Predicting Goal-directed Attention Control Using Inverse-Reinforcement Learning

Understanding how goals control behavior is a question ripe for interrogation by new methods from machine learning. These require large and labeled datasets to train models. To annotate large-scale image dataset with observed search fixations, we collected 16,184 fixations people searching either microwaves or clocks in of 4,366 images (MS-COCO). We then used this behaviorally-annotated the learning method inverse-reinforcement (IRL) learn target-specific reward functions policies these two target goals. Finally, learned predict 60 behavioral searchers (clock = 30, microwave 30) disjoint test kitchen scenes depicting both clock (thus controlling differences low-level contrast). found that IRL model predicted efficiency fixation-density maps using multiple metrics. Moreover, revealed patterns suggest, not just attention guidance features, but also scene context (e.g., along walls clocks). Using psychologically-meaningful principle reward, it possible visual features goal-directed control.