Partially trained neural networks based on partition of unity

The authors propose partially trained neural networks (PTNNs) where only a part of connection weights are trained at a time to improve generalization, learning speed, computational time and incremental learning capabilities. PTNNs are composed of many small neural network fractions and firing neurons. The firing neuron fires a fraction of the PTNN depending on input patterns. The main features of the PTNN are partial update of weights, self-determined network size, no corruption of the old learning, reduced computational time, reduced connections, and fast convergence for a complicated problem. Simulations reported that the learning speed and computational time of PTNNs were superior to those of the standard neural networks for a complicated continuous function and the two-spiral problem