Connectionists: New paper on why modules evolve, and how to evolve modular neural networks

[Jaap Murre]

An approach I found was interesting but which has not been pursued very
much is a 'four-step' approach to modelling:

  1. genetic algorithm -> 2. neural network architecture -> 3. exposure to
patterns in environment -> 4. test for generalization

This corresponds very roughly to

  1. evolution -> 2. structured (e.g., modular) brain (coarse structure) ->
3. developmental phase (fine structure) -> 4. behavior

When we tried this approach, we were surprised by the interesting neural
architectures we obtained in a number recognition task. The evolved
architectures showed better generalization than the ones we had thought up
ourselves and they showed better performance (faster development and better
generalization). Also, many of the efficient architectures incorporated
elements we find in the visual system, such as coarse-grained versus
fine-grained processing (cf. dorsolateral vs ventral stream) and extensive
recurrent connectivity whereby higher areas feed back to lower areas.

Happel, B. L. M., & Murre, J. M. J. (1994). Design and evolution of modular
neural-network architectures. Neural Networks, 7(6-7), 985-1004.

We also did a mathematical analyses of the feasibility of modular neural
networks in terms of wiring length and brain volume, aiming to incorporate
as much as possible the quantitative neuroanatomy known in the early
nineties. We found that such architectures would indeed fit into our skull,
whereas say a sparse random architecture would not.

Murre, J. M. J., & Sturdy, D. P. F. (1995). The connectivity of the brain:
Multi-level quantitative analysis. Biological Cybernetics, 73, 529-545.

A surprising outcome to us in this study was that when working from first
principles, in large brain structures (above a certain number of neurons),
having a architecture with a cortex and white matter underneath is actually
more efficient (less volume) than when white and grey matter are mixed
together. Intuitively, a cortex architecture seems to waste space and
increase wiring length but this does not seem to be the case, which may
explain why large brains use it so much.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://mailman.srv.cs.cmu.edu/mailman/private/connectionists/attachments/20130227/ce05cdd2/attachment-0001.html>