New approximation method for smooth error backpropagation in a quantron network

In this work, we propose a new approximation method to perform error backpropagation in a quantron network while avoiding the silent neuron problem that usually affects networks of realistic neurons. In our experiments, we train quantron networks to solve the XOR problem and other nonlinear classification problems. We achieve this while using less parameters than the number necessary to solve the same problems with networks of perceptrons or spiking neurons.