Storage Capacity of Ram-based Neural Networks: Pyramids

Recently the authors developed a modular approach to assess the storage capacity of RAM-based neural networks 1] that can be applied to any architecture. It is based on collisions of information during the learning process. It has already been applied to the GNU architecture. In this paper, the technique is applied to the pyramid. The results explain practical problems reported in the literature and agree with experimental data and theoretical results obtained by architecture-speciic techniques developed by other research groups. The successful application of that modular approach to the pyramid and GNU architectures | two out of the three most common ones | show it to be a useful tool for RAM-based neural networks speciication for practical applications.