Training multilayer networks with discrete activation functions

Efficient training of multilayer networks with discrete activation functions is a subject of considerable ongoing research. The use of these networks greatly reduces the complexity of the hardware implementation, provides tolerance to noise and improves the interpretation of the internal representations. Methods available in the literature mainly focus on two-state (binary) nodes and try to train these networks by approximating the gradient and modifying appropriately the gradient descent. However, they exhibit slow convergence speed and low possibility of success compared to networks with continuous activations. In this work, we propose an evolution-motivated approach, which is eminently suitable for networks with discrete output states and compare its performance with four other methods