Thank you for attending the 2021 Showcase Symposium - Day 2 (December 8) or viewing the recording on Youtube .
Below are questions from our audience that were answered with written responses by our speakers. Additional Q&A is at the end of each talk in the event. We invite you to use this thread to ask additional questions or follow up on the conversation more generally.
Question: Cindy Poo. My layperson question. there are 4 odors. is there influence because of a rat odor preference? How does that weigh in their behaviors? Thanks! (super interesting talk. Just saw polar bears and learning how their olfactory sense is 1000x stronger than humans.
Cindy Poo: Great question. Yes, we would expect that inherent odor preference would affect how animals behave in this kind of task. This is why we chose odors that have neutral valence for the animal. This is an important control here, so that we know animals won’t be biased.
Question: Hi Cindy, great talk! How do you ensure that odorant concentration isn’t spatially biased at a given time point during the task / is it possible spatial coding in piriform cortex is actually encoding gradients in odor intensity?
Cindy Poo: Thank you for this question. In our task, odors are only available when rats have their noses inside the odor port (we transiently inject odors into a constant flow of clean filtered air). So there are no spatial odor gradients available when their noses are not inside the port. But the question of how odor gradients can be used during more natrualistic behaviors to help guide navigation is a very interesting question!
Question: great talk Cindy! How does previous reward (t-1 trial) bias the spatial computation and decision making?
Cindy Poo: thank you for the great question. It is possible that computations in brain regions such as Orbitofrontal cortex, or perhaps Olfactory tubercle (ventral striatum) would be involved in updating the value of different spatial locations. We did look at the effect of reward on the piriform spatial map (correct and error trial analysis), and we found that rewards did not affect the spatial map in piriform.
Question: Hi Cindy, this is an awesome talk. 1) Did you find splitter cells (i.e., place cells firing only during certain trajectory when away from the olfactory cues)
Cindy Poo: very interesting question! We actually have not looked at splitter cells, but this is a great suggest. we should! We are also interested in looking at sequences in piriform such as replay/replay type activity in the future.
Question: Hi Cindy, 2) for those neurons time-locked to sniffing only, why was there a coherence prior to the sniffing onset?
Cindy Poo: they are coherent to respiration before the active odor sampling because they tend to be coupled even during baseline breathing, and not just active odor sampling. I hope this answers your question — let me know if this is unclear.
Question from Staci Sorensen, Allen Institute staff: Can you get local morphology, Mike? You mentioned that you can’t get long-range projections… Thanks for the nice talk!
Mike Taormina: Hi Staci, I’m afraid the answer is that we don’t know. Prior to adopting an expanded hydrogel tissue processing, we didn’t really have much hope for this. The short time that we’ve had access to this higher quality image data hasn’t afforded the opportunity to work with a researcher who is more knowledgeable in extracting cell morphology from image stacks.
Question from Marina Garrett, Allen Institute staff: Hi Lucas, wonderful talk. I am wondering if you could speculate a bit about the role of the SI and DB cholinergic projections. One specific question is why the DB projections would target visual cortex but not other sensory regions like barrel cortex? You could expect different cholinergic projections to target sensory vs. association areas, but what you see is a posterior vs. anterior specificity. Do you think this is related to the different timescales of activity across cortex?
Lucas Pinto: Hi Marina, thanks, and great question! I chose to highlight the AP gradient but it’s more complex than that — there is also an ML gradient and even a laminar gradient. My general speculation is that these projections are recruited to change dynamics specifically in subsets of cortical areas. I do think the clearest dichotomy between posterior/anterior regions is in timescales (and in fact ACh changes integration scales both at the network and single-cell level). There might also be implications for specific changes in bottom-up sensory processing vs. more predictive coding schemes that favor top-down information flows. Hopefully this makes sense! Happy to elaborate.
Question from Forrest Collman, Allen Institute staff: Kanaka - Did I understand you correctly that you said that these inter-areal currents are not accessible experimentally… wouldn’t recording fluorescence from axonal arbors labeled as coming from one area in another area be the equivalent experimental measurement to validate against? Or is there some aspect that this would not capture (other than its hard/impossible to do this for all area sinks/sources simultaneously).
Kanaka Rajan: You could test the currents certainly, exactly as you said. But without knowing all the sources and sinks, as you said, it would not be a complete picture. RNNs like this can tell you where to look for such currents and what timescales they vary. The other thing I didnt mention but wish I did is: currents come with signs, so it is possible to tell whether the currents are excitatory or inhibitory, before you do the experiment to block or hyperactivate a source.
Question: Do you train separate models for early vs late conditions? Also, how do you generate training samples for your multti-region RNN (small timesteps or the whole timeseries at once?)
Kanaka Rajan: No, we use a thing we call chaining… the data are divided into epochs and then RNNs are trained based on initializing to the finally trained state from the previous epoch.
Question from Farzaneh Najafi, Allen Institute staff: Thanks for a great talk. Did the RNN help with revealing where the Raphe suppression during the inescapable stress come from? Is it related to interactions with Habenula or Telencephaon?
Kanaka Rajan: Raphe suppression came partly from the ffwd interactions from the Hb, and from the dorsal thalamus, which didn’t make it into the talk
Question from Tom Chartrand, Allen Institute staff: Great talk Kanaka! I’m wondering if you can explore a broader solution space when fitting the RNNs, rather than a single solution, and use that to assess the robustness of the predictions?
Kanaka Rajan: Can you tell me what sort of broader solution space you had in mind? One thing we are doing now is to train a single RNN to match data from a shocked fish AND control fish… you see, there is only one connectivity map brainwide. If a shocked fish is not shocked, it’s a control fish… So that is a broader, yet more consistent solution.
Scott Linderman, Next Generation Leader: Thanks for the nice talk, Kanaka! Technical question: it seems like the number of neurons could be much larger than the number of timesteps. In that case, the individual weights would be underspecified, but you might still be able to accurately estimate the average weights between areas or the state space projections, like you showed. But in that regime, do you see added value from the single-unit RNN as compared to fitting a low-d state space model directly?
Kanaka Rajan: These data come from a 45 min long experiment sampled at 1Hz, so the number of timepoints is larger than the N = 10-40000 neurons… in general though, this is our first order pass - next up, i want to see how well the RNN built on control fish generalizes to data from a shocked fish, and vice versa. The weight profile can’t vary wildly between shock and ctrl condition, you see.