Backpropagation converges for multi-layered networks and linearly-separable patterns

Summary form only given. Backpropagation can fail to discover the optimal solution, since it can get stuck in local minima. In this paper it is proved that it converges provided that the patterns are linearly separable. This is also true for networks with hidden units. In this case, the experience gained in several experiments shows that multilayered networks surpass perceptrons in generalization to new examples