August 2014
Volume 14, Issue 10
Free
Vision Sciences Society Annual Meeting Abstract  |   August 2014
Predicting the task from eye movements using multivariate pattern analysis
Author Affiliations
  • Grigori Yourganov
    Department of Psychology, University of South Carolina, Columbia, SC, USA
  • Marc Berman
    Department of Psychology, University of South Carolina, Columbia, SC, USA
  • John Henderson
    Department of Psychology, University of South Carolina, Columbia, SC, USA
Journal of Vision August 2014, Vol.14, 765. doi:https://doi.org/10.1167/14.10.765
  • Views
  • Share
  • Tools
    • Alerts
      ×
      This feature is available to authenticated users only.
      Sign In or Create an Account ×
    • Get Citation

      Grigori Yourganov, Marc Berman, John Henderson; Predicting the task from eye movements using multivariate pattern analysis. Journal of Vision 2014;14(10):765. https://doi.org/10.1167/14.10.765.

      Download citation file:


      © ARVO (1962-2015); The Authors (2016-present)

      ×
  • Supplements
Abstract

In our study, we investigated the use of multivariate pattern analysis to predict a subject's viewing task from their eye movements. Previous research indicates that prediction is possible if the visual stimuli are different across tasks (Henderson, Shinkareva, Wang, Luke, & Olejarczyk, 2013). However, one study suggests that if the same stimuli are used in all tasks, prediction is no better than random guessing (Greene, Liu, & Wolfe, 2012). To investigate this question further, we recorded eye movements from 72 subjects performing 3 tasks on the same set of real-world visual scenes: 1) a visual search task, 2) a scene memorization task, and 3) an aesthetic evaluation task. A set of classifiers (linear and nonlinear, univariate and multivariate) was used to predict the task for a particular trial using 7 features of eye movements recorded during this trial (number of fixations; mean, standard deviation, and skewness of fixation durations and of saccade amplitudes). All classifiers were successful in predicting the task when trained on the other trials recorded from the same subject. Linear classifiers, particularly Fisher's linear discriminant (FLD), were also successful in predicting the task for each subject when the training data came from the other subjects (mean prediction accuracy of FLD was 56%, with random guessing corresponding to 33%). These results suggest that the homoscedastic multivariate Gaussian model effectively captures predictive eye movement information that is specific to a task and generalizes across subjects. For each pair of tasks, we computed the loadings of each of our Z-scored features to determine the importance of each feature in predicting the task. The two most relevant features, on average, were the number of fixations and the mean saccade amplitude. The skewness of saccade amplitudes was the least important, but still contributed to prediction.

Meeting abstract presented at VSS 2014

×
×

This PDF is available to Subscribers Only

Sign in or purchase a subscription to access this content. ×

You must be signed into an individual account to use this feature.

×