The maximum likelihood estimation learning neural network
C.C. Chiang, H.C. Fu · 2002
Summary form only given, as follows. The authors propose a three-layer neural network using the maximum likelihood estimation method as the training rules. The proposed network generates hidden neuron units dynamically during the training phase. The simulation results show two exciting properties in the proposed neural network; high-speed learning and small network size. For the two-spiral problem, the number of training epochs required ranges from 21 to 30, with an average of 25 epochs. The total number of dynamically created hidden units ranges from 41 to 55 with an average of 48 units. For 2-b XOR problems and the 10-b contiguity problem, only two training epochs and two hidden units are required.>