Predictive approaches for investigating brain activity underlying successful learning

• Author / Creator

  Chakravarty, Sucheta

• Successful learning is of vital importance to human cognition. Accordingly, researchers have been interested to understand brain-activity signals that support it. However, traditional analysis of brain activity is based on planned comparisons and descriptive methods, which can both overestimate brain activity by overfitting it, and also underestimate the behavioural relevance of brain-activity measures by ignoring the subtle multivariate patterns. Complementary to the traditional analysis, here I used predictive approaches that can offer a stronger framework for finding behaviourally-relevant brain activity by testing predictions for unseen/future observations. Two learning situations were considered; one, where participants studied lists of words followed by old/new recognition tests for target (studied) and lure (new) items, the other involved trial-and-error learning of the stimulus-response rules for a large set of words, driven by reward feedback. For both learning paradigms, I asked if brain activity present when participants studied the material explained subsequent variability in the learning outcomes. First, I tested if features of brain-activity signals present during the study phase, as identified by previous planned-comparisons based investigations, could withstand tests of predictions for learning outcomes at the level of individual trials. For both tasks, this produced a small but significant success across a large number of participants. Next, I asked if data-driven multivariate pattern analysis of the study-phase activity produced better predictions for the learning outcomes. The multivariate pattern analysis achieved a small significant success for the item-recognition task, but it was under-powered for the trial-and-error learning task and produced non-significant results. Taken together, for both tasks, the contribution of the study-phase activity to later variability in learning outcomes was overall small, indicating that other important predictors may be missing. To test this suggestion, I further investigated brain activity during the test phase of the item recognition task. Following a similar approach as study, first I used features of previously-identified brain-activity signals, to predict the memory outcomes, which achieved modest success. Then, I conducted multivariate pattern analysis of test-phase activity, which was still modest but predicted significantly better than the individual signals at study or test phases, as well as the multivariate activity at study. Further, combining brain-activity features from both study and test phases led to similar size of predictions as that for the test-phase only. Thus, test-phase activity predicted memory outcomes more directly. Also, study-phase activity did not contribute to memory-variance that was not shared by test-phase activity. The multivariate pattern analysis also offered additional important insights. Across investigations, performance of the multivariate classifiers was positively correlated with participants' performance, and was meaningfully large for better-performing participants. This could suggest that brain activity for better-performing participants has a greater task-relevance, which is picked up by the classifiers, leading to better predictions. The multivariate pattern analysis of test-phase activity for item recognition also showed that depending on the time it takes to reach a decision, memory judgments could be driven by either a unitary, integrated signal or two independent sources of evidence; suggesting a way to reconcile the existing debate on single- versus dual-process theory. Overall, these investigations showed that predictive approaches can be used to characterize as well as to quantify the contributions of different brain-activity signals in explaining the variability in learning outcomes. While the classifier approach could be built upon to improve the classification rates, the overall modest predictions could also suggest that successful learning depends on other factors that are not reflected in the brain activity during the study- or test phases of the item recognition task, or during the feedback processing of the trial-and-error learning task. Accordingly, future investigations will need to identify and include these factors into the predictive analysis incrementally, in order to reach a more comprehensive explanation of learning behaviour.

• Subjects / Keywords

  • memory
  • recognition
  • machine learning
  • reinforcement learning
  • ERP

• Graduation date

  Fall 2021

• Type of Item

  Thesis

• Degree

  Doctor of Philosophy

• DOI

  https://doi.org/10.7939/r3-0tk9-t970

• License

  This thesis is made available by the University of Alberta Libraries with permission of the copyright owner solely for non-commercial purposes. This thesis, or any portion thereof, may not otherwise be copied or reproduced without the written consent of the copyright owner, except to the extent permitted by Canadian copyright law.