Emergent Rate-Based Dynamics in Duplicate-Free Populations of Spiking Neurons

Area of Science:

• Computational neuroscience
• Artificial intelligence
• Machine learning

Background:

• Spiking neural networks (SNNs) are biologically inspired computational models.
• Recurrent neural networks (RNNs) are widely used for sequential data processing.
• Classical mean-field theory suggests SNNs approximate RNNs only with neuron duplicates.

Purpose of the Study:

• To investigate if SNNs can approximate RNN dynamics without relying on neuron duplicates.
• To explore the underlying mechanisms enabling such convergence in large SNNs.

Main Methods:

• Utilized a disordered network model to ensure the absence of duplicate neurons.
• Applied principles of the concentration of measure phenomenon.
• Analyzed the convergence of SNNs to RNNs in a duplicate-free scenario.

Main Results:

• Demonstrated that duplicate-free SNNs can converge to RNN dynamics.
• Showcased the concentration of measure phenomenon as a key mechanism for this convergence.
• Established that large SNNs can exhibit rate-based dynamics without neuron redundancy.

Conclusions:

• The study reveals a general mechanism for the emergence of rate-based dynamics in large SNNs.
• SNNs can effectively approximate RNNs even in the absence of highly correlated neuronal inputs.
• This finding broadens the understanding of SNN capabilities and their relationship to traditional neural networks.