Abstract

The method of generalized projections is applied to the multilayer feed-forward neural-net problem to derive a new learning algorithm.¹ Unfortunately, strict adherence to the learning rule requires the solution of coupled, nonlinear algebraic equations. By permitting a simplification, a new learning rule results, which produces an implementable algorithm. We call this learning rule the projection-method learning rule (PMLR). We apply the PMLR to two well-known pattern-recognition problems, neither of which is solvable by a linear-discriminant scheme. The PMLR is compared with the error-back-propagation learning rule (BPLR) and is shown to converge faster than the BPLR for the problems being considered.

© 1990 Optical Society of America