Partially affine invariant training using dense transform matrices
Melvin Robinson, MICHAEL T. MANRY · 2013
The concept of equivalent networks is reviewed as a method for testing algorithms for affine invariance. Partial affine invariance is defined and introduced to first order training through the development of linear transforms of the hidden layer's net function vector. The resulting two-step training algorithm has convergence properties that are comparable to Levenberg-Marquardt, but with fewer multiplies per iteration.