Gradient-Free Neural Network Training on the Edge