← @ElDuvelle Twitter archive
El Duvelle 🌍
@ElDuvelle
The rodent hippocampus is amazing: place cells, time cells, reward cells...
Did you know about the puzzling #SplitterCells?
We sought to understand them in an experimental & computational review on temporal context & latent state models!
Preprint: https://psyarxiv.com/9z4wr/
1/25
06/08/2022, 17:56:36
Favs: 304
Retweets: 82
linkEl Duvelle 🌍
@ElDuvelle
2. What are "splitter cells"?
AKA trajectory-dependent cells, differential activity, context-dependent firing, journey coding... these neurons 'split' their activity, at the same place, depending on a non-sensory parameter such as current trajectory. Recognize this?
06/08/2022, 17:56:38
Favs: 10
Retweets: 0
linkEl Duvelle 🌍
@ElDuvelle
3. Here's what a real SplitterCell sounds like, from @rmgrieves et al.' great 2016 paper (https://elifesciences.org/articles/15986)
Notice how strikingly different the firing in the start box is, depending on the future trajectory!
(🔉sound = spikes from 1 splitter cell)
06/08/2022, 17:56:45
Favs: 13
Retweets: 1
linkEl Duvelle 🌍
@ElDuvelle
4. Why is this puzzling? While classical place cell firing can be mostly explained with a combination of sensory / self-motion cues, the firing of splitter cells appears to represent non-sensory, latent variables, such as the context or current/ past/ future trajectory.
06/08/2022, 17:56:46
Favs: 8
Retweets: 0
linkEl Duvelle 🌍
@ElDuvelle
5. Indeed, a robust property of splitter cells only revealed in some cleverly designed tasks is that they can encode the past (= retrospective splitters) or the future (=prospective splitters) trajectory.
How can the brain (of a rat or mouse!!) do this? 🤔
06/08/2022, 17:56:47
Favs: 8
Retweets: 1
linkEl Duvelle 🌍
@ElDuvelle
6. First, to better understand splitter cells & their sometimes unstable firing at the single cell level, we can look at the 'neural activity space' (here, plotting principal components of the ensemble activity)
This view helps distinguish classical place cells from splitters:
06/08/2022, 17:56:47
Favs: 9
Retweets: 0
linkEl Duvelle 🌍
@ElDuvelle
7. So where does this activity come from?
One explanation for retrospective firing is to combine info about the current location (classical place cell) and a memory trace of the recent past - suggested by the 'Temporal Context Model':
(e.g. @HasselmoMichael, @jeremyRmanning)
06/08/2022, 17:56:48
Favs: 6
Retweets: 1
linkEl Duvelle 🌍
@ElDuvelle
8. "but that doesn't explain prospective splitters!" - you say.
Don't panic! The closely related Successor Representation (SR) provides a prediction of the upcoming future, sufficient to generate prospective splitters.
(e.g. @neuro_kim, @criticalneuro)
06/08/2022, 17:56:49
Favs: 5
Retweets: 0
linkEl Duvelle 🌍
@ElDuvelle
9. The temporal context model (including SR) explains not only retrospective & prospective splitters, but also their different spatial distributions: retrospective towards the start of central stem, prospective towards the choice point.
06/08/2022, 17:56:49
Favs: 5
Retweets: 0
linkEl Duvelle 🌍
@ElDuvelle
10. One cool property of temporal context is that it explains why adding a delay to a continuous alternation task 'shifts' the splitter activity earlier in the maze, mostly to the delay phase!
06/08/2022, 17:56:50
Favs: 4
Retweets: 0
linkEl Duvelle 🌍
@ElDuvelle
11. However, the temporal context view does not explain all splitter signal properties: for example its variations with learning or task demands, with some tasks not finding splitters at all.
To make sense of these, we turn to the "latent state inference models"...
06/08/2022, 17:56:51
Favs: 3
Retweets: 0
linkEl Duvelle 🌍
@ElDuvelle
12. Latent (or hidden) state inference models discover the structure of the task and classify experience into discrete task states (e.g. @dileeplearning, @behrenstimb, @gershbrain)
They explain better the learning/task-demands related properties mentioned above 👍
06/08/2022, 17:56:52
Favs: 7
Retweets: 1
linkEl Duvelle 🌍
@ElDuvelle
13. Some properties remain puzzling, for example retrospective splitters are consistently more numerous than prospective splitters, across studies (on average ~2x more retrospective cells)
Check Table 1 for a summary and let us know what you think!
06/08/2022, 17:56:52
Favs: 4
Retweets: 0
linkEl Duvelle 🌍
@ElDuvelle
14. Main conclusion:
From this review, our best explanation for splitter cells is that they arise from both the temporal context and latent state models working cooperatively!
This shouldn't really surprise us: after all, the hippocampus likes to mix and match…
06/08/2022, 17:56:53
Favs: 5
Retweets: 1
linkEl Duvelle 🌍
@ElDuvelle
15. Take-home messages:
- splitter cells are more complicated than they look!
- focusing on this very specific topic and paradigm allows to figure out some principles of the hippocampal function, and the brain
- many avenues of research on splitters are still unexplored
06/08/2022, 17:56:53
Favs: 7
Retweets: 0
linkEl Duvelle 🌍
@ElDuvelle
16. This thread only scratches the surface, many points could be developed. Our review might have some answers: enjoy the read!
Please don't hesitate to ask for clarifications or make comments on here, and let us know about any forgotten or misinterpreted studies! 🙏
06/08/2022, 17:56:54
Favs: 5
Retweets: 0
linkEl Duvelle 🌍
@ElDuvelle
17. Most credit for this review goes to @mattmizumi who supervised this work and immensely contributed with original ideas, lots of writing and very nice figures!
We were also lucky to have the splitter cell expert and literature wizard @rmgrieves as a precious co-author.
06/08/2022, 17:56:55
Favs: 5
Retweets: 0
linkEl Duvelle 🌍
@ElDuvelle
18. Note 1: We list some untested predictions of models, and suggest some informative experimental paradigms, to understand better the splitter signal... including a maze like this 👀
06/08/2022, 17:56:56
Favs: 7
Retweets: 0
linkEl Duvelle 🌍
@ElDuvelle
19. Note 2: In biological splitter examples, you might notice how position or other parameters sometimes change with the past/future trajectory. Taking those into account in task design (see below for a quite extreme example) or analysis is crucial!
➡️ see our Box 3
06/08/2022, 17:56:57
Favs: 4
Retweets: 0
linkEl Duvelle 🌍
@ElDuvelle
20. Note 3: If, like me at the start of this review, you thought splitter cells were mostly present in dorsal hippocampal CA1 (and maybe also prefrontal cortex / nucleus reuniens)... think again!
➡️see our Box 1
The end!
06/08/2022, 17:56:58
Favs: 5
Retweets: 0
linkEl Duvelle 🌍
@ElDuvelle
21. Still there? Are you ready for some splitter cell trivia?
(I'll post the answers later)
Question 1: how many separate experimental studies on splitter cells do you think exist?
06/08/2022, 17:56:59
Favs: 2
Retweets: 0
linkEl Duvelle 🌍
@ElDuvelle
22. Q2: Which paper first separately analyzed CA3 splitter cells?
1. Ferbinteanu & Shapiro 2003
2. Bahar & Shapiro 2012
3. Ito et al., 2015
06/08/2022, 17:57:00
Favs: 2
Retweets: 0
linkEl Duvelle 🌍
@ElDuvelle
23. Who first analyzed splitter cells using a population approach?
(to our knowledge)
06/08/2022, 17:57:01
Favs: 2
Retweets: 0
linkEl Duvelle 🌍
@ElDuvelle
24. Q4: Which study first demonstrated prospective & retrospective splitters?
06/08/2022, 17:57:01
Favs: 3
Retweets: 0
linkEl Duvelle 🌍
@ElDuvelle
25. Q5: Who first used the term "splitter cell"
1) informally
2) in a published form?
Open question!
The end (for real)... Hope you enjoyed this thread, looking forward to your feedback on the review! 😃
(link reminder: https://psyarxiv.com/9z4wr/)
06/08/2022, 17:57:02
Favs: 6
Retweets: 0
link