Digital hardware realization of a hyper basis function network for on-line learning

The proposed paper describes a digital neural network hardware realization performing a hyper basis function network for function approximation. Both, learning and recall of the network are implemented in hardware to achieve a high performance network calculation. This opens the use of the function approximator to applications with real-time learning requirements for on-line learning. The presented hardware uses a flexible network structure, i.e. the number of basis functions is not fixed in advance, but they are integrated into the network during learning depending on the learning data set. Thus, we have a good approximation result by using a minimal number of basis functions.