Probing the structure-function relationship with neural networks constructed by solving a system of linear equations.

Area of Science:

• Computational neuroscience
• Systems neuroscience
• Neural modeling

Background:

• Neural network models are crucial for understanding brain function, bridging cellular, circuit, and behavioral levels.
• Traditional models rely on fitting numerous parameters using optimization algorithms to match experimental data.
• Existing methods face challenges in separating dynamic features from structural and algorithmic aspects.

Purpose of the Study:

• To propose an inverse method for constructing neural network models by inferring parameters from network dynamics.
• To decouple dynamical properties (firing rate, correlation) from structural connectivity and task-solving algorithms.
• To offer a new approach for investigating the relationship between neural network structure and function.

Main Methods:

• Inverting the fitting process: deriving network parameters from dynamics instead of fitting parameters to data.
• Utilizing firing state transitions and network dynamics to construct a system of linear equations.
• Ensuring consistency by relating network firing states and membrane potentials.
• Applying the method to a sequence memory task to analyze structure-function relationships.

Main Results:

• Successfully generated neural networks by deriving parameters from dynamics.
• The resulting networks exhibited connectivity and firing statistics consistent with experimental findings.
• Demonstrated the ability to uncouple dynamic and structural features of the model.
• Provided insights into the structure-function relationship within the context of a sequence memory task.

Conclusions:

• The proposed inverse method offers a complementary approach to traditional neural network construction for brain modeling.
• This technique facilitates a clearer understanding of how neural network structure relates to its function.
• The method successfully recapitulated experimental observations in a sequence memory task.
• This approach is valuable for advancing computational neuroscience and neural modeling.