Do not update in backpropagation

Currently, the update is done in the forward pass, this is done for simplicity purposes when developping 0.0.1 but we should override the update() torch method.