Chunking Structures and Brains

I have been thinking about is something I read a long time ago about the key to efficiently absorbing information was to get good at pick out and internalising salient features, and specifically that this and the capability difference between the novice and the master (slow/fast, dumb/smart) in a given field (It was some public digestible take on the research. Ring any bells?). These thoughts were provoked by Jeff Moser's post on becoming a grand master software developer via an active chunking methodology. It is a great post, although I am less interested in the notions as an approach toward being an awesome developer, and more interested in the broader notions as they relate to efficient associative systems and data structures. Specifically, Jeff's coverage of hierarchical chunking suggests (to me) at procedures toward automated abstraction along the lines of a transition from sub-symbolic toward symbolic meaningful for a specific context (the grand visions of vast hierarchical Artificial Neural Networks or Genetic Programs).

Toward this end, I was thinking about the TED talk by Neuroanatomist Jill Bolte Taylor on her experiences studying her own massive stroke 12 years ago. It was a great talk, although I was fascinated by her initial description of the brain. Specifically, she described the two hemispheres in terms of the analytical left focused on association and anticipation, and the right concerned with focusing and integrating sensory input (she phrased it all much better than I can, watch the video). Anyway, I was interested in the computational properties of such a system, specifically if such assumptions are made in the design of a toy system, how do you get the two parts to cooperate to do useful things. I was also thinking along the lines of how such a system would get started, specifically associations cannot be made without a knowledge base. The usefulness of the associations (in terms of forward and backward propagation of the associations for memory reuse and anticipation) would be biased by the extent and accuracy of the association-drive knowledge base. Further, could such knowledge could be built up using a hierarchical chunking method starting with sub-symbolic cases until associations start being useful - symbolic in the sense that they or their effects are meaningful in the world.

Anyway, I was mulling on this and came across the notion that children's memory may be more reliable than that of adults because of the lack of prior association bias (they are more specific, and call it Fuzzy Trace Theory). These ideas were familiar to me because of the similar features you see in adaptive systems like the immune system (something I've learned a few things about), where something like the finite capability of novel antibody generation is used up as in the specific pathogenic exposures of a hosts environment (along the lines of time of exposure bias in the case of Antigenic Sin, and the ongoing biasing of the capability of the repertoire).