Strong and localized recurrence controls dimensionality of neural activity across brain areas

Abstract

The dimensionality of a network's collective activity is the number of modes into which it is organized. This quantity is of great interest in neural coding: small dimensionality suggests a compressed neural code and possibly high robustness and generalizability, while high dimensionality suggests expansion of input features to enable flexible downstream computation. Here, for recurrent neural circuits operating in the ubiquitous balanced regime, we show how dimensionality arises mechanistically via perhaps the most basic property of neural circuits: a single number characterizing the net strength of their connectivity. Our results combine novel theoretical approaches with new analyses of high-density neuropixels recordings and high-throughput synaptic physiology datasets. The analysis of electrophysiological recordings identifies bounds on the dimensionality of neural responses across brain regions, showing that it is on the order of hundreds -- striking a balance between high and low-dimensional codes. Furthermore, focusing on the visual stream, we show that dimensionality expands from primary to deeper visual areas and similarly within an area from layer 2/3 to layer 5. We interpret these results via a novel theoretical result which links dimensionality to a single measure of net connectivity strength. This requires calculations that extend beyond traditional mean-field approaches to neural networks. Our result suggests that areas across the brain operate in a strongly coupled regime where dimensionality is under sensitive control by net connectivity strength; moreover, we show how this net connectivity strength is regulated by local connectivity features, or synaptic motifs. This enables us to interpret changes in dimensionality in terms of changes in coupling among pairs and triplets of neurons. Analysis of large-scale synaptic physiology datasets from both mouse and human cortex then reveal the presence of synaptic coupling motifs capable of substantially regulating this dimensionality.

Perception and propagation of activity through the cortical hierarchy is determined by neural variability

“…5e-h, Extended Data Fig. 10) , and found that neural dynamics in S1 are strongly recurrent ( R 0.964 ± 0.008), similar to other cortical areas in rodents 36, 39, 44 . Importantly, we also observed that lower recurrence R prior to stimulation predicts higher detection probability (p 0.002, Fig.…”

“…Population variance is one measure of neural variability, which can arise from recurrent neural activity 36 . We quantified the recurrence R in order to test theories 65, 67, 68 about how it impacts signal propagation.…”

“…5a,b, Extended Data Fig. 10d-e) as this likely arose from connections between recorded neurons 13, 14, 16, 36 . We explored two plausible circuit mechanisms that could give rise to the previously observed increased variability prior to miss trials.…”

“…5c,d, p 0.365, Wilcoxon signed-rank test). Second, neural circuits are characterised by strong recurrence 38, 39 , which is known to give rise to trial-to-trial variability 40 ; moreover, the strength of recurrence can determine local amplification and encoding of external stimuli 36, 41–43 . We measured the strength of recurrence in S1 ( R ) as previously described 36 (Fig.…”

Local Connectivity and Synaptic Dynamics in Mouse and Human Neocortex

“…We quantified the ratio of connected pairs with and without bidirectional connections and observed that reciprocal connections were 3 to 5 times more common than expected for a randomly connected network (red line) in E→E, E→I, I→E, and I→I connections (Fig 2A, bottom row). Further analysis of higher order connectivity motifs in our data and their impact on cortical computation is pursued in a parallel study (Dahmen et al, 2020).…”