Revealing interpretable object representations from human behavior