1 February 2017: How do retrieval dynamics drive learning? Insights from fMRI and computational models

Ken Norman
Princeton Neuroscience Institute
Princeton University

About 10 years ago, our lab built a neural network model of how competitive neural dynamics drive learning (Norman, Newman, & Perotte, 2005; Norman et al., 2006; Norman, Newman, & Detre, 2007). This model predicts a U-shaped relationship between neural activation and learning, whereby strong activation causes strengthening of synaptic connections, moderate activation causes weakening of synaptic connections, and lower levels of activation result in no change to synaptic connections. To test this prediction, we have run several studies where we use pattern classifiers (applied to fMRI and EEG data) to track the activation of the competing memories, and then we relate these competitive dynamics to subsequent memory performance. If — according to the classifier — a memory activates to a moderate degree, our theory predicts that it will be weakened and, through this, it will subsequently become harder to retrieve. I will present evidence in support of this prediction from studies using several different paradigms. I will also present more recent data showing how competitive dynamics, coupled with interleaved learning, can result in differentiation of competing memories. I will conclude by discussing possible applications to clinical phenomena.