[Paper Review] On the ground state of spiking network activity in mammalian cortex

Jens Wilting, Viola Priesemann|arXiv (Cornell University)|Apr 20, 2018

Neural dynamics and brain function60 references3 citations

TL;DR

This study resolves a long-standing debate in cortical network dynamics by introducing a subsampling-invariant estimator that reveals in vivo spiking activity in mammalian cortex consistently occupies a unique, narrow 'reverberating state'—distinct from both asynchronous irregular and critical states. The authors validate a generic model tuned to this state, enabling accurate predictions of single-neuron, pairwise, and population-level properties, and infer previously inaccessible network parameters such as timescale and input strength.

ABSTRACT

Electrophysiological recordings of spiking activity are limited to a small number of neurons. This spatial subsampling has hindered characterizing even most basic properties of collective spiking in cortical networks. In particular, two contradictory hypotheses prevailed for over a decade: the first proposed an asynchronous irregular state, the second a critical state. While distinguishing them is straightforward in models, we show that in experiments classical approaches fail to correctly infer network dynamics because of subsampling. Deploying a novel, subsampling-invariant estimator, we find that in vivo dynamics do not comply with either hypothesis, but instead occupy a narrow state consistently across multiple mammalian species and cortical areas. A generic model tuned to this reverberating state predicts single neuron, pairwise, and population properties. With these predictions we first validate the model and then deduce network properties that are challenging to obtain experimentally, like the network timescale and strength of cortical input.

Motivation & Objective

• To resolve the longstanding contradiction between the asynchronous irregular and critical state hypotheses in cortical network dynamics.
• To overcome the limitations of classical inference methods that fail under spatial subsampling in electrophysiological recordings.
• To identify a universal dynamical state governing spiking activity across diverse mammalian cortical areas and species.
• To develop a model calibrated to this state that accurately predicts single-neuron, pairwise, and population-level activity statistics.
• To infer experimentally inaccessible network properties such as network timescale and strength of cortical input.

Proposed method

• Development of a novel subsampling-invariant estimator to reliably infer network dynamics despite limited recording of neurons.
• Application of the estimator to in vivo electrophysiological data from multiple mammalian species and cortical areas.
• Parameter tuning of a generic spiking network model to match the empirically inferred dynamics, yielding predictions across multiple scales.
• Validation of model predictions against observed single-neuron, pairwise, and population-level statistics from experimental data.
• Use of model predictions to infer hidden network properties such as timescale and input strength, which are difficult to measure directly.

Experimental results

Research questions

• RQ1 What dynamical state do in vivo cortical networks actually occupy, given the limitations of spatial subsampling in recordings?
• RQ2 How can we reliably infer network dynamics when only a small fraction of neurons are recorded?
• RQ3 Does the observed activity conform to either the asynchronous irregular or critical state hypotheses, or does it represent a distinct state?
• RQ4 Can a generic model calibrated to the empirically inferred state accurately predict single-neuron, pairwise, and population-level activity statistics?
• RQ5 What are the underlying network properties—such as timescale and input strength—that govern this state, and can they be inferred from data?

Key findings

• The in vivo dynamics of cortical networks across multiple mammalian species and cortical areas consistently occupy a unique, narrow reverberating state, distinct from both asynchronous irregular and critical states.
• Classical inference methods fail under subsampling, leading to incorrect classification of network dynamics, which explains the long-standing controversy.
• A generic spiking network model tuned to the reverberating state accurately predicts single-neuron firing rates, pairwise correlations, and population activity patterns.
• The model's predictions are validated against empirical data, confirming the robustness and accuracy of the inferred state.
• The model enables inference of previously inaccessible network properties, including a network timescale and strength of cortical input, which are challenging to measure directly in experiments.