Using machine learning as a research tool in experimental psychology

Abstract

This dissertation evaluates how methodologies from machine learning can be applied in experimental psychology to gain new insights from neurophysiological data. Using two examples from memory psychology, the experimentally collected EEG data are evaluated once with classical group-level statistics and once with classification methods from the field of machine learning. The combination of the results of both methods shows that new insights can be gained that will profitably advance research in experimental psychology. The use of new methodologies in this area is necessary because conventional group-level statistics have problems that have spread extensively in science and have had serious consequences, especially in the replication crisis that started in the year 2000. The benefits of machine learning can help to alleviate these problems. In comparison to the use of group-level statistics alone, the combination of both methods allows data to be evaluated equally at both group and single-subject levels in order to obtain a complete picture of the data. Also, the information of individual regions can be compared and evaluated with that of an entire association of sensors collecting data. In this way, underlying patterns can also be considered. The addition of machine learning also enables explorative data analysis, which is not yet feasible in the area of group statistics.\\\\ In concrete terms, the application of machine learning techniques has made it possible to refine the characterization of executive functions and to draw up new hypotheses regarding episodic memory. Of great importance were methodologies that make the operation of machine learning processes transparent. This allowed the application to be legitimized and the results to be interpreted for a specific purpose. Furthermore, the comparison of the behavioral accuracy and the accuracy of the machine learning process was particularly valuable. In this comparison it could be shown that there is not necessarily a connection between the visual processing of an image and its active recognition. Both case studies were able to show in a representative manner which possibilities arise from the use of machine learning methods and thus present new findings which would not have been possible without the application of machine learning in this context.