| A new neural network learning algorithm, called the Umbrella Algorithm, is developed and analysed. Its generalization, which does not exhibit over-specialisation, is observed in the EXOR problem and in an artificial data discrimination (Toy Data) problem. The learning time is found to be about 1/15 of backpropagation learning time for the parity problem. The algorithm is applied to cosmic high energy gamma ray detection and is further used for determining the spectral index of the Markarian 421 source. |
