model of learning and memory in the brain that will supposedly be general across different tasks (i.e. it works for vision processing, audio, motor stuff, etc). What he has come up with is called an HTM, and it is a hierarchical network of nodes that use belief propagation. Here is a video from 2003 where Hawkins lays out some of his ideas about the brain and his belief that intelligence is better defined as an ability to predict the future. This video is from 2006 and is a presentation were Hawkins describes the theory and implementation of the HTM system.
I think the idea of the HTM system is very cool. I really think we should be looking to the brain for ideas on good learning and memory systems and doing it hierarchically makes a lot of sense. The system Hawkins shows seems to be like a sort of hierarchical structure of Self-Organizing Maps (SOM), which is a method for unsupervised learning and clustering. His method connects these different nodes using belief propagation so they all agree on the same thing. And he has some unspecified method for detecting temporal sequences. But otherwise it seems a lot like a SOM, grouping similar sets of inputs together and being able to provide the closest matching set from memory. The results he shows on a simple object recognition task look very much like something you would see from a SOM or neural network. It would be cool to see some results where the temporal sequencing or the hierarchical structure come into play.
Even though I think the results of the work are a bit weak so far, I still think it is pretty cool. I’m still trying to figure out what direction I want to go in for my research and this gives me some ideas. I would definitely like to work on something motivated by or inspired by the brain. And it should be general enough to be applicable to many different problems. It will be interesting to see what kind of results Hawkins gets out of the HTM as it gets extended and applied to more interesting problems.