Optimizing Learning using Machine Teaching

• In Event: 2-159 - Applications of Cognitive Science Research: Contributions to Education Across Levels of Analysis

Integrative Statement

Developmental research offers insight into how children learn at different stages of development. One challenge for the field is how to take results from learning research to design more effective instruction and teaching. In this talk we describe an approach that combines formal models of learning with the concept of “machine optimized teaching” (Zhu et al., 2018). In machine teaching, a computer algorithm determines the best order and timing of information presentation to improve an individual’s learning. Such algorithms are tailored to reflect the general principles of human learning that have been identified in developmental studies as well as individual differences in the path that particular students take on the way to mastery.
This paper introduces machine teaching algorithms in an accessible manner, discussing their potential advantages in a developmental research setting as well as applications to educational technology. In particular, we focus on a framework from computer science called Partially Observable Markov Decision Processes (POMDPs) (Kaelbling et al., 1998). In a POMDP, an agent makes observations to form a belief about the current state of the world and acts to optimize some objective (such as earning a reward). Interestingly, we can model the problem of teaching as a POMDP: A teacher, uncertain about the latent knowledge state of her students, infers student knowledge from noisy behavior to choose the best pedagogical action that optimizes performance along some measure of mastery (Rafferty et al., 2016).
A teaching POMDP consists of 1) a learner model and 2) a cost function. For instance, a model of the student as a Bayesian learner would describe how a student’s knowledge state changes after various learning episodes occur (e.g., studying flashcards versus taking a test) and how that knowledge is reflected in the student’s performance on tests (Tenenbaum & Griffiths, 2001). The cost function shapes the motivation behind action selection; for example, the teacher might attempt to maximize the student’s exam scores or minimize class time needed to teach a concept. Together these components allow one to define the optimal sequence of teaching actions to minimize the cost function.
We will also present one application of machine teaching that we have been developing in our lab that explores the applications of neural data to assessment and instruction. In the project, adults attempt to memorize foreign language vocabulary while undergoing fMRI, and recall is assessed at a later date. Results from behavioral (N=150) and fMRI (N=20) subjects demonstrate that a model of memory incorporating both behavioral and neural observations best predicts whether an individual word-pair will be remembered or forgotten. We show how the learner model from this work can be used to devise POMDP-optimized teaching sequences that aim to improve a learner’s long-term retention of the target materials. To help conference attendees integrate these methods in their own research, code will be shared online.

Authors

• Pamela J Osborn Popp, NYU Center for Neural Science
  Presenting Author

• Todd Gureckis, New York University
  Non-Presenting Author