Introducing the OMNI Project: Optimizing Memory using Neural Information

The lab is excited to announce the OMNI (Optimizing Memory using Neural Information) project, a new endeavor aimed at leveraging both behavioral and neural data to develop a system that improves how people learn and retain new information. OMNI is supported by a National Science Foundation grant to Todd Gureckis (PI) and Lila Davachi (co-PI).

Why improve memory?

Our ability to learn new information is essential for school, work, and life. Given the immense amount of information we need to memorize (as well as the limited amount of time we have for study or practice) we often have to make choices about how and when to study.

Most of the time, decisions about what to study and how to study it are left to the learner’s discretion, but research has shown that people may not be able to accurately predict what they know or what they will remember (Benjamin, Bjork, & Schwartz, 1998; Kornell & Metcalfe, 2006). It’s therefore worthwhile to explore methods of optimizing human learning, such that a learner acquires the maximum amount of new information in the shortest amount of time.

The idea of optimizing human memory is not a new one: over the past several decades, a number of projects have developed intelligent tutoring systems, or “cognitive tutors” that adapt the way material is presented to learners, often using insights from cognitive science and psychology (Atkinson, 1972; The best of these tutoring systems have been based on computational models that combine knowledge of memory dynamics (e.g. that memories decay over time; some things are easier to learn than others) with an individual’s behavior (e.g. responses on a memory test) to determine how study sessions should be optimally structured.

The OMNI approach:

We hope to broaden and expand these existing model-based tutoring systems by drawing on recent developments in cognitive neuroscience, In particular, we aim to make use of subsequent memory signals: neural activity produced at the time of study that differs according to whether material was remembered or forgotten later on (Davachi, Mitchell, & Wagner, 2003; Sanquist et al., 1980). These signals reveal something about how well a studied item has been learned, and thus have evident value for the development of a tutoring system.

Our project involves using fMRI (and later other signals like EEG) to identify subsequent memory signals and then incorporating them into a comprehensive computational cognitive model. This model will make use of new advances in machine learning and hierarchical Bayesian modeling to infer the latent “state” of each studied item (i.e., learned or unlearned) at each point in time. We will be able to more accurately characterize an individual’s learning progress, giving us a better picture of what they know and what they will remember. More specifically, it will take the form of a Hidden Markov Model — a type of statistical model that is particularly well suited to joining cognitive models with noisy yet informative neural data (Rabiner, 1989). By integrating both neural and behavioral data into the same modeling framework, we will be able to more accurately characterize an individual’s learning progress, giving us a better picture of what they know and what they will remember.

In the final phase of the OMNI project, this model will be used to construct a new tutoring system, which will customize sequences of study items for individual learners by removing items that are already known and re-presenting those that need additional study—ultimately leading to more effective learning overall.

Keep track of our progress

For updates on the OMNI project, as well as links to our released data, publications, modeling code, and more, head to the project website! This website will collect all the resources for the project in a easy to access place.

Tags: , , , , , ,