Extending the Functional Subnetwork Approach to a Generalized Linear Integrate-and-Fire Neuron Model.

Area of Science:

• Computational neuroscience
• Neural engineering
• Neuromorphic computing

Background:

• Engineering neural networks for specific tasks is challenging due to complex architecture and parameter tuning.
• Existing methods often rely on global optimization or machine learning, which can be computationally intensive.
• The Functional Subnetwork Approach (FSA) was previously developed for tuning non-spiking neural networks.

Purpose of the Study:

• To extend the Functional Subnetwork Approach (FSA) for tuning networks of spiking neurons.
• To enable direct assembly and tuning of spiking neural networks based on intended function.
• To provide an analytical method for parameter tuning without global optimization or machine learning.

Main Methods:

• Demonstrated fundamental similarities between generalized linear integrate-and-fire (GLIF) neuron dynamics and non-spiking leaky integrator models.
• Derived analytical expressions showing functional parallels between spiking and non-spiking neuron/synapse properties.
• Applied the extended FSA to model a neuromuscular reflex pathway using both spiking and non-spiking components.

Main Results:

• Established analytical expressions for steady-state and transient spiking frequencies analogous to non-spiking neuron voltages.
• Showed that spiking synapse average conductance parallels non-spiking synapse conductance.
• Successfully modeled a neuromuscular reflex pathway, demonstrating that a single non-spiking neuron can represent average spiking neuron activity.

Conclusions:

• The extended FSA provides an analytical method for tuning spiking neural networks.
• This approach facilitates the construction of large-scale spiking neural networks, particularly for neuromorphic hardware.
• The method allows for models incorporating both spiking and non-spiking units, offering flexibility in network design.