Development of Active Function Learning

• In Event: 2-105 - The Child as an Active Learner: Influences of Social-models, Curiosity, Strategy, and Executive Function

Integrative Statement

Humans constantly have to make sense of functional relationships in everyday life. For example, drivers have to learn how far they can drive with a certain amount of gas in the tank. Previous work has established several characteristics of function learning in adults. For instance, they often expect functional relationships to be linear and have a positive slope, and have difficulty learning non-linear functions. This research has traditionally focused on passive learning, where inputs are randomly chosen or pre-selected by the experimenter before being presented to participants. However, in many real-world scenarios, people actively seek out information, manipulating input features to observe how they impact the outcome. How do people actively select inputs to learn how they relate to a continuous output? How do these active learning strategies develop, and is active learning more effective than passive learning for these kinds of tasks?

We developed a novel multiple-feature function learning task to investigate active function learning in adults (n=61), 7-8-year-olds (n=73), and 10-11-year-olds (n=63). Participants were presented with 27 cards, displayed on a tablet, each depicting a monster together with its score for three features (friendly, cheeky, and funny; Figure 1). The goal was to learn the relationship between the monsters' features and the number of magic fruits picked by each monster (criterion). Participants were assigned to one of two learning conditions: active learners could select 22 cards to observe their criterion, whereas passive learners observed the criterion of 22 randomly selected cards. Participants then completed two tests assessing how well they had learned the function and could generalize their knowledge. We predicted that performance would increase with age, and that active learners would perform better than passive learners at all ages, in line with previous literature (e.g., Markant & Gureckis, 2016; Ruggeri et al., 2018).

First, as expected, we found a developmental improvement in performance. A computational analysis of the adults’ learning data revealed that their strategies were best-described by an active learning model based on a linear regression paired with an optimistic (upper-confidence bound) sampling strategy. This suggests that adults adapted well to the underlying linear function, and sought to learn both about the underlying function and how to obtain high criterion values, a strategy that also characterizes information search in exploration-exploitation dilemmas (Wu et al., 2017). Preliminary computational analyses of children’s data suggest that younger children’s active learning strategies may be less systematic than older children’s and adults’. Older children may also be best-fit by a linear regression, but might employ different sampling strategies than adults. Finally, our results revealed no difference in performance between the active and passive conditions for any age group. This may be explained by a ceiling effect due to the high number of training observations (22 out of 27 cards were observed). It may also be that the function itself was too easy to learn, which may have masked any potential benefits of active learning. We are currently exploring these interpretations with follow-up studies.

Authors

• Angela Jones, Max Planck Institute for Human Development
  Presenting Author

• Eric Schulz, Harvard University
  Non-Presenting Author

• Björn Meder, Max-Planck-Institute for Human Development Berlin
  Non-Presenting Author

• Azzurra Ruggeri, Max-Planck-Institute for Human Development Berlin; Technical University Munich (School of Education)
  Non-Presenting Author