A competitive learning of three-layer neural networks

Summary form only given, as follows. A competitive learning algorithm called geometrical expansion learning (GEL) was proposed to train a three-layer neural network for an arbitrary function in discrete space. The most significant difference between GEL and backpropagation learning (BPL) is that GEL always guarantees the convergence, while the convergence of BPL is not known. Moreover, GEL automatically determines the required number of neurons in a hidden layer, which varies depending on the given training patterns. Also, the learning speed of GEL is much faster than that of BPL