Recognizing the fact that “most of modern machine learning can be viewed as consistent with Michie’s weak criterion,” the authors of this paper are motivated to work on Michie’s ultra-strong criterion, which “requir[es] the learner to teach the hypothesis to a human” and evaluate their performance.

The paper defines an “operational definition of the comprehensibility of a logic program,” including inspection time, predicate recognition, and so on, besides the public and private predicate symbols. Further, these definitions are extended for the ultra-strong machine learning study; experimental hypotheses are given.

For the comprehensibility and predicate study, the authors are interested in the “interactions between predicate use and complexity of rules,” for example, whether the use of additional predicates can be helpful or whether results depend on the ability of a person to assign a specific meaning to the predicate. Their results show that both the complexity of the program and anonymous predicate symbols affect the degree of comprehensibility. With the goal of showing how “[inductive logic programming, ILP] learned classification rules can result in operational effectiveness for humans,” the second experiment shows that participants were not able to “learn the relational concept on their own ... but they were able to apply the relational definition provided by the ILP system correctly.”

The paper concludes with an interesting discussion on “human learning, teaching, and verbal interaction.”