A simple strategy for building and training multilayer neural networks

Summary form only given, as follows. A multilayer neural network building algorithm is proposed: the offset algorithm. The algorithm has two basic steps-a growth step in which two hidden layers are built and a pruning step in which any redundant units are removed. During the growth step, the first hidden layer is built by adding units to offset errors, as they are needed, until zero error convergence is achieved. The problem of mapping these internal representations onto the desired output is the n-parity problem. Thus, the second hidden layer is built by a geometrical design procedure with no learning. After the pruning step, the final architecture can have one or two hidden layers. The offset strategy is simple and could easily be turned into hardware. The possibilities of its VLSI implementation have also been investigated