Tobias Thomas^1,2 (), Constantin A. Rothkopf^1,2; ¹Technische Universität Darmstadt, ²Hessian Center for Artificial Intelligence (hessian.AI)

The human visual system relies on saccadic eye movements to bring information from the environment to the high acuity region of the retina. The decision-making process underlying each shift has predominantly been studied to investigate the effect of stimulus or task properties. Yet, in recent years, the idea that gaze shifts themselves might vary in cost has gained prominence, with research both qualitatively and quantitatively describing these effects, as well as demonstrating that they influence behavior in different contexts. In this work, we propose two preference elicitation experiments that enable us to quantify the effects of most previously discussed spatial saccadic features. These include the amplitude, direction, change in direction, and orbital eye position. In both experiments, subjects repeatedly revealed their preferences by making a forced choice gaze shift to one of two targets on the screen. While the first experiment included forced return saccades to the center of the field of view to ensure that all decisions started there, the second experiment generated a carefully designed sequence of gaze targets across all parameters, thereby additionally investigating the influence of the orbital eye position. All subjects showed a linear preference for smaller saccades; however, for most other features, we found idiosyncratic behavior that was highly consistent across experiments and across days. To quantify the individual differences in preferences and facilitate gaze prediction, we inferred parametric cost functions using a random utility model. The inferred costs of the model align with the empirically observed preferences for all saccadic features, and the model is able to predict approximately 80% of the decisions across participants. In summary, using two preference elicitation experiments designed to investigate factors contributing to saccadic costs, along with a random utility choice model, allows for the characterization of individual differences in gaze selection.

Acknowledgements: The authors acknowledge the support of the cluster project ‘The Third Wave of AI’ as part of the Excellence Program of the Hessian Ministry of Higher Education, Science, Research and Art