Assessing Human Understanding of Machine Learning Models

Aus SDQ-Institutsseminar
Vortragende(r) Moritz Renftle
Vortragstyp Masterarbeit
Betreuer(in) Holger Trittenbach
Termin Fr 5. Juni 2020
Kurzfassung To deploy an ML model in practice, a stakeholder needs to understand the behaviour and implications of this model. To help stakeholders develop this understanding, researchers propose a variety of technical approaches, so called eXplainable Artificial Intelligence (XAI). Current XAI approaches follow very task- or model-specific objectives. There is currently no consensus on a generic method to evaluate most of these technical solutions. This complicates comparing different XAI approaches and choosing an appropriate solution in practice. To address this problem, we formally define two generic experiments to measure human understanding of ML models. From these definitions we derive two technical strategies to improve understanding, namely (1) training a surrogate model and (2) translating inputs and outputs to effectively perceivable features. We think that most existing XAI approaches only focus on the first strategy. Moreover, we show that established methods to train ML models can also help stakeholders to better understand ML models. In particular, they help to mitigate cognitive biases. In a case study, we demonstrate that our experiments are practically feasible and useful. We suggest that future research on XAI should use our experiments as a template to design and evaluate technical solutions that actually improve human understanding.