| |
Abstract:
Some neural network architectures are unique in that their
implementation is considered to be a good reflection of the neural
substrate. Consequently, it has been assumed that the behavioral
deficits noted following brain injury can be modeled by "lesioning"
a neural network. Unfortunately the validity of this assumption is
contingent upon network architecture, size, and the method by which
the artificial lesion is created. In order to determine the extent
to which artificial lesions actually do reflect natural lesions,
simulations were carried out to determine: (1) differential effects
of lesion locus, in particular, how lesioning a pathway differs
from lesioning a unit pool, as well as how globally applied lesions
relate to focal lesions, and (2) whether the method used to apply
the lesion might have a differential impact on network performance,
i.e. lesioning the network through the application of noise as
opposed to by zeroing connection weights. Simulations were carried
out on a fully interconnected Hopfield network comprised of 100
orthographic and phonological units, and 200 semantic units. The
simulations found: (1) lesioning via zeroing results in crippling
damage, even at low levels of damage, e.g. 10%, (2) pathways may be
more resilient to damage, depending on whether alternative
undamaged pathways exist, (3) lesioning via the addition of noise
gives a much greater degree of control over the extent of damage,
and (4) focal and global network lesions using noise can be used to
model patient performance for patients suffering from either focal
lesions or global dementing processes, respectively.
|
