How do you approach debugging in a machine learning model?