A Human-Markov Chain Monte Carlo Method For Investigating Facial Expression Categorization

This paper demonstrates how a human-Markov Chain Monte Carlo (MCMC) method can be used to investigate models of facial expression categorization. Data were collected from four participants. At each step participants were asked to select a representation from a pair, that most resembled a particular emotional state; this was repeated iteratively. As such, they formed a component in the MCMC process. The representations were line drawn facial images with 10 nodes and four degrees of freedom. The judgements formed samples for a set of interleaved Markov Chains. These were mapped to a twodimensional plane using Generalized Discriminant Analysis. We contrast the results of the MCMC task with those of a second discrimination task. Estimates of the distributions along each of the four dimensions showed that for the outer eyebrow and lip corner variables one of the categories could be discriminated with confidence. The average examples from both MCMC and discrimination tasks were both plausible. However, the MCMC method allowed for greater sampling from areas of high interest. Finally, we show that a naive Bayes classifier trained on the MCMC data can be used to successfully predict human classification in a discrimination task.