Efficient and robust coding in heterogeneous recurrent networks.

Area of Science:

• Computational Neuroscience
• Neural Coding
• Systems Neuroscience

Background:

• Cortical networks exhibit significant neuronal property heterogeneity.
• Traditional coding models often assume homogeneous neuronal populations (excitatory and inhibitory).

Purpose of the Study:

• To analytically derive a class of recurrent spiking neural networks for optimal online tracking of continuous inputs.
• To investigate the functional differences between homogeneous and heterogeneous neuronal networks within a predictive coding framework.

Main Methods:

• Derivation of a recurrent network model based on linear spike decoding and mean-squared error minimization.
• Development of a predictive coding framework unifying encoding and decoding processes.
• Analysis of network performance with homogeneous versus heterogeneous populations of neurons with diverse properties.

Main Results:

• Heterogeneous networks, incorporating 'type 1' and 'type 2' neurons, arise naturally from the derived framework.
• Networks composed of heterogeneous neuronal populations demonstrate increased efficiency and robustness against correlated noise.
• Predicted distinct correlation patterns between neuron types (integrators, resonators) and differential coherence of 'type 2' neurons with network activity.

Conclusions:

• Heterogeneous neuronal populations are crucial for efficient and robust information processing in cortical networks.
• The predictive coding framework provides a unified approach to understanding neural encoding and decoding.
• The study offers testable predictions for experimental validation of neuronal heterogeneity's role in network function.