Cluster Learning

Prior to training neural networks require an initial value for their synaptic weights. These values may be negative, positive or even zero, but, depending on the initial values the network error - the difference between the actual and the desired network output - is located at a particular position at the hyper dimensional error surface. If that location is far away from the global goal, e.g. there are several local minima in between the actual position at the surface and the global minimum, the chances are very likely to trap into one of the local minima.

To overcome that problem neural networks are several times initialized and trained, each time with different initialisation values for the weights chosen randomly. The process of cluster learning is designed to overcome that general time consuming process of re-initialisation and re-training several networks with identical architectures but different initialisation values.

cVision automatically creates and trains an arbitrary number of networks at the same time but with different random initializations and stores the preferred network, e.g. the network that yields the best solution after training is completed. The particular experts to be trained in the cluster can be selected from an expert repository, the number of networks trained in the expert cluster is arbitrary.