Emergence of functionally differentiated structures via mutual information minimization in recurrent neural networks

Area of Science:

• Computational Neuroscience
• Artificial Intelligence
• Information Theory

Background:

• Brain function relies on specialized regions, a process modeled by artificial neural networks.
• Understanding functional differentiation is key to complex system analysis.

Purpose of the Study:

• To develop a novel method for inducing functional differentiation in recurrent neural networks.
• To investigate the relationship between functional and structural modularity during network development.

Main Methods:

• Utilized mutual information neural estimation to minimize information between neural subgroups.
• Applied the method to working memory and chaotic signal separation tasks.
• Analyzed network performance, correlation patterns, and synaptic weight matrices.

Main Results:

• Mutual information minimization led to high task performance and clear functional modularity.
• Functional differentiation emerged earlier than structural modularity.
• Results suggest functional specialization drives structural reorganization.

Conclusions:

• Information-theoretic principles may govern the emergence of specialized functions and modular structures.
• Functional differentiation precedes structural changes in developing neural networks.
• The proposed method offers insights into artificial and biological brain development.