Though this end-to-end nature also makes it harder to spot biases and errors. There is some work that first trains with deep learning and then tries to infer a simpler model that is understandable for humans.