Associative Memory
- I've spent a lot of time working with associative memory in
one guise or another.
- It occurred to me that I'd like to make use of a large number
of neurons available in SpiNNaker. So, I thought I could cache
out a large symbolic associative memory to neurons. Wordnet has
117,000 words, with an inheritence hierachy.
- I figured out a way to implement. You calculate the weights
so that it spreads up the hierarchy. Each concept has 10 neurons
so it's not too big, and connections are small
- This works in NEST and SpiNNaker on a small scale.
- I'm currently exploring Prepostional Phrase Attachment Ambiguity
resolution as a task. Learning is off chip, but is based on
firing behaviour.
- The learning off chip (or off NEST) uses co-firing behaviour; it
occurred to me that LTP/D takes more than a few ms, so any rule
that does that is inaccurate. I could implement this on chip, with
updates say once a day for long-time learning.
- I also did some associative memory stuff with neural learning
using NEST, writing my own compensatory synapses and neurons.
- I also implemented a small scale generic associative memory, where
the associations are learned.
- I think I'm building up a nice suite of systems and tools for exploring
associative memory.
- We put out a paper on two simple neuro-cognitive models of associative
memory. It's in the cogsci offshoot IC in Cognitive Modelling in
Madison in June. It models the Stroop task (no learning) and a
a question answering task from Quillian and Collins showing
human like semantic net behaviour. We're also putting out a better
Stroop paper to SGAI for Cambridge in December.
