| |
Abstract:
Abstract: Neural maps have been shown to possess complex and
interesting computational properties. This is very exciting,
because this class of algorithms have been used to model the
organization of primary sensory areas, as well as developmental
disorders related to perceptual difficulties. But can the
computational principles of self-organization be of help in
modelling superior psychological functions? In this study we set
out to investigate neural maps from the standpoint of learning
theory. This puts our simulations on a solid footing, i.e. it
allows us to specify with precision what computational principles
are involved (bias vs. variance modulation, or approximation vs.
regularization trade-off). Armed with these theoretical results we
propose an original model of the influence of attention on learning
that we apply to the observations of Merzenich and Recanzano on the
cortex of macaques. The changes in the cortical maps brought about
by attention-modulated learning are recast in learning-theoretic
terms as approximation control. The thrust of our poster is the
suggestion that the filtering functions related to attention solve
a fundamental problem of learning. By delimiting a portion of the
input that must be apprehended with precision, they limit the
complexity of a learning problem that would otherwise be impossible
to solve.
|
