Transformation of back-propagation networks in multiresolution learning

We proposed to train a backpropagation network using the multiresolution learning method. We first train a backpropagation network with input at a coarsest resolution, then gradually refine the input into finer resolution. The problem is to transform the connection weights of the backpropagation network from a coarsest level into a finest level. The objective of it is to improve the convergence rate of the networks. We evaluate two schemes for this network transformation. One is the wavelet approach and the other is the average splitting approach. Experimental results demonstrated the ability of this approach