Understanding Changes in Hippocampal Representations by Measuring Memories with Natural Language Processing

PROJECT SUMMARY
The hippocampus plays an essential role in encoding long-term episodic memories. However, because
many of the experiences we encode share similar features (people, locations, objects), a critical challenge
for the episodic memory system is to prevent interference or confusion between these memories. Recent
human neuroimaging studies have revealed that highly similar events can trigger a “repulsion” of
corresponding representations within the hippocampus such that nearly identical events are associated with
markedly different activity patterns. Critically, there is evidence that hippocampal repulsion is adaptive in
that it is associated with reduced memory interference. However, a fundamental open question is whether
or how hippocampal repulsion impacts the actual contents of memories. Addressing this question requires
methods for precisely characterizing potentially subtle differences in behavioral and neural expressions of
memory content. In this proposal, I will leverage Natural Language Processing (NLP) algorithms to
transform measures of verbal recall into text embeddings (i.e., numerical vectors) within a multidimensional
semantic space. These text embeddings will allow me to quantify the similarity of memories for highly similar
natural scene images. Additionally, I will gain new training in advanced fMRI methods and computational
analyses that will allow me to characterize and relate behavioral expressions of memory to corresponding
representations within the hippocampus. My central hypothesis is that repulsion of hippocampal
representations will be associated with the exaggeration of differences between similar scene stimuli when
they are verbally recalled. This hypothesis and the feasibility of my approach is supported by a preliminary
study I have conducted which validates that NLP methods are sensitive to subtle distortions in how similar
scene images are remembered. In Aim 1, using NLP methods and a behavioral memory paradigm, I will
test the hypothesis that distortions in memory content are explained by a targeted “movement” of competing
memories away from each other in a high-dimensional semantic space. In Aim 2, I will test the hypothesis
that changes in memory content (measured by NLP methods) are predicted by the degree of repulsion of
hippocampal representations. In addition to supporting my training with new neuroimaging and
computational methods, this project will yield important new insight into how the hippocampus resolves
interference between similar memories. Moreover, the specific combination of techniques and approaches
that I will employ have the potential to open up new avenues of research in the field of episodic memory. In
summary, this research will support my long-term objective of developing innovative methods to understand
how the hippocampus supports the efficient storage of episodic memories.

Grant Number: 5F31MH135686-03
NIH Institute/Center: NIH
Principal Investigator: Anisha Babu