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

[Tsvi Achler]

The definitions of modularity as stated have some room for ambiguity.  To
be able to discuss these ideas in a common language it is important to have
the definitions be crystal clear.  I think it would be a good to open up a
discussion to settle on a common definition of modularity.

I suggest a definition based on fixed points of networks.
Fixed points are solutions of the network and ultimately the reason for the
networks. In supervised networks the fixed points are designed to be an
optimized cluster, average, or ideal data points learned through training. In
unsupervised networks the fixed points are the basis functions learned .
The actual origin (or purpose) of the fixed points do not matter in the
test.

The test is: if two networks can be separated from a single network and
their combined  fixed points remain the same then they are modular.  If the
fixed points change, they are not.

-Tsvi
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mailman.srv.cs.cmu.edu/pipermail/connectionists/attachments/20130403/2489dba3/attachment.html>