Sparse model-based estimation of functional dependence in high-dimensional field and spike multiscale networks

Area of Science:

• Neuroscience
• Computational Neuroscience
• Machine Learning

Background:

• Brain activity is encoded across multiple scales, from neuronal spikes to local field potentials (LFPs).
• Modeling these multiscale networks requires understanding complex spike-field dependencies.
• High-dimensional data makes traditional learning methods prone to overfitting.

Purpose of the Study:

• To develop a sparse model-based estimation algorithm for learning multiscale network dependencies.
• To address challenges in learning spike-field dependencies in high-dimensional neural recordings.
• To create more accurate multiscale encoding models of brain activity.

Main Methods:

• Developed a multiscale encoding model using a point process for neuronal spikes, with firing rates dependent on LFP network features and behavior.
• Formulated a constrained optimization problem with an L1 penalty to learn spike-field dependencies, maximizing likelihood.
• Applied the Akaike Information Criterion (AIC) to enforce sparsity in dependency parameters.

Main Results:

• The sparse algorithm improved spike prediction accuracy compared to models without dependencies.
• Fewer dependency parameters were identified compared to standard methods, reducing spurious dependencies.
• Incorporating LFP features from all electrodes enhanced spike prediction on individual electrodes.
• The algorithm revealed distance, brain region, and frequency band-dependent spike-field network patterns.

Conclusions:

• The developed algorithm effectively studies functional dependencies in high-dimensional spike-field networks.
• This approach leads to more accurate and robust multiscale encoding models.
• The method offers a powerful tool for analyzing complex neural data.