15 November 1995, 12:00, WeH 7220


	    Rehearsal and pseudorehearsal as solutions to
		 the catastrophic forgetting problem

		Anthony Robins (U. Otago, New Zealand)

How do you incorporate new information in to an existing artificial
neural network (ANN)?  In most ANNs learning new information results in
the "catastrophic forgetting" (significant disruption or complete loss
of) old information.  This problem makes typical ANNs very inflexible
once trained.  In this talk we describe a solution to this problem based
on rehearsal, the retraining of old information as the new information
is added.  This forces new information to be integrated in to the
information already stored in a network rather than simply overwriting
it.  We also show that it is possible to achieve this effect even when
we don't have the old information (original training population) to
rehearse.  This "pseudorehearsal" mechanism approximates the old
information by sampling the network with randomly constructed
"pseudoitems".  Both rehearsal and pseudorehearsal methods increase the
flexibility of ANN learning by allowing new information to be integrated
into an existing network with minimum disruption of old information.