Marino PaganDec 1, 2022
David Sussillo

1/
On the bird site, I said that Marino and the Brody lab’s paper is a major advance. @NicoleCRust asked why. I’m sure Marino and Carlos have their own bits that excite each of them the most. This thread explains what excites me the most, as a peripheral contributor.

In short, I’m super stoked on the work because it’s a significant *indirect* validation of the line attractor hypothesis for frontal cortices and for computation through neural population dynamics in general.

Here goes nothing!

Dec 1, 2022 at 4:22amWeb
Show threadDavid SussilloDec 1, 2022
2/
The rats have to integrate audio pulses, which are clicks that vary in location (left or right) or frequency (high or low). The animal is cued to pay attention to either location or frequency. This is in direct analogy with “Mante, Sussillo” where the monkeys had to integrate random dots and pay attention to either color or motion.
Show threadDavid SussilloDec 1, 2022
3/
So if you believed in things like line attractors in the brain, and that both the relevant and irrelevant information make it into these frontal areas, then the input that is integrated has to project onto the selection vector (SV, left eigenvector) of the line attractor (LA, right eigenvector) while irrelevant information has to be orthogonal to the SV.
Show threadDavid SussilloDec 1, 2022
4/
While alignment of input and SV determines whether information is integrated, alignment of input and LA determines *how fast* information is integrated. If an input is highly aligned to the line attractor, then information is integrated basically immediately. If an input is aligned with the SV, but only partially with the LA, then it will require recurrent decay dynamics of all the other modes to ultimately get the input onto the LA, which is a slower process.
Show threadDavid SussilloDec 1, 2022
5/
A trial will have a beginning, middle, and end. If you have a slow integration of a click, and the click happens at the end of the trial, it won’t be integrated because the recurrent dynamics haven't had time to place the evidence from the SV onto the LA. Thus late clicks that use dynamics to integrate *shouldn't influence the rat's behavior.* If you have fast integration of a click, then clicks at the end of the trial should be integrated and *should influence the rat's behavior.*
Show threadDavid SussilloDec 1, 2022
6/
Due to the pulsatile nature of the inputs, Marino was able to develop a "neural kernel" that estimates the weight of clicks at the beginning, middle, and end of the trial from neural data. What he showed is that there is quite a bit of variability in these kernels on a rat-by-rat basis. (In RNNs, his methods for these neural kernels essentially tracked how information is integrated, so it's an indirect measure of things like the angle of the input with the line attractor.)
Show threadDavid SussilloDec 1, 2022
7/
Of course, Marino recorded the choices of the rats, so there were behavioral measurements as well. Marino was able to estimate a "behavioral kernel", that is how a click at the beginning, middle or end of a trial influenced the rats’ behaviors (choices). These kernels did not use neural data at all and are based solely on a rat’s choices.
Show threadDavid SussilloDec 1, 2022

8/
So what Marino showed: neural kernels that indirectly measure alignment of the input with the LA correlate strongly on a rat-by-rat basis with behavioral kernels.

I.e, if the neural kernel implies high alignment of the input and line attractor, then the behavioral kernel showed that late clicks are important to the rat.

If the neural kernel implied low alignment of the input with the LA, then the behavioral kernel showed that the late clicks aren't as important to the rat's behavior.

Show threadDavid SussilloDec 1, 2022

9/
That's why I'm so stoked!

It's indirect, but powerful evidence of a number of things I talk about all the time that aren't yet proved.
1. Line attractors.

2. Non-normal dynamics: the utility of selection vectors and the fact that they may not be aligned to the line attractor.

3. Computation through neural population dynamics in the first place! State-space considerations using fixed point structures and linearizations strongly relate to heterogenous behaviors across a population of rats.

Show threadDavid SussilloDec 1, 2022
10/
Plus the work generalizes "Mante,Sussillo"! We only found cases where the RNNs used context-dependent dynamics to integrate information on a selection vector. But Marino showed there's an entire family of solutions. Some rats use recurrent dynamics, and some do not. But either way, the irrelevant information always makes it to the frontal cortex and it's really about a family of solutions in how integrating relevant information happens, not in some upstream sensory processing region.
Show threadDavid SussilloDec 1, 2022

11/
Crazy bonus! We can link computational through neural population dynamics to across-subject behavioral variability; that’s wild!

END

PS If you understood any of that, well Amen to you and me both! 😆 😍

Show threadDavid SussilloDec 1, 2022
12/
Here’s the link to the preprint
https://www.biorxiv.org/content/10.1101/2022.11.28.518207v1?rss=1
Show threadSimon SchultzDec 1, 2022
@SussilloDavid nice explanation. 👍
Show threadNicole RustDec 1, 2022
@SussilloDavid
Really, really exciting. Thank you for this explanation. A masterclass explanation here!
Show threadMitchell TillmanDec 1, 2022
@SussilloDavid For those not on the birdsite, can you please link to this paper? Sounds like a fascinating step forward in motor control!
Show threadCarlos BrodyDec 1, 2022

Thank you @SussilloDavid for such great posts about Marino's @marinopagan work!

(yes, I know David is a coauthor --he's awesome to work with!-- but it's still really nice! 😊👍🏽)