Reward produces learning of a consciously inaccessible feature