Date of Award

Fall 2022

Document Type

Dissertation

Degree Name

Doctor of Philosophy (PhD)

Department

Psychology

First Advisor

Turk-Browne, Nicholas

Abstract

Every day is a new day, but much of our experience is richly structured and predictable. From the scenery of our walks to the stereotyped nature of conversation, the vast majority of our experiences are ones that we've encountered before. Human memory is thus shaped by the predictable, structured nature of experiences. Our memories contain not only details about unique experiences (episodic memory), such as your tenth birthday, but also commonalities extracted across experience (statistical learning), such as what tends to happen at a birthday party. Notably, these two kinds of memory require opposing computations: episodic memories are formed rapidly using separated representations to minimize interference, whereas statistical learning occurs gradually using overlapping representations to identify commonalities. Despite these opposing computational requirements, a single brain region, the hippocampus has been implicated for both kinds of learning. This poses a fundamental conflict: How does a single system (i.e., the hippocampus) resolve these two opposing processes at once? The aim of this dissertation is to address this question, asking how episodic memory and statistical learning interact in the brain and behavior. First, I present a novel paradigm which allows for the simultaneous encoding of episodic memories and statistical regularities. Participants were exposed to a sequence of trial-unique scene images, in which certain scene categories reliably followed one another. Thus, participants could form an episodic memory for each individual scene image, but also abstract the high-level regularities of which scene categories predict one another. Behaviorally, I found evidence of competition between these memory systems, as indicated by worse episodic memory for statistically-predictive category images. In a complementary fMRI study, I identified the hippocampus as the seat of this trade-off between statistical prediction and episodic memory: the extent to which the hippocampus exhibited evidence of prediction for the upcoming category was negatively associated with participants’ memories for those predictive categories. Next, I extend this work, exploiting the temporal resolution of intracranial EEG to examine how these interactions unfold over time. Using neural entrainment and multivariate pattern similarity metrics, I found evidence that electrodes in visual cortex exhibited rapid learning of high-level temporal regularities, abstracting over idiosyncratic, episodic detail. I also observed a similar trade-off between prediction and memory as in the prior chapter: trial-by-trial evidence of prediction (in visual electrodes) related to worse memory for predictive items, but better memory for predictable items. After establishing this competitive relationship, I extend this framework by considering some of the factors which may bias the hippocampus towards one encoding strategy or the other. Specifically, I ask how acute stress differentially influences episodic encoding and statistical learning. Based on rodent work, in combination with neural network models, I hypothesized that acute stress may enhance statistical learning, while impairing episodic encoding. Indeed, initial analyses show that acute stress enhances behavioral evidence of statistical learning. Acute stress time-dependently modulated episodic encoding, with enhancements in episodic memory when acute stress occurred ~15 minutes prior to learning. Collectively, this work provides insight into how two fundamental memory processes interact with one another. In combining across sensitive behavioral measures, the anatomical precision of fMRI, and the temporal resolution of EEG, I aim to comprehensively elucidate how we learn from and remember past experience, and how those past experiences are leveraged to shape future experiences and future memories. Specifically, the first two chapters provide evidence that statistical learning and episodic memory compete, insofar as statistically-learned predictions actively interfere with episodic encoding. These findings suggest that statistical learning may fundamentally --- and adaptively --- constrain episodic encoding, such that the mind prioritizes exploiting learned regularities over re-encoding information which is already in memory. The final chapter extends this body of work, aiming to understand how these interactions are modulated by acute stress; in doing so, this work also presents a novel role for stress in enhancing some forms of hippocampal learning. This work also has broader implications for our understanding of how memory is organized in the brain. I discuss other examples of brain regions which house competing memory computations, and present a revision of the Multiple Memory Systems theory, which considers why co-localizing competing memory systems within a single brain region may in fact be adaptive.

Download

Share

COinS