Abstract
Successful prediction may be the main objective of any cognitive system. We investigate how humans solve the underconstrained problem of spatial extrapolation of complex 2D curves.
Contour continuation problems are usually modeled based on local variational principles such as minimizing total curvature or variation of curvature and have only been applied to a very limited set of curves classes. Singh&Fulvio (2006) showed that only a probabilistic model of these constraints can explain the continuation of arcs of Euler spirals.
We investigate whether the same model generalizes to more complex curves such as polynomials or regular curves, how much higher-order information is used and how the extrapolation is influenced by the experimental design. Subjects report either the position of a single curve dot behind a semi-circular occluder or draw the distribution of possible extrapolants.
Our results show that subjects report multiple curves for more complex curves if they are given the chance (e.g., by drawing distributions or by probability matching). We approach the extrapolation problem from a more global Bayesian perspective and show that the results are more consistent with Bayesian model averaging rather than purely local models.
Manish Singh, Roland Fleming