Growing smaller networks with the tiling algorithm

Within the framework of multilayer feedforward networks, detailed design of the network-number of layers and the number of units within each layer-has long been a concern. The trial and error process of adding units and/or layers to the network is successful in sofar as it often achieves near-optimal performance but it is very time consuming. Constructive algorithms would be the ideal solution if the resulting networks generalised as well as their trial-and-error counterparts. We examine the tiling algorithm of Mezard and Nadal (1989) in greater detail. By choosing various alternative strategies for training the ancillary units in each layer, we are able to reduce the size of the network and improve generalisation