An attempt is made to study learning in neural networks with local minima. For small learning parameters η, the transition time from one mimimum to another is asymptotically given by exp(η̃/η), with η̃, a constant independent of η, called the reference learning parameter. A general scheme to calculate the reference learning parameter is presented. This scheme is valid for a large class of learning rules.

• Received 5 May 1992

DOI:https://doi.org/10.1103/PhysRevA.46.5221