CEU Electronic Theses and Dissertations, 2022
| Author | Stanciu, Oana |
|---|---|
| Title | The Feasibility of Human Teaching by Sampling and the Challenges of Learning from Teachers |
| Summary | From a young age, humans efficiently utilize sparse observable samples to make reliable inferences about a highly complex, stochastic and ever-changing world. Undoubtedly, the social grounding of human inductive learning, especially through teaching by more knowledgeable and helpful others, contributes greatly to the success of human learners. However, while there is tremendous interest in the development of intelligent tutoring systems for educational applications, and significant theoretical and applied advancements have been made in the burgeoning field of machine teaching, experimental work in cognitive science has focused almost exclusively on investigating the behaviors of learners, largely overlooking teachers. This thesis investigated one of the main ways in which humans teach, both in formal and informal settings: by offering examples. Since explicitly generating samples for the purpose of teaching others is undeniably a normatively hard problem, we explored possible limitations faced by teachers due to the abstraction and complexity of the task (Chapter 2) and having a good model of the learner (Chapter 3). From the learner's perspective, we examined whether learners could effectively learn from sampled data (Chapter 2) and assess imperfect teachers (Chapter 4). In Chapter 2, replications and extensions of two teaching games (prototype and category boundary teaching) lead to diverging results in terms of the optimality of teachers, which we attributed to differences in the level of abstraction of the task and the complexity of the stimuli. In turn, learners were also limited in their ability to effectively adapt to the data generative process, specifically when producing estimates based on autocorrelated samples, a known feature of samples produced by humans. In Chapter 3, we proposed that teachers can overcome one of the more challenging aspects of teaching - building an adequate model of how learners make inferences - by engaging in the experience of learning, and especially active learning (given the computational similarity). We present evidence that prior active learning facilitates teaching from two different category teaching experiments. Lastly, in Chapter 4, we proposed that confidence may be a useful pedagogical signal. For instance, explicit confidence statements could allow teachers to monitor the progress of their students, and help learners to choose better teachers. Focusing on the learners' perspective, a first experiment found, in agreement with previous work, that humans preferred more informative and better calibrated advisers as future collaborators (even when compared with overconfident advisers). Interestingly, a second experiment showed a dissociation between partner preferences and decision making. Specifically, learners did not optimally use information about the relative metacognitive skills of informants to improve their decision making. In light of the results presented, we suggest that while rational pedagogical models can be a useful computational-level description of teaching solutions in some limited domains, they are unlikely to provide a close account of flexible, on the fly teaching behaviors without significant modifications that account for the important limitations facing teachers. |
| Supervisor | Fiser, Jozsef |
| Department | Cognitive Science PhD |
| Full text | https://www.etd.ceu.edu/2022/stanciu_oana.pdf |
Visit the CEU Library.
© 2007-2021, Central European University