Background: Xenopus tadpoles are good at avoiding collisions (here’s my video of the process). If collisions are slow enough, they just correct their course a bit, not even breaking a sweat. For faster collisions, they respond later, and with a startle.https://youtu.be/i1g97CNUMK8
-
-
Prikaži ovu nit
-
We know which part of the brain they mostly use to avoid collisions (optic tectum), but we still don’t know how! It is clearly an emergent property, as messing with network activity in any way ruins looming selectivity. But beyond that: a mystery.pic.twitter.com/1edtN7wxNp
Prikaži ovu nit -
So this time we went for high-speed Ca imaging, to hunt for looming-selective circuits! I tried three different stimuli, and different cells had different selectivity (those colorful dots on the right). Looming responses were more salient than non-looming. Sweet!pic.twitter.com/S5thXLRNbm
Prikaži ovu nit -
As tectum is retinotopic, I could see looming stimuli gradually “unroll” (not with a naked eye, but on average). Also there were ensembles. Still neither observation helped to explain looming selectivity.pic.twitter.com/TWmOyIrgFo
Prikaži ovu nit -
But here’s the thing: transmission in the tectum is relatively slow, and I could hope to see not just correlations in activation of diff cells, but _propagation_ of signals through the tectum! I used transfer entropy to identify connections…pic.twitter.com/qTKjSBiuKI
Prikaži ovu nit -
…And ended up with 30 weighted oriented graphs of about similar sizes (~100 cells in each, as that’s what fit in the field of view). Half from younger tadpoles, half from older ones. And for each node (cell) I know whether it’s looming selective. How to compare them?pic.twitter.com/VXQF6Xfr7E
Prikaži ovu nit -
First, I looked at a bunch of network properties. Some changed in development (degree distribution, modularity), some didn’t; many were decidedly non-random (based on constrained rewiring). That’s a good sign, but still did not help to explain looming.pic.twitter.com/Vjis38vPae
Prikaži ovu nit -
What about the
#centrality of looming-selective cells? I really hoped to see them acting as “integrators” within the network, and indeed there are some indications of it (higher Katz centrality, higher in-degree, stronger activity overall), but the effects were small.pic.twitter.com/8GIe6UoD3Q
Prikaži ovu nit -
So I decided to build a model, and study all same stuff in the model. If the model matches all other properties we observed in the tectum, them maybe we can trust it, and use it as an “extension” of biological experiments!
Prikaži ovu nit -
The model consisted of neurons, governed by spike-time-dependent plasticity, homeostatic intrinsic plasticity, synaptic competition, and subjected to pattern visual stimuli from a “fake retina”. And then I analyzed model results _exactly_ the same way I did it for tadpoles.pic.twitter.com/RzHBc3yH0J
Prikaži ovu nit -
The results are cool, but not as clear as I hoped. (There’s a table in the paper, with a list of “atomic statements”, and either TRUE or FALSE for each statement, for both bio experiments and the model). The model learned to detect looming stimuli, and not quite as I expected.pic.twitter.com/AAzF2roFxg
Prikaži ovu nit -
It looks that it developed “synfire chains” of neurons connected to each other in the same sequence in which they are activated during a loom. And then responses of these synfire chains are combined by cells with high Katz centrality. And even this is so tricky to demonstrate!pic.twitter.com/CsQD3DYCeL
Prikaži ovu nit -
To sum up, the paper hast lots of fun analyses and results (essentially, it's 2 papers stacked one upon another: first a full experimental one, then a model). And 15 pages of “Methods”! And though I did not fully close the question of how tadpoles detect impeding collisions...
Prikaži ovu nit -
...I show that synfire chains are a very likely candidate. Yet even in a model, where I supposedly have full control over both development and analysis, it feels that looming detection is a bit too distributed to pinpoint the mechanism, or describe it in one sentence!
Prikaži ovu nit -
All code for the paper, and all summary data is available on github https://github.com/khakhalin/Ca-Imaging-and-Model-2018 … I plan to release raw and final reconstructions, raw data, and some model runs as well; just need to find a way to do i best, and pick some good formats.
Prikaži ovu nit -
And also, as it’s a single-author paper, I would Really Appreciate any feedback you might have! If you could spare any! What to fix? What to drop? Should I split it in two? Please let me know what you think, before I submit it anywhere! (khakhalin at bard dot edu). Thank you!
Prikaži ovu nit -
Finally, some trivia: it took me 5 years to write this thing :) I recorded the data in 2014, immediately before leaving
@AizenmanLab and joining@BardCollege, and could not process it for 2 years. So happy to finally share it all with the world!Prikaži ovu nit
Kraj razgovora
Novi razgovor -
Čini se da učitavanje traje već neko vrijeme.
Twitter je možda preopterećen ili ima kratkotrajnih poteškoća u radu. Pokušajte ponovno ili potražite dodatne informacije u odjeljku Status Twittera.