Biologically motivated neuronal models have become popular in auto-associative recurrent networks due to their ability to solve the binding problem and to segment complex scenes in to previously stored components. Most approaches only use simple Hebbian learning, which works best for orthogonal patterns. This paper presents a learning algorithm based on perceptron learning, which enhances the storage capability in such neural networks and allows correlated patterns. As these iterative learning algorithms allow weights to grow arbitrarily, the amount of network input may also grow arbitrarily and can cause de-synchronization. We therefore incorporate a method to ensure a constant network input for trained patterns while facilitating the switching from one attractor to a different one when a sequence of patterns is generated.