Abstract
Demonstrations of top-down effects in perception typically concern objects, events, and relations that are present in a scene. However, there is a class of relations that depend on what fails to be present: causal relations. Eye-tracking evidence reported in (Gerstenberg et al., 2017) suggests that subjects strategically deploy attention to non-actual events in service of causal judgment, but not when their task merely requires tracking actual outcomes. We have developed a computational model in the ARCADIA framework (Bridewell & Bello, 2016), that captures the differential patterns of attentional deployment reported in Gerstenberg et al. In Gerstenberg's task, a ball B moves towards a gate G. Meanwhile, a second ball A collides with B, with B subsequently entering (or missing) G. The model is interrogated either on whether B entered G or on A's causal contribution to B's entering G. For the causal judgement, the model initiates a coarse sweep of spatial attention along B's pre-collision trajectory, checking whether it intersects G. If the intersection is judged to be partial, finer-grained sweeps are attempted. For the outcome judgement, such attentional sweeps are not needed. The model generates ratings for each judgement that closely match human ratings, r(16) = .93, r(16) = .92, both ps < .001. Because eye movements in the model follow covert attention shifts, the model generates predictive saccades along B's pre-collision trajectory when it is making a causal judgement. Consistent with human behavior, the number of predictive saccades tracks the fineness of attentional sweeping required to determine whether B will enter G. These saccades correspond to what would have been the case for the B-G relationship prior to collision, and thus can represent non-actual events that inform causal judgment. Future research will explore the attentional strategies and associated eye movements used to verify other causal and spatial relations.
Meeting abstract presented at VSS 2018