Abstract
The human visual system has a remarkable ability to decide between multiple targets for fixation in complex perceptual environments. Evolution has refined this process to be both rapid and cheap, allowing over 100,000 saccades to be made every day. Previous work on modeling visual decision making emphasized value maximization and saliency. A recent study by Navalpakkam et al. combined these strategies, suggesting that expected value (EV) computed from the probability of target presence and the magnitude of target reward offered the best performance in modeling human behavior. However, EV strategies are often insufficient in modeling human preferences as individuals tend to exhibit a risk averse (RA) decision making strategy due to decreasing marginal utility for most types of reward. We propose an alternative model for visual decision making, maximizing utility as opposed to value under the decreasing marginal utility assumption. To test our model, we asked 10 UCSD graduate students to participate in a visual decision making experiment. Each trial consisted of a central fixation display, followed by a brief 500 ms presentation of two targets. Subjects were instructed to saccade to the target they perceive will maximize their reward during the stimulus presentation, and a feedback screen is displayed. Each target had the same expected value, but different variance. Risk averse subjects should choose the lower variance target, while the EV strategy will tend to choose the target with the highest recent average reward (and hence can perform above chance, even though the long term EV is equal). Our results show that the risk averse model significantly outperforms the expected value model (p <0.0001) in predicting human fixation location. This suggests that the dynamic decision-making of eye-movements are influenced not only by the expected value of reward, but also the variance of the reward distribution.
Meeting abstract presented at VSS 2012