Constructing functional models from biophysically-detailed neurons.

Area of Science:

• Computational neuroscience
• Neural engineering
• Cognitive modeling

Background:

• Brain models aim to bridge low-level neural details with high-level behaviors.
• Improving biological plausibility and functional capacity are key goals in neural modeling.

Purpose of the Study:

• To develop a method for training biologically-detailed spiking neural networks (SNNs) that realize cognitively relevant dynamical systems.
• To investigate the relationship between biophysical mechanisms and functional capabilities in neural models.

Main Methods:

• Introduced the oracle-supervised Neural Engineering Framework (osNEF) for training SNNs.
• Utilized four distinct neuron models (LIF, Izhikevich, nonlinear point, pyramidal cell reconstructions) and various synaptic models.
• Trained networks to perform cognitive tasks including communication, multiplication, harmonic oscillation, and gated working memory.

Main Results:

• osNEF networks successfully exhibited target dynamics, accounting for neuron model nonlinearities.
• Performance was comparable across different neuron and synaptic models, with variance related to complexity.
• A working memory model demonstrated performance and forgetting rates consistent with animal data from delayed match-to-sample tasks (DMTST).

Conclusions:

• osNEF enables the training of functional brain models using biologically-detailed components.
• The framework opens new research avenues for exploring the link between neural biophysics and cognitive function.
• This approach advances the development of more realistic and capable computational neuroscience models.