Of half a mind Thoughts on learning and uncertainty

Psychonomics 2016 - Adaptive exploration in decisions from experience

Imagine that you’re on Amazon trying to decide which of two products to purchase. One way to learn about the options before you buy is to look at reviews from other people who have bought the same items. Each review gives you a glimpse into the relative value of each option, and you can explore each item (i.e., keep on reading reviews) for as long as you like until you feel ready to hit the buy button.

So, how many reviews do you decide to read for each product? How does the variability in the ratings affect how long you explore—for example, is there a difference between seeing a string of solid 4-star reviews as opposed to a mixture of 5-star and 3-star reviews? How does your exploration change when searching for a relatively mundane product like a power adapter as opposed to a major purchase like a computer?1

At Psychonomics this week I’m presenting some new results from a project where we try to understand how people adapt their exploration in response to these kinds of environmental factors, including the variability in the outcomes they experience and the rewards that are at stake. The experiment described in the poster below was designed to test the predictions of a sequential sampling model that accounts for this adaptive exploration and its effects on how people make choices. Sequential sampling models are widely used to model the relationship between choices and response times in a number of decision making tasks2, and our results suggest that similar mechanisms can account for the way that people sample experiences from the environment prior to making a choice.

Click on the image below to get the PDF of the poster:


  1. I’m using product reviews as an illustrative example of experience-based decision making in the wild (our experiment was more stripped down), but some recent research has looked exactly at how sampling review ratings to learn about products impacts choice (see here).

  2. See this recent blog post for a nice introduction to sequential sampling models as applied to choice RT.

CogSci 2016 - Hypothesis generation and self-directed learning

At CogSci this year I presented some new work on hypothesis generation and its role in self-directed learning (specifically, the learning of categorical rules). This project explores the idea that hypothesis generation is a bottleneck in active learning, as is suggested by a lot of work in classic cognitive tasks as well as studies from educational psychology. The results give an example of how biases in the hypothesis generation process can have a big impact on whether people can efficiently learn different kinds of rules.

For more on this project, see the paper in the proceedings and the poster below (PDF).


Enhanced memory as a common effect of active learning (new paper)

I have a new paper out in Mind, Brain, and Education, coauthored with Azzu Ruggeri, Todd Gureckis, and Fei Xu. In the paper we argue that one reason active learning improves outcomes relative to passive conditions (like lectures) is that exercising control leads to enhanced memory (in particular, episodic memory for the events experienced during learning).

In many lab-based tasks, having control over learning (e.g., controlling the selection or pacing of new information) often leads to better memory compared to yoked conditions in which such control is absent but the same information is experienced. By surveying a range of research that has employed this kind of active/yoked design in different domains, we outline how improved memory can arise from a number of different mechanisms depending on the kind of control afforded to the active learner. Thinking about real-world learning activities in terms of this basic dimension of control (which is just one part of what people typically think of as “active learning”) may help us understand when and why active learning tends to produce better outcomes than traditional instruction.

A couple of things struck me while researching the paper:

  • Some form of episodic memory is important to many theories of conceptual learning or semantic memory – but episodic memory is not really a focus in educational research. Perhaps unsurprisingly, educational psychologists seem to be primarily interested in measures of conceptual learning: does the student remember the definition of a term, or how to solve a certain kind of physics problem? As far as I can tell there isn’t much work examining memory for the details of the learning experience (e.g., the activities completed, example problems that were encountered, conversations during class, etc.) and how it relates to other performance measures.

  • There’s a similar gap in lab-based research related to inquiry-based learning (including work on category and causal learning which usually shows an advantage for active control over yoked observation). This work has largely focused on the value of the information that active learners select (e.g., do they ask questions that are useful? Can they come up with experiments that lead to unconfounded evidence?). But it seems likely that the same set of memory-based mechanisms could play a role, particularly in real-world inquiry learning settings that involve richer materials and forms of interaction than the typical lab-based experiment.

Writing the paper helped me pinpoint some of these big open questions about how memory and control interact during learning. Of course, despite our best efforts it’s inevitable that we missed some existing work that would be useful. If you know of something relevant or have any other comments about the paper, let me know!

Active learning workshop at CogSci

Next month I’ll be participating in a full-day workshop at CogSci, entitled “Active learning: Cognitive development, education, and computational models.” The workshop will feature lots of terrific researchers from the CogSci community who have been thinking about active learning from many angles. Come by if you are in Philly on Wednesday!