🚨🚨🚨NEW PREPRINT🚨🚨🚨
A big mystery in brain research is what are the neural mechanisms that drive individual differences in higher order cognitive processes. Here we present a new theoretical and experimental framework, in collaboration with Vincent Tang, Mikio Aoi, Jonathan Pillow, Valerio Mante, @SussilloDavid and Carlos Brody.
https://www.biorxiv.org/content/10.1101/2022.11.28.518207v1
1/16
First, we trained rats to perform a task requiring context-dependent evidence accumulation towards a decision (like the monkeys in Mante et al., 2013).
Rats were presented with a train of auditory pulses, and were cued to selectively accumulate location (ignoring frequency) or to accumulate frequency (ignoring location).
2/16
Using an automated, high-throughput training procedure we trained 20 rats to solve the task with high performance, collecting more than 120,000 trials for each rat!
While rats performed the task, we recorded neural responses in frontal cortex to study population activity.
3/16
First, was neural activity in rats similar to monkeys, where irrelevant evidence was not gated before reaching the Frontal Eye Field?
The neural trajectories we observed in rat Frontal Orienting Fields were strikingly similar 🤯, leading us to the same conclusion!
4/16
Then, we wanted to generalize this result, and study the entire space of possible network solutions, so we turned to math!🤓
Following the footsteps of Mante et al., 2013, we studied the properties of linearized system dynamics around fixed points.
5/16
Rearranging the formula for differential evidence integration revealed that any network solving the task uses a combination of three distinct components, i.e. under assumptions supported by both monkey and rat data, a triangle constitutes the space of all possible solutions.
6/16
Can we tell apart different mechanisms using neural data collected during the task? 🤔
Yes!!! We developed a new analysis that leverages the high statistical power provided by our pulsatile stimulus to retrieve population trajectories evoked by single pulses of evidence!
7/16
We then developed a new method to engineer recurrent neural networks (RNNs) to implement different mechanisms, and we applied the analysis to their dynamics as they solved the task.
The pulse-evoked trajectories clearly distinguish RNNs using different mechanisms.
8/16
Why did we develop a new method to build RNNs, you might ask, instead of simply training RNNs using backprop and studying those solutions?
It turns out that RNNs trained using backprop only find a small subset of the larger space of solutions!
9/16
Ok, now that we have validated our analysis using RNNs, let’s apply it to data from rat brains!🧠
The result? Different individual rats display different pulse-evoked neural trajectories, very similar to those produced by RNNs implementing different mechanisms!🤯
10/16
Our theory also predicts a specific behavioral “fingerprint”. The context-dependent behavioral kernels should be a reflection on the time axis of the pulse responses.
When measuring these behavioral kernels using logistic regression, that’s exactly what we observe in RNNs.
11/16
Remarkably, it works out also in rats! The variability of neural pulse-evoked trajectories is highly correlated with the variability of behavioral kernels, strongly suggesting that both measurements reflect the individual variability of the underlying mechanisms!
12/16
In conclusion, our results provide a new experimentally-supported theoretical framework to analyze biological and artificial systems performing flexible decision-making, opening the door to the study of individual variability in neural computations underlying higher cognition!
13/16
First of all, I can't thank enough my wonderful Mentor, Carlos Brody, for being an inexhaustible source of guidance, wisdom and support throughout this project! 🙏 🙏 🙏
Thanks also to Vincent Tang, Mikio Aoi, @SussilloDavid , Valerio Mante and Jonathan Pillow for a super-fun collaboration!
Finally, thanks to the Simons Collaboration on the Global Brain (SCGB) and the Simons Foundation Autism Research Initiative (SFARI) for supporting my research!
14/16
We are super excited about two future directions for this research.
First, we plan to use high-density probes to record large neural populations across brain areas as rats solve the task, and to directly study recurrent dynamics in simultaneously recorded neural populations using latent-based approaches (e.g. LFADS and PLNDE).
15/16
Second, with the support of a SFARI Bridge to Independence Award, I am extremely excited to use this task, and these Systems and Computational approaches to study the role of genetic mutations in cognitive flexibility in rat models of autism!
https://sfari.org/2021/08/19/sfari-2021-bridge-to-independence-fellows-announced/
16/16
@marinopagan @SussilloDavid
This is really exciting. Congratulations!
Here we've been talking a lot about complex dynamical systems, and I have so many questions.
First one: Over on the bird site, @SussilloDavid wrote, "Hell, I’ll go bigger. This work provides strong indirect evidence for the computation thru dynamics framework."
Can you please unpack that just a bit more @SussilloDavid ?
@PessoaBrain @DrYohanJohn @manlius @kordinglab @neuralengine @cogneurophys
@complexsystems
@NicoleCRust @marinopagan @SussilloDavid @DrYohanJohn @manlius @kordinglab @neuralengine @cogneurophys @complexsystems
I'm fan #1 as I've said before!
However, as @DrYohanJohn would say, what would be the alternative?? It has to be dynamics! The point I'm making is about the quote of course, not the specifics of the paper.
In 2022 we should be more comfortable with dynamics...
Perhaps we’re in need a better name then. The point is that we are beginning to make sense of complex cortical dynamics in higher order areas using rnns and understanding *how* dynamical systems compute.
The name I guess is historical in that vision has long been the predominant paradigm in systems neuroscience. While of course all neurons are massively recurrently connected, the high level systems view used feed forward networks to approximate function, from H&B to H-Max style models.
Here is an example of a computation through dynamics analysis of context dependent processing of English sentiment. It’s worth reading if you wanna know what we (I?) mean!
This problem could *easily* be solved by a feed forward network (no dynamics), so CtD isn’t vacuous fluff, even if it’s confusing good faith folks.
As for brains, feed forward is still the dominant approximation, so CtD means something there too.
http://proceedings.mlr.press/v119/maheswaranathan20a/maheswaranathan20a.pdf
Of course it's not fluff. I guess I never thought people would think forward architectures are a possible way forward... 😅
I grew up in a world of dynamics where bidirectional interactions are basic building blocks.
Imo feed-forward systems can't work in more general settings/behaviors.
Perceptual processing is not a passive process of "observing the world" but at the very least works in a tight perception-action cycle, if not even more heavily action based.
I.e., external world signals are integrated with ongoing processing!
We need more "naturalistic" experiments to reveal this more fully, which is still very far except in rodents.
@SussilloDavid @cogneurophys @complexsystems
Curious fact from some of us a bit older: visual responses were considered to be basically the same in the anesthetized animal as in awake!
I bet younger generations will find this a bit surprising if not shocking.
Ok, I exaggerate. Important computational principles (mechanisms?) were discovered this way, though they may not be capturing natural brain function.@neuralengine @kordinglab @NicoleCRust @marinopagan @SussilloDavid @PessoaBrain @DrYohanJohn @manlius @cogneurophys @complexsystems We found evidence that individuality in odor-choice assays can arise in the sensory periphery https://www.biorxiv.org/content/10.1101/2021.12.24.474127
We think of it as differences in odor-coding rather than dynamics, but I suppose some brief integration is needed to realize small individual differences in firing rates.
@neuralengine @kordinglab @marinopagan @SussilloDavid @PessoaBrain @DrYohanJohn @manlius @cogneurophys @complexsystems
I suppose it's only fair for me to copy the second part of that post as well, which elaborates a bit: "By reasoning about input projections onto *putative* line attractors, Marino could explain a surprising amount of behavioral heterogeneity.
Destined to be a classic."
Still looking forward to the unpack now that @SussilloDavid gets at least 500 characters.
@kordinglab @NicoleCRust @marinopagan
I’d distinguish computation without dynamics as feed forward computation. y = f(g(h(x))) isn’t dynamical imo x_n = f(f(f(…(x_0)))) is imo
As vision has historically been the paradigm, the top equation has been the predominant approximation in neuro.
Recurrence (bottom) brings a different set of formalisms and appears be useful in other, more traditionally ignored areas.
@NicoleCRust @SussilloDavid @kordinglab
In Mante/Sussillo, monkey data suggested that irrelevant info is not gated before reaching frontal cortex. To understand how computations in frontal cortex can selectively accumulate relevant evidence, they trained RNNs and found that most RNNs solved the task using context-dependent recurrent dynamics. This led to the idea that context-dependent recurrent dynamics is a candidate solution for how the brain solves the task.
1/3
@NicoleCRust @SussilloDavid @kordinglab
In our work, we asked whether other solutions exist. Our key theory result is that, under assumptions supported by both monkey and rat data, all networks solving the task use a combination of three components. This led us to multiple insights: 1) Most RNNs trained with gradient descent converge to a single solution, context-dependent recurrent dynamics (which we call selection vector modulation), but the full space of solutions is larger than that!
2/3
@NicoleCRust @SussilloDavid @kordinglab
2) Using the math of our theory result (Eq. 1 and Eq. 2 in the paper) we can now engineer RNNs spanning the full space of solutions. 3) Different solutions have anatomical implications (i.e. impose constraints on network architecture), as well as implications on the resulting neural dynamics and behavior. 4) A specific link exists between neural dynamics and behavioral integration, which was crucial to explain individual variability in our rat data.
3/3
@marinopagan @NicoleCRust @SussilloDavid @kordinglab
Sorry to butt in but maybe this is a good time to raise this point I've wondered for a while. In this task colour and motion are separate and so it might be less surprising that there is no attention gating or suppression to pfc. In tasks where features are more integrated or not orthogonal, it would make more sense to suppress irrelevant info. And of course others have shown enhancement of relevant vs irrelevant in pfc.
@marinopagan @NicoleCRust @SussilloDavid @kordinglab
I guess I wonder how general it is to fit to this task? NNs in my experiences don't show "attention" effects automatically just from training.
@marinopagan @NicoleCRust @SussilloDavid @kordinglab
Attention effects meaning units increase or decrease firing rate or modulate tuning curves,which often also affects the strength of the stimulus rep (eg lower decoding accuracy).
To show increase or decreasing in firing rate some tasks have a neutral condition where both are attended.
@marinopagan @NicoleCRust @SussilloDavid @kordinglab
Great stuff thanks. Have to check it out!
@marinopagan @Rob_Mok @NicoleCRust @SussilloDavid @kordinglab
There’s good work in cog control on this q https://pubmed.ncbi.nlm.nih.gov/18760657/
But, I think generalization is a little overrated. The brain is optimizing for local conditions, so it will use task-appropriate strategies when necessary.
Conflict between competing neural representations is thought to serve as an internal signal for the recruitment of 'cognitive control', which resolves conflict by biasing information processing in line with current task demands. Because conflict can occur at different levels of stimulus and response …
@marinopagan @SussilloDavid
@PessoaBrain @DrYohanJohn @manlius @kordinglab @neuralengine @cogneurophys
@complexsystems
On first read, this reminds me of a beautiful 2004 paper from Eve Marder's lab (https://doi.org/10.1038/nn1352). Even within a relatively simple dynamical system, there are many ways to achieve the same goals. (Later work showed there are also many different ways for the system to break down under stress)
I'll be re-reading this when I have more time; there's a lot to take in.
@lucaspinto @SussilloDavid @carlosbrody
Thank you Lucas!!!
Marino Pagan
Nicole Rust
Luiz Pessoa
David Sussillo
kari hoffman
Konrad Kording
Simon Schultz
Ben de Bivort
Rob Mok
harrison ritz
Mike Schmidt
Lucas Pinto