Computational, Neural, and Behavioral Studies of Competition-Dependent Learning

Princeton University, Princeton NJ

Investigators

Linked publications, trials & patents

Abstract

The overarching goal of this renewal application is to understand how stored memories change as a function of experience, and how these changes support our ability to discriminate similar stimuli. The inability to make such distinctions can lead to mundane errors and memory interference, but also devastating cognitive, semantic, communication, and developmental deficits. We focus on the challenging scenario when the differences between stimuli are subtle and the stimuli appear at different times, requiring memory. The specific goal of this proposal is to understand the interacting roles of the hippocampus and cortex in this scenario. Recent studies have found that the key to behavioral discrimination is differentiation of hippocampal representations — a phenomenon whereby the representations of similar stimuli are pushed apart beyond the orthogonality of pattern separation, to the point of anti-correlation. Crucially, existing computational models of the hippocampus do not exhibit differ- entiation and thus cannot provide a satisfactory account of memory-based discrimination learning. The major significance of this proposal is to remedy this critical gap, by formulating and testing a new neurocomputational theory that explains the causes of hippocampal differentiation and its consequences for discrimination behavior. Aim 1 breaks new theoretical ground by incorporating our theory of how unsupervised learning gives rise to differentiation (developed in the prior funding period) into a biologically grounded neural network model of hip- pocampus and its interactions with cortex, which will be used to explain existing results and to generate novel neural and behavioral predictions about discrimination learning. Aim 2 tests these predictions with fMRI. A key prediction is that the hippocampus has a time-limited role in discrimination learning, supporting activation of unique features of similar stimuli in visual cortex after (but not before) differentiation, then relinquishing this role after consolidation. This prediction and others are tested using new fMRI decoding methods based on deep learning to quantify how strongly visual cortex represents unique features of paired stimuli, and how this relates to hippocampal activity. Aim 3 combines innovative experimental approaches with greater causal leverage to further test model predictions: real-time closed-loop fMRI to manipulate cortical inputs to the hippocampus; in- tracranial EEG (iEEG) to measure the temporal precedence of hippocampal signals driving representation of unique features in visual cortex; and responsive neurostimulation (RNS) to transiently disrupt hippocampal pro- cessing in patients with chronic implants. In summary: This proposal tightly couples theoretical neural network modeling with experimental tests of model predictions using advanced correlational (fine-grained fMRI decoding, consolidation-related EEG signals) and causal (real-time fMRI, iEEG, RNS) human neuroscience methods. This work provides pivotal basic science insights into how the hippocampus supports learning. These insights may facilitate the development of evidence-based learning paradigms that improve discrimination behavior, with po- tential application to new treatments for psychiatric and neurological disorders.