On initialization of neural network

The paper concerns the problem of initialization in neural networks. We focus on backpropagation networks with one hidden layer. The initialization of the weights is crucial: if the network is incorrectly initialized, it converges to local minima. So, the classical random initialization appears as a very bad solution. If we consider the problem's specificity and the non linearity of sigmoids, improvements can be very significant. We propose a new initialization scheme based on the search for an explicit solution to the problem. We study two cases with different hypotheses. Simulation results are presented showing that these original initializations allow us to avoid local minima, to reduce training time, to obtain a better generalization and to estimate the network size.