OMELET (Optimizing Human Learning with Machine Teaching)


A long standing but elusive goal in machine-aided education has been to exploit cognitive models of human learning to select teaching or practice experiences for students that will efficiently lead them toward the desired knowledge state. We show how contemporary optimization methods allow theorists to discover, for any implemented learning model and desired outcome, an optimal teaching set that produces the desired outcome most efficiently given the learning model. Finally, we aim to test whether the optimal teaching set can speed human learning.


Our goal is to develop and test optimal teaching or practice problems for any learning task, taking arithmetic as an example domain, with state of the art search techniques in machine teaching and computational cognitive models of human learning.

The animation shows an example of a hill climbing algorithum to search for our optimal training set.
In the left panel, each dot is the accuracy of the model. When the accuracy increases the line advances to the right, showing the best-so-far training set. 
The right panel shows which training sets are being compared. Each dot is one of 136 possible problems.
Human performance is 39%. Just by choosing the right training set, we can boost performance to 55%


Potential Implications:

The potential implication is to boost learning in important educational domains by designing and offering good lessens or exercises to learners with machine teaching and cognitive modeling techniques.

Transferable Skills:

Rui Meng, Educational Psychology Graduate Student has developed skills such as:

  • Cognitive Modeling, such as building computational cognitive models and searching for parameters of computational cognitive models
  • Scientific Communication, such as oral presentation and scientific writing
  • Cross-Discipline Collaboration, understanding knowledge from other disciplines and explaining knowledge from my own discipline

Ayon Sen, Computer Sciences Graduate Student has developed skills such as:

  • Cognitive modeling like building recurrent neural network, searching for proper hyperparameters of said model
  • Human-Computer Interaction e.g., how problems should be presented to humans through a computer
  • Scientific Communication, such as oral presentation and scientific writing.

Industry Collaboration

We are currently working with Carnegie Learning, an education technology company focused on math curriculum and software, to help develop optimal visual representations and practice problems for fractions learning. We will help build a computational cognitive model for fraction learning and then search for optimal visual representations and practice problems with state of art searching techniques in machine teaching. Finally, the optimal training set will be tested with real learners to examine the learning enhancement effect.

Our project could also have implications in other disciplines apart from mathematics. The approach we use can be applied to any learning task for which there is a computational cognitive model, such as chemistry and so on.


Thank you to our OMELET project team:

Martha Alibali, Psychology Professor  |  Sarah Brown, Psychology PhD Candidate  |  Chuck Kalish, Educational Psychology Professor   |   Percival Matthews, Educational Sciences Professor   |   Rui Meng, Educational Psychology & LUCID Graduate Student   |    Tim Rogers, Psychology Professor  |   Ayon Sen, Computer Sciences & LUCID Graduate Student   |   Jerry Zhu, Computer Sciences Professor