Making deep neural networks right for the right scientific reasons by interacting with their explanations