Improving Parameter Recovery for a Bayesian Model of Verticality Perception

Thumbnail Image
Issue Date
Journal Title
Journal ISSN
Volume Title
Alberts et al. 2016 proposed a Bayesian model designed to decompose verticality perception into constituent components whose contribution on the integrated percept can then be quantified using a psychometric experiment called the rod-and-frame task. Potentially, the model can be utilized as a diagnostic instrument in a clinical setting to determine the effectiveness of treatments related to vestibular disorders. As of now, the number of trials that is needed to estimate the model’s parameters using the rod-and-frame task deems this intention infeasible. This thesis examined the ability of ReLU based fully connected feed forward neural networks to approximate the Bayesian model in the upright condition and replace it in an adaptive stimulus selection procedure to investigate whether the model’s parameters could be recovered using an online evaluation of stimuli while at the same time remaining computationally cheap. It could be shown that the networks performed well on training data but significantly worse on unfamiliar data. Nevertheless, parameters could be partially recovered using a parameter setting familiar to the network. The average trial time of 2.14s that was measured during the procedure suggests the feasibility of a dynamical stimulus evaluation approach within adaptive stimulus selection.
Faculteit der Sociale Wetenschappen