Multitask learning of a biophysically-detailed neuron model.

Area of Science:

• Computational Neuroscience
• Artificial Intelligence in Neuroscience
• Neural Circuit Simulation

Background:

• Understanding the human brain requires integrated research across multiple biological scales.
• Biophysically-detailed neuron models are crucial for studying neural circuits but are computationally intensive.
• Existing artificial neural networks (ANNs) can predict some neuron behaviors but are limited to specific compartments.

Purpose of the Study:

• To develop a novel method for predicting membrane potentials in all compartments of biophysically-detailed neuron models.
• To accelerate neuron model simulations beyond current ANN capabilities.
• To enable broader comparisons with experimental data and pave the way for predicting brain activity signals.

Main Methods:

• Utilized enhanced state-of-the-art multitask learning (MTL) architectures.
• Trained ANNs to simultaneously predict membrane potentials for every neuron compartment.
• Developed a challenging benchmark for MTL due to large datasets and complex data correlations.

Main Results:

• Achieved simulation speeds up to two orders of magnitude faster than classical methods.
• Enabled simultaneous prediction of membrane potentials across all neuron compartments.
• Provided a foundation for predicting local field potentials (LFPs) and electroencephalogram (EEG) signals.

Conclusions:

• The novel MTL approach significantly accelerates the simulation of detailed neuron models.
• Predicting all dendritic voltages captures neuron electrophysiology comprehensively.
• This work facilitates comparison with diverse experimental recordings and advances large-scale neural simulations.