Read on Twitter
Delighted to share our review of hippocampal pattern similarity studies, written with the brilliant @jess_robin_ @rosanna_olsen @morganbarense and Morris Moscovitch. For a quick walk-through, thread below:
https://www.sciencedirect.com/science/article/abs/pii/S0149763420304838
https://www.sciencedirect.com/science/article/abs/pii/S0149763420304838
The hippocampus is one of our collective favourite regions to discuss from computational, theoretical, behavioural perspectives. But our question was quite simple: What have we learned about the hippocampus from fMRI studies of multi-voxel patterns? 1/
The idea for this paper evolved from conversations @jess_robin_ and I have had over the years, centered around the question of why we sometimes see different patterns of data despite similar experimental designs. What determines the degree of pattern similarity? 2/
There are two common ideas. 1: the hippocampus efficiently separates memory traces, supporting memory for individual experiences. 2: the hippocampus integrates across related experiences. Let’s take an example of events taking place in two different cities. 3/
If there’s a representation for each individual experience, we should be able to detect some similarity between encoding & retrieval. This idea is widely accepted. In this paper we call this within-stimulus similarity. 4/
If related event representations are integrated, we should see increased similarity for those events relative to unrelated ones. Conversely, similarity should be lower if the representations are differentiated. Interestingly, both kinds of findings have been reported. 5/
It may be that these effects are driven by hippocampal anatomy. The aHPC and subfield CA1 have been implicated in integration, and the pHPC and DG in differentiation of related events. Further complicating this is the fact that subfield distribution varies along the axis 6/
We thought the best way to approach this conundrum would be to gather every study we could find that used pattern similarity analyses in the hippocampus. We found about 100 relevant papers and categorized them based on comparisons of interest. 7/
We generally found increased similarity for related, relative to unrelated stimuli. A fair number, however, also found decreased similarity, or a mixed pattern of findings. Many of these examined the same subregions, so we don’t think it’s fully driven by the choice of ROIs. 8/
We discuss these studies at length, and try to reconcile these findings depending on the stimulus, timescale & ROI. But the findings suggest that the current behavioural goals matter most, and crucially shape the representational similarity structure in the hippocampus. 9/
Future work could parametrically manipulate the degree of overlap in the encoded content and task demands to test whether, and to what extent, we can ‘push’ representations around. We discuss a few of these avenues. 10/
Thanks so much to coauthors @morganbarense, @rosanna_olsen and Morris! We would like to thank the two thoughtful and constructive reviewers whose feedback has improved this paper, as well as the editor. This has been a positive and fun experience throughout. 11/
Finally, on a personal note: the writing of this paper started in late 2018, and between @jess_robin_ and me, spanned 4 jobs and 4 cities. It’s been a constant reminder of the best aspects of science and I feel very fortunate.
Very happy to discuss any of the above further, and interested to hear feedback/thoughts!
