Machine Explanations and Human Understanding
Explanations are hypothesized to improve human understanding of machine learning models. However, the fundamental questions of what it means to understand a model for humans and how explanations impact this understanding remain unclear.
Empirical studies have found mixed and even negative results. add example here.
In our latest work, we formalize the concept of human understanding in the context of human-AI cooperation and provide a formal framework to analyze how humans improve their understanding with explanations, using causal diagrams.
Measuring human understanding
With a thorough literature review, we conclude three key concepts of human understanding that researchers used to measure in a various of empirical studies:
- Task decision boundary
- Model decision boundary
- Model error
