INFERRING SALIENT MOTIFS DURING HUMAN AND ANIMAL LEARNING
Discovery of learning motifs in reward-based learning experiments
Identification of learning motifs that separate between test and control subjects
Understanding the biological mechanisms in the brain activity during learning necessities deriving the behavioural motifs and juxtaposing them with the brain activity. In this project, we analyse the behavioural data of subjects (animals/humans) during reward-based learning experiments and derive their learning motifs. We define a learning motif as a behavioural policy that the subject exhibits during parts of the learning process. We develop a Granger causality-based method to derive these learning motifs. And since learning is an evolving process, we use a hidden Markov model to capture the changes in the learning motifs followed by the subject over time. We also build on the hidden Markov model to identify salient learning motifs followed by a group of subjects. The evaluation of our proposed methods is not a trivial task due to the absence of the ground truth of the learning motifs. Therefore, we also propose a validation scheme that does not require ground truth to validate the derived learning motifs.
What we want to achieve
Our Project Goals
Inferring the learning motifs followed by a subject in reward-based learning experiments
During reward-based learning experiments, the responses to a current stimulus are affected by prior responses to the stimulus. We develop methods that build on Granger causality to capture the dependencies between the past and the future observations/responses.
Identifying stimulus features and actions related to the learning task.
In reward-based learning, the subject is exposed to stimuli that are defined by a set of features. We develop methods that identify features selected by the subject as relevant to the learning task and actions that are also triggered as a response to the stimuli. We model the learning motif as a set of Granger causality relationships between the stimuli features and the observable actions.
Capturing changes in the derived learning motifs over time.
We fit a hidden Markov model on the sequence of learning motifs represented as a set of Granger causality relationships. We use the Viterbi algorithm to map the sequence of the learning motifs that are derived over time for each subject to a sequence of latent states using the hidden Markov model. These latent states reflect the change in the derived learning motifs.
Inferring salient learning motifs that are followed by a group of subjects.
We jointly model the sequences of learning motifs derived from a group of subjects using one hidden Markov model. The latent learning states by the hidden Markov model reflect the average learning motifs followed by this group of subjects.
Developing a validation scheme for the derived learning motifs.
It is difficult to acquire the ground truth of the learning motifs because animals cannot articulate these learning motifs and this also applies to humans in complex learning tasks. We propose a validation scheme that does not require ground truth. In every learning task, we have groups of subjects that are known to apply different learning motifs. We build a classification model that uses the derived learning motifs as predictive features to separate these different groups and compare the performance to the one achieved by another baseline model that uses the raw behavioral data alone as features. If the classification model that uses the derived learning motifs outperforms the baseline model, then we accept the derived learning motifs to be valid.
Project Team
Noor Jamaludeen
Felix Kuhn
Prof. Dr. Stefan Remy
Prof. Dr. Myra Spiliopoulou
Publications
Inferring Salient Motifs during Learning Experiments
2023 IEEE 36th International Symposium on Computer-Based Medical Systems (CBMS)Support our research
Become a study participant!
How do our brains, our bodies and our environment interact? How do physical illnesses affect our mental performance? And why are we more efficient on some days than others?
We would like to get to the bottom of these questions together with you. Register now and take part in exciting studies.
