Opening the black box of neural nets: case studies in stop/top discrimination