Modularity: a natural choice for improving generalization.
To appear in: Proceedings of
WSES'2002 International conference on Neural Networks applications.
Leonardo Franco
Cognitive Neuroscience Sector
SISSA - Trieste - Italy
and
Jose M. Jerez
Depto. de lenguajes y Cs. de la Computacion
Univ. de Malaga
Malaga - SPAIN
One of the principal motivations for constructing artificial
modular
neural networks comes from biological neural systems,
where modularity can be observed at different levels of organization. Starting
from some theoretical results that roughly state that
modular architectures generalize better than their monolithic counterparts,
we test through simulations this assertion on three problems: the parity
function
and other two ``real world" problems as face expression
recognition and diabetes diagnosis. We also analyze how the size
of the networks influences the generalization ability. From the results
we extract some general recommendations on how to build and train modular
architectures.