Bayesian Teaching of Image Categories

AbstractA large portion of human learning occurs in pedagogical contexts where there is a teacher and a learner. This framework of teaching and learning has been formalized in recent work as Cooperative inference, where both the teacher and learner engage in a process of recursive reasoning (Shafto, Goodman, & Griffiths, 2014; Yang et al., 2018). A special of this is Bayesian teaching, which focuses on selecting helpful examples to teach a learner some target model. While these approaches have been successful in capturing pedagogical behavior in a variety of contexts, they have been limited by their tractability to relatively simple domains. One of the open questions regarding Bayesian teaching is whether it can scale to teach from naturalistic domains with more interesting datasets. In this work, we show how to apply Bayesian teaching to teach human participants categories learned by a machine learning model from a dataset of images with category labels. How well the categories are taught is measured by how well the participants can predict the behavior of the target machine learning model. Our results demonstrate that manipulating the goodness of teaching examples influences the how well participants responses match the predictions of the target model.


Return to previous page