Aligned explanations in neural networks