Bridging the expectation-reality gap in machine learning