Researchers have debated whether instructional-based teach-ing or exploration-based active learning is better for decadeswith unsatisfying results. A main obstacle is the difficulty inprecisely controlling and characterizing the pedagogical meth-ods used and the learning conditions in empirical studies. Toaddress this, we leveraged existing computational models ofteaching and active learning to formalize the methods and thelearning process. We compared the two pedagogical methodsin a concept-learning framework and investigated their effec-tiveness under various scenarios. Our results show that whenthe learner and teacher are conceptually aligned, teaching isat least as effective as, and often much more effective than ac-tive learning, but when the alignment is broken, active learningcan yield moderate improvement over teaching. We concludeby discussing our results’ implications for the debate and theprospects of bringing computational models to bear on com-plex real-world problems that are resistant to simple experi-mental investigation.

Hundreds of millions of people learn a second language (L2).1When learning a specific L2, there are common errors for na-tive speakers of a given L1 language, suggesting specific ef-fects of L1 on L2 learning. Nevertheless, language instruc-tion materials are designed based only on L2. We developa computational model that mimics the behavior of a non-native speaker of a specific language to provide a deeper un-derstanding of the problem of learning a second language. Weuse a Naive Bayes to model prepositional choices in English(L2) by native Mandarin (L1) speakers. Our results show thatboth correct and incorrect responses can be explained by thelearner’s L1 information. Moreover, our model predicts incor-rect choices with no explicit training data of non-native mis-takes. Our results thus provide a new medium to analyze anddevelop tools for L2 teaching.

Theory-of-Mind inference is natural for humans but poses significant computational challenges. The core difficulty can be traced back to the exponential growth of paths to consider in planning given a mental state. In this paper we tackle this problem in a search-and-rescue task implemented in Minecraft. Our goal is to infer differences in knowledge from participants' continuous-time trajectory. By abstracting the spatio-temporal state space and the reward function together, we surface natural decision points, on which we compare the participants' behavior to myopic rational agents of varying knowledgeability. Collectively, the abstraction and rational agent analysis yield successful inference of participants' knowledge states and reveal distinct patterns of their exploratory behavior.

Humans learn from other knowledgeable informants whochoose data to foster learning. Mathematical models of teach-ing and learning have formalized this process of learning fromhelpful others. While these approaches have been successful incapturing teaching and learning in a variety of contexts, theyhave been limited to relatively simple domains. One of theopen questions regarding Bayesian teaching is whether it canscale to teach from naturalistic domains with more interestingdatasets. In this work, we show how to apply Bayesian teach-ing to teach human participants categories learned by a super-vised machine learning model. The effectiveness of teaching ismeasured by how well the participants can predict the behaviorof the target machine learning model. Our results demonstratethat Bayesian teaching can be applied to naturalistic domains,show that the best sets of examples according to the modelyield better learning, and suggest avenues for improving ourability to automate teaching of image categories.

The enormous scale of the available information and productson the Internet has necessitated the development of algorithmsthat intermediate between options and human users. These al-gorithms do not select information at random, but attempt toprovide the user with relevant information. In doing so, thealgorithms may incur potential negative consequences relatedto, for example, “filter bubbles.” Building from existing al-gorithms, we introduce a parametrized model that unifies andinterpolates between recommending relevant information andactive learning. In a concept learning paradigm, we illustratethe trade-offs of optimizing prediction and recommendation,show that there is a broad parameter region of stable perfor-mance that optimizes for both, identify a specific regime thatis most robust to human variability, and identify the cause ofthis optimized performance. We conclude by discussing im-plications for the cognitive science of concept learning and thepractice of machine learning in the real world.