Looking inside machine learning black boxes