This paper goes way beyond place cells and grid cells:
“Furthermore, by sorting the behavioral data according to the network states, we found that each state corresponded to a distinct behavior, irrespective of the position of the mouse. We termed these behaviors‘Rearing’,‘Turning’,‘Start run’,‘End run’, and ‘Drinking’. We confirmed this clustering-based classification using manual labeling of the mouse’s behavior”
I consider this little bit to be very important:
“The internal structure was conserved across mice, allowing using one animal’s data to decode another animal’s behavior. Thus, the internal structure of neuronal activity itself enables reconstructing internal representations and discovering new behavioral variables hidden within a neural code.”
This is an interesting read. I’m not quite through with it, but this is a bit misleading:
I am not sure the actions in the PFC scatter plot were even labeled, so it’s not even clear what those actions are. I would expect there to be some similarity simply because all mice have pretty much the same anatomy, environments, and very likely genetics in laboratory settings. But I certainly don’t think this implies you can invoke one mouses “affordances” by using another mouses affordance representations as a trigger.
The distribution in the hippocampus is NOT the thing that is being shown - it is that the activity pattern can be dimension reduced using the techniques outlined in the article. The article makes it clear that the information that is encoded is mixed together and has to be extracted to get any useful meaning from it.
The reduction has variability as described, but the general location in the decoded state-space indicates behavior.
That may work for hippocampus, but the “Mouse 3 / Comp. 1” blob is not derived from hippocampus, it is derived from anterior dorsal nucleus of the thalamus (ADn) and the post subiculum (PoS) during food foraging (they pulled this from the REM study portion of the paper). To me, it does not fit with the rest of the graphic. I’m confused about why it is shown. I don’t think your arrow from this image to the “unsupervised decoder - Mouser 1” is correct.
You are correct, the ensemble that they are using (CA1/hippocampus) for the track position is called out in figure one, frames A though D. So much for dashing off a quick sketch when I should be working.
I maintain that once you specify the source of the cell ensemble to the either mPFC or CA1 from the hippocampus, (depending on if you are explaining either behavior or position) the rest of the flow is correct as described.
Figure 5, block 2, is the clearest illustration of the process of reading from one mouse and predicting what that will mean in a second mouse.
Figure 2, A through F, (source ACC/mPFC) is the clearest in showing how the parts of the derived lower dimension model matches to the behavior.
In figure 5 they do a poor job of trying to describe the variability of the measurement over time and between mice. I see that they are trying to relate how the variability is not all that important in showing the main information when comparing between mice, (the track model is offered as an example) but the relation to the two comp 2 images mostly distracts and does not add to the message in the rest of that figure. Yes, there is variability but the relation to the variability in the track data is not clear.