Towards a unified model of cortical computation I: Data compression and data reconstruction
architecture using dynamic feedback
T. Fomin, J. Körmendy-Rácz and A. Lõrincz
Neural Network World
7,
121--136
(1997)
Abstract
A dynamic connectionist data compression and
reconstruction (DCR) network is
introduced. The network features fast learning capabilities, dynamic
feedback of the output to the input, and apparent competition.
It is shown that the data reconstruction procedure of the DCR network
is equivalent to Wittmeyer's iterative method. Comparisons
with a soft competition network, the Hebbian and anti-Hebbian network,
and with principal component analysis
demonstrate the superiority of the DCR network in terms of
learning time since the network exhibits
similar reconstruction abilities to the other networks and can make use
of but does not require a slow tuning procedure. It is demonstrated
that the DCR network can be added on top of other networks
to improve reconstruction performance.
