Sparse low-order interaction network underlies a highly correlated and learnable neural population code.

Area of Science:

• Neuroscience
• Computational Neuroscience
• Information Theory

Background:

• Neural information processing relies on population activity patterns.
• Understanding complex neural codes is challenging due to high dimensionality and inter-neuron dependencies.
• Existing models struggle with large neuronal groups, especially in sensory systems like the retina.

Purpose of the Study:

• To investigate the structure of population neural codes in large neuronal groups.
• To determine if higher-order interactions are necessary for accurate neural coding.
• To explore the organizational principles of neural interaction networks.

Main Methods:

• Analysis of ~100 retinal neurons responding to natural stimuli.
• Development and application of a novel model to capture higher-order interactions.
• Investigation of interaction network properties, including sparsity, hierarchy, and modularity.

Main Results:

• Pairwise models become insufficient for large neuronal populations (~100 neurons).
• A sparse, low-order interaction network underlies population neural codes.
• Higher-order interactions can be learned effectively using a novel sparse modeling approach.
• The neural interaction network exhibits hierarchical and modular organization, suggesting scalability.

Conclusions:

• The neural code in large populations is characterized by sparse, low-order interactions.
• Hierarchical and modular organization facilitates scalability in neural coding.
• Learnability appears to be a fundamental feature of neural codes, enabling efficient information processing.