Author ORCID Identifier
Date of Award
Executive Doctorate in Business (EDB)
Pamela S. Ellen
Advances in “deep learning” algorithms have led to intelligent systems that provide automated classifications of unstructured data. Until recently these systems could not provide the reasons behind a classification. This lack of “explainability” has led to resistance in applying these systems in some contexts. An intensive research and development effort to make such systems more transparent and interpretable has proposed and developed multiple types of explanation to address this challenge. Relatively little research has been conducted into how humans process these explanations. Theories and measures from areas of research in social cognition were selected to evaluate attribution of mental processes from intentional systems theory, measures of working memory demands from cognitive load theory, and self-efficacy from social cognition theory. Crowdsourced natural disaster damage assessment of aerial images was employed using a written assessment guideline as the task. The “Wizard of Oz” method was used to generate the damage assessment output of a simulated agent. The output and explanations contained errors consistent with transferring a deep learning system to a new disaster event. A between-subjects experiment was conducted where three types of natural language explanations were manipulated between conditions. Counterfactual explanations increased intrinsic cognitive load and made participants more aware of the challenges of the task. Explanations that described boundary conditions and failure modes (“hedging explanations”) decreased agreement with erroneous agent ratings without a detectable effect on cognitive load. However, these effects were not large enough to counteract decreases in self-efficacy and increases in erroneous agreement as a result of providing a causal explanation. The extraneous cognitive load generated by explanations had the strongest influence on self-efficacy in the task. Presenting all of the explanation types at the same time maximized cognitive load and agreement with erroneous simulated output. Perceived interdependence with the simulated agent was also associated with increases in self-efficacy; however, trust in the agent was not associated with differences in self-efficacy. These findings identify effects related to research areas which have developed methods to design tasks that may increase the effectiveness of explanations.
Dougherty, Sean, "Partnering People with Deep Learning Systems: Human Cognitive Effects of Explanations." Dissertation, Georgia State University, 2019.