Connectivity Structure and Dynamics of Nonlinear Recurrent Neural Networks

Area of Science:

• Computational Neuroscience
• Network Science
• Systems Neuroscience

Background:

• Recurrent neural networks (RNNs) often assume simple, independent connections.
• Real neural circuits display complex connectivity, including structured singular values and vectors.
• This structure deviates from standard assumptions and may significantly impact network dynamics.

Purpose of the Study:

• To develop a theoretical framework for analyzing the impact of structured connectivity on RNN dynamics.
• To understand how biological connectivity properties shape high-dimensional collective activity.
• To identify key parameters governing network activity dimensions and timescales.

Main Methods:

• Introduction of the random-mode model, a random-matrix ensemble.
• Utilizing singular-value decomposition to control spectral properties and mode overlaps.
• Application of a novel path-integral calculation for deriving analytical expressions.

Main Results:

• Connectivity structure significantly shapes collective activity, even when invisible at the single-neuron level.
• The dimension of activity is determined by coupling variance and effective rank of the coupling matrix.
• Structured overlaps in the Drosophila connectome were identified and incorporated, further influencing dynamics.

Conclusions:

• Biological connectivity structure is crucial for understanding RNN dynamics.
• The random-mode model provides a powerful tool for analyzing complex neural networks.
• Future research can leverage these findings to better interpret large-scale neural data.