Mark Ho, Postdoc, UC Berkeley
Title: Communicative Intent and Interactive Teaching
Abstract: People often attempt to communicate about the world or teach others how to act in the world. Understanding the mechanisms underlying this capacity requires characterizing complex temporal dependencies between a conveyer of information, a receiver of information, and their shared environment. For example, when a person is actively teaching, their behavior will be based on the structure of the environment, the history of interaction, and even their model of the learner. In this talk, I will present several related computational models and experiments that examine the mechanisms of interactive teaching. In particular, I focus on two types of non-verbal teaching interactions that have been studied in the psychological literature. The first is teaching by evaluative feedback, in which a person uses rewards and punishments to influence another’s behavior. Much previous work assumes social rewards operate like non-social rewards (e.g. according to principles of reward-maximization). However, I will discuss computational analyses and empirical evidence indicating that people do not use rewards in this manner. Rather, when using rewards and punishments, people assume that learners can interpret their feedback communicatively. Second, I will discuss work on communicative demonstrations, in which a demonstrator leverages an observer’s capacity for theory of mind in order to convey their beliefs and goals while interacting with the environment. Our experimental results reveal that people will modify their behaviors in systematic ways that are suboptimal for realizing non-communicative goals but reflect planning in the belief space of an observer. A common thread across these two lines of research is that the teacher’s communicative intent and learner’s recognition of that intent is crucial for understanding teaching interactions. This indicates that evaluative feedback and demonstration are closely related to other forms of communication and teaching such as teaching by example or by verbal instruction. Moreover, it has implications for how learners (both natural and artificial) can best benefit from interactive teaching.