Long-overdue thread on our latest work using the IBL data to reveal the shared organizational principles of the neural code in the cortex.
A systematic analysis of categoricality 🧱 and dimensionality 📐 of the neural code across 40+ cortical regions.
doi.org/10.1101/202...
👇 1/n
Rarely categorical, always high-dimensional: how the neural code changes along the cortical hierarchy
A long-standing debate in neuroscience concerns whether individual neurons are organized into functionally distinct populations that encode information differently ("categorical" representations) and the implications for neural computation. Here, we systematically analyzed how cortical neurons encode cognitive, sensory, and movement variables across 43 cortical regions during a complex task (14,000+ units from the International Brain Laboratory public Brainwide Map data set) and studied how these properties change across the sensory-cognitive cortical hierarchy. We found that the structure of the neural code was scale-dependent: on a whole-cortex scale, neural selectivity was categorical and organized across regions in a way that reflected their anatomical connectivity. However, within individual regions, categorical representations were rare and limited to primary sensory areas. Remarkably, the degree of categorical clustering of neural selectivity was inversely correlated to the dime
On a larger scale, the brain is clearly functionally and anatomically organized. However, many studies at single-neuron resolution show a complex and seemingly disorganized code, especially in cognitive areas. How do we reconcile these two seemingly conflicting perspectives? 2/n
To answer this, we developed a set of analysis pipelines to systematically study the structure of neural representations from two perspectives: (a) single neuron selectivity (b) representational geometry - and a mathematical theory to understand their mutual relation. 3/n
What does "structure" mean in these two spaces? In the conditions space, neurons could form functionally distinct clusters (categorical representations); in the neural space, conditions could form low/high-dimensional geometries with different computational properties. 4/n
We analyzed the neural representations of cognitive, sensory, and movement variables in 43 mouse cortical regions (15000+ cells, IBL BrainWide data set) and compared them with anatomical information of cortical connectivity (Allen Atlas). 5/n
To study single-neuron responses, we developed a reduced-rank regression model (RRR model), which captures well time-varying neural activity in an interpretable set of parameters, giving an 8-dimensional embedding (8 variables) for every single neuron. 6/n
We started from a large anatomical scale, studying the avg. selectivity for each region. The more regions are anatomically connected, the more similar their selectivities are. Also, we can decode the region from single neuron response profiles. Well-connected regions are harder to decode. 7/n
What about categorical clustering? We developed a pipeline that (1) finds the best clusters (2) computes quality (silhouette) (3) compares to uni-modal null model. We found that a few regions are better than the null. Most notably, they are all low-hierarchy ones (eg, VISp). 8/n
What are the computational implications of categorical clusters? Intuitively, clusters reduce the dimensionality of the data (correlations). This constrains the geometry in the activity space since PR(X) = PR(X^T), limiting the flexibility typical of high-dim representations. 9/n
We studied the relation between clusters and geometry in a mathematical model where participation ratio (PR), a measure of dimensionality, can be computed analytically from Gaussian clusters - PR depends on # conditions (M), # clusters (k), and cluster quality. 10/n
Importantly, M is the number of independent conditions: those that are discriminable from each other in the neural activity. We developed an iterative algorithm to isolate the independent conditions in the data - finding that cognitive regions encode more conditions than sensory ones. 11/n
Using this M we were able to verify our theory, which accurately and quantitatively predicted the relationship between clustering and dimensionality. Importantly, clustering and dimensionality are inversely correlated in the data, with PR increasing along the hierarchy. 12/n
Finally, we computed the Shattering Dimensionality (SD) - a measure of coding flexibility (fraction of linearly solvable classification problems on conditions in the activity space). When considering independent conditions, SD was maximal in all areas, including sensory ones! 13/n
While large-scale analyses sound relatively simple on paper (one hammer many nails), it is incredibly complex to organize so much data in a coherent way. This was a big effort from an amazing team: big thanks to Shuqi Wang (co-1st), Sam Muscinelli, Liam Paninski,
@stefanofusi.bsky.social! /fin
And, of course, a shoutout to
@intlbrainlab.bsky.social and
@alleninstitute.bsky.social for making their data available and making this project possible. They are world-leading examples of what open science should look like, and I’m sure there are many works like ours to come!
Make sure to check out our new version of the preprint
doi.org/10.1101/2024... as well as the wonderful article from Holly Barker on The Transmitter, where we were interviewed us on this very issue:
www.thetransmitter.org/neural-codin...
Peace out ✌️
Most neurons in mouse cortex defy functional categories
The majority of cells in the cerebral cortex are unspecialized, according to an unpublished analysis—and scientists need to take care in naming neurons, the researchers warn.
