The design of multi-layer perceptrons using building blocks

A building-block approach for constructing large backpropagation (BP) neural networks is described. This results in considerably less training time than conventional BP, which starts from random initial weights. Unlike previous approaches, this approach involves the mapping of conventional algorithms onto neural network structures. This has several benefits. First, it produces alternative parallel structures for implementation of conventional signal processing algorithms. Second, it produces a good set of initial weights for BP training. An example is given in which a randomized initial weight network fails to learn, but the assembled network succeeds