Extreme learning machine with initialized hidden weight

The Extreme Learning Machine (ELM) is a recent training method for feedforward neural networks. Its main advantage is a faster and simpler training procedure when it is compared with traditional global search optimization method. It is achieved by using a least square solution for the output layer and random initialization for hidden layer. In this way only one solution is attained. In this sense, a question arises: is the random initialization method really an efficient for ELM? The present work studies the influence of more sophisticated methods of initialization, in terms of performance and complexity.