What learning goals do instructors of machine learning courses for non-majors find difficult to teach? The answer is goals corresponding to higher levels in a generally used taxonomy of educational goals. Such higher-level goals include relating real life to computer input/output, evaluating performance according to the bias-variance tradeoff, and developing a model from scratch. In contrast, course aspects corresponding to lower-level goals, such as the functioning of a linear regression algorithm, are easier to teach.

The authors recruited ten faculty members from a variety of institutions and conducted structured telephone interviews. The transcripts were then coded and sorted. It became apparent from the sorting that the clusters related to learning goals aligned with several learning taxonomies. The most useful fit was with the structure of observed learning (SOLO) taxonomy. SOLO is based on four stages of learning: unistructural (example: describing an algorithm), multistructural (example: interpreting an algorithm), relational (example: evaluating performance), and extended abstract (example: communicating performance). The coded data was then rearranged and the rearranged clusters matched the SOLO taxonomy.

Incidentally, the conclusion that parts of a course that correspond to higher stages are harder to learn is validated by the results of a student survey reported in the literature. Sulmont et al.’s paper should be useful to instructors planning a course, especially one for non-majors, although the conclusion seems to be more general. It contains much information and is clear.