also the network could have weird results for dropout values around 20%-30% depending on how the robustness was 'learned'