Making the ineffable explicit: estimating the information employed for face classifications