I went to a talk by David MacKay today on distributed phase codes for associative memories. One of his demos used a program called Dasher, which is a text entry system with a novel interface that could be used by disabled people. It’s baffling at first, but I imagine it becomes quite intuitive after a while. There are some demos on the website — it’s definitely worth checking out.
As far as the talk went, I have to admit I was a little lost, since everything I know about Hebbian learning and associative memories could fit on a 4×6 index card, probably. The main point of the talk was that utilizing inter-neuron spike times (or phases) and coincidence detectors that look for spatiotemporal spike pattens (with given delays) can produce structures that learn several patterns within the same neurons and can recall multiple patterns simultaneously. It’s an interesting idea, but the presentation was math-poor, so I ended up with very little idea about the “pattern capacity” of these memories, the effects of noise (and how it was modeled). These prosaic engineering questions weren’t really the focus of the talk, however, but maybe I’ll do the back-of-the-envelope calculations later.