Biases for learning from teaching

Nicholas SearcyUniversity of Louisville
Patrick ShaftoUniversity of Louisville

Abstract

According to previous accounts, teaching is the helpful sampling of examples according to a learner's known biases. Using the domain of Boolean concepts, we show that biases are necessary, there is no single rational bias, and teaching is not possible when the teacher does not know the learner's bias. Taken together, these results suggest that teaching via sampling would be either ineffective or impossible for Boolean concepts. We offer an alternative account of teaching based on cooperation and the teacher's omission of irrelevant features. The result is a model of teaching that is computationally efficient, effective in concept spaces with infinitely many features, and suggestive of a natural concept representation based on cooperation.

Files

Biases for learning from teaching (151 KB)



Back to Table of Contents