Asking creative questions is a hallmark of human cognition. In comparison, machine learning systems that attemptto mimic this ability are still extremely limited (e.g., current chatbots ask questions based on preprogrammed routines). In thepresent work, we developed a computational model of question generation. Based on a corpus of questions collected from onlineparticipants playing an information-seeking game, we designed a “grammar of questions.” The grammar is powerful enough torepresent all human questions we collected and thus defines the “question space.” Given a particular context (game scenario),people are more likely to ask (generate) some questions that others. Our computational model predicts these likelihoods, thatis, a probability distribution over the question space. In addition, the model can generalize to novel contexts. Key modelingredients are informativity, compositionality, and length of a question.

We present an information-theoretic approach to modeling childrens performance in a deductive reasoning game. Ourapproach takes cognitive limitations into account to model the interpretability of feedback that children receive during thegame. We use data of thousands of children, 5 to 12 years of age, from a popular online educational learning system. In theDeductive Mastermind game the player seeks to identify a hidden code that consists of a sequence of colors. The playersees a series of proposed codes together with corresponding feedback providing partial information about the similarity ofeach proposal and the hidden code. In Deductive Mastermind games, the proposals are set up such that deductive reasoningleads to a single possible hidden code. The games vary in code length, the number of possible colors, and the number ofproposals, resulting in game difficulties of various degrees.

Recently, both psychology and machine learning have explored the ability of learners to ask questions. However, much of this work has focused on a single type of question: a “label query”. When making a label query, the learner selects an unfamiliar (unlabeled) item and requests a label for it (e.g., ”What is this?”). We hypothesized that people often prefer much richer types of questions (e.g., feature queries: “Is this feature relevant?”, demonstration queries: “Can I see an example of a ladybug?”, etc.). To study this behavior, we had people play a simple game where they generated natural language questions to determine a hidden configuration of objects. We compute the normative value of these rich questions as measured by modelbased analyses (e.g., information gain). A second experiment evaluates the ability of human observers to judge this value when the demands of question generation are removed.

The study of question asking in humans and machines hasgained attention in recent years. A key aspect of question ask-ing is the ability to select good (informative) questions froma provided set. Machines—in particular neural networks—generally struggle with two important aspects of question ask-ing, namely to learn from the answer to their selected ques-tion and to flexibly adjust their questioning to new goals. Inthe present paper, we show that people are sensitive to both ofthese aspects and describe a unified Bayesian account of ques-tion asking that is capable of similar ingenuity. In the first ex-periment, we predict people’s judgments when adjusting theirquestion-asking towards a particular goal. In the second ex-periment, we predict people’s judgments when deciding whatfollow-up question to ask. An alternative model based on su-perficial features, such as the existence of certain key wordsin the questions, was not able to capture these judgments to areasonable degree.

What are the major topics of the Cognitive Science Societyconference? How have they changed over the years? To an-swer these questions, we applied an unsupervised learning al-gorithm known as dynamic topic modeling (Blei & Lafferty,2006) to the 2000–2017 Proceedings of the Cognitive Sci-ence Society. Unlike traditional topic models, a dynamic topicmodel is sensitive to the temporal context of documents andcan characterize the evolution of each topic across years. Us-ing this model, we identify historical trends in the popularity oftopics over time, and shifts in word use within topics indicativeof changing focuses within the field. We also measure the cor-relation across topics, and use the model to highlight the topicstructure of particular papers and labs. We believe dynamictopic models present an important tool towards understandingCognitive Science as it continues to grow and evolve over time

This paper investigates the emergence and development ofchildren’s ability to adapt their information search to differentgoals. In Study 1, 3- to 7-year-olds had to decide whether tostudy the arms or legs of two monsters to predict which wouldsucceed at a throwing vs. jumping challenge. Children’s abil-ity to adaptively select the relevant piece of information andtailor their search to the given goal increased with age, surpass-ing chance level around 4;6. Study 2 investigated additionaladaptation to distributions of, e.g., long arms in the search do-main. Preliminary results confirm the observed developmentaltrend in search adaptiveness and effectiveness, suggesting anability to tailor information search to the relevant distributionsin the environment. These studies provide first insights intothe development of adaptive information search given complexgoals, deepening our understanding of this key aspect of learn-ing, judgment and decision-making.

The ability to ask questions during learning is a key aspect ofhuman cognition. While recent research has suggested com-mon principles underlying human and machine “active learn-ing,” the existing literature has focused on relatively simpletypes of queries. In this paper, we study how humans constructrich and sophisticated natural language queries to search for in-formation in a large yet computationally tractable hypothesisspace. In Experiment 1, participants were allowed to ask anyquestion they liked in natural language. In Experiment 2, par-ticipants were asked to evaluate questions that they did not gen-erate themselves. While people rarely asked the most informa-tive questions in Experiment 1, they strongly preferred moreinformative questions in Experiment 2, as predicted by an idealBayesian analysis. Our results show that rigorous information-based accounts of human question asking are more widely ap-plicable than previously studied, explaining preferences acrossa diverse set of natural language questions.

How do people learn complex rules? We introduce a novel paradigm called ”Track-A-Mole”, in which participants have tolearn about and predict the moves of a cartoon mole, whose movements are generated by graphical programs. Our resultsshow that participants can learn to predict richly structured programs, and often require only few observations to do so,showing rapid learning and early insights about the underlying patterns. Moreover, we found that how learnable a programis can be predicted by features related to its complexity and compressibility. Finally, participants also show interestingpatterns of generalizations, assuming more parsimonious rules first and then gradually adjusting their predictions to morecomplex regularities, as well as matching their predictions to the general direction of movements and producing sensi-ble errors. These results extend our understanding of complex rule learning and open up future opportunities to modelsequential pattern predictions as graphical program induction.