Learning pop-out detection: building representations for conflicting target-distractor relationships