A Non-spiking Neuron Model With Dynamic Leak to Avoid Instability in Recurrent Networks

Area of Science:

• Neuroscience
• Computational Neuroscience
• Systems Neuroscience

Background:

• Recurrent neuronal circuitry is widespread in the brain, comprising both excitatory and inhibitory connections.
• Potential instability in recurrent networks can lead to unreliable internal representations and predispose to conditions like epilepsy.
• Understanding the inherent stability of these networks is crucial for comprehending brain function.

Purpose of the Study:

• To assess the inherent stability properties of recurrent neuronal networks.
• To investigate the role of a "dynamic leak" in stabilizing network activity.
• To determine if dynamic leak improves fidelity and prevents spurious signal generation.

Main Methods:

• Utilized a linear summation, non-spiking neuron model with and without a dynamic leak.
• Compared neuron model fidelity to spiking neuron models across various input frequencies.
• Constructed fully connected recurrent networks with random synaptic weights and drove them with pseudorandom sensory inputs.
• Analyzed the impact of dynamic leak on network activity, stability, and spurious signal generation.

Main Results:

• The non-spiking neuron model with dynamic leak demonstrated higher input-output fidelity than spiking models.
• Recurrent networks without dynamic leak generated spurious high-frequency self-amplifying components.
• The addition of dynamic leak consistently eliminated these spurious high-frequency signals across networks.
• Network stability with dynamic leak scaled effectively with network size, conduction delays, and input density.

Conclusions:

• Neuronal dynamic leak acts as a low-pass filter, removing spurious high-frequency noise from synaptic inputs.
• This filtering mechanism protects recurrent neuronal circuitry from self-induced instability.
• Dynamic leak enables the brain to utilize recurrent circuitry robustly across diverse network sizes and configurations.