Generalization error for a bar-counting multilayer neural network
August Romeo · Network Computation in Neural Systems · 1993
I investigate a supervised process of learning from examples in two-dimensional feature detectors, implemented by feed-forward multilayer neural networks. In this paper, the method is applied, as a special example, to the construction of systems capable of distinguishing T from C shapes irrespective of translations and rotations modulo π/2. The generalization function is calculated, and the average learning and generalization curves are obtained by Monte Carlo simulations. Different behaviours-apparently continuous decay and discontinuous phase transition-show up, depending on the design employed. The properties of the nets are in accordance with those of similar models proposed in the literature.