Estimation of neural network model parameters from local field potentials (LFPs).

Area of Science:

• Systems Neuroscience
• Computational Neuroscience
• Computational Neuroscience Modeling

Background:

• Traditional systems neuroscience models are often descriptive, correlating neural activity with input.
• Mechanistic models, common in physics, require experimental validation, typically using neuronal spikes.
• Validating mechanistic cortical network models has historically relied on high-frequency neuronal spike data.

Purpose of the Study:

• Investigate the utility of the low-frequency local field potential (LFP) signal for validating mechanistic cortical network models.
• Determine if LFPs can accurately estimate synaptic connection weights within neural networks.
• Assess the potential of LFPs for inferring properties of computational neuroscience models.

Main Methods:

• Utilized a Brunel network model with excitatory and inhibitory integrate-and-fire neurons.
• Computed network-generated LFPs via a hybrid scheme, replaying spikes onto multicompartmental neurons.
• Applied convolutional neural networks (CNNs) to analyze LFP power spectra for parameter estimation.

Main Results:

• All three key network parameters were accurately estimated from stationary LFP signals.
• Convolutional neural networks demonstrated high precision in parameter inference from LFP power spectra.
• The findings indicate a strong correlation between LFP characteristics and underlying network parameters.

Conclusions:

• Local field potentials (LFPs) are a viable signal for validating mechanistic cortical network models.
• LFP analysis, particularly using CNNs, can reliably estimate synaptic connection weights.
• This approach offers a promising alternative to spike-based validation in computational neuroscience.