Ever since then, a multitude of work has been conducted regarding the effect of task-set on oculomotor behavior (for a review, see
Henderson & Hollingworth, 1998;
Vo & Wolfe, 2015). Indeed, Yarbus’ first discoveries have been replicated many times over the years. These studies have shown that oculomotor measures, such as fixations and saccades, may vary as a function of task set. Since then, the field of decoding the underlying cognitive mechanisms behind these task-induced differences has been influenced greatly by advancements in pattern classification. In other fields of research, state-of-the-art machine learning methods for pattern classification have been applied to a wide range of research domains, as well as practical applications (
Kotsiantis, Zaharakis, & Pintelas, 2007). Using a more data driven perspective, recent work has shown the application of such models in decoding the underlying cognitive mechanisms behind differences induced by different viewing tasks. Early data-driven approaches at such decoding models were not deemed successful (
Greene, Liu, & Wolfe, 2012). However, Borji and Itti found that it was indeed possible to decode the observers’ task by eye-movement features by expanding the search for a better model (
Borji and Itti, 2014). Additionally, Henderson and colleagues have reported similar results during photograph, scene, and text viewing by using a naïve Bayes classifier model (
Henderson et al., 2013). Similarly, other work has propagated probabilistic models that enable the modeling of time dimensions in eye-tracking data (
Haji-Abolhassani, & Clark, 2014;
MacInnes, Hunt, Clarke, & Dodd, 2018). Additionally,
Tseng, Cameron, Pari, Reynolds, Munoz, and Itti (2013) have shown promising results applying a Support Vector Machine classifier on viewing behavior in an attempt at classifying neurological disorders. Finally, it seems that human observers are also able to classify different trial conditions. In a study by
Bahle, Mills, and Dodd (2017), human observers were able to correctly classify conditions above chance when viewing images overlaid with oculomotor metrics (e.g., fixation locations and durations, scan paths) although to a very low degree of proficiency. However, the results here imply that humans themselves may not be able to parse task information from eye movements alone, further emphasizing the importance of machine-learning approaching for decoding cognitive mechanisms.