Like a lot of geeks, I've been interested in how the brain works for most of my life. Artificial Intelligence was always one of my interests within computing (and part of what got me interested in linguistics at a very early age).
Within my linguistics research, I've always been interested in models that are biologically plausible so it was a huge delight to read Jeff Hawkins' On Intelligence back in early 2005 and find a theory that was biologically-based and believable from a linguistics point of view. One prominent psycholinguist told me in 2006 that it was one of the most promising theories he'd ever read.
After reading the book, I promptly went out and built a library (as I am wont to do) of about 20 books on general neuroscience, computational neuroscience and the relationship between the brain and language. I started thinking about how to implement the ideas and, after reading some of Jeff's and Dileep George's early papers, augmented the library further with books on Bayesian networks, belief propagation, etc.
When Jeff and Dileep started Numenta and eventually released an early version of their Hierarchical Temporal Memory (HTM) platform in Python, I was particularly excited to try it out, in particular applying it to linguistics. I started the htm-ling mailing list to gather other people interested in applying HTM to models of language. It turned out to be hard to get word out to other people interested in HTM and linguistics, however.
I never got very far with Numenta's code, mostly because there were just too many other things I was working on.
But then a couple of months ago, I found out Numenta was running a workshop / conference. I thought it would be an excellent opportunity for me to (a) get back up to speed with what Numenta was doing and how to use their NuPIC platform; (b) meet other people interested in applying HTM to linguistics.
So a couple of weeks ago, I attended the first Numenta HTM Workshop. I had a great time. It was great to meet Jeff and the rest of the team. Dileep's talk on the algorithms in NuPIC was particularly helpful to me in understanding how things work.
There were a number of people who expressed an interest in the application to linguistics so in the evening I ran a BOF. None of the attendees (as far as I could tell) were linguists by training so I didn't really get to talk too technically from a linguistics perspective. The boost to the mailing list membership hasn't created any more discussion yet either.
But I am still hopeful that an HTM-like approach (whether in the form of NuPIC or some other implementation) might be useful in building biologically-plausible models of language processing.
The original post had 2 comments I'm in the process of migrating over.