Using flexible neural trees to seed backpropagation

Abstract

Neural networks are a powerful computational architecture for modeling data, but optimizing the connection weights can be very difficult. Flexible neural trees (FNTs) are good at finding a globally near-optimal network to fit a dataset, using evolutionary algorithms and particle swarm optimization. We show that putting the two methods together can yield very good results. The FNT solution can be embedded into a larger neural network that is then optimized using backpropagation. The combination of the two methods outperforms either method alone.