Likelihood Based Evaluation for Scanpath Models in Scene Viewing

Seminar | May 30 | 12-1:30 p.m. | 560 Evans Hall

 Heiko Schütt, University of Tübingen

 Neuroscience Institute, Helen Wills

When observers view natural scenes their eye movements show elaborate statistical patterns beyond the fixation density over an image. An important approach to understand these patterns --- and thus ultimately to understand how humans choose where to look at --- is to build models which generate full scanpaths in natural scenes, i.e. a sequence of fixation locations. There are a multitude of different approaches to build scanpath models and to evaluate them statistically, however. Therefore, unifying and improving the statistical analysis of such models is essential. In my talk I will show that a likelihood can be calculated directly for virtually all scanpath models. Using our recent SceneWalk model as an example, I will illustrate how likelihood enables better model fitting including Bayesian inference to obtain reliable parameter estimates and corresponding credible intervals. Using hierarchical models, inference is even possible for individual observers. Furthermore, the likelihood can be used to compare different models. As an example I will show that the SceneWalk model produces more exact predictions than any model could by predicting only a static fixation density the way saliency models do. Additionally, the likelihood based evaluation differentiates model variants, which produced indistinguishable predictions on hitherto used statistics. Beyond the application to scanpath models, a direct computation of the likelihood might be an interesting approach for any other models which predict sequentially dependent human behaviour.

 nrterranova@berkeley.edu